REAL-TIME CHOLANGIOSCOPY ARTIFICIAL INTELLIGENCE-ASSISTED DIAGNOSIS OF BILIARY TRACT MALIGNANCY: A PROSPECTIVE FEASIBILITY TRIAL

Introduction: Indeterminate biliary strictures persist as a diagnostic challenge. Diagnostic tools include brush cytology (BC) and forceps biopsy (FB); however, these modalities have poor sensitivity for detecting malignancy. Artificial intelligence (AI) has the promise of providing objective interpretation of cholangioscopy video analysis and has been demonstrated to have improved performance for detecting biliary tract malignancy compared to other diagnostic modalities. This clinical trial aimed to assess the feasibility and performance of a point-of-care, real-time cholangioscopy AI system.
Methods: A previously developed AI was directly uploaded into a computer system without further retraining. The system receives footage directly from a cholangioscope console, processes the video stream, applies AI to the incoming images, and then outputs real-time predictions for the endoscopist. The system provides an overall assessment as to whether the entire patient video contains malignancy (AI-CADx) (Figure 1A-C). From a single center that was not involved in training the AI, consecutive patients undergoing cholangioscopy procedures were approached for enrollment. The system was deployed if the patient consented and if cholangioscopy was utilized. Once deployed, the system was positioned within view of the endoscopist. The per-patient performance of the AI-CADx was compared to that of BC and FB for the diagnosis of biliary strictures.
Results: A total of 26 cholangioscopy examinations for 20 patients were performed during the trial. Three different endoscopists utilized the cholangioscopy AI system. Of the examinations performed, 20 were for evaluation of strictures (10 malignant, 6 benign, and 4 requiring follow-ups for definitive diagnosis) and 6 for treatment of choledocholithiasis. In addition, biliary stents were removed prior to cholangioscopy in 61.5% of cases. For classifying strictures, the AI-CADx was 92.3% accurate (95% CI: 64.0-99.8%), BC was 62.5% accurate (95% CI: 24.5-91.5%), and FB was 75.0% accurate (95% CI: 34.9-96.8%) (Figure 2). In addition, the AI correctly predicted that all patients undergoing evaluation of choledocholithiasis did not have malignancy. Overall, the real-time AI-CADx was 94.4% accurate (95% CI: 72.7-99.9%) in classifying patients as having biliary tract malignancy or not. The AI provided real-time predictions at 8 frames/second.
Discussion: This prospective clinical trial demonstrates the feasibility and accuracy of real-time cholangioscopy AI system for the evaluation of biliary pathology. Larger multicenter trials are needed to confirm the performance benefits of the AI compared to BC and FB. In the future, predictions of a cholangioscopy-AI could be included as part of the ensemble of diagnostic tests used to evaluate patients with indeterminate biliary strictures.