PD42-09 URINE CELL IMAGE ANALYSIS USING A DEEP LEARNING MODEL

Abstract

You have accessJournal of UrologyBladder Cancer: Basic Research & Pathophysiology III (PD42)1 Apr 2020PD42-09 URINE CELL IMAGE ANALYSIS USING A DEEP LEARNING MODEL Masatomo Kaneko*, Keisuke Tsuji, Keiichi Masuda, Kengo Ueno, Kohei Henmi, Shota Nakagawa, Ryo Fujita, Kensho Suzuki, Yuichi Inoue, Hikaru Shindo, Eiichi Konishi, Tetsuro Takamatsu, and Osamu Ukimura Masatomo Kaneko*Masatomo Kaneko* More articles by this author , Keisuke TsujiKeisuke Tsuji More articles by this author , Keiichi MasudaKeiichi Masuda More articles by this author , Kengo UenoKengo Ueno More articles by this author , Kohei HenmiKohei Henmi More articles by this author , Shota NakagawaShota Nakagawa More articles by this author , Ryo FujitaRyo Fujita More articles by this author , Kensho SuzukiKensho Suzuki More articles by this author , Yuichi InoueYuichi Inoue More articles by this author , Hikaru ShindoHikaru Shindo More articles by this author , Eiichi KonishiEiichi Konishi More articles by this author , Tetsuro TakamatsuTetsuro Takamatsu More articles by this author , and Osamu UkimuraOsamu Ukimura More articles by this author View All Author Informationhttps://doi.org/10.1097/JU.0000000000000922.09AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: Urine cytology is a noninvasive and inexpensive screening test, however, the sensitivity is low with respect to low grade urothelial cancer. In addition, urine cytology has poor interobserver reproducibility. In order to reduce the workload of both pathologists and cytotechnologists and to improve the reproducibility and accuracy, there is a great need for tools for urine cytology. Recently, with the advance of Artificial Intelligence (AI) technology using a convolutional neural network algorithm for image analysis, some studies reported promising results. In urine cytology, image analysis using AI is expected to be a method that not only solves the above mentioned problem but also contributes to reduction in the burden on cytologists and pathologists. In this study, we developed AI for automated urine cell image analysis using a deep learning model in collaboration with Kyocera Communication System Co., Ltd. and Rist Co., Ltd. METHODS: We collected a total of 195 consecutive urine samples which were obtained from patients eventually histologically diagnosed with urothelial cancer in our hospital from January 2016 to December 2017. The urine samples was classified as inadequate in 1 sample (0.5%), negative for high grade urothelial carcinoma (HGUC) in 77 samples (39.5%), atypical urothelial cells in 36 samples (18.5%), suspicious for HGUC in 17 samples (8.7%) and malignant in 64 cases (32.8%) which equivalent to HGUC, low-grade urothelial neoplasm, and other malignancies according to The Paris System. Histological grading were low grade in 90 cases (46.2%) and high grade in 105 cases (53.8%). Two experienced cytotechnologists independently evaluated urine cytology slides, and 4637 cell images whose cell evaluation between 2 observers were concordant were collected. 3128 benign cells were labeled as negative cells and 398 atypical cells and 1111 malignant cells were labeled as positive. The labeled dataset was used as ground truth. We developed a convolutional neural network algorithm which classify a cell image into positive or negative, and split the dataset into 80% for training and validation, and 20% for test data. After 5-Fold cross validation, the receiver operator characteristic (ROC) analysis was used to evaluate the binary classification model. RESULTS: The area under the ROC curve was 0.984, the average accuracy was 0.943, the sensitivity was 92.0%, and the specificity was 95.7%. CONCLUSIONS: We developed urine cell image analysis algorithm performing highly accurate result. The algorithm can be useful to assist cytotechnologists and pathologists. Source of Funding: None. © 2020 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 203Issue Supplement 4April 2020Page: e830-e830 Advertisement Copyright & Permissions© 2020 by American Urological Association Education and Research, Inc.MetricsAuthor Information Masatomo Kaneko* More articles by this author Keisuke Tsuji More articles by this author Keiichi Masuda More articles by this author Kengo Ueno More articles by this author Kohei Henmi More articles by this author Shota Nakagawa More articles by this author Ryo Fujita More articles by this author Kensho Suzuki More articles by this author Yuichi Inoue More articles by this author Hikaru Shindo More articles by this author Eiichi Konishi More articles by this author Tetsuro Takamatsu More articles by this author Osamu Ukimura More articles by this author Expand All Advertisement PDF downloadLoading ...