Combining Patch-Based CNN Models with Hierarchical Shapley Explanations for Breast Cancer Diagnosis

Abstract

Purpose: Accurate automated breast cancer diagnosis from mammography remains a challenging task due to the small size and subtle nature of breast lesions in relation to the large dimensions of mammographic images. Additionally, the lack of explainability in deep learning-based classification models limits their clinical applicability. This study aims to develop a CNN-based automated breast cancer diagnosis model while integrating a novel Hierarchical Shapley (h-Shap) method to improve model explainability by identifying and visualizing key image regions influencing classification outcomes.Methods: The publicly available CBIS-DDSM mammography dataset, comprising 1,131 patients with 1,355 abnormalities, was utilized. Each image was resized to 1500×1000 pixels and uniformly segmented into smaller patches for localized analysis. A CNN-based EfficientNet-B0 deep learning model was trained to classify whether each patch contained an abnormality. The dataset was split into training, validation, testing sets in an 7:1:2 ratio. The h-Shap method was employed to compute the contribution of individual image regions to classification decisions, allowing for a hierarchical decomposition of feature importance. Heatmaps were generated to align model predictions with clinically relevant features. Results: The proposed model achieved an overall classification accuracy of 83.43% on the test set. The confusion matrix indicated a true positive rate (TPR) of 29.62% and a true negative rate (TNR) of 90.16%, highlighting challenges in detecting subtle abnormalities. Notably, 35.20% of correctly classified positive samples exhibited tumor regions accurately identified and highlighted through h-Shap visualizations, demonstrating the method’s effectiveness in improving model explainability. Conclusion: This study integrates the EfficientNet-B0 CNN model with the h-Shap method to enhance the explainability of automated breast cancer diagnosis in mammography. The results demonstrate the model’s potential in detecting abnormalities while providing visual explanations. Future work will focus on improving sensitivity and refining explainability techniques for more reliable clinical application.