AI-based prediction of molecular aberrations in prostate cancer using digital pathology: a systematic review

Abstract

Molecular diagnostics for homologous repair and mismatch repair deficiencies are valuable in metastatic prostate cancer, as they are targetable by poly ADP-ribose polymerase or immune checkpoint inhibition. Molecular diagnostics are rarely used in prostate cancer as they are complex and expensive, and the incidence of the relevant molecular aberrations is low. To address these limitations, image-based artificial intelligence algorithms have been developed to predict molecular aberrations from hematoxylin and eosin slides beyond the pathologist's visual detection. This systematic review assesses the advancements of image-based artificial intelligence algorithms predicting molecular aberrations in prostate cancer pathology and their potential in clinical practice. After screening 4121 articles, 20 articles were identified and assessed using the QUADAS-2 criteria. Nine algorithms, focusing on specific molecular aberrations in prostate cancer, reached a mean area under the curve of 0.78 (range 0.67 - 0.91). When focusing on the thus far clinically relevant specific molecular aberrations, the AI algorithms predicting BRCA, homologous repair deficiency, and mismatch repair deficiency achieved an area under the curve of 0.79, 0.84, and 0.72 on internal validation. Due to the lack of molecularly tested image data, most studies (17/20) used The Cancer Genome Atlas, and only five studies performed external validation. Our review shows that image-based artificial intelligence algorithms could be a pre-screening molecular diagnostic tool, particularly with the recent shift toward clinically more relevant molecular aberrations. Nonetheless, the artificial intelligence algorithms remain in the development stage due to the limited availability of molecularly tested pathology image data needed for proper external validation.