Deep learning for identification of fasciculation from muscle ultrasound images. Issue 5 (29th May 2019)

• Record Type: Journal Article
• Authors: Nodera, Hiroyuki; Takamatsu, Naoko; Yamazaki, Hiroki; Satomi, Ryutaro; Osaki, Yusuke; Mori, Atsuko; Izumi, Yuishin; Kaji, Ryuji
• Is Part Of: Neurology and clinical neuroscience. Volume 7:Issue 5(2019:Sep.)
• Page Start: 267
• Page End: 275
• Publication Date: 2019-05-29
• Subjects: amyotrophic lateral sclerosis -- artificial neural network -- background subtraction -- deep learning -- fasciculation; Neurology -- Periodicals; Neurosciences -- Periodicals; 616.8
• DOI: 10.1111/ncn3.12307
• Languages: English
• ISSNs: 2049-4173

Abstract

Abstract: Background: Detection of fasciculation plays a significant role in the early diagnosis of amyotrophic lateral sclerosis (ALS). Although ultrasound (US) has been reported to be superior to needle electromyography (EMG) and visual inspection in terms of detection rates, other similar movements could lower the reliability of the result. Artificial intelligence, such as deep learning, promises to enhance the classification of visual data. Aim: To classify fasciculation and its mimics by deep learning. Methods: Using an 11‐MHz linear‐array transducer, muscle US was recorded from the biceps brachii and tibialis anterior muscles. Fasciculation for patients with ALS, and movements of voluntary muscles or the recording probe of healthy individuals were recorded. Background subtraction was performed to obtain binary images to infer movements. Deep learning was performed using five different networks with and without pre‐trained weights. Results: Three groups of images were divided into training, validation, and test (fasciculation: N = 1473; voluntary movement: N = 861; probe movement: N = 1626). The accuracy of detection with pre‐trained weights (fine‐tuning) ranged from 0.959 to 1.0. The best accuracy was obtained by VGG16/19 convolutional neural networks and the ResNet‐152 network. Accuracy of prediction was considerably lower without the pre‐trained weights. The mean white pixels (inferring movements) were lower in fasciculation than in the voluntary and probe movement groups; however, the non‐fasciculation groups showed similar pixel counts ( P = 0.95). Conclusions: US accurately distinguished between fasciculation and its mimics by deep learning. Pixel counting could be a reliable quantitative method to detect fasciculation.