Brain Age Estimation Based on Brain Magnetic Resonance Imaging of Healthy Adults

Abstract

Recent studies have shown that age-related brain changes can be used to accurately predict the chronological age in healthy people by training machine learning methods on neuroimaging data. This prediction can be used to determine whether the individual shows signs of abnormal brain aging which can facilitate early diagnoses of diseases associated with abnormal brain aging. In this dissertation, two different methods are developed to predict the chronological age in healthy people from their brain magnetic resonance imaging (MRI) scans with the aim to achieve state-of-the-art accuracy. The two methods are 3D convolutional neural networks (CNN) and support vector regression (SVR), both trained to find the most accurate mapping from the brain data of healthy individuals to the chronological age. In this work, the grey matter (GM) volume images are extracted from the T1-weighted MRI scans firstly, then the extracted GM volume images are used in both methods. The 3D CNN model is trained on the extracted GM volume data directly, whereas the SVR model is trained on a dimensionally reduced GM volume data. By using a proper network architecture, the 3D CNN model achieved a state-of- the-art accuracy with a mean absolute error (MAE) of 3.69 years on the holdout test set. Moreover, the proposed 3D CNN model shows significant improvements in the execution time and GPU resources compared with the previous state-of-the-art similar brain age prediction system. On the other hand, the brain age prediction SVR model is trained using only 100 reduced features from the GM images and predicted the age accurately with an MAE of 5.99 years. This is not the best prediction accuracy compared with that of other similar methods but is comparable. Besides, in our work, experiments show that the SVR model’s accuracy continuously improves as the number of dimensionally reduced features increase. The proposed 3D CNN model has the potential to be used as a tool for early diagnosis of brain disorders based on age-prediction differences, which represents abnormal brain changes, thus facilitating early treatment or preventative intervention.