Image-based artificial intelligence algorithms for molecular diagnostics in prostate pathology: How far are we? A systematic review.

Abstract

193 Background: Guidelines recommend molecular diagnostics for homologous and mismatch repair (MMR) defects in men with metastatic castration-resistant prostate cancer. However, in clinical practice, routine testing is hindered because it is costly, time-consuming and not available in all laboratories. Moreover, many patients need to be tested to identify the few with these genetic aberrations. To address these challenges, image-based artificial intelligence (AI) algorithms have been developed to predict molecular alterations from hematoxylin and eosin (HE) slides that are beyond the pathologist's visual detection. This review aims to assess the advancements of image-based AI algorithms in molecular diagnostics in prostate pathology and their potential in daily clinical practice. Methods: Through a search query, 3213 articles were identified and screened by two reviewers. In total, 12 articles were selected for this review and were assessed using the QUADAS-2 criteria. Results: Out of 12 articles, 5 focused exclusively on a specific set of preselected genes in prostate cancer, while the other articles examined broader ranges of genomic or transcriptomic changes, with some covering various cancer types. Training datasets consisted of up to 451 prostate cancer slides and most studies included only slides from radical prostatectomies. The majority made use of the cancer genome atlas program (TCGA). The best performance reported was 0.91 area under the curve (AUC) on internal and AUC of 0.89 on external validation. However, only 5 of the 12 studies performed external validation and none looked into the application of their algorithm in a clinical setting. Conclusions: Our review shows that in concept, AI is reasonably capable of detecting molecular changes based on HE slides of prostate pathology with reasonable predictive power. This has been shown in different specific genetic alterations, as well as in genome- or transcriptome-wide alterations. The predictive power of AI may however not be good enough to forgo molecular testing in patients predicted negative. Furthermore, these systems have only reached the developmental stage, leaving their potential performance in a clinical setting yet to be determined.