Impact of User's Background Knowledge and Polyp Characteristics in Colonoscopy with Computer-Aided Detection

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Interactions between human endoscopists and computer-aided detection (CADe) systems are crucial for improving colonoscopy outcomes.
• User experience and polyp characteristics significantly influence the effectiveness of CADe in clinical practice.

Purpose of the Study:

• To investigate how user experience and polyp characteristics affect the performance of human-CADe interactions.
• To evaluate the added value of CADe assistance in polyp detection based on histopathology and detection difficulty.

Main Methods:

• A CADe system utilizing YOLOv4 was developed and trained on a large dataset of polyp and sessile serrated lesion (SSL) images.
• The performance of CADe-assisted polyp detection was assessed through a computerized test module involving 18 participants with varying levels of colonoscopy experience (nurses, fellows, experts).
• Statistical analysis, including odds ratios and confidence intervals, was used to quantify the impact of CADe assistance on detection accuracy and sensitivity.

Main Results:

• CADe assistance improved overall polyp detection accuracy from 69.7% to 77.7% (OR, 1.88).
• The positive impact of CADe was most significant for nurses (OR, 1.97) and least for experts (OR, 1.42).
• While CADe enhanced detection of adenomas and easy polyps, it showed decreased sensitivity for SSLs and difficult-to-detect polyps, especially for less experienced users.

Conclusions:

• CADe systems offer a significant boost in polyp detection rates during colonoscopies.
• The effectiveness of CADe is modulated by the endoscopist's experience level, with greater benefits observed in less experienced individuals.
• Careful consideration of CADe system performance with challenging lesions and diverse user expertise is essential for optimizing its clinical utility.