Human-machine interaction in optical polyp diagnosis: decision-making after CADx polyp diagnosis

Area of Science:

• Gastroenterology
• Medical Informatics

Background:

• Artificial intelligence (AI)-based computer-aided diagnosis (CADx) systems are increasingly utilized for colorectal polyp classification during colonoscopy.
• The effectiveness of these AI systems hinges on endoscopists' interpretation of AI-generated predictions.

Purpose of the Study:

• To evaluate endoscopist decision-making when presented with AI CADx outputs for colorectal polyps.
• To compare endoscopist decisions against histopathology as the ground truth.

Main Methods:

• Secondary analysis of a prospective study (NCT06822816) involving polyps ≤10 mm.
• Inclusion of cases with documented AI CADx output and endoscopist diagnosis.
• Assessment of primary outcome: accuracy of accepted AI CADx diagnoses; secondary outcomes: reclassification patterns after AI output rejection.

Main Results:

• The study included 754 polyps from 367 patients.
• Accepted AI CADx diagnoses were accurate for 89.1% of neoplastic and 68.7% of hyperplastic polyp predictions.
• Endoscopists accepted incorrect AI classifications more often for neoplastic (64.9%) than hyperplastic (26.5%) polyps. Rejection of correct AI neoplastic and hyperplastic predictions led to misclassification as sessile serrated lesions (SSLs) in 23.4% and 29.5% of cases, respectively.

Conclusions:

• Endoscopists demonstrate a diagnostic bias towards neoplasia, frequently misclassifying hyperplastic polyps as SSLs.
• This bias, while potentially enhancing safety, may lead to increased surveillance colonoscopies and healthcare costs.