A deep learning model for brain age prediction using minimally preprocessed T1w images as input.

Area of Science:

• Neuroimaging
• Machine Learning
• Radiology

Background:

• Several models predict biological brain age using structural MRI and machine learning.
• Existing methods often require extensive preprocessing, limiting accessibility in research.

Purpose of the Study:

• To develop and validate a simplified convolutional neural network (CNN)-based model for biological brain age prediction.
• To enhance the accessibility and implementation of brain age prediction models in research settings.

Main Methods:

• Utilized a multicohort dataset of 17,296 T1-weighted MRIs from cognitively healthy individuals (age 32.0-95.7 years).
• Employed hold-out and cross-validation approaches (CNN1-4) with varying training data inclusion and skull-stripped images.
• Validated generalisability using two external datasets with different populations and MRI characteristics.

Main Results:

• Trained CNN models achieved mean absolute errors (MAE) between 2.67 and 3.08 years.
• Model performance on external datasets was comparable to existing literature.
• Salience maps indicated periventricular, temporal, and insular regions as key predictors of brain age.

Conclusions:

• The developed CNN model demonstrates good performance with minimal preprocessing (rigid image registration).
• The model offers a simplified and accessible tool for researchers studying aging and age-related disorders.
• Biological brain age can serve as a valuable metric for age correction in neuroimaging studies.