A Prediction Model of Microsatellite Status from Histology Images

Machine learning approaches have received sufficient attention in tumor detection in histopathology, and the very recent researches show their potential in extraction of molecular information for biomarker prediction. However, as we can only obtain label information of the whole slide, the patch-wise prediction classification results are simply summarized to reach the final diagnosis in previous work. In this paper, we develop a novel framework to precisely predict biomarker from hematoxylin and eosin (H&E) stained histology slides, where microsatellite instability (MSI) status in colorectal cancer is used as a case study. We develop a patch-wise binary classifier to detect tumor tissue as biomarker is tightly associated with tumor tissues. To obtain a precise predication of MSI status, a noise-robust convolutional neural network is trained by relabeling patch-wise output iteratively and feed back as input information. We employ a distillation framework for the dataset relabeling task. We also design a mathematical algorithm to sort out representative patches towards the final MSI status prediction. The model is evaluated by a large patient cohort from The Cancer Genome Atlas (TCGA), and tested on the state-of-the-art deep learning device NVIDIA GPU TeslaTM V100. The experimental results demonstrate the improved reliability in MSI prediction from histology images.