Biomarkers.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Gerontology

Background:

• Brain age models assess brain health and neurodegeneration by calculating the brain age gap (BAG).
• Deep learning (DL) models offer advanced capabilities for brain age prediction.
• Scanner variability poses challenges for longitudinal brain health studies.

Purpose of the Study:

• To develop DL-based, scanner-invariant brain age models for aging and dementia research.
• To enable reliable longitudinal follow-up despite potential changes in scanning equipment.

Main Methods:

• Utilized T1 scans from 3374 cognitively unimpaired (CU) participants in the Mayo Clinic Study of Aging (MCSA).
• Developed two DenseNet models: Model-A (trained on a subset) and Model-B (trained on all CU).
• Incorporated histogram matching and scanner type as inputs, testing on cross-vendor data.

Main Results:

• Including scanner type as input reduced mean age prediction differences between scanners (Model-A: 2.17 years, Model-B: 1.71 years).
• Model-B demonstrated lower Mean Absolute Error (MAE) for predicting brain age in MCI participants (4.06 years) compared to Model-A (5.72 years).
• Longitudinal prediction accuracy increased with age for both models.

Conclusions:

• Training DL models on more variable data (all CU participants) improves prediction accuracy and reduces scanner-related errors.
• Histogram matching and scanner input create scanner-invariant brain age estimation.
• The developed models enhance the reliability of longitudinal brain age estimation in aging and dementia studies.