Medical Image Segmentation Using Triple Cross-Attention Graph Agnostic Multi Scale U-Net

Breast cancer (BC) is one of the most prevalent cancers in women. Breast Cancer Pathological Image Segmentation (BCPIS) is an important process for assisting medical experts in calculating tumor area and providing treatment suggestions. However, achieving accurate semantic segmentation is still a challenging task for this technology. This work introduces Triple Cross-Attention Graph-Agnostic Multi-Scale U-Net (TC-GAMU-Net), a new deep-learning architecture that combines graph-based learning, crossattention mechanisms, and multi-scale feature extraction to improve segmentation performance. The Triple CrossAttention module captures local and global dependencies, while the Graph-Agnostic framework provides robustness across imaging modalities. Moreover, the Multi-Scale U-Net architecture enables segmentation of fine-grained and largescale structures, allowing accurate boundary delineation. Experimental assessments on benchmark BC histopathology datasets reveal TC-GAMU-Net better than other models with accuracy of 0.987. This research demonstrates the significance of cross-attention and graph-based learning in accurate segmentation to support enhanced Computer-Aided Diagnosis (CAD) systems in BC detection.