PD56-08 A DEEP LEARNING–BASED ALGORITHM FOR AUTOMATED DETECTION AND CLASSIFICATION OF PROSTATE LESIONS ON BIPARAMETRIC MAGNETIC RESONANCE IMAGING
Sherif Mehralivand,Dong Yang,Daguang Xu,Ziyue Xu,Holger Roth,Stephanie Harmon,Samira Masoudi,Thomas Sanford,Deepak Kesani,Nathan Lay,Bradford Wood,Peter Choyke,Peter Pinto,Baris Turkbey
DOI: https://doi.org/10.1097/ju.0000000000002090.08
IF: 7.6
2021-01-01
The Journal of Urology
Abstract:You have accessJournal of UrologyProstate Cancer: Detection & Screening VI (PD56)1 Sep 2021PD56-08 A DEEP LEARNING–BASED ALGORITHM FOR AUTOMATED DETECTION AND CLASSIFICATION OF PROSTATE LESIONS ON BIPARAMETRIC MAGNETIC RESONANCE IMAGING Sherif Mehralivand, Dong Yang, Daguang Xu, Ziyue Xu, Holger Roth, Stephanie Harmon, Samira Masoudi, Thomas Sanford, Deepak Kesani, Nathan Lay, Bradford Wood, Peter Choyke, Peter Pinto, and Baris Turkbey Sherif MehralivandSherif Mehralivand More articles by this author , Dong YangDong Yang More articles by this author , Daguang XuDaguang Xu More articles by this author , Ziyue XuZiyue Xu More articles by this author , Holger RothHolger Roth More articles by this author , Stephanie HarmonStephanie Harmon More articles by this author , Samira MasoudiSamira Masoudi More articles by this author , Thomas SanfordThomas Sanford More articles by this author , Deepak KesaniDeepak Kesani More articles by this author , Nathan LayNathan Lay More articles by this author , Bradford WoodBradford Wood More articles by this author , Peter ChoykePeter Choyke More articles by this author , Peter PintoPeter Pinto More articles by this author , and Baris TurkbeyBaris Turkbey More articles by this author View All Author Informationhttps://doi.org/10.1097/JU.0000000000002090.08AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: The Prostate Imaging-Reporting and Data System (PI-RADS) version 2.1 has become the standard in risk categorization of prostate lesions. However, there is significant inter-reader variability. Recent advancements in Deep Learning (DL) have led to improved diagnostic algorithms in medical imaging. However, a common problem of DL algorithms is their single-center training and testing which leads to institution-related overfitting. We sought to train and test a DL-based algorithm for automated detection and PI-RADS classification on prostate biparametric MRI using a diverse multi-institutional patient dataset. METHODS: Overall, 1247 patients scans from two different institutions were used with 204 scans derived from a publicly available dataset (Prostate X Challenge, Nijmegen, The Netherlands) and 1043 scans from our institution. All scans were reevaluated by a single genitourinary radiologist with > 14 years of experience and > 1500 reads per year. The model consists of three cascaded 3D Deep Neural Networks: 1. Segmentation of the whole prostate, 2. Detection and segmentation of suspicious prostate lesions and 3. PI-RADS classification using 3D U-Net and 3D ResNet architectures. Detection was assessed by sensitivity and false positive rate, classification by accuracy and segmentation congruence by Dice similarity coefficients. RESULTS: The 1247 patient scans were subdivided into training and validation set (n=1097, 88%) for 5-fold cross-validation and an independent test set (n=150, 12%). In the test set, the model correctly detected 139/194 lesions leading to a 71.65% detection sensitivity. This came at the cost of 229 false positives leading to a mean false positive rate of 1.65 false positives per true positive lesion (Figure 1). Correct classification among PI-RADS categories was distributed as follows: PI-RADS category 2 0/23 (0%), PI-RADS category 3 15/38 (39.47%), PI-RADS category 4 41/91 (45.05%), PI-RADS category 5 35/39 (89.74%). Mean lesion segmentation congruence (Dice) was 0.385 (range 0.357-0.387). CONCLUSIONS: We developed a DL-based 3D algorithm with promising detection and PI-RADS classification performance for prostate lesions on biparametric MRI. This system can be used as an assistant tool for radiologists during the reading process. Source of Funding: This project is based on a Cooperative Research and Development Agreement (CRADA) between the National Institutes of Health and the NVIDIA Corporation. This project has been funded in whole or in part with federal funds from the National Cancer Institute, National Institutes of Health, under Contract No. HHSN261200800001E. The content of this publication does not necessarily reflect the views or policies of the Department of Health and Human Services, nor does mention of trade names, commercial products, or organizations imply endorsement by the U.S. Government © 2021 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 206Issue Supplement 3September 2021Page: e1006-e1007 Advertisement Copyright & Permissions© 2021 by American Urological Association Education and Research, Inc.MetricsAuthor Information Sherif Mehralivand More articles by this author Dong Yang More articles by this author Daguang Xu More articles by this author Ziyue Xu More articles by this author Holger Roth More articles by this author Stephanie Harmon More articles by this author Samira Masoudi More articles by this author Thomas Sanford More articles by this author Deepak Kesani More articles by this author Nathan Lay More articles by this author Bradford Wood More articles by this author Peter Choyke More articles by this author Peter Pinto More articles by this author Baris Turkbey More articles by this author Expand All Advertisement PDF downloadLoading ...