Semantic Segmentation of Gastrointestinal Tract using UNet Model with ResNet 18 Backbone

Abstract

Every year, the number of people diagnosed with gastrointestinal (GI) cancer rise. There will be about 5 million new patients with GI cancer this year. Radiation treatment is the gold standard in the biomedical sector for treating gastrointestinal cancer. An X-ray beam is directed towards the tumor while protecting surrounding healthy tissue in radiation treatment. Manually segmenting healthy organs can be time-consuming and error-prone for an oncologist. Therefore, a technology that can autonomously separate the GI tract's healthy organs is required. With the help of the automated system, radio oncologists can perform treatments more quickly. This study proposes using a deep learning-based UNet model with ResNet as an encoder to distinguish between the many healthy GI tract organs. UW Madison's gastrointestinal (GI) database was used for the implementation. There are 38496 MRI scans stored in the database using run-length encoding (RLE). We have calculated the model loss, IoU coefficient, and dice coefficient to assess the quality of the suggested model.