Comparison of automated and manual quantification methods for neuromelanin-sensitive MRI in Parkinson's disease

Abstract

Neuromelanin-sensitive magnetic resonance imaging quantitative analysis methods have provided promising biomarkers that can noninvasively quantify degeneration of the substantia nigra in patients with Parkinson's disease. However, there is a need to systematically evaluate the performance of manual and automated quantification approaches. We evaluate whether spatial, signal-intensity, or subject specific abnormality measures using either atlas based or manually traced identification of the substantia nigra better differentiate patients with Parkinson's disease from healthy controls using logistic regression models and receiver operating characteristics. Inference was performed using bootstrap analyses to calculate 95% confidence interval bounds. Pairwise comparisons were performed by generating 10,000 permutations, refitting the models, and calculating a paired difference between metrics. Thirty-one patients with Parkinson's disease and 22 healthy controls were included in the analyses. Signal intensity measures significantly outperformed spatial and subject specific abnormality measures, with the top performers exhibiting excellent ability to differentiate patients with Parkinson's disease and healthy controls (balanced accuracy = 0.89; area under the curve = 0.81; sensitivity =0.86; and specificity = 0.83). Atlas identified substantia nigra metrics performed significantly better than manual tracing metrics. These results provide clear support for the use of automated signal intensity metrics and additional recommendations. Future work is necessary to evaluate whether the same metrics can best differentiate atypical parkinsonism, perform similarly in de novo and mid-stage cohorts, and serve as longitudinal monitoring biomarkers.