Application Research Of Artificial Intelligence Model Based On YOLOv5s Algorithm To Assist Endoscopic Ultrasound Diagnosis Of Common Bile Duct Stones

Objective: Endoscopic ultrasound(EUS)is currently one of the most sensitive imaging methods for diagnosing common bile duct stones(CBDS).In this study,we developed an artificial intelligence(AI)model based on YOLOv5 s algorithm for recognizing CBDS in EUS examination videos,and explored the possibility and application potential of AI-assisted endoscopic ultrasound for diagnosing common bile duct stones through human-machine testing and data processing.Methods: The EUS images containing CBDS were retrospectively collected from Shengjing Hospital of China Medical University from January 2011 to February 2022 via endoscopic ultrasound,and after labeling the stones,they were randomly divided into training set,validation set and test set.The single-stage target detection algorithm YOLOv5 s was selected to develop the EUS-AI image model,and the obtained optimal model was used to construct the EUS-AI video model.EUS examination videos diagnosed as CBDS-positive and CBDS-negative by endoscopic ultrasound were retrospectively collected from March 2022 to January 2023,from which the CBDS-positive and negative video clips were edited for human-machine testing.The test was performed in three groups: physician group,AI group,and physician combined with AI group,and the physician group included four experienced endoscopic sonographers.The Kaplan-Meier method was used to process the test data,and the diagnostic accuracy was compared by log-rank test,and the cumulative diagnostic analysis curve and cumulative misdiagnosis risk curve were plotted.Results: A total of 2900 EUS images containing CBDS were screened and divided into1990 images in the training set,579 images in the validation set and 331 images in the test set.On the test set,the precision of the EUS-AI image model to detect CBDS was0.915,the recall was 0.87,and the m AP@.5 was 0.855.Sixty video clips containing CBDS and 53 video clips containing the common bile duct were randomly selected,totaling 113 cases used in human-machine test.Data processing and analysis:(1)The accuracy of the physician group was 86.3% and that of the AI group was 81.4%,with no statistically significant difference(P=0.179),showing that the diagnostic accuracy of the EUS-AI video model was not significantly different compared with that of human physicians,and the diagnostic performance was reliable;(2)the accuracy of the first 60 cases in the physician group was 89.6% and that of the last 53 cases was 83.0%,with a statistically significant difference(P=0.012),indicating that as the number of diagnostic cases increased,the accumulated fatigue caused a decrease in the accuracy of physician diagnosis and an increase in the risk of misdiagnosis;(3)the accuracy of the physician combined with AI group was 96.7% in the first 60 cases and 95.3% in the last 53 cases,with no statistically significant difference(P=0.311),indicating that AI-assisted diagnosis could eliminate the misdiagnosis caused by physician fatigue.(4)the accuracy was 86.3% in the physician group and 96.0% in the physician combined with AI group,with a statistically significant difference(P<0.001),indicating that AI-assisted diagnosis can effectively improve the accuracy rate of physicians.Conclusion: The diagnostic performance of the EUS-AI video model we developed is reliable,and the diagnostic accuracy is not significantly different from that of experienced endoscopic ultrasonographers.The model can effectively reduce misdiagnosis caused by physician fatigue,assist endoscopic ultrasonographers in diagnosing common bile duct stones,and improve the accuracy of diagnosis.This study demonstrates that artificial intelligence has the ability to assist physicians in improving the accuracy of EUS diagnosis of CBDS and has promising clinical applications.