Dynamic multi-scale deep learning with mixture of experts for differentiating iNPH and PSP using MRI.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Idiopathic normal pressure hydrocephalus (iNPH) and progressive supranuclear palsy (PSP) share overlapping symptoms like gait issues and cognitive decline, making differential diagnosis difficult.
• Accurate differentiation is crucial for appropriate patient management and treatment strategies.

Purpose of the Study:

• To develop and validate a novel multi-scale deep learning framework for distinguishing iNPH from PSP.
• To integrate global and local magnetic resonance imaging (MRI) features using a mixture of experts (MoE) mechanism to improve diagnostic accuracy.

Main Methods:

• A framework combining a 3D CNN for global features and a 2.5D recurrent CNN for region-specific features (ventricles, sulci, midbrain, Sylvian fissures).
• A mixture of experts (MoE) mechanism was employed to dynamically weight global and local features.
• Model performance was evaluated using fivefold cross-validation on T1-weighted MRI from 118 patients (53 iNPH, 65 PSP).

Main Results:

• The MoE model achieved high accuracy (0.983), recall (0.985), precision (0.986), and an AUC of 1.000.
• The MoE approach outperformed traditional methods and single-branch deep learning models.
• Interpretability was enhanced through Grad-CAM visualizations, showing focus on disease-specific regions.

Conclusions:

• The MoE framework provides a robust and interpretable method for differentiating iNPH from PSP using integrated MRI features.
• This AI-driven approach reduces reliance on subjective assessments and shows potential for wider clinical use.
• Further validation with larger, multicenter datasets is recommended.