Self-supervised Visual Transformers for Breast Cancer Diagnosis

Owing to the growing incidences, breast cancer is considered the world's most prevalent cancer both in terms of morbidity and mortality rates. While breast cancer treatment is quite effective, higher mortality is caused due to the delayed diagnosis. Mammogram image processing using artificial intelligence techniques can be promising in the early diagnosis and Convolutional Neural Networks (CNN) have shown some potential in image classification and semantic segmentation. However, working with unlabeled image data poses challenges and while the manual labeling is not efficient, pre-trained CNN s also perform poorly on the actual medical images. In this paper, we propose to use transformer-based networks in the field of computer vision. Previously published research also suggests a predominance of transformer-based vision models over the convolutional models. Bootstrap Your Own Latent (BYOL), which is Self-supervised learning (SSL) algorithm, is proposed in this paper to process mammograms images for diagnostic purposes. BYOL can handle unlabeled image data and improve the deep learning mechanism for data analytics. UNet networks with Xavier initializations have been considered for the baseline to compare with their performances when implemented with and without BYOL. It is shown that the overall network performance improves when they are trained using the proposed BYOL algorithm.