Deep Learning-Based Detection of Malignant Bile Duct Stenosis in Fluoroscopy Images of Endoscopic Retrograde Cholangiopancreatography.

Introduction: The accurate distinction between benign and malignant biliary strictures (BS) poses a significant challenge. Despite the use of bile duct biopsies and brush cytology via endoscopic retrograde cholangiopancreaticography (ERCP), the results remain suboptimal. Single-operator cholangioscopy can enhance the diagnostic yield in BS, but its limited availability and high costs are substantial barriers. Convolutional Neural Network (CNN)-based systems may improve the diagnostic process and enhance reproducibility. Therefore, we assessed the feasibility of using deep learning to differentiate BS using fluoroscopy images during ERCP.

Methods: We conducted a retrospective review of adult patients (n=251) from three university centers in Germany (Leipzig, Dresden, Halle) who underwent ERCP. We developed and evaluated a deep learning-based model using fluoroscopy images. The performance of the classifier was evaluated by measuring the area under the receiver operating characteristic curve (AUROC) and we utilized saliency map analyses to understand the decision-making process of the model.

Results: In cross-validation experiments, malignant BS were detected with a mean AUROC of 0.89 ± 0.03. The test set of the Leipzig cohort demonstrated an AUROC of 0.90. In two independent external validation cohorts (Dresden, Halle), the deep learning-based classifier achieved an AUROC of 0.72 and 0.76, respectively. The artificial intelligence model's predictions identified plausible characteristics within the fluoroscopy images.

Conclusion: By using a deep learning model, we were able to discriminate malignant BS from benign biliary conditions. The application of artificial intelligence enhances the diagnostic yield of malignant BS and should be validated in a prospective design.