Causal Prototype-Guided Multi-Instance Learning for Explainable Breast Cancer Diagnosis in Mammogram Images

Abstract

Breast cancer remains a leading cause of mortality among women worldwide, making early and accurate diagnosis through mammography critically important. However, conventional deep learning approaches for mammogram analysis often operate as black-box models, limiting interpretability and clinical trust. This study proposes a Causal Prototype-Guided Multi-Instance Learning (CP-MIL) framework for explainable breast cancer diagnosis from mammogram images. The method decomposes each mammogram into multiple region-level instances, extracts clinically meaningful handcrafted features, and learns benign and malignant prototypes to guide diagnostic reasoning. Counterfactual causal analysis is incorporated to quantify the contribution of each region to the final prediction, enabling transparent and region-specific explanations aligned with radiological practice. Experimental evaluation on benchmark mammography datasets demonstrates that the proposed framework achieves competitive classification performance with an accuracy of 98.9%, outperforming several existing deep learning and transfer learning models. Visualization results, including segmentation masks, attention heatmaps, and causal influence distributions, confirm the model’s ability to highlight diagnostically relevant regions while maintaining robustness and low overfitting risk. Overall, the CP-MIL framework provides a reliable, interpretable, and clinically meaningful solution for automated breast cancer detection, supporting radiologists in early diagnosis and improving decision-making in medical imaging applications.