Automated Categorization of Parkinsonian Syndromes Using Magnetic Resonance Imaging in a Clinical Setting

Area of Science:

• Neurology
• Radiology
• Artificial Intelligence

Background:

• Machine learning (ML) algorithms show promise in differentiating parkinsonian syndromes using MRI data.
• Clinical validation of these ML algorithms in routine practice is lacking.

Purpose of the Study:

• To evaluate the accuracy of an ML algorithm for categorizing parkinsonian syndromes.
• The algorithm was trained on a research cohort and tested on an independent clinical cohort.

Main Methods:

• 322 subjects (94 controls, 119 Parkinson's disease, 51 PSP, 35 MSA-P, 23 MSA-C) were analyzed.
• MRI volumetry and DTI metrics from 13 brain regions were used as input for a supervised ML algorithm.
• Data normalization techniques were applied to mitigate scanner-dependent effects.

Main Results:

• High classification accuracies (balanced accuracies: 0.840-0.983) were achieved for PD-PSP, PD-MSA-C, and PSP-MSA-C using volumetry.
• Classification performance was lower for PD-MSA-P and MSA-C-MSA-P comparisons.
• DTI metrics improved with control-based normalization but were less effective than volumetry alone or combined.

Conclusions:

• Volumetry-based ML enables accurate classification of early-stage parkinsonism in clinical settings.
• The study demonstrates the feasibility of using ML with MRI data across different scanners for diagnosing parkinsonian syndromes.