Variational Auto encoders for Improved Breast Cancer Classification

Abstract

Among women breast cancer is the second leading cause of cancer. Emergence of Artificial Intelligence(AI) in the medical care leads to good survival rate by diagnosing and effective prognosis of the breast cancer patients. Scientific findings show that supervised deep learning model is highly dependent on the size of the training set, which must be manually labeled by experienced radiologists. The freely available biomedical imaging dataset are small in size mostly. In addition, obtaining large medical image files is difficult due to privacy and legal reasons. Thus the deep learning models tend to overfit and fail to provide a general result. Also the recent studies indicate that early detection of breast cancer (BC) is crucial to achieve favorable treatment results and reduce associated mortality. Data augmentation is the most widely used approach to address the aforementioned problem. Our proposed architecture consist of the pectoral muscle removal of the dataset, then the variational auto encoders used for data augmentation and then the U-Net and its varients used for breast cancer classification. The dice similarity score obtained for the model after the pectoral muscle segmentation is 97.53% and the classification rate accuracy is 98.5% by use of the Attention Residual U-Net than the other two models.