Improving brain age prediction with anatomical feature attention-enhanced 3D-CNN

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning models excel at predicting brain age from structural MRI (sMRI).
• Traditional anatomical features from sMRI have been underutilized in current deep learning approaches.
• Accurate brain age prediction is crucial for understanding cognitive decline and neurological disorders.

Purpose of the Study:

• To develop an attention-based network that integrates both anatomical and deep convolutional features from sMRI data.
• To improve the accuracy and efficiency of brain age prediction models.
• To leverage anatomical feature attention (AFA) for enhanced feature extraction.

Main Methods:

• An attention-based network incorporating an anatomical feature attention (AFA) module was designed.
• A fully convolutional network was employed to simplify deep convolutional feature extraction and reduce memory usage.
• The proposed model was evaluated on eight public sMRI datasets comprising 2501 participants.

Main Results:

• The proposed fusion model achieved a mean absolute error (MAE) of 2.20 years in brain age prediction.
• The model demonstrated superior performance compared to several widely-used deep learning models.
• The inclusion of the AFA module significantly improved the overall performance of the deep learning models.

Conclusions:

• The integration of anatomical and deep convolutional features offers a promising approach for accurate brain age prediction from sMRI.
• The developed attention-based network provides an effective method for capturing salient anatomical information.
• This approach has potential applications in clinical diagnosis and monitoring of age-related neurological conditions.