A Comparison of Gastrointestinal Image Segmentation by U-Net Based on Ensemble of U-Net Models and Single Model

More than 1 million people worldwide are diagnosed with gastrointestinal cancer each year. About half of these patients are eligible for radiation therapy. Radiation oncologists strive to administer potent doses of radiation by directing beams of X-rays towards the tumor while ensuring the avoidance of the gastrointestinal tract and stomach [1]. The present objective is to direct X-ray beams towards malignant tumors while circumventing the gastrointestinal tract and stomach. This research presents a method for dividing the gastrointestinal organs (small the intestinal tract, the large intestinal tract, and stomach) into distinct sections. This technique aims to assist radiation oncologists in treating cancer patients with more speed and precision [1]. In this paper, the multi-path U-NET 2.5D model is assembled with U-NET 3D model for segmentation. In addition, among the proposed models, EfficientNet-B0 and EfficientNet-B5 are selected as the backbone models of U-Net to achieve high parameter and computational efficiency and more profound feature extraction capability. Finally, the Dice coefficient and Hausdorff distance are weighted into the evaluation index, where the Dice coefficient weighted 0.6, and the Hausdorff distance weighted 0.4. The final results show that the U-Net2.5D model (with EfficientNet-B5 as the backbone) has the best indicator when integrated with the U-Net3D model, reaching 0.8762. They are improved by 0.0244 compared to U-Net2.5D(EfficientNet-B5) alone.