Encoder-Decoder Variant Analysis for Semantic Segmentation of Gastrointestinal Tract Using UW-Madison Dataset.

Abstract

The gastrointestinal (GI) tract, an integral part of the digestive system, absorbs nutrients from ingested food, starting from the mouth to the anus. GI tract cancer significantly impacts global health, necessitating precise treatment methods. Radiation oncologists use X-ray beams to target tumors while avoiding the stomach and intestines, making the accurate segmentation of these organs crucial. This research explores various combinations of encoders and decoders to segment the small bowel, large bowel, and stomach in MRI images, using the UW-Madison GI tract dataset consisting of 38,496 scans. Encoders tested include ResNet50, EfficientNetB1, MobileNetV2, ResNext50, and Timm_Gernet_S, paired with decoders UNet, FPN, PSPNet, PAN, and DeepLab V3+. The study identifies ResNet50 with DeepLab V3+ as the most effective combination, assessed using the Dice coefficient, Jaccard index, and model loss. The proposed model, a combination of DeepLab V3+ and ResNet 50, obtained a Dice value of 0.9082, an IoU value of 0.8796, and a model loss of 0.117. The findings demonstrate the method's potential to improve radiation therapy for GI cancer, aiding radiation oncologists in accurately targeting tumors while avoiding healthy organs. The results of this study will assist healthcare professionals involved in biomedical image analysis.