Histopathology-based artificial intelligence algorithms for the prediction of prostate cancer metastasis after radical prostatectomy

Abstract

Background and objective Multimodal artificial intelligence (AI) algorithms have been validated to predict prostate cancer (PCa) metastasis using combined histopathology and clinical-pathologic parameters in clinical trial cohorts. Here, we used purely histopathology-based AI algorithms to predict the probability of lethal PCa in surgically treated population- or hospital-based cohorts, comparing with genomic classifiers and standard clinical risk tools. Methods This study included representative whole slide images (WSIs) of radical prostatectomy (RP) and needle biopsy samples, or tissue microarrays (TMAs) constructed from RP specimens across five surgically treated PCa cohorts. A concatenated feature-based classification system using histopathologic data from each image generated an AI risk score for metastasis. Key findings and limitations In Cox models for time to metastasis, an AI risk score from prostatectomy WSIs showed similar performance (C-index: 0.81–0.85) to the Decipher or Prolaris genomic classifiers (C-index: 0.72–0.80) in testing cohorts. A modified TMA AI score analyzing ∼1 mm2 of prostatectomy tumor tissue from a nationwide study of 1351 patients had a C-index of 0.71 (95% confidence interval [CI]: 0.67–0.75). In a needle biopsy cohort followed for metastasis after prostatectomy, the TMA AI score had a C-index of 0.74 (95% CI: 0.70–0.79), and models combined with the Cancer of the Prostate Risk Assessment (CAPRA) score showed improved performance (C-index: 0.83 [95% CI: 0.82–0.87]) compared with CAPRA alone (C-index: 0.79 [95% CI: 0.73–0.84]). Limitations include relatively older cohorts, with many biopsies performed before the adoption of modern magnetic resonance imaging–guided techniques. Conclusions and clinical implications This study is among the first to show that histopathology-based AI algorithms applied to small samples of tumor tissue can predict the risk of lethal PCa. These algorithms perform comparably to commonly used genomic classifiers, and their predictive performance is enhanced when combined with clinicopathologic variables.