Brain Age Prediction Using Multi-Hop Graph Attention Combined with Convolutional Neural Network.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Convolutional Neural Networks (CNNs) are widely used for brain age prediction from MRI scans.
• Existing CNNs primarily capture local image features, neglecting crucial connective information between distant brain regions.

Purpose of the Study:

• To develop a novel module that integrates local and global brain connectivity for enhanced brain age prediction.
• To improve the accuracy of biological age estimation from brain MRI data.

Main Methods:

• A multi-hop graph attention (MGA) module was proposed and integrated into a CNN architecture (sSE-ResNet18).
• The MGA module converts CNN feature maps into graph data, models multi-hop connections using Markov chains, and re-converts to updated feature maps.
• The model (MGA-sSE-ResNet18) was trained and validated on 2788 3D T1-weighted MR images from healthy subjects.

Main Results:

• The MGA-sSE-ResNet18 model achieved a mean absolute error (MAE) of 2.822 years.
• A high Pearson's correlation coefficient (PCC) of 0.968 was obtained, indicating strong agreement between predicted and actual brain ages.
• The proposed model outperformed four established CNNs and two representative brain age prediction models.

Conclusions:

• The multi-hop graph attention (MGA) module significantly enhances CNN-based brain age prediction by incorporating global brain connectivity.
• The MGA-sSE-ResNet18 model demonstrates superior accuracy and potential for clinical applications in neuroimaging.
• This approach offers a promising direction for more precise and reliable brain age estimation.