Feature integration of [18F]FDG PET brain imaging using deep learning for sensitive cognitive decline detection

Area of Science:

• Neuroimaging
• Machine Learning
• Medical Diagnostics

Background:

• Accurate diagnosis of cognitive decline (CD), including early Alzheimer's disease, is crucial for timely intervention.
• Positron emission tomography (PET) reveals metabolic brain changes in CD but faces limitations due to cost and radiation.
• A data-efficient framework is needed to enhance PET's clinical utility by integrating multi-scale data.

Purpose of the Study:

• To propose a novel framework integrating multi-scale PET representations (voxel-level and region-level) for improved cognitive decline detection.
• To develop a data-efficient method that overcomes limitations of traditional PET imaging.
• To enhance diagnostic accuracy for cognitive decline using advanced machine learning techniques.

Main Methods:

• Extracted voxel-level features using CNN and PCANet from [¹⁸F]FDG PET scans.
• Derived region-level features using DNN from standardized uptake value ratio measurements.
• Integrated voxel- and region-level features via direct concatenation and applied ensemble machine learning models.
• Validated models using 5-fold cross-validation on 252 participants from the Alzheimer's Disease Neuroimaging Initiative.

Main Results:

• The integrated DNN-CNN model achieved the highest classification accuracy of 0.87 ± 0.05.
• This integrated approach showed a 6.33% improvement in accuracy and reduced standard deviation compared to DNN-only models.
• Significant improvements were observed in Recall (0.77 to 0.88) and F1-Score (0.82 to 0.88).
• Model predictions correlated significantly with Mini-Mental State Examination (MMSE) scores, outperforming MMSE-based classification.

Conclusions:

• Combining multi-scale PET representations with deep learning significantly improves cognitive decline classification performance.
• The proposed framework enhances sensitivity to cognitive decline, outperforming single-representation models.
• Multi-scale FDG-PET representations show potential for machine learning-based detection of cognitive decline.