SER-UNet: A Network for Gastrointestinal Image Segmentation

Abstract

Cancers of the digestive tract include esophageal tumors, gastric tumors, and intestinal tumors. Radiation oncologists try to deliver high doses of radiation using X-rays directed at the tumor while avoiding the stomach and intestine, but the complex manual labeling of the gut is time-consuming and inaccurate. Using deep learning can help automate the segmentation process, and this method of segmenting the stomach and intestine will lead to faster treatment. It will allow more patients to be treated more effectively. Thus, we propose a network model for GI segmentation that uses a residual network with a fused channel attention mechanism as an encoder for the U-Net model, combined with a U-Net decoder and a feature fusion architecture to achieve pixel-level classification and segmentation of images. In our experiments, we choose IOU as the model evaluation index, and the higher the IOU, the better the performance of the model. The experimental results show that the IOU of our model is improved by 1.8% to 2.5% compared with other models, which outperforms other models in the GI segmentation task.