Abstract:In the clinical diagnosis of breast cancer, how to segment breast tumors accurately and automatically is always a challenging task. With the advances in deep learning, automatic segmentation methods can reduce the workload for radiologists. In this paper, we design an automatic 2D breast tumor segmentation method based on a fully convolutional network named M-Net combined with a phase-based active contour (PBAC) model. The M-Net is comprised of multi-scale input and output, along with V-shape convolutional network. An improved structure of the residual functions named the shortcut connection is utilized to accelerate the training process. The results of the M-Net are fairly rough as a result of blurry boundaries and size variations of breast tumors. Therefore, we apply the PBAC model to optimize the segmentation details using the M-Net output as initialization. The M-Net is trained end-to-end from scratch using a dataset which contains 110 training images and 28 test images. The experimental results of our M-Net + PBAC method achieve a Dice similarity coefficient score of 96.89%, and a Hausdorff distance (HD) of 6.96 pixels, which are the best segmentation performance compared with U-Net and V-Net.

Combining a Fully Convolutional Network and an Active Contour Model for Automatic 2D Breast Tumor Segmentation from Ultrasound Images.