Prediction of Microsatellite Instability From Gastric Histological Images Based on Residual Attention Networks With Non-Local Modules

Abstract

Gastric cancer can be classified into different subtypes according to their genetic expression. Microsatellite instability (MSI) is one of these subtypes and an important clinical marker for prognosis and consideration for immunotherapy. Since genetic testing is relatively expensive and laborious, this study tackles the challenge of using deep neural networks (DNNs) to identify MSI based on analyzing histomorphologic features of gastric whole-slide images (WSIs). A two-stage patch-wise framework was proposed, which first differentiates the tumor regions from normal, then predicts MSI status from the tumorous patches. The proposed deep learning architecture enhances the residual attention network with non-local modules and visual context fusion modules, thereby allowing both local fine-grained details and coarse long-range dependencies to be captured. Image post-processing procedures were also proposed to better align the region segmentation with pathologist annotations. The model was applied to a three-way classification task, namely normal tissue, microsatellite stable (MSS), and MSI, using a private dataset gathered by Chang Gung Memorial Hospital and achieved 91.95% slide-wise accuracy. We also studied the feasibility of transfer learning by fine tuning on the TCGA-STAD public dataset, where we attained a high accuracy of 96.53% and an AUC of 0.99, outperforming previous literature.

Keywords

• Image post-processing
• microsatellite instability
• non-local neural networks
• residual attention network
• whole slide image