Classification of Parkinson's disease by deep learning on midbrain MRI

Area of Science:

• Neuroimaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Quantitative susceptibility mapping (QSM) and susceptibility map-weighted imaging (SMWI) enable accurate nigrosome-1 (N1) evaluation, crucial for Parkinson's disease (PD) diagnosis.
• Neuromelanin-sensitive (NMS) MRI can enhance automated N1 analysis by quantifying neuromelanin content, potentially improving PD detection.

Purpose of the Study:

• To compare the diagnostic performance of four approaches for Parkinson's disease classification.
• Evaluate a quantitative "QSM-NMS" composite marker, two deep learning (DL) models (morphological abnormality and volume using SMWI), and neuroradiological evaluation of N1 using SMWI.

Main Methods:

• Retrospective analysis of 3T midbrain MRI data from 82 PD patients and 107 healthy controls.
• Utilized T2*-SWI multi-echo-GRE for QSM/SMWI and NMS-MRI. Diagnostic performance was assessed using Area Under the Curve (AUC).
• Correlated imaging measures with clinical parameters (severity, duration, levodopa dose) using Spearman-Rho or Kendall-Tau-Beta correlation.

Main Results:

• Excellent classification performance was achieved by the QSM-NMS composite marker (AUC=0.94), N1 SMWI abnormality (AUC=0.92), N1 SMWI volume (AUC=0.90), and neuroradiologist evaluation (AUC=0.98).
• Identified common reasons for misclassification including asymmetry, re-slicing, pulsation artifacts, and thin N1.
• Significant correlation found between the substantia nigra QSM-NMS composite measure and levodopa dosing (rho=-0.303, p=0.006).

Conclusions:

• A quantitative QSM-NMS marker and DL-based PD classification algorithms demonstrate excellent performance using midbrain MRI.
• These findings support the clinical utility of these advanced imaging techniques for PD diagnosis.
• Further validation in earlier-stage PD cohorts is recommended to confirm clinical utility.