Esophagus tumor segmentation using fully convolutional neural network and graph cut

Abstract

The development of Esophagus radiation treatment plan demands accurate Esophagus tumor segmentation. However, such task was often prevented by random distribution and weak boundaries of Esophagus tumors on CT images. To address these challenges, we develop a novel framework based on the combination of Fully Convolutional Neural Network (FCN) and graph cut algorithms. FCN is utilized to establish an Esophagus tumor classifier on the training dataset with expert-labeled tumor regions. When segmenting Esophagus tumors on the test dataset, the tumor probability maps are first estimated. Graph cut is next used to extract the actual tumor regions by enforcing the spatial constraints. 87 CT sequences were selected as the validation dataset, and 3-fold cross-validation was performed to evaluate the segmentation accuracy. Tumor volume overlap between ground-truth and segmentation results was only 71% by exploiting FCN alone, while it was improved to 80% by combining graph cut algorithm. These promising results suggest that the combination of FCN and graph cut can accurately segment Esophagus tumors, which has a great potential to reduce human burden in contouring tumor regions as well as improve the accuracy of radiation treatment planning.