Recent publications have shown that training supervised regression methods on MRI brain imaging can be used to predict the brain age of an individual with high precision. We can use these predictions to detect diseases associated with abnormal brain ageing where the predicted age does not match the chronological age.
In this paper, we develop a convolutional neural network to predict brain age accurately. The architecture of the model is a simplified adaptation of the VGG architecture. The network is trained on healthy grey-matter segmented images and applied to clinical T1-weighted MRIs.
The model is trained on a publicly available healthy dataset and applied to a clinical dataset consisting of Schizophrenia, Parkinson’s Disease, and Post-Traumatic Stress Disorder patients. We demonstrated bias in brain age prediction, and we corrected it to improve the reliability of the results. Our BrainAge model obtained a mean absolute error (MAE) of 4.03 years and 0.96 R2 on the healthy dataset after correcting the bias. We used transfer learning to apply the BrainAge model to the clinical data and compared the brain age delta (predicted age – chronological age) for each condition. The results were not statistically significant p<.05, meaning that the brain age delta does not indicate abnormal brain ageing in this instance.