Interpretable Lesion-Guided Correlation Learning for Mammography Computer-Aided Diagnosis with Swin Transformer

Abstract

Deep learning-based computer-aided diagnosis (CAD) systems for mammography have achieved remarkable performance in recent years, evolving from convolutional neural networks to transformer-based architectures. However, it remains an open challenge how to effectively incorporate lesion-level clinical information into model training to regulate feature learning, such that both diagnostic accuracy and interpretability can be improved in a principled manner. In this work, we propose a lesion-guided correlation learning framework that incorporates lesion location information as a structural constraint during training. Instead of relying on additional segmentation networks or multi-pathway architectures, our method directly regularizes the classification representation by modeling the correlation between model attention and lesion-related cues derived from training annotations. The proposed framework is instantiated on a Swin Transformer for mammography classification. By explicitly constraining how the classifier utilizes lesion-related information, the proposed approach encourages the model to focus on clinically relevant regions without introducing extra supervision pathways. On the TOMPEI-CMMD MLO-view dataset, the proposed method achieved an AUC of $0.8766 \pm$ 0.0020, outperforming the standard Swin Transformer (0.8666 ± 0.0041), the segmentation-assisted multitask model (0.8649 ± 0.0106), and the pseudo map-guided variant $(0.8680 \pm 0.0110)$. For interpretability, the proposed correlation map achieved a Soft IoU of $0.0986 \pm 0.0741$ with lesion masks, corresponding to a 72.4% relative improvement over Grad-CAM on the standard Swin Transformer. These results indicate that incorporating lesion-level clinical information as a correlation-based training constraint provides an effective inductive bias for improving both explainability and discriminative representation learning in mammography CAD.