OBJECTIVE To develop an auto-categorization system based on machine learning for three-dimensional magnetic resonance cholangiopancreatography (3D MRCP) to detect choledocholithiasis from healthy and symptomatic individuals. METHODS 3D MRCP sequences from 254 cases with common bile duct (CBD) stones and 251 cases with normal CBD were enrolled to train the 3D Convolutional Neural Network (3D-CNN) model. Then 184 patients from three different hospitals (91 with positive CBD stone and 93 with normal CBD) were prospectively included to test the performance of 3D-CNN. RESULTS With a cutoff value of 0.2754, 3D-CNN achieved the sensitivity, specificity, and accuracy of 94.51%, 92.47%, and 93.48%, respectively. In the receiver operating characteristic curve analysis, the area under the curve (AUC) for the presence or absence of CBD stones was 0.974 (95% CI, 0.940-0.992). There was no significant difference in sensitivity, specificity, and accuracy between 3D-CNN and radiologists. In addition, the performance of 3D-CNN was also evaluated in the internal test set and the external test set, respectively. The internal test set yielded an accuracy of 94.74% and AUC of 0.974 (95% CI, 0.919-0.996), and the external test set yielded an accuracy of 92.13% and AUC of 0.970 (95% CI, 0.911-0.995). CONCLUSIONS An artificial intelligence-assisted diagnostic system for CBD stones was constructed using 3D-CNN model for 3D MRCP images. The performance of 3D-CNN model was comparable to that of radiologists in diagnosing CBD stones. 3D-CNN model maintained high performance when applied to data from other hospitals.