Clinical Validation of AI-Assisted Evaluation of Indeterminate Biliary Strictures in Digital-Single Operator...

Biliary strictures (BS) are a significant challenge, with malignant strictures frequently diagnosed at advanced stages, limiting curative options. Digital single-operator cholangioscopy (D-SOC) enables high-resolution, direct visualization of the bile duct, yet with suboptimal accuracy. Artificial intelligence (AI) has shown promise for detection and differentiation of BS in frame-level analysis and small clinical series. This study aimed to validate a deep learning model for AI-assisted D-SOC image analysis.

This multicenter study included 135 D-SOC examinations from 129 patients (61 with malignant BS) across 14 centers in the United States, Brazil, Spain, Colombia, Australia, and Saudi Arabia. For each examination, up to 25 clinically relevant frames were selected and uploaded to a web-based platform for AI analysis. The model performed both detection and differentiation of BS: Detection was assessed by comparing AI-generated bounding boxes with expert-defined annotations using intersection-over-union (IoU), while differentiation was benchmarked against histopathology. Performance metrics included accuracy, sensitivity, specificity, and positive and negative predictive values.

At the patient level, malignant BS were identified with 86.0% accuracy, 84.1% sensitivity, and 85.7% specificity, with an area under the curve of 0.904. The model demonstrated robust detection performance, achieving a mean IoU of 70.3%. Performance was maintained across demographic variables and centers.

This first multicentric validation study demonstrates real-world performance of AI-assisted D-SOC analysis across multiples continents and devices, with robust accuracy for BS detection and differentiation. These findings support AI as an adjunctive tool in D-SOC, enhancing a more accurate evaluation of patients with indeterminate BS.