PD03-04 SUPPORT SYSTEM OF CYSTOSCOPIC DIAGNOSIS FOR BLADDER CANCER BASED ON ARTIFICIAL INTELLIGENCE SURPASSES UROLOGISTS

Abstract

You have accessJournal of UrologyBladder Cancer: Non-invasive I (PD03)1 Apr 2020PD03-04 SUPPORT SYSTEM OF CYSTOSCOPIC DIAGNOSIS FOR BLADDER CANCER BASED ON ARTIFICIAL INTELLIGENCE SURPASSES UROLOGISTS Atsushi Ikeda*, Yuta Kochi, Hirokazu Nosato, Takahiro Kojima, Hidenori Sakanashi, Masahiro Murakawa, and Hiroyuki Nishiyama Atsushi Ikeda*Atsushi Ikeda* More articles by this author , Yuta KochiYuta Kochi More articles by this author , Hirokazu NosatoHirokazu Nosato More articles by this author , Takahiro KojimaTakahiro Kojima More articles by this author , Hidenori SakanashiHidenori Sakanashi More articles by this author , Masahiro MurakawaMasahiro Murakawa More articles by this author , and Hiroyuki NishiyamaHiroyuki Nishiyama More articles by this author View All Author Informationhttps://doi.org/10.1097/JU.0000000000000823.04AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: Cystoscopy is the primary diagnostic modality for bladder cancer. However, the level of experience of urologists may affect the accuracy of the diagnosis. We developed an artificial intelligence (AI)-based system to improve the detection of tumors that may otherwise be missed in regular cystoscopy. In our system, we employ the principle of deep learning, specifically transfer learning, to enable anomaly detection using gastroscopic images by extracting features that apply to cystoscopic images. We previously demonstrated that the system was capable of differentiating images of normal tissue from those that contain papillary or flat lesions irrespective of their size or location. In the present study, we evaluated the performance of the AI system by having medical students and physicians with different levels of experience assess cystoscopic images that were previously examined by the AI system for the diagnosis of bladder cancer. METHODS: We created a test dataset containing 422 cystoscopic images (tumor image: 87, normal image: 335). Besides, we developed software that enables images to be shown continuously in random order. Using the software, we asked an observer to determine the presence of a tumor in each image. A total of 53 observers, which consisted of 36 medical students, four junior residents, four beginner urologists, and nine expert urologists, assessed the images. Based on the results, we calculated the diagnostic accuracy for each group. For the AI system, we applied a stepwise transfer learning in convolutional neural networks using 1680 cystoscopic images and 8,728 gastroscopic images of the cardia. Using the same test dataset, receiver operating characteristic (ROC) curves were constructed to assess the diagnostic accuracy. RESULTS: The mean scores for Youden’s index (sensitivity + specificity - 1) for medical students, junior residents, beginner urologists, and expert urologists were 0.687, 0.787, 0.906, and 0.914, respectively. For the AI system, the maximum score for Youden’s index was 0.930. The median time to diagnosis was 634 for medical students and urologists and 5 seconds for the AI system. CONCLUSIONS: Our study demonstrated that the diagnostic accuracy was correlated with the level of experience of the study subjects. Also, the AI system had an equivalent or higher level of diagnostic performance to that of expert urologists. Source of Funding: This work was supported in part by JSPS KAKENHI Grant Number JP16678973 and JP17K16775. © 2020 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 203Issue Supplement 4April 2020Page: e73-e73 Advertisement Copyright & Permissions© 2020 by American Urological Association Education and Research, Inc.MetricsAuthor Information Atsushi Ikeda* More articles by this author Yuta Kochi More articles by this author Hirokazu Nosato More articles by this author Takahiro Kojima More articles by this author Hidenori Sakanashi More articles by this author Masahiro Murakawa More articles by this author Hiroyuki Nishiyama More articles by this author Expand All Advertisement PDF downloadLoading ...