Fully automated detection of pancreaticobiliary maljunction based on 3D magnetic resonance cholangiopancreatography using deep learning.
Authors
Abstract
Pancreaticobiliary maljunction (PBM) is closely related to biliary tract cancer. The early detection of PBM is crucial but challenging. We aim to develop and validate a deep learning model for fully automatically detection of PBM using magnetic resonance cholangiopancreatography (MRCP) images. Clinical and imaging data from patients who underwent MRCP examinations from January 2020 to December 2021 were retrospectively collected. A total of 200 patients (100 with confirmed PBM and 100 non-PBM controls) were enrolled. The dataset was randomly divided into training, validation, and test sets in a 6:2:2 ratio. The training and validation sets were used to train YOLOv5 and Inception-ResNetV2. The test set was used to evaluate the accuracy of the model. The PBM group had a higher proportion of females (62% vs. 42%, P = .01) and biliary tract cancer (11% vs. 3%, P = .03) than controls. The optimal model achieved a mean absolute error (MAE) of 6.2% and an F1 score of 93.0%, with sensitivity, specificity, positive predictive value, and negative predictive value of 95.0%, 92.5%, 92.7%, and 94.9%, respectively. The proposed deep learning model demonstrates high accuracy in automated PBM detection on MRCP images, offering potential to improve diagnostic efficiency and reduce interobserver variability.