Aims Accurate distinction between benign and malignant biliary strictures (BS) is challenging.
The use of bile duct biopsies and brush cytology via endoscopic retrograde cholangiopancreaticography
(ECRP) remains suboptimal. Single-operator cholangioscopy increases the diagnostic
yield in BS but has limited availability and high costs. Convolutional neural network
(CNN)-based systems have the potential to assist in the diagnostic process and improve
reproducibility. Thus, we assessed the feasibility of using deep learning to differentiate
BS out of 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 an ERCP. We developed and
evaluated a deep learning-based model (DenseNet) by means of fluoroscopy images. We
measured the area under the receiver operating characteristic curve (AUROC) to evaluate
the performance of the classifier and used saliency maps analyses to understand the
decision-making process of the model.
Results In cross-validation (Leipzig cohort), malignant BS were detected with an mean AUROC
of 0.88±0.02. On two independent external validation cohorts (Dresden, Halle), the
of the deep learned based classifier reached a mean AUROC of of 0.71+±+0.04 and 0.74+±+0.07,
respectively. The artificial intelligence model's predictions identify plausible
characteristics within the fluoroscopy images.
Conclusions By using a deep learning model, we were able to discriminate malignant BS from benign
biliary conditions. Artificial intelligence improves the diagnostic yield of malignant
BS and needs to be validated in prospective design.
