MP72-16 DETECTION AND GRADING OF BLADDER CANCER ON HISTOLOGICAL SLIDES USING DEEP LEARING

Résumé fourni par la source

You have accessJournal of UrologyBladder Cancer: Non-invasive III (MP72)1 Apr 2020MP72-16 DETECTION AND GRADING OF BLADDER CANCER ON HISTOLOGICAL SLIDES USING DEEP LEARING Ilaria Jansen*, Marit Lucas, Judith Bosschieter, Sybren L. Meijer, Ton G. van Leeuwen, Henk A. Marquering, Jakko A. Nieuwenhuijzen, Daniel M. de Bruin, and Cemile D. Savci-Heijink Ilaria Jansen*Ilaria Jansen* More articles by this author , Marit LucasMarit Lucas More articles by this author , Judith BosschieterJudith Bosschieter More articles by this author , Sybren L. MeijerSybren L. Meijer More articles by this author , Ton G. van LeeuwenTon G. van Leeuwen More articles by this author , Henk A. MarqueringHenk A. Marquering More articles by this author , Jakko A. NieuwenhuijzenJakko A. Nieuwenhuijzen More articles by this author , Daniel M. de BruinDaniel M. de Bruin More articles by this author , and Cemile D. Savci-HeijinkCemile D. Savci-Heijink More articles by this author View All Author Informationhttps://doi.org/10.1097/JU.0000000000000952.016AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: Histological grade is an important predictor for recurrence and progression in non-muscle invasive bladder cancer (NMIBC). However, it is subjected to low interobserver agreement. Therefore, the aim of this study is to automatically detect and grade urothelial cell carcinoma (UCC) of the bladder on histological slides using deep learning. METHODS: IRB gave approval. Histological glass slides of patients with NMIBC who underwent a transurethral resection of bladder tumor (TURBT) between 2000-2016 in three hospitals in the Netherlands were included. Three pathologists independently reviewed all glass slides, assigning the WHO’04 grade. After digitizing, an expert observer manually annotated all the slides with supervision of a uropathologist. At first, a deep learning network was trained for the detection of urothelium. The second network used the segmentations of the first network to grade the UCC according to the WHO’04 grading system into low-grade and high-grade. The consensus score of the three pathologist was used as a reference. The outcome of the automated grading was compared with the consensus and the individual grading of the three pathologists. RESULTS: Altogether we included 328 tissue samples of 232 patients. The urothelium was accurately detected in 93% of the slides. False positive or false negative regions were detected in 13% of cases. Inaccurate classification was principally in regions with inflammation or in slides with extensive color loss. Surprisingly, more urothelium was detected by the network than initially delineated by the observer. This was seen in areas with non-diagnostic urothelium due to mechanical artefacts (Figure 1). Automated grading was done correctly in 76% of the low-grade and 71% of the high grade according to the consensus reading. Moderate (κ=0.48 ± 0.14 se) agreement was seen when between the automated grading and the consensus. Agreement between pathologists was fair (κ=0.35 ± 0.13 se, κ=0.38 ± 0.11 se) and moderate (κ=0.52 ± 0.13 se). CONCLUSIONS: The automated grading of UCC of the bladder using deep learning is as good as the interobserver agreement of pathologist. To overcome the inconsistency in grading of bladder cancer, we aim at using patient outcome to train a deep learning network in the future. Source of Funding: This study has been funded by ITEA3. Grant number: ITEA151003. © 2020 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 203Issue Supplement 4April 2020Page: e1080-e1081 Advertisement Copyright & Permissions© 2020 by American Urological Association Education and Research, Inc.MetricsAuthor Information Ilaria Jansen* More articles by this author Marit Lucas More articles by this author Judith Bosschieter More articles by this author Sybren L. Meijer More articles by this author Ton G. van Leeuwen More articles by this author Henk A. Marquering More articles by this author Jakko A. Nieuwenhuijzen More articles by this author Daniel M. de Bruin More articles by this author Cemile D. Savci-Heijink More articles by this author Expand All Advertisement PDF downloadLoading ...

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.