Computational analysis of pathological image enables interpretable prediction for microsatellite instability

Objective Microsatellite instability (MSI) is associated with several tumor types and its status has become increasingly vital in guiding patient treatment decisions. However, in clinical practice, distinguishing MSI from its counterpart is challenging since the diagnosis of MSI requires additional genetic or immunohistochemical tests. In this study, we aimed to establishe an interpretable pathological image analysis strategies to help medical experts to identify MSI automatically. Design Three cohorts of Haematoxylin and eosin-stained whole-slide images from 1033 patients with different tumor types were collected from The Cancer Genome Atlas. These images were preprocessed and tessallated into small tiles. A image-level interpretable deep learning model and a feature-level interpretable random forest model were built up on these files. Results Both models performed well in the three datasets and achieved image-level and feature-level interpretability repectively. Importantly, both from the image-level and feature-level interpretability, color features and texture characteristics are shown to contribute the most to the MSI prediction. Based on them, we established an interpretable classification framework. Therefore, the classification models under the proposed framework can serve as an efficient tool for predicting the MSI status of patients. Conclusion This study establishes a interpretable classification framework to for predicting the MSI status of patients and provide more insights to pathologists with clinical understanding.