Breast cancer classification based on microcalcifications using dual branch vision transformer fusion

基于微钙化的乳腺癌分类:采用双分支视觉变换融合技术

阅读:1

Abstract

Breast microcalcifications (MCs) are among the earliest and most challenging indicators of breast cancer, owing to their subtle morphological characteristics and the structural variability between CC and MLO mammographic projections. Current deep learning architectures frequently lose micro-level lesion details while attempting to model cross-view continuity, limiting their reliability in clinical diagnosis. This paper introduces a novel anatomy-aware Multiview transformer architecture designed to preserve microcalcification-scale texture patterns and enforce cross-view anatomical consistency. This approach enables a degree of view-coherent feature learning not previously achieved in mammography analysis. A key innovation is the development of pathology-oriented, view-congruent transformer representations that maintain the fine morphological features of MCs and preserve structural integrity across projections, directly addressing long-standing challenges in dual-view interpretation. The proposed framework was evaluated on the CBIS-DDSM dataset and demonstrated strong diagnostic performance, achieving an accuracy of 96.80, a precision of 97.10, a recall of 96.40, an F1-score of 96.70, and an AUC of 0.982. It also performed effectively in clinically challenging subgroups, obtaining the highest class-wise F1-scores for No Damage (0.9792), Matrix Cracking (0.9565), and Pleomorphic MCs (0.9608). Moreover, visual explanation techniques consistently highlighted clinically relevant calcification clusters, enhancing both interpretability and clinical reliability. These results establish a new representational paradigm for dual-view mammography. The proposed framework advances automated microcalcification assessment by integrating fine-grained lesion preservation with anatomically aligned Multiview reasoning.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。