Deep learning captures selective features for discrimination of microsatellite instability from pathologic tissue slides of gastric cancer

Microsatellite instability (MSI) status is an important prognostic marker for various cancers. Furthermore, because immune checkpoint inhibitors are much more effective in tumors with high level of MSI (MSI‐H), MSI status is routinely tested in multiple cancer types. Therefore, many studies have tested the feasibility of deep learning (DL)‐based prediction of MSI status from hematoxylin and eosin (H&E)‐stained tissue slides. In the present study, we attempted a fully automated classification of MSI status in gastric cancer (GC) tissue slides. For frozen and formalin‐fixed paraffin‐embedded (FFPE) GC tissues from The Cancer Genome Atlas (TCGA), the areas under the curves (AUCs) for the receiver operating characteristic (ROC) curves were 0.893 and 0.902, respectively. The classifier trained with the TCGA FFPE tissues performed well on an external validation Asian FFPE cohort, with an AUC of 0.874. However, the DL‐based classifier seems incompatible with cancers from different organs because morphologic features of MSI‐H tissues are different. Analysis of histomorphologic features of MSI‐H GC tissues suggested that MSI‐H GC could largely be divided into two groups: intestinal type tumors with moderate to poor differentiation and diffuse type mucinous tumors. However, the recognizable morphologic features cannot completely explain the good performance of the DL‐based classifier. These results indicate that DL could automatically learn the optimal features for discrimination of MSI status in GC tissue slides. This study demonstrated the potential of a DL‐based MSI classifier as a screening tool for definitive cases. What's new? Microsatellite instability (MSI) status is an important prognostic and treatment response indicator for various cancers. By applying a fully automated process for the prediction of MSI status from tissue slides, here the authors achieved state‐of‐the‐art performance in both training and external validation cohorts for gastric cancer. The deep learning model could also reveal morphological features of MSI‐high tissues. This study shows the potential of automated deep learning for complementary MSI testing in gastric cancer using tissue slides, which could decrease the costs and save time during clinical assessment of gastric cancer patients.