Autonomous Artificial Intelligence Versus AI Assisted Human Optical Diagnosis

Overview

Computer-aided image-enhanced endoscopy can predict the nature of colorectal polyps with over 90% accuracy. This technology uses artificial intelligence (AI) to analyze video recordings of polyps, learning to make diagnoses in real-time. This means that doctors can get immediate predictions about small polyps during the procedure, reducing the need for separate pathology exams and saving costs, ultimately improving patient care. Human and AI interactions are complex and a framework to reap synergistic effects CADx systems when used by humans to harness optimal performance needs to be established. AI solutions in medicine are usually developed to be used as assistive devices, however, then they rely on humans to correct AI errors. Optical polyp diagnosis is a complex task. Non experts usually achieve diagnostic accuracy in 70-80%. CADx systems have a similar diagnostic accuracy when used autonomously. Clinical evaluation of CADx systems showed that CADx assisted OD performs equally to the operator performance when using non CADx assisted OD. To harness a benefit of clinical CADx implementation we would have to find a way that synergies between human and CADx come into play to eliminate cases in which CADx assisted and/ or human OD results in low diagnostic accuracy and also addresses the problem of serrated polyp recognition.

Our study hypothesis is that for CADx implementation, instead of using the high/low confidence framework, identifying cases with suboptimal diagnostic accuracy could be facilitated through identifying cases in which CADx and endoscopist disagreed in their diagnosis. Eliminating such cases might separate out cases with low accuracy when using CADx assisted OD. Since endoscopists have a high sensitivity but low specificity for serrated polyp OD, this framework will also allow us to implement a strategy to adequately manage serrated polyps found in the cohort.