BreakHis-IB: Leakage-Checked Splits Across Magnifications for Robust BreakHis Evaluation

BreakHis is a widely used breast cancer histopathology dataset containing H&E-stained microscopy images across four magnifications (40\u00d7\/100\u00d7\/200\u00d7\/400\u00d7) and eight histological subtypes spanning benign and malignant classes. Reported results on BreakHis often vary due to (i) class and subtype imbalance, (ii) magnification bias, and (iii) split protocols that risk patient leakage. We present BreakHis-IB, a derivative resource that provides patient-held-out evaluation protocols with a subtype-balanced-per-magnification test set, a companion natural-frequency test split, and leakage-checked train\/validation partitions. We release split CSVs, a split-builder notebook\/script, and a fairness notebook that reports per-subtype and per-magnification metrics (macro-F1, balanced accuracy, sensitivity gaps) with optional stain-normalization toggles (Macenko\/Vahadane). BreakHis-IB does not redistribute images; instead, it enables users to recreate the exact splits locally from the official dataset. This standardization improves comparability and supports inclusivity-oriented evaluation in computational pathology.