Integrating Artificial Intelligence into Endoscopic Decision-Making: ANN-Assisted CT for Common Bile Duct Stones

Background Computed tomography (CT) is widely used in the initial evaluation of suspected common bile duct (CBD) stones, but limited sensitivity often necessitates additional endoscopic procedures. We developed and validated an artificial neural network (ANN) to enhance CT interpretation and assessed its potential to support endoscopic decision-making. Methods We An ANN model integrating UNETR for segmentation and ResNet50 for classification was trained to detect CBD stones on CT. Patients who underwent abdominal CT for suspected CBD stones between March 2018 and June 2023 at Hallym University Kangnam Sacred Heart Hospital were included. A retrospective derivation cohort (n = 830) was used for model training, and a prospective validation cohort (n = 225) for testing, with endoscopic retrograde cholangiopancreato- graphy (ERCP) serving as the reference standard. ANN performance was compared with that of expert radiologists and trainee radiologists with ANN assistance. Multivariate analysis evaluated clinical factors influencing diagnostic accuracy, and heatmap visualization assessed interpretability relevant to endoscopic decision-making. Results The ANN achieved diagnostic accuracy comparable to expert radiologists (93.3% vs. 93.8%). When assisting trainees, accuracy improved from 82.2% (AUC 0.82) to 91.1% (AUC 0.91), approaching expert performance (93.8%; AUC 0.94). Stone type and bile duct diameter > 10 mm significantly increased ANN detection rates. Heatmap visualization confirmed the plausibility of ANN predictions in both clearly identifiable lesions and indeterminate CT findings, improving interpretability for endoscopic decision-making. Conclusions The ANN achieved expert-level diagnostic accuracy for detecting CBD stones. By enhancing CT interpretation, it may optimize ERCP indications, reduce unnecessary invasive procedures, and improve training for less-experienced clinicians. Prospective validation and integration into multimodal endoscopic workflows are warranted.