Multi-slice representational learning of convolutional neural network for Alzheimer's disease classification using...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Alzheimer's Disease (AD) is a progressive neurodegenerative disorder affecting individuals over 65.
• Accurate and early diagnosis is critical for managing AD, especially given the challenges in advanced stages.
• Existing deep learning methods for AD classification require large datasets and can be sensitive to environmental variations in medical images.

Purpose of the Study:

• To propose a novel deep learning-based method for Alzheimer's Disease diagnosis.
• To develop a model that is less sensitive to dataset variations for external validation.
• To utilize F-18 fluorodeoxyglucose positron emission tomography/computed tomography (FDG-PET/CT) imaging for AD classification.

Main Methods:

• A deep learning network was developed for AD diagnosis using FDG-PET/CT scans.
• The model incorporated a Global Average Pooling (GAP) layer, comparing its performance to a fully connected layer.
• The model was trained and validated on both a proprietary dataset and the publicly available Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.

Main Results:

• The proposed network achieved an accuracy of 86.09% and 91.02% on the proprietary and ADNI datasets, respectively.
• Sensitivity and specificity were also high, reaching 80.00% and 92.96% on the proprietary dataset, and 87.93% and 93.57% on the ADNI dataset.
• The GAP layer demonstrated statistically significant superior performance (p < 0.01) compared to the fully connected layer. Crucially, no statistically significant differences in performance were observed between the two datasets (p > 0.05).

Conclusions:

• The proposed deep learning model effectively classifies Alzheimer's Disease using FDG-PET/CT imaging.
• The model's ability to learn AD-specific features from the posterior cingulate cortex (PCC) was confirmed.
• The model's consistent performance across different datasets highlights its robustness and potential for widespread clinical application.