다발경화증과 시신경척수염범주질환의 감별진단을 위한 뇌 자기공명영상에 대한 기계학습

Abstract

Introduction: Distinguishing between neuromyelitis optica spectrum disorder (NMOSD) and multiple sclerosis (MS) is important because their treatments differ, and disease-modifying treatments for MS can worsen NMOSD. Brain magnetic resonance imaging (MRI) is one of the most important diagnostic tools used to differentiate between the two diseases. To date, considerable effort has been put in the identification of the brain MRI characteristics that enable the differentiation between MS and NMOSD. Machine learning has been studied as a method for identifying medical images. The study aimed to implement a supervised machine-learning method to perform differential diagnosis of MS and NMOSD by using brain MRIs.

Methods: Fluid-attenuated inversion recovery (FLAIR) MRIs were acquired from patients with relapsing-remitting MS (RRMS) and NMOSD with aquaporin-4 immunoglobulin G (AQP4-IgG) admitted at the Asan Medical Center, Seoul, Korea, between 2005 and 2017. FLAIR MRIs were used for a machine-learning method based on the combination of lesion frequency analysis for feature voxel selection and support vector machines (SVM) for classification algorithm. Diagnostic performance of machine learning was compared to that of two neurologists with more than two years of clinical experience in demyelinating disease.

Results: Final analysis included 746 and 292 MRIs from 172 patients with RRMS and 97 patients with NMOSD with AQP4-IgG, respectively. Lesion frequency analysis found that lesions adjacent to the lateral ventricle and in the inferior temporal lobe were frequently observed in RRMS, and dorsal medulla, cerebral peduncle/internal capsule, and corpus callosum lesions were frequently observed in NMOSD. The performance of SVM was 57.5% sensitivity, 78.4% specificity, and 63.3% accuracy, which showed a fair level of agreement with human raters (Cohen’s κ, rater A=0.279, rater B=0.262).

Conclusion: Machine learning using brain MRI data could discern RRMS and NMOSD with comparable accuracy to that of clinicians, encouraging the application of machine learning-aided diagnosis in clinical practice.