Multi-view Unet for Automated GI Tract Segmentation

Abstract

Gastrointestinal cancer has shown a high incidence trend in recent years, affecting the health of millions of people. During the treatment of gastrointestinal cancer, radiation oncologists have to deliver x-ray beams pointed toward the tumor and meanwhile avoid the stomach and intestines. In this paper, we propose an automated process for gastro-intestinal tract (GI Tract) image segmentation. The 2D segmentation methods are time-saving while not good at modeling the multidimensional information of medical images. The 3D methods take advantage of both temporal and spatial information. However, they are relatively time-consuming and dependent on computing resources. To combine the advantages of the above methods, we propose an improved 2.5D method for GI Tract image segmentation. In order to effectively utilize the association of adjacent images, we explore as well as fuse different 2.5D data generation approaches and propose a 2.5D feature fusion method with adjacent weighting. To make up for the impact of low spatial-temporal resolution of traditional methods on the results, we propose a method for fusing 2.5D and 3D results. Our method fuses plenty of representation processes by embedding multi-dimensional convolutions deeply into basic modules, making the selections of representation processes flexible, thus balancing representation for sparse inter-slice information and dense intra-slice information adaptively. It also takes the advantages of the high segmentation accuracy of the easily recognized regions in 2.5D view as well as the high smoothness of 3D organ contour representations, thus obtaining better explanatory and more accurate modeling for target regions. Comprehensive experiments are performed on a public GI Tract dataset, experiments show that the 2.5D fusion method improves 0.36% on dice and 0.12% on Jaccard compared with the 2.5D method without feature fusion. The results fusion method improves 0.007 scores compared with 2.5D and 0.009 scores compared with 3D on the kaggle open test set.