Diagnostic accuracy of deep learning for microsatellite instability prediction in gastric cancer: A systematic review and meta-analysis

A meta-analysis posted on medRxiv, from researchers at the Affiliated Hospital of Qinghai University, pooled ten studies of deep learning applied to H&E-stained pathology images to predict microsatellite instability in gastric cancer. Pooled sensitivity was 0.78, specificity 0.85 and summary ROC AUC 0.89; hospital-derived cohorts outperformed TCGA-derived ones (sensitivity 0.86 vs 0.71; diagnostic odds ratio 39.8 vs 11.5). The authors call for standardised multicentre, independent external and prospective studies before clinical use.