Sync from GitHub ncclab-sustech/DCA main

Source: https://github.com/ncclab-sustech/DCA@3b717c902b8ff8b153293e441dfac2f716d254a0

.gitattributes

@@ -0,0 +1,10 @@
+DCA/data/fmri/101915_FWHM6.nii.gz filter=lfs diff=lfs merge=lfs -text
+DCA/data/fmri/101915_raw.nii.gz filter=lfs diff=lfs merge=lfs -text
+AtlaScore/downstream/fc_data/fc_data.zip filter=lfs diff=lfs merge=lfs -text
+DCA/data/swin_model_epoch_30.pth filter=lfs diff=lfs merge=lfs -text
+DCA/data/fmri/101915_FWHM3.nii.gz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1 @@
+__pycache__/

AtlaScore/downstream/demo.ipynb

@@ -0,0 +1,269 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "71ead501",
+   "metadata": {},
+   "source": [
+    "This notebook demonstrates how to run AtlaScore downstream tasks on the proposed **DCA100 atlas**.\n",
+    "\n",
+    "For DCA100, we provide precomputed functional connectivity (FC) features required to run the downstream evaluations. You can directly execute the section *run AtlaScore downstream tasks* to evaluate the performance of DCA100.\n",
+    "\n",
+    "For other atlases, you will need to first download the required NIfTI data, extract the region-wise time series, and compute FC features. These steps are provided in the *get nii data* and *get FC data* sections.\n",
+    "\n",
+    "Note: Preprocessed data from the ADNI dataset will be shared via OneDrive after the paper is accepted. The corresponding download link and notebook updates will be made available at that time. Until then, attempting to run `get_fc_ADNI` or `AD_diagnosis` for other atlases will result in errors (as the required data is not yet accessible)."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "cb56d181",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import numpy as np\n",
+    "import downstream\n",
+    "import zipfile"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "79e7c8c5",
+   "metadata": {},
+   "source": [
+    "#### get nii data"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "3eda46a0",
+   "metadata": {},
+   "source": [
+    "##### HCP"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "785975f0",
+   "metadata": {},
+   "source": [
+    "To access and download Human Connectome Project (HCP) data locally for use with AtlaScore, you must first obtain the necessary credentials. Visit the official HCP website at [https://db.humanconnectome.org](https://db.humanconnectome.org), create an account, and request data access. Once your request is approved, you will be able to retrieve your `access_key` and `secret_key` from your account.\n",
+    "\n",
+    "These credentials are required to authenticate and download data using the provided scripts. Please insert your keys into the corresponding variables in the notebook.\n",
+    "\n",
+    "Downloading the required HCP data typically takes around 15 hours and requires over 500 GB of disk space."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5d2cd2d1",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# access_key = 'your_access_key'\n",
+    "# secret_key = 'your_secret_key'\n",
+    "\n",
+    "# for subj in np.loadtxt('./docs/HCP_subjlist.txt', dtype = str):\n",
+    "\n",
+    "#     downstream.get_sub_HCP_rfMRI(subject = subj, access_key = access_key, secret_key = secret_key)\n",
+    "#     downstream.get_sub_HCP_tfMRI(subject = subj, access_key = access_key, secret_key = secret_key)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "13c96145",
+   "metadata": {},
+   "source": [
+    "##### ABIDE"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "375a6ff8",
+   "metadata": {},
+   "source": [
+    "To use the ABIDE dataset for functional connectivity analysis, you will need to download the preprocessed data. The dataset includes resting-state fMRI scans from multiple sites and subjects.\n",
+    "\n",
+    "Please note that downloading the full ABIDE dataset may take up to 9 hours and requires approximately 90 GB of disk space. Make sure you have sufficient bandwidth and storage before starting the download."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "2924faf6",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# downstream.get_ABIDE()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "fe6a85a6",
+   "metadata": {},
+   "source": [
+    "##### ADNI"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "8a427acd",
+   "metadata": {},
+   "source": [
+    "The ADNI dataset contains multimodal neuroimaging data, including structural and functional MRI. For our analysis, we used preprocessed data derived from ADNI scans.\n",
+    "\n",
+    "The total size of the processed dataset is approximately 70 GB. This preprocessed version will be made available for download via OneDrive after the paper is accepted for publication."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "df3d8336",
+   "metadata": {},
+   "source": [
+    "#### get FC data"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "c12ae2e8",
+   "metadata": {},
+   "source": [
+    "##### HCP"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "45082e6b",
+   "metadata": {},
+   "source": [
+    "To run the analysis on your own atlas, please modify the `atlas_name` and `atlas_loc` arguments in the corresponding functions."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "051d4cbb",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# for subj in np.loadtxt('./docs/HCP_subjlist.txt', dtype = str):\n",
+    "\n",
+    "#     downstream.get_fc_HCP_rfMRI(subject = subj, atlas_name = 'DCA100', atlas_loc = './docs/DCA100.nii.gz')\n",
+    "#     downstream.get_fc_HCP_tfMRI(subject = subj, atlas_name = 'DCA100', atlas_loc = './docs/DCA100.nii.gz')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "5f831e97",
+   "metadata": {},
+   "source": [
+    "##### ABIDE"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "156aeebf",
+   "metadata": {},
+   "source": [
+    "To run the analysis on your own atlas, please modify the `atlas_name` and `atlas_loc` arguments in the corresponding functions."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b568852b",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# downstream.get_fc_ABIDE(atlas_name = 'DCA100', atlas_loc = './docs/DCA100.nii.gz')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "0b5eeb0c",
+   "metadata": {},
+   "source": [
+    "##### ADNI"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "021a9c9c",
+   "metadata": {},
+   "source": [
+    "To run the analysis on your own atlas, please modify the `atlas_name` and `atlas_loc` arguments in the corresponding functions."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4cf92856",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# downstream.get_fc_ADNI(atlas_name = 'DCA100', atlas_loc = './docs/DCA100.nii.gz')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "b4faef6a",
+   "metadata": {},
+   "source": [
+    "#### run AtlaScore downstream tasks"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "835bd40c",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "--- DCA100 downstream report ---\n",
+      "Gender classification: 0.666±0.080\n",
+      "Fluid intelligence: 0.491±0.082\n",
+      "Cognitive task (7-way): 0.869±0.062\n",
+      "Cognitive task (24-way): 0.452±0.030\n",
+      "Autism diagnosis: 0.655±0.054\n",
+      "AD diagnosis: 0.387±0.077\n",
+      "FC stability: 0.650±0.045\n",
+      "Fingerprinting: 0.696±0.201\n",
+      "Age group classification: 0.452±0.136\n",
+      "Crystallized intelligence: 0.472±0.095\n",
+      "General intelligence: 0.442±0.104\n",
+      "Autism cross-site: 0.662±0.068\n"
+     ]
+    }
+   ],
+   "source": [
+    "with zipfile.ZipFile('./fc_data/fc_data.zip', 'r') as zip_file: zip_file.extractall('./fc_data/')\n",
+    "downstream.downstream_all(atlas_name = 'DCA100')"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "connectome",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.2"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

AtlaScore/downstream/docs/DCA100.nii.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55cb73621697360e6107394f51fc0606367c1c43a88d96c8693c0d810d91d8d1
+size 132151

AtlaScore/downstream/docs/HCP_subjlist.txt

@@ -0,0 +1,100 @@
+100206
+100307
+100408
+100610
+101006
+101107
+101309
+101915
+102008
+102109
+102311
+102513
+102614
+102715
+102816
+103010
+103111
+103212
+103414
+103515
+103818
+104012
+104416
+104820
+105014
+105115
+105216
+105620
+105923
+106016
+106319
+106521
+106824
+107018
+107321
+107422
+107725
+108020
+108121
+108222
+108323
+108525
+108828
+109123
+109325
+109830
+110007
+110411
+110613
+111009
+111211
+111312
+111413
+111514
+111716
+112112
+112314
+112516
+112920
+113215
+113316
+113619
+113922
+114217
+114318
+114419
+114621
+114823
+114924
+115017
+115219
+115320
+115724
+115825
+116524
+116726
+117021
+117122
+117324
+117930
+118023
+118124
+118225
+118528
+118730
+118831
+118932
+119025
+119126
+119732
+119833
+120111
+120212
+120414
+120515
+120717
+121416
+121618
+121921
+122317

AtlaScore/downstream/docs/behavior_ABIDE.csv

@@ -0,0 +1,872 @@
+SUB_ID,SITE_ID,FILE_ID,DX_GROUP,DSM_IV_TR,AGE_AT_SCAN,SEX,HANDEDNESS_CATEGORY
+50003,PITT,Pitt_0050003,1,1,24.45,1,R
+50004,PITT,Pitt_0050004,1,1,19.09,1,R
+50005,PITT,Pitt_0050005,1,1,13.73,2,R
+50006,PITT,Pitt_0050006,1,1,13.37,1,L
+50007,PITT,Pitt_0050007,1,1,17.78,1,R
+50008,PITT,Pitt_0050008,1,1,32.45,1,R
+50010,PITT,Pitt_0050010,1,1,35.2,1,L
+50011,PITT,Pitt_0050011,1,1,16.93,1,L
+50012,PITT,Pitt_0050012,1,1,21.48,1,R
+50013,PITT,Pitt_0050013,1,1,9.33,1,R
+50014,PITT,Pitt_0050014,1,1,14.2,1,R
+50015,PITT,Pitt_0050015,1,1,14.2,1,R
+50016,PITT,Pitt_0050016,1,1,21.82,1,R
+50020,PITT,Pitt_0050020,1,1,20.83,1,R
+50022,PITT,Pitt_0050022,1,1,16.99,1,R
+50023,PITT,Pitt_0050023,1,1,12.64,2,R
+50024,PITT,Pitt_0050024,1,1,22.64,1,R
+50025,PITT,Pitt_0050025,1,1,31.72,1,R
+50026,PITT,Pitt_0050026,1,1,16.28,1,R
+50027,PITT,Pitt_0050027,1,1,12.24,1,R
+50028,PITT,Pitt_0050028,1,1,12.74,1,R
+50030,PITT,Pitt_0050030,2,0,25.12,1,-9999
+50031,PITT,Pitt_0050031,2,0,12.92,1,R
+50032,PITT,Pitt_0050032,2,0,19.8,1,R
+50033,PITT,Pitt_0050033,2,0,12.15,1,R
+50034,PITT,Pitt_0050034,2,0,14.77,1,R
+50035,PITT,Pitt_0050035,2,0,17.36,1,R
+50036,PITT,Pitt_0050036,2,0,13.28,2,R
+50037,PITT,Pitt_0050037,2,0,19.16,1,R
+50038,PITT,Pitt_0050038,2,0,13.95,2,R
+50039,PITT,Pitt_0050039,2,0,30.66,1,R
+50040,PITT,Pitt_0050040,2,0,24.6,1,R
+50041,PITT,Pitt_0050041,2,0,27.92,1,R
+50042,PITT,Pitt_0050042,2,0,33.17,1,R
+50043,PITT,Pitt_0050043,2,0,13.78,1,R
+50044,PITT,Pitt_0050044,2,0,17.13,1,R
+50045,PITT,Pitt_0050045,2,0,15.7,1,R
+50046,PITT,Pitt_0050046,2,0,20.65,1,R
+50047,PITT,Pitt_0050047,2,0,15.35,1,R
+50048,PITT,Pitt_0050048,2,0,11.81,1,R
+50049,PITT,Pitt_0050049,2,0,15.82,2,R
+50050,PITT,Pitt_0050050,2,0,14.37,1,L
+50051,PITT,Pitt_0050051,2,0,12.83,1,R
+50052,PITT,Pitt_0050052,2,0,33.24,1,R
+50053,PITT,Pitt_0050053,1,1,12.03,1,R
+50054,PITT,Pitt_0050054,2,0,9.44,1,R
+50056,PITT,Pitt_0050056,1,1,13.59,1,R
+50057,PITT,Pitt_0050057,1,1,12.62,2,R
+50059,PITT,Pitt_0050059,2,0,21.22,2,R
+50060,PITT,Pitt_0050060,2,0,20.34,1,R
+50102,OLIN,Olin_0050102,2,0,14,1,R
+50103,OLIN,Olin_0050103,2,0,14,1,R
+50104,OLIN,Olin_0050104,2,0,16,1,R
+50105,OLIN,Olin_0050105,2,0,17,1,R
+50106,OLIN,Olin_0050106,2,0,10,1,R
+50107,OLIN,Olin_0050107,2,0,21,1,R
+50109,OLIN,Olin_0050109,2,0,20,1,R
+50111,OLIN,Olin_0050111,2,0,14,1,R
+50112,OLIN,Olin_0050112,2,0,17,1,R
+50113,OLIN,Olin_0050113,2,0,15,2,R
+50114,OLIN,Olin_0050114,2,0,20,2,R
+50115,OLIN,Olin_0050115,2,0,23,1,L
+50116,OLIN,Olin_0050116,2,0,15,1,R
+50117,OLIN,Olin_0050117,2,0,21,1,L
+50118,OLIN,Olin_0050118,1,0,15,1,R
+50119,OLIN,Olin_0050119,1,0,19,2,R
+50121,OLIN,Olin_0050121,1,0,20,1,R
+50123,OLIN,Olin_0050123,1,0,17,1,R
+50124,OLIN,Olin_0050124,1,0,18,1,R
+50125,OLIN,Olin_0050125,1,0,24,2,R
+50127,OLIN,Olin_0050127,1,0,14,2,R
+50128,OLIN,Olin_0050128,1,0,16,1,R
+50129,OLIN,Olin_0050129,1,0,12,1,R
+50130,OLIN,Olin_0050130,1,0,18,1,R
+50131,OLIN,Olin_0050131,1,0,21,1,L
+50132,OLIN,Olin_0050132,1,0,16,1,R
+50134,OLIN,Olin_0050134,1,0,18,1,L
+50135,OLIN,Olin_0050135,1,0,12,1,L
+50142,OHSU,OHSU_0050142,1,-9999,13.99,1,L
+50143,OHSU,OHSU_0050143,1,-9999,13.79,1,R
+50144,OHSU,OHSU_0050144,1,-9999,10.22,1,R
+50145,OHSU,OHSU_0050145,1,-9999,10.75,1,R
+50146,OHSU,OHSU_0050146,1,-9999,8,1,R
+50147,OHSU,OHSU_0050147,1,-9999,11.35,1,R
+50148,OHSU,OHSU_0050148,1,-9999,12.65,1,R
+50149,OHSU,OHSU_0050149,1,-9999,12.32,1,R
+50150,OHSU,OHSU_0050150,1,-9999,9.42,1,R
+50152,OHSU,OHSU_0050152,1,-9999,9.73,1,R
+50153,OHSU,OHSU_0050153,1,-9999,9.72,1,R
+50156,OHSU,OHSU_0050156,1,-9999,15.23,1,R
+50157,OHSU,OHSU_0050157,2,-9999,10.5,1,R
+50158,OHSU,OHSU_0050158,2,-9999,11.55,1,R
+50159,OHSU,OHSU_0050159,2,-9999,10.29,1,R
+50160,OHSU,OHSU_0050160,2,-9999,11.45,1,R
+50161,OHSU,OHSU_0050161,2,-9999,9.15,1,R
+50162,OHSU,OHSU_0050162,2,-9999,8.94,1,R
+50163,OHSU,OHSU_0050163,2,-9999,9.4,1,R
+50164,OHSU,OHSU_0050164,2,-9999,8.86,1,R
+50167,OHSU,OHSU_0050167,2,-9999,10.08,1,R
+50168,OHSU,OHSU_0050168,2,-9999,9.69,1,R
+50169,OHSU,OHSU_0050169,2,-9999,11.99,1,R
+50170,OHSU,OHSU_0050170,2,-9999,10.53,1,R
+50171,OHSU,OHSU_0050171,2,-9999,10.64,1,R
+50182,SDSU,SDSU_0050182,1,2,16.61,1,R
+50183,SDSU,SDSU_0050183,1,1,14.14,1,R
+50184,SDSU,SDSU_0050184,1,2,17.15,1,R
+50186,SDSU,SDSU_0050186,1,2,12.13,1,R
+50187,SDSU,SDSU_0050187,1,2,16.48,1,R
+50188,SDSU,SDSU_0050188,1,-9999,15.4,1,R
+50189,SDSU,SDSU_0050189,1,2,16.64,1,R
+50190,SDSU,SDSU_0050190,1,2,13.99,1,R
+50193,SDSU,SDSU_0050193,2,0,14.75,1,R
+50194,SDSU,SDSU_0050194,2,0,13.35,1,L
+50195,SDSU,SDSU_0050195,2,0,12.58,1,R
+50196,SDSU,SDSU_0050196,2,0,12.91,1,R
+50198,SDSU,SDSU_0050198,2,0,15.28,1,R
+50199,SDSU,SDSU_0050199,2,0,15.53,1,R
+50200,SDSU,SDSU_0050200,2,0,13.88,1,R
+50201,SDSU,SDSU_0050201,2,0,13.77,2,R
+50202,SDSU,SDSU_0050202,2,0,12.08,1,R
+50203,SDSU,SDSU_0050203,2,0,16.05,1,L
+50204,SDSU,SDSU_0050204,2,0,14.86,1,R
+50205,SDSU,SDSU_0050205,2,0,14.1,2,R
+50206,SDSU,SDSU_0050206,2,0,8.67,2,R
+50208,SDSU,SDSU_0050208,2,0,13.78,2,R
+50210,SDSU,SDSU_0050210,2,0,16.6,1,R
+50213,SDSU,SDSU_0050213,2,0,12.38,2,R
+50214,SDSU,SDSU_0050214,2,0,14.76,1,R
+50215,SDSU,SDSU_0050215,2,0,16.88,2,R
+50217,SDSU,SDSU_0050217,2,0,12.9,1,R
+50232,TRINITY,Trinity_0050232,1,3,23.08,1,R
+50233,TRINITY,Trinity_0050233,1,3,20.75,1,R
+50234,TRINITY,Trinity_0050234,1,3,15.66,1,R
+50236,TRINITY,Trinity_0050236,1,3,19.16,1,R
+50237,TRINITY,Trinity_0050237,1,3,21.42,1,R
+50239,TRINITY,Trinity_0050239,1,1,17.03,1,R
+50240,TRINITY,Trinity_0050240,1,2,15.05,1,R
+50241,TRINITY,Trinity_0050241,1,1,17.17,1,R
+50243,TRINITY,Trinity_0050243,1,1,13.75,1,R
+50245,TRINITY,Trinity_0050245,1,1,21.58,1,R
+50247,TRINITY,Trinity_0050247,1,2,12,1,R
+50248,TRINITY,Trinity_0050248,1,1,13.58,1,R
+50249,TRINITY,Trinity_0050249,1,1,13.83,1,R
+50250,TRINITY,Trinity_0050250,1,1,13.91,1,R
+50251,TRINITY,Trinity_0050251,1,1,17.25,1,R
+50252,TRINITY,Trinity_0050252,1,2,15.25,1,R
+50253,TRINITY,Trinity_0050253,1,2,18.01,1,R
+50254,TRINITY,Trinity_0050254,1,2,19.41,1,R
+50255,TRINITY,Trinity_0050255,1,2,14.41,1,R
+50257,TRINITY,Trinity_0050257,2,0,14.91,1,R
+50259,TRINITY,Trinity_0050259,2,0,19.83,1,R
+50260,TRINITY,Trinity_0050260,2,0,20.33,1,R
+50261,TRINITY,Trinity_0050261,2,0,20.08,1,R
+50262,TRINITY,Trinity_0050262,2,0,21.33,1,R
+50263,TRINITY,Trinity_0050263,2,0,19.08,1,R
+50264,TRINITY,Trinity_0050264,2,0,25.66,1,R
+50265,TRINITY,Trinity_0050265,2,0,12.25,1,R
+50266,TRINITY,Trinity_0050266,2,0,15.91,1,R
+50267,TRINITY,Trinity_0050267,2,0,17.5,1,R
+50268,TRINITY,Trinity_0050268,2,0,14.41,1,R
+50269,TRINITY,Trinity_0050269,2,0,14.66,1,R
+50270,TRINITY,Trinity_0050270,2,0,19.33,1,R
+50271,TRINITY,Trinity_0050271,2,0,19.66,1,R
+50272,UM_1,UM_1_0050272,1,1,14.2,1,R
+50273,UM_1,UM_1_0050273,1,1,16.8,1,R
+50274,UM_1,UM_1_0050274,1,1,14.2,1,R
+50275,UM_1,UM_1_0050275,1,1,11.5,1,L
+50276,UM_1,UM_1_0050276,1,1,16.8,2,-9999
+50278,UM_1,UM_1_0050278,1,1,11.3,2,R
+50282,UM_1,UM_1_0050282,1,1,15.2,1,
+50284,UM_1,UM_1_0050284,1,2,11.2,2,R
+50285,UM_1,UM_1_0050285,1,1,12.8,2,R
+50287,UM_1,UM_1_0050287,1,1,14.4,1,L
+50289,UM_1,UM_1_0050289,1,2,12.9,1,L
+50290,UM_1,UM_1_0050290,1,1,14,1,R
+50291,UM_1,UM_1_0050291,1,1,13.9,1,R
+50292,UM_1,UM_1_0050292,1,1,12.6,1,R
+50293,UM_1,UM_1_0050293,1,1,12.4,1,R
+50294,UM_1,UM_1_0050294,1,2,12.6,1,R
+50295,UM_1,UM_1_0050295,1,1,11.5,1,R
+50297,UM_1,UM_1_0050297,1,-9999,15.9,1,R
+50298,UM_1,UM_1_0050298,1,1,12.8,1,R
+50300,UM_1,UM_1_0050300,1,3,9.7,2,R
+50301,UM_1,UM_1_0050301,1,1,16.1,1,L
+50302,UM_1,UM_1_0050302,1,-9999,10.4,2,R
+50304,UM_1,UM_1_0050304,1,1,11,1,-9999
+50308,UM_1,UM_1_0050308,1,1,8.9,1,-9999
+50310,UM_1,UM_1_0050310,1,1,9.7,1,R
+50312,UM_1,UM_1_0050312,1,1,11.2,1,R
+50314,UM_1,UM_1_0050314,1,2,16.1,1,R
+50315,UM_1,UM_1_0050315,1,1,13.4,1,R
+50318,UM_1,UM_1_0050318,1,1,9.2,1,L
+50319,UM_1,UM_1_0050319,1,1,15,2,R
+50320,UM_1,UM_1_0050320,1,1,18.6,1,R
+50321,UM_1,UM_1_0050321,1,1,18,2,R
+50324,UM_1,UM_1_0050324,1,1,13.8,1,R
+50325,UM_1,UM_1_0050325,1,2,13.1,1,R
+50327,UM_1,UM_1_0050327,2,0,15.3,1,L
+50329,UM_1,UM_1_0050329,2,0,17.1,1,R
+50330,UM_1,UM_1_0050330,2,0,16.1,1,R
+50331,UM_1,UM_1_0050331,2,0,17.2,1,L
+50332,UM_1,UM_1_0050332,2,0,10.4,1,R
+50333,UM_1,UM_1_0050333,2,0,10.2,1,R
+50334,UM_1,UM_1_0050334,2,0,11,1,R
+50335,UM_1,UM_1_0050335,2,0,17.8,1,R
+50336,UM_1,UM_1_0050336,2,0,14.3,2,R
+50337,UM_1,UM_1_0050337,2,0,11.8,1,R
+50338,UM_1,UM_1_0050338,2,0,13.5,2,R
+50339,UM_1,UM_1_0050339,2,0,18,1,R
+50340,UM_1,UM_1_0050340,2,0,16.1,2,R
+50341,UM_1,UM_1_0050341,2,0,18.2,2,R
+50342,UM_1,UM_1_0050342,2,0,12.6,1,R
+50343,UM_1,UM_1_0050343,2,0,13.8,2,R
+50344,UM_1,UM_1_0050344,2,0,18.2,1,R
+50345,UM_1,UM_1_0050345,2,0,17.3,1,R
+50346,UM_1,UM_1_0050346,2,0,18.9,1,R
+50347,UM_1,UM_1_0050347,2,0,15.6,1,R
+50348,UM_1,UM_1_0050348,2,0,19.2,2,R
+50349,UM_1,UM_1_0050349,2,0,18.9,1,R
+50350,UM_1,UM_1_0050350,2,0,12.4,1,R
+50351,UM_1,UM_1_0050351,2,0,16.8,1,R
+50352,UM_1,UM_1_0050352,2,0,15.5,1,L
+50353,UM_1,UM_1_0050353,2,0,13.1,2,R
+50354,UM_1,UM_1_0050354,2,0,14,2,R
+50355,UM_1,UM_1_0050355,2,0,10.9,1,R
+50356,UM_1,UM_1_0050356,2,0,17.2,2,L
+50357,UM_1,UM_1_0050357,2,0,9.5,2,R
+50358,UM_1,UM_1_0050358,2,0,9.8,1,R
+50359,UM_1,UM_1_0050359,2,0,10.4,1,R
+50360,UM_1,UM_1_0050360,2,0,13.9,1,R
+50361,UM_1,UM_1_0050361,2,0,18.3,2,-9999
+50362,UM_1,UM_1_0050362,2,0,11.2,1,L
+50363,UM_1,UM_1_0050363,2,0,9.8,1,R
+50364,UM_1,UM_1_0050364,2,0,10.8,1,R
+50365,UM_1,UM_1_0050365,2,0,15.6,1,L
+50366,UM_1,UM_1_0050366,2,0,8.2,1,-9999
+50367,UM_1,UM_1_0050367,2,0,10.8,1,-9999
+50368,UM_1,UM_1_0050368,2,0,17.9,1,R
+50369,UM_1,UM_1_0050369,2,0,14.8,2,R
+50370,UM_1,UM_1_0050370,2,0,12.8,1,R
+50372,UM_1,UM_1_0050372,2,0,10.2,1,R
+50373,UM_1,UM_1_0050373,2,0,17.2,2,R
+50374,UM_1,UM_1_0050374,2,0,9.7,2,R
+50375,UM_1,UM_1_0050375,2,0,11.6,2,R
+50376,UM_1,UM_1_0050376,2,0,9.2,1,R
+50377,UM_1,UM_1_0050377,2,0,11.5,1,R
+50379,UM_1,UM_1_0050379,2,0,14.3,2,R
+50380,UM_1,UM_1_0050380,2,0,17.4,2,L
+50381,UM_1,UM_1_0050381,2,0,16.5,1,R
+50382,UM_2,UM_2_0050382,2,0,28.8,1,R
+50383,UM_2,UM_2_0050383,2,0,14.4,1,R
+50385,UM_2,UM_2_0050385,2,0,17.8,1,R
+50386,UM_2,UM_2_0050386,2,0,17.8,1,R
+50387,UM_2,UM_2_0050387,2,0,14,1,R
+50388,UM_2,UM_2_0050388,2,0,26.8,1,R
+50390,UM_2,UM_2_0050390,2,0,14.5,1,R
+50391,UM_2,UM_2_0050391,2,0,17.2,1,L
+50397,UM_2,UM_2_0050397,1,1,14.7,1,-9999
+50399,UM_2,UM_2_0050399,1,2,17.4,1,R
+50402,UM_2,UM_2_0050402,1,1,13.1,1,R
+50403,UM_2,UM_2_0050403,1,2,12.8,1,R
+50404,UM_2,UM_2_0050404,1,1,15.8,1,R
+50405,UM_2,UM_2_0050405,1,1,16.6,1,R
+50406,UM_2,UM_2_0050406,1,1,15.2,1,R
+50407,UM_2,UM_2_0050407,1,1,16.6,2,R
+50408,UM_2,UM_2_0050408,1,1,14.8,1,R
+50410,UM_2,UM_2_0050410,1,1,16.1,1,R
+50411,UM_2,UM_2_0050411,1,2,14,1,R
+50412,UM_2,UM_2_0050412,1,1,13.2,1,R
+50413,UM_2,UM_2_0050413,1,1,13.1,1,L
+50414,UM_2,UM_2_0050414,2,0,13.8,2,R
+50415,UM_2,UM_2_0050415,2,0,17.6,1,R
+50416,UM_2,UM_2_0050416,2,0,14.8,1,R
+50417,UM_2,UM_2_0050417,2,0,13.3,1,L
+50418,UM_2,UM_2_0050418,2,0,14.9,1,R
+50419,UM_2,UM_2_0050419,2,0,15.8,1,R
+50421,UM_2,UM_2_0050421,2,0,15.4,1,R
+50422,UM_2,UM_2_0050422,2,0,14.8,1,R
+50424,UM_2,UM_2_0050424,2,0,17.9,1,R
+50425,UM_2,UM_2_0050425,2,0,16.6,1,R
+50426,UM_2,UM_2_0050426,2,0,13.6,1,R
+50427,UM_2,UM_2_0050427,2,0,15.8,1,R
+50428,UM_2,UM_2_0050428,2,0,15.5,1,R
+50433,USM,USM_0050433,2,0,18.7379,1,
+50434,USM,USM_0050434,2,0,18.2615,1,
+50435,USM,USM_0050435,2,0,16.104,1,
+50436,USM,USM_0050436,2,0,13.8097,1,
+50437,USM,USM_0050437,2,0,14.9432,1,
+50438,USM,USM_0050438,2,0,15.0253,1,
+50439,USM,USM_0050439,2,0,22.601,1,
+50440,USM,USM_0050440,2,0,23.948,1,
+50441,USM,USM_0050441,2,0,27.5975,1,
+50442,USM,USM_0050442,2,0,26.8528,1,
+50443,USM,USM_0050443,2,0,17.0705,1,
+50444,USM,USM_0050444,2,0,24.9884,1,
+50445,USM,USM_0050445,2,0,18.1383,1,
+50446,USM,USM_0050446,2,0,27.3238,1,
+50447,USM,USM_0050447,2,0,12.9391,1,
+50448,USM,USM_0050448,2,0,10.4668,1,
+50449,USM,USM_0050449,2,0,18.2177,1,
+50453,USM,USM_0050453,2,0,8.7721,1,
+50463,USM,USM_0050463,2,0,28.3669,1,
+50466,USM,USM_0050466,2,0,39.2526,1,
+50467,USM,USM_0050467,2,0,19.7591,1,
+50468,USM,USM_0050468,2,0,39.3949,1,
+50469,USM,USM_0050469,2,0,28.5613,1,
+50470,USM,USM_0050470,2,0,9.9165,1,
+50477,USM,USM_0050477,1,1,20.1807,1,
+50480,USM,USM_0050480,1,1,29.0897,1,
+50481,USM,USM_0050481,1,1,17.1663,1,
+50482,USM,USM_0050482,1,1,27.4114,1,
+50483,USM,USM_0050483,1,1,21.4346,1,
+50485,USM,USM_0050485,1,1,23.5948,1,
+50486,USM,USM_0050486,1,1,17.7878,1,
+50487,USM,USM_0050487,1,1,25.7972,1,
+50488,USM,USM_0050488,1,1,25.4511,1,
+50490,USM,USM_0050490,1,1,19.6277,1,
+50491,USM,USM_0050491,1,1,26.9487,1,
+50492,USM,USM_0050492,1,1,32.5503,1,
+50493,USM,USM_0050493,1,1,35.7098,1,
+50494,USM,USM_0050494,1,1,37.7769,1,
+50496,USM,USM_0050496,1,1,24.3231,1,
+50497,USM,USM_0050497,1,1,34.4613,1,
+50498,USM,USM_0050498,1,1,21.1307,1,
+50499,USM,USM_0050499,1,1,25.4346,1,
+50500,USM,USM_0050500,1,1,17.2375,1,
+50501,USM,USM_0050501,1,1,17.7057,1,
+50502,USM,USM_0050502,1,1,32.9582,1,
+50503,USM,USM_0050503,1,1,28.6708,1,
+50504,USM,USM_0050504,1,1,17.6454,1,
+50507,USM,USM_0050507,1,1,28.1095,1,
+50509,USM,USM_0050509,1,1,16.7337,1,
+50510,USM,USM_0050510,1,1,14.1437,1,
+50514,USM,USM_0050514,1,1,21.4018,1,
+50515,USM,USM_0050515,1,1,15.8522,1,
+50516,USM,USM_0050516,1,1,17.1143,1,
+50518,USM,USM_0050518,1,1,18.6585,1,
+50519,USM,USM_0050519,1,1,16.3778,1,
+50520,USM,USM_0050520,1,1,17.6728,1,
+50521,USM,USM_0050521,1,1,18.4175,1,
+50523,USM,USM_0050523,1,1,17.3251,1,
+50524,USM,USM_0050524,1,1,12.334,1,
+50525,USM,USM_0050525,1,1,32.8487,1,
+50526,USM,USM_0050526,1,1,50.2231,1,
+50527,USM,USM_0050527,1,3,18.4148,1,
+50528,USM,USM_0050528,1,1,11.3539,1,
+50529,USM,USM_0050529,1,1,42.3354,1,
+50530,USM,USM_0050530,1,1,18.245,1,
+50531,USM,USM_0050531,1,1,28.0903,1,
+50532,USM,USM_0050532,1,1,16.9199,1,
+50551,YALE,Yale_0050551,2,0,15.92,1,R
+50552,YALE,Yale_0050552,2,0,12.75,1,R
+50555,YALE,Yale_0050555,2,0,14.42,2,R
+50557,YALE,Yale_0050557,2,0,14.25,2,R
+50558,YALE,Yale_0050558,2,0,13.75,2,R
+50561,YALE,Yale_0050561,2,0,13.33,1,R
+50563,YALE,Yale_0050563,2,0,13.92,2,R
+50565,YALE,Yale_0050565,2,0,11,2,R
+50568,YALE,Yale_0050568,2,0,13.92,1,R
+50569,YALE,Yale_0050569,2,0,16.66,2,R
+50570,YALE,Yale_0050570,2,0,16.66,1,L
+50571,YALE,Yale_0050571,2,0,11,1,R
+50572,YALE,Yale_0050572,2,0,15.75,2,R
+50573,YALE,Yale_0050573,2,0,14.17,1,R
+50574,YALE,Yale_0050574,2,0,10.92,1,R
+50575,YALE,Yale_0050575,2,0,14,1,R
+50576,YALE,Yale_0050576,2,0,8.42,2,R
+50577,YALE,Yale_0050577,2,0,15.25,1,R
+50578,YALE,Yale_0050578,2,0,12.33,1,L
+50601,YALE,Yale_0050601,1,3,14.33,1,L
+50602,YALE,Yale_0050602,1,2,10,1,R
+50603,YALE,Yale_0050603,1,1,13.25,1,L
+50604,YALE,Yale_0050604,1,1,14.42,2,R
+50606,YALE,Yale_0050606,1,3,16.42,1,R
+50607,YALE,Yale_0050607,1,2,17.17,1,R
+50608,YALE,Yale_0050608,1,3,16.58,1,R
+50612,YALE,Yale_0050612,1,2,11.83,1,R
+50613,YALE,Yale_0050613,1,2,12.5,1,L
+50614,YALE,Yale_0050614,1,1,17.75,1,R
+50615,YALE,Yale_0050615,1,2,8.58,2,R
+50616,YALE,Yale_0050616,1,3,11.83,1,R
+50619,YALE,Yale_0050619,1,1,16.5,2,R
+50620,YALE,Yale_0050620,1,3,12.75,2,R
+50621,YALE,Yale_0050621,1,2,16.58,2,R
+50622,YALE,Yale_0050622,1,3,9.92,1,R
+50623,YALE,Yale_0050623,1,3,13.66,2,R
+50624,YALE,Yale_0050624,1,3,11.08,2,R
+50625,YALE,Yale_0050625,1,3,7,1,L
+50626,YALE,Yale_0050626,1,3,11.08,1,L
+50627,YALE,Yale_0050627,1,3,9.5,2,R
+50628,YALE,Yale_0050628,1,3,14.42,1,R
+50642,CMU,CMU_a_0050642,1,1,33,1,R
+50644,CMU,CMU_b_0050644,1,1,19,2,R
+50647,CMU,CMU_a_0050647,1,1,27,1,R
+50648,CMU,CMU_b_0050648,1,1,31,2,R
+50649,CMU,CMU_a_0050649,1,1,22,1,R
+50654,CMU,CMU_a_0050654,1,1,24,1,L
+50656,CMU,CMU_a_0050656,2,0,28,2,R
+50659,CMU,CMU_a_0050659,2,0,27,1,R
+50664,CMU,CMU_a_0050664,2,0,21,1,R
+50665,CMU,CMU_a_0050665,2,0,33,1,R
+50669,CMU,CMU_b_0050669,2,0,30,2,R
+50682,LEUVEN_1,Leuven_1_0050682,2,0,23,1,R
+50683,LEUVEN_1,Leuven_1_0050683,2,0,24,1,R
+50685,LEUVEN_1,Leuven_1_0050685,2,0,23,1,R
+50686,LEUVEN_1,Leuven_1_0050686,1,1,19,1,R
+50687,LEUVEN_1,Leuven_1_0050687,2,0,22,1,R
+50688,LEUVEN_1,Leuven_1_0050688,2,0,21,1,R
+50689,LEUVEN_1,Leuven_1_0050689,1,1,21,1,R
+50690,LEUVEN_1,Leuven_1_0050690,1,1,22,1,R
+50691,LEUVEN_1,Leuven_1_0050691,2,0,22,1,R
+50692,LEUVEN_1,Leuven_1_0050692,2,0,22,1,R
+50693,LEUVEN_1,Leuven_1_0050693,1,1,22,1,R
+50694,LEUVEN_1,Leuven_1_0050694,1,1,19,1,R
+50695,LEUVEN_1,Leuven_1_0050695,1,1,19,1,R
+50696,LEUVEN_1,Leuven_1_0050696,1,1,20,1,R
+50697,LEUVEN_1,Leuven_1_0050697,1,1,19,1,R
+50698,LEUVEN_1,Leuven_1_0050698,2,0,28,1,R
+50699,LEUVEN_1,Leuven_1_0050699,2,0,21,1,R
+50700,LEUVEN_1,Leuven_1_0050700,1,1,23,1,L
+50701,LEUVEN_1,Leuven_1_0050701,2,0,18,1,R
+50702,LEUVEN_1,Leuven_1_0050702,1,1,18,1,R
+50703,LEUVEN_1,Leuven_1_0050703,2,0,29,1,R
+50704,LEUVEN_1,Leuven_1_0050704,1,1,29,1,R
+50705,LEUVEN_1,Leuven_1_0050705,1,1,24,1,R
+50706,LEUVEN_1,Leuven_1_0050706,2,0,22,1,L
+50707,LEUVEN_1,Leuven_1_0050707,2,0,22,1,R
+50708,LEUVEN_1,Leuven_1_0050708,1,1,32,1,R
+50709,LEUVEN_1,Leuven_1_0050709,2,0,25,1,R
+50711,LEUVEN_1,Leuven_1_0050711,1,1,19,1,R
+50722,LEUVEN_2,Leuven_2_0050722,2,0,13.8,2,L
+50723,LEUVEN_2,Leuven_2_0050723,2,0,13.8,2,L
+50724,LEUVEN_2,Leuven_2_0050724,2,0,15.1,1,R
+50725,LEUVEN_2,Leuven_2_0050725,2,0,14.2,1,L
+50726,LEUVEN_2,Leuven_2_0050726,2,0,16.6,1,R
+50728,LEUVEN_2,Leuven_2_0050728,2,0,15.6,1,R
+50730,LEUVEN_2,Leuven_2_0050730,2,0,12.3,2,R
+50731,LEUVEN_2,Leuven_2_0050731,2,0,16.2,1,R
+50733,LEUVEN_2,Leuven_2_0050733,2,0,14.7,1,R
+50735,LEUVEN_2,Leuven_2_0050735,2,0,14.3,2,R
+50737,LEUVEN_2,Leuven_2_0050737,2,0,12.2,1,R
+50738,LEUVEN_2,Leuven_2_0050738,2,0,12.8,1,R
+50739,LEUVEN_2,Leuven_2_0050739,2,0,15.2,1,R
+50740,LEUVEN_2,Leuven_2_0050740,2,0,12.4,1,R
+50741,LEUVEN_2,Leuven_2_0050741,2,0,14.2,1,R
+50742,LEUVEN_2,Leuven_2_0050742,2,0,16.9,1,R
+50743,LEUVEN_2,Leuven_2_0050743,1,1,14.2,2,R
+50744,LEUVEN_2,Leuven_2_0050744,1,1,12.5,2,R
+50745,LEUVEN_2,Leuven_2_0050745,1,1,13,1,R
+50748,LEUVEN_2,Leuven_2_0050748,1,1,14.2,1,R
+50749,LEUVEN_2,Leuven_2_0050749,1,1,15.3,2,L
+50750,LEUVEN_2,Leuven_2_0050750,1,1,16.8,1,R
+50751,LEUVEN_2,Leuven_2_0050751,1,1,12.1,1,R
+50752,LEUVEN_2,Leuven_2_0050752,1,1,14.7,1,L
+50754,LEUVEN_2,Leuven_2_0050754,1,1,12.3,1,R
+50755,LEUVEN_2,Leuven_2_0050755,1,1,12.2,1,R
+50756,LEUVEN_2,Leuven_2_0050756,1,1,14.9,1,R
+50757,LEUVEN_2,Leuven_2_0050757,1,1,14.3,1,R
+50772,KKI,KKI_0050772,2,0,12.77,1,R
+50773,KKI,KKI_0050773,2,0,10.84,1,R
+50774,KKI,KKI_0050774,2,0,10.64,1,R
+50775,KKI,KKI_0050775,2,0,11,1,R
+50776,KKI,KKI_0050776,2,0,9.3,1,R
+50777,KKI,KKI_0050777,2,0,8.39,1,R
+50778,KKI,KKI_0050778,2,0,9.73,2,R
+50780,KKI,KKI_0050780,2,0,9.82,2,R
+50781,KKI,KKI_0050781,2,0,9.3,1,R
+50782,KKI,KKI_0050782,2,0,10.18,1,R
+50783,KKI,KKI_0050783,2,0,10.4,2,R
+50786,KKI,KKI_0050786,2,0,8.75,1,R
+50790,KKI,KKI_0050790,2,0,8.83,2,R
+50791,KKI,KKI_0050791,1,1,10.18,1,R
+50792,KKI,KKI_0050792,1,2,8.2,2,Mixed
+50796,KKI,KKI_0050796,1,2,11.99,2,R
+50797,KKI,KKI_0050797,1,2,12.54,1,R
+50798,KKI,KKI_0050798,1,1,10.17,2,Mixed
+50799,KKI,KKI_0050799,1,1,11.14,1,R
+50800,KKI,KKI_0050800,1,2,11.76,1,R
+50801,KKI,KKI_0050801,1,1,11.27,1,R
+50803,KKI,KKI_0050803,1,2,8.45,1,R
+50807,KKI,KKI_0050807,1,1,10.65,1,Mixed
+50812,KKI,KKI_0050812,2,0,9.97,2,R
+50814,KKI,KKI_0050814,2,0,8.46,1,L
+50816,KKI,KKI_0050816,2,0,9.73,1,Mixed
+50817,KKI,KKI_0050817,2,0,9.97,1,Mixed
+50818,KKI,KKI_0050818,2,0,11.79,1,R
+50820,KKI,KKI_0050820,2,0,8.87,2,R
+50821,KKI,KKI_0050821,2,0,11.17,1,R
+50822,KKI,KKI_0050822,2,0,12.43,1,R
+50823,KKI,KKI_0050823,1,2,11.37,1,R
+50824,KKI,KKI_0050824,1,2,10.29,1,R
+50952,NYU,NYU_0050952,1,1,8.74,2,
+50954,NYU,NYU_0050954,1,1,14.75,2,
+50955,NYU,NYU_0050955,1,1,12.64,2,
+50956,NYU,NYU_0050956,1,1,14.06,2,
+50957,NYU,NYU_0050957,1,1,14.75,2,
+50958,NYU,NYU_0050958,1,1,10.23,2,
+50959,NYU,NYU_0050959,1,2,22.81,2,
+50960,NYU,NYU_0050960,1,2,38.76,2,
+50961,NYU,NYU_0050961,1,2,25.02,2,
+50962,NYU,NYU_0050962,1,2,24.41,2,
+50964,NYU,NYU_0050964,1,1,12.75,1,
+50965,NYU,NYU_0050965,1,3,9.25,1,
+50966,NYU,NYU_0050966,1,3,15.14,1,
+50967,NYU,NYU_0050967,1,3,9.74,1,
+50968,NYU,NYU_0050968,1,3,9.95,1,
+50969,NYU,NYU_0050969,1,2,7.61,1,
+50970,NYU,NYU_0050970,1,2,8.9,1,
+50972,NYU,NYU_0050972,1,2,13.95,1,
+50973,NYU,NYU_0050973,1,2,17.88,1,
+50974,NYU,NYU_0050974,1,2,11.56,1,
+50976,NYU,NYU_0050976,1,2,14.65,1,
+50977,NYU,NYU_0050977,1,2,7.13,1,
+50978,NYU,NYU_0050978,1,2,9.58,1,
+50979,NYU,NYU_0050979,1,2,9.37,1,
+50981,NYU,NYU_0050981,1,1,12.96,1,
+50982,NYU,NYU_0050982,1,1,9.46,1,
+50983,NYU,NYU_0050983,1,1,10.48,1,
+50984,NYU,NYU_0050984,1,1,13.2,1,
+50985,NYU,NYU_0050985,1,1,13.09,1,
+50986,NYU,NYU_0050986,1,1,8.53,1,
+50987,NYU,NYU_0050987,1,1,8.56,1,
+50988,NYU,NYU_0050988,1,1,14.21,1,
+50989,NYU,NYU_0050989,1,1,9.78,1,
+50990,NYU,NYU_0050990,1,1,13.71,1,
+50991,NYU,NYU_0050991,1,1,10.9,1,
+50992,NYU,NYU_0050992,1,1,10.66,1,
+50993,NYU,NYU_0050993,1,1,10.87,1,
+50994,NYU,NYU_0050994,1,1,15.66,1,
+50995,NYU,NYU_0050995,1,1,16.74,1,
+50996,NYU,NYU_0050996,1,1,15.81,1,
+50997,NYU,NYU_0050997,1,1,16.27,1,
+50999,NYU,NYU_0050999,1,1,14.53,1,
+51000,NYU,NYU_0051000,1,1,13.33,1,
+51001,NYU,NYU_0051001,1,1,10.71,1,
+51002,NYU,NYU_0051002,1,1,7.228,1,
+51003,NYU,NYU_0051003,1,1,8.51,1,
+51006,NYU,NYU_0051006,1,1,11.11,1,
+51007,NYU,NYU_0051007,1,1,11.92,1,
+51008,NYU,NYU_0051008,1,1,12.37,1,
+51009,NYU,NYU_0051009,1,1,11.03,1,
+51010,NYU,NYU_0051010,1,1,8.53,1,
+51011,NYU,NYU_0051011,1,1,8.79,1,
+51012,NYU,NYU_0051012,1,1,9.79,1,
+51013,NYU,NYU_0051013,1,1,7.15,1,
+51014,NYU,NYU_0051014,1,1,11.005,1,
+51015,NYU,NYU_0051015,1,2,29.98,1,
+51016,NYU,NYU_0051016,1,2,22.99,1,
+51017,NYU,NYU_0051017,1,2,22.04,1,
+51018,NYU,NYU_0051018,1,2,20.25,1,
+51019,NYU,NYU_0051019,1,2,22.29,1,
+51020,NYU,NYU_0051020,1,1,28.58,1,
+51021,NYU,NYU_0051021,1,1,23.66,1,
+51023,NYU,NYU_0051023,1,1,20.62,1,
+51024,NYU,NYU_0051024,1,1,39.1,1,
+51025,NYU,NYU_0051025,1,1,19.64,1,
+51026,NYU,NYU_0051026,1,1,18.58,1,
+51027,NYU,NYU_0051027,1,1,22.98,1,
+51028,NYU,NYU_0051028,1,1,29.18,1,
+51029,NYU,NYU_0051029,1,1,26.516,1,
+51030,NYU,NYU_0051030,1,3,7.29,1,
+51032,NYU,NYU_0051032,1,1,7.15,1,
+51033,NYU,NYU_0051033,1,1,9.98,1,
+51034,NYU,NYU_0051034,1,1,10.65,1,
+51035,NYU,NYU_0051035,1,1,10.27,1,
+51036,NYU,NYU_0051036,2,0,8.04,2,
+51038,NYU,NYU_0051038,2,0,8.26,2,
+51039,NYU,NYU_0051039,2,0,8.5,2,
+51040,NYU,NYU_0051040,2,0,8.52,2,
+51041,NYU,NYU_0051041,2,0,8.9,2,
+51042,NYU,NYU_0051042,2,0,8.91,2,
+51044,NYU,NYU_0051044,2,0,10.86,2,
+51045,NYU,NYU_0051045,2,0,11.56,2,
+51046,NYU,NYU_0051046,2,0,11.56,2,
+51047,NYU,NYU_0051047,2,0,12.1,2,
+51048,NYU,NYU_0051048,2,0,12.44,2,
+51049,NYU,NYU_0051049,2,0,13.22,2,
+51050,NYU,NYU_0051050,2,0,13.25,2,
+51051,NYU,NYU_0051051,2,0,14.06,2,
+51052,NYU,NYU_0051052,2,0,14.18,2,
+51053,NYU,NYU_0051053,2,0,14.38,2,
+51054,NYU,NYU_0051054,2,0,14.42,2,
+51055,NYU,NYU_0051055,2,0,15.95,2,
+51056,NYU,NYU_0051056,2,0,17.31,2,
+51057,NYU,NYU_0051057,2,0,21.58,2,
+51058,NYU,NYU_0051058,2,0,22.11,2,
+51059,NYU,NYU_0051059,2,0,22.76,2,
+51060,NYU,NYU_0051060,2,0,22.74,2,
+51061,NYU,NYU_0051061,2,0,27.03,2,
+51062,NYU,NYU_0051062,2,0,27.76,2,
+51063,NYU,NYU_0051063,2,0,29.13,2,
+51064,NYU,NYU_0051064,2,0,7.26,1,
+51065,NYU,NYU_0051065,2,0,10.52,1,
+51066,NYU,NYU_0051066,2,0,18.59,1,
+51067,NYU,NYU_0051067,2,0,23.65,1,
+51068,NYU,NYU_0051068,2,0,31.78,1,
+51069,NYU,NYU_0051069,2,0,8.15,1,
+51070,NYU,NYU_0051070,2,0,7.29,1,
+51072,NYU,NYU_0051072,2,0,11.91,1,
+51073,NYU,NYU_0051073,2,0,12.49,1,
+51074,NYU,NYU_0051074,2,0,12.81,1,
+51075,NYU,NYU_0051075,2,0,14.2,1,
+51076,NYU,NYU_0051076,2,0,16.93,1,
+51077,NYU,NYU_0051077,2,0,17.3,1,
+51078,NYU,NYU_0051078,2,0,6.47,1,
+51079,NYU,NYU_0051079,2,0,7.19,1,
+51080,NYU,NYU_0051080,2,0,8.01,1,
+51081,NYU,NYU_0051081,2,0,8.82,1,
+51082,NYU,NYU_0051082,2,0,8.88,1,
+51083,NYU,NYU_0051083,2,0,8.93,1,
+51084,NYU,NYU_0051084,2,0,9.23,1,
+51085,NYU,NYU_0051085,2,0,9.81,1,
+51086,NYU,NYU_0051086,2,0,10.46,1,
+51087,NYU,NYU_0051087,2,0,10.52,1,
+51088,NYU,NYU_0051088,2,0,10.73,1,
+51089,NYU,NYU_0051089,2,0,10.76,1,
+51090,NYU,NYU_0051090,2,0,11.03,1,
+51091,NYU,NYU_0051091,2,0,11.32,1,
+51093,NYU,NYU_0051093,2,0,11.69,1,
+51094,NYU,NYU_0051094,2,0,12.1,1,
+51095,NYU,NYU_0051095,2,0,12.23,1,
+51096,NYU,NYU_0051096,2,0,13.18,1,
+51097,NYU,NYU_0051097,2,0,13.583,1,
+51098,NYU,NYU_0051098,2,0,13.628,1,
+51099,NYU,NYU_0051099,2,0,13.72,1,
+51100,NYU,NYU_0051100,2,0,14.36,1,
+51101,NYU,NYU_0051101,2,0,14.425,1,
+51102,NYU,NYU_0051102,2,0,14.653,1,
+51103,NYU,NYU_0051103,2,0,14.66,1,
+51104,NYU,NYU_0051104,2,0,15.27,1,
+51105,NYU,NYU_0051105,2,0,15.284,1,
+51106,NYU,NYU_0051106,2,0,15.42,1,
+51107,NYU,NYU_0051107,2,0,15.53,1,
+51109,NYU,NYU_0051109,2,0,16.13,1,
+51110,NYU,NYU_0051110,2,0,16.28,1,
+51111,NYU,NYU_0051111,2,0,16.55,1,
+51112,NYU,NYU_0051112,2,0,19.228,1,
+51113,NYU,NYU_0051113,2,0,22.38,1,
+51114,NYU,NYU_0051114,2,0,22.48,1,
+51116,NYU,NYU_0051116,2,0,25.34,1,
+51117,NYU,NYU_0051117,2,0,28.203,1,
+51118,NYU,NYU_0051118,2,0,29.02,1,
+51122,NYU,NYU_0051122,2,0,11.92,1,
+51123,NYU,NYU_0051123,2,0,12.07,1,
+51124,NYU,NYU_0051124,2,0,12.67,1,
+51126,NYU,NYU_0051126,2,0,16.31,1,
+51127,NYU,NYU_0051127,2,0,16.55,1,
+51128,NYU,NYU_0051128,2,0,16.88,1,
+51129,NYU,NYU_0051129,2,0,17.7,1,
+51130,NYU,NYU_0051130,2,0,19.13,1,
+51131,NYU,NYU_0051131,2,0,19.73,1,
+51132,TRINITY,Trinity_0051132,2,0,24.83,1,R
+51133,TRINITY,Trinity_0051133,2,0,12.04,1,R
+51134,TRINITY,Trinity_0051134,2,0,16.83,1,R
+51135,TRINITY,Trinity_0051135,2,0,13.75,1,R
+51136,TRINITY,Trinity_0051136,2,0,12.66,1,R
+51137,TRINITY,Trinity_0051137,2,0,13.25,1,R
+51138,TRINITY,Trinity_0051138,2,0,12.66,1,R
+51139,TRINITY,Trinity_0051139,2,0,19.5,1,R
+51140,TRINITY,Trinity_0051140,2,0,15.75,1,R
+51141,TRINITY,Trinity_0051141,2,0,15.91,1,R
+51142,TRINITY,Trinity_0051142,2,0,14.91,1,R
+51146,NYU,NYU_0051146,2,0,20.02,1,
+51147,NYU,NYU_0051147,2,0,20.04,1,
+51148,NYU,NYU_0051148,2,0,20.3,1,
+51149,NYU,NYU_0051149,2,0,20.56,1,
+51150,NYU,NYU_0051150,2,0,22.77,1,
+51151,NYU,NYU_0051151,2,0,23.08,1,
+51152,NYU,NYU_0051152,2,0,23.35,1,
+51153,NYU,NYU_0051153,2,0,26.17,1,
+51154,NYU,NYU_0051154,2,0,30.08,1,
+51155,NYU,NYU_0051155,2,0,30.78,1,
+51156,NYU,NYU_0051156,2,0,21.15,1,
+51159,NYU,NYU_0051159,2,0,12.81,1,
+51161,STANFORD,Stanford_0051161,1,-9999,9.0301,1,R
+51162,STANFORD,Stanford_0051162,1,-9999,8.9199,2,L
+51163,STANFORD,Stanford_0051163,1,-9999,9.0821,1,R
+51164,STANFORD,Stanford_0051164,1,-9999,8.4216,2,Ambi
+51168,STANFORD,Stanford_0051168,1,-9999,11.2628,1,L
+51169,STANFORD,Stanford_0051169,1,-9999,9.7775,1,L
+51170,STANFORD,Stanford_0051170,1,-9999,10.8652,2,R
+51171,STANFORD,Stanford_0051171,1,-9999,12.9363,1,R
+51173,STANFORD,Stanford_0051173,1,-9999,7.5284,1,R
+51177,STANFORD,Stanford_0051177,1,-9999,11.6982,1,R
+51178,STANFORD,Stanford_0051178,1,-9999,11.4326,1,R
+51179,STANFORD,Stanford_0051179,1,-9999,11.6543,1,R
+51180,STANFORD,Stanford_0051180,2,-9999,7.7502,2,R
+51181,STANFORD,Stanford_0051181,2,-9999,9.3149,1,R
+51182,STANFORD,Stanford_0051182,2,-9999,10.6927,1,R
+51183,STANFORD,Stanford_0051183,2,-9999,8.2656,1,R
+51184,STANFORD,Stanford_0051184,2,-9999,9.1068,2,R
+51185,STANFORD,Stanford_0051185,2,-9999,8.2464,1,R
+51187,STANFORD,Stanford_0051187,2,-9999,8.3313,1,R
+51188,STANFORD,Stanford_0051188,2,-9999,8.9747,2,R
+51189,STANFORD,Stanford_0051189,2,-9999,9.2081,1,R
+51192,STANFORD,Stanford_0051192,2,-9999,10.2218,2,R
+51194,STANFORD,Stanford_0051194,2,-9999,12.3477,1,R
+51197,STANFORD,Stanford_0051197,2,-9999,12.4298,1,R
+51198,STANFORD,Stanford_0051198,2,-9999,12.2546,1,Ambi
+51201,UCLA_1,UCLA_1_0051201,1,1,13.52,1,R
+51202,UCLA_1,UCLA_1_0051202,1,1,11.56,1,R
+51203,UCLA_1,UCLA_1_0051203,1,1,13.37,1,R
+51204,UCLA_1,UCLA_1_0051204,1,1,14.57,1,R
+51205,UCLA_1,UCLA_1_0051205,1,1,17.94,1,R
+51206,UCLA_1,UCLA_1_0051206,1,1,15.78,1,R
+51207,UCLA_1,UCLA_1_0051207,1,1,14.11,2,R
+51208,UCLA_1,UCLA_1_0051208,1,1,16.98,1,L
+51210,UCLA_1,UCLA_1_0051210,1,1,16.56,1,R
+51211,UCLA_1,UCLA_1_0051211,1,1,11.27,1,R
+51212,UCLA_1,UCLA_1_0051212,1,1,15.66,1,L
+51214,UCLA_1,UCLA_1_0051214,1,1,10.41,1,R
+51215,UCLA_1,UCLA_1_0051215,1,1,11.56,2,R
+51216,UCLA_1,UCLA_1_0051216,1,1,10.54,1,L
+51217,UCLA_1,UCLA_1_0051217,1,1,15.22,1,R
+51218,UCLA_1,UCLA_1_0051218,1,1,16.87,1,R
+51219,UCLA_1,UCLA_1_0051219,1,1,13.93,2,R
+51220,UCLA_1,UCLA_1_0051220,1,1,14.74,1,R
+51221,UCLA_1,UCLA_1_0051221,1,1,14.58,1,R
+51222,UCLA_1,UCLA_1_0051222,1,1,11.81,1,R
+51223,UCLA_1,UCLA_1_0051223,1,1,13.31,1,R
+51224,UCLA_1,UCLA_1_0051224,1,1,12.99,1,R
+51225,UCLA_1,UCLA_1_0051225,1,1,12.33,1,R
+51226,UCLA_1,UCLA_1_0051226,1,1,9.26,2,R
+51228,UCLA_1,UCLA_1_0051228,1,1,8.36,2,R
+51229,UCLA_1,UCLA_1_0051229,1,1,14.59,1,R
+51230,UCLA_1,UCLA_1_0051230,1,1,17.53,2,R
+51231,UCLA_1,UCLA_1_0051231,1,1,10.28,1,R
+51234,UCLA_1,UCLA_1_0051234,1,1,10.91,1,R
+51235,UCLA_1,UCLA_1_0051235,1,1,10.67,1,R
+51236,UCLA_1,UCLA_1_0051236,1,1,12.42,1,R
+51237,UCLA_1,UCLA_1_0051237,1,1,17.4,1,R
+51239,UCLA_1,UCLA_1_0051239,1,1,13.67,1,R
+51240,UCLA_1,UCLA_1_0051240,1,1,14.96,1,R
+51241,UCLA_1,UCLA_1_0051241,1,1,10.9,1,R
+51248,UCLA_1,UCLA_1_0051248,1,1,13.5,1,R
+51249,UCLA_1,UCLA_1_0051249,1,1,8.49,1,R
+51250,UCLA_1,UCLA_1_0051250,2,0,14.88,1,R
+51251,UCLA_1,UCLA_1_0051251,2,0,12.36,1,R
+51252,UCLA_1,UCLA_1_0051252,2,0,10.99,1,R
+51253,UCLA_1,UCLA_1_0051253,2,0,11.8,1,L
+51254,UCLA_1,UCLA_1_0051254,2,0,14.44,1,L
+51255,UCLA_1,UCLA_1_0051255,2,0,14.8,1,R
+51256,UCLA_1,UCLA_1_0051256,2,0,15.04,1,R
+51257,UCLA_1,UCLA_1_0051257,2,0,13.38,1,R
+51260,UCLA_1,UCLA_1_0051260,2,0,13.66,1,R
+51261,UCLA_1,UCLA_1_0051261,2,0,17.79,1,R
+51262,UCLA_1,UCLA_1_0051262,2,0,11.56,1,R
+51264,UCLA_1,UCLA_1_0051264,2,0,13.46,2,R
+51265,UCLA_1,UCLA_1_0051265,2,0,15.92,1,R
+51266,UCLA_1,UCLA_1_0051266,2,0,14.03,1,R
+51267,UCLA_1,UCLA_1_0051267,2,0,11.25,2,R
+51268,UCLA_1,UCLA_1_0051268,2,0,17.78,1,R
+51269,UCLA_1,UCLA_1_0051269,2,0,14.29,1,R
+51271,UCLA_1,UCLA_1_0051271,2,0,12.09,1,R
+51272,UCLA_1,UCLA_1_0051272,2,0,13.41,1,R
+51273,UCLA_1,UCLA_1_0051273,2,0,12.68,1,R
+51275,UCLA_1,UCLA_1_0051275,2,0,15.74,1,R
+51276,UCLA_1,UCLA_1_0051276,2,0,13.85,1,R
+51277,UCLA_1,UCLA_1_0051277,2,0,12.01,1,R
+51278,UCLA_1,UCLA_1_0051278,2,0,9.21,1,L
+51279,UCLA_1,UCLA_1_0051279,2,0,13.82,2,R
+51280,UCLA_1,UCLA_1_0051280,2,0,9.5,1,R
+51281,UCLA_1,UCLA_1_0051281,2,0,11.83,1,R
+51291,UCLA_2,UCLA_2_0051291,1,1,16.47,1,R
+51292,UCLA_2,UCLA_2_0051292,1,1,12.24,1,R
+51293,UCLA_2,UCLA_2_0051293,1,1,13.08,1,R
+51294,UCLA_2,UCLA_2_0051294,1,1,11.7,1,R
+51295,UCLA_2,UCLA_2_0051295,1,1,10.04,1,L
+51297,UCLA_2,UCLA_2_0051297,1,1,14.27,1,L
+51298,UCLA_2,UCLA_2_0051298,1,1,10.57,1,R
+51299,UCLA_2,UCLA_2_0051299,1,1,14.77,1,R
+51300,UCLA_2,UCLA_2_0051300,1,1,14.08,1,L
+51301,UCLA_2,UCLA_2_0051301,1,1,13.28,1,R
+51302,UCLA_2,UCLA_2_0051302,1,1,10.85,1,R
+51303,UCLA_2,UCLA_2_0051303,2,0,13,2,R
+51304,UCLA_2,UCLA_2_0051304,2,0,10.53,1,R
+51305,UCLA_2,UCLA_2_0051305,2,0,12.64,2,R
+51306,UCLA_2,UCLA_2_0051306,2,0,11.66,1,R
+51307,UCLA_2,UCLA_2_0051307,2,0,12.15,1,R
+51308,UCLA_2,UCLA_2_0051308,2,0,9.79,1,R
+51309,UCLA_2,UCLA_2_0051309,2,0,11.9,1,R
+51311,UCLA_2,UCLA_2_0051311,2,0,11.95,1,R
+51313,UCLA_2,UCLA_2_0051313,2,0,13.36,1,R
+51315,UCLA_2,UCLA_2_0051315,2,0,13.63,1,R
+51318,MAX_MUN,MaxMun_a_0051318,1,2,19,1,R
+51319,MAX_MUN,MaxMun_a_0051319,1,2,33,1,R
+51320,MAX_MUN,MaxMun_a_0051320,1,2,31,1,R
+51321,MAX_MUN,MaxMun_a_0051321,1,2,45,2,R
+51322,MAX_MUN,MaxMun_b_0051322,1,2,29,1,L
+51323,MAX_MUN,MaxMun_b_0051323,1,2,30,1,R
+51325,MAX_MUN,MaxMun_b_0051325,1,2,32,2,L
+51326,MAX_MUN,MaxMun_b_0051326,1,2,42,1,R
+51327,MAX_MUN,MaxMun_b_0051327,1,2,58,1,R
+51328,MAX_MUN,MaxMun_c_0051328,1,2,22,1,R
+51329,MAX_MUN,MaxMun_d_0051329,1,2,18,1,R
+51330,MAX_MUN,MaxMun_d_0051330,1,2,45,1,R
+51331,MAX_MUN,MaxMun_d_0051331,1,2,35,1,R
+51332,MAX_MUN,MaxMun_c_0051332,2,0,23,1,R
+51333,MAX_MUN,MaxMun_c_0051333,2,0,24,1,R
+51334,MAX_MUN,MaxMun_c_0051334,2,0,27,1,R
+51335,MAX_MUN,MaxMun_c_0051335,2,0,23,1,R
+51336,MAX_MUN,MaxMun_c_0051336,2,0,26,1,R
+51338,MAX_MUN,MaxMun_c_0051338,2,0,30,1,R
+51339,MAX_MUN,MaxMun_c_0051339,2,0,29,1,R
+51340,MAX_MUN,MaxMun_c_0051340,2,0,29,1,R
+51341,MAX_MUN,MaxMun_c_0051341,2,0,29,1,R
+51342,MAX_MUN,MaxMun_c_0051342,2,0,23,1,R
+51343,MAX_MUN,MaxMun_c_0051343,2,0,30,1,R
+51344,MAX_MUN,MaxMun_c_0051344,2,0,22,1,R
+51345,MAX_MUN,MaxMun_c_0051345,2,0,35,1,R
+51346,MAX_MUN,MaxMun_c_0051346,2,0,25,1,R
+51347,MAX_MUN,MaxMun_c_0051347,2,0,24,1,R
+51349,MAX_MUN,MaxMun_d_0051349,1,2,11,1,R
+51350,MAX_MUN,MaxMun_d_0051350,1,1,11,1,R
+51351,MAX_MUN,MaxMun_d_0051351,1,1,11,1,R
+51354,MAX_MUN,MaxMun_d_0051354,1,2,12,1,R
+51356,MAX_MUN,MaxMun_d_0051356,2,0,10,1,R
+51357,MAX_MUN,MaxMun_d_0051357,2,0,11,1,R
+51359,MAX_MUN,MaxMun_d_0051359,2,0,13,1,R
+51360,MAX_MUN,MaxMun_d_0051360,2,0,7,1,R
+51361,MAX_MUN,MaxMun_d_0051361,2,0,16,1,R
+51362,MAX_MUN,MaxMun_a_0051362,2,0,33,1,R
+51363,MAX_MUN,MaxMun_a_0051363,2,0,32,1,R
+51364,MAX_MUN,MaxMun_a_0051364,2,0,26,1,R
+51365,MAX_MUN,MaxMun_a_0051365,2,0,23,1,R
+51369,MAX_MUN,MaxMun_a_0051369,2,0,32,1,R
+51370,MAX_MUN,MaxMun_a_0051370,2,0,32,2,R
+51373,MAX_MUN,MaxMun_a_0051373,2,0,46,1,R
+51461,CALTECH,Caltech_0051461,1,4,37.7,1,R
+51463,CALTECH,Caltech_0051463,1,1,20.2,2,R
+51464,CALTECH,Caltech_0051464,1,1,20.9,1,Ambi
+51465,CALTECH,Caltech_0051465,1,1,20.2,1,R
+51473,CALTECH,Caltech_0051473,1,1,21.2,1,R
+51477,CALTECH,Caltech_0051477,2,0,42.5,1,R
+51479,CALTECH,Caltech_0051479,2,0,20,2,R
+51480,CALTECH,Caltech_0051480,2,0,20.8,2,R
+51481,CALTECH,Caltech_0051481,2,0,27.9,1,R
+51482,CALTECH,Caltech_0051482,2,0,21.1,2,R
+51484,CALTECH,Caltech_0051484,2,0,23.6,1,R
+51487,CALTECH,Caltech_0051487,2,0,17,1,R
+51488,CALTECH,Caltech_0051488,2,0,23.3,1,L
+51491,CALTECH,Caltech_0051491,2,0,56.2,1,R
+51493,CALTECH,Caltech_0051493,2,0,29.2,2,R
+51556,SBL,SBL_0051556,2,0,20,1,
+51557,SBL,SBL_0051557,2,0,26,1,
+51558,SBL,SBL_0051558,2,0,27,1,
+51559,SBL,SBL_0051559,2,0,27,1,
+51560,SBL,SBL_0051560,2,0,27,1,
+51562,SBL,SBL_0051562,2,0,37,1,
+51563,SBL,SBL_0051563,2,0,38,1,
+51564,SBL,SBL_0051564,2,0,39,1,
+51565,SBL,SBL_0051565,2,0,39,1,
+51566,SBL,SBL_0051566,2,0,41,1,
+51567,SBL,SBL_0051567,2,0,34,1,
+51568,SBL,SBL_0051568,2,0,37,1,
+51569,SBL,SBL_0051569,2,0,36,1,
+51570,SBL,SBL_0051570,2,0,42,1,
+51572,SBL,SBL_0051572,1,2,28,1,
+51573,SBL,SBL_0051573,1,3,30,1,L
+51574,SBL,SBL_0051574,1,2,30,1,
+51576,SBL,SBL_0051576,1,3,31,1,
+51577,SBL,SBL_0051577,1,3,31,1,
+51578,SBL,SBL_0051578,1,2,33,1,
+51579,SBL,SBL_0051579,1,3,41,1,
+51580,SBL,SBL_0051580,1,3,42,1,
+51582,SBL,SBL_0051582,1,2,31,1,
+51583,SBL,SBL_0051583,1,2,35,1,
+51584,SBL,SBL_0051584,1,2,49,1,
+51585,SBL,SBL_0051585,1,1,27,1,
+51606,MAX_MUN,MaxMun_a_0051606,1,2,29,2,R
+51607,MAX_MUN,MaxMun_a_0051607,1,2,26,1,R

AtlaScore/downstream/docs/behavior_ADNI.csv

@@ -0,0 +1,268 @@
+Subject ID,Sex,Research Group
+003_S_6258,M,MCI
+003_S_6259,M,CN
+003_S_6264,M,AD
+003_S_6678,F,MCI
+003_S_6833,F,AD
+006_S_6209,M,CN
+006_S_6234,F,CN
+006_S_6657,M,MCI
+006_S_6672,M,MCI
+006_S_6682,M,MCI
+006_S_6689,M,AD
+009_S_6212,F,CN
+009_S_6286,F,CN
+009_S_6402,M,MCI
+011_S_6303,M,AD
+012_S_6503,M,MCI
+013_S_6768,F,AD
+013_S_6975,M,AD
+014_S_6145,F,CN
+014_S_6522,M,CN
+014_S_6935,F,CN
+014_S_6944,F,MCI
+018_S_6207,F,CN
+018_S_6351,F,CN
+019_S_6585,F,AD
+019_S_6630,M,MCI
+019_S_6668,F,MCI
+019_S_6712,M,AD
+020_S_6566,M,CN
+021_S_6914,M,CN
+021_S_6940,M,CN
+022_S_6013,F,AD
+022_S_6069,F,CN
+022_S_6716,M,MCI
+027_S_6001,F,CN
+027_S_6648,M,AD
+027_S_6733,M,AD
+027_S_6793,M,MCI
+027_S_6849,M,AD
+027_S_6965,M,AD
+027_S_7024,M,CN
+032_S_6055,M,MCI
+032_S_6600,F,AD
+032_S_6602,M,AD
+032_S_6701,F,CN
+032_S_6717,M,CN
+032_S_6855,M,AD
+035_S_6195,F,CN
+035_S_6200,M,CN
+035_S_6306,M,CN
+035_S_6480,F,MCI
+035_S_6551,F,CN
+035_S_6650,M,AD
+035_S_6660,F,AD
+035_S_6730,M,CN
+035_S_6950,F,CN
+035_S_6967,F,CN
+035_S_6999,M,CN
+035_S_7001,M,AD
+035_S_7073,M,MCI
+035_S_7105,M,MCI
+035_S_7120,M,AD
+037_S_6115,F,CN
+041_S_6354,F,CN
+041_S_6447,F,CN
+053_S_7109,F,AD
+057_S_6746,M,AD
+057_S_6869,M,AD
+070_S_6229,M,MCI
+070_S_7078,M,MCI
+082_S_6197,F,CN
+082_S_6415,F,CN
+082_S_6690,F,AD
+094_S_6250,F,CN
+094_S_6269,F,CN
+098_S_6601,M,AD
+098_S_6655,M,AD
+098_S_6658,F,AD
+098_S_6734,F,CN
+099_S_6097,M,CN
+100_S_6308,M,CN
+100_S_6349,M,CN
+100_S_6713,F,AD
+109_S_6213,M,CN
+109_S_6363,M,MCI
+109_S_6373,F,MCI
+109_S_6376,F,MCI
+109_S_6405,F,CN
+114_S_6063,F,CN
+114_S_6347,M,AD
+114_S_6368,M,AD
+114_S_6595,M,AD
+116_S_6100,F,AD
+116_S_6458,M,CN
+116_S_6537,F,CN
+116_S_6543,M,AD
+116_S_6750,M,CN
+123_S_6118,F,CN
+123_S_6825,F,AD
+123_S_6891,M,AD
+126_S_6683,M,AD
+126_S_6721,M,AD
+126_S_6724,M,MCI
+126_S_7015,M,MCI
+126_S_7060,M,MCI
+126_S_7083,M,MCI
+127_S_6173,F,CN
+127_S_6330,M,CN
+127_S_6433,M,AD
+127_S_6512,M,MCI
+127_S_6549,M,AD
+129_S_6288,F,CN
+129_S_6457,M,CN
+129_S_6763,F,AD
+129_S_6784,M,AD
+129_S_6852,M,MCI
+130_S_6072,F,AD
+130_S_6111,F,CN
+130_S_6469,F,CN
+130_S_6558,F,CN
+130_S_6611,M,MCI
+131_S_6616,M,MCI
+131_S_7032,M,MCI
+135_S_6284,M,AD
+135_S_6389,M,AD
+135_S_6411,M,CN
+135_S_6544,F,MCI
+135_S_6545,F,AD
+135_S_6586,M,MCI
+135_S_6622,M,MCI
+135_S_6687,F,AD
+135_S_6840,F,AD
+135_S_7003,F,AD
+141_S_6061,F,CN
+141_S_6075,M,MCI
+141_S_6787,F,MCI
+141_S_6964,F,MCI
+168_S_6049,F,CN
+168_S_6065,F,CN
+168_S_6108,M,CN
+168_S_6142,F,AD
+168_S_6321,M,CN
+168_S_6426,F,MCI
+168_S_6591,F,MCI
+168_S_6735,F,AD
+168_S_6754,M,AD
+168_S_6827,M,AD
+168_S_6828,M,AD
+168_S_6843,F,AD
+168_S_6851,F,MCI
+168_S_6874,F,MCI
+168_S_6921,M,AD
+168_S_6938,M,AD
+177_S_6420,F,CN
+301_S_6501,F,CN
+301_S_6592,M,AD
+301_S_6615,F,MCI
+301_S_6811,F,MCI
+305_S_6438,F,CN
+305_S_6498,M,MCI
+305_S_6810,M,AD
+305_S_6850,M,AD
+305_S_6877,M,MCI
+305_S_6881,F,AD
+941_S_6017,M,MCI
+941_S_6068,M,MCI
+136_S_4993,F,AD
+136_S_4932,M,MCI
+136_S_4848,M,MCI
+136_S_4836,M,MCI
+136_S_4727,F,CN
+136_S_4726,F,CN
+136_S_4433,F,CN
+136_S_4408,M,MCI
+136_S_4269,F,CN
+131_S_5138,M,AD
+130_S_5231,F,AD
+130_S_5059,M,AD
+130_S_5006,F,AD
+130_S_4997,F,AD
+130_S_4990,F,AD
+130_S_4984,F,AD
+130_S_4982,F,AD
+130_S_4971,M,AD
+130_S_4925,M,MCI
+130_S_4817,M,MCI
+130_S_4730,F,AD
+130_S_4660,F,AD
+130_S_4641,F,AD
+130_S_4605,F,MCI
+130_S_4589,F,AD
+130_S_4542,F,MCI
+130_S_4352,M,CN
+130_S_4343,M,CN
+130_S_4294,F,MCI
+100_S_5246,F,CN
+100_S_5106,M,AD
+100_S_5091,M,CN
+100_S_4970,M,MCI
+100_S_4556,M,MCI
+100_S_4511,M,CN
+100_S_4469,M,CN
+053_S_5208,M,AD
+053_S_5070,M,AD
+053_S_4578,F,CN
+031_S_4721,M,MCI
+031_S_4590,F,MCI
+031_S_4496,F,CN
+031_S_4474,M,CN
+031_S_4218,M,CN
+031_S_4203,F,MCI
+031_S_4194,M,MCI
+031_S_4042,M,MCI
+031_S_4032,F,CN
+031_S_4029,M,MCI
+031_S_4024,F,AD
+031_S_4021,M,CN
+019_S_5019,F,AD
+019_S_5012,M,AD
+019_S_4835,M,CN
+019_S_4549,M,AD
+019_S_4548,M,MCI
+019_S_4477,F,AD
+019_S_4367,F,CN
+019_S_4293,M,MCI
+019_S_4252,F,AD
+018_S_5240,F,AD
+018_S_5074,F,AD
+018_S_4889,M,MCI
+018_S_4733,M,AD
+018_S_4696,F,AD
+018_S_4400,M,CN
+018_S_4349,F,CN
+018_S_4313,F,CN
+018_S_4257,M,CN
+013_S_5071,M,AD
+013_S_4985,F,MCI
+013_S_4731,F,CN
+013_S_4616,M,CN
+013_S_4595,M,MCI
+013_S_4580,F,CN
+013_S_4395,F,MCI
+013_S_4236,F,MCI
+010_S_5163,M,AD
+010_S_4442,F,CN
+010_S_4135,F,MCI
+006_S_4960,M,MCI
+006_S_4867,M,AD
+006_S_4713,M,MCI
+006_S_4546,M,AD
+006_S_4515,M,MCI
+006_S_4485,M,CN
+006_S_4449,F,CN
+006_S_4363,F,MCI
+006_S_4357,F,CN
+006_S_4346,M,MCI
+006_S_4192,M,AD
+006_S_4153,M,AD
+006_S_4150,M,CN
+002_S_4270,F,CN
+002_S_4264,F,CN
+002_S_4262,F,CN
+002_S_4251,M,MCI
+002_S_4229,M,MCI
+002_S_4219,F,MCI
+002_S_4213,F,CN
+002_S_4171,M,MCI

AtlaScore/downstream/docs/behavior_HCP.csv

@@ -0,0 +1,101 @@
+Subject,Release,Acquisition,Gender,Age,Age_Group,MMSE_Score,PMAT24_A_CR,CogFluidComp_AgeAdj,CogFluidComp_AgeAdj_Group,CogCrystalComp_AgeAdj,CogCrystalComp_AgeAdj_Group,CogTotalComp_AgeAdj,CogTotalComp_AgeAdj_Group
+100206,S900,Q11,M,26-30,1,30,20,130.07,2,113.89,1,133.76,2
+100307,Q1,Q01,F,26-30,1,29,17,107.88,1,97.42,1,106.85,1
+100408,Q3,Q03,M,31-35,2,30,7,113.77,1,134.57,2,133.96,2
+100610,S900,Q08,M,26-30,1,30,23,115.63,2,149.19,2,153.36,2
+101006,S500,Q06,F,31-35,2,28,11,72.02,0,82.93,0,72.15,0
+101107,S500,Q06,M,22-25,0,29,14,106.57,1,118.74,2,120.13,2
+101309,S500,Q06,M,26-30,1,29,15,90.01,1,110.72,1,100.32,1
+101915,Q3,Q04,F,31-35,2,29,21,87.99,1,117.73,2,103.22,1
+102008,S500,Q06,M,22-25,0,30,20,112.08,1,129.31,2,131.66,2
+102109,S1200,Q13,M,26-30,1,29,21,93.65,1,117.69,2,109.51,1
+102311,S500,Q06,F,26-30,1,30,20,93.44,1,107.45,1,101.53,1
+102513,S900,Q10,M,26-30,1,29,18,72.82,0,120.41,2,92.5,1
+102614,S1200,Q12,M,22-25,0,29,18,139.94,2,98.7,1,134.26,2
+102715,S1200,Q12,M,26-30,1,29,22,104.28,1,124.02,2,122.45,2
+102816,Q3,Q03,F,26-30,1,29,17,102.9,1,75.59,0,86.38,1
+103010,S1200,Q12,M,22-25,0,28,19,87.61,1,100.28,1,91.51,1
+103111,S500,Q06,M,26-30,1,26,21,100.56,1,117.11,2,113.45,1
+103212,S1200,Q13,M,31-35,2,26,15,94.11,1,124.11,2,114.23,1
+103414,Q2,Q02,F,22-25,0,29,17,84.22,0,99.49,1,88.44,1
+103515,Q1,Q02,F,26-30,1,27,8,77.63,0,71.99,0,69.84,0
+103818,Q1,Q01,F,31-35,2,28,16,95.9,1,124.18,2,115.58,2
+104012,MEG2,Q07,F,26-30,1,27,10,79.7,0,87.94,1,79.29,0
+104416,S900,Q09,F,31-35,2,30,23,131.74,2,125.37,2,143.78,2
+104820,S500,Q06,F,36+,2,27,15,101.58,1,77.78,0,86.97,1
+105014,S500,Q05,F,26-30,1,30,21,85.35,1,123.5,2,107.29,1
+105115,Q2,Q02,M,31-35,2,29,22,105.98,1,116.28,2,116.5,2
+105216,Q3,Q03,M,26-30,1,29,22,141.25,2,127.97,2,153.36,2
+105620,S900,Q08,F,31-35,2,30,22,136.5,2,120.16,2,143.87,2
+105923,MEG2,Q07,F,31-35,2,30,22,94.52,1,119.01,2,110.18,1
+106016,Q3,Q04,F,31-35,2,30,22,108.46,1,107.77,1,112.08,1
+106319,Q3,Q03,M,26-30,1,30,24,107.24,1,132.49,2,131.22,2
+106521,S500,Q06,F,26-30,1,30,23,131.41,2,134.58,2,153.36,2
+106824,S1200,Q13,M,22-25,0,30,22,124.28,2,121.7,2,135.54,2
+107018,S900,Q09,F,26-30,1,28,10,87.55,1,121.64,2,107.92,1
+107321,S500,Q04,F,22-25,0,30,23,133.33,2,128.4,2,153.36,2
+107422,S500,Q07,M,22-25,0,29,24,126.89,2,139.39,2,153.36,2
+107725,S900,Q08,F,31-35,2,29,10,101.54,1,80.79,0,88.74,1
+108020,S1200,Q12,M,22-25,0,30,21,108.5,1,106.25,1,112.78,1
+108121,S500,Q04,F,26-30,1,30,16,115.5,2,111.32,1,122.49,2
+108222,S900,Q07,M,31-35,2,29,22,134.37,2,124.34,2,146.49,2
+108323,S500,Q04,F,26-30,1,30,13,119.88,2,106.66,1,124.12,2
+108525,S500,Q05,M,22-25,0,30,18,67.64,0,87.61,1,72.18,0
+108828,S500,Q05,M,31-35,2,29,20,83.74,0,113.24,1,96.69,1
+109123,S500,Q06,M,31-35,2,30,22,138.61,2,123.43,2,153.36,2
+109325,S500,Q07,M,31-35,2,28,19,112.13,1,89.24,1,103.1,1
+109830,S900,Q08,F,31-35,2,28,,76.61,0,128.72,2,104.97,1
+110007,S900,Q11,F,31-35,2,30,18,108.12,1,95.69,1,104.22,1
+110411,Q2,Q02,M,31-35,2,29,16,122.5,2,124.37,2,134.04,2
+110613,S900,Q11,M,26-30,1,30,17,88.25,1,122.21,2,109.09,1
+111009,S900,Q03,F,26-30,1,30,20,,,,,,
+111211,S1200,Q12,F,26-30,1,30,19,107.41,1,122.72,2,123.48,2
+111312,Q1,Q01,F,31-35,2,28,17,118.47,2,78.3,0,101.2,1
+111413,S500,Q04,F,26-30,1,28,13,110.99,1,108.65,1,117.02,2
+111514,MEG2,Q02,M,31-35,2,29,17,,,,,,
+111716,Q3,Q04,F,31-35,2,29,22,92.62,1,124.93,2,113.39,1
+112112,S900,Q11,M,26-30,1,30,21,94.23,1,122.33,2,113.93,1
+112314,S900,Q11,F,26-30,1,27,18,101.73,1,110.66,1,109.66,1
+112516,S900,Q08,F,31-35,2,29,11,97.49,1,95.85,1,95.5,1
+112920,S900,Q10,M,31-35,2,30,17,115.57,2,115.82,2,122.8,2
+113215,Q3,Q03,F,26-30,1,30,17,132.08,2,125.77,2,146.88,2
+113316,S1200,Q12,M,22-25,0,29,20,95.84,1,113.65,1,107.75,1
+113619,Q2,Q02,F,31-35,2,30,15,100.19,1,123.11,2,118.51,2
+113922,S500,Q04,M,31-35,2,30,15,134.06,2,91.33,1,125.45,2
+114217,S900,Q08,F,26-30,1,30,21,128.13,2,127.38,2,141.91,2
+114318,S900,Q06,F,22-25,0,27,16,113.23,1,113.67,1,122.26,2
+114419,S500,Q07,M,31-35,2,30,13,113.1,1,111.34,1,118.24,2
+114621,S900,Q09,M,26-30,1,30,15,89.35,1,112.92,1,101.67,1
+114823,S900,Q07,F,31-35,2,28,22,134.48,2,133.1,2,153.36,2
+114924,Q1,Q01,M,26-30,1,29,4,86.17,1,75.08,0,76.08,0
+115017,S900,Q06,F,31-35,2,29,15,95.82,1,107.64,1,102.13,1
+115219,S900,Q09,M,31-35,2,28,18,119.95,2,126.58,2,133.55,2
+115320,Q2,Q02,F,31-35,2,28,14,94.78,1,87.1,1,88.56,1
+115724,S1200,Q12,F,22-25,0,29,10,113.45,1,82.77,0,102.85,1
+115825,S900,Q10,M,22-25,0,27,9,76.5,0,70.59,0,68.82,0
+116524,S500,Q05,M,26-30,1,29,13,85.74,1,74.77,0,75.57,0
+116726,S900,Q08,M,26-30,1,29,19,131.22,2,123.89,2,143.4,2
+117021,S1200,Q12,F,26-30,1,29,23,116.82,2,125.18,2,130.46,2
+117122,Q1,Q01,F,26-30,1,30,24,127.07,2,116.33,2,131.87,2
+117324,Q3,Q03,F,22-25,0,29,12,77.63,0,85.67,1,76.33,0
+117930,S900,Q10,F,31-35,2,30,13,88.12,1,122.75,2,108.49,1
+118023,S900,Q07,F,26-30,1,30,19,99.19,1,101.19,1,102.25,1
+118124,S900,Q07,F,31-35,2,28,16,92.55,1,84.34,0,84.66,0
+118225,S900,Q08,M,26-30,1,29,23,115.22,2,126.2,2,132.22,2
+118528,S500,Q04,F,26-30,1,30,8,106.92,1,107.59,1,112.46,1
+118730,Q2,Q03,M,22-25,0,30,10,96.74,1,99.71,1,99.08,1
+118831,S1200,Q13,M,22-25,0,27,17,109.26,1,137.29,2,135.38,2
+118932,Q1,Q02,M,26-30,1,29,19,99.3,1,113.38,1,110.42,1
+119025,S1200,Q12,M,26-30,1,30,20,102.59,1,131.98,2,127.91,2
+119126,S900,Q11,F,22-25,0,30,17,108.5,1,94.54,1,105.89,1
+119732,S900,Q10,F,31-35,2,28,22,78.61,0,117.14,2,95.4,1
+119833,Q1,Q01,F,26-30,1,29,12,105.97,1,93.51,1,102.68,1
+120111,Q3,Q03,F,26-30,1,28,14,88.41,1,119.68,2,107.01,1
+120212,Q1,Q01,F,31-35,2,29,22,117.91,2,123.06,2,130.39,2
+120414,S1200,Q12,F,26-30,1,27,12,69.33,0,104.91,1,80.31,0
+120515,S500,Q04,F,26-30,1,30,21,87.13,1,95.96,1,88.94,1
+120717,S900,Q09,F,31-35,2,29,20,128.67,2,123.49,2,138.26,2
+121416,S900,Q08,M,26-30,1,29,20,74.66,0,116.24,2,91.2,1
+121618,S500,Q05,M,31-35,2,27,11,93.24,1,122.8,2,112.11,1
+121921,S900,Q07,M,31-35,2,30,21,86.02,1,119.97,2,103.44,1
+122317,Q3,Q04,M,31-35,2,30,18,99.89,1,87.97,1,92.99,1

AtlaScore/downstream/downstream.py

@@ -0,0 +1,740 @@
+import os
+from boto3.session import Session
+import nibabel as nib
+from nilearn.image import resample_to_img
+from nilearn.datasets import fetch_abide_pcp
+import numpy as np
+import pandas as pd
+from scipy.signal import detrend
+from scipy.stats import zscore
+import shutil
+from sklearn.model_selection import StratifiedKFold, KFold
+from sklearn.decomposition import PCA
+from sklearn.svm import SVC
+from sklearn.metrics import accuracy_score
+from sklearn.preprocessing import StandardScaler
+
+
+
+def get_sub_HCP_rfMRI(subject, access_key, secret_key, addr = './nii_data'):
+
+    bucketName = 'hcp-openaccess'
+    prefix = 'HCP_1200/'
+    session = Session(aws_access_key_id = access_key, aws_secret_access_key = secret_key)
+    bucket = session.resource('s3').Bucket(bucketName)
+    os.makedirs(os.path.join(addr, 'HCP_rfMRI'), exist_ok = True)
+
+    for run in ['REST1_LR', 'REST1_RL', 'REST2_LR', 'REST2_RL']:
+
+        source_addr = '{}{}/MNINonLinear/Results/rfMRI_{}/rfMRI_{}_hp2000_clean.nii.gz'.format(prefix, subject, run, run)
+        target_addr = os.path.join(addr, 'HCP_rfMRI/{}_rfMRI_{}_hp2000_clean.nii.gz'.format(subject, run))
+
+        if not os.path.exists(target_addr):
+            try: bucket.download_file(source_addr, target_addr)
+            except: continue
+
+
+
+def get_sub_HCP_tfMRI(subject, access_key, secret_key, addr = './nii_data'):
+
+    bucketName = 'hcp-openaccess'
+    prefix = 'HCP_1200/'
+    session = Session(aws_access_key_id = access_key, aws_secret_access_key = secret_key)
+    bucket = session.resource('s3').Bucket(bucketName)
+    os.makedirs(os.path.join(addr, 'HCP_tfMRI'), exist_ok = True)
+
+    for task in ['WM', 'GAMBLING', 'MOTOR', 'LANGUAGE', 'SOCIAL', 'RELATIONAL', 'EMOTION']:
+
+        if task == 'WM': subtask_list = ['0bk_body', '0bk_faces', '0bk_places', '0bk_tools', '2bk_body', '2bk_faces', '2bk_places', '2bk_tools']
+        elif task == 'GAMBLING': subtask_list = ['win', 'loss']
+        elif task == 'MOTOR': subtask_list = ['cue', 'lf', 'rf', 'lh', 'rh', 't']
+        elif task == 'LANGUAGE': subtask_list = ['story', 'math']
+        elif task == 'SOCIAL': subtask_list = ['mental', 'rnd']
+        elif task == 'RELATIONAL': subtask_list = ['relation', 'match']
+        else: subtask_list = ['fear', 'neut']
+
+        source_addr = '{}{}/MNINonLinear/Results/tfMRI_{}_LR/tfMRI_{}_LR.nii.gz'.format(prefix, subject, task, task)
+        target_addr = os.path.join(addr, 'HCP_tfMRI/{}_tfMRI_{}_LR.nii.gz'.format(subject, task))
+
+        if not os.path.exists(target_addr):
+            try: bucket.download_file(source_addr, target_addr)
+            except: continue
+        
+        for subtask in subtask_list:
+
+            subtask_source_addr = '{}{}/MNINonLinear/Results/tfMRI_{}_LR/EVs/{}.txt'.format(prefix, subject, task, subtask)
+            subtask_target_addr = os.path.join(addr, 'HCP_tfMRI/{}_tfMRI_{}_{}_ev.txt'.format(subject, task, subtask))
+
+            if not os.path.exists(subtask_target_addr):
+                try: bucket.download_file(subtask_source_addr, subtask_target_addr)
+                except: continue
+
+
+
+def get_ABIDE(addr = './nii_data'):
+
+    fetch_abide_pcp(data_dir = addr, band_pass_filtering = True)
+    os.remove(os.path.join(addr, 'README.md'))
+    os.remove(os.path.join(addr, 'ABIDE_pcp/Phenotypic_V1_0b_preprocessed1.csv'))
+    shutil.move(os.path.join(addr, 'ABIDE_pcp/cpac/filt_noglobal'), os.path.join(addr, 'ABIDE'))
+    shutil.rmtree(os.path.join(addr, 'ABIDE_pcp'))
+
+
+
+def get_fc_HCP_rfMRI(subject, atlas_name, atlas_loc, nii_addr = './nii_data', fc_addr = './fc_data'):
+
+    os.makedirs(os.path.join(fc_addr, 'HCP_rfMRI', atlas_name), exist_ok = True)
+
+    for run in ['REST1_LR', 'REST1_RL', 'REST2_LR', 'REST2_RL']:
+    
+        rfmri_img = nib.load(os.path.join(nii_addr, 'HCP_rfMRI', '{}_rfMRI_{}_hp2000_clean.nii.gz'.format(subject, run)))
+        rfmri_data = rfmri_img.get_fdata(); X = rfmri_data.reshape(-1, rfmri_data.shape[-1])
+        atlas = nib.load(atlas_loc); atlas = resample_to_img(atlas, rfmri_img, interpolation = 'nearest')
+        labels = atlas.get_fdata().astype(int).flatten()
+        time_series = np.stack([X[labels == label, :].mean(axis = 0) for label in np.sort(np.unique(labels[labels > 0]))], axis = 1)
+        time_series = zscore(detrend(time_series, axis = 0, type = 'linear'), axis = 0, ddof = 1)
+
+        for i in range(int(X.shape[1]/300)):
+
+            fc = np.corrcoef(time_series[i*300:(i+1)*300].T); n = fc.shape[0]; fc = fc[np.triu_indices(n, k = 1)]
+            if not np.isnan(fc).any(): np.savez_compressed(os.path.join(fc_addr, 'HCP_rfMRI', '{}/{}_{}_{}.npz'.format(atlas_name, subject, run, i)), fc)
+
+
+
+def get_fc_HCP_tfMRI(subject, atlas_name, atlas_loc, nii_addr = './nii_data', fc_addr = './fc_data'):
+
+    os.makedirs(os.path.join(fc_addr, 'HCP_tfMRI', atlas_name), exist_ok = True)
+
+    for task in ['WM', 'GAMBLING', 'MOTOR', 'LANGUAGE', 'SOCIAL', 'RELATIONAL', 'EMOTION']:
+
+        if task == 'WM': subtask_list = ['0bk_body', '0bk_faces', '0bk_places', '0bk_tools', '2bk_body', '2bk_faces', '2bk_places', '2bk_tools']
+        elif task == 'GAMBLING': subtask_list = ['win', 'loss']
+        elif task == 'MOTOR': subtask_list = ['cue', 'lf', 'rf', 'lh', 'rh', 't']
+        elif task == 'LANGUAGE': subtask_list = ['story', 'math']
+        elif task == 'SOCIAL': subtask_list = ['mental', 'rnd']
+        elif task == 'RELATIONAL': subtask_list = ['relation', 'match']
+        else: subtask_list = ['fear', 'neut']
+
+        tfmri_img = nib.load(os.path.join(nii_addr, 'HCP_tfMRI', '{}_tfMRI_{}_LR.nii.gz'.format(subject, task)))
+        tfmri_data = tfmri_img.get_fdata(); X = tfmri_data.reshape(-1, tfmri_data.shape[-1])
+        atlas = nib.load(atlas_loc); atlas = resample_to_img(atlas, tfmri_img, interpolation = 'nearest')
+        labels = atlas.get_fdata().astype(int).flatten()
+        time_series = np.stack([X[labels == label, :].mean(axis = 0) for label in np.sort(np.unique(labels[labels > 0]))], axis = 1)
+        time_series = zscore(detrend(time_series, axis = 0, type = 'linear'), axis = 0, ddof = 1)
+
+        fc = np.corrcoef(time_series.T); n = fc.shape[0]; fc = fc[np.triu_indices(n, k = 1)]
+        if not np.isnan(fc).any(): np.savez_compressed(os.path.join(fc_addr, 'HCP_tfMRI', '{}/{}_{}_LR.npz'.format(atlas_name, subject, task)), fc)
+
+        timepoint = np.arange(X.shape[-1]) * 0.72
+
+        for subtask in subtask_list:
+
+            if not os.path.exists(os.path.join(nii_addr, 'HCP_tfMRI', '{}_tfMRI_{}_{}_ev.txt'.format(subject, task, subtask))): continue
+
+            task_timepoint = np.loadtxt(os.path.join(nii_addr, 'HCP_tfMRI', '{}_tfMRI_{}_{}_ev.txt'.format(subject, task, subtask))).reshape(-1, 3)
+            subtask_time_series = np.vstack([time_series[(timepoint > task_timepoint[i, 0] - 1e-3) & (timepoint < task_timepoint[i, 0] + task_timepoint[i, 1] + 1e-3)] for i in range(task_timepoint.shape[0])])
+            fc = np.corrcoef(subtask_time_series.T); n = fc.shape[0]; fc = fc[np.triu_indices(n, k = 1)]
+            if not np.isnan(fc).any(): np.savez_compressed(os.path.join(fc_addr, 'HCP_tfMRI', '{}/{}_{}_LR_{}.npz'.format(atlas_name, subject, task, subtask)), fc)
+
+
+
+def get_fc_ABIDE(atlas_name, atlas_loc, nii_addr = './nii_data', fc_addr = './fc_data'):
+
+    os.makedirs(os.path.join(fc_addr, 'ABIDE', atlas_name), exist_ok = True)
+
+    for filename in os.listdir(os.path.join(nii_addr, 'ABIDE')):
+
+        prefix = filename[:-20]
+
+        rfmri_img = nib.load(os.path.join(nii_addr, 'ABIDE', filename))
+        rfmri_data = rfmri_img.get_fdata(); X = rfmri_data.reshape(-1, rfmri_data.shape[-1])
+        atlas = nib.load(atlas_loc); atlas = resample_to_img(atlas, rfmri_img, interpolation = 'nearest')
+        labels = atlas.get_fdata().astype(int).flatten()
+        time_series = np.stack([X[labels == label, :].mean(axis = 0) for label in np.sort(np.unique(labels[labels > 0]))], axis = 1)
+        time_series = zscore(detrend(time_series, axis = 0, type = 'linear'), axis = 0, ddof = 1)
+        fc = np.corrcoef(time_series.T); n = fc.shape[0]; fc = fc[np.triu_indices(n, k = 1)]
+        if not np.isnan(fc).any(): np.savez_compressed(os.path.join(fc_addr, 'ABIDE', '{}/{}.npz'.format(atlas_name, prefix)), fc)
+
+
+
+def get_fc_ADNI(atlas_name, atlas_loc, nii_addr = './nii_data', fc_addr = './fc_data'):
+
+    os.makedirs(os.path.join(fc_addr, 'ADNI', atlas_name), exist_ok = True)
+
+    for prefix in os.listdir(os.path.join(nii_addr, 'ADNI')):
+
+        rfmri_img = nib.load(os.path.join(nii_addr, 'ADNI', prefix, 'Filtered_4DVolume.nii'))
+        rfmri_data = rfmri_img.get_fdata(); X = rfmri_data.reshape(-1, rfmri_data.shape[-1])
+        atlas = nib.load(atlas_loc); atlas = resample_to_img(atlas, rfmri_img, interpolation = 'nearest')
+        labels = atlas.get_fdata().astype(int).flatten()
+        time_series = np.stack([X[labels == label, :].mean(axis = 0) for label in np.sort(np.unique(labels[labels > 0]))], axis = 1)
+        time_series = zscore(detrend(time_series, axis = 0, type = 'linear'), axis = 0, ddof = 1)
+        fc = np.corrcoef(time_series.T); n = fc.shape[0]; fc = fc[np.triu_indices(n, k = 1)]
+        if not np.isnan(fc).any(): np.savez_compressed(os.path.join(fc_addr, 'ADNI', '{}/{}.npz'.format(atlas_name, prefix)), fc)
+
+
+
+def fc_stability(atlas_name, fc_addr = './fc_data'):
+
+    subjlist = np.loadtxt('./docs/HCP_subjlist.txt', dtype = str)
+    addr = os.path.join(fc_addr, 'HCP_rfMRI', atlas_name)
+
+    res_list = []
+    for sub in subjlist:
+        if len([file for file in os.listdir(addr) if sub in file]) > 1:
+            fc_corr = np.corrcoef(np.array([np.load(os.path.join(addr, file))['arr_0'] for file in os.listdir(addr) if sub in file]))
+            res_list.append(np.mean(fc_corr[np.triu_indices(fc_corr.shape[0], k = 1)]))
+    
+    return np.array(res_list)
+
+
+
+def fingerprinting(atlas_name, fc_addr = './fc_data'):
+
+    addr = os.path.join(fc_addr, 'HCP_rfMRI', atlas_name)
+    subjlist = np.loadtxt('./docs/HCP_subjlist.txt', dtype = str)
+    subjlist = [sub for sub in subjlist if [file for file in os.listdir(addr) if sub in file] != []]
+
+    reference = [np.load(os.path.join(addr, [file for file in os.listdir(addr) if sub in file][0]))['arr_0'] for sub in subjlist]
+    res_list = []
+
+    for sub in subjlist:
+        if len([file for file in os.listdir(addr) if sub in file]) == 1: continue
+        file_list = [file for file in os.listdir(addr) if sub in file][1:]
+        fc_list = [np.load(os.path.join(addr, file))['arr_0'] for file in file_list]
+        count = 0
+        for idx in range(len(file_list)):
+            temp = np.array([np.corrcoef([fc_list[idx], ref])[0, 1] for ref in reference])
+            if subjlist[np.where(temp == np.max(temp))[0][0]] == sub: count += 1
+        res_list.append(count/len(fc_list))
+
+    return np.array(res_list)
+
+
+
+def age_group_classification(atlas_name, fc_addr = './fc_data'):
+
+    subjlist = np.loadtxt('./docs/HCP_subjlist.txt', dtype = str)
+    behavior = pd.read_csv('./docs/behavior_HCP.csv').to_dict(orient = 'list')
+
+    addr = os.path.join(fc_addr, 'HCP_rfMRI', atlas_name)
+
+    sub_X = []; sub_y = []; mat_X = []; mat_y = []; mat_master = []
+    for sub in subjlist:
+        target = behavior['Age_Group'][behavior['Subject'].index(int(sub))]
+        if np.isnan(target): continue
+        sub_X.append(int(sub)); sub_y.append(target)
+        file_list = [np.load(os.path.join(addr, file))['arr_0'] for file in os.listdir(addr) if sub in file]
+        for f in file_list: mat_X.append(f); mat_y.append(target); mat_master.append(int(sub))
+    sub_X = np.array(sub_X); sub_y = np.array(sub_y); mat_X = np.array(mat_X); mat_y = np.array(mat_y); mat_master = np.array(mat_master)
+
+    skf = StratifiedKFold(n_splits = 10, shuffle = True, random_state = 0)
+    acc_list = []
+
+    for trainsub_idx, testsub_idx in skf.split(sub_X, sub_y):
+
+        train_sub, test_sub = sub_X[trainsub_idx], sub_X[testsub_idx]
+        trainmat_idx = np.array([(train_sub == master).any() for master in mat_master]); testmat_idx = np.array([(test_sub == master).any() for master in mat_master])
+        X_train, X_test = mat_X[trainmat_idx], mat_X[testmat_idx]; y_train, y_test = mat_y[trainmat_idx], mat_y[testmat_idx]
+
+        if X_train.shape[1] > 100:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+            pca = PCA(n_components = 100, random_state = 0); X_train_pca = pca.fit_transform(X_train_scaled); X_test_pca = pca.transform(X_test_scaled)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_pca, y_train)
+
+            y_pred = clf.predict(X_test_pca)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+        else:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_scaled, y_train)
+
+            y_pred = clf.predict(X_test_scaled)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+    return np.array(acc_list)
+
+
+
+def gender_classification(atlas_name, fc_addr = './fc_data'):
+
+    subjlist = np.loadtxt('./docs/HCP_subjlist.txt', dtype = str)
+    behavior = pd.read_csv('./docs/behavior_HCP.csv').to_dict(orient = 'list')
+
+    addr = os.path.join(fc_addr, 'HCP_rfMRI', atlas_name)
+
+    sub_X = []; sub_y = []; mat_X = []; mat_y = []; mat_master = []
+    for sub in subjlist:
+        target = int(behavior['Gender'][behavior['Subject'].index(int(sub))] == 'M')
+        if np.isnan(target): continue
+        sub_X.append(int(sub)); sub_y.append(target)
+        file_list = [np.load(os.path.join(os.path.join(addr, file)))['arr_0'] for file in os.listdir(addr) if sub in file]
+        for f in file_list: mat_X.append(f); mat_y.append(target); mat_master.append(int(sub))
+    sub_X = np.array(sub_X); sub_y = np.array(sub_y); mat_X = np.array(mat_X); mat_y = np.array(mat_y); mat_master = np.array(mat_master)
+
+    skf = StratifiedKFold(n_splits = 10, shuffle = True, random_state = 0)
+    acc_list = []
+
+    for trainsub_idx, testsub_idx in skf.split(sub_X, sub_y):
+
+        train_sub, test_sub = sub_X[trainsub_idx], sub_X[testsub_idx]
+        trainmat_idx = np.array([(train_sub == master).any() for master in mat_master]); testmat_idx = np.array([(test_sub == master).any() for master in mat_master])
+        X_train, X_test = mat_X[trainmat_idx], mat_X[testmat_idx]; y_train, y_test = mat_y[trainmat_idx], mat_y[testmat_idx]
+
+        if X_train.shape[1] > 100:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+            pca = PCA(n_components = 100, random_state = 0); X_train_pca = pca.fit_transform(X_train_scaled); X_test_pca = pca.transform(X_test_scaled)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_pca, y_train)
+
+            y_pred = clf.predict(X_test_pca)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+        else:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_scaled, y_train)
+
+            y_pred = clf.predict(X_test_scaled)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+    return np.array(acc_list)
+
+
+
+def fluid_intelligence(atlas_name, fc_addr = './fc_data'):
+
+    subjlist = np.loadtxt('./docs/HCP_subjlist.txt', dtype = str)
+    behavior = pd.read_csv('./docs/behavior_HCP.csv').to_dict(orient = 'list')
+
+    addr = os.path.join(fc_addr, 'HCP_rfMRI', atlas_name)
+
+    sub_X = []; sub_y = []; mat_X = []; mat_y = []; mat_master = []
+    for sub in subjlist:
+        target = behavior['CogFluidComp_AgeAdj_Group'][behavior['Subject'].index(int(sub))]
+        if np.isnan(target): continue
+        sub_X.append(int(sub)); sub_y.append(target)
+        file_list = [np.load(os.path.join(addr, file))['arr_0'] for file in os.listdir(addr) if sub in file]
+        for f in file_list: mat_X.append(f); mat_y.append(target); mat_master.append(int(sub))
+    sub_X = np.array(sub_X); sub_y = np.array(sub_y); mat_X = np.array(mat_X); mat_y = np.array(mat_y); mat_master = np.array(mat_master)
+
+    skf = StratifiedKFold(n_splits = 10, shuffle = True, random_state = 0)
+    acc_list = []
+
+    for trainsub_idx, testsub_idx in skf.split(sub_X, sub_y):
+
+        train_sub, test_sub = sub_X[trainsub_idx], sub_X[testsub_idx]
+        trainmat_idx = np.array([(train_sub == master).any() for master in mat_master]); testmat_idx = np.array([(test_sub == master).any() for master in mat_master])
+        X_train, X_test = mat_X[trainmat_idx], mat_X[testmat_idx]; y_train, y_test = mat_y[trainmat_idx], mat_y[testmat_idx]
+
+        if X_train.shape[1] > 100:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+            pca = PCA(n_components = 100, random_state = 0); X_train_pca = pca.fit_transform(X_train_scaled); X_test_pca = pca.transform(X_test_scaled)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_pca, y_train)
+
+            y_pred = clf.predict(X_test_pca)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+        else:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_scaled, y_train)
+
+            y_pred = clf.predict(X_test_scaled)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+    return np.array(acc_list)
+
+
+
+def crystallized_intelligence(atlas_name, fc_addr = './fc_data'):
+
+    subjlist = np.loadtxt('./docs/HCP_subjlist.txt', dtype = str)
+    behavior = pd.read_csv('./docs/behavior_HCP.csv').to_dict(orient = 'list')
+
+    addr = os.path.join(fc_addr, 'HCP_rfMRI', atlas_name)
+
+    sub_X = []; sub_y = []; mat_X = []; mat_y = []; mat_master = []
+    for sub in subjlist:
+        target = behavior['CogCrystalComp_AgeAdj_Group'][behavior['Subject'].index(int(sub))]
+        if np.isnan(target): continue
+        sub_X.append(int(sub)); sub_y.append(target)
+        file_list = [np.load(os.path.join(addr, file))['arr_0'] for file in os.listdir(addr) if sub in file]
+        for f in file_list: mat_X.append(f); mat_y.append(target); mat_master.append(int(sub))
+    sub_X = np.array(sub_X); sub_y = np.array(sub_y); mat_X = np.array(mat_X); mat_y = np.array(mat_y); mat_master = np.array(mat_master)
+
+    skf = StratifiedKFold(n_splits = 10, shuffle = True, random_state = 0)
+    acc_list = []
+
+    for trainsub_idx, testsub_idx in skf.split(sub_X, sub_y):
+
+        train_sub, test_sub = sub_X[trainsub_idx], sub_X[testsub_idx]
+        trainmat_idx = np.array([(train_sub == master).any() for master in mat_master]); testmat_idx = np.array([(test_sub == master).any() for master in mat_master])
+        X_train, X_test = mat_X[trainmat_idx], mat_X[testmat_idx]; y_train, y_test = mat_y[trainmat_idx], mat_y[testmat_idx]
+
+        if X_train.shape[1] > 100:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+            pca = PCA(n_components = 100, random_state = 0); X_train_pca = pca.fit_transform(X_train_scaled); X_test_pca = pca.transform(X_test_scaled)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_pca, y_train)
+
+            y_pred = clf.predict(X_test_pca)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+        else:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_scaled, y_train)
+
+            y_pred = clf.predict(X_test_scaled)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+    return np.array(acc_list)
+
+
+
+def general_intelligence(atlas_name, fc_addr = './fc_data'):
+
+    subjlist = np.loadtxt('./docs/HCP_subjlist.txt', dtype = str)
+    behavior = pd.read_csv('./docs/behavior_HCP.csv').to_dict(orient = 'list')
+
+    addr = os.path.join(fc_addr, 'HCP_rfMRI', atlas_name)
+
+    sub_X = []; sub_y = []; mat_X = []; mat_y = []; mat_master = []
+    for sub in subjlist:
+        target = behavior['CogTotalComp_AgeAdj_Group'][behavior['Subject'].index(int(sub))]
+        if np.isnan(target): continue
+        sub_X.append(int(sub)); sub_y.append(target)
+        file_list = [np.load(os.path.join(addr, file))['arr_0'] for file in os.listdir(addr) if sub in file]
+        for f in file_list: mat_X.append(f); mat_y.append(target); mat_master.append(int(sub))
+    sub_X = np.array(sub_X); sub_y = np.array(sub_y); mat_X = np.array(mat_X); mat_y = np.array(mat_y); mat_master = np.array(mat_master)
+
+    skf = StratifiedKFold(n_splits = 10, shuffle = True, random_state = 0)
+    acc_list = []
+
+    for trainsub_idx, testsub_idx in skf.split(sub_X, sub_y):
+
+        train_sub, test_sub = sub_X[trainsub_idx], sub_X[testsub_idx]
+        trainmat_idx = np.array([(train_sub == master).any() for master in mat_master]); testmat_idx = np.array([(test_sub == master).any() for master in mat_master])
+        X_train, X_test = mat_X[trainmat_idx], mat_X[testmat_idx]; y_train, y_test = mat_y[trainmat_idx], mat_y[testmat_idx]
+
+        if X_train.shape[1] > 100:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+            pca = PCA(n_components = 100, random_state = 0); X_train_pca = pca.fit_transform(X_train_scaled); X_test_pca = pca.transform(X_test_scaled)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_pca, y_train)
+
+            y_pred = clf.predict(X_test_pca)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+        else:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_scaled, y_train)
+
+            y_pred = clf.predict(X_test_scaled)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+    return np.array(acc_list)
+
+
+
+def cognitive_task_7way(atlas_name, fc_addr = './fc_data'):
+
+    subjlist = np.loadtxt('./docs/HCP_subjlist.txt', dtype = str)
+    addr = os.path.join(fc_addr, 'HCP_tfMRI', atlas_name)
+    task = ['WM', 'GAMBLING', 'MOTOR', 'LANGUAGE', 'SOCIAL', 'RELATIONAL', 'EMOTION']
+
+    mat_X = []; mat_y = []; mat_master = []
+    for sub in subjlist:
+        for t in task:
+            if not os.path.exists(os.path.join(addr, '{}_{}_LR.npz'.format(sub, t))): continue
+            mat_X.append(np.load(os.path.join(addr, '{}_{}_LR.npz'.format(sub, t)))['arr_0']); mat_y.append(task.index(t)); mat_master.append(sub)
+    mat_X = np.array(mat_X); mat_y = np.array(mat_y); mat_master = np.array(mat_master)
+
+    kf = KFold(n_splits = 10, shuffle = True, random_state = 0)
+    acc_list = []
+
+    for trainsub_idx, testsub_idx in kf.split(subjlist):
+
+        train_sub, test_sub = subjlist[trainsub_idx], subjlist[testsub_idx]
+        trainmat_idx = np.array([(train_sub == master).any() for master in mat_master]); testmat_idx = np.array([(test_sub == master).any() for master in mat_master])
+        X_train, X_test = mat_X[trainmat_idx], mat_X[testmat_idx]; y_train, y_test = mat_y[trainmat_idx], mat_y[testmat_idx]
+
+        if X_train.shape[1] > 100:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+            pca = PCA(n_components = 100, random_state = 0); X_train_pca = pca.fit_transform(X_train_scaled); X_test_pca = pca.transform(X_test_scaled)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_pca, y_train)
+
+            y_pred = clf.predict(X_test_pca)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+        else:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_scaled, y_train)
+
+            y_pred = clf.predict(X_test_scaled)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+    return np.array(acc_list)
+
+
+
+def cognitive_task_24way(atlas_name, fc_addr = './fc_data'):
+
+    subjlist = np.loadtxt('./docs/HCP_subjlist.txt', dtype = str)
+    addr = os.path.join(fc_addr, 'HCP_tfMRI', atlas_name)
+    task = {
+        '0bk_body': 'WM', '0bk_faces': 'WM', '0bk_places': 'WM', '0bk_tools': 'WM', '2bk_body': 'WM', '2bk_faces': 'WM', '2bk_places': 'WM', '2bk_tools': 'WM', 
+        'win': 'GAMBLING', 'loss': 'GAMBLING', 
+        'cue': 'MOTOR', 'lf': 'MOTOR', 'rf': 'MOTOR', 'lh': 'MOTOR', 'rh': 'MOTOR', 't': 'MOTOR', 
+        'story': 'LANGUAGE', 'math': 'LANGUAGE', 
+        'mental': 'SOCIAL', 'rnd': 'SOCIAL', 
+        'relation': 'RELATIONAL', 'match': 'RELATIONAL', 
+        'fear': 'EMOTION', 'neut': 'EMOTION'
+    }
+
+    mat_X = []; mat_y = []; mat_master = []
+    for sub in subjlist:
+        for t in list(task.keys()):
+            if not os.path.exists(os.path.join(addr, '{}_{}_LR_{}.npz'.format(sub, task[t], t))): continue
+            mat_X.append(np.load(os.path.join(addr, '{}_{}_LR_{}.npz'.format(sub, task[t], t)))['arr_0']); mat_y.append(list(task.keys()).index(t)); mat_master.append(sub)
+    mat_X = np.array(mat_X); mat_y = np.array(mat_y); mat_master = np.array(mat_master)
+
+    kf = KFold(n_splits = 10, shuffle = True, random_state = 0)
+    acc_list = []
+
+    for trainsub_idx, testsub_idx in kf.split(subjlist):
+
+        train_sub, test_sub = subjlist[trainsub_idx], subjlist[testsub_idx]
+        trainmat_idx = np.array([(train_sub == master).any() for master in mat_master]); testmat_idx = np.array([(test_sub == master).any() for master in mat_master])
+        X_train, X_test = mat_X[trainmat_idx], mat_X[testmat_idx]; y_train, y_test = mat_y[trainmat_idx], mat_y[testmat_idx]
+
+        if X_train.shape[1] > 100:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+            pca = PCA(n_components = 100, random_state = 0); X_train_pca = pca.fit_transform(X_train_scaled); X_test_pca = pca.transform(X_test_scaled)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_pca, y_train)
+
+            y_pred = clf.predict(X_test_pca)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+        else:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_scaled, y_train)
+
+            y_pred = clf.predict(X_test_scaled)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+    return np.array(acc_list)
+
+
+
+def autism_diagnosis(atlas_name, fc_addr = './fc_data'):
+
+    behavior = pd.read_csv('./docs/behavior_ABIDE.csv').to_dict(orient = 'list')
+    addr = os.path.join(fc_addr, 'ABIDE', atlas_name)
+
+    mat_X = []; mat_y = []
+    for file in os.listdir(addr):
+        target = behavior['DX_GROUP'][behavior['FILE_ID'].index(file[:-4])]
+        if np.isnan(target): continue
+        mat_X.append(np.load(os.path.join(addr, file))['arr_0']); mat_y.append(target)
+    mat_X = np.array(mat_X); mat_y = np.array(mat_y)
+
+    skf = StratifiedKFold(n_splits = 10, shuffle = True, random_state = 0)
+    acc_list = []
+
+    for trainmat_idx, testmat_idx in skf.split(mat_X, mat_y):
+
+        X_train, X_test = mat_X[trainmat_idx], mat_X[testmat_idx]; y_train, y_test = mat_y[trainmat_idx], mat_y[testmat_idx]
+
+        if X_train.shape[1] > 100:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+            pca = PCA(n_components = 100, random_state = 0); X_train_pca = pca.fit_transform(X_train_scaled); X_test_pca = pca.transform(X_test_scaled)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_pca, y_train)
+
+            y_pred = clf.predict(X_test_pca)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+        else:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_scaled, y_train)
+
+            y_pred = clf.predict(X_test_scaled)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+    return np.array(acc_list)
+
+
+
+def autism_cross_site(atlas_name, fc_addr = './fc_data'):
+
+    behavior = pd.read_csv('./docs/behavior_ABIDE.csv').to_dict(orient = 'list')
+    addr = os.path.join(fc_addr, 'ABIDE', atlas_name)
+
+    mat_X = []; mat_y = []; mat_master = []
+    for file in os.listdir(addr):
+        target = behavior['DX_GROUP'][behavior['FILE_ID'].index(file[:-4])]
+        site = behavior['SITE_ID'][behavior['FILE_ID'].index(file[:-4])]
+        if np.isnan(target): continue
+        mat_X.append(np.load(os.path.join(addr, file))['arr_0']); mat_y.append(target); mat_master.append(site)
+    mat_X = np.array(mat_X); mat_y = np.array(mat_y); mat_master = np.array(mat_master, dtype = str)
+
+    acc_list = []
+
+    for holdout_site in np.unique(mat_master):
+
+        trainmat_idx = np.array([(master != holdout_site) for master in mat_master]); testmat_idx = np.array([(master == holdout_site) for master in mat_master])
+        X_train, X_test = mat_X[trainmat_idx], mat_X[testmat_idx]; y_train, y_test = mat_y[trainmat_idx], mat_y[testmat_idx]
+
+        if X_train.shape[1] > 100:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+            pca = PCA(n_components = 100, random_state = 0); X_train_pca = pca.fit_transform(X_train_scaled); X_test_pca = pca.transform(X_test_scaled)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_pca, y_train)
+
+            y_pred = clf.predict(X_test_pca)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+        else:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_scaled, y_train)
+
+            y_pred = clf.predict(X_test_scaled)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+    return np.array(acc_list)
+
+
+
+def AD_diagnosis(atlas_name, fc_addr = './fc_data'):
+
+    behavior = pd.read_csv('./docs/behavior_ADNI.csv').to_dict(orient = 'list')
+    addr = os.path.join(fc_addr, 'ADNI', atlas_name)
+
+    mat_X = []; mat_y = []
+    for file in os.listdir(addr):
+        target = behavior['Research Group'][behavior['Subject ID'].index(file[:-4])]
+        if target == 'CN': target_idx = 0
+        elif target == 'MCI': target_idx = 1
+        else: target_idx = 2
+        mat_X.append(np.load(os.path.join(addr, file))['arr_0']); mat_y.append(target_idx)
+    mat_X = np.array(mat_X); mat_y = np.array(mat_y)
+
+    skf = StratifiedKFold(n_splits = 10, shuffle = True, random_state = 0)
+    acc_list = []
+
+    for trainmat_idx, testmat_idx in skf.split(mat_X, mat_y):
+
+        X_train, X_test = mat_X[trainmat_idx], mat_X[testmat_idx]; y_train, y_test = mat_y[trainmat_idx], mat_y[testmat_idx]
+
+        if X_train.shape[1] > 100:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+            pca = PCA(n_components = 100, random_state = 0); X_train_pca = pca.fit_transform(X_train_scaled); X_test_pca = pca.transform(X_test_scaled)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_pca, y_train)
+
+            y_pred = clf.predict(X_test_pca)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+        else:
+
+            scaler = StandardScaler(); X_train_scaled = scaler.fit_transform(X_train); X_test_scaled = scaler.transform(X_test)
+
+            clf = SVC(kernel = 'linear', class_weight = 'balanced')
+            clf.fit(X_train_scaled, y_train)
+
+            y_pred = clf.predict(X_test_scaled)
+            acc_list.append(accuracy_score(y_test, y_pred))
+
+    return np.array(acc_list)
+
+
+
+def downstream_all(atlas_name, fc_addr = './fc_data'):
+
+    print('--- {} downstream report ---'.format(atlas_name))
+
+    res = gender_classification(atlas_name, fc_addr)
+    print('Gender classification: {:.3f}±{:.3f}'.format(np.mean(res), np.std(res)))
+
+    res = fluid_intelligence(atlas_name, fc_addr)
+    print('Fluid intelligence: {:.3f}±{:.3f}'.format(np.mean(res), np.std(res)))
+
+    res = cognitive_task_7way(atlas_name, fc_addr)
+    print('Cognitive task (7-way): {:.3f}±{:.3f}'.format(np.mean(res), np.std(res)))
+
+    res = cognitive_task_24way(atlas_name, fc_addr)
+    print('Cognitive task (24-way): {:.3f}±{:.3f}'.format(np.mean(res), np.std(res)))
+
+    res = autism_diagnosis(atlas_name, fc_addr)
+    print('Autism diagnosis: {:.3f}±{:.3f}'.format(np.mean(res), np.std(res)))
+
+    res = AD_diagnosis(atlas_name, fc_addr)
+    print('AD diagnosis: {:.3f}±{:.3f}'.format(np.mean(res), np.std(res)))
+
+    res = fc_stability(atlas_name, fc_addr)
+    print('FC stability: {:.3f}±{:.3f}'.format(np.mean(res), np.std(res)))
+
+    res = fingerprinting(atlas_name, fc_addr)
+    print('Fingerprinting: {:.3f}±{:.3f}'.format(np.mean(res), np.std(res)))
+
+    res = age_group_classification(atlas_name, fc_addr)
+    print('Age group classification: {:.3f}±{:.3f}'.format(np.mean(res), np.std(res)))
+
+    res = crystallized_intelligence(atlas_name, fc_addr)
+    print('Crystallized intelligence: {:.3f}±{:.3f}'.format(np.mean(res), np.std(res)))
+
+    res = general_intelligence(atlas_name, fc_addr)
+    print('General intelligence: {:.3f}±{:.3f}'.format(np.mean(res), np.std(res)))
+
+    res = autism_cross_site(atlas_name, fc_addr)
+    print('Autism cross-site: {:.3f}±{:.3f}'.format(np.mean(res), np.std(res)))

AtlaScore/downstream/fc_data/fc_data.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2635230a3252baf502aa451b5d5289cc1c6d5fc41c476faba3a916eb4c0e707e
+size 215604832

AtlaScore/similarity/compute_adj.py

@@ -0,0 +1,67 @@
+import numpy as np
+import nibabel as nib
+from scipy.ndimage import generate_binary_structure, binary_dilation
+
+
+def calculate_label_adjacency_bg(atlas_file, exclude_zero=True):
+    if atlas_file.endswith('nii.gz'):
+        nii = nib.load(atlas_file)
+        data = nii.get_fdata().astype(np.int32)
+    else:
+        data = np.load(atlas_file)
+    
+    unique_labels = np.unique(data)
+    bg_label = np.min(unique_labels)
+    print('background value: ', bg_label)
+
+    if exclude_zero:
+        unique_labels = unique_labels[unique_labels != bg_label]
+    
+    adjacency_dict = {label: set() for label in unique_labels}
+    
+    struct = generate_binary_structure(3, 3)
+    
+    for label in unique_labels:
+        mask = (data == label)
+        
+        dilated = binary_dilation(mask, structure=struct)
+        border = dilated & ~mask  
+        
+        adjacent_labels = np.unique(data[border])
+        if exclude_zero:
+            adjacent_labels = adjacent_labels[adjacent_labels != 0]
+        
+        adjacency_dict[label].update(adjacent_labels)
+    
+    adjacency_dict = {k: list(v) for k, v in adjacency_dict.items()}
+    
+    return adjacency_dict
+
+
+
+def get_adjacent_voxels(atlas_file, adjacency_dict):
+    if atlas_file.endswith('nii.gz'):
+        nii = nib.load(atlas_file)
+        data = nii.get_fdata().astype(np.int32)
+    else:
+        data = np.load(atlas_file)
+    label_voxels = {}
+    for label in adjacency_dict.keys():
+        label_voxels[label] = np.array(np.where(data == label)).T
+    
+    adj_voxel_dict = {}
+    
+    print(f"total of {len(adjacency_dict)} label adjacency relationship:")
+    for label, neighbors in adjacency_dict.items():
+        print(f"Processing label {label},  {len(neighbors)} neighbors")
+        
+        adjacent_voxels = []
+        for n in neighbors:
+            if n in label_voxels:
+                adjacent_voxels.extend(label_voxels[n])
+
+        adj_voxel_dict[label] = np.array(adjacent_voxels) if adjacent_voxels else np.array([])
+        
+        print(f"Label {label} neighbor: {len(adj_voxel_dict[label])}")
+    
+    return adj_voxel_dict

AtlaScore/similarity/eva.py

@@ -0,0 +1,131 @@
+import nibabel as nib
+import numpy as np
+import pyzstd,io
+import pickle
+import cupy as cp
+from compute_adj import calculate_label_adjacency_bg, get_adjacent_voxels
+
+def load_zstd_file(file_path):
+        with open(file_path, 'rb') as f:
+            decompressed_data = pyzstd.decompress(f.read())
+            buffer = io.BytesIO(decompressed_data)
+            return np.load(buffer)
+        
+
+def load_dict_pickle(file_path):
+    with open(file_path, 'rb') as f:
+        return pickle.load(f)
+        
+
+def evaluate(input_file, label_file, adj_dict):
+
+    if input_file.endswith('.zst'):
+        input_matrix = load_zstd_file(input_file)
+    else:
+        input_matrix = nib.load(input_file).get_fdata()
+    input_matrix = input_matrix[:,:,:,:300]
+    # print(input_matrix.shape) # (96, 96, 96, 60)
+
+    # img_label = nib.load(label_file)
+    if label_file.endswith('nii.gz'):
+        label_matrix = nib.load(label_file)
+        label_matrix = label_matrix.get_fdata()
+    else:
+        label_matrix = np.load(label_file)
+    # print(f'Unique labels in {label_file}:', np.unique(label_matrix))
+
+    unique_labels = np.unique(label_matrix)
+    bg = np.min(unique_labels)
+    print('background value: ', bg)
+    unique_labels = unique_labels[unique_labels > bg]
+    print('roi number:',len(unique_labels))
+    fc_means = {}
+    voxel_counts = {}
+    silhouette_scores = {}
+
+    for label in unique_labels:
+        voxel_indices = np.argwhere(label_matrix == label)
+        if voxel_indices.shape[0] < 2:
+            fc_means[label] = np.nan
+            voxel_counts[label] = len(voxel_indices)
+            silhouette_scores[label] = np.nan
+            continue
+        time_series = input_matrix[voxel_indices[:, 0], voxel_indices[:, 1], voxel_indices[:, 2], :]
+        n_voxels = time_series.shape[0]
+
+        fc_mean, fc_mat = calculate_fc_mean(time_series)
+        fc_means[label] = fc_mean
+        voxel_counts[label] = len(voxel_indices)
+
+        dissimilarity = 1 - fc_mat
+        w_i = np.nanmean(dissimilarity, axis=1)
+        adj_voxel_indices = adj_dict[label]
+        if adj_voxel_indices.shape[0] == 0:
+            silhouette_scores[label] = np.nan
+            continue
+        adj_time_series = input_matrix[adj_voxel_indices[:, 0], adj_voxel_indices[:, 1], adj_voxel_indices[:, 2], :]
+
+        time_series_gpu = cp.asarray(time_series)
+        adj_time_series_gpu = cp.asarray(adj_time_series)
+        time_series_std = (time_series_gpu - cp.mean(time_series_gpu, axis=1, keepdims=True)) / cp.std(time_series_gpu, axis=1, keepdims=True)
+        adj_time_series_std = (adj_time_series_gpu - cp.mean(adj_time_series_gpu, axis=1, keepdims=True)) / cp.std(adj_time_series_gpu, axis=1, keepdims=True)
+        cross_fc_gpu = cp.dot(time_series_std.astype(cp.float32), adj_time_series_std.T.astype(cp.float32)) / time_series_gpu.shape[1]
+        cross_fc = cross_fc_gpu.get()
+        cp.get_default_memory_pool().free_all_blocks()
+        cross_dissimilarity = 1 - cross_fc
+        b_i = np.nanmean(cross_dissimilarity, axis=1)
+        silhouette = np.zeros(n_voxels)
+        valid_mask = (w_i != 0) | (b_i != 0)
+        silhouette[valid_mask] = (b_i[valid_mask] - w_i[valid_mask]) / np.maximum(w_i[valid_mask], b_i[valid_mask])
+        silhouette_scores[label] = np.nanmean(silhouette)
+        
+    weighted_sum_fc = 0.0
+    valid_voxels = 0 
+    for label in fc_means:
+        if not np.isnan(fc_means[label]):  
+            weighted_sum_fc += fc_means[label] * voxel_counts[label]
+            valid_voxels += voxel_counts[label]
+    weighted_mean_fc = weighted_sum_fc / valid_voxels if valid_voxels > 0 else np.nan
+    
+    weighted_sum_silhouette = 0.0
+    valid_voxels = 0
+    for label in silhouette_scores:
+        if not np.isnan(silhouette_scores[label]):
+            weighted_sum_silhouette += silhouette_scores[label] * voxel_counts[label]
+            valid_voxels += voxel_counts[label]
+    silhouette_avg = weighted_sum_silhouette / valid_voxels if valid_voxels > 0 else np.nan
+
+    print(f'Homogeneity: {weighted_mean_fc:.3f}\n')
+    print(f'Silhouette Score: {silhouette_avg: .3f}\n')
+
+    return weighted_mean_fc, silhouette_avg
+
+def calculate_fc_mean(time_series):
+    if time_series.shape[0] < 2:
+        return np.nan, np.nan
+    
+    time_series_gpu = cp.asarray(time_series)
+    fc_matrix = cp.corrcoef(time_series_gpu)
+    fc_matrix = fc_matrix.get()
+    
+    if fc_matrix.ndim != 2 or fc_matrix.shape[0] != fc_matrix.shape[1]:
+        return np.nan, np.nan 
+    
+    lower_triangle_indices = np.tril_indices(fc_matrix.shape[0], -1)
+    lower_triangle_values = fc_matrix[lower_triangle_indices]
+    
+    positive_fc_values = lower_triangle_values[lower_triangle_values > 0]
+    mean_fc = np.mean(positive_fc_values) if positive_fc_values.size > 0 else 0
+    # mean_fc = np.nanmean(lower_triangle_values)
+    return mean_fc, fc_matrix
+
+    
+
+if __name__ == "__main__":
+    # fmri data
+    data_file_path = r'/path/to/your/data'
+    # your atlas
+    label_file_path = r'/path/to/your/atlas'
+    adjacency = calculate_label_adjacency_bg(label_file_path)
+    adj_dict = get_adjacent_voxels(label_file_path, adjacency)
+    homogeneity, silhouette = evaluate(data_file_path, label_file_path, adj_dict)

AtlaScore/similarity/eva_DCBC.py

@@ -0,0 +1,246 @@
+import scipy
+from scipy.sparse import csr_matrix
+import torch as pt
+import cupy as cp
+import numpy as np
+import os
+import pickle
+from pathlib import Path
+import nibabel as nib
+
+
+def compute_var_cov_pt(data, cond='all', mean_centering=True):
+    """ Compute the variance and covariance for a given data matrix.
+        (PyTorch GPU version)
+
+    Args:
+        data: subject's connectivity profile, shape [N * k]
+                     N - the size of vertices (voxel)
+                     k - the size of activation conditions
+        cond: specify the subset of activation conditions to evaluation
+              (e.g condition column [1,2,3,4]), if not given, default to
+              use all conditions
+        mean_centering: boolean value to determine whether the given subject
+                        data should be mean centered
+
+    Returns: cov - the covariance matrix of current subject data, shape [N * N]
+             var - the variance matrix of current subject data, shape [N * N]
+    """
+    if mean_centering:
+        data = data - pt.mean(data, dim=1, keepdim=True)  # mean centering
+    # specify the condition index used to compute correlation, otherwise use all conditions
+    if cond != 'all':
+        data = data[:, cond]
+    elif cond == 'all':
+        data = data
+    else:
+        raise TypeError("Invalid condition type input! cond must be either 'all'"
+                        " or the column indices of expected task conditions")
+    k = data.shape[1]
+    cov = pt.matmul(data, data.T) / (k - 1)
+    # sd = data.std(dim=1).reshape(-1, 1)  # standard deviation
+    sd = pt.sqrt(pt.sum(data ** 2, dim=1, keepdim=True) / (k - 1))
+    var = pt.matmul(sd, sd.T)
+    return cov, var
+
+def compute_var_cov_np(data, cond='all', mean_centering=True):
+    """ Compute the variance and covariance for a given data matrix.
+        (Numpy CPU version)
+
+    Args:
+        data: subject's connectivity profile, shape [N * k]
+                     N - the size of vertices (voxel)
+                     k - the size of activation conditions
+        cond: specify the subset of activation conditions to evaluation
+              (e.g condition column [1,2,3,4]), if not given, default to
+              use all conditions
+        mean_centering: boolean value to determine whether the given subject
+                        data should be mean centered
+
+    Returns: cov - the covariance matrix of current subject data, shape [N * N]
+             var - the variance matrix of current subject data, shape [N * N]
+    """
+    if mean_centering:
+        mean = data.mean(axis=1)
+        data = data - mean[:, np.newaxis]  # mean centering
+    else:
+        data = data
+
+    # specify the condition index used to compute correlation,
+    # otherwise use all conditions
+    if cond != 'all':
+        data = data[:, cond]
+    elif cond == 'all':
+        data = data
+    else:
+        raise TypeError("Invalid condition type input! cond must be either 'all'"
+                        " or the column indices of expected task conditions")
+
+    k = data.shape[1]
+    sd = np.sqrt(np.sum(np.square(data), axis=1) / (k-1))  # standard deviation
+    sd = np.reshape(sd, (sd.shape[0], 1))
+    var = np.matmul(sd, sd.transpose())
+    cov = np.matmul(data, data.transpose()) / (k-1)
+    return cov, var
+
+def compute_var_cov_cp(data, cond='all', mean_centering=True):
+    """ Compute the variance and covariance for a given data matrix.
+        (Numpy CPU version)
+
+    Args:
+        data: subject's connectivity profile, shape [N * k]
+                     N - the size of vertices (voxel)
+                     k - the size of activation conditions
+        cond: specify the subset of activation conditions to evaluation
+              (e.g condition column [1,2,3,4]), if not given, default to
+              use all conditions
+        mean_centering: boolean value to determine whether the given subject
+                        data should be mean centered
+
+    Returns: cov - the covariance matrix of current subject data, shape [N * N]
+             var - the variance matrix of current subject data, shape [N * N]
+    """
+    if mean_centering:
+        mean = data.mean(axis=1)
+        data = data - mean[:, cp.newaxis]  # mean centering
+    else:
+        data = data
+
+    # specify the condition index used to compute correlation,
+    # otherwise use all conditions
+    if cond != 'all':
+        data = data[:, cond]
+    elif cond == 'all':
+        data = data
+    else:
+        raise TypeError("Invalid condition type input! cond must be either 'all'"
+                        " or the column indices of expected task conditions")
+    data = cp.asarray(data)
+    k = data.shape[1]
+    sd = cp.sqrt(cp.sum(cp.square(data), axis=1) / (k-1))  # standard deviation
+    sd = cp.reshape(sd, (sd.shape[0], 1))
+    var = cp.matmul(sd, sd.transpose())
+    cov = cp.matmul(data, data.transpose()) / (k-1)
+    return cov, var
+
+def delete_rows_csr(mat, indices):
+    """
+    Remove the rows denoted by ``indices`` form the CSR sparse matrix ``mat``.
+    """
+    if not isinstance(mat, scipy.sparse.csr_matrix):
+        raise ValueError("works only for CSR format -- use .tocsr() first")
+    indices = list(indices)
+    mask = np.ones(mat.shape[0], dtype=bool)
+    mask[indices] = False
+    return mat[mask]
+
+
+def delete_cols_csr(mat, indices):
+    """
+    Remove the cols denoted by ``indices`` form the CSR sparse matrix ``mat``.
+    """
+    if not isinstance(mat, scipy.sparse.csr_matrix):
+        raise ValueError("works only for CSR format -- use .tocsr() first")
+    indices = list(indices)
+    mask = np.ones(mat.shape[1], dtype=bool)
+    mask[indices] = False
+    return mat[:, mask]
+
+
+def ecbc(data, dist, parcellation):
+        """
+        The public function that handle the main DCBC evaluation routine
+
+        :param parcellation: The cortical parcellation to evaluate
+        :return: dict T that contain all needed DCBC evaluation results
+        """
+        maxDist = 35
+        binWidth = 1
+        weighting = True
+        numBins = int(np.floor(maxDist / binWidth))
+        h = 'L'
+
+        D = dict()
+
+        dist = csr_matrix(dist)
+        dist = dist.tocsr()
+        
+        print(f'Evaluating {h} hemisphere for subject')
+
+        # remove nan value and medial wall from subject data
+        nanIdx = np.union1d(np.unique(np.where(np.isnan(data))[0]), np.where(parcellation == 0)[0])
+        data = np.delete(data, nanIdx, axis=0)
+        cov, var = compute_var_cov_cp(data)  # This line can be changed to use compute_corr()
+
+        # remove the nan value and medial wall from dist file
+        this_dist = delete_rows_csr(dist, nanIdx)
+        this_dist = delete_cols_csr(this_dist, nanIdx)
+        row, col, distance = scipy.sparse.find(this_dist)
+
+        parcellation = cp.asarray(parcellation)
+        row = cp.asarray(row)
+        col = cp.asarray(col)
+        distance = cp.asarray(distance)
+
+        # making parcellation matrix without medial wall and nan value
+        par = cp.delete(parcellation, nanIdx, axis=0)
+        num_within, num_between, corr_within, corr_between = [], [], [], []
+        for i in range(numBins):
+            inBin = cp.where((distance > i * binWidth) & (distance <= (i + 1) * binWidth))[0]
+
+            # lookup the row/col index of within and between vertices
+            within = cp.where((par[row[inBin]] == par[col[inBin]]) == True)[0]
+            between = cp.where((par[row[inBin]] == par[col[inBin]]) == False)[0]
+
+            # retrieve and append the number of vertices for within/between in current bin
+            num_within = cp.append(num_within, within.shape[0])
+            num_between = cp.append(num_between, between.shape[0])
+
+            # Compute and append averaged within- and between-parcel correlations in current bin
+            this_corr_within = cp.nanmean(cov[row[inBin[within]], col[inBin[within]]]) / cp.nanmean(var[row[inBin[within]], col[inBin[within]]])
+            this_corr_between = cp.nanmean(cov[row[inBin[between]], col[inBin[between]]]) / cp.nanmean(var[row[inBin[between]], col[inBin[between]]])
+            corr_within = cp.append(corr_within, this_corr_within)
+            corr_between = cp.append(corr_between, this_corr_between)
+
+            del inBin
+
+        if weighting:
+            weight = 1/(1/num_within + 1/num_between)
+            weight = weight / cp.sum(weight)
+            DCBC = cp.nansum(cp.multiply((corr_within - corr_between), weight))
+        else:
+            DCBC = cp.nansum(corr_within - corr_between)
+            weight = cp.nan
+
+        D = {
+            "binWidth": binWidth,
+            "maxDist": maxDist,
+            "hemisphere": h,
+            "num_within": num_within,
+            "num_between": num_between,
+            "corr_within": corr_within,
+            "corr_between": corr_between,
+            "weight": weight,
+            "DCBC": DCBC
+        }
+
+        return D
+
+
+
+if __name__ == "__main__":
+    hemisphere = 'L'
+    geodesic = np.load('/path/to/your/dist/file') # npz file
+    dist = geodesic['arr_0']
+    print(f'Geodesic distance shape: {dist.shape}')
+
+    parcellation = nib.load('/path/to/your/atlas/on/surface').darrays[0].data
+    print(f'Parcellation shape: {parcellation.shape}')
+
+    data = nib.load('/path/to/your/fmri/on/surface')
+    surf = [x.data for x in data.darrays]
+    surf = np.array(surf).T
+    print(f'FMRI data shape: {surf.shape}')
+    
+    dcbc = ecbc(surf, dist, parcellation)
+    print(dcbc['DCBC'])

DCA/abalation_fmri.py

@@ -0,0 +1,81 @@
+import os
+import glob
+import numpy as np
+import nibabel as nib
+import torch
+import torch.nn.functional as F
+from sklearn.cluster import KMeans
+from scipy.optimize import linear_sum_assignment
+from scipy.sparse import csr_matrix, diags, eye
+from scipy.sparse.linalg import eigsh
+from utils import process_nifti_file,large_constrcut,build_spatial_graph_and_weights,sparse_spectral_clustering
+
+# ------------------------ Main ------------------------
+if __name__=='__main__':
+    data_dir = r'./data/fmri'
+    mask_dir = r'./data/mask'
+    # out_graph_dir = [todo: your out path]
+    # out_km_dir    = [todo: your out path]
+    os.makedirs(out_graph_dir, exist_ok=True)
+    os.makedirs(out_km_dir, exist_ok=True)
+
+    with open(r'./data/sub_test.txt','r') as f:
+        subs = [s.strip() for s in f if s.strip()]
+
+    for subj in subs:
+        print('Subject', subj)
+        mask = nib.load(f'{mask_dir}/{subj}.nii.gz').get_fdata().astype(int)
+        dis_mask = (mask==1)|(mask==11)
+        dis_mask = large_constrcut(dis_mask)
+        vol = nib.load(f'{data_dir}/{subj}_FWHM3.nii.gz').get_fdata().astype(np.float32)
+        vol = np.transpose(vol,(3,0,1,2))  # (T,X,Y,Z)
+        T,D,H,W = vol.shape[0], *vol.shape[1:]
+        ts = vol.reshape(T, -1).T  # (N, T)
+        valid_idx = np.where(dis_mask.flatten())[0]
+        feats = ts[valid_idx]
+
+        import itertools
+        mask = torch.tensor(dis_mask)
+        valid_idx = torch.nonzero(mask == 1)
+        index_map = -torch.ones(D, H, W, dtype=torch.long)
+        index_map[mask == 1] = torch.arange(valid_idx.shape[0])  
+
+        offsets = torch.tensor(
+            [
+                [dx, dy, dz]
+                for dx, dy, dz in itertools.product((-1, 0, 1), repeat=3)
+                if not (dx == dy == dz == 0)         
+            ],
+            dtype=torch.int8,                         
+        )
+
+        edge_list = []
+        for i, coord in enumerate(valid_idx):
+            for offset in offsets:
+                neighbor = coord + offset
+                x, y, z = neighbor.tolist()
+                if 0 <= x < D and 0 <= y < H and 0 <= z < W:
+                    j = index_map[x, y, z].item()
+                    if j >= 0:
+                        edge_list.append([i, j])
+
+        ei = torch.tensor(edge_list).t().contiguous() # [2, E]
+
+        ew = build_spatial_graph_and_weights(ei, torch.tensor(feats)) 
+        
+        valid = np.nonzero(dis_mask.flatten())[0]
+
+        labels_graph = sparse_spectral_clustering(torch.tensor(ei), torch.tensor(ew), len(valid), k=100)
+        out_vol = np.zeros(D*H*W, dtype=int)
+        out_vol[valid] = labels_graph+1
+        out3 = out_vol.reshape(D,H,W)
+        np.save(f'{out_graph_dir}/{subj}_graph.npy', out3)
+
+        # Direct KMeans clustering
+        km = KMeans(n_clusters=100, n_init=10)
+        labels_km = km.fit_predict(feats)
+        out_vol_km = np.zeros(D*H*W, dtype=int)
+        out_vol_km[valid] = labels_km+1
+        out3_km = out_vol_km.reshape(D,H,W)
+        np.save(f'{out_km_dir}/{subj}_kmeans.npy', out3_km)
+

DCA/data/data_preparation.ipynb

@@ -0,0 +1,107 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "e38996d5",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import numpy as np\n",
+    "import nibabel as nib\n",
+    "from nibabel.processing import resample_from_to"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "ac2a1dd8",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "affine_target = np.array([\n",
+    "    [-2,  0,  0,   96], \n",
+    "    [ 0,  2,  0, -112],\n",
+    "    [ 0,  0,  2,  -90],\n",
+    "    [ 0,  0,  0,    1]\n",
+    "])"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "eee0a15e",
+   "metadata": {},
+   "source": [
+    "##### prepare fmri data"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "e1da73fd",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "fmri_img = nib.load('./MNI152_fmri_data.nii.gz') # TBD\n",
+    "\n",
+    "fmri_data = fmri_img.get_fdata(); affine = fmri_img.affine; header = fmri_img.header.copy()\n",
+    "fmri_cropped = fmri_data[..., :300]; header.set_data_shape(fmri_cropped.shape); cropped_img = nib.Nifti1Image(fmri_cropped, affine, header)\n",
+    "resampled_fmri = resample_from_to(cropped_img, ((96, 96, 96, 300), affine_target), order = 3) # if fmri data has 2mm isotropic spatial resolution, use order = 0 for quicker resampling\n",
+    "\n",
+    "nib.save(resampled_fmri, './prepared_rfmri_data.nii.gz') # TBD"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "c54ff8b2",
+   "metadata": {},
+   "source": [
+    "##### prepare tissue mask"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "795e404d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "aseg_img = nib.load('./MNI152_aparc+aseg.nii.gz') # TBD\n",
+    "aseg_data = aseg_img.get_fdata()\n",
+    "\n",
+    "res = np.zeros_like(aseg_data)\n",
+    "res += 1 * ((aseg_data >= 1000) & (aseg_data <= 1035)).astype(int); res += 11 * ((aseg_data >= 2000) & (aseg_data <= 2035)).astype(int)\n",
+    "res += 2 * (aseg_data == 2).astype(int); res += 12 * (aseg_data == 41).astype(int)\n",
+    "for i in [10, 11, 12, 13, 17, 18, 26, 27, 28, 75, 96]: res += 3 * (aseg_data == i).astype(int)\n",
+    "for i in [49, 50, 51, 52, 53, 54, 58, 59, 60, 76, 97]: res += 13 * (aseg_data == i).astype(int)\n",
+    "res += 20 * ((aseg_data >= 251) & (aseg_data <= 255)).astype(int)\n",
+    "\n",
+    "new_img = nib.Nifti1Image(res, affine = aseg_img.affine)\n",
+    "resampled_img = resample_from_to(new_img, ((96, 96, 96), affine_target), order = 0)\n",
+    "\n",
+    "nib.save(resampled_img, './prepared_tissue_mask.nii.gz') # TBD"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "connectome",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.2"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

DCA/data/fmri/101915_FWHM3.nii.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1412cf0d0c3e618de05a5ca9134fd7598ba0efab0ccd922c69dbd22ff0eb467
+size 246573686

DCA/data/fmri/101915_FWHM6.nii.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab71fa6a8f329425daaa8f14a8fba7296a6eb2136944886dcb733c275aa54463
+size 245852861

DCA/data/fmri/101915_raw.nii.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07836b9803a8dbede7e0fe226b1301d486c8a15a92fa5cb9b6c8951156f82c3d
+size 254082341

DCA/data/mask/101915_tissue_mask.nii.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf712ff0ab0838aff49771feee9fc90854246b0bb7216d9edd36a4fdbb0868ab
+size 151082

DCA/data/sub_test.txt

@@ -0,0 +1 @@
+101915

DCA/main.py

@@ -0,0 +1,243 @@
+from __future__ import print_function, division
+import argparse
+import os
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:128"
+os.environ["OPENBLAS_NUM_THREADS"] = "16"
+os.environ["OMP_NUM_THREADS"] = "16"
+os.environ["GOTO_NUM_THREADS"] = "16"
+
+os.environ["NUMEXPR_NUM_THREADS"] = "16"
+os.environ['CUDA_VISIBLE_DEVICES'] = "5"
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+import torch.optim as optim
+import itertools
+
+import argparse
+from utils import large_constrcut,build_spatial_graph_and_weights,sparse_spectral_clustering,match_labels,validate
+from model import DCA,MySwinUNETR
+import time
+import nibabel as nib
+import torch.nn as nn
+import warnings 
+import argparse
+from torch.nn.parameter import Parameter
+
+warnings.filterwarnings("ignore")
+
+
+def train_DCA(swin,origin,background_mask,evaluation_data,data_dir):
+
+    model = DCA(
+        swin,
+        n_z=args.n_z,
+        n_clusters=args.n_clusters,
+        background_mask=background_mask,
+        ).to(device)
+    start = time.time()        
+
+    for name, param in model.swinunetr.named_parameters():
+        if 'c3d' in name or 'out' in name:
+            param.requires_grad = True
+        else:
+            param.requires_grad = False
+
+    optimizer = optim.Adam([
+        {'params': [
+            p for n, p in model.swinunetr.named_parameters() 
+            if p.requires_grad
+        ],
+        'lr': args.lr},
+    ])
+
+    data = torch.from_numpy(origin).float() 
+    data = data.unsqueeze(0).to(device)  
+    data.requires_grad_(True)  
+    
+    D, H, W = background_mask.shape
+    mask = torch.tensor(background_mask).to(device)
+    valid_idx = torch.nonzero(mask == 0)
+    index_map = -torch.ones(D, H, W, dtype=torch.long)
+    index_map[mask == 0] = torch.arange(valid_idx.shape[0])  
+    # 6-neighbour
+    # offsets = torch.tensor([
+    #     [0, 0, 1], [0, 0, -1],
+    #     [0, 1, 0], [0, -1, 0],
+    #     [1, 0, 0], [-1, 0, 0],
+    # ]).to(device)
+
+    # 26-neighbour
+    offsets = torch.tensor(
+        [
+            [dx, dy, dz]
+            for dx, dy, dz in itertools.product((-1, 0, 1), repeat=3)
+            if not (dx == dy == dz == 0)           # exclude the centre voxel
+        ],
+        dtype=torch.int8,                   
+    ).to(device) 
+
+    edge_list = []
+    for i, coord in enumerate(valid_idx):
+        for offset in offsets:
+            neighbor = coord + offset
+            x, y, z = neighbor.tolist()
+            if 0 <= x < D and 0 <= y < H and 0 <= z < W:
+                j = index_map[x, y, z].item()
+                if j >= 0:
+                    edge_list.append([i, j])
+
+    edge_index = torch.tensor(edge_list).t().contiguous().to(device)  # [2, E]
+    
+    
+    model.train()
+    print('start tunning')
+    best_h = 0
+    best_s = 0
+    for epoch in range(args.epoch):
+        optimizer.zero_grad()
+        x_bar, tmp_s, z,hidden = model(data,epoch)
+
+        ## valid flatten
+        mask_flat = background_mask.flatten()  
+        valid_mask = mask_flat == 0             
+        valid_mask = torch.tensor(valid_mask).to(device)
+        
+        hidden_permuted = hidden.permute(0, 2, 3, 4, 1) 
+        if not hidden_permuted.is_contiguous():
+            hidden_permuted = hidden_permuted.contiguous()
+        hidden_flat = hidden_permuted.view(-1, hidden.shape[1])  
+        recon_permuted = x_bar.permute(0, 2, 3, 4, 1)
+        if not recon_permuted.is_contiguous():
+            recon_permuted = recon_permuted.contiguous()
+        recon_flat = recon_permuted.view(-1, x_bar.shape[1])
+
+        valid_indices = torch.nonzero(valid_mask, as_tuple=True)[0] 
+        valid_voxels = torch.index_select(hidden_flat, dim=0, index=valid_indices)
+        valid_recon = torch.index_select(recon_flat, dim=0, index=valid_indices)
+
+        edge_weight = build_spatial_graph_and_weights(
+            edge_index,  # [96, 96, 96]
+            valid_voxels
+        )
+
+        labels = sparse_spectral_clustering(edge_index.detach().cpu(), edge_weight.detach().cpu(), (torch.max(edge_index)+1).item(), args.n_clusters)
+
+        if epoch > 0:
+            labels = match_labels(prev_labels, labels, args.n_clusters)
+        prev_labels = labels.copy()
+
+        kl_loss = model.total_loss(data, x_bar, tmp_s, target=torch.tensor(labels).long().to(device))
+        
+        loss = kl_loss
+        print('epoch:',epoch)
+        # print('reconstr_loss',recon)
+        print('kl_loss',kl_loss) 
+        
+        loss.backward()
+        optimizer.step()
+        label_3d = np.full(background_mask.shape, -1)  # Initialize with -1 to mark background areas
+        # y_pred = target_distribution(tmp_s).argmax(1).cpu()
+        label_3d[background_mask == 0] = labels 
+        if args.vali :
+            homogeneity = validate(data_dir, label_3d+1) #input numpy
+            if best_h < homogeneity:
+                best_h = homogeneity
+                bset_s = label_3d
+            label_3d = bset_s
+
+    # edge_weight = build_spatial_graph_and_weights(
+    #     edge_index, 
+    #     valid_voxels
+    # )
+    # labels = weighted_bfs_connected_clustering(edge_index.detach(), edge_weight.detach(), (torch.max(edge_index)+1).item(), args.n_clusters)
+
+    end = time.time() 
+    print('Running time: ', end-start)
+    return label_3d
+    
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser(
+        description='DCA training',
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+
+    parser.add_argument('--lr', type=float, default=0.005)
+    parser.add_argument('--n_clusters','-k', default=100, type=int) #38
+    parser.add_argument('--n_z', default=256, type=int)
+    parser.add_argument('--epoch','-e', default=8, type=int)
+    parser.add_argument('--dis_mask',default=None)
+    parser.add_argument('--vali','-v',default=True) # For small Memory, choose False and set smaller epoch about 8 to avoid overfit.
+
+
+    args = parser.parse_args()
+    print(args)
+    args.cuda = torch.cuda.is_available()
+    print("use cuda: {}".format(args.cuda))
+    print("gpu counts:", torch.cuda.device_count())
+    torch.cuda.empty_cache()
+    device = torch.device("cuda" if args.cuda else "cpu")
+
+    subj_list_file = r'data/sub_test.txt'
+
+    with open(subj_list_file, 'r') as f:
+        subj_ids = f.read().splitlines()
+    
+    # import random
+    # subj_ids_copy = subj_ids[:]  #shuffle
+    # random.shuffle(subj_ids_copy)
+    
+    # Loop over each subject ID in the list
+    for idx, subj_id in enumerate(subj_ids):
+        k_cluster = args.n_clusters
+        args.n_clusters = k_cluster
+        print(f"/n[ {idx+1}/{len(subj_ids)} ]  Processing subject {subj_id} …")
+
+        result_file = fr'results/demo/hcp_{subj_id}_{k_cluster}.npy'
+        if os.path.isfile(result_file):
+            print(f"  ->  {result_file} already exists, skip.")
+            continue                   
+
+        data_dir = fr'data/fmri/{subj_id}_FWHM3.nii.gz'
+        dis_mask_dir = fr'data/mask/101915_tissue_mask.nii.gz'  
+
+        dis_mask = nib.load(dis_mask_dir).get_fdata()
+        dis_mask = (dis_mask==1) | (dis_mask==11)
+        # 1: left grey matter
+        # 2: left white matter
+        # 3: left subcortex
+        # 11: right grey matter
+        # 12: right white matter
+        # 13: right subcortex
+        # 20: corpus callosum
+        dis_mask = large_constrcut(dis_mask)  #processing isolate points
+
+        original_volume = nib.load(data_dir).get_fdata().astype('float32')
+        original_volume = np.transpose(original_volume, (3, 0, 1, 2))
+        evaluation_data = original_volume
+
+        background_mask = (dis_mask == 0)
+        background_mask = np.broadcast_to(background_mask, original_volume.shape)
+        nb_vals = original_volume[~background_mask]
+        if nb_vals.min() != nb_vals.max():
+            mean, std = nb_vals.mean(), nb_vals.std()
+            original_volume[~background_mask] = (nb_vals - mean) / std
+        original_volume[background_mask] = 0
+        normalized_original = original_volume
+
+        checkpoint_path = 'data/swin_model_epoch_30.pth'
+        model = MySwinUNETR(img_size=(96,96,96), in_channels=300,
+                            out_channels=300, feature_size=48,
+                            emb_size=256, use_checkpoint=False).to(device)
+        if os.path.isfile(checkpoint_path):
+            ckpt = torch.load(checkpoint_path, map_location=device)
+            model.load_state_dict(ckpt['model_state_dict'])
+            print(f"  -> checkpoint loaded ({checkpoint_path})")
+
+        label_3d = train_DCA(model,normalized_original,background_mask[0],evaluation_data,data_dir)
+
+
+        np.save(result_file, label_3d+1)
+        print(f"  -> saved {result_file}")

DCA/model.py

@@ -0,0 +1,88 @@
+import torch
+
+import torch.nn as nn
+from torch.nn.parameter import Parameter
+import torch.nn.functional as F
+import torch.optim as optim
+#   pip install monai
+from monai.networks.nets import SwinUNETR
+import importlib.util
+spec = importlib.util.spec_from_file_location('SwinUNETR', 'swin_unetr.py')
+my_module = importlib.util.module_from_spec(spec)
+spec.loader.exec_module(my_module)
+SwinUNETR = my_module.SwinUNETR
+
+class MySwinUNETR(SwinUNETR):
+    def forward(self, x_in):
+        hidden_states_out = self.swinViT(x_in, self.normalize)
+        enc0 = self.encoder1(x_in)
+        enc1 = self.encoder2(hidden_states_out[0])
+        enc2 = self.encoder3(hidden_states_out[1])
+        enc3 = self.encoder4(hidden_states_out[2])
+        dec4 = self.encoder10(hidden_states_out[4])
+        dec3 = self.decoder5(dec4, hidden_states_out[3])
+        dec2 = self.decoder4(dec3, enc3)
+        dec1 = self.decoder3(dec2, enc2)
+        dec0 = self.decoder2(dec1, enc1)
+        out = self.decoder1(dec0, enc0)
+        out = self.c3d(out)
+        logits = self.out(out)
+        return logits, out
+
+class DCA(nn.Module):
+    def __init__(self,
+                 model,
+                 n_z,
+                 n_clusters,
+                 background_mask,
+                 ):
+        super(DCA, self).__init__()
+        # self.pretrain_path = pretrain_path
+        self.n_clusters = n_clusters
+
+        self.swinunetr = model
+        self.background_mask = background_mask
+        self.D = Parameter(torch.Tensor(n_clusters, n_z),requires_grad=True)
+        nn.init.xavier_uniform_(self.D)
+
+     
+    def forward(self, x, epoch):
+        x_bar, hidden = self.swinunetr(x)
+        z = hidden
+        B, C, D, H, W = z.shape
+        z_perm = z.permute(0,2,3,4,1).reshape(-1, C)  # (B*D*H*W, C)
+        mask_flat = torch.from_numpy(self.background_mask).flatten().to(z.device)  # (D*H*W,)
+        z_features = z_perm[~mask_flat.repeat(B)]
+        feature_distances = torch.cdist(z_features, self.D)
+        s = - feature_distances
+        return x_bar, s, z_features,z
+
+    def orthogonality_loss(self,D):
+    
+        K = D.size(0)
+        G = D @ D.t()                     # shape [K, K]
+        G_off = G - torch.eye(K, device=D.device, dtype=D.dtype)
+        loss = (G_off.sum() / (K * (K - 1))).sqrt() 
+        return loss
+
+    def masked_mse_loss(self, pred, target, mask):
+        mask = torch.from_numpy(mask)
+        if mask.dim() != pred.dim():  # (H, W, D)
+            mask = mask.unsqueeze(0)  # Add batch dimension to mask (1, C, H, W, D)
+            mask = mask.expand(pred.size(1), -1, -1, -1)  # Repeat mask for each batch (B, C, H, W, D)
+            mask = mask.unsqueeze(0)  # Add batch dimension to mask (1, C, H, W, D)
+            mask = mask.expand(pred.size(0), -1, -1, -1, -1)
+        diff = pred - target
+        masked_diff = diff[~mask]  # Select only the non-background areas (where mask is False)
+        if masked_diff.numel() == 0:
+            return torch.tensor(0.0, requires_grad=True, device=pred.device)
+        return torch.mean(masked_diff**2)
+
+    def total_loss(self, x, x_bar, pred, target):
+
+        loss1 = F.cross_entropy(pred, target) #Since target is hard label, minimizing the KL divergence is equivalent to minimizing the cross‐entropy.
+
+        # reconstr_loss =  self.masked_mse_loss(x_bar, x, self.background_mask)    
+        # orth_loss = self.orthogonality_loss(self.D)
+        return loss1
+

DCA/req_trim.txt

@@ -0,0 +1,10 @@
+scikit-learn
+tensorboard
+monai
+matplotlib
+torch_geometric
+nibabel
+pyzstd
+pandas
+einops
+cupy-cuda12x

DCA/requirements.txt

@@ -0,0 +1,75 @@
+absl-py==2.2.2
+aiohappyeyeballs==2.6.1
+aiohttp==3.11.18
+aiosignal==1.3.2
+async-timeout==5.0.1
+attrs==25.3.0
+certifi==2025.4.26
+charset-normalizer==3.4.2
+contourpy==1.3.0
+cycler==0.12.1
+einops==0.8.1
+filelock==3.18.0
+fonttools==4.58.0
+frozenlist==1.6.0
+fsspec==2025.3.2
+grpcio==1.71.0
+idna==3.10
+importlib_metadata==8.7.0
+importlib_resources==6.5.2
+Jinja2==3.1.6
+joblib==1.5.0
+kiwisolver==1.4.7
+Markdown==3.8
+MarkupSafe==3.0.2
+matplotlib==3.9.4
+monai==1.4.0
+mpmath==1.3.0
+multidict==6.4.3
+networkx==3.2.1
+nibabel==5.3.2
+numpy==1.26.4
+nvidia-cublas-cu12==12.6.4.1
+nvidia-cuda-cupti-cu12==12.6.80
+nvidia-cuda-nvrtc-cu12==12.6.77
+nvidia-cuda-runtime-cu12==12.6.77
+nvidia-cudnn-cu12==9.5.1.17
+nvidia-cufft-cu12==11.3.0.4
+nvidia-cufile-cu12==1.11.1.6
+nvidia-curand-cu12==10.3.7.77
+nvidia-cusolver-cu12==11.7.1.2
+nvidia-cusparse-cu12==12.5.4.2
+nvidia-cusparselt-cu12==0.6.3
+nvidia-nccl-cu12==2.26.2
+nvidia-nvjitlink-cu12==12.6.85
+nvidia-nvtx-cu12==12.6.77
+packaging==25.0
+pandas==2.2.3
+pillow==11.2.1
+propcache==0.3.1
+protobuf==6.30.2
+psutil==7.0.0
+pyparsing==3.2.3
+python-dateutil==2.9.0.post0
+pytz==2025.2
+pyzstd==0.17.0
+requests==2.32.3
+scikit-learn==1.6.1
+scipy==1.13.1
+six==1.17.0
+sympy==1.14.0
+tensorboard==2.19.0
+tensorboard-data-server==0.7.2
+threadpoolctl==3.6.0
+torch==2.7.0
+torch-geometric==2.6.1
+torchaudio==2.7.0
+torchvision==0.22.0
+tqdm==4.67.1
+triton==3.3.0
+typing_extensions==4.13.2
+tzdata==2025.2
+urllib3==2.4.0
+Werkzeug==3.1.3
+yarl==1.20.0
+zipp==3.21.0

DCA/results/demo/hcp_101915_100.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:951f1c252353e01b0df1f3f04b469595393cedce07bc693c158ccf067ffeb7c4
+size 7078016

DCA/swin_unetr.py

@@ -0,0 +1,1060 @@
+# Copyright (c) MONAI Consortium
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#     http://www.apache.org/licenses/LICENSE-2.0
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+import itertools
+from collections.abc import Sequence
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+from torch.nn import LayerNorm
+
+from monai.networks.blocks import MLPBlock as Mlp
+from monai.networks.blocks import PatchEmbed, UnetOutBlock, UnetrBasicBlock, UnetrUpBlock
+from monai.networks.layers import DropPath, trunc_normal_
+from monai.utils import ensure_tuple_rep, look_up_option, optional_import
+
+rearrange, _ = optional_import("einops", name="rearrange")
+
+__all__ = [
+    "SwinUNETR",
+    "window_partition",
+    "window_reverse",
+    "WindowAttention",
+    "SwinTransformerBlock",
+    "PatchMerging",
+    "PatchMergingV2",
+    "MERGING_MODE",
+    "BasicLayer",
+    "SwinTransformer",
+]
+
+
+class SwinUNETR(nn.Module):
+    """
+    Swin UNETR based on: "Hatamizadeh et al.,
+    Swin UNETR: Swin Transformers for Semantic Segmentation of Brain Tumors in MRI Images
+    <https://arxiv.org/abs/2201.01266>"
+    """
+
+    def __init__(
+        self,
+        img_size: Sequence[int] | int,
+        in_channels: int,
+        out_channels: int,
+        depths: Sequence[int] = (2, 2, 2, 2),
+        num_heads: Sequence[int] = (3, 6, 12, 24),
+        feature_size: int = 24,
+        norm_name: tuple | str = "instance",
+        drop_rate: float = 0.0,
+        attn_drop_rate: float = 0.0,
+        dropout_path_rate: float = 0.0,
+        normalize: bool = True,
+        use_checkpoint: bool = False,
+        spatial_dims: int = 3,
+        downsample="merging",
+        use_v2=False,
+        emb_size=256
+    ) -> None:
+        """
+        Args:
+            img_size: dimension of input image.
+            in_channels: dimension of input channels.
+            out_channels: dimension of output channels.
+            feature_size: dimension of network feature size.
+            depths: number of layers in each stage.
+            num_heads: number of attention heads.
+            norm_name: feature normalization type and arguments.
+            drop_rate: dropout rate.
+            attn_drop_rate: attention dropout rate.
+            dropout_path_rate: drop path rate.
+            normalize: normalize output intermediate features in each stage.
+            use_checkpoint: use gradient checkpointing for reduced memory usage.
+            spatial_dims: number of spatial dims.
+            downsample: module used for downsampling, available options are `"mergingv2"`, `"merging"` and a
+                user-specified `nn.Module` following the API defined in :py:class:`monai.networks.nets.PatchMerging`.
+                The default is currently `"merging"` (the original version defined in v0.9.0).
+            use_v2: using swinunetr_v2, which adds a residual convolution block at the beggining of each swin stage.
+
+        Examples::
+
+            # for 3D single channel input with size (96,96,96), 4-channel output and feature size of 48.
+            >>> net = SwinUNETR(img_size=(96,96,96), in_channels=1, out_channels=4, feature_size=48)
+
+            # for 3D 4-channel input with size (128,128,128), 3-channel output and (2,4,2,2) layers in each stage.
+            >>> net = SwinUNETR(img_size=(128,128,128), in_channels=4, out_channels=3, depths=(2,4,2,2))
+
+            # for 2D single channel input with size (96,96), 2-channel output and gradient checkpointing.
+            >>> net = SwinUNETR(img_size=(96,96), in_channels=3, out_channels=2, use_checkpoint=True, spatial_dims=2)
+
+        """
+
+        super().__init__()
+
+        img_size = ensure_tuple_rep(img_size, spatial_dims)
+        patch_size = ensure_tuple_rep(2, spatial_dims)
+        window_size = ensure_tuple_rep(7, spatial_dims)
+
+        if spatial_dims not in (2, 3):
+            raise ValueError("spatial dimension should be 2 or 3.")
+
+        for m, p in zip(img_size, patch_size):
+            for i in range(5):
+                if m % np.power(p, i + 1) != 0:
+                    raise ValueError("input image size (img_size) should be divisible by stage-wise image resolution.")
+
+        if not (0 <= drop_rate <= 1):
+            raise ValueError("dropout rate should be between 0 and 1.")
+
+        if not (0 <= attn_drop_rate <= 1):
+            raise ValueError("attention dropout rate should be between 0 and 1.")
+
+        if not (0 <= dropout_path_rate <= 1):
+            raise ValueError("drop path rate should be between 0 and 1.")
+
+        if feature_size % 12 != 0:
+            raise ValueError("feature_size should be divisible by 12.")
+
+        self.normalize = normalize
+
+        self.swinViT = SwinTransformer(
+            in_chans=in_channels,
+            embed_dim=feature_size,
+            window_size=window_size,
+            patch_size=patch_size,
+            depths=depths,
+            num_heads=num_heads,
+            mlp_ratio=4.0,
+            qkv_bias=True,
+            drop_rate=drop_rate,
+            attn_drop_rate=attn_drop_rate,
+            drop_path_rate=dropout_path_rate,
+            norm_layer=nn.LayerNorm,
+            use_checkpoint=use_checkpoint,
+            spatial_dims=spatial_dims,
+            downsample=look_up_option(downsample, MERGING_MODE) if isinstance(downsample, str) else downsample,
+            use_v2=use_v2,
+        )
+
+        self.encoder1 = UnetrBasicBlock(
+            spatial_dims=spatial_dims,
+            in_channels=in_channels,
+            out_channels=feature_size,
+            kernel_size=3,
+            stride=1,
+            norm_name=norm_name,
+            res_block=True,
+        )
+
+        self.encoder2 = UnetrBasicBlock(
+            spatial_dims=spatial_dims,
+            in_channels=feature_size,
+            out_channels=feature_size,
+            kernel_size=3,
+            stride=1,
+            norm_name=norm_name,
+            res_block=True,
+        )
+
+        self.encoder3 = UnetrBasicBlock(
+            spatial_dims=spatial_dims,
+            in_channels=2 * feature_size,
+            out_channels=2 * feature_size,
+            kernel_size=3,
+            stride=1,
+            norm_name=norm_name,
+            res_block=True,
+        )
+
+        self.encoder4 = UnetrBasicBlock(
+            spatial_dims=spatial_dims,
+            in_channels=4 * feature_size,
+            out_channels=4 * feature_size,
+            kernel_size=3,
+            stride=1,
+            norm_name=norm_name,
+            res_block=True,
+        )
+
+        self.encoder10 = UnetrBasicBlock(
+            spatial_dims=spatial_dims,
+            in_channels=16 * feature_size,
+            out_channels=16 * feature_size,
+            kernel_size=3,
+            stride=1,
+            norm_name=norm_name,
+            res_block=True,
+        )
+
+        self.decoder5 = UnetrUpBlock(
+            spatial_dims=spatial_dims,
+            in_channels=16 * feature_size,
+            out_channels=8 * feature_size,
+            kernel_size=3,
+            upsample_kernel_size=2,
+            norm_name=norm_name,
+            res_block=True,
+        )
+
+        self.decoder4 = UnetrUpBlock(
+            spatial_dims=spatial_dims,
+            in_channels=feature_size * 8,
+            out_channels=feature_size * 4,
+            kernel_size=3,
+            upsample_kernel_size=2,
+            norm_name=norm_name,
+            res_block=True,
+        )
+
+        self.decoder3 = UnetrUpBlock(
+            spatial_dims=spatial_dims,
+            in_channels=feature_size * 4,
+            out_channels=feature_size * 2,
+            kernel_size=3,
+            upsample_kernel_size=2,
+            norm_name=norm_name,
+            res_block=True,
+        )
+        self.decoder2 = UnetrUpBlock(
+            spatial_dims=spatial_dims,
+            in_channels=feature_size * 2,
+            out_channels=feature_size,
+            kernel_size=3,
+            upsample_kernel_size=2,
+            norm_name=norm_name,
+            res_block=True,
+        )
+
+        self.decoder1 = UnetrUpBlock(
+            spatial_dims=spatial_dims,
+            in_channels=feature_size,
+            out_channels=feature_size,
+            kernel_size=3,
+            upsample_kernel_size=2,
+            norm_name=norm_name,
+            res_block=True,
+        )
+        self.c3d = nn.Conv3d(feature_size, emb_size, kernel_size=1)
+        self.out = UnetOutBlock(spatial_dims=spatial_dims, in_channels=emb_size, out_channels=out_channels)
+
+    def load_from(self, weights):
+        with torch.no_grad():
+            self.swinViT.patch_embed.proj.weight.copy_(weights["state_dict"]["module.patch_embed.proj.weight"])
+            self.swinViT.patch_embed.proj.bias.copy_(weights["state_dict"]["module.patch_embed.proj.bias"])
+            for bname, block in self.swinViT.layers1[0].blocks.named_children():
+                block.load_from(weights, n_block=bname, layer="layers1")
+            self.swinViT.layers1[0].downsample.reduction.weight.copy_(
+                weights["state_dict"]["module.layers1.0.downsample.reduction.weight"]
+            )
+            self.swinViT.layers1[0].downsample.norm.weight.copy_(
+                weights["state_dict"]["module.layers1.0.downsample.norm.weight"]
+            )
+            self.swinViT.layers1[0].downsample.norm.bias.copy_(
+                weights["state_dict"]["module.layers1.0.downsample.norm.bias"]
+            )
+            for bname, block in self.swinViT.layers2[0].blocks.named_children():
+                block.load_from(weights, n_block=bname, layer="layers2")
+            self.swinViT.layers2[0].downsample.reduction.weight.copy_(
+                weights["state_dict"]["module.layers2.0.downsample.reduction.weight"]
+            )
+            self.swinViT.layers2[0].downsample.norm.weight.copy_(
+                weights["state_dict"]["module.layers2.0.downsample.norm.weight"]
+            )
+            self.swinViT.layers2[0].downsample.norm.bias.copy_(
+                weights["state_dict"]["module.layers2.0.downsample.norm.bias"]
+            )
+            for bname, block in self.swinViT.layers3[0].blocks.named_children():
+                block.load_from(weights, n_block=bname, layer="layers3")
+            self.swinViT.layers3[0].downsample.reduction.weight.copy_(
+                weights["state_dict"]["module.layers3.0.downsample.reduction.weight"]
+            )
+            self.swinViT.layers3[0].downsample.norm.weight.copy_(
+                weights["state_dict"]["module.layers3.0.downsample.norm.weight"]
+            )
+            self.swinViT.layers3[0].downsample.norm.bias.copy_(
+                weights["state_dict"]["module.layers3.0.downsample.norm.bias"]
+            )
+            for bname, block in self.swinViT.layers4[0].blocks.named_children():
+                block.load_from(weights, n_block=bname, layer="layers4")
+            self.swinViT.layers4[0].downsample.reduction.weight.copy_(
+                weights["state_dict"]["module.layers4.0.downsample.reduction.weight"]
+            )
+            self.swinViT.layers4[0].downsample.norm.weight.copy_(
+                weights["state_dict"]["module.layers4.0.downsample.norm.weight"]
+            )
+            self.swinViT.layers4[0].downsample.norm.bias.copy_(
+                weights["state_dict"]["module.layers4.0.downsample.norm.bias"]
+            )
+
+    def forward(self, x_in):
+
+        hidden_states_out = self.swinViT(x_in, self.normalize)
+        enc0 = self.encoder1(x_in)
+        enc1 = self.encoder2(hidden_states_out[0])
+        enc2 = self.encoder3(hidden_states_out[1])
+        enc3 = self.encoder4(hidden_states_out[2])
+        dec4 = self.encoder10(hidden_states_out[4])
+        dec3 = self.decoder5(dec4, hidden_states_out[3])
+        dec2 = self.decoder4(dec3, enc3)
+        dec1 = self.decoder3(dec2, enc2)
+        dec0 = self.decoder2(dec1, enc1)
+        out = self.decoder1(dec0, enc0)
+        out = self.c3d(out)
+        logits = self.out(out)
+        return logits
+
+
+def window_partition(x, window_size):
+    """window partition operation based on: "Liu et al.,
+    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
+    <https://arxiv.org/abs/2103.14030>"
+    https://github.com/microsoft/Swin-Transformer
+
+     Args:
+        x: input tensor.
+        window_size: local window size.
+    """
+    x_shape = x.size()
+    if len(x_shape) == 5:
+        b, d, h, w, c = x_shape
+        x = x.view(
+            b,
+            d // window_size[0],
+            window_size[0],
+            h // window_size[1],
+            window_size[1],
+            w // window_size[2],
+            window_size[2],
+            c,
+        )
+        windows = (
+            x.permute(0, 1, 3, 5, 2, 4, 6, 7).contiguous().view(-1, window_size[0] * window_size[1] * window_size[2], c)
+        )
+    elif len(x_shape) == 4:
+        b, h, w, c = x.shape
+        x = x.view(b, h // window_size[0], window_size[0], w // window_size[1], window_size[1], c)
+        windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size[0] * window_size[1], c)
+    return windows
+
+
+def window_reverse(windows, window_size, dims):
+    """window reverse operation based on: "Liu et al.,
+    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
+    <https://arxiv.org/abs/2103.14030>"
+    https://github.com/microsoft/Swin-Transformer
+
+     Args:
+        windows: windows tensor.
+        window_size: local window size.
+        dims: dimension values.
+    """
+    if len(dims) == 4:
+        b, d, h, w = dims
+        x = windows.view(
+            b,
+            d // window_size[0],
+            h // window_size[1],
+            w // window_size[2],
+            window_size[0],
+            window_size[1],
+            window_size[2],
+            -1,
+        )
+        x = x.permute(0, 1, 4, 2, 5, 3, 6, 7).contiguous().view(b, d, h, w, -1)
+
+    elif len(dims) == 3:
+        b, h, w = dims
+        x = windows.view(b, h // window_size[0], w // window_size[1], window_size[0], window_size[1], -1)
+        x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(b, h, w, -1)
+    return x
+
+
+def get_window_size(x_size, window_size, shift_size=None):
+    """Computing window size based on: "Liu et al.,
+    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
+    <https://arxiv.org/abs/2103.14030>"
+    https://github.com/microsoft/Swin-Transformer
+
+     Args:
+        x_size: input size.
+        window_size: local window size.
+        shift_size: window shifting size.
+    """
+
+    use_window_size = list(window_size)
+    if shift_size is not None:
+        use_shift_size = list(shift_size)
+    for i in range(len(x_size)):
+        if x_size[i] <= window_size[i]:
+            use_window_size[i] = x_size[i]
+            if shift_size is not None:
+                use_shift_size[i] = 0
+
+    if shift_size is None:
+        return tuple(use_window_size)
+    else:
+        return tuple(use_window_size), tuple(use_shift_size)
+
+
+class WindowAttention(nn.Module):
+    """
+    Window based multi-head self attention module with relative position bias based on: "Liu et al.,
+    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
+    <https://arxiv.org/abs/2103.14030>"
+    https://github.com/microsoft/Swin-Transformer
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        window_size: Sequence[int],
+        qkv_bias: bool = False,
+        attn_drop: float = 0.0,
+        proj_drop: float = 0.0,
+    ) -> None:
+        """
+        Args:
+            dim: number of feature channels.
+            num_heads: number of attention heads.
+            window_size: local window size.
+            qkv_bias: add a learnable bias to query, key, value.
+            attn_drop: attention dropout rate.
+            proj_drop: dropout rate of output.
+        """
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = head_dim**-0.5
+        mesh_args = torch.meshgrid.__kwdefaults__
+
+        if len(self.window_size) == 3:
+            self.relative_position_bias_table = nn.Parameter(
+                torch.zeros(
+                    (2 * self.window_size[0] - 1) * (2 * self.window_size[1] - 1) * (2 * self.window_size[2] - 1),
+                    num_heads,
+                )
+            )
+            coords_d = torch.arange(self.window_size[0])
+            coords_h = torch.arange(self.window_size[1])
+            coords_w = torch.arange(self.window_size[2])
+            if mesh_args is not None:
+                coords = torch.stack(torch.meshgrid(coords_d, coords_h, coords_w, indexing="ij"))
+            else:
+                coords = torch.stack(torch.meshgrid(coords_d, coords_h, coords_w))
+            coords_flatten = torch.flatten(coords, 1)
+            relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]
+            relative_coords = relative_coords.permute(1, 2, 0).contiguous()
+            relative_coords[:, :, 0] += self.window_size[0] - 1
+            relative_coords[:, :, 1] += self.window_size[1] - 1
+            relative_coords[:, :, 2] += self.window_size[2] - 1
+            relative_coords[:, :, 0] *= (2 * self.window_size[1] - 1) * (2 * self.window_size[2] - 1)
+            relative_coords[:, :, 1] *= 2 * self.window_size[2] - 1
+        elif len(self.window_size) == 2:
+            self.relative_position_bias_table = nn.Parameter(
+                torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)
+            )
+            coords_h = torch.arange(self.window_size[0])
+            coords_w = torch.arange(self.window_size[1])
+            if mesh_args is not None:
+                coords = torch.stack(torch.meshgrid(coords_h, coords_w, indexing="ij"))
+            else:
+                coords = torch.stack(torch.meshgrid(coords_h, coords_w))
+            coords_flatten = torch.flatten(coords, 1)
+            relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]
+            relative_coords = relative_coords.permute(1, 2, 0).contiguous()
+            relative_coords[:, :, 0] += self.window_size[0] - 1
+            relative_coords[:, :, 1] += self.window_size[1] - 1
+            relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+
+        relative_position_index = relative_coords.sum(-1)
+        self.register_buffer("relative_position_index", relative_position_index)
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        trunc_normal_(self.relative_position_bias_table, std=0.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask):
+        b, n, c = x.shape
+        qkv = self.qkv(x).reshape(b, n, 3, self.num_heads, c // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]
+        q = q * self.scale
+        attn = q @ k.transpose(-2, -1)
+        relative_position_bias = self.relative_position_bias_table[
+            self.relative_position_index.clone()[:n, :n].reshape(-1)  # type: ignore
+        ].reshape(n, n, -1)
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()
+        attn = attn + relative_position_bias.unsqueeze(0)
+        if mask is not None:
+            nw = mask.shape[0]
+            attn = attn.view(b // nw, nw, self.num_heads, n, n) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, n, n)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn).to(v.dtype)
+        x = (attn @ v).transpose(1, 2).reshape(b, n, c)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class SwinTransformerBlock(nn.Module):
+    """
+    Swin Transformer block based on: "Liu et al.,
+    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
+    <https://arxiv.org/abs/2103.14030>"
+    https://github.com/microsoft/Swin-Transformer
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        window_size: Sequence[int],
+        shift_size: Sequence[int],
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        drop: float = 0.0,
+        attn_drop: float = 0.0,
+        drop_path: float = 0.0,
+        act_layer: str = "GELU",
+        norm_layer: type[LayerNorm] = nn.LayerNorm,
+        use_checkpoint: bool = False,
+    ) -> None:
+        """
+        Args:
+            dim: number of feature channels.
+            num_heads: number of attention heads.
+            window_size: local window size.
+            shift_size: window shift size.
+            mlp_ratio: ratio of mlp hidden dim to embedding dim.
+            qkv_bias: add a learnable bias to query, key, value.
+            drop: dropout rate.
+            attn_drop: attention dropout rate.
+            drop_path: stochastic depth rate.
+            act_layer: activation layer.
+            norm_layer: normalization layer.
+            use_checkpoint: use gradient checkpointing for reduced memory usage.
+        """
+
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        self.use_checkpoint = use_checkpoint
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim,
+            window_size=self.window_size,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            attn_drop=attn_drop,
+            proj_drop=drop,
+        )
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(hidden_size=dim, mlp_dim=mlp_hidden_dim, act=act_layer, dropout_rate=drop, dropout_mode="swin")
+
+    def forward_part1(self, x, mask_matrix):
+        x_shape = x.size()
+        x = self.norm1(x)
+        if len(x_shape) == 5:
+            b, d, h, w, c = x.shape
+            window_size, shift_size = get_window_size((d, h, w), self.window_size, self.shift_size)
+            pad_l = pad_t = pad_d0 = 0
+            pad_d1 = (window_size[0] - d % window_size[0]) % window_size[0]
+            pad_b = (window_size[1] - h % window_size[1]) % window_size[1]
+            pad_r = (window_size[2] - w % window_size[2]) % window_size[2]
+            x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b, pad_d0, pad_d1))
+            _, dp, hp, wp, _ = x.shape
+            dims = [b, dp, hp, wp]
+
+        elif len(x_shape) == 4:
+            b, h, w, c = x.shape
+            window_size, shift_size = get_window_size((h, w), self.window_size, self.shift_size)
+            pad_l = pad_t = 0
+            pad_b = (window_size[0] - h % window_size[0]) % window_size[0]
+            pad_r = (window_size[1] - w % window_size[1]) % window_size[1]
+            x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
+            _, hp, wp, _ = x.shape
+            dims = [b, hp, wp]
+
+        if any(i > 0 for i in shift_size):
+            if len(x_shape) == 5:
+                shifted_x = torch.roll(x, shifts=(-shift_size[0], -shift_size[1], -shift_size[2]), dims=(1, 2, 3))
+            elif len(x_shape) == 4:
+                shifted_x = torch.roll(x, shifts=(-shift_size[0], -shift_size[1]), dims=(1, 2))
+            attn_mask = mask_matrix
+        else:
+            shifted_x = x
+            attn_mask = None
+        x_windows = window_partition(shifted_x, window_size)
+        attn_windows = self.attn(x_windows, mask=attn_mask)
+        attn_windows = attn_windows.view(-1, *(window_size + (c,)))
+        shifted_x = window_reverse(attn_windows, window_size, dims)
+        if any(i > 0 for i in shift_size):
+            if len(x_shape) == 5:
+                x = torch.roll(shifted_x, shifts=(shift_size[0], shift_size[1], shift_size[2]), dims=(1, 2, 3))
+            elif len(x_shape) == 4:
+                x = torch.roll(shifted_x, shifts=(shift_size[0], shift_size[1]), dims=(1, 2))
+        else:
+            x = shifted_x
+
+        if len(x_shape) == 5:
+            if pad_d1 > 0 or pad_r > 0 or pad_b > 0:
+                x = x[:, :d, :h, :w, :].contiguous()
+        elif len(x_shape) == 4:
+            if pad_r > 0 or pad_b > 0:
+                x = x[:, :h, :w, :].contiguous()
+
+        return x
+
+    def forward_part2(self, x):
+        return self.drop_path(self.mlp(self.norm2(x)))
+
+    def load_from(self, weights, n_block, layer):
+        root = f"module.{layer}.0.blocks.{n_block}."
+        block_names = [
+            "norm1.weight",
+            "norm1.bias",
+            "attn.relative_position_bias_table",
+            "attn.relative_position_index",
+            "attn.qkv.weight",
+            "attn.qkv.bias",
+            "attn.proj.weight",
+            "attn.proj.bias",
+            "norm2.weight",
+            "norm2.bias",
+            "mlp.fc1.weight",
+            "mlp.fc1.bias",
+            "mlp.fc2.weight",
+            "mlp.fc2.bias",
+        ]
+        with torch.no_grad():
+            self.norm1.weight.copy_(weights["state_dict"][root + block_names[0]])
+            self.norm1.bias.copy_(weights["state_dict"][root + block_names[1]])
+            self.attn.relative_position_bias_table.copy_(weights["state_dict"][root + block_names[2]])
+            self.attn.relative_position_index.copy_(weights["state_dict"][root + block_names[3]])  # type: ignore
+            self.attn.qkv.weight.copy_(weights["state_dict"][root + block_names[4]])
+            self.attn.qkv.bias.copy_(weights["state_dict"][root + block_names[5]])
+            self.attn.proj.weight.copy_(weights["state_dict"][root + block_names[6]])
+            self.attn.proj.bias.copy_(weights["state_dict"][root + block_names[7]])
+            self.norm2.weight.copy_(weights["state_dict"][root + block_names[8]])
+            self.norm2.bias.copy_(weights["state_dict"][root + block_names[9]])
+            self.mlp.linear1.weight.copy_(weights["state_dict"][root + block_names[10]])
+            self.mlp.linear1.bias.copy_(weights["state_dict"][root + block_names[11]])
+            self.mlp.linear2.weight.copy_(weights["state_dict"][root + block_names[12]])
+            self.mlp.linear2.bias.copy_(weights["state_dict"][root + block_names[13]])
+
+    def forward(self, x, mask_matrix):
+        shortcut = x
+        if self.use_checkpoint:
+            x = checkpoint.checkpoint(self.forward_part1, x, mask_matrix)
+        else:
+            x = self.forward_part1(x, mask_matrix)
+        x = shortcut + self.drop_path(x)
+        if self.use_checkpoint:
+            x = x + checkpoint.checkpoint(self.forward_part2, x)
+        else:
+            x = x + self.forward_part2(x)
+        return x
+
+
+class PatchMergingV2(nn.Module):
+    """
+    Patch merging layer based on: "Liu et al.,
+    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
+    <https://arxiv.org/abs/2103.14030>"
+    https://github.com/microsoft/Swin-Transformer
+    """
+
+    def __init__(self, dim: int, norm_layer: type[LayerNorm] = nn.LayerNorm, spatial_dims: int = 3) -> None:
+        """
+        Args:
+            dim: number of feature channels.
+            norm_layer: normalization layer.
+            spatial_dims: number of spatial dims.
+        """
+
+        super().__init__()
+        self.dim = dim
+        if spatial_dims == 3:
+            self.reduction = nn.Linear(8 * dim, 2 * dim, bias=False)
+            self.norm = norm_layer(8 * dim)
+        elif spatial_dims == 2:
+            self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+            self.norm = norm_layer(4 * dim)
+
+    def forward(self, x):
+        x_shape = x.size()
+        if len(x_shape) == 5:
+            b, d, h, w, c = x_shape
+            pad_input = (h % 2 == 1) or (w % 2 == 1) or (d % 2 == 1)
+            if pad_input:
+                x = F.pad(x, (0, 0, 0, w % 2, 0, h % 2, 0, d % 2))
+            x = torch.cat(
+                [x[:, i::2, j::2, k::2, :] for i, j, k in itertools.product(range(2), range(2), range(2))], -1
+            )
+
+        elif len(x_shape) == 4:
+            b, h, w, c = x_shape
+            pad_input = (h % 2 == 1) or (w % 2 == 1)
+            if pad_input:
+                x = F.pad(x, (0, 0, 0, w % 2, 0, h % 2))
+            x = torch.cat([x[:, j::2, i::2, :] for i, j in itertools.product(range(2), range(2))], -1)
+
+        x = self.norm(x)
+        x = self.reduction(x)
+        return x
+
+
+class PatchMerging(PatchMergingV2):
+    """The `PatchMerging` module previously defined in v0.9.0."""
+
+    def forward(self, x):
+        x_shape = x.size()
+        if len(x_shape) == 4:
+            return super().forward(x)
+        if len(x_shape) != 5:
+            raise ValueError(f"expecting 5D x, got {x.shape}.")
+        b, d, h, w, c = x_shape
+        pad_input = (h % 2 == 1) or (w % 2 == 1) or (d % 2 == 1)
+        if pad_input:
+            x = F.pad(x, (0, 0, 0, w % 2, 0, h % 2, 0, d % 2))
+        x0 = x[:, 0::2, 0::2, 0::2, :]
+        x1 = x[:, 1::2, 0::2, 0::2, :]
+        x2 = x[:, 0::2, 1::2, 0::2, :]
+        x3 = x[:, 0::2, 0::2, 1::2, :]
+        x4 = x[:, 1::2, 0::2, 1::2, :]
+        x5 = x[:, 0::2, 1::2, 0::2, :]
+        x6 = x[:, 0::2, 0::2, 1::2, :]
+        x7 = x[:, 1::2, 1::2, 1::2, :]
+        x = torch.cat([x0, x1, x2, x3, x4, x5, x6, x7], -1)
+        x = self.norm(x)
+        x = self.reduction(x)
+        return x
+
+
+MERGING_MODE = {"merging": PatchMerging, "mergingv2": PatchMergingV2}
+
+
+def compute_mask(dims, window_size, shift_size, device):
+    """Computing region masks based on: "Liu et al.,
+    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
+    <https://arxiv.org/abs/2103.14030>"
+    https://github.com/microsoft/Swin-Transformer
+
+     Args:
+        dims: dimension values.
+        window_size: local window size.
+        shift_size: shift size.
+        device: device.
+    """
+
+    cnt = 0
+
+    if len(dims) == 3:
+        d, h, w = dims
+        img_mask = torch.zeros((1, d, h, w, 1), device=device)
+        for d in slice(-window_size[0]), slice(-window_size[0], -shift_size[0]), slice(-shift_size[0], None):
+            for h in slice(-window_size[1]), slice(-window_size[1], -shift_size[1]), slice(-shift_size[1], None):
+                for w in slice(-window_size[2]), slice(-window_size[2], -shift_size[2]), slice(-shift_size[2], None):
+                    img_mask[:, d, h, w, :] = cnt
+                    cnt += 1
+
+    elif len(dims) == 2:
+        h, w = dims
+        img_mask = torch.zeros((1, h, w, 1), device=device)
+        for h in slice(-window_size[0]), slice(-window_size[0], -shift_size[0]), slice(-shift_size[0], None):
+            for w in slice(-window_size[1]), slice(-window_size[1], -shift_size[1]), slice(-shift_size[1], None):
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+    mask_windows = window_partition(img_mask, window_size)
+    mask_windows = mask_windows.squeeze(-1)
+    attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+    attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+    return attn_mask
+
+
+class BasicLayer(nn.Module):
+    """
+    Basic Swin Transformer layer in one stage based on: "Liu et al.,
+    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
+    <https://arxiv.org/abs/2103.14030>"
+    https://github.com/microsoft/Swin-Transformer
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        depth: int,
+        num_heads: int,
+        window_size: Sequence[int],
+        drop_path: list,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = False,
+        drop: float = 0.0,
+        attn_drop: float = 0.0,
+        norm_layer: type[LayerNorm] = nn.LayerNorm,
+        downsample: nn.Module | None = None,
+        use_checkpoint: bool = False,
+    ) -> None:
+        """
+        Args:
+            dim: number of feature channels.
+            depth: number of layers in each stage.
+            num_heads: number of attention heads.
+            window_size: local window size.
+            drop_path: stochastic depth rate.
+            mlp_ratio: ratio of mlp hidden dim to embedding dim.
+            qkv_bias: add a learnable bias to query, key, value.
+            drop: dropout rate.
+            attn_drop: attention dropout rate.
+            norm_layer: normalization layer.
+            downsample: an optional downsampling layer at the end of the layer.
+            use_checkpoint: use gradient checkpointing for reduced memory usage.
+        """
+
+        super().__init__()
+        self.window_size = window_size
+        self.shift_size = tuple(i // 2 for i in window_size)
+        self.no_shift = tuple(0 for i in window_size)
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+        self.blocks = nn.ModuleList(
+            [
+                SwinTransformerBlock(
+                    dim=dim,
+                    num_heads=num_heads,
+                    window_size=self.window_size,
+                    shift_size=self.no_shift if (i % 2 == 0) else self.shift_size,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    drop=drop,
+                    attn_drop=attn_drop,
+                    drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                    norm_layer=norm_layer,
+                    use_checkpoint=use_checkpoint,
+                )
+                for i in range(depth)
+            ]
+        )
+        self.downsample = downsample
+        if callable(self.downsample):
+            self.downsample = downsample(dim=dim, norm_layer=norm_layer, spatial_dims=len(self.window_size))
+
+    def forward(self, x):
+        x_shape = x.size()
+        if len(x_shape) == 5:
+            b, c, d, h, w = x_shape
+            window_size, shift_size = get_window_size((d, h, w), self.window_size, self.shift_size)
+            x = rearrange(x, "b c d h w -> b d h w c")
+            dp = int(np.ceil(d / window_size[0])) * window_size[0]
+            hp = int(np.ceil(h / window_size[1])) * window_size[1]
+            wp = int(np.ceil(w / window_size[2])) * window_size[2]
+            attn_mask = compute_mask([dp, hp, wp], window_size, shift_size, x.device)
+            for blk in self.blocks:
+                x = blk(x, attn_mask)
+            x = x.view(b, d, h, w, -1)
+            if self.downsample is not None:
+                x = self.downsample(x)
+            x = rearrange(x, "b d h w c -> b c d h w")
+
+        elif len(x_shape) == 4:
+            b, c, h, w = x_shape
+            window_size, shift_size = get_window_size((h, w), self.window_size, self.shift_size)
+            x = rearrange(x, "b c h w -> b h w c")
+            hp = int(np.ceil(h / window_size[0])) * window_size[0]
+            wp = int(np.ceil(w / window_size[1])) * window_size[1]
+            attn_mask = compute_mask([hp, wp], window_size, shift_size, x.device)
+            for blk in self.blocks:
+                x = blk(x, attn_mask)
+            x = x.view(b, h, w, -1)
+            if self.downsample is not None:
+                x = self.downsample(x)
+            x = rearrange(x, "b h w c -> b c h w")
+        return x
+
+
+class SwinTransformer(nn.Module):
+    """
+    Swin Transformer based on: "Liu et al.,
+    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
+    <https://arxiv.org/abs/2103.14030>"
+    https://github.com/microsoft/Swin-Transformer
+    """
+
+    def __init__(
+        self,
+        in_chans: int,
+        embed_dim: int,
+        window_size: Sequence[int],
+        patch_size: Sequence[int],
+        depths: Sequence[int],
+        num_heads: Sequence[int],
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        drop_rate: float = 0.0,
+        attn_drop_rate: float = 0.0,
+        drop_path_rate: float = 0.0,
+        norm_layer: type[LayerNorm] = nn.LayerNorm,
+        patch_norm: bool = False,
+        use_checkpoint: bool = False,
+        spatial_dims: int = 3,
+        downsample="merging",
+        use_v2=False,
+    ) -> None:
+        """
+        Args:
+            in_chans: dimension of input channels.
+            embed_dim: number of linear projection output channels.
+            window_size: local window size.
+            patch_size: patch size.
+            depths: number of layers in each stage.
+            num_heads: number of attention heads.
+            mlp_ratio: ratio of mlp hidden dim to embedding dim.
+            qkv_bias: add a learnable bias to query, key, value.
+            drop_rate: dropout rate.
+            attn_drop_rate: attention dropout rate.
+            drop_path_rate: stochastic depth rate.
+            norm_layer: normalization layer.
+            patch_norm: add normalization after patch embedding.
+            use_checkpoint: use gradient checkpointing for reduced memory usage.
+            spatial_dims: spatial dimension.
+            downsample: module used for downsampling, available options are `"mergingv2"`, `"merging"` and a
+                user-specified `nn.Module` following the API defined in :py:class:`monai.networks.nets.PatchMerging`.
+                The default is currently `"merging"` (the original version defined in v0.9.0).
+            use_v2: using swinunetr_v2, which adds a residual convolution block at the beginning of each swin stage.
+        """
+
+        super().__init__()
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.patch_norm = patch_norm
+        self.window_size = window_size
+        self.patch_size = patch_size
+        self.patch_embed = PatchEmbed(
+            patch_size=self.patch_size,
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None,  # type: ignore
+            spatial_dims=spatial_dims,
+        )
+        self.pos_drop = nn.Dropout(p=drop_rate)
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]
+        self.use_v2 = use_v2
+        self.layers1 = nn.ModuleList()
+        self.layers2 = nn.ModuleList()
+        self.layers3 = nn.ModuleList()
+        self.layers4 = nn.ModuleList()
+        if self.use_v2:
+            self.layers1c = nn.ModuleList()
+            self.layers2c = nn.ModuleList()
+            self.layers3c = nn.ModuleList()
+            self.layers4c = nn.ModuleList()
+        down_sample_mod = look_up_option(downsample, MERGING_MODE) if isinstance(downsample, str) else downsample
+        for i_layer in range(self.num_layers):
+            layer = BasicLayer(
+                dim=int(embed_dim * 2**i_layer),
+                depth=depths[i_layer],
+                num_heads=num_heads[i_layer],
+                window_size=self.window_size,
+                drop_path=dpr[sum(depths[:i_layer]) : sum(depths[: i_layer + 1])],
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                norm_layer=norm_layer,
+                downsample=down_sample_mod,
+                use_checkpoint=use_checkpoint,
+            )
+            if i_layer == 0:
+                self.layers1.append(layer)
+            elif i_layer == 1:
+                self.layers2.append(layer)
+            elif i_layer == 2:
+                self.layers3.append(layer)
+            elif i_layer == 3:
+                self.layers4.append(layer)
+            if self.use_v2:
+                layerc = UnetrBasicBlock(
+                    spatial_dims=3,
+                    in_channels=embed_dim * 2**i_layer,
+                    out_channels=embed_dim * 2**i_layer,
+                    kernel_size=3,
+                    stride=1,
+                    norm_name="instance",
+                    res_block=True,
+                )
+                if i_layer == 0:
+                    self.layers1c.append(layerc)
+                elif i_layer == 1:
+                    self.layers2c.append(layerc)
+                elif i_layer == 2:
+                    self.layers3c.append(layerc)
+                elif i_layer == 3:
+                    self.layers4c.append(layerc)
+
+        self.num_features = int(embed_dim * 2 ** (self.num_layers - 1))
+
+    def proj_out(self, x, normalize=False):
+        if normalize:
+            x_shape = x.size()
+            if len(x_shape) == 5:
+                n, ch, d, h, w = x_shape
+                x = rearrange(x, "n c d h w -> n d h w c")
+                x = F.layer_norm(x, [ch])
+                x = rearrange(x, "n d h w c -> n c d h w")
+            elif len(x_shape) == 4:
+                n, ch, h, w = x_shape
+                x = rearrange(x, "n c h w -> n h w c")
+                x = F.layer_norm(x, [ch])
+                x = rearrange(x, "n h w c -> n c h w")
+        return x
+
+    def forward(self, x, normalize=True):
+        x0 = self.patch_embed(x)
+        x0 = self.pos_drop(x0)
+        x0_out = self.proj_out(x0, normalize)
+        if self.use_v2:
+            x0 = self.layers1c[0](x0.contiguous())
+        x1 = self.layers1[0](x0.contiguous())
+        x1_out = self.proj_out(x1, normalize)
+        if self.use_v2:
+            x1 = self.layers2c[0](x1.contiguous())
+        x2 = self.layers2[0](x1.contiguous())
+        x2_out = self.proj_out(x2, normalize)
+        if self.use_v2:
+            x2 = self.layers3c[0](x2.contiguous())
+        x3 = self.layers3[0](x2.contiguous())
+        x3_out = self.proj_out(x3, normalize)
+        if self.use_v2:
+            x3 = self.layers4c[0](x3.contiguous())
+        x4 = self.layers4[0](x3.contiguous())
+        x4_out = self.proj_out(x4, normalize)
+        return [x0_out, x1_out, x2_out, x3_out, x4_out]

DCA/utils.py

@@ -0,0 +1,213 @@
+import numpy as np
+import torch
+import torch.nn as nn
+import matplotlib.pyplot as plt
+import nibabel as nib
+from collections import defaultdict
+from scipy.sparse.linalg import svds
+import torch.nn.functional as F
+import torch.optim as optim
+from sklearn.metrics import normalized_mutual_info_score as nmi_score
+from sklearn.cluster import SpectralClustering
+
+from scipy.sparse import csr_matrix
+from scipy.sparse.linalg import eigsh
+from sklearn.cluster import KMeans
+from torch_geometric.utils import to_scipy_sparse_matrix
+from scipy.sparse import identity
+from scipy.sparse.linalg import lobpcg
+from sklearn.cluster import AgglomerativeClustering
+
+from scipy import ndimage
+from scipy.optimize import linear_sum_assignment
+
+
+def weighted_bfs_connected_clustering(edge_index, edge_weight, num_nodes, k, seed=None):
+
+    if seed is not None:
+        np.random.seed(seed)
+        random.seed(seed)
+
+    if torch.is_tensor(edge_index):
+        edge_index = edge_index.cpu().numpy()
+    if torch.is_tensor(edge_weight):
+        edge_weight = edge_weight.cpu().numpy()
+
+    graph = defaultdict(list)
+    E = edge_index.shape[1]
+    for idx in range(E):
+        i, j = int(edge_index[0, idx]), int(edge_index[1, idx])
+        w = float(edge_weight[idx])
+        graph[i].append((j, w))
+        graph[j].append((i, w))
+
+    labels = np.full(num_nodes, -1, dtype=int)
+    max_size = int(np.ceil(num_nodes / k))
+    cluster_sizes = [0] * k
+    heaps = [ [] for _ in range(k) ]
+
+    seeds = np.random.choice(num_nodes, k, replace=False)
+    for cid, seed_node in enumerate(seeds):
+        labels[seed_node] = cid
+        cluster_sizes[cid] += 1
+        for nbr, w in graph[seed_node]:
+            if labels[nbr] == -1:
+                heapq.heappush(heaps[cid], (-w, nbr))
+
+    assigned = k  
+
+    while assigned < num_nodes:
+        made_progress = False
+        for cid in range(k):
+            if cluster_sizes[cid] >= max_size:
+                continue
+            heap = heaps[cid]
+            while heap:
+                neg_w, node = heapq.heappop(heap)
+                if labels[node] == -1:
+                    labels[node] = cid
+                    cluster_sizes[cid] += 1
+                    assigned += 1
+                    for nbr, w in graph[node]:
+                        if labels[nbr] == -1:
+                            heapq.heappush(heap, (-w, nbr))
+                    made_progress = True
+                    break
+        if not made_progress:
+            break
+
+    for node in range(num_nodes):
+        if labels[node] == -1:
+            nbr_labels = [labels[nbr] for nbr, _ in graph[node] if labels[nbr] != -1]
+            if nbr_labels:
+                cid = min(nbr_labels, key=lambda x: cluster_sizes[x])
+            else:
+                cid = int(np.argmin(cluster_sizes))
+            labels[node] = cid
+            cluster_sizes[cid] += 1
+
+    return labels
+
+
+
+def match_labels(y_true, y_pred, n_clusters):
+
+    cost = np.zeros((n_clusters, n_clusters), dtype=int)
+    for i in range(n_clusters):
+        for j in range(n_clusters):
+            cost[i, j] = np.sum((y_true == i) & (y_pred == j))
+    row_ind, col_ind = linear_sum_assignment(-cost)
+    mapping = dict(zip(col_ind, row_ind))
+    return np.vectorize(lambda x: mapping.get(x, x))(y_pred)
+
+
+
+def large_constrcut(mask):
+
+    binary_mask = (mask > 0).astype(int)
+
+    structure = ndimage.generate_binary_structure(3, 1)  
+    labeled_mask, num_labels = ndimage.label(binary_mask, structure=structure)
+
+    region_sizes = ndimage.sum(binary_mask, labeled_mask, index=np.arange(1, num_labels + 1))
+
+    if num_labels == 0:
+        largest_label = 0
+    else:
+        largest_label = np.argmax(region_sizes) + 1  
+    large_mask = (labeled_mask == largest_label).astype(int)     
+    print('total voxels:', len(np.where(large_mask>0)[0]))
+    return large_mask
+
+
+def build_spatial_graph_and_weights(edge_index,features):
+
+    src, tgt = edge_index
+    feat_src = features[src]
+    feat_tgt = features[tgt]
+    x_centered = feat_src - feat_src.mean()
+    y_centered = feat_tgt - feat_tgt.mean()
+    edge_weight = F.cosine_similarity(x_centered, y_centered)
+    return edge_weight
+
+
+
+def sparse_spectral_clustering(edge_index, edge_weight, num_nodes, k=2):
+
+    adj_sparse = to_scipy_sparse_matrix(
+        edge_index, edge_weight, num_nodes=num_nodes
+    )  
+
+    degree = np.array(adj_sparse.sum(axis=1)).flatten()  
+    D = csr_matrix((degree, (np.arange(num_nodes), np.arange(num_nodes))), shape=(num_nodes, num_nodes))
+    L = D - adj_sparse  
+
+    _, eigenvectors = eigsh(L.astype(np.float64), k=k, which='LM',sigma=0, maxiter=10000) 
+
+    kmeans = KMeans(n_clusters=k)
+    labels = kmeans.fit_predict(eigenvectors)
+    return labels
+
+
+
+def validate(input_file, label_file):
+
+    if input_file.endswith('.zst'):
+        input_matrix = load_zstd_file(input_file)
+    elif input_file.endswith('.npy'):
+        input_matrix = np.load(input_file)
+    else:
+        input_matrix = nib.load(input_file).get_fdata()
+    input_matrix = input_matrix[:,:,:,:300]
+
+    label_matrix = label_file
+
+    unique_labels = np.unique(label_matrix)
+    unique_labels = unique_labels[unique_labels > 0]
+    print('roi number:',len(unique_labels))
+    fc_means = {}
+    voxel_counts = {}
+    silhouette_scores = {} 
+
+    for label in unique_labels:
+        voxel_indices = np.argwhere(label_matrix == label)
+        if voxel_indices.shape[0] < 2:
+            fc_means[label] = np.nan
+            voxel_counts[label] = len(voxel_indices)
+            silhouette_scores[label] = np.nan
+            continue
+        time_series = input_matrix[voxel_indices[:, 0], voxel_indices[:, 1], voxel_indices[:, 2], :]
+        n_voxels = time_series.shape[0]
+
+        fc_mean, fc_mat = calculate_fc_mean(time_series)
+        fc_means[label] = fc_mean
+        voxel_counts[label] = len(voxel_indices)
+
+    weighted_sum_fc = 0.0
+    valid_voxels = 0 
+    for label in fc_means:
+        if not np.isnan(fc_means[label]):  
+            weighted_sum_fc += fc_means[label] * voxel_counts[label]
+            valid_voxels += voxel_counts[label]
+    weighted_mean_fc = weighted_sum_fc / valid_voxels if valid_voxels > 0 else np.nan
+
+
+    return weighted_mean_fc
+
+
+def calculate_fc_mean(time_series):
+    if time_series.shape[0] < 2:
+        return np.nan, np.nan
+    
+    fc_matrix = np.corrcoef(time_series)
+    
+    if fc_matrix.ndim != 2 or fc_matrix.shape[0] != fc_matrix.shape[1]:
+        return np.nan, np.nan  
+    
+    lower_triangle_indices = np.tril_indices(fc_matrix.shape[0], -1)
+    lower_triangle_values = fc_matrix[lower_triangle_indices]
+    
+    # positive_fc_values = lower_triangle_values[lower_triangle_values > 0]
+    # mean_fc = np.mean(positive_fc_values) if positive_fc_values.size > 0 else 0
+    mean_fc = np.mean(lower_triangle_values)
+    return mean_fc, fc_matrix

README.md

@@ -0,0 +1,74 @@
+
+# DCA: Graph-Guided Deep Embedding Clustering for Brain Atlases 
+<a href="https://arxiv.org/abs/2509.01426"><img src="https://img.shields.io/badge/Paper-Arxiv-darkred.svg" alt="Paper"></a>
+<a href="https://polyformproject.org/licenses/noncommercial/1.0.0/"><img src="https://img.shields.io/badge/License-PolyForm--NC--1.0.0-blueviolet.svg" alt="License: PolyForm Noncommercial 1.0.0"></a>
+
+This repository contains the official implementation of our **NeurIPS 2025** paper (5554, Poster) [**DCA: Graph-Guided Deep Embedding Clustering for Brain Atlases**](https://arxiv.org/abs/2509.01426). The method integrates pretraining and spatial graph-based constraints to generate anatomically and functionally meaningful brain atlases.
+
+![](fig.png)
+
+
+## DCA Usage
+
+0. Dependencies
+
+We recommend using `Python == 3.9.21`.
+
+All required packages are listed in `DCA/req_trim.txt`. 
+
+
+1. Prepare Data
+   
+
+Use `DCA/data/data_preparation.ipynb` to prepare both fMRI and mask data: fMRI inputs should be resting-state scans normalized to the MNI152 space with at least 300 TRs (note that if TR ≠ 0.72 s, it is recommended to resample temporally to 0.72 s beforehand, as this step is not included in the notebook), and mask inputs should be the corresponding FreeSurfer `aparc+aseg.nii.gz` already registered in MNI152 space.
+
+Place preprocessed 4D fMRI volumes in `DCA/data/fmri/`, we have placed a demo fMRI.
+
+Place your ROI masks in `DCA/data/mask/`. This implementation supports customization for gray matter, white matter, and subcortex-specific atlases. We have placed a demo mask.
+
+Ensure data/sub_test.txt contains the list of subject IDs (one per line), we have placed a demo text.
+
+The pretrained model (swin_model_epoch_30.pth) is automatically loaded if present. We conduct pre-training using monai (https://monai.io/) and customize some functions through DCA/swin_unetr.py
+
+
+2. Run DCA
+
+```bash
+python main.py
+```
+
+This will generate subject-level brain parcellations using the provided pretrained model. Results will be saved to results/demo/.
+
+ Command-line Options
+
+You can customize key inference settings via arguments in `main.py`. The main options are:
+
+- `-k`, `--n_clusters`: Number of parcels to generate (default: `100`)
+- `-e`, `--epoch`: Maximum training epochs (default: `8`)
+- `-v`, `--vali`: Whether to keep the best atlas based homogeneity (default: `True`)
+
+ Validation requires more computing resources. If `--vali` is set to `False`, we recommend using `--epoch < 10` to avoid overfitting.
+
+
+3. Results
+
+Output can be found in data/results/demo
+
+## AtlaScore Usage
+### Downstream
+For all the operation instructions, please see `demo.ipynb`.
+
+### Similarity
+- Provide fMRI data `shape=(x,y,z,t)` and the atlas file `shape=(x,y,z)` to be evaluated, modify the paths in `eva.py`, and run the command `python eva.py`.
+- To evaluate DCBC, you must provide the following files mapped to cortical surface vertices: fMRI data, vertex distance file, and parcellation file. Refer to [Zhi et al.](https://github.com/DiedrichsenLab/DCBC) for more detailed methodological instructions.
+
+## Citations
+If you find our work useful for your research, please consider citing our paper:
+```bibtex
+@article{wang2025dca,
+  title={DCA: Graph-Guided Deep Embedding Clustering for Brain Atlases},
+  author={Wang, Mo and Peng, Kaining and Tang, Jingsheng and Wen, Hongkai and Liu, Quanying},
+  journal={arXiv preprint arXiv:2509.01426},
+  year={2025}
+}
+```

Release/DCA100.nii.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55cb73621697360e6107394f51fc0606367c1c43a88d96c8693c0d810d91d8d1
+size 132151

Release/DCA200.nii.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:922ea7b5774ba2ba553767e084efcaef5f96df2921b24db046c5f17716163b9a
+size 142767

Release/DCA360.nii.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac8661c8883211591f562ce7f1c0e97655bd2db35154d0f23b0d8f5630aba056
+size 152187

Release/DCA400.nii.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4225d5c96e71db74bf5400936047a433e7f500d74c1e3adcaa11e53e921406e3
+size 154316

Release/DCA500.nii.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:771969400101648c0d8b36909999520750136143e8544c2fda8d7e4da31895ba
+size 158566