Deep learning architectures for Parkinson's disease detection by using multi-modal features.

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Early diagnosis of Parkinson's disease (PD) is critical for effective management and treatment.
• Delayed diagnosis can negatively impact patient outcomes.
• Multi-modal feature analysis for PD diagnosis is an area of ongoing research.

Purpose of the Study:

• To evaluate the diagnostic accuracy of deep learning models using multi-modal features for Parkinson's disease detection.
• To compare two deep learning frameworks: feature-level and modal-level integration of data.
• To assess the utility of combining neuroimaging (T1 MRI, SPECT) and biological (CSF) data for PD classification.

Main Methods:

• Two deep learning frameworks were developed: feature-level and modal-level.
• The feature-level framework integrated all T1 MRI, SPECT, and CSF features.
• The modal-level framework used ReliefF to reduce T1 MRI features before integration with SPECT and CSF data.

Main Results:

• Models achieved high diagnostic accuracy on an imbalanced dataset (73 PD, 59 healthy).
• The feature-level framework using Convolutional Neural Networks (CNN) reached a maximum accuracy of 93.33%.
• The modal-level framework using CNN achieved a maximum accuracy of 92.38%.

Conclusions:

• Multi-modal feature analysis, despite its complexity, is effective for classifying individuals with Parkinson's disease.
• This approach can significantly aid clinicians in accurate PD diagnosis.
• Combining diverse data types enhances diagnostic capabilities beyond single-modality approaches.