The full MIMIC-CXR-JPG release is ~600 GB, but the imaging battery only needs a few hundred studies per arm, and the arm sizes nest (blind 100 subset of cascade 150 subset of referee 300 subset of solo/cross-modal 600, with 35 nested inside the 600 for the NIH-matched comparison). This adds tooling to select that nested subset deterministically and download only the images it references.
experiments/mimic_cxr_image/build_subset.py: select (rank studies via budget.subsample_cases, write per-arm manifests + download_list.txt) and verify (check images resolve, write sha256 provenance).- Reproducibility = deterministic list (from seed) + pinned per-arm manifests + provenance checksums + a standard-tool
wget one-liner. No bespoke fetcher, no credentials in the repo, mirroring how the NIH lane pins its data.
- All images per study; finding-positive studies by default. Manifest CSVs are gitignored (they embed credentialed report text; regenerate from the seed).
Part of #295. Unblocks #92.
The full MIMIC-CXR-JPG release is ~600 GB, but the imaging battery only needs a few hundred studies per arm, and the arm sizes nest (blind 100 subset of cascade 150 subset of referee 300 subset of solo/cross-modal 600, with 35 nested inside the 600 for the NIH-matched comparison). This adds tooling to select that nested subset deterministically and download only the images it references.
experiments/mimic_cxr_image/build_subset.py:select(rank studies viabudget.subsample_cases, write per-arm manifests +download_list.txt) andverify(check images resolve, write sha256 provenance).wgetone-liner. No bespoke fetcher, no credentials in the repo, mirroring how the NIH lane pins its data.Part of #295. Unblocks #92.