Pd27-01 Development of a Machine Learning (Ml) Model to Automatically and Precisely Identify Kidney Stones From Ureteroscopy Video Recordings

You have accessJournal of UrologySurgical Technology & Simulation: Artificial Intelligence II (PD27)1 May 2024PD27-01 DEVELOPMENT OF A MACHINE LEARNING (ML) MODEL TO AUTOMATICALLY AND PRECISELY IDENTIFY KIDNEY STONES FROM URETEROSCOPY VIDEO RECORDINGS Galen Cheng, Jixuan Leng, Junfei Liu, Jiebo Luo, Haohan Wang, Scott Quarrier, and Rajat Jain Galen ChengGalen Cheng , Jixuan LengJixuan Leng , Junfei LiuJunfei Liu , Jiebo LuoJiebo Luo , Haohan WangHaohan Wang , Scott QuarrierScott Quarrier , and Rajat JainRajat Jain View All Author Informationhttps://doi.org/10.1097/01.JU.0001008580.58088.27.01AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: The use of intraoperative video data from ureteroscopy allows for analysis for quality assurance and performance review. Currently, automated analysis of such data is limited. Recent advances in machine-learning methods and processing power allow for rapid analysis of images. However, the data must be segmented to precisely identify the object of analysis amidst the background noise. Manual segmentation is time and resource-intensive. The Segment Anything Model (SAM) has been previously used for analysis of natural images and video. In this study, we developed and tested a ML model that can automatically, accurately and efficiently identify stones from endoscopic images for further analysis. METHODS: Two high volume stone surgeons performed ureteroscopy for kidney stones. Representative video was recorded of the stone, and high quality individual frames were extracted from the videos. The correct location of the stone was determined by human tracing of the stone for each of the images tested–this tracing was considered "ground truth." A machine learning model using SAM and Unet, a neural network generated raw masks that outlined kidney stones in the images. Post-processing techniques enhanced the precision of these outlines. 5-fold cross validation was done to get an average precision score for segmentation. RESULTS: 1677 images were extracted from 78 videos. 1341 images were used for training and 336 images for testing. The ML model accurately segmented and classified 94.83% of the images, i.e. 5% of the image was misclassified on average. Dice coefficient (0.9129) and Intersection over Union (0.8486) confirmed good segmentation performance of the ML model. CONCLUSIONS: Here we demonstrate the development of an automated ML model that accurately identifies kidney stones from "natural" endoscopic images. This reduces resources required for segmentation and thus lowers the barrier for further analysis of endoscopic images. For example, we have developed a related ML model for prediction of stone chemical composition. These experiments are in early stages and require substantial further refinement before they can be routinely used. Download PPT Source of Funding: None © 2024 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 211Issue 5SMay 2024Page: e550 Advertisement Copyright & Permissions© 2024 by American Urological Association Education and Research, Inc.Metrics Author Information Galen Cheng More articles by this author Jixuan Leng More articles by this author Junfei Liu More articles by this author Jiebo Luo More articles by this author Haohan Wang More articles by this author Scott Quarrier More articles by this author Rajat Jain More articles by this author Expand All Advertisement PDF downloadLoading ...