MP43-03 MACHINE LEARNING IMPROVES DETECTION OF CLINICALLY SIGNIFICANT PROSTATE CANCER IN EQUIVOCAL LESIONS ON MPMRI

Abstract

You have accessJournal of UrologyProstate Cancer: Detection & Screening IV (MP43)1 Sep 2021MP43-03 MACHINE LEARNING IMPROVES DETECTION OF CLINICALLY SIGNIFICANT PROSTATE CANCER IN EQUIVOCAL LESIONS ON MPMRI Maxime D. Rappaport, Indrani Bhattacharya, Richard E. Fan, Mirabela Rusu, and Geoffrey A. Sonn Maxime D. RappaportMaxime D. Rappaport More articles by this author , Indrani BhattacharyaIndrani Bhattacharya More articles by this author , Richard E. FanRichard E. Fan More articles by this author , Mirabela RusuMirabela Rusu More articles by this author , and Geoffrey A. SonnGeoffrey A. Sonn More articles by this author View All Author Informationhttps://doi.org/10.1097/JU.0000000000002064.03AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: PIRADS version 2.1 is a standardized method for evaluating prostate multiparametric MRI (mpMRI). While a PIRADS score of 5 is highly specific for clinically significant prostate cancer (csPCa), false positives are very common in lesions scored PIRADS 3 or 4. These lesions present a dilemma to urologists seeking to maximize csPCa detection, while minimizing unnecessary biopsies. Radiomics, a method of extracting quantitative features from medical images, has the potential to differentiate equivocal lesions containing csPCa from those containing benign tissue. We aimed to develop a machine leaning (ML) algorithm that identifies csPCa within PIRADS 3 and 4 lesions to reduce unnecessary biopsies. METHODS: We identified 133 patients with PIRADS 3 or 4 lesions on mpMRI who underwent subsequent targeted prostate biopsy. Ground truth for presence of csPCa within a lesion was a targeted biopsy core with Gleason grade group ≥2. We created a data pipeline (Fig 1A) that involved processing mpMRI lesions, identifying relevant radiomic features, and training logistic regression models using 5-fold cross validation (CV). We selected ten radiomic features (Fig 1B) and two clinical values, age and PSA density, to train our models to detect csPCa. Each model was tested on an independent test set of 31 PIRADS 3 and 4 lesions and evaluated by the area under the curve (AUC) statistic (Fig 1C). Finally, the utility of using our classifier results to determine lesions requiring biopsy was evaluated by the proportion of false negative cases in our predictions. RESULTS: Our binary classifier models achieved a mean AUC of 0.86 ± 0.06 (mean ± standard deviation) and a csPCa detection accuracy of 75 ± 13%. The models had a sensitivity of 0.78 ± 0.11, specificity of 0.75 ± 0.15, positive predictive value of 0.58 ± 0.14, and negative predictive value of 0.89 ± 0.07. Using our classifiers to select lesions for targeted biopsy would have reduced the number of PIRADS 3 and 4 lesions requiring biopsy by 59 ± 9%, while missing csPCa in only 6.5 ± 3% of those lesions. CONCLUSIONS: We developed a binary classifier for csPCa detection in equivocal lesions (PIRADS 3 or 4) based on radiomic and clinical features. With further refinement and external validation, ML models have the potential to help urologists decide which lesions warrant targeted biopsy and which can be safely observed. Source of Funding: None © 2021 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 206Issue Supplement 3September 2021Page: e782-e782 Advertisement Copyright & Permissions© 2021 by American Urological Association Education and Research, Inc.MetricsAuthor Information Maxime D. Rappaport More articles by this author Indrani Bhattacharya More articles by this author Richard E. Fan More articles by this author Mirabela Rusu More articles by this author Geoffrey A. Sonn More articles by this author Expand All Advertisement Loading ...