Small bowel capsule endoscopy examination and open access database with artificial intelligence: The SEE-artificial...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Small bowel capsule endoscopy (CE) generates extensive image data, posing challenges for manual interpretation.
• Artificial intelligence (AI) offers potential for automating image classification in CE.
• Developing effective AI models for CE requires robust datasets and validated algorithms.

Purpose of the Study:

• To create a comprehensive dataset of small bowel CE images.
• To develop and validate an object detection AI model for identifying disease lesions in CE.
• To explore the challenges and potential of AI in assisting small bowel CE interpretation.

Main Methods:

• Extracted 18,481 images from 523 small bowel CE procedures.
• Annotated 12,320 images with 23,033 disease lesions, creating a dataset with normal images.
• Developed an object detection AI model using YOLO v5 and validated its performance.

Main Results:

• The AI model achieved approximately 91% sensitivity across 12 annotation types.
• Achieved a high area under the receiver operating characteristic curve of 0.98 for certain annotations.
• Demonstrated variable detection quality depending on the specific annotation type.

Conclusions:

• An object detection AI model using YOLO v5 shows potential for effective reading assistance in small bowel CE.
• The developed dataset, AI model weights, and demonstration are publicly available for further research.
• Continued improvement of AI models is anticipated for enhanced small bowel CE analysis.