Noda Yoshikatsu, Sekiguchi Kenji, Matoba Shun, Suehiro Hirotomo, Nishida Katsuya, Matsumoto Riki
Division of Neurology, Kobe University Graduate School of Medicine, 7-5-1 Kusunoki-cho, Chuo-ku, Kobe, 650-0017, Japan.
Department of Neurology, National Hospital Organization Hyogo Chuo National Hospital, 1314 Ohara, Sanda 669-1592, Japan.
Clin Neurophysiol Pract. 2024 Aug 19;9:242-248. doi: 10.1016/j.cnp.2024.08.003. eCollection 2024.
Many artificial intelligence approaches to muscle ultrasound image analysis have not been implemented on usable devices in clinical neuromuscular medicine practice, owing to high computational demands and lack of standardised testing protocols. This study evaluated the feasibility of using real-time texture analysis to differentiate between various pathological conditions.
We analysed 17,021 cross-sectional ultrasound images of the biceps brachii of 75 participants, including 25 each with neurogenic disorders, myogenic disorders, and healthy controls. The size and location of the regions of interest were randomly selected to minimise bias. A random forest classifier utilising texture features such as Dissimilarity and Homogeneity was developed and deployed on a mobile PC, enabling real-time analysis.
The classifier distinguished patients with an accuracy of 81 %. Echogenicity and Contrast from the Co-Occurrence Matrix were significant predictive features. Validation on 15 patients achieved accuracies of 78 %/93 % per image/patient over 15-second videos, respectively. The use of a mobile PC facilitated real-time estimation of the underlying pathology during ultrasound examination, without influencing procedures.
Real-time automatic texture analysis is feasible as an adjunct for the diagnosis of neuromuscular disorders.
Artificial intelligence using texture analysis with a light computational load supports the semi-quantitative evaluation of neuromuscular ultrasound.
由于计算需求高且缺乏标准化测试方案，许多用于肌肉超声图像分析的人工智能方法尚未在临床神经肌肉医学实践中的可用设备上实现。本研究评估了使用实时纹理分析区分各种病理状况的可行性。
我们分析了75名参与者肱二头肌的17021张横截面超声图像，其中包括25名患有神经源性疾病、25名患有肌源性疾病和25名健康对照者。随机选择感兴趣区域的大小和位置以尽量减少偏差。开发了一种利用诸如差异性和同质性等纹理特征的随机森林分类器，并将其部署在移动个人电脑上，实现实时分析。
该分类器区分患者的准确率为81%。共生矩阵中的回声性和对比度是显著的预测特征。对15名患者进行验证，在15秒视频中，每张图像/每名患者的准确率分别达到78%/93%。使用移动个人电脑有助于在超声检查期间实时估计潜在病理情况，而不影响操作过程。
实时自动纹理分析作为神经肌肉疾病诊断的辅助手段是可行的。
使用纹理分析且计算负荷轻的人工智能支持神经肌肉超声的半定量评估。
