Towards full integration of explainable artificial intelligence in colon capsule endoscopy's pathway

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Colon capsule endoscopy (CCE) shows promise for colorectal disease diagnosis but lags behind optical colonoscopy (OC) due to manual image analysis challenges.
• Current limitations include poor bowel preparation, logistical issues, and the time-consuming manual review of CCE images.
• Integrating AI aims to bridge this gap by automating critical image processing tasks.

Purpose of the Study:

• To develop and evaluate a comprehensive AI system for the autonomous detection, localization, characterization, and size estimation of findings in CCE.
• To enhance the efficiency and accuracy of CCE analysis, facilitating its routine clinical integration.
• To reduce unnecessary follow-up procedures by accurately identifying significant colorectal polyps.

Main Methods:

• Development of explainable deep neural networks (DNNs) for polyp detection, characterization, and size estimation within CCE image sequences.
• Training and validation of algorithms on extensive, unaugmented image datasets, including normal mucosa and various polyp types.
• Implementation of a multi-stage AI pipeline for autonomous image processing and data integration into the CCE workflow.

Main Results:

• The polyp detection DNN achieved high sensitivity ([Formula: see text]) and specificity ([Formula: see text]), with an excellent negative predictive value ([Formula: see text]).
• The characterization DNN accurately classified polyps as neoplastic or non-neoplastic with [Formula: see text] sensitivity and [Formula: see text] specificity.
• The size estimation DNN demonstrated high accuracy ([Formula: see text]) in polyp segmentation.

Conclusions:

• Explainable AI can automate key aspects of CCE image analysis, significantly improving diagnostic performance.
• This AI-driven approach moves CCE towards seamless integration into clinical practice, comparable to OC.
• Automated analysis reduces the burden of manual review and aids in optimizing patient management by identifying clinically significant findings.