Microsatellite instability (MSI) is a variation caused by the damage of deoxyribonucleic acid (DNA) mismatch repair. The MSI status can be used as an important biomarker to evaluate whether cancer patients can adapt to immunotherapy. Previously, the gold-standard method for MSI determination was a PCR test or immunohistochemical detection, but in recent years, research has proved that deep learning can detect MSI in tumor samples on routine histology slides faster and more cheaply than molecular assays. In this study, we used a number of self-supervised methods including ResNest, Transformer and other deep network structures on the Cancer Genome Atlas (TCGA) pathological image datasets to detect MSI status. The test results show that all methods received an area under the curve (AUC) of over 0.90. These results go beyond the performance of a single classical network structure on the datasets, which not only proves the superiority of the self-supervision training method in such research, but also that the relevant experimental performance of the transformer structure is more accurate than weight random initialization and the ImageNet migration model. Overall, our proposed self-supervised methods make up for the prior knowledge gap between natural image and medical image pre-training knowledge.

Microsatellite instability;Self-supervised;Deep learning;Pathological image