Explainable Deep Fusion for Breast Cancer Classification in Histopathology Images

Abstract

Breast cancer diagnosis from histopathology images is accurate but labor-intensive and subject to interobserver variability. Deep learning has shown strong performance on digital pathology benchmarks; however, limited transparency, magnification variability, and patient-level data leakage can reduce clinical trust and robustness. This paper proposes an explainable deep fusion framework for magnification-robust breast histopathology classification on the BreakHis dataset. In the experiments, all folds are generated at the patient level, and model selection is performed in two stages: first, single backbones and backbone-pair fusion models are compared without attention; second, attention modules are evaluated on the selected fusion pair. The results identify VGG16ResNet18 fusion without attention as the nal configuration, achieving an accuracy of 0.8890, an F1-score of 0.8888, and an AUC of 0.9326 under mixed-magnication, patient-level 5-fold cross-validation. Although the improvement over the strongest single-backbone baseline is modest and not statistically significant in the fold-level Wilcoxon test, an image-level discordant-case analysis shows that the proposed fusion model corrects more EfficientNetV2 errors than it introduces. To support transparency, Grad-CAM-based case analysis is used to examine corrected, newly misclassified, and persistently misclassified examples, with LIME and SHAP retained as complementary explanation tools when generated for selected cases. These results suggest that feature-level fusion can provide useful gains for some difficult histopathology images while highlighting remaining limitations caused by ambiguous tissue morphology, staining variation, magnification-dependent texture changes, and attention to non-diagnostic regions.