You have accessJournal of UrologyBladder Cancer: Non-invasive IV (MP73)1 Apr 2020MP73-19 DEEP LEARNING FOR HISTOPATHOLOGY AND CLINICAL CHARACTERISTICS BASED RECURRENCE PREDICTION IN NON-MUSCLE INVASIVE BLADDER CANCER PATIENTS Marit Lucas*, Ilaria Jansen, Jorg R. Oddens, C. Dilara Savci-Heijink, Ton G. van Leeuwen, Daniel M. de Bruin, and Henk A. Marquering Marit Lucas*Marit Lucas* More articles by this author , Ilaria JansenIlaria Jansen More articles by this author , Jorg R. OddensJorg R. Oddens More articles by this author , C. Dilara Savci-HeijinkC. Dilara Savci-Heijink More articles by this author , Ton G. van LeeuwenTon G. van Leeuwen More articles by this author , Daniel M. de BruinDaniel M. de Bruin More articles by this author , and Henk A. MarqueringHenk A. Marquering More articles by this author View All Author Informationhttps://doi.org/10.1097/JU.0000000000000959.019AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: The EORTC and CUETO are used to assess the risk for recurrence in non-muscle invasive bladder cancer (NMIBC). Low concordance rates with recurrence are reported for those tools, and could be explained a.o. by the subjectivity of the histopathological parameters used in the prediction. By the use of deep learning, we want to objectify the recurrence prediction. This deep learning approach aims to combine clinical characteristics, including histopathological parameters, with the histopathology images. In this abstract, we report on the first steps taken in the development of this deep learning approach. METHODS: All patients that underwent a trans-urethral resection of bladder tumor in the Amsterdam UMC, location AMC, between 2000 and 2018 were included in this study. Histopathological slides of NMIBC patients were retrieved from the archive and digitized by the Philips Ultrafast scanner. A three-stage deep learning approach is proposed to assess the risk of recurrence. The first stage is the detection of the urothelium on the histopathology slides. The second stage selects regions within the urothelium which are most predictive for recurrence. The last step is a deep learning network which combines those regions with the clinical characteristics. The short (1 yr)-term recurrence dataset contained 395 patients, with 26% of recurred patients. The long (5 yr)-term dataset contained 306 patients, with recurrence in 63%. Missing data in clinical characteristics were imputed using a random forest. The three step deep learning approach will be compared with a multivariable logistic regression model based on clinical characteristics only. RESULTS: The first step from the deep learning approach has been completed and resulted in an automated and accurate detection of urothelium. The results from the other two stages will be presented at the AUA Annual Meeting in May. Multivariable logistic regression based on clinical characteristics only shows low predictive power, with a sensitivity of 11% and 72%, specificity of 96% and 35% and AUC of 0.60 and 0.62 for the prediction of recurrence within 1 yr and 5 yr, respectively. CONCLUSIONS: Accurate detection of urothelium is achieved with deep learning, and is the first step in order to create a deep learning approach for recurrence prediction. The next steps consists of the automated detection of regions on histopathology slides predictive for recurrence and at last the incorporation of clinical data into the recurrence prediction of NMIBC patients. Source of Funding: Cure for Cancer Foundation, ITEA3 (ITEA151003) © 2020 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 203Issue Supplement 4April 2020Page: e1129-e1129 Advertisement Copyright & Permissions© 2020 by American Urological Association Education and Research, Inc.MetricsAuthor Information Marit Lucas* More articles by this author Ilaria Jansen More articles by this author Jorg R. Oddens More articles by this author C. Dilara Savci-Heijink More articles by this author Ton G. van Leeuwen More articles by this author Daniel M. de Bruin More articles by this author Henk A. Marquering More articles by this author Expand All Advertisement PDF downloadLoading ...