Deep Learning-Based Multi-Class Breast Cancer Histopathological Image Classification: A Review

Abstract

Breast cancer histopathology is a challenging computer-vision domain because diagnostically relevant evidence is distributed across cells, tissue architecture, staining patterns, and multiple microscopic scales. The availability of digital histopathology benchmarks has enabled a transition from handcrafted image descriptors toward machine learning, convolutional neural networks, transfer learning, attention mechanisms, and increasingly multimodal artificial intelligence. This review examines the development of multi-class breast histopathological image classification with particular attention to the problem addressed by an InceptionResNetV2-CBAM framework using the BreakHis benchmark. The review organizes the field across dataset design, class taxonomy, image preprocessing, class balancing, data augmentation, convolutional feature extraction, transfer learning, attention-based feature refinement, evaluation methodology, explainability, and translational considerations. It also considers how recent breast-cancer AI research is expanding into ensemble learning, multimodal fusion, subgroup analysis, biosensing, spectroscopy, and specialized architectures. The literature indicates that high internal accuracy should not be interpreted independently from the experimental protocol. Patient-wise partitioning, training-only resampling, external validation, calibration, probability-based ROC/AUC, and component-level ablation are increasingly important for determining whether a high-performing benchmark system generalizes beyond its development data. The reviewed evidence suggests that the next generation of breast histopathology systems will need to combine strong representation learning with methodological rigor, interpretable error analysis, and validation across heterogeneous institutions and acquisition conditions. The review concludes by defining a practical research agenda for robust multi-class breast histopathological image classification and by positioning the InceptionResNetV2-CBAM design as one representative integrated architecture within that broader research landscape.