X-Pruning: a dual-stream information fusion mammography diagnosis network based on pruned transformer and cross-attention mechanism.

Background: Breast cancer remains the most prevalent malignancy among women globally. While early detection through mammography is crucial for improving survival rates, automating this process poses significant computational challenges. Specifically, detecting small malignant lesions within high-resolution images and effectively integrating complementary information from standard craniocaudal (CC) and mediolateral oblique (MLO) views are difficult tasks. Therefore, the aim of this study is to develop a novel, highly efficient dual-view mammography diagnostic network, termed X-Pruning, designed to overcome these computational and integration challenges. Methods: The proposed X-Pruning framework addresses multi-view mammography by combining pruned transformer blocks (PTBs) with cross-attention fusion. The model employs parallel PTBs to process standard CC and MLO views. By implementing a dynamic pruning strategy based on window importance within the transformer layers, the network focuses computational resources specifically on suspicious regions. This reduces the heavy burden of self-attention calculations while strictly preserving full-image scale information to enhance the detection of small lesions. Additionally, a novel cross-attention fusion module was developed and integrated into the network to facilitate interactive information exchange between the CC and MLO views, enabling comprehensive multi-view feature integration. Results: The X-Pruning framework was evaluated on two widely recognized public datasets: Vindr-Mammo and Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM). Experimental results demonstrated that the proposed model consistently outperformed several mainstream baseline architectures. Specifically, X-Pruning achieved area under the curve (AUC) scores of 0.812 on the Vindr-Mammo dataset and 0.792 on the CBIS-DDSM dataset. Furthermore, these performance improvements were achieved while significantly reducing overall computational demands compared to standard models. Conclusions: The X-Pruning network successfully resolves the trade-off between high diagnostic accuracy and computational efficiency in automated mammogram analysis. By intelligently allocating computational resources to critical lesion areas and effectively fusing cross-view information, this framework demonstrates robust diagnostic capabilities. These advancements highlight the clinical potential of X-Pruning to serve as an efficient, automated tool for enhancing the early detection and diagnosis of breast cancer.