You have accessJournal of UrologyProstate Cancer: Localized: Surgical Therapy I (MP19)1 Apr 2020MP19-20 STATE-OF-THE-ART WEAKLY SUPERVISED AUTOMATED CLASSIFICATION OF PROSTATE CANCER TISSUE MICROARRAYS VIA DEEP LEARNING: CAN SUFFICIENT ACCURACY BE ACHIEVED WITHOUT MANUAL PATCH LEVEL ANNOTATION? Sami-Ramzi Leyh-Bannurah*, Ulrich Wolffgang, Jonathan Schmitz, Veronique Ouellet, Feryel Azzi, Zhe Tian, Burkhard Helmke, Markus Graefen, Lars Budäus, Pierre I. Karakiewicz, Dominique Trudel, and Fred Saad Sami-Ramzi Leyh-Bannurah*Sami-Ramzi Leyh-Bannurah* More articles by this author , Ulrich WolffgangUlrich Wolffgang More articles by this author , Jonathan SchmitzJonathan Schmitz More articles by this author , Veronique OuelletVeronique Ouellet More articles by this author , Feryel AzziFeryel Azzi More articles by this author , Zhe TianZhe Tian More articles by this author , Burkhard HelmkeBurkhard Helmke More articles by this author , Markus GraefenMarkus Graefen More articles by this author , Lars BudäusLars Budäus More articles by this author , Pierre I. KarakiewiczPierre I. Karakiewicz More articles by this author , Dominique TrudelDominique Trudel More articles by this author , and Fred SaadFred Saad More articles by this author View All Author Informationhttps://doi.org/10.1097/JU.0000000000000852.020AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: A huge barrier to implement computational pathology grading constitutes great effort of manual patch annotation within digitized tissue images for training and validation purpose of state-of-the-art machine learning approaches. We hypothesized that comprehensive data quantity and rigorous multi-expert review enables automated Gleason grading of radical prostatectomy tissue microarray sports without such time-consuming and error-prone manual patch-level-annotation. METHODS: Therefore, we relied on double-reviewed 9.557 40x magnified and digitized H&E stained TMA spots from 1,512 prostate cancer individuals, who underwent radical prostatectomy in a Canadian multi-center cohort. Benign vs. prostate cancer tissue was present in 49.2 vs. 50.8%. Data was randomly sampled into 3:1 training:validation sets. For the task at hand, a custom CNN model was trained and compared to transfer learning based in pretrained existing CNN architectures (Xception, VGG, Inception, Densenet) RESULTS: We achieved a maximum model accuracy for classification into benign vs. tumor of 85.9% based on a custom convolutional neural network (CNN) architecture. Utilizing “Xception” as an example of a pre-trained CNN architecture for transfer-learning, yielded considerably lower accuracy of 67% CONCLUSIONS: Our findings showed promising accuracy. To our knowledge, our study is one of very few performed without manual patch annotation, with focus on per-patient analyses. In turn, our approach might be more easily implemented in pathology clinical workflows. As with all deep learning approaches, thorough validation is essential, with particular focus to dissolve classification discrepancies. Interestingly, less complex deep learning models with custom architecture performed better that pre-trained architectures. For clinical implantation, this means, that future implementation might need less computational resources for the task at hand. Source of Funding: None © 2020 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 203Issue Supplement 4April 2020Page: e306-e306 Advertisement Copyright & Permissions© 2020 by American Urological Association Education and Research, Inc.MetricsAuthor Information Sami-Ramzi Leyh-Bannurah* More articles by this author Ulrich Wolffgang More articles by this author Jonathan Schmitz More articles by this author Veronique Ouellet More articles by this author Feryel Azzi More articles by this author Zhe Tian More articles by this author Burkhard Helmke More articles by this author Markus Graefen More articles by this author Lars Budäus More articles by this author Pierre I. Karakiewicz More articles by this author Dominique Trudel More articles by this author Fred Saad More articles by this author Expand All Advertisement PDF downloadLoading ...