Advancing Gastrointestinal Tract Cancer Treatment: A Comparative Study of State-of-the-Art Deep Learning Models for MRI Segmentation

This paper investigates the performance of state-of-the-art deep learning models for segmenting the gastrointestinal (GI) tract in Magnetic Resonance Imaging (MRI) images, a critical step in radiotherapy planning for GI tract cancers. Manual segmentation of the GI tract is time-consuming, error-prone, and complicated by anatomical complexity and patient movement. To automate this process, we explore various deep learning techniques for their potential to provide accurate and efficient segmentation. Specifically, this study compares Feature Pyramid Networks (FPN), LinkNet, Pyramid Scene Parsing Networks (PSPNet), and traditional U-Net architectures and their variants. Using the publicly available UW Madison GI Tract Image Segmentation dataset, which includes 38,496 MRI scans with corresponding RLE-encoded masks, we evaluate architectures such as U-Net with different encoders (EfficientNetB1, MobileNetV2, and ResNet50), U-Net++ with EfficientNetB1, PSP-Net (EfficientNetB1 and ResNet34), FPN (EfficientNetB0 and EfficientNetB1) and LinkNet with EfficientNetB1. The models were trained using a standardized protocol over 30 epochs, with performance assessed through the Dice and Jaccard/IoU coefficients. Results indicate that the models achieved Dice scores ranging from 0.86 to 0.89, with no single model consistently out-performing the others. However, U-Net variations with residual connections demonstrated the most robust performance overall. This study provides valuable insights into the application of deep learning for medical image segmentation. It lays the groundwork for further research to enhance segmentation accuracy in GI cancer treatment, ultimately contributing to improved patient care and expedited treatment planning.