Enhancing early Parkinson's disease detection through multimodal deep learning and explainable AI: insights from the...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Parkinson's Disease (PD) is a common neurodegenerative disorder affecting millions globally.
• Early detection of PD is crucial for effective management and treatment.

Purpose of the Study:

• To investigate multimodal deep learning for prodromal Parkinson's Disease detection.
• To integrate imaging and clinical data using a novel co-learning approach.
• To enhance diagnostic accuracy through Explainable AI (XAI) techniques.

Main Methods:

• Utilized the Parkinson's Progression Markers Initiative (PPMI) dataset.
• Developed a joint co-learning framework for multimodal data fusion in deep neural networks.
• Employed 3D Convolutional Neural Networks (CNNs) with Excitation Networks (EN) and Vision Transformers (ViT).
• Applied XAI methods like Integrated Gradients and Attention Heatmaps for model interpretability.

Main Results:

• DenseNet with EN demonstrated superior performance, significantly improving accuracy when clinical data was included.
• XAI analysis revealed DenseNet's focus on critical brain regions (right temporal, left pre-frontal) for prodromal PD.
• ViT highlighted the lateral ventricles, potentially linked to cognitive decline in early PD stages.

Conclusions:

• The proposed multimodal deep learning framework effectively aids in early PD diagnosis and subtype prediction.
• Identified brain regions and lateral ventricles show promise as early biomarkers for Parkinson's Disease.
• Findings support the development of advanced diagnostic tools for precision medicine in neurodegenerative diseases.