Differentiation of neurodegenerative parkinsonian syndromes by volumetric magnetic resonance imaging analysis and support vector machine classification

Background: Clinical differentiation of parkinsonian syndromes is still challenging. Objectives: A fully automated method for quantitative MRI analysis using atlas-based volumetry combined with support vector machine classification was evaluated for differentiation of parkinsonian syndromes in a multicenter study. Methods: Atlas-based volumetry was performed on MRI data of healthy controls (n = 73) and patients with PD (204), PSP with Richardson's syndrome phenotype (106), MSA of the cerebellar type (21), and MSA of the Parkinsonian type (60), acquired on different scanners. Volumetric results were used as input for support vector machine classification of single subjects with leave-one-out cross-validation. Results: The largest atrophy compared to controls was found for PSP with Richardson's syndrome phenotype patients in midbrain (−15%), midsagittal midbrain tegmentum plane (−20%), and superior cerebellar peduncles (−13%), for MSA of the cerebellar type in pons (−33%), cerebellum (−23%), and middle cerebellar peduncles (−36%), and for MSA of the parkinsonian type in the putamen (−23%). The majority of binary support vector machine classifications between the groups resulted in balanced accuracies of >80%. With MSA of the cerebellar and parkinsonian type combined in one group, support vector machine classification of PD, PSP and MSA achieved sensitivities of 79% to 87% and specificities of 87% to 96%. Extraction of weighting factors confirmed that midbrain, basal ganglia, and cerebellar peduncles had the largest relevance for classification. Conclusions: Brain volumetry combined with support vector machine classification allowed for reliable automated differentiation of parkinsonian syndromes on single-patient level even for MRI acquired on different scanners.