Accurate brain-age models for routine clinical MRI examinations

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Radiology

Background:

• Convolutional neural networks (CNNs) accurately predict chronological age from MRI scans.
• Existing brain-age models have limitations in clinical utility due to scan type, generalization, and processing time.
• There is a need for brain-age tools suitable for routine clinical examinations.

Purpose of the Study:

• To develop a brain-age prediction framework for routine clinical head MRI examinations.
• To create a tool that is fast, accurate, and generalizable across different scanners and hospitals.
• To assess the clinical relevance of brain-age predictions in identifying age-related atrophy.

Main Methods:

• A deep learning classifier was used to generate a dataset of 23,302 'radiologically normal for age' head MRI scans.
• Brain-age prediction was performed using minimally processed axial T2-weighted and diffusion-weighted scans.
• Model performance was evaluated for speed, accuracy (MAE), and generalizability across hospitals and scanner vendors.

Main Results:

• The framework achieved fast (<5s) and accurate brain-age prediction (MAE < 4 years).
• The model demonstrated good generalizability between hospitals and scanner vendors (Δ MAE < 1 year).
• Patients with radiologically confirmed age-excessive atrophy showed systematically higher brain-predicted age (mean difference +5.89 years) compared to controls (p < 0.0001).

Conclusions:

• The developed brain-age framework is feasible for real-time screening during routine hospital MRI examinations.
• The tool can automatically detect older-appearing brains, assisting in clinical decision-making.
• This approach has implications for early neurodegeneration detection, patient care, and optimizing MRI data collection.