DB M-Net: An Efficient Segmentation Network for Esophagus and Esophageal Tumor in Computed Tomography Images

Abstract

Esophageal cancer is one of the diseases afflicting human beings. Automatic segmentation of esophagus and esophageal tumor from computed tomography (CT) images is a challenging problem, which can assist in the diagnosis of esophageal cancer. In this paper, DB M-Net is proposed for the segmentation of esophagus and esophageal tumor from CT images, which combines M-Net modified from U-Net with an approximate function for binarization called differentiable binarization (DB). We construct the multi-scale input layers and the multi-level output layers in the network to facilitate features fusion, and DB is performed to enhance the robustness. Fewer parameters are applied in our DB M-Net but the network achieves a better performance. The experiments are based on the dataset of 2,219 slices from 16 CT scans, which show our DB M-Net outperforms other existing algorithms.