Harmonized multi-modal medical imaging train / validation / test corpus and healthy-tissue reference statistics (chest X-ray, CT, brain MRI)

Description

This dataset provides a harmonized, quality-controlled, multi-modal medical imaging corpus assembled from five public sources spanning three modalities: chest X-ray (NIH ChestX-ray14), CT (DeepLesion, LIDC-IDRI), and brain MRI (BraTS-Africa, OpenBHB). All images have been standardized to a consistent 2D schema (256x256, uint8) with unified metadata fields (patient ID, acquisition site, split), following dataset-specific quality-control procedures including DICOM series completeness/modality auditing, HU intensity-offset correction and clinical windowing for CT, and brain-tissue-occupancy slice filtering for volumetric MRI-to-2D conversion. Also included are precomputed healthy-tissue reference statistics (Sref_mri.npz, Sref_cxr.npz): mean and covariance of unsupervised tissue-composition ratios, computed from confirmed-healthy image subsets (NIH ChestX-ray14 No Finding cases; OpenBHB control subjects), intended as a plausibility baseline for anomaly detection and explainable generative modeling research. Due to file size, the corpus is split across two access points: validation and test partitions plus the Sref statistics are hosted directly in this repository; training partitions (~18GB combined) are provided via a direct external link, as file size exceeds this repository's storage limit. All partitions are additionally fully reproducible from the cited public source repositories using the processing pipeline documented in the accompanying Data in Brief article, which is the authoritative long-term reference for reproducibility. Files hosted in this repository (Mendeley Data): val.zip - validation partitions for all five datasets: chestxray14_val.h5 (chest X-ray, 14-class pathology labels), deeplesion_val.h5 (CT, lesion-type labels), brats2d_val.h5 (brain MRI, 4-channel T1n/T1c/T2f/T2w, tumor category labels), lidcidri2d_val.h5 (CT, lung-windowed), openbhb2d_val.h5 (brain MRI, 62 real acquisition sites, control diagnosis) test.zip - test partitions for the same five datasets, identical structure to val.zip sref.zip - Sref_mri.npz, Sref_cxr.npz: healthy-tissue reference statistics Files hosted via external link (not archival; see also Related Links): train.zip - training partitions for all five datasets: chestxray14_train.h5, deeplesion_train.h5, brats2d_train.h5, lidcidri2d_train.h5, openbhb2d_train.h5 Link: https://drive.google.com/file/d/1xCCiGICA3DEPIAKMCDdIb88EBty0i2gX/view?usp=drive_link

Steps to reproduce

1. Source data was downloaded from: NIH ChestX-ray14 (NIH Clinical Center), DeepLesion (NIH Clinical Center), BraTS-Africa (TCIA, collection "PKG - BraTS-Africa"), LIDC-IDRI (TCIA, via nbia-data-retriever), OpenBHB (Hugging Face Hub, snapshot_download). 2. Each source was audited for quality: LIDC-IDRI series were filtered to Modality=="CT" with >=30 slices (excludes scout/localizer images); OpenBHB subjects were filtered by the quasiraw-corr quality metric (>=0.5) and verified against disk availability. 3. CT images (DeepLesion) were corrected for a 16-bit storage offset (true HU = pixel - 32768) and windowed per-image using the DICOM_windows field from the source metadata. LIDC-IDRI CT volumes were converted to true HU using rescale slope/intercept, then a fixed lung window (center -600, width 1500) was applied per slice. 4. MRI volumes (BraTS-Africa, OpenBHB) were converted from 3D to 2D by selecting axial slices where >5% of pixels exceeded a near-zero background threshold, then normalizing via 1st-99th percentile clipping. 5. All images were resized to 256x256 and stored as uint8 in HDF5, with patient-level 70/15/15 train/validation/test splits (train not included in this deposit). 6. Sref statistics were computed by clustering pixel intensities of confirmed-healthy images into 4 tissue classes via k-means, then averaging the resulting tissue-ratio vectors across up to 5000 sampled images per modality.

Institutions

• Université Hassan 1er

  Casablanca-Settat, Settat

Categories

Related Links

Licence