NEXT GENERATION SEQUENCING AND EXPERIMENTAL MYOLOGY: P.385Physician-Level muscle disease classifier for computer-aided diagnostics with neural networks

Abstract

Even in the current era of molecular biology, muscle pathology still plays a central role in diagnosing muscle disease. Among these diseases, muscular dystrophy (MD) is a group of progressive hereditary muscle diseases for which no curative therapy is available, and immune-mediated necrotizing myopathy (IMNM) is a subclass of myositis that can be treated by immune therapy. Whereas MD and IMNM are both pathologically characterized by the presence of necrotic and regenerating fibers and often look similar. Therefore, the differentiation of these conditions is a hot topic in myology. To develop and evaluate an algorithm for computer-aided muscle histopathological diagnosis that uses deep learning. We chose two disease categories, MD and IMNM, and aimed to distinguish between these conditions. We prepared 11-layer convolutional neural networks and 1977 (MD: 885, IMNM: 1092) H&E-stained muscle section images classified in real diagnosis. Of these, 190 (MD: 78, IMNM: 112) images were used as a test set and the others (MD: 807, IMNM: 980) for training. After training our model with the training dataset, we evaluated its classification performance. For comparison, seven physicians also classified the same test set. After training of the model, the AUC, the area under the ROC (Receiver operating characteristic) curve, of our model was 0.94. In terms of accuracy, the classifier achieved better results, 88% on average, than seven physicians specializing in muscle disease among whom the best accuracy was 78%, indicating that our model outperforms physicians at classifying the two muscle diseases on H&E images. There are a number of limitations in this study. For example, we used only H&E images and the classification (diagnosis) was made between only two disease conditions. Nevertheless, our results suggest the deep neural network model can make pathological diagnosis better than physicians at least under certain conditions.