Provenance
Datasets
Public, deidentified, permissively licensed biomedical image datasets. Every field below is read from data/datasets.yaml, which is itself sourced from research/data_provenance.md.
PathMNIST (histopathology)
MedMNIST v2 (Yang et al.), derived from Kather et al. NCT-CRC-HE-100K + CRC-VAL-HE-7K
Known biases
- stain/scanning batch variation between centers
- class imbalance (background/adipose overrepresented)
- patch labels derived from whole-slide pathologist annotation, not independent per-patch review
Citation: See research/literature_matrix.csv rows for Kather et al. (2016/2018/2019) and Yang et al. MedMNIST v1/v2
BloodMNIST (blood cell microscopy)
MedMNIST v2 (Yang et al.), derived from Acevedo et al. peripheral blood cell dataset
Known biases
- class imbalance (platelets and immature granulocytes are minority classes)
- single acquisition protocol per source paper, but this cannot be independently re-verified from the released images alone
Citation: See research/literature_matrix.csv row for Acevedo et al. and Yang et al. MedMNIST v1/v2
Patient/specimen-level leakage
PathMNIST's train/val pool (NCT-CRC-HE-100K) and test pool (CRC-VAL-HE-7K) come from different clinical centers -- a real, source-verified split used as the leakage boundary for every evaluation on that dataset. BloodMNIST exposes no donor identifier in its public release; independence there is assumed at the image level only, which is a documented limitation, not a guarantee.