Learning patterns of the ageing brain in MRI using deep convolutional networks.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Gerontology

Background:

• Brain morphological changes occur with normal aging and neurodegenerative diseases.
• Age-related brain changes are complex, exhibiting subtle, nonlinear, and heterogeneous patterns.
• Machine learning models can identify subtle patterns in brain appearance despite individual variability.

Purpose of the Study:

• To predict chronological age using a 3D convolutional neural network (CNN) architecture.
• To investigate the relationship between brain age prediction errors and clinical measurements.
• To explore the impact of image registration on CNN-based age prediction.

Main Methods:

• Developed a 3D CNN for brain age prediction using 12,802 T1-weighted MRI images for training and 6,885 for testing from the UK Biobank.
• Analyzed correlations between predicted brain age errors (ΔBrainAge) and UK Biobank clinical data.
• Examined the relationship between ΔBrainAge and image-derived phenotypes (IDPs) from multiple imaging modalities.
• Investigated the effect of nonlinear image registration on CNN performance.

Main Results:

• The CNN achieved competitive performance in chronological age prediction.
• ΔBrainAge showed significant correlations with various clinical measurements in both male and female groups.
• Correlations between ΔBrainAge and IDPs were consistent with known aging patterns.
• Nonlinear image registration artifacts were found to negatively impact CNNs' ability to detect subtle aging indicators.

Conclusions:

• CNNs, trained on UK Biobank MRI data, can predict brain age and identify deviations linked to clinical factors.
• The study highlights the importance of appropriate image preprocessing, cautioning against using nonlinearly registered images.
• Future longitudinal analysis of ΔBrainAge may reveal predictive value for health outcomes.