chicagoypark/Second_Baseline_Measures

Second Baseline Measures — segmentation predictions

Raw predicted segmentation masks from baseline runs of a multimodal medical segmentation evaluation. Each run is archived as a single .zip under preds/ (one archive instead of tens of thousands of loose files, so the repo stays fast to clone and browse).

Contents

Path	Uncompressed	Files	Description
preds/mvg-mm-trainfixed_iter0000095.zip	3.68 GiB	60,080	mvg-mm-trainfixed predictions, checkpoint iter 95 (35.5 MiB compressed)
preds/mvg-mm-trainfixed_iter0000095/run_hparams.json	1 KB	1	Same file as inside the archive, kept loose so the run config is readable without downloading the archive

---

mvg-mm-trainfixed_iter0000095

Predictions from the mvg-mm-trainfixed multimodal segmentation baseline over held-out BraTS, AMOS and CHAOS slices.

Run configuration

Field	Value
Method	mvg-mm-trainfixed
Task	segmentation
Checkpoint	MVG_Multimodal/joint_2gpu_huge/run_main/checkpoint-95.pth
Model size	huge
Tile size	256
Prompt policy	train_fixed, frac = 0.2
Datasets	amos, chaos (+ four BraTS modalities)
Metrics reported	dice, iou, nsd@1px
Run window	2026-07-10 23:18:01 → 23:31:50

Fixed in-context prompt slice per modality (from run_hparams.json):

Dataset	Modality	Prompt subject	Prompt z
AMOS	CT	ct/amos_0001	56
AMOS	MRI	mri/amos_0507	22
CHAOS	CT	ct/1	30
CHAOS	T1DUAL	mr/1	13
CHAOS	T2SPIR	mr/1	15

Splits

Split	Masks	Cases	Description
m0_t1n	14,470	240	BraTS T1n
m1_t1c	14,470	240	BraTS T1c
m2_t2w	14,470	240	BraTS T2w
m3_t2f	14,470	240	BraTS T2f (FLAIR)
amos_ct	1,217	45	AMOS CT
amos_mri	700	9	AMOS MRI
chaos_ct	184	3	CHAOS CT
chaos_t1dual	45	3	CHAOS T1DUAL
chaos_t2spir	43	3	CHAOS T2SPIR
Total	60,069

Archive layout

mvg-mm-trainfixed_iter0000095/
├── run_hparams.json          # run configuration (table above)
├── input_manifest.csv        # case, modality, input_sha256 — all splits
├── m0_t1n/
│   ├── BraTS-GLI-01309-000_z031.npy
│   ├── ...
│   └── input_hashes.csv      # per-split case → input sha256
├── m1_t1c/  m2_t2w/  m3_t2f/
├── amos_ct/  amos_mri/
└── chaos_ct/  chaos_t1dual/  chaos_t2spir/

Predictions are one .npy per axial slice, named <case>_z<NNN>.npy:

• shape (256, 256)
• dtype uint8
• values {0, 1} — binary foreground mask

The input_sha256 columns identify the exact input slice each prediction was produced from, so predictions can be re-aligned to inputs without relying on filenames.

Loading

Read a single mask straight out of the archive — no need to extract all 3.68 GiB:

import io, zipfile
import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "chicagoypark/Second_Baseline_Measures",
    "preds/mvg-mm-trainfixed_iter0000095.zip",
    repo_type="dataset",
)

with zipfile.ZipFile(path) as zf:
    member = "mvg-mm-trainfixed_iter0000095/m0_t1n/BraTS-GLI-01309-000_z031.npy"
    mask = np.load(io.BytesIO(zf.read(member)))

print(mask.shape, mask.dtype, np.unique(mask))  # (256, 256) uint8 [0 1]

Iterate over a whole split:

with zipfile.ZipFile(path) as zf:
    names = [n for n in zf.namelist()
             if n.startswith("mvg-mm-trainfixed_iter0000095/amos_ct/") and n.endswith(".npy")]
    for n in sorted(names):
        mask = np.load(io.BytesIO(zf.read(n)))

Or extract everything:

hf download chicagoypark/Second_Baseline_Measures \
    preds/mvg-mm-trainfixed_iter0000095.zip --repo-type=dataset --local-dir .
unzip preds/mvg-mm-trainfixed_iter0000095.zip

Integrity

Archive	preds/mvg-mm-trainfixed_iter0000095.zip
SHA-256	6ee73b228b4ecef8f1085ff89c3a8c7c54622d5d5e2a87f1e16b30391f508840
Compressed	37,230,334 bytes (35.5 MiB)
Uncompressed	3,950,220,056 bytes (3.68 GiB)
Entries	60,080 files (+ 10 directory entries)
Compression	deflate (zip -6), 106×

Every member was CRC32-verified bit-identical against the source tree before upload.

License and source data

These are derived model outputs (binary masks) computed over the public BraTS, AMOS and CHAOS datasets. Use of the underlying imaging is governed by each source dataset's own license and data use agreement; please cite and comply with those when using these predictions.