Explainable Deep Learning-Based Breast Cancer Classification Using Histopathological Images: A Comparative Study of CNN and Vision Transformer Architectures

This study presents a rigorous framework for improving the accuracy and clinical trustworthiness of computer-aided breast cancer diagnosis. The primary research contributions are organized below:

Comprehensive Architectural Evaluation: The work conducts a detailed performance comparison of six leading deep learning models—DenseNet121, EfficientNetB0, ResNet50, Xception, MobileNetV3, and Vision Transformer (ViT-B/16)—on the BreakHis histopathological dataset, providing a clear benchmark for binary breast cancer classification.

Strict Prevention of Data Leakage: By adopting a patient-wise data splitting strategy rather than a standard image-wise split, the study ensures that models are evaluated on completely unseen patients. This prevents artificially inflated performance metrics and provides a realistic measure of how the models will generalize in actual clinical practice.

Integration of Explainable AI (XAI): To overcome the traditional “black-box” limitation of deep learning, the framework integrates Grad-CAM++ to generate visual explanations of the model predictions. This highlights the exact cellular areas driving the classification, offering crucial interpretability for medical practitioners.

Optimization for Resource-Constrained Regions: The study successfully identifies lightweight, highly efficient pipelines (such as EfficientNetB0 and Xception) that balance strong diagnostic accuracy with low computational footprints, making them highly suitable for deployment in infrastructure-limited environments like Bangladesh.