Automated identification of dementia using FDG-PET imaging

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Parametric FDG-PET imaging shows promise for automated dementia diagnosis.
• Current methods for characterizing and differentiating dementia patterns have limitations.

Purpose of the Study:

• To develop and evaluate a hybrid feature extraction, selection, and classification approach (GA-MKL) for differentiating dementia syndromes.
• To improve the accuracy of automated identification of Alzheimer's disease (AD) and frontotemporal dementia (FTD) compared to existing methods.

Main Methods:

• Extracted three feature groups (average level, spatial variation, asymmetry) from 116 cortical volumes of FDG-PET scans.
• Utilized a genetic algorithm (GA) for optimal feature selection.
• Employed a multikernel learning (MKL) machine for classification with automatic kernel weight estimation.

Main Results:

• The GA-MKL algorithm achieved 94.62% accuracy in separating dementia types (AD/FTD) from normal controls.
• Demonstrated superior performance compared to two state-of-the-art dementia identification algorithms.
• Showed strong agreement between the automated technique and clinical diagnoses.

Conclusions:

• The proposed GA-MKL approach offers a highly accurate and reliable method for automated dementia diagnosis using FDG-PET.
• This automated technique shows significant potential for clinical application in differentiating dementia syndromes.