Differential diagnosis between Parkinson's disease (PD) and atypical parkinsonism, such as multiple system atrophy (MSA), can be difficult, especially in the early stages of the disease. Deep learning using neural networks (NNs) makes possible the prediction of the diagnosis using various types of biomarkers, unlike conventional linear statistics. We aimed to differentiate the Parkinson's variant of MSA (MSA-P) from PD both in the early stages by clinical utilization of NN analyses before the hot cross-bun and putaminal rim imaging features of MSA appeared. Analysis by NN involved the data of voxel-based morphometry (VBM) that indicate morphological changes and magnetic resonance spectroscopy (MRS) that indicate qualitative changes. VBM analysis showed that compared with PD patients, MSA-P patients showed atrophy in the superior cerebellar peduncle, middle cerebellar peduncle, cerebellar hemisphere, pons, midbrain, and putamen, but not in the globus pallidus. Proton MRS on the globus pallidus in the diseased hemisphere, lacking atrophy as observed with VBM, revealed decreased neurons and gliosis in both groups. Clinical differentiation of MSA-P from PD using NN analysis, involved measuring the prediction potential using the area under the receiver operator characteristic (ROC) curves (AUC). Using both VBM and MRS data, NNs contributed adequately to the clinical diagnosis.
Read full abstract