Predicting the Age of the Brain with Minimally Processed T1-weighted MRI Data

Abstract

AbstractIn the last few years, several models trying to calculate the biological brain age have been proposed based on structural magnetic resonance imaging scans (T1-weighted MRIs, T1w), using multivariate methods and artificial intelligence. We aimed to develop and validate a convolutional neural network (CNN) model for brain age prediction (PBA) using minimally processed T1w MRIs. Our model only requires one preprocessing step (i.e., image registration to MNI space), which is an advantage in comparison with previous methods that require more preprocessing steps. We used a multi-cohort dataset of cognitively healthy individuals comprising 16734 MRIs for training and evaluation. To validate our model and its interpretability, we used a multivariate model, Orthogonal Projections to Latent Structures (OPLS), which uses brain segmented cortical thicknesses and volumes. We trained and evaluated the models with the same dataset, and systematically investigated how predictions of the CNN model differ from those of the OPLS model. The validation of our model was made by testing an external dataset. The CNN and the OPLS model achieved a mean absolute error (MAE) in the testing dataset of 3.04 and 4.81 years, respectively. The model’s performance in the external dataset was in the typical range of MAE found in the literature for testing sets. The CNN model revealed similar image patterns when grouped by chronological age (CA) and CNN predicted age. No significant differences were found between the oldest and youngest quartiles of age predictions by the CNN in a validation cohort of individuals with CA of 70 years old. Sensitivity maps analysis revealed that the age prediction is based mainly on the ventricles and other CSF spaces, which have been shown in the literature to reflect aging and are in accordance with the most important regions for the prediction in the OPLS model. While both the CNN and the OPLS model demonstrated acceptable performance metrics on a hold-out test set, individual predictions differed substantially, with brain age patterns of the CNN model being more comparable to the chronological age. In conclusion, our CNN model showed results comparable to the literature, using minimally processed images, which may facilitate the future implementation of brain age prediction in research and clinical settings.