From Infancy to Aging: Precise Brain Age Estimation via Hybrid CoTResNet3D and CrossViT Models on T1-Weighted Imaging

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomarker Development

Background:

• Accurate brain age estimation from structural MRI is crucial for assessing neurodevelopment and identifying neurological disorder risks.
• Current models face challenges in lifespan generalizability, cross-center consistency, and temporal reliability.
• A large-scale, multi-center dataset is needed to overcome these limitations.

Purpose of the Study:

• To develop a robust and generalizable brain age estimation model across the entire lifespan.
• To address limitations in current models, including heterogeneity, cross-center generalizability, and temporal reliability.
• To introduce a novel hybrid deep learning architecture for enhanced feature extraction and global dependency modeling.

Main Methods:

• Curated a large-scale, multi-center T1-weighted MRI dataset (22,271 scans, 0-96 years) from 17 cohorts for training and 10 for validation.
• Proposed ResNet-CrossViT, a hybrid architecture combining 3D Contextual Transformer-ResNet (CoTResNet3D) and CrossVision Transformer (CrossViT).
• Evaluated model performance on internal, external cross-center, longitudinal, and test-retest datasets.

Main Results:

• Achieved a mean absolute error (MAE) of 2.72 years and maximal MAE (mMAE) of 5.10 years on the internal test set.
• Demonstrated strong cross-center generalizability (MAE = 4.19 years) and superior longitudinal consistency (MAdE = 3.68).
• Exhibited excellent test-retest reliability (Intraclass Correlation Coefficient, ICC = 0.994).

Conclusions:

• ResNet-CrossViT provides a precise, generalizable, and reliable framework for brain age estimation.
• The hybrid architecture effectively captures local and global brain structural information.
• This framework advances neurodevelopmental and aging research using neuroimaging biomarkers.