Conference Paper Published

Abstract

By applying deep learning on 3D MR images, global brain age measures have been formulated by predicting the chronological age. However, little attention has been paid to regional brain age measures, which may capture the cognitive functions supported by individual brain regions. Using a sample of 1,703 T1-weighted images from the Human Connectome Projects (HCP-Young adult and HCP-Aging), we constructed 3D convolutional neural networks to estimate the regional brain ages of 18 cortical regions (9 on each cerebral hemisphere). Two hypotheses were tested: (1) on the cerebral cortex, there are higher associations of regional brain ages between homo-topic than non-homotopic region-pairs; and (2) regional brain ages are directly associated with both crystallized and fluid intelligence even after chronological age is factored out. Three methods were implemented to ameliorate the "regression to the mean" (RttM) problem, ubiquitous in brain age studies, including Cole’s method, de Lange’s method, and our proposed "percentile matching" method, in which the prediction of the testing data is corrected for by mapping the distributional properties of the predicted value and ground truth using the validation data. While our results supported both hypotheses and confirmed the desirable symmetry and association properties, consistent evidence was found only when the regional brain ages were corrected for by either Cole’s method or the percentile matching method. Between these two methods, the latter method yielded smaller mean absolute error and root mean square error, along with larger Spearman’s correlation. Our results highlight the potential of regional brain ages and the percentile matching method in brain age research.Clinical relevance—Regional brain ages exhibit symmetry properties and associations with cognitive functions, and they may represent promising biomarkers of brain ageing.