TransUNet for Precise and Robust GI tract Segmentation in MRI Images

Abstract

Medical image processing is revolutionizing Gastrointestinal (GI) cancer radiation therapy by enabling precise targeting of high X-ray beam dosages to the tumors while protecting the surrounding healthy organs. However, the manual segmentation of healthy organs at risk from MR-Linac images is a time-consuming process that can delay treatment and increase patient suffering. Deep learning-based medical image processing algorithms have the potential to automate the segmentation of organs at risk, thereby improving the accuracy and efficiency of GI cancer radiation therapy. Every day, the MR-Linac, a cutting-edge MRI technology, tracks the ever-changing positions of tumors. This advanced tool empowers medical professionals to fine-tune cancer treatments with remarkable precision. This study presents a TransUNet model, a combination of transformer architecture with Convolutional Neural Networks (CNNs). TransUNet achieves remarkable results in segmenting and labeling different regions within images by integrating the spatial comprehension of CNNs with the self-attention mechanisms of transformers. Our research compares the TransUNet model with various combinations of loss functions. The model outperforms with Dice+BCE loss function.