Hybrid diffusion tensor imaging feature-based AD classification.

Area of Science:

• Neuroimaging
• Machine Learning
• Medical Diagnostics

Background:

• Alzheimer's disease (AD) detection remains challenging in clinical settings.
• Machine learning models offer significant potential to assist in AD diagnosis.

Purpose of the Study:

• To develop and evaluate a novel detection model for improved accuracy in diagnosing Alzheimer's disease.
• To enhance diagnostic capabilities for clinicians through advanced computational methods.

Main Methods:

• Brain networks were constructed from diffusion tensor and T1w images of AD patients (n=98) and normal controls (n=100).
• Graph theory was employed to extract 9 types of features from 3D brain networks.
• Selected features underwent Pearson correlation analysis, and machine learning classifiers (Random Forest, SVM, CNN) were trained and tested.

Main Results:

• A support vector machine (SVM) model trained with Degree Centrality features achieved the highest accuracy of 98% with 96% sensitivity and 100% specificity.
• A Random Forest classifier using Shortest Path Length (SPL) features reached 90% accuracy, 92% sensitivity, and 88% specificity.
• A Convolutional Neural Network (CNN) model also achieved 90% accuracy with 72% sensitivity and 94% specificity.

Conclusions:

• The developed machine learning models significantly improve the accuracy of Alzheimer's disease detection.
• This approach avoids biases associated with direct dimensionality reduction from high-dimensional data, offering a more robust diagnostic tool.