Exploiting macro- and micro-structural brain changes for improved Parkinson's disease classification from MRI data.

Area of Science:

• Neuroimaging
• Machine Learning
• Neurology

Background:

• Parkinson's disease (PD) is a common neurodegenerative disorder.
• Early and accurate diagnosis of PD is critical for effective management and improving patient outcomes.
• Diagnosing PD, especially in its early stages, presents significant challenges.

Purpose of the Study:

• To develop and validate an explainable deep learning model for classifying Parkinson's disease (PD) using multimodal magnetic resonance imaging (MRI) data.
• To enhance the interpretability of the classification model by identifying key brain regions contributing to PD diagnosis.
• To leverage a large, diverse dataset for robust model training and evaluation.

Main Methods:

• Utilized a large dataset of 1264 multimodal MRI scans (T1-weighted and diffusion-tensor imaging) from 611 PD patients and 653 healthy controls.
• Employed a convolutional neural network (CNN) trained on pre-processed imaging data and demographic information.
• Implemented SmoothGrad saliency maps for model explainability, identifying critical brain regions for PD classification.

Main Results:

• Achieved a high classification performance with ROC-AUC of 0.89, 80.8% accuracy, 82.4% specificity, and 79.1% sensitivity on the test set.
• Saliency maps indicated that diffusion tensor imaging (DTI) metrics, particularly fractional anisotropy, were more influential than T1-weighted MRI.
• Identified key brain regions, including the brainstem, thalamus, amygdala, hippocampus, and cortical areas, as important for PD classification.

Conclusions:

• The developed explainable deep learning model accurately classifies PD patients from healthy controls using multimodal MRI.
• Micro-structural brain changes, as detected by DTI, play a significant role in the progression of Parkinson's disease.
• The model's explainability provides clinically relevant insights into the neuroimaging biomarkers of PD.