Computer-Aided Diagnosis of Colorectal Polyps: Clinical Usefulness and Limitations

Area of Science:

• Gastroenterology and Artificial Intelligence
• Medical Imaging and Diagnostics

Background:

• Computer-aided diagnosis (CADx) systems are being developed to assist in the real-time optical characterization of colorectal polyps during colonoscopy.
• These systems aim to improve diagnostic accuracy, consistency, and potentially reduce costs, supporting both expert and non-expert endoscopists.

Purpose of the Study:

• To critically evaluate the clinical utility and limitations of CADx systems in colorectal polyp diagnosis.
• To assess CADx performance metrics, influencing factors, and barriers to clinical adoption.

Main Methods:

• Narrative review of existing literature on CADx systems for colorectal polyp characterization.
• Focus on two primary clinical strategies: 'resect and discard' and 'leave in situ'.
• Analysis of key performance metrics (PPV, NPV, sensitivity, specificity) and influencing factors like training data and human-AI interaction.

Main Results:

• CADx system performance in clinical settings is highly variable and often falls below established benchmarks.
• Clinician trust and explainability issues lead to underutilization of accurate CADx predictions.
• Challenges persist in differentiating specific lesion types (e.g., sessile serrated vs. hyperplastic polyps) due to data limitations and ground truth issues.

Conclusions:

• While CADx offers potential benefits for diagnostic confidence and decision support, its widespread clinical integration is hindered by variable performance, trust deficits, and technical limitations.
• Addressing human-AI interaction, improving system transparency, and refining models for diverse lesion types are crucial for future adoption.
• Further validation is needed for emerging applications, such as predicting colorectal cancer invasion depth.