MPE-DiGA-Net: A physics-aware and evidence-driven dual-view mammography framework for trustworthy breast cancer diagnosis

Abstract

The high incidence of breast cancer and the urgent need for effective screening pose a significant challenge to global public health. Constructing a complete diagnostic evidence chain using dual-view mammography has become a critical approach to support precision diagnosis and treatment. However, due to the high density of breast tissue and non-rigid compression during imaging, dual-view collaborative diagnosis faces four fundamental challenges at the information-processing level: the “semantic gap” between physical mechanisms and deep representations; “background masking” caused by dense glandular tissue; “evidence misalignment” resulting from non-rigid soft-tissue deformation; and the “lack of risk perception” inherent in deterministic decision-making. To address these challenges, we propose MPE-DiGA-Net, a physics-aware and evidence-driven trustworthy diagnosis framework. First, a Physics-Aware Multi-scale Feature Cooperative Embedding (MPE) mechanism introduces clinical physical priors into the feature evolution space, effectively alleviating the semantic opacity of purely data-driven representations. Second, a Density-Adaptive Feature Discrimination (DA-FDM) module dynamically reweights feature channels according to glandular density. This dynamically overcomes the “needle-in-a-haystack” dilemma in dense backgrounds. Third, a fusion architecture based on Interactive Graph Attention (DiGA) establishes a semantic-level soft alignment mechanism in the topological space, compensating for non-rigid anatomical displacements caused by differences in mechanical compression. Finally, an Evidence-Driven Uncertainty Quantification (EDL) mechanism is introduced to perceive the epistemic risk of out-of-distribution samples in real time, avoiding overconfident decisions and ensuring diagnostic reliability in complex clinical environments. Experimental results on the Chinese Mammography Database (CMMD) demonstrate that MPE-DiGA-Net achieves an accuracy of 90.10% (AUC = 0.9450). It significantly outperforms current state-of-the-art (SOTA) methods across multiple key metrics. This study validates the consistency between the model’s decision logic and radiological priors, effectively bridging the trust gap between artificial intelligence and clinicians. It further establishes the practical value of the proposed framework as a reliable auxiliary reference integrated into physician-led workflows, providing important technical support for the trustworthy development of precision breast cancer diagnosis systems in resource-limited settings.