From a deep learning model back to the brain-Identifying regional predictors and their relation to aging

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Gerontology

Background:

• Brain age prediction from structural MRI is linked to neurodegenerative diseases and mortality.
• Existing methods for identifying brain regions contributing to age prediction (explanation maps) are often noisy and unreliable.
• Understanding brain aging processes requires reliable identification of contributing brain structures.

Purpose of the Study:

• To develop a robust deep learning framework for chronological age prediction from structural MRI.
• To create reliable, population-based "explanation maps" identifying brain regions crucial for age prediction.
• To investigate the role of specific brain structures, like cerebrospinal fluid cavities, in brain aging.

Main Methods:

• A deep learning framework utilizing a convolutional neural network (CNN) ensemble was developed.
• A novel inference scheme was implemented to combine subject-specific prediction contributions into population-based explanation maps.
• The framework was trained and evaluated on a lifespan dataset of 10,176 T1-weighted brain MRI scans.

Main Results:

• The model achieved a mean absolute error of 3.07 years and a correlation of r=0.98 between chronological and predicted age.
• Population-based explanation maps revealed cerebrospinal fluid cavities as highly influential in age prediction.
• The method demonstrated increased replicability of explanation maps, aligning with voxel-based morphometry studies.

Conclusions:

• The developed deep learning framework provides accurate chronological age prediction from brain MRI.
• The novel inference scheme generates reliable, population-based insights into brain aging mechanisms.
• Cerebrospinal fluid cavities are highlighted as significant indicators of brain aging, offering new avenues for research.