Edge-Guided Cross-Modal Fusion Network for Multi-Resolution Breast Cancer Segmentation in Smart Digital Pathology

Accurate segmentation of carcinoma in situ and invasive carcinoma in Whole Slide Images (WSIs) is crucial for improving breast cancer diagnostics in smart healthcare systems. Existing methods that rely solely on Hematoxylin and Eosin (H&E) staining lack molecular boundary-specific markers and struggle with resolution limitations. To address these challenges, we propose a breast cancer segmentation framework that fuses multi-resolution semantic features from H&E images with edge information from Cytokeratin 5/6 (CK5/6) immunohistochemical staining. The model integrates three modules: a multi-resolution semantic segmentation branch, an edge detection module aligned with H&E images, and a multi-scale fusion module. By combining multi-modal information and selectively zooming in on key regions, the method enhances the diagnostic process of medical practitioners, making the system more accurate and suitable for deployment in an Internet of Medical Things (IoMT) platform. Evaluations on the Breast Cancer Semantic Segmentation (BCSS) and the Chinese People's Liberation Army (PLA) General Hospital datasets show segmentation similarity coefficients of 81.28% and 93.16%, respectively. This approach offers an effective solution for user-facing digital pathology systems and supports clinical decision-making in consumer-centric smart healthcare.