391 How accurate is the PSA test? A prevalence study of disturbed PSA values in a tertiary referral hospital

Abstract

INTRODUCTION AND OBJECTIVES: Prostate cancers (PC) associated with overt metastases at diagnosis (stage M1) or that progress to this stage are lethal. Although pathologic features of most M1 cases are indistinguishable from high-grade, non-metastatic PC (stage M0), profound disparities in outcome are apparent. More than 70% of M1 patients progress to castration-resistant prostate cancer (CRPC) and death within 5 years compared to 100% survival of highgrade M0 cases. Improved tools that detect features of lethal high-grade disease in prostate needle biopsies (PNBXs) will enable early identification of patients at greatest risk of metastatic progression. We have applied novel quantitative imaging (QI) technology and validated machine learning software tools to extract histomorphometric features from PNBX specimens that distinguish M1 and M0 cancers. This 00computer vision00 approach can be applied to routine diagnostic biopsies with the goal of detecting lethal PC as early as possible and improving patient stratification for entry into clinical trials. METHODS: We created a diverse, annotated biorepository with detailed clinical information corresponding to PNBX tissue from 427 patients with high-grade M0 or M1 PC. Retrospective analysis was used to assemble cohorts matched for age, race, and Gleason grade that were either M0 (n1⁄48) or M1 (n1⁄410). Archival PNBX slides stained with hematoxylin and eosin were quality checked and digitally scanned at 40X magnification. Fifteen image tiles were collected from each slide and two classifiers were constructed to assign each image tile to the M0 or M1 groups. QI analysis was used to investigate a set of features, including Fractal Dimensions (FD), Lacunarity (LA) [1st classifier], and Nuclear Features (NF) [2nd classifier] in M0 and M1 PC. RESULTS: Approximately 332 image tiles from the M0-RRP (n1⁄4122) and M1-ADT (n1⁄4210) cases were converted to nuclear masks. Using LA and FD tests, a trained classifier achieved 70% accuracy in distinguishing images from M1 and M0 cases, in which LA features were significantly different. We also identified 63 NFs differentially expressed in the M0 and M1 cases after random selection of 30% of nuclei and calculation of 1,500 NFs using partial least squares method. With this method, a classifier trained with the selected NFs to characterize nuclear intensity, texture, size, and shape achieved an accuracy of 75.5% in M0-RRP cases and 74.8% in M1-ADT cases. CONCLUSIONS: This innovative cohort of patients at the extremes of high-grade PC provides the opportunity to generate and test the performance of novel QI algorithms.