Lightweight and Interpretable Hybrid CNN-Transformer Model for Multiclass Breast Cancer Classification in Histopathological Images
Аннотация
Accurate and interpretable classification of breast cancer (BC) histopathological subtypes is vital for supporting clinical decision-making. This study aims to develop an efficient and explainable deep learning framework capable of magnification-independent analysis of histopathological images, addressing the challenges of dataset variability and lack of interpretability. We propose HistoMixFormer, a hybrid CNN–Transformer architecture that integrates a lightweight ConvMixer block, adaptive attention mechanism, and a multi-method interpretability module combining SHAP, Grad-CAM++, and LIME. The model processes all images within a unified magnification-independent pipeline. Performance was evaluated on four benchmark datasets—BreakHis, BACH, Camelyon16, and TCGA-BRCA—using accuracy, macro F1-score, and computational efficiency. Quantitative interpretability assessment was conducted using Pointing Game accuracy and Intersection over Union (IoU). HistoMixFormer achieved 92.85% accuracy on BreakHis, 90.12% on BACH, and 88.47% on Camelyon16, while requiring only 3.7 GFLOPs and 10.3 ms per image. External validation confirmed robust cross-domain generalization across independent datasets. Interpretability analyses demonstrated strong correspondence between attention maps and expert-annotated tumor regions. Ablation experiments validated the incremental benefits of magnification unification and hybrid attention fusion. HistoMixFormer establishes a unified, interpretable, and computationally efficient approach for BC histopathology classification. By achieving high accuracy, balanced generalization, and transparent decision reasoning, it holds strong potential for real-world clinical deployment in digital pathology.