Diffusion tensor imaging for the differential diagnosis of Parkinsonism by machine learning.

Background: There are currently no specific tests for either idiopathic Parkinson's disease or Parkinson-plus syndromes. The study aimed to investigate the diagnostic performance of features extracted from the whole brain using diffusion tensor imaging concerning parkinsonian disorders.

Methods: The retrospective data yielded 625 participants (average age: 61.4 ± 8.2, men/women: 313/312; healthy controls/idiopathic Parkinson's disease/multiple system atrophy/progressive supranuclear palsy: 219/286/51/69) between 2008 and 2017. Diffusion-weighted images were obtained using a 3T MR scanner. The 90th, 50th, and 10th percentiles of fractional anisotropy and mean/axial/radial diffusivity from each parcellated brain area were recorded. Statistical analysis was evaluated based on the features extracted from the whole brain, as determined using discriminant function analysis and support vector machine. 20% of the participants were used as an independent blind dataset with 5 times cross-verification. Diagnostic performance was evaluated by the sensitivity and the F1 score.

Results: Diagnoses were accurate for distinguishing idiopathic Parkinson's disease from healthy control and Parkinson-plus syndromes (87.4 ± 2.1% and 82.5 ± 3.9%, respectively). Diagnostic F1 scores varied for Parkinson-plus syndromes with 67.2 ± 3.8% for multiple system atrophy and 71.6 ± 3.5% for progressive supranuclear palsy. For early and late detection of idiopathic Parkinson's disease, the diagnostic performance was 79.2 ± 7.4% and 84.4 ± 6.9%, respectively. The diagnostic performance was 68.8 ± 11.0% and 52.5 ± 8.9% in early and late detection to distinguish different Parkinson-plus syndromes.

Conclusions: Features extracted from diffusion tensor imaging of the whole brain can provide objective evidence for the diagnosis of healthy control, idiopathic Parkinson's disease, and Parkinson-plus syndromes with fair to very good diagnostic performance.