OMICs AND AI APPROACHES FOR MUSCLE DISEASES: O.08 Deep convolutional-neural-network can differentiate twelve major muscular diseases better than human

Abstract

Muscle pathology plays an important role in the diagnosis of muscle disease even in the molecular era, as most muscle diseases have been classified or even partly defined by pathological features. Nevertheless, availability of experts is limited especially in underserved areas of the world. This study aims to establish the algorithm to assist muscle pathology diagnosis using deep convolutional neural networks (CNNs), which can differentiate major muscle diseases only by hematoxylin and eosin (H&E)-stained pathological images. Target diseases were dystrophinopathy (DYST), limb-girdle muscular dystrophy (LGMD) 2A (LG2A), LGMD2B (LG2B), Ullrich congenital muscular dystrophy (CMD) (UCMD), Fukuyama CMD (FCMD), congenital myopathy (CM), GNE myopathy (GNEM), dermatomyositis (DM), inclusion body myositis (IBM), immune-mediated necrotizing myopathy (IMNM), antisynthetase syndrome (ASS), and neuropathic disease (NP). All images to train CNNs were taken from slides with CCD cameras via light microscopes. A total of 4041 images were obtained from 1400 slides. The algorithm consists of two steps: 1) to differentiate between idiopathic inflammatory myopathy (IIM) and other conditions, and 2) to subclassify each category. For the first step, images of DM, IBM, IMNM, and ASS were combined to create the IIM group. Meanwhile, those of DYST, FCMD, LG2A, LG2B, UCMD, CM, GNEM, and NP were combined to create another group. At the second step, four subtypes of IIM were analyzed. The first step achieved 0.996 area under the curve (AUC). The accuracy of the CNN and the best pathologist of all were 96.9% and 93.8%, respectively, indicating the CNN outperformed muscle specialists. In the second step, the CNNs classified four subtypes (DM, IBM, IMNM, and ASS) with high AUCs: 0.953 (IMNM), 0.969 (IBM), 0.965 (DM), and 0.944 (ASS). The algorithm used only a few H&E images obtained by CCD, which does not need a huge financial investment. This system should be helpful in clinical settings especially in underserved regions, where experts are not easily available.