Adapting Segment Anything Model for Ulcer Segmentation in Inflammatory Bowel Disease

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Inflammatory Bowel Disease (IBD) affects millions, requiring accurate assessment of mucosal injury via colonoscopy.
• Current methods for assessing IBD-related ulceration can be subjective and time-consuming.

Purpose of the Study:

• To adapt the Segment Anything Model (SAM) for accurate ulcer segmentation in IBD.
• To evaluate the performance of a tailored SAM approach against existing convolutional neural network models.

Main Methods:

• Utilized a Parameter-Efficient Transfer Learning (PETL) approach with SAM for ulcer segmentation.
• Incorporated binary cross-entropy loss with boundary loss for enhanced segmentation accuracy.
• Compared the developed model against UNet-based convolutional neural network models.

Main Results:

• The SAM-based model demonstrated superior performance, particularly with limited training data.
• Achieved Dice scores of 61.92% (colon) and 55.60% (ileum), IoU scores of 75.18% (colon) and 77.00% (ileum).
• Obtained 95%HD scores of 10.07 (colon) and 7.74 (ileum), indicating precise boundary detection.

Conclusions:

• The proposed SAM-based approach significantly enhances the efficiency and accuracy of IBD diagnosis and assessment.
• Computer vision techniques show great potential for improving clinical evaluation of gastrointestinal disorders.
• This method offers a promising tool for objective and reliable IBD mucosal injury evaluation.