Segmentation of Breast Cancer Ultrasound Images

Abstract

Breast cancer is the most common cancer occurring in women. Breast-conserving surgery is a desirable choice for an early-stage breast cancer. An intra-operative margin assessment of excised breast lesion tissue can help avoid additional surgeries. An essential problem in intra-operative margin assessment is how to extract an accurate boundary of the excised lesion automatically and quickly. To solve this problem, we segment breast cancer ultrasound (US) images and then generate boundaries based on the segmentation results.

In this research, we propose a new convolutional neural network model, named IU-Net, to segment breast cancer US images. IU-Net combines inception blocks and the well-known U-Net model. We train IU-Net with US images and corresponding manually segmented images provided by Dr. Jeffery Carson and his research group of the Lawson Health Research Institute, London, Ontario, Canada. We also apply an autoencoder in training IU-Net. The experimental results show that IU-Net achieves slightly more accurate results than U-Net and uses 3.8x fewer parameters than U-Net.