Twoviewdensity-Pvt: A Vision Transformer System for Accurate Bi-Rads Classification in Breast Cancer Grading From Mammogram Images

In this paper, a novel vision transformer-based system, TwoViewDensity-PVT, is designed for classifying Breast Cancer (BC) into four BI-RADS categories from mammogram images. It aims to enhance BC grading accuracy and consistency by offering an objective categorization method. The pipeline workflow is inaugurated by identifying Regions of Interest (ROIs) in craniocaudal (CC) and mediolateral oblique (MLO) views, which are then processed by transformers to discern BI-RADS-associated radiomic features. To assist the transformers in differentiating both global and local radiomics, such as mass density, we integrated Improved Receptive Field Block (IRFB) (i.e., multiscale kernel) to streamline this process. TwoViewDensity-PVT derives its name from its dual-view approach and effective identification of mass density relevant to BI-RADS grading. Rigorous testing on 4368 patients across four BI-RADS classes yielded a system sensitivity of 83.89% and specificity of 92.13%.