Datasets:

YongchengYAO
/

MedVision

v1.3.0 adds MSWAL: 484 abdominal CT cases with 42 MedVision configurations. The release provides segmentation and detection plans for gallstones and kidney stones, plus biometry plans for five tumour and lesion labels. It also fixes CT-window selection for several newly introduced label names, so their tumour/lesion figures render in the intended soft-tissue window.

export MedVision_PLANNER_VERSION=latest   # resolves to 1.3.0

Summary

	Change	In one line	Action
Major
1	MSWAL dataset	484 abdominal CT cases, 42 configurations, and five biometry tasks	update to `1.3.0` or `latest` to use it
2	Reproducible download paths	build from the upstream source or use the pinned preprocessed mirror	none
Minor
3	CT figure normalization	new cancer, cyst, and stone labels use the intended CT windows	none
4	Catalogue and validation updates	configuration catalogue grows from 950 to 992 entries	none

MSWAL dataset

MSWAL contributes 484 abdominal CT cases. The upstream dataset.json names a 210-case test split, but those cases were not uploaded; MedVision therefore plans a reproducible split of the available cohort using seed 1024 and a 0.7 training ratio:

Split	Cases
Train	338
Test	146
Total	484

The dataset adds 42 configurations:

Family	Configurations	Scope
Mask-Size	6	segmentation size benchmarks
Box-Size	6	detection size benchmarks
Tumor-Lesion-Size	30	five labels across three anatomical planes
Total	42

The five biometry labels are liver tumour, kidney tumour, pancreatic cancer, liver cyst, and kidney cyst (labels 3–7). Gallstone and kidney stone (labels 1–2) are included for segmentation and detection only.

Download paths

MSWAL has two supported preparation routes:

Route	Source	Pinned revision	Notes
Raw build	`zhaodongwu/MSWAL`	`62c286b0`	`download_raw.py` copies headers, converts images to `uint16`, and reorients to RAS+ before planning
Fast download	`YongchengYAO/MSWAL-Lite`	`39fb50b6`	`download_fast.py` retrieves the prepared images and masks

The raw route uses the 484 uploaded imagesTr cases and applies the split above. Reorienting to RAS+ during download ensures image arrays and generated annotations share the same coordinate frame.

CT figure normalization

LABEL_MAP_REGROUP now recognizes pancreatic cancer, liver cyst, gallstone, and kidney stone. Therefore, we can use the intended soft-tissue Hounsfield-unit window for image normalization.

Catalogue and validation

The release registers MSWAL in MedVision.py, including its annotation index, biometry family, package mapping, and version notes. The package version and release frontier are now 1.3.0.

The published catalogue now contains 992 configurations, up from 950. Validator expectations were updated to 75 annotation-resolution pairs across 31 datasets. The release was checked with:

test_annotation_resolution: 440 / 440 checks passed
test_tl_ack_gate: 16 / 16 checks passed

For maintainers

The MSWAL build recipe is registered in scripts/gen-annotations/dataset_specs.py; use the raw route when rebuilding the dataset from source. The package is exposed through medvision_ds.datasets, and the preprocessed archive, benchmark plans, landmarks, and regenerated figures are published in Datasets/MSWAL.zip.

Release v1.3.0