Accuracy is Not Enough: Multi-Head Gated Attention Multiple Instance Learning Model for Explainable Breast Cancer Diagnosis

Attention-based models are inherently interpretable owing to their ability to highlight relevant features, yet their explanations are rarely rigorously validated. We introduce a multi-head gated attention multiple instance learning (MH-GAMIL) mechanism for image classification, designed to provide disentangled attention via separated heads, and compare it with established models like gated attention (GAMIL), clustering constrained attention (CLAM), and dual-stream attention (DSMIL) to assess both predictive performance and explanation effectiveness. The study used two mammography datasets: TOMPEI-CMMD, comprising 2436 images, and a non-public dataset from the Medical University of Gdańsk (MUG) with 2422 images. The datasets differ in resolution, pixel depth, spacing, and acquisition devices. Breast cancer classification performance was evaluated using accuracy, macro precision, macro recall, and macro F1-score, while explanation quality was assessed with Relevance Rank Accuracy (RRA) and an adapted Effective Sparseness. GAMIL and MH-GAMIL consistently outperformed CLAM and DSMIL, achieving 76–78% intra-site and 57–67% cross-site accuracy, though accuracy was not our main focus. MH-GAMIL achieved the highest localization performance, with the positive head reaching an average RRA of up to 0.452 and moderate sparsity (0.361–0.565) across all training–testing permutations. DSMIL showed moderate localization (RRA 0.115–0.269, sparsity 0.256–0.440), while CLAM exhibited high sparsity but almost no correct localization (RRA as low as 0.005). Single-head GAMIL performed poorly in both metrics. Notably, MH-GAMIL and DSMIL maintained reasonable RRA even under cross-site evaluation, demonstrating robustness to domain shift. Our study highlights empirical evaluation is crucial to ensure attention-based models provide meaningful, trustworthy insights rather than giving the illusion of decision interpretability.