Adapting Segment Anything Model for Ulcer Segmentation in Inflammatory Bowel Disease.

Inflammatory Bowel Disease (IBD) is a widespread gastrointestinal disorder characterized by inflammation and ulceration, affecting millions of individuals. Mucosal injury assessment through colonoscopy plays a crucial role in determining disease severity and therapeutic efficacy in IBD. This study draws inspiration from the Segment Anything Model (SAM) framework and tailors it for ulcer segmentation using a Parameter-Efficient Transfer Learning (PETL) approach and incorporating binary cross-entropy loss with boundary loss. Comparative assessments against state-of-the-art UNet-based convolutional neural network models underscore SAM's superior performance, especially when faced with limited training samples. Our ulcer segmentation models demonstrate Dice, IoU, and 95%HD scores of 61.92±0.69, 75.18±0.38, and 10.07±0.05 in the colon dataset, and 55.60±0.82, 77.00±0.67, and 7.74±0.23 in the ileum dataset, respectively. This study highlights the potential of our approach to significantly enhance the efficiency and accuracy of IBD diagnosis and assessment through cutting-edge computer vision techniques.