Prediction of Microsatellite Instability in Colorectal Cancer Based on Deep Learning

Abstract

In this paper, a method for detecting microsatellite instability (MSI) status from colorectal cancer biopsy samples by deep learning with a convolutional neural network (CNN) using a pathological Whole Slide Image (WSI) is proposed. The proposed model has a mechanism to add weights to feature maps obtained from images of multiple sizes and resolutions as inputs. The MSI detection for a specimen is performed by adding a process to expand the cancer cell area and a confidence measure in consideration for the characteristic of the biopsy samples. In addition, the model used was changed according to the size of the specimen. The method used in the previous study detected the MSI status on the patch with an accuracy of 76.0%, whereas the method used in this study achieved an accuracy of 88.3%.