Objective: To develop a multi-sequence MRI-based machine learning model differentiating PD from MSA.

Background: MRI protocols capable of distinguishing PD from MSA are greatly needed.

Method: A template-based deep gray matter (DGM) mapping method was applied to automatically segment seven bilateral deep gray matter nuclei: caudate nucleus, putamen, globus pallidus, red nucleus, substantia nigra, subthalamic nucleus, and dentate nucleus. Neuromelanin-related features of the bilateral substantia nigra pars compacta (SNpc) were extracted using a midbrain template method. The brainstem, midbrain, and pons were automatically segmented from T1-weighted images using an open-source deep learning model. Brain extraction was performed using the HD-BET tool for volume normalization. A multi-step pipeline was developed for feature processing and modeling. Radiomics features from QSM and R2* images, NM-MRI features (including volume, mean signal intensity, and relative contrast ratio of the substantia nigra pars compacta), and normalized brainstem, midbrain, and pons volumes were extracted and standardized via z-score normalization. Redundant features were removed by excluding one of any feature pair with a Pearson correlation coefficient >0.99. The Least Absolute Shrinkage and Selection Operator (LASSO) algorithm was then applied for feature selection and preliminary model construction for the differentiation of MSA vs. PD.

Results: Two datasets derived from STrategically Acquired Gradient Echo (STAGE) imaging were analyzed: one with a slice thickness of 1.34 mm (n=517; 471 PD patients,46 MSA patients) was used for model training and internal testing, while the other with a slice thickness of 2 mm (n=168; 155 PD, 13 MSA patients) was employed for external validation. For the classification of MSA versus PD, a LASSO model was constructed using 9 selected features. The model yielded AUC values of 0.937 (95% confidence interval [CI], 0.891–0.974), 0.842 (95% CI, 0.719–0.943), and 0.815 (95% CI, 0.623–0.971) in the training, internal test, and external test cohorts, respectively (Figure 1). The normalized pontine volume was identified as the most important feature for distinguishing MSA from PD.

Conclusion: Distinguishing PD from MSA using multi-parameter MRI data can achieve high sensitivity and specificity.

« Back to 2026 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/multi-sequence-mri-based-machine-learning-model-for-differentiating-parkinson-disease-pd-from-multiple-system-atrophy-msa/