Breast Tumor Segmentation using Deep Learning by U-Net Network

Abstract

The methods of image analysis are important for the segmentation and detection of breast tumors, where a reliable diagnosis will be supported by portraying crucial morphological.  This paper presents the segmentation of a breast tumor using a deep convolutional neural network by U-net. The structure includes a contractile route to capture the background and an a-symmetric expansion path that provides precise localization. It trains end-to-end from a small number of images. This segmentation algorithm is applied to breast tumours within a Region of Interest (ROI) with a High-Power Field (HPF) in biopsy and its corresponding ground-truth, where it has been labeled by pathologists with benign and malignant. The suggested segmentation model produces binary masks as realistic as possible.  It supplies an Intersection Over Union (IOU) of 68% and attains an overall accuracy as high as 91%, which shows it performs better than the current works.