Prospective assessment of pancreatic ductal adenocarcinoma diagnosis from endoscopic ultrasonography images with the assistance of deep learning

BackgroundEndosonographers are highly dependent on the diagnosis of pancreatic ductal adenocarcinoma (PDAC). The objectives of this study were to develop a deep-learning radiomics (DLR) model based on endoscopic ultrasonography (EUS) images for identifying PDAC and to explore its true clinical benefit. MethodsA retrospective data set of EUS images that included PDAC and benign lesions was used as a training cohort (N = 368 patients) to develop the DLR model, and a prospective data set was used as a test cohort (N = 123 patients) to validate the effectiveness of the DLR model. In addition, seven endosonographers performed two rounds of reader studies on the test cohort with or without DLR assistance to further assess the clinical applicability and true benefits of the DLR model. ResultsIn the prospective test cohort, DLR exhibited an area under the receiver operating characteristic curves of 0.936 (95% confidence interval [CI], 0.889-0.976) with a sensitivity of 0.831 (95% CI, 0.746-0.913) and 0.904 (95% CI, 0.820-0.980), respectively. With DLR assistance, the overall diagnostic performance of the seven endosonographers improved: one endosonographer achieved a significant expansion of specificity (p = .035,) and another achieved a significant increase in sensitivity (p = .038). In the junior endosonographer group, the diagnostic performance with the help of the DLR was higher than or comparable to that of the senior endosonographer group without DLR assistance. ConclusionsA prospective test cohort validated that the DLR model based on EUS images effectively identified PDAC. With the assistance of this model, the gap between endosonographers at different levels of experience narrowed, and the accuracy of endosonographers expanded.