Prediction of microsatellite instability in colorectal cancer with deep learning

In this study, a deep learning-based model was proposed for the prediction of MSI in colorectal cancer using tissue slides directly stained with hematoxylin and eosin (H&E). With this proposed model, a classification model was created using deep learning structure, convolutional neural networks (CNN), and transfer learning methods, and the models were evaluated with comparative analysis. The image data used in the model were obtained from the Kaggle website open access data. In the study, MSI and MSS (microsatellite stable) classification was performed using 150,000 unique image patches of the colorectal cancer H&E stained histological image dataset for 80% training and 20% testing. Considering the performance results obtained, it was seen that the VGG19 model provided the highest classification performance among nine different pre-trained models (VGG19, MobileNet, ResNet50…), with an accuracy of 90.6%, precision of 88.6%, the sensitivity of 93.1% and AUC 90.6%.