Presentation Authors: Nathan Ing, Fangjin Huang, Eric Miller, Arash Amighi*, Michael Lewis, Isla Garraway, Arkadiusz Gertych, Beatrice Knudsen, Los Angeles, CA
Introduction: Conventional histological analysis cannot differentiate diagnostic prostate needle biopsies (PNBX) from patients with localized disease (stage M0) versus those with de novo metastases (stage M1). Digital image analysis (DIA) is a new precision medicine platform that shows promise in advancing diagnostic histopathology. Our team has developed and deployed a novel DIA risk of metastasis (RoM) classifier that systematically interrogates standard-of-care PNBX slides to determine M-stage and likelihood of clinical metastatic progression.
Methods: We collected diagnostic PNBX from 163 prostate cancer cases (85 M0 and 78 M1) that contained high grade cancer foci. Pathological annotation was performed by two independent pathologists. Following slide digitization at 40X, image tiles (approximately 30 image tiles per case, n = 6501) were used to extract 62 handcrafted (HC) and 64 autoencoder (AE) nuclear features, as well as 1234 tissue-based (TI) features. A Z-score normalization of features was performed. The nuclear and tissue architecture features were combined and used in a Leave One Out cross validation with 3 models: elastic net, gradient boosted decision trees, and a multi-layer perceptron. RoM scores were generated for each tile, and cases were classified into M0, M1, or "uncertain" based on the variance of RoM scores. Clinical progression was predicted using the model.
Results: After color normalization, 273,103 total nuclei and 6501 tiles were analyzed and values extracted for HC, AE, and TI features. Of the three models tested, the E-net model performed the best with an AUC of 0.885 from a cross-validation experiment. Of the 163 cases, 80% were assigned to M0 or M1 status, while 20% were placed into the "uncertain" category. Fifteen from 78 M1 cases were misclassified as M0 cases, while the other 63 cases were correctly classified. The classifier predicted progression to metastatic disease with an area under the curve (AUC) of 0.75. A second validation cohort of 59 more cases is underway.
Conclusions: DIA technology and machine learning software enabled extraction of 1360 features that differentiate high-grade M0 from M1 prostate cancer. Initial results show promise in using this classifier to prognosticate progression based solely on diagnostic PNBX. DIA technology has the potential to serve as a cost-effective clinical tool for confirming M-stage and predicting metastatic progression. When combined with additional clinical variables, the accuracy of the prediction can be further improved.
Source of Funding: PCF Challenge Award, the STOP Cancer Foundation, AUA Summer Medical Student Reserach Fellowship: Herbert Brendler, MD Research Fund, DOD PC131996, PCF-Movember GAP1 Unique TMAs Project, Prostate Cancer Foundation (PCF) Creativity Award, Jean Perkins Found