Physician-Level Aggregated Classifier for Genetic Muscle Disorders.

Muscle histopathology is the one of the most important diagnostic methods in the diagnosis of muscle diseases [1]. Nevertheless, it is a highly specialized field and only a limited number of experts are available in the world. Not surprisingly, significant number of cases are undiagnosed in underserved areas. We therefore intended to establish a computer-aided diagnostic system which should be helpful in this domain. This study aims to develop a multi-class classifier with deep learning for computer-aided muscle histopathological diagnosis. We chose five genetic muscle disease categories (dystrophinopathy, limb-girdle muscular dystrophy 2A (LGMD2A), limb-girdle muscular dystrophy 2B (LGMD2B), Ullrich congenital muscular dystrophy (UCMD), and Fukuyama-type congenital muscular dystrophy (FCMD)) as targets and aimed to distinguish between these diseases. We developed a new classifier, which we call the aggregated classifier, that improves the classification accuracy of existing classifiers to deal with histopathological images. The classifier achieved better classification accuracy than not only the existing classifier but also eight physicians who have been trained for muscle histopathology for variable duration. This result suggests that causative genes may possibly be precisely predicted only by hematoxylin-eosin (H&E) stained images without genetic testing.