Self-supervised Learning in Histopathology: New Perspectives for Prostate Cancer Grading

The prostate carcinoma (PCa) is the second most common cause of cancer-deaths among men. To estimate the appropriate therapy pathway after diagnosis, the Gleason score (GS) has been established as an international measure. While the GS has been proven to be a good tool for tumour assessment, it naturally suffers from subjectivity. Especially for cancers of lower to medium severity, this leads to inter- and intra observer variability and a remarkable amount of over- and under therapy. The PCa thus is in the focus of various research works, that aim to improve the grading procedure. With recently emerging AI technologies, solutions have been proposed to automate the GS-based PCa-grading while keeping predictions consistent. Current solutions, however, fail to handle data variability arising from preparation differences among hospitals and typically require a large amount of annotated data, which is often not available. Thus, in this paper, we propose self-supervised learning (SSL) as a new perspective for AI-based PCa grading. Using several thousand PCa cases, we demonstrate that SSL may be a feasible alternative for analysing histopathological samples and pretraining grading models. Our SSL-pretrained models extract features related to the Gleason grades (GGs), and achieve competitive accuracy for PCa downstream classification.