EAMNet: an Alzheimer's disease prediction model based on representation learning.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• ¹⁸F-FDG PET scans assess brain lesion metabolism, crucial for predicting Alzheimer's disease (AD).
• Extracting relevant features and managing noise in PET images are significant challenges for accurate AD prediction.
• Existing methods struggle with the complexity of PET image analysis for early disease detection.

Purpose of the Study:

• To develop an innovative deep learning model, the efficient adaptive multiscale network (EAMNet), for predicting Alzheimer's disease (AD) using PET image slices.
• To enhance feature extraction and noise compensation in PET imaging for improved diagnostic accuracy.
• To enable earlier intervention and treatment planning for patients at risk of AD.

Main Methods:

• Implemented an efficient convolutional strategy to optimize the receptive field and reduce computational complexity in PET image analysis.
• Introduced a channel attention module for adaptive weight allocation, mitigating the impact of spatial noise on classification.
• Utilized skip connections to effectively merge multi-scale lesion information, aligning with clinical relevance.

Main Results:

• The EAMNet model demonstrated high classification performance on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.
• Achieved accuracy rates of 97.66% (AD vs. NC), 96.32% (AD vs. MCI), 95.23% (MCI vs. NC), and 95.68% (AD vs. MCI vs. NC).
• Visual analysis confirmed that the network's focus aligns with regions of interest identified by clinical experts.

Conclusions:

• The proposed EAMNet significantly advances the accurate prediction and classification of Alzheimer's disease using¹⁸F-FDG PET imaging.
• EAMNet outperforms existing state-of-the-art algorithms in AD classification tasks.
• The developed method offers a promising tool for early diagnosis and patient stratification in Alzheimer's disease research.