A Vision Transformer Approach for Breast Cancer Classification in Histopathology

Abstract

Breast cancer is the second-leading cause of cancer-related deaths in women worldwide. Early detection through regular screenings and self-examinations is vital to increasing survival rates especially through histopathology image analysis, which helps determine the extent of tumor invasion, the presence of metastasis, and the aggressiveness of the cancer cells. This research paper introduces a novel methodology for classifying Breast Cancer (BC) from histopathology slides using a Vision Transformer (ViT) model. The approach involves gathering a dataset of annotated high-quality histopathology slides, followed by a meticulous pre-processing phase. The ViT model is trained on multiple resolutions of Regions of Interest (ROIs), and a majority voting mechanism is employed for decision-making. Additionally, post-processing techniques, including region growing and fast-marching level set, are applied to enhance the prediction. The proposed framework achieves outstanding results, with the best performance obtained at a ROI scale of 1,024 pixels. At this scale, the model achieves an impressive accuracy of 99.42%, precision and recall of 98.86% and 98.84%, respectively, and a balanced accuracy of 99.23%. The research contributes to the advancement of computer-aided diagnosis in histopathology, particularly in BC classification.