Abstract—Medical image processing has significantly transformed the landscape of healthcare, particularly in the diagnosis and treatment of various diseases. Gastrointestinal (GI) cancer has emerged as a rapidly growing concern, with an estimated 5 million new cases reported annually. To achieve this precision, healthcare professionals now leverage the cutting-edge Magnetic Resonance Imaging (MRI) modality known as MR-Linacs, which provides a daily view of tumor positions. However, a bottleneck in this process arises during the manual segmentation of healthy organs at risk, such as the stomach and intestines, from the obtained medical images. This task, performed by radiologists, is time-consuming and can significantly prolong treatment times, consequently amplifying the suffering of patients. Thus, automation of the GI tract segmentation helps oncologists seamlessly. Our research proposes a model that automates the segmentation of the GI tract. This study presents a U-Net model that segments the stomach and intestines from MRI scans. The dataset, sourced from the UW-Madison Carbone Cancer Center, comprises RLE-encoded masks for training annotations alongside 16-bit grayscale PNG images. Each case consists of multiple scan slices, split either by time or the entire case. Our approach used various combinations of loss functions on U-Net to improve the accuracy and efficiency of the automated segmentation of the gastrointestinal tract. Our model achieves high accuracy with the Dice+BCE loss function when compared with other loss functions. On the training dataset, the U-Net model with Dice+BCE loss function received the highest dice score of 0.9082 and IOU score of 0.8594. On the validation dataset, the model obtained a dice score of 0.8974 and an Intersection over Union (IoU) score of 0.8181. This research contributes to addressing the challenges associated with manual GI tract segmentation, offering a viable solution through automated segmentation using deep learning techniques.
 
Keywords—Magnetic Resonance Imaging (MRI), combined loss functions, U-Net, UW-Madison Carbone cancer center, gastro-intestinal tract segmentation


Copyright © 2024 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC BY-NC-ND 4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.