A Self-Attention Enhanced Method for Non-Invasive Breast Cancer Grading on Magnetic Resonance Imaging

Breast cancer represents the most prevalent malignancy and the leading cause of cancer-related mortality among women globally. Current grading systems rely on conventional pathological criteria necessitating invasive procedures for tissue or cellular acquisition, which are labor-intensive, subjective, and invasive. This paper introduces a non-invasive end-to-end deep learning framework for breast cancer grading. It marks the inaugural attempt to deploy deep learning methodologies on breast cancer magnetic resonance imaging (MRI), providing evidence for imaging biomarkers of breast cancer. Addressing the fine-grained and multi-scale lesions inherent in datasets, we propose a feature fusion approach based on self-attention mechanisms. Experimental results validate the feasibility of exploring breast cancer biomarkers on MRI through deep learning methodologies, with the novel feature fusion strategy enhancing the extraction of salient pixel information and augmenting classification accuracy.