Deep learning and capsule endoscopy: automatic panendoscopic detection of protruding lesions

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Capsule endoscopy (CE) enables minimally invasive panendoscopic evaluation of the gastrointestinal tract.
• Conventional CE reading is time-consuming and prone to errors.
• Detecting protruding lesions (polyps, tumors) is crucial for diagnosis.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN)-based algorithm for automatic detection of protruding lesions in CE exams.
• To enhance the diagnostic accuracy and efficiency of capsule endoscopy through artificial intelligence.

Main Methods:

• A multicentric retrospective study involving 1245 CE exams and over 191 million frames.
• Data included 52,717 frames with protruding lesions, validated by triple assessment.
• A CNN algorithm was trained and tested using an exam-split design with fivefold cross-validation.

Main Results:

• The CNN achieved a sensitivity of 79.7% and specificity of 96.5% on the test set.
• Positive Predictive Value (PPV) was 81.5%, and Negative Predictive Value (NPV) was 96.0%.
• Overall accuracy of the algorithm reached 93.7%.

Conclusions:

• This study reports the first CNN for detecting protruding lesions across the GI tract in AI-enhanced capsule panendoscopy.
• AI integration can improve CE diagnostic accuracy and potentially increase access to this minimally invasive procedure.
• Further prospective studies are required to validate these findings for clinical implementation.