Abstract: PD71-09
Sources of Funding: DOD PC131996, PCF-Movember GAP1 Unique TMAs Project, Prostate Cancer Foundation (PCF) Creativity Award, Jean Perkins Foundation, NIH/NCI P01 CA098912-09, NIH R01CA131255 and P50CA092131, Stephen Spielberg Team Science Award.

Introduction

Prostate cancer (PC) with de novo bone metastases (M1) has a 5-year survival of 28%. Pathological features of primary M1 tumors are generally indistinguishable from those of high-grade localized (M0) cases, however 5-year survival for M0 PC is nearly 100%. Digital image analysis is an evolving &[Prime]OMICS&[Prime] platform for biomarker development that can be applied to diagnostic histopathology. We hypothesize that novel software analysis tools and machine learning can systematically interrogate digitized prostate needle biopsy (PNBX) slides to extract histomorphometric features that identify discrepant architecture and nuclear texture of M0 and M1 tumors. Herein, algorithms that measure these features were developed in a training set of digital images and then validated in an independent patient cohort._x000D_ _x000D_

Methods

We created a biorepository of diagnostic PNBX specimens from 2150 PC patients from the Greater Los Angeles VA Healthcare System between 2000 and 2016. The biorepository was mined to create a matched cohort of M0 (n=44) and M1 (n=61) cases. Slides were digitized at 40X magnification and two pathologists annotated all cancer foci. ~30 image tiles were obtained from each case (n=2857) and 88 features were extracted. Segmentation based fractal texture descriptors (SFTA), Gabor (GF), grey level run length (GLRL), and nuclear texture (CP) features were used to train a classifier to distinguish M0 from M1 tiles.

Results

After conversion of M0 and M1 image tiles to digital nuclear masks, training features were used to classify nuclear texture or tissue architecture. The majority vote from nuclear classification was transferred to the tile level and the majority classification of tiles was used to classify each case. For tissue architecture, 45 STFA and 60 Gabor features classified M1 and M0 cases with an accuracy of 71.8% and 80%, respectively. For nuclear features, 44 GLRL and 8 CP classified M0 versus M1 cases with an accuracy of 63% and 75.4%, respectively. A classifier trained with a combined 88 features achieved 86% accuracy in distinguishing M1 from M0 cases.

Conclusions

We applied digital imaging technology and machine learning to extract 88 novel features that accurately differentiate high-grade M0 from M1 PC. The quantification of tissue architecture and nuclear morphology provides an orthogonal approach for biomarker development, which can be applied to prognostication and potential treatment decisions in patients with high-risk localized or metastatic PC._x000D_

Funding

DOD PC131996, PCF-Movember GAP1 Unique TMAs Project, Prostate Cancer Foundation (PCF) Creativity Award, Jean Perkins Foundation, NIH/NCI P01 CA098912-09, NIH R01CA131255 and P50CA092131, Stephen Spielberg Team Science Award.