PD27-07 ENHANCING PRECISION AND EFFICIENCY: DEEP LEARNING IN BLADDER TUMOR DETECTION
Zixing Ye,Yingjie Li,Zhigang Ji,Jianyang Lv,Hongbo Zhang,Yujiao Sun,Chengqing He
DOI: https://doi.org/10.1097/01.ju.0001008580.58088.27.07
2024-01-01
Abstract:You have accessJournal of UrologySurgical Technology & Simulation: Artificial Intelligence II (PD27)1 May 2024PD27-07 ENHANCING PRECISION AND EFFICIENCY: DEEP LEARNING IN BLADDER TUMOR DETECTION Zixing Ye, Yingjie Li, Zhigang Ji, Jianyang Lv, Hongbo Zhang, Yujiao Sun, and Chengqing He Zixing YeZixing Ye , Yingjie LiYingjie Li , Zhigang JiZhigang Ji , Jianyang LvJianyang Lv , Hongbo ZhangHongbo Zhang , Yujiao SunYujiao Sun , and Chengqing HeChengqing He View All Author Informationhttps://doi.org/10.1097/01.JU.0001008580.58088.27.07AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: Thorough cystoscopy is critical for early lesion detection and tumor recurrence surveillance. However, conventional examination based on visual observation and subjective judgment often leads to inconsistency and may overlook potential malignancy. To improve the precision and efficiency, we adopted a strategy to integrate an innovative deep learning algorithm with intelligent bladder tumor detection and validated its potency in different identifiability groups in this multi-center study. METHODS: Videos of 99 patients who underwent cystoscopy or transurethral resection of bladder tumor in three Chinese hospitals over a period of 6 months were analyzed. A total of 106 video clips comprising 35268 frames with suspicious bladder lesions (including non-invasive urothelial neoplasia and invasive urothelial carcinoma) were collected and randomly divided into training and test sets in an 8:2 ratio. Within the test set, 6101 frames were further categorized into high-resolution group and 1153 frames into low-resolution group based on recognizability. Manual annotation of bladder lesion outlines was performed, and HRNet-V2, a remarkable semantic segmentation model was employed for subsequent training and validation. Measurements of performance included sensitivity, precision, and mean Dice coefficient. RESULTS: The overall sensitivity and precision of the test set were 90.1% and 88.7%, with 78.4% mean Dice coefficient. In the high-resolution group, the sensitivity and precision were 92.9% and 91.6%, while in the low-resolution group, they were 74.4% and 73.1%, respectively. The mean Dice coefficients for the two groups were 82.1% and 52.2%, respectively. CONCLUSIONS: Deep learning achieved excellent performance in bladder lesion detection in terms of high sensitivity and precision, particularly in high-resolution group. Additional training with more clinical information will contribute to improved diagnosis efficiency and personalized treatment decision. Download PPT Source of Funding: None © 2024 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 211Issue 5SMay 2024Page: e553 Advertisement Copyright & Permissions© 2024 by American Urological Association Education and Research, Inc.Metrics Author Information Zixing Ye More articles by this author Yingjie Li More articles by this author Zhigang Ji More articles by this author Jianyang Lv More articles by this author Hongbo Zhang More articles by this author Yujiao Sun More articles by this author Chengqing He More articles by this author Expand All Advertisement PDF downloadLoading ...