Add BRAX_Dataset for the BRAX Brazilian chest X-ray dataset

[AmlanMishra2004]

Summary

Adds BRAX_Dataset, a loader for BRAX v1.1.0 (Reis et al., Scientific Data 2022, https://doi.org/10.1038/s41597-022-01608-8): 40,967 chest radiographs from Hospital Israelita Albert Einstein, São Paulo, labeled from Brazilian Portuguese reports with a Portuguese adaptation of the CheXpert labeler. It is one of the few non-US/European CXR datasets and a natural external test set for CheXpert/MIMIC-trained models, but xrv has no loader for it.

Because the labels use CheXpert's encoding, BRAX_Dataset follows CheX_Dataset/MIMIC_Dataset closely:

• Same 13 pathologies in the same order (sorted, then "Pleural Effusion" renamed to "Effusion"), so it merges with CheXpert/MIMIC without relabelling.
• -1 (uncertain) becomes NaN; "No Finding" zeroes every other label except Support Devices.
• imgpath is the release's images/ PNG folder, csvpath is master_spreadsheet_update.csv. Images are 8-bit grayscale PNGs (normalize(maxval=255)).
• views filters on ViewPosition; the ~27% of rows with no ViewPosition get "UNKNOWN" via limit_to_selected_views.
• csv gets the usual patientid, age_years (BRAX gives 5-year age groups, with the oldest as the string "85 or more", mapped to 85), sex_male, sex_female.

License / access: BRAX is under the PhysioNet Credentialed Health Data License 1.5.0 (credentialed account, CITI training, signed DUA). This PR ships no BRAX data or labels; csvpath is required, like MIMIC_Dataset. The docstring states the license and access requirements.

Counts: the paper and PhysioNet page say 19,351 patients and 24,959 studies, but the official master_spreadsheet_update.csv has 18,442 distinct PatientIDs and 23,276 distinct AccessionNumbers (the image count, 40,967, matches). The docstring therefore only states the image count.

Two BRAX-specific choices worth flagging:

• max_aspect_ratio=10: 12 labeled images in the release are thin strips (12-118 px on one side, 19:1 to 213:1 aspect ratio) rather than chest radiographs; the bytes match the release's own SHA256SUMS.txt, so this is in the source data. Every other image is below 9:1. Two of them also exceed Pillow's decompression-bomb pixel limit. They are dropped by default using the CSV's Rows/Columns; max_aspect_ratio=None keeps them. This is similar in spirit to the known-bad image list in PC_Dataset, but doesn't hardcode filenames.
• unique_patients uses drop_duplicates("PatientID") rather than groupby("PatientID").first(). groupby().first() takes the first non-null value per column, so for a patient with several images the resulting row can combine labels from different images (e.g. image 1 is "No Finding", image 2 has Cardiomegaly=1, and the merged row gets both). drop_duplicates keeps one real image. Happy to switch to groupby().first() for consistency with the other loaders if you prefer.

Testing

• 2 new tests in tests/test_dataloaders.py using a small synthetic master_spreadsheet_update.csv: default view/unique-patient filtering, the aspect-ratio filter (and turning it off), UNKNOWN views, -1 to NaN, No Finding zeroing, that a second image's labels don't leak into a patient's row, "85 or more" ages, and image loading from the images/ path.
• Full tests/test_dataloaders.py suite passes (23/23).
• Verified against the real BRAX v1.1.0 label CSV: views=["*"], unique_patients=False loads 40,955 images (40,967 without the aspect filter), views=["PA"] loads 13,958 images / 12,640 unique patients, and all views give 18,442 unique patients (every patient ID in the CSV). Labels matched an independent loader image-for-image across ["*"], ["PA"] and ["PA", "AP"]. A real release PNG loads as (1, 1649, 1024) in [-1024, 1024].
• pep8.sh selection (autopep8 --select=E1,E2,E3,W1,W2) reports nothing on the new code.

Made with Cursor