Deep Learning to Differentiate Parkinsonian Syndromes Using Multimodal Magnetic Resonance Imaging: A Proof-of-Concept...

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Differentiating multiple system atrophy (MSA) from Parkinson's disease (PD) is clinically challenging, particularly in early stages.
• Deep learning (DL) and magnetic resonance imaging (MRI) offer potential for automated diagnostic assistance.

Purpose of the Study:

• To assess the feasibility of a 3D convolutional neural network (CNN) using multimodal, multicentric MRI data.
• To differentiate MSA and its variants (MSA-P, MSA-C, MSA-PC) from PD.

Main Methods:

• Retrospective collection of MRI data from three French MSA reference centers.
• Quantitative MRI maps (gray matter density [GD] and mean diffusivity [MD]) were input into a 3D CNN.
• Classification models were trained for PD vs. MSA, PD vs. MSA-C&PC, and PD vs. MSA-P differentiation.

Main Results:

• The study included 92 MSA patients and 64 PD patients.
• The best classification accuracy (0.88 ± 0.03) was achieved for PD/MSA differentiation using combined GD-MD maps.
• Activation maps identified key regions like the putamen and cerebellum in MSA pathophysiology.

Conclusions:

• The developed 3D CNN approach shows promise for an efficient, MRI-based, and user-independent diagnostic tool.
• This tool could aid in differentiating parkinsonian syndromes in clinical practice.