HistoBreastCAD: an integrated deep learning framework for segmentation and multiclass classification of whole-slide breast histopathology images

Breast cancer remains a major cause of cancer-related morbidity and mortality worldwide, and histopathological examination remains the diagnostic gold standard. This study presents HistoBreastCAD, an integrated deep learning framework that combines tissue-region segmentation, four-class histopathology classification, spatial mapping of predictions, and interactive whole-slide image (WSI) review. Segmentation was evaluated using WSI-grouped five-fold cross-validation, while classification used source-image-grouped five-fold cross-validation to prevent source-level data leakage. The final DeepLabV3+ L0 segmentation configuration achieved an out-of-fold micro-Dice of 0.8390, IoU of 0.7226, precision of 0.8278, and recall of 0.8505. For classification, EfficientNetB0 was the strongest individual model, while a frozen equal-weight soft-voting ensemble of EfficientNetB0 and ResNet18-GAP-attention achieved an accuracy of 0.8283 ± 0.0320 and macro-F1 of 0.8282 ± 0.0316, outperforming EfficientNetB0 in all five validation folds. An exploratory reader study using the 100 microscopy images from the held-out BACH test set showed moderate interobserver agreement between two pathologists (Cohen’s κ = 0.5105; n = 95), whereas model-to-reader concordance was lower. The ensemble did not increase reader concordance relative to EfficientNetB0. HistoBreastCAD therefore demonstrates the feasibility of integrating established deep learning components into a source-separated WSI-oriented digital pathology workflow. However, external multi-institutional validation, adjudicated reader studies, and formal explainability assessment are required before claims of clinical generalizability or utility can be supported.