Prospective Video-Based Study Assessing Effect of Computer-Assisted Optical Diagnosis on Distinguishing Serrated,...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Current computer-aided diagnosis (CADx) systems lack training for sessile-serrated lesions (SSL), misclassifying them as hyperplastic polyps (HP) or tubular adenomas (TA).
• This misclassification by CADx may impact endoscopists' diagnostic accuracy for SSL.

Purpose of the Study:

• To evaluate the impact of current CADx systems, not specifically trained for SSL, on the diagnostic accuracy of endoscopists in recognizing SSL.
• To assess how CADx performance on HP and TA influences endoscopist accuracy for these lesions.

Main Methods:

• 17 endoscopists reviewed 115 polyp videos/images (33% SSL, 33% HP, 34% TA).
• Polyp presentations were randomized, with 50% shown with a CADx diagnosis and 50% without.
• CADx performance was evaluated: it identified 52.6% of SSL as HP and 47.4% as TA, while correctly identifying 68.4% of HP and 82.1% of TA.

Main Results:

• Endoscopist accuracy for SSL recognition did not significantly differ with or without CADx display (64.1% vs. 64.9%, p > 0.05).
• Diagnostic accuracy for HP and TA improved when endoscopists received correct CADx output but worsened with incorrect CADx output.
• CADx misclassified SSL as HP (52.6%) or TA (47.4%).

Conclusions:

• CADx systems not trained for SSL recognition do not significantly alter endoscopist accuracy for these lesions, even with incorrect diagnostic suggestions.
• Accurate CADx outputs for HP and TA can improve endoscopist diagnostic accuracy for those specific polyp types.