A deep learning-based linear system for liver station in EUS (with video)

Background and Objectives: With the widespread adoption of endoscopic techniques, the integration of EUS into hepatology has become an inevitable trend. This study aimed to develop a system that can conduct a comprehensive inspection of liver segments through scanning imaging at 8 stations. Methods: A standardized EUS examination protocol for the liver comprises 8 stations. We developed a deep learning-based station classification model using a dataset of 9403 EUS liver images. Internal validation was performed on 2780 EUS images and 26 EUS video clips. To further evaluate the model’s clinical utility, a human–machine competition involving 340 liver images was conducted. The same set of 26 video clips in internal video validation was also employed for a crossover study to evaluate the system’s effectiveness in reducing the difficulty of interpreting liver EUS images. Results: For station classification, the model achieved an accuracy of 92.52% on image validation and a per-frame accuracy of 85.98% on video validation. In the human–machine comparison, the overall classification accuracy reached 88.53%, which was comparable to the average accuracy of the EUS experts (84.56%, P = 0.064). In the crossover study, trainees’ station recognition accuracy improved significantly, with an increase of 16.36% (95% confidence interval, 11.43%–21.29%; P < 0.05). Conclusions: The system demonstrates high accuracy in liver station localization and achieves a classification performance comparable to EUS experts. It shows promise for facilitating standardized hepatic EUS examinations and may contribute to shortening the learning curve for trainees in the future.