Classification of Microsatellite Instability Status in Slide-level Annotated Colorectal Tumors by Weakly Supervised Deep Learning

Abstract

Microsatellite instability (MSI) is a typical pathogenesis of colorectal cancer, and its status is helpful to the diagnosis of related diseases such as Lynch syndrome. In this paper, we propose an adaptive clustering-constrained attention multiple instance learning model based on weak supervision, which can classify the MSI state of H&E- stained images slide-level labeled at a low cost, and achieve an average AUC of 0.91 and an average ACC of 0.83 on a mixed dataset. On the basis of related work, the model integrates the instance classifiers with the bag classifier, and optimizes the clustering algorithm in the MSI classification scene, reducing the complexity of the model, while improving the final accuracy.