Patient-Disjoint Evaluation for Trustworthy and Explainable Breast Cancer Histopathology Classification

Abstract

Breast cancer diagnosis from histopathological images increasingly relies on deep learning, with recent studies reporting classification accuracies above 95% on the benchmark BreaKHis dataset. However, much of this literature uses image-level splits that let the same patient's images appear in both training and test sets a source of inflation that is not merely statistical but potentially clinically misleading if taken as evidence of real-world reliability and rarely pairs high accuracy with a rigorous account of model interpretability. This study addresses that gap with an explainable deep learning pipeline for binary (benign/malignant) BreaKHis classification under a patient-disjoint split. An EfficientNetB0 backbone, pretrained on ImageNet, was fine-tuned in two phases with class-weighted loss and Reinhard-based stain normalization. Evaluated under this leakage-safe protocol, the baseline model achieved a test-set balanced accuracy of 0.765 markedly lower than the above-95% figures reported elsewhere, yet a more trustworthy generalization estimate together with an F1-score of 0.777 for the malignant class, a ROC-AUC of 0.849, and a PR-AUC of 0.940, with the best checkpoint selected before overfitting deepened during fine-tuning. A regularization ablation against a more heavily regularized variant showed a narrower generalization gap but lower test performance, confirming the unregularized configuration as the more suitable final model. Grad-CAM was applied to the model's final convolutional layer to visualize the regions driving individual predictions, supporting qualitative inspection of correct and misclassified cases. These findings argue for combining leakage-aware evaluation with explainability as a joint, rather than separate, requirement for trustworthy histopathological classification models.