Automatic detection of various abnormalities in capsule endoscopy videos by a deep learning-based system: a...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Capsule endoscopy (CE) reading requires efficient screening tools for detecting gastrointestinal abnormalities.
• Current methods like QuickView mode have limitations in identifying diverse pathologies.
• Deep convolutional neural networks (CNNs) show potential for automated analysis of CE images.

Purpose of the Study:

• To develop and evaluate a CNN-based system for detecting various abnormalities in capsule endoscopy.
• To compare the diagnostic performance of the developed CNN system against the existing QuickView mode.

Main Methods:

• A CNN system was trained on a large dataset of 66,028 CE images.
• The system's detection capabilities were assessed using an independent test set of 379 small-bowel CE videos from multiple institutions.
• Performance was evaluated based on detection rates for specific abnormalities (mucosal breaks, angioectasia, protruding lesions, blood content) per patient.

Main Results:

• The CNN system achieved a significantly higher overall detection rate for abnormalities per patient (99%) compared to QuickView mode (89%).
• Specific detection rates for the CNN were 100% for mucosal breaks, 97% for angioectasia, 99% for protruding lesions, and 100% for blood content.
• QuickView mode showed lower detection rates for mucosal breaks (91%), protruding lesions (80%), and blood content (96%).

Conclusions:

• A CNN-based system was successfully developed and validated for detecting various abnormalities in multicenter CE videos.
• The CNN system demonstrates superior performance and can serve as an effective alternative high-level screening tool to QuickView mode.