Pyramid Feature Iterative Fusion: A Cross-Scale Fusion Algorithm for Enhanced Analysis of H&E Images in HER2-Positive Breast Cancer

For HER2-positive breast cancer patients, the presence of distant metastasis is a crucial prognostic element. Early detection can notably assist in devising treatment strategies, thereby improving patient prognosis. Pathological Hematoxylin and Eosin (H&E) whole-slide images (WSIs) provide a rich array of prognosis-related histological features. By conducting an exhaustive multi-scale analysis, a hierarchical model embodying core tumor trait can be established. Currently, prevalent techniques in multi-scale H&E image analysis entail concatenating feature vectors and applying attention mechanisms to calculate feature correlations. Nonetheless, due to insufficient cross-scale correlation constraints, these methods might generate redundant information. This paper introduces Pyramid Feature Iterative Fusion (PFIF), a groundbreaking cross-scale fusion algorithm for H&E images. PFIF resolves the issues of redundancy and reduces noise following fusion. By integrating a backward feature fusion module, it emulates human evaluation of H&E images across scales. It amalgamates structural semantic features from a low-magnification view with texture features from a highmagnification view, thereby conveying key complementary details across resolutions. Training this model with H&E image data from Chongqing University's Cancer Hospital resulted in an AUC rate improvement of approximately 3.69.