Artificial Intelligence-Based Localization of Small Bowel Anatomic Transition Zones in Crohn's Disease Using Capsule...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Small bowel capsule endoscopy (SBCE) is crucial for diagnosing Crohn's disease (CD).
• Interpreting SBCE videos is time-consuming, necessitating AI assistance.
• Automated segmentation of the small bowel in CD patients presents challenges.

Purpose of the Study:

• To assess an AI model's ability to detect the pylorus and ileocolonic junction in SBCE videos of CD patients.
• To evaluate the accuracy and limitations of AI-driven landmark detection in this context.

Main Methods:

• A ResNet-18 convolutional neural network was employed for frame classification (stomach, SB, colon).
• Viterbi-based, anatomically constrained segmentation was utilized.
• The model was trained on nearly 600,000 images and validated using five-fold cross-validation on 60 CD patient videos.

Main Results:

• The AI model achieved a median localization error of 1 second for the pylorus and approximately 5 minutes for the ileocolonic junction.
• Both junctions were localized within a 1-minute discrepancy in 43.3% of videos compared to expert annotations.
• AI detection of the ileocolonic junction preceded expert annotation in over half of cases, potentially impacting lesion detection.

Conclusions:

• The AI model successfully localizes both small bowel junctions in nearly half of CD SBCE videos.
• Pylorus localization was consistent, but ileocolonic junction performance was modest.
• Inadequate bowel cleanliness and capsule movement dynamics pose significant challenges for automated segmentation.