Breast Cancer Prediction Using Auto-Encoders

Abstract

Breast cancer is one of the most common forms of cancer that is diagnosed in most women and some rare cases even men. In recent years, breast cancer survival rates have increased significantly, due to factors such as earlier detection. Treatment for breast cancer largely depends on identifying the type of mass of tissue formed, which is known as a tumor. If normal cells grow in an uncontrollable manner the tumor is called benign (non-cancerous). But, if the cells’ growth is out of control and their behavior is abnormal, then the tumor is called malignant (cancerous). During the invasive, (i.e. curable) stage of cancer, only 10–15% part of the breast contains cancerous cells. Therefore, it is difficult to diagnose it using mammography. However, the development of machine learning techniques has allowed early detection of breast cancer in clinical trials. This paper presents a comparative study of different machine learning techniques for detecting breast cancer. It also presents an auto-encoder model that performs breast cancer detection in an unsupervised manner. It attempts to identify a compact representation of features that are strongly related to breast cancer. The techniques are tested on the Breast Cancer Wisconsin (Diagnostic) Dataset, which is publicly available on Kaggle. The auto-encoder beats its competitors with a precision and recall of 98.4%.