MedHEB-Bench

This repository hosts the full MedHEB-Bench benchmark — a multimodal medical embedding / retrieval benchmark spanning 2D imaging, 3D imaging, and biomedical text. A small reviewer-facing sample (≈ 2.5 GB) is available separately at MEDHEB/MedHEB-Bench_Sample.

Repository layout

.
├── 2D_Task/       
├── 2D_Images/     
├── 3D_Task/       
├── 3D_Images/      
├── Text_Task/      
├── Train_Data/     
└── Preprocess/   

JSON record format

Test JSON (under 2D_Task/, 3D_Task/, Text_Task/):

{
  "qry_inst": "Find the same category image with the provided image:",
  "qry_text": "",
  "qry_img_path": "6cd606dc52e9.png",
  "tgt_text": ["", "", "", "", ""],
  "tgt_img_path": [
    "82bb8a01935f.png",
    "87b671c6d4c5.png",
    "..."
  ]
}

Train JSONL (under Train_Data/) — chat-style records with messages / images / positive_messages:

{
  "messages": [
    {"role": "system", "content": "Represent the given image with the following question:"},
    {"role": "user",   "content": "<image>What is the diabetic retinopathy severity level observed in this retinal image?"}
  ],
  "images": ["./2D_Images/APTOS/train_images/0024cdab0c1e.png"],
  "positive_messages": [[{"role": "user", "content": "Mild"}]]
}

Image paths in test JSONs resolve relative to the matching sub-folder under 2D_Images/ / 3D_Images/. Train JSONL paths are relative to the repository root.

Restricted datasets — image archives NOT redistributed

Image archives for the following datasets are not redistributed here because their upstream license / data-use agreements restrict re-hosting. A LICENSE_NOTICE.md placeholder remains in each folder pointing to the official source. Task-definition JSONs are kept so the benchmark structure is visible.

• 2D_Images/ChestImagrome/ — Chest-ImaGenome (PhysioNet)
• 2D_Images/chexpert_plus/ — CheXpert-Plus (Stanford AIMI)
• 2D_Images/MIMIC-CXR-Report/ — MIMIC-CXR (PhysioNet)
• 2D_Images/MIMIC-CXR-T/ — MIMIC-CXR-LT (PhysioNet)
• 2D_Images/MIMIC-CXR-VQA/ — MIMIC-CXR-VQA (PhysioNet)
• 2D_Images/VindrCXR/ — VinDr-CXR (PhysioNet)
• 2D_Images/VindrMammo/ — VinDr-Mammo (PhysioNet)
• 2D_Images/RadImageNet-VQA/ — RadImageNet-VQA (HuggingFace)
• 3D_Images/CT_RATE/ — CT-RATE (HuggingFace)
• 3D_Images/MRNet/ — MRNet (Stanford ML Group)

Dataset Download

All test sets used in our benchmark can be downloaded through the links provided in the table. Please note that, for some datasets, we use redistributed versions rather than the original official download sources: Kvasir, LC25000, and MRNet are obtained from Kaggle, while PublicHealthQA, CirrMRI600, and VQA_RAD are obtained from Hugging Face.

Dataset	Dim	Modality	Download
MIMIC-CXR-LT	2D	Chest X-ray
ISIC-2019	2D	Skin disease image
Brain-Tumor-MRI	2D	Brain MRI
PathMNIST	2D	Colon Pathology
ChestMNIST(NIH-ChestXray14)	2D	Chest X-ray
DermaMNIST(HAM10000)	2D	Skin Disease image
OCTMNIST(OCT)	2D	Optical Coherence Tomography
TissueMNIST(BBBC051)	2D	Kidney Cortex Microscope
BloodMNIST	2D	Blood Cell image
OrganAMNIST(LiTS)	2D	Abdominal CT scans(arxial)
OrganCMNIST(LiTS)	2D	Abdominal CT scans(Coronal)
OrganSMNIST(LiTS)	2D	Abdominal CT scans(Sagittal)
Kvasir	2D	Gastrointestinal (GI) tract
APTOS-2019	2D	Optical Coherence Tomography
PanNuke	2D	Pathdology
VinDr-CXR	2D	Chest X-ray
VinDr-Mammo	2D	Breast X-ray
Chest-imagenome	2D	Chest X-ray
Kvasir-SEG	2D	Gastrointestinal (GI) tract
USforKidneyMSBench(CT2USforKidneySeg)	2D	Ultrasound
UltrasoundNerveMSBench	2D	Ultrasound
UWSkinCancerMSBench	2D	Skin Image
Isic2018MSBench(ISIC 2018 Challenge)	2D	Skin Image
PolypGenMSBench	2D	Gastrointestinal (GI) tract
MIMIC-CXR	2D	Chest X-ray
USData	2D	Ultrasound
Path-VQA	2D	Pathdology
PMC-VQA	2D	Multi_modalities
ROCO-VQA	2D	Multi_modalities
MedPIX-VQA	2D	Multi_modalities
RadImageNet-VQA	2D	Multi_modalities
VQA_RAD	2D	Multi_modalities
MIMIC-CXR-VQA	2D	Chest X-ray
MedMCQA	Text	Biological text
PubMedQA	Text	Biological text
MedicalQARetrieval	Text	Biological text
PublicHealthQA	Text	Biological text
MMDental	3D	Dental CBCT
CT-RATE	3D	Chest CT
MRNet	3D	Knee CT
CirrMRI600	3D	3D liver MRI
OrganMNIST(LiTS)	3D	Abdominal CT scans
NoduleMNIST	3D	Chest CT
SynapseMNIST(MitoEM dataset)	3D	Pathdology
RadGenomme-Chest-CT	3D	Chest CT
M3D	3D	Multi_modalities
BraTS-PED	3D	Brain MRI
SynthRAD2023	3D	Brain and Pelvis MRI,CT
HaN-Seg	3D	Head and Neck MRI,CT
Chexpert_plus	2D	Chest X-ray
Retinal Disease Classification	2D	Optical Coherence Tomography
LC25000	2D	Pathlogy
OmniMedVQA	2D	Multi_modalities
BraTS-MEN	3D	Brain MRI

Dataset Preprocess

The following datasets are supported and should be processed using the provided scripts in Preprocess/:

• BraTS-MEN
• BraTS-PED
• CT-RATE
• ChirrMRI600
• HaN-Seg
• MedMNIST
• MRNet
• RadGenomme-Chest-CT
• SynthRAD
• PanNuke
• PublicHealthQA
• ROCO-QA
• ChestImagenome
• Gastrointestinal
• Skin-Image
• Ultrasound
• VinDr-CXR
• VinDr-Mammo

3D_Task

1. BraTS-MEN

• a. Preprocess/3D_Task/BraTS-MEN/cut_slides.py

Description: cut_slides.py — This script converts the original t1n.nii.gz files from each BraTS-MEN case into 2D T1-weighted MRI slices. We follow the test split defined by the impression-text annotations in RadGenome-Brain_MRI. INPUT_ROOT should be set to ./BraTS-MEN-Train.

2. BraTS-PED

• a. Preprocess/3D_Task/BraTS-PED/cut_slides.py

Description: cut_slides.py — This script converts the original t1n.nii.gz and t2w.nii.gz files from each BraTS-PED case into 2D T1-weighted and T2-weighted MRI slices. INPUT_ROOT should be set to ./BraTS2023-PED.

3. CT-RATE

• a. Preprocess/3D_Task/CT-RATE/cut_slides.py
• b. Preprocess/3D_Task/CT-RATE/CT2Xray.py

Description: cut_slides.py — This script converts the original 3D CT volumes from each case into 2D CT slices. CT2Xray.py — This script converts the original 3D CT volumes into 2D chest X-ray projection images. INPUT_ROOT should be set to ./CT_DATA/data_volumes/dataset.

4. ChirrMRI600

• a. Preprocess/3D_Task/ChirrMRI600/cut_slides_split.py

Description: ChirrMRI600 contains liver MRI cases with different cirrhosis grades and healthy controls. We label healthy cases as 'No liver cirrhosis'. Cirrhosis_T1_slide_folders.csv and Cirrhosis_T2_slide_folders.csv define the split after adding the healthy cases. cut_slides_split.py converts 3D liver MRI volumes into 2D slices and organizes them according to the CSV split. Place the two CSV files in the original data folder and set this folder as the input path.

5. HaN-Seg

• a. Preprocess/3D_Task/HaN-Seg/cut_slides.py

Description: cut_slides.py — This script converts the original 3D CT volumes and MRI volumes into 2D slices, the input root shoud be set to ./HaN-Seg/set_1.

6. MNIST

• a. Preprocess/3D_Task/MNIST/unzip_cut_slides.py

Description: unzip_cut_slides.py — This script extracts the downloaded Zenodo .npz files and converts the original 3D .npy files into 2D slices. It supports preprocessing for all 3D MedMNIST classification datasets. The input root set to the download npz files.

7. MRNet

• a. Preprocess/3D_Task/MRNet/cut_slides.py

Description: cut_slides.py — This script converts each 3D knee MRI volume into 2D slices from the axial, coronal, and sagittal views, with 10 slices extracted for each view. We map MRNet labels as follows: abnormal = 0 is labeled as healthy, while cases with abnormal = 1, acl = 0, and meniscus = 0 are labeled as other diseases.

8. RadGenomme-Chest-CT

• a. Preprocess/3D_Task/RadGenomme-Chest-CT/cut_slides.py

Description: cut_slides.py — This script converts the original 3D CT volumes from each case into 2D CT slices. The input root should be set to ./RadGenomme-Chest-CT/dataset.

9. SynthRAD

• a. Preprocess/3D_Task/SynthRAD/cut_slides.py

Description: cut_slides.py — This script converts the ct.nii.gz and mr.nii.gz files from each SynthRAD2023 Task 1 brain and pelvis case into 2D slices. For brain and pelvis data, set the input roots to ./SynthRAD2023/Task1/brain and ./SynthRAD2023/Task1/pelvis, respectively.

Image_Classification

1. MNIST

• a. Preprocess/2D_Task/MNIST/unzip_MNIST.py

Description: unzip_MNIST.py — This script processes all 2D MedMNIST classification datasets. It extracts the downloaded Zenodo .npz files, organizes images according to the official train/validation/test split, and further separates them into class-specific folders. The input root should be set to the path of the downloaded .npz files.

2. PanNuke

• a. Preprocess/2D_Task/PanNuke/Preprocess.py

Description: Preprocess.py — This script merges images from the three .parquet files and organizes them into class-specific folders. PanNuke_split.csv defines the dataset split and should be placed in the downloaded data folder. The input root should be set to the directory containing the three .parquet files.

VQA

1. ROCO_QA

• a. Preprocess/2D_Task/ROCO_QA/unzip.py

Description: unzip.py — This script extracts the five .parquet files. The input root should be set to ./ROCO-QA/data, and FILE_ORDER should match the order specified in the script.

Vision_Grounding

1. ChestImagenome

• a. Preprocess/2D_Task/ChestImagenome/Preprocess.py

Description: Preprocess.py — This script crops the original images according to the official bounding-box annotations. gold_crop is used as the benchmark test set, while silver_sample_crop is used as the training set. The input root should be set to ./Chest_imagenome.

2. Gastrointestinal

• a. Preprocess/2D_Task/Gastrointestinal/Preprocess.py

Description: Preprocess.py — This script converts segmentation masks from PolypGenMSBench and Kvasir-SEG into bounding boxes and cropped images. The split for Kvasir-SEG is defined in split.csv, which should be placed in the data folder. PolypGenMSBench is automatically downloaded and processed. For Kvasir-SEG, set the image root to ./Kvasir-SEG/images and the annotation file to ./Kvasir-SEG/kvasir_bboxes.json.

3. PanNuke

• a. Preprocess/2D_Task/PanNuke_VG/Preprocess.py

Description: Preprocess.py — This script merges images from the three .parquet files and converts segmentation masks into bounding-box crops. The input root should be set to ./PanNuke/data.

4. Skin_Image

• a. Preprocess/2D_Task/Skin_Image/Preprocess.py

Description: Preprocess.py — This script converts segmentation masks from ISIC2018MSBench and UWSkinCancerMSBench into bounding boxes and cropped images. Both datasets are automatically downloaded and processed.

5. Ultrasound

• a. Preprocess/2D_Task/Ultrasound/Preprocess.py

Description: Preprocess.py — This script converts segmentation masks from USforKidneyMSBench and UltrasoundNerveMSBench into bounding boxes and cropped images. Both datasets are automatically downloaded and processed.

6. VindrCXR

• a. Preprocess/2D_Task/VindrCXR/Preprocess.py

Description: Preprocess.py — This script crops the images and follows the official dataset split. The input root should be set to ./VinDr-CXR.

7. VindrMammo

• a. Preprocess/2D_Task/VindrMammo/Preprocess.py

Description: Preprocess.py — This script crops the images and follows the official dataset split. The input root should be set to ./VindrMammo.

Dataset Arrangement

After downloading and preprocessing the datasets, please organize the data according to the folder structure below so that it is compatible with the evaluation code.

MIMIC-CXR Data Setup

For the MIMIC-CXR-based datasets, you only need to download the original MIMIC-CXR image files once. The same downloaded files can then be reused across multiple sub-datasets by creating symbolic links.

The related datasets include:

• MIMIC-CXR-Report
• MIMIC-CXR-T
• MIMIC-CXR-VQA
• ChestImagrome

After downloading MIMIC-CXR, you can do:

cd 2D_Images/(ChestImagrome, MIMIC-CXR-Report, MIMIC-CXR-VQA, MIMIC-CXR-T)
ln -sfn "../mimic_original/2.0.0/files" "files"

for d in p10 p11 p12 p13 p14 p15 p16 p17 p18 p19; do
    ln -sfn "../mimic_original/2.0.0/files/$d" "$d"
done

Organize your data under a single root directory (DATA_BASEDIR):

DATA_BASEDIR/
│
├── 2D_Task/
│   ├── APTOS/
│   │   ├── APTOS_test.json
│   │   ├── APTOS_i2i_test.json
│   │   └── APTOS_t2i_test.json
│   ├── BloodMNIST/
│   │   ├── BloodMNIST_test.json
│   │   ├── BloodMNIST_i2i_test.json
│   │   └── BloodMNIST_t2i_test.json
│   ├── Brain-Tumor-MRI/
│   │   ├── Brain_Tumor_MRI_test.json
│   │   ├── Brain_Tumor_MRI_i2i_test.json
│   │   └── Brain_Tumor_MRI_t2i_test.json
│   ├── ChestImagrome/
│   │   └── Chest_imagenome_test.json
│   ├── ChestMNIST/
│   │   ├── ChestMNIST_test.json
│   │   ├── ChestMNIST_i2i_test.json
│   │   └── ChestMNIST_t2i_test.json
│   ├── ChexpertPlus/
│   │   ├── chexpert_plus_cls_new.json
│   │   ├── chexpert_plus_i2t_new.json
│   │   └── chexpert_plus_t2i_new.json
│   ├── DermaMNIST/
│   │   ├── DermaMNIST_test.json
│   │   ├── DermaMNIST_i2i_test.json
│   │   └── DermaMNIST_t2i_test.json
│   ├── Gastrointestinal/
│   │   └── gastrointestinal_test.json
│   ├── ISIC-2019/
│   │   ├── ISIC_2019_test.json
│   │   ├── ISIC_2019_i2i_test.json
│   │   └── ISIC_2019_t2i_test.json
│   ├── Kvasir/
│   │   ├── Kvasir_test.json
│   │   ├── Kvasir_i2i_test.json
│   │   └── Kvasir_t2i_test.json
│   ├── lc25000/
│   │   ├── LC25000_cls_new.json
│   │   ├── LC25000_i2i_new.json
│   │   └── LC25000_t2i_new.json
│   ├── MedPIX/
│   │   └── MedPix_test.json
│   ├── MIMIC-CXR-Report/
│   │   └── MIMIC-CXR_test.json
│   ├── MIMIC-CXR-T/
│   │   ├── mimic-cxr-lt-test.json
│   │   ├── mimic-cxr-lt-i2i-test.json
│   │   └── mimic-cxr-lt-t2i-test.json
│   ├── MIMIC-CXR-VQA/
│   │   └── mimic_test.json
│   ├── OCTMNIST/
│   │   ├── OCTMNIST_test.json
│   │   ├── OCTMNIST_i2i_test.json
│   │   └── OCTMNIST_t2i_test.json
│   ├── OmniMedVQA_vqa/
│   │   └── OmniMedVQA_vqa_new.json
│   ├── OrganAMNIST/
│   │   ├── OrganAMNIST_test.json
│   │   ├── OrganAMNIST_i2i_test.json
│   │   └── OrganAMNIST_t2i_test.json
│   ├── OrganCMNIST/
│   │   ├── OrganCMNIST_test.json
│   │   ├── OrganCMNIST_i2i_test.json
│   │   └── OrganCMNIST_t2i_test.json
│   ├── OrganSMNIST/
│   │   ├── OrganSMNIST_test.json
│   │   ├── OrganSMNIST_i2i_test.json
│   │   └── OrganSMNIST_t2i_test.json
│   ├── PanNuke/
│   │   ├── pannuke_i2i_test.json
│   │   └── pannuke_t2i_test.json
│   ├── PanNuke_VG/
│   │   └── PanNuke_test.json
│   ├── PathMNIST/
│   │   ├── PathMNIST_test.json
│   │   ├── PathMNIST_i2i_test.json
│   │   └── PathMNIST_t2i_test.json
│   ├── PathVQA/
│   │   └── Path-VQA_test.json
│   ├── PMC/
│   │   └── PMC-VQA_test.json
│   ├── RadImageNet-VQA/
│   │   └── RadLmageNet_test.json
│   ├── REtinal/
│   │   ├── Retinal_i2i_new.json
│   │   └── Retinal_t2i_new.json
│   ├── ROCO-VQA/
│   │   └── ROCO_test.json
│   ├── SkinLesion/
│   │   └── skin_test.json
│   ├── TissueMNIST/
│   │   ├── TissueMNIST_test.json
│   │   ├── TissueMNIST_i2i_test.json
│   │   └── TissueMNIST_t2i_test.json
│   ├── UltrasoundNerve/
│   │   └── Ultrasound_test.json
│   ├── USData/
│   │   └── USData_test.json
│   ├── VindrCXR/
│   │   └── VinDr-CXR_test.json
│   ├── VindrMammo/
│   │   └── VindrMammo_test.json
│   └── VQA-RAD/
│       └── VQA_test.json
│
├── 2D_Images/
│   ├── APTOS/train_images/
│   ├── BloodMNIST/output_images/test/
│   ├── Brain-Tumor-MRI/datasets/Testing/
│   ├── ChestImagrome/
│   ├── ChestMNIST/output_images/test/
│   ├── chexpert_plus/png/
│   ├── DermaMNIST/output_images/test/
│   ├── Gastrointestinal/
│   ├── ISIC-2019/test/ISIC_2019_Test_Input/
│   ├── kvasir-dataset/datasets/test/
│   ├── lc25000/lung_colon_image_set/Test_Set/
│   ├── MedPIX/
│   ├── MIMIC-CXR-Report/
│   ├── MIMIC-CXR-T/
│   ├── MIMIC-CXR-VQA/
│   ├── OCTMNIST/output_images/test/
│   ├── OmniMedVQA_vqa/test/
│   ├── OrganAMNIST/output_images/test/
│   ├── OrganCMNIST/output_images/test/
│   ├── OrganSMNIST/output_images/test/
│   ├── PanNuke/test_images/
│   ├── PanNuke_VG/
│   ├── PathMNIST/output_images/test/
│   ├── PathVQA/test_images/
│   ├── PMC/
│   ├── RadImageNet-VQA/test/
│   ├── REtinal/Test_Set/Test/
│   ├── ROCO-VQA/
│   ├── SkinLesion/
│   ├── TissueMNIST/output_images/test/
│   ├── UltrasoundNerve/
│   ├── USData/
│   ├── VindrCXR/
│   ├── VindrMammo/
│   └── VQA-RAD/test/
│
├── 3D_Task/
│   ├── BraTS2023/
│   │   ├── bratsped_t1_to_t2_test.json
│   │   └── bratsped_t2_to_t1_test.json
│   ├── BraTS_MEN/
│   │   └── BraTS-MEN-i2t_new.json
│   ├── ChirrMRI600/
│   │   ├── Chirr_test.json
│   │   ├── Chirr_i2i_test.json
│   │   └── Chirr_t2i_test.json
│   ├── CirrMRI600_Cross/
│   │   ├── ChirrMRI_test_t1_to_t2.json
│   │   └── ChirrMRI_test_t2_to_t1.json
│   ├── CT_RATE/
│   │   ├── CT_RATE_test.json
│   │   ├── CT_RATE_vqa_test.json
│   │   ├── CT_RATE_t2i_test.json
│   │   └── CT_RATE_i2i_test.json
│   ├── HaN-Seg/
│   │   └── hanseg_mri_to_ct_test.json
│   ├── M3D/
│   │   ├── m3d_3dqa_test.json
│   │   ├── m3d_i2t_test.json
│   │   └── m3d_t2i_test.json
│   ├── MMDental/
│   │   └── MMDental_i2t_test.json
│   ├── MRNet/
│   │   ├── MRNet_test.json
│   │   ├── MRNet_i2i_test.json
│   │   └── MRNet_t2i_test.json
│   ├── NoduleMNIST/
│   │   └── Nodule_test.json
│   ├── Organ3dMNIST/
│   │   ├── Organ_test.json
│   │   ├── Organ_i2i_test.json
│   │   └── Organ_t2i_test.json
│   ├── RadGen_CT/
│   │   ├── RGCC_i2t_test.json
│   │   └── RGCC_test_VQA.json
│   ├── SynapseMNIST/
│   │   └── Synapse_test.json
│   └── SynthRAD/
│       ├── brain_ct_to_mri_test.json
│       ├── brain_mri_to_ct_test.json
│       ├── pelvis_ct_to_mri_test.json
│       └── pelvis_mri_to_ct_test.json
│
├── 3D_Images/
│   ├── BraTS2023/
│   ├── BraTS_MEN/BraTS-MEN-Test/
│   ├── ChirrMRI600/Cirrhosis_T1_slide/
│   ├── CirrMRI600_Cross/
│   ├── CT_RATE/
│   │   ├── valid_fixed_sliced/
│   │   └── valid_fixed_drr/
│   ├── HaN-Seg/
│   ├── M3D/ct_quizze/
│   ├── MMDental/cbct_png/
│   ├── MRNet/valid_images/
│   ├── NoduleMNIST/test_slides/
│   ├── Organ3dMNIST/test_slides/
│   ├── RadGen_CT/valid_preprocessed_sliced/
│   ├── SynapseMNIST/test_slides/
│   └── SynthRAD/images/
│
└── Text_Task/
    ├── ChexpertPlus/
    │   └── chexpert_plus_f2i_new.json
    ├── MedicalQARetrieval/
    │   └── MedicalQARetrieval_test.json
    ├── MedMCQA/
    │   └── medmcqa_test.json
    ├── MIMIC_f2i/
    │   └── MIMIC_f2i_test.json
    ├── MMDental/
    │   └── MMDental_t2t_test.json
    ├── PublicHealthQA/
    │   └── PublicHealthQA_test.json
    └── PubMedQA/
        └── PubMedQA_test.json

License

This benchmark aggregates data drawn from many publicly available medical imaging and text datasets. Each sub-dataset retains the license and data-use terms of its original source. Users are responsible for complying with the upstream license of any sub-dataset they use.