Toward high resolution scoring system for HER2 expression in breast cancer based on pathological images via deep learning

New human epidermal growth factor receptor 2 (HER2)-positive breast cancer drugs are clinically active in HER2 low-expressing breast cancer. However, current molecular marker techniques (IHC and FISH) are not able to subdivide the low expressing status of HER2 breast cancer. In response to the specified challenge, this paper develops a high-resolution clinical scoring system based on Transformer models and Fourier Transform. It autonomously identifies invasive tumor regions and accurately classifies HER2 tiles. This system introduces granular categorization guidelines for the IHC 0 to IHC 1+ and IHC 2+ statuses, effectively simulating the clinical diagnostic process. In tests conducted on whole slide images (WSIs) from 128 patients, the system achieved accuracies of 98.78% in invasive tumor region identification and 96.32% in HER2 tile classification, with both metrics exceeding an AUC of 99%. Among the models tested, the multi-scale training approach and the Vision Transformer (ViT) with Fourier transform techniques demonstrated the most impressive performance. The proposed system enhances the reliability of diagnosing low HER2 expression status, providing more precise assessments and guidance for patients’ prognosis and HER2 targeted therapies, potentially significantly improving treatment outcomes.