Mp07-16 Computer Vision Analysis of Upper Tract Urothelial Carcinoma to Predict High vs Low Grade Pathology

You have accessJournal of UrologySurgical Technology & Simulation: Artificial Intelligence I (MP07)1 May 2024MP07-16 COMPUTER VISION ANALYSIS OF UPPER TRACT UROTHELIAL CARCINOMA TO PREDICT HIGH VS LOW GRADE PATHOLOGY Bryn M. Launer, Daiwei Lu, Ipek Oguz, and Nicholas L. Kavoussi Bryn M. LaunerBryn M. Launer , Daiwei LuDaiwei Lu , Ipek OguzIpek Oguz , and Nicholas L. KavoussiNicholas L. Kavoussi View All Author Informationhttps://doi.org/10.1097/01.JU.0001008728.41882.d7.16AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: Current treatment strategies for upper tract urothelial carcinoma (UTUC) depend on pathologic grade. Tumor pathology is usually determined via endoscopic biopsy, which can be challenging due to instrumentation and visualization limitations. We sought to implement an artificial intelligence-based computer vision model to evaluate videos of upper tract tumors during flexible ureteroscopy (fURS) and automatically distinguish between low-grade (LG) and high-grade (HG) pathology. METHODS: We collected 20 separate videos from 17 patients undergoing fURS for biopsy and treatment of UTUC tumors. The tumors were classified by pathology (LG or HG) based on microscopic pathologic evaluation. We extracted individual frames from the fURS videos at 30 FPS and developed a deep convolutional neural network to predict tumor pathology. Eighty percent of the frames (N=20,323) were used to train a deep convolutional neural network to predict pathologic outcome and 20% of frames (N=4,064) were reserved for model testing. Model performance of frame classification with a threshold of 0.5 was evaluated via accuracy, sensitivity, specificity and the area under the receiver operating curve (AUC-ROC). RESULTS: Forty percent (8/20) of the videos demonstrated LG UTUC tumors on pathologic analysis, with 60% (12/20) demonstrating HG tumors. The overall mean duration of the video clips for training was 16s±9. Mean video times of videos depicting LG and HG tumors were similar (17s vs. 16s). The computer vision model demonstrated good accuracy for frame classification by pathology (0.72), and a sensitivity and specificity of 0.72 and 0.73, respectively. The model showed fair performance across all thresholds of frame classification with an AUC-ROC of 0.67. CONCLUSIONS: Computer vision models show feasibility of distinguishing LG and HG UTUC pathology during fURS. Future model optimization with a more robust video dataset could provide intraoperative feedback and streamline treatment pathways. Download PPT Source of Funding: VISE physician in residence program (Nick Kavoussi). NIH R21 1R21DK133742 (Nick Kavoussi and Ipek Oguz). Training Program for Innovative Engineering Research in Surgery and Intervention Project Number 3T32EB021937 (Daiwei Lu) © 2024 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 211Issue 5SMay 2024Page: e111 Advertisement Copyright & Permissions© 2024 by American Urological Association Education and Research, Inc.Metrics Author Information Bryn M. Launer More articles by this author Daiwei Lu More articles by this author Ipek Oguz More articles by this author Nicholas L. Kavoussi More articles by this author Expand All Advertisement PDF downloadLoading ...