MP35-02 COMPUTER-EXTRACTED FEATURES OF NUCLEAR AND GLANDULAR MORPHOLOGY FROM DIGITAL H&E TISSUE IMAGES PREDICT PROSTATE CANCER BIOCHEMICAL RECURRENCE AND METASTASIS FOLLOWING RADICAL PROSTATECTOMY

Abstract

You have accessJournal of UrologyProstate Cancer: Markers I1 Apr 2018MP35-02 COMPUTER-EXTRACTED FEATURES OF NUCLEAR AND GLANDULAR MORPHOLOGY FROM DIGITAL H&E TISSUE IMAGES PREDICT PROSTATE CANCER BIOCHEMICAL RECURRENCE AND METASTASIS FOLLOWING RADICAL PROSTATECTOMY Patrick Leo, Anna Gawlik, Guangjing Zhu, Michael Feldman, Sanjay Gupta, Robert Veltri, and Anant Madabhushi Patrick LeoPatrick Leo More articles by this author , Anna GawlikAnna Gawlik More articles by this author , Guangjing ZhuGuangjing Zhu More articles by this author , Michael FeldmanMichael Feldman More articles by this author , Sanjay GuptaSanjay Gupta More articles by this author , Robert VeltriRobert Veltri More articles by this author , and Anant MadabhushiAnant Madabhushi More articles by this author View All Author Informationhttps://doi.org/10.1016/j.juro.2018.02.1115AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookTwitterLinked InEmail INTRODUCTION AND OBJECTIVES Radical prostatectomy (RP), resection of the entire prostate, is the most common surgical technique for prostate cancer. There is a risk of biochemical recurrence (BCR) and metastasis following RP, both of which raise mortality and affect treatment plans. Patients likely to experience BCR and metastasis are candidates for aggressive therapy. We present a computerized approach to predict BCR and metastasis using hematoxylin and eosin stained tissue microarrays (TMAs). METHODS Seven TMAs were collected, containing 0.6mm diameter tumor region samples from N=260 RP specimens with 5-year follow-up data (210 BCR, 50 non-BCR, 132 metastasis, 128 non-metastasis). BCR models were trained on half the patients of three TMAs, for a 24 non-BCR, 30 BCR training set, and 180 BCR, 26 non-BCR testing set. Metastasis models used training and testing sets of 66 metastasis, 64 non-metastasis spots each. On each spot, nuclei and lumens were automatically segmented and 216 features of architecture, shape, and disorder were extracted. 26 Haralick texture features were extracted from the entire spot. We trained random forest models to predict BCR and metastasis using nuclei and texture (CNBCR, CNMet), lumen (CLBCR, CLMet), and all features (CLNBCR, CLNMet). Models used the top 10 features identified on the training set by minimum redundancy maximum relevance (mRMR) selection. We compared these models to Gleason score (CGBCR, CGMet), and Kattan (CKBCR, KMet) and Stephenson (CSBCR, CSMet) nomograms. RESULTS Table 1 shows the top 3 features for each model. Table 2 shows the area under the receiver operating characteristic curve (AUC) of each method. The lumen-nuclei-texture feature model performed best. CONCLUSIONS Our methods predicted BCR and metastasis better than gold standard methods. This could potentially help identify patients who would receive added benefit from adjuvant therapy. © 2018FiguresReferencesRelatedDetails Volume 199Issue 4SApril 2018Page: e446-e447 Advertisement Copyright & Permissions© 2018MetricsAuthor Information Patrick Leo More articles by this author Anna Gawlik More articles by this author Guangjing Zhu More articles by this author Michael Feldman More articles by this author Sanjay Gupta More articles by this author Robert Veltri More articles by this author Anant Madabhushi More articles by this author Expand All Advertisement Advertisement PDF downloadLoading ...