Brain Age Estimation from MRI Using a Two-Stage Cascade Network with Ranking Loss

Abstract

As age increases, human brains will be aged, and people tend to experience cognitive decline with a higher risk of neuro-degenerative disease and dementia. Recently, it was reported that deep neural networks, e.g., 3D convolutional neural networks (CNN), are able to predict chronological age accurately in healthy people from their T1-weighted magnetic resonance images (MRI). The predicted age, called as “brain age” or “brain predicted age”, could be a biomarker of the brain ageing process. In this paper, we propose a novel 3D convolutional network, called as two-stage-age-net (TSAN), for brain age estimation from T1-weighted MRI data. Compared with the state-of-the-art CNN by Cole et al., TSAN has several improvements: 1) TSAN uses a two-stage cascade architecture, where the first network is to estimate a discretized age range, then the second network is to further estimate the brain age more accurately; 2) Besides using the traditional mean square error (MSE) loss between chronological and estimated ages, TSAN considers two additional novel ranking losses, based on paired samples and a batch of samples, for regularizing the training process; 3) TSAN uses densely connected paths to combine feature maps with different scales; 4) TSAN considers gender labels as input features for the network, considering brains of male and female age differently. The proposed TSAN was validated in three public datasets. The experiments showed that TSAN could provide accurate brain age estimation in healthy subjects, yielding a mean absolute error (MAE) of 2.428, and a Pearson’s correlation coefficient (PCC) of 0.985, between the estimated and the chronological ages.

Co-first authors—Z. Liu and J. Cheng contributed equally.