Brain Age Prediction of Children Using Routine Brain MR Images via Deep Learning

Abstract

Predicting brain age of children accurately and quantitatively can give help in brain development analysis and brain disease diagnosis. Traditional methods to estimate brain age based on 3D magnetic resonance (MR), T1 weighted imaging (T1WI), and diffusion tensor imaging (DTI) need complex preprocessing and extra scanning time, decreasing clinical practice, especially in children. This research aims at proposing an end-to-end AI system based on deep learning to predict the brain age based on routine brain MR imaging. We spent over 5 years enrolling 220 stacked 2D routine clinical brain MR T1-weighted images of healthy children aged 0 to 5 years old and randomly divided those images into training data including 176 subjects and test data including 44 subjects. Data augmentation technology, which includes scaling, image rotation, translation, and gamma correction, was employed to extend the training data. A 10-layer 3D convolutional neural network (CNN) was designed for predicting the brain age of children and it achieved reliable and accurate results on test data with a mean absolute deviation (MAE) of 67.6 days, a root mean squared error (RMSE) of 96.1 days, a mean relative error (MRE) of 8.2%, a correlation coefficient (R) of 0.985, and a coefficient of determination (R ²) of 0.971. Specially, the performance on predicting the age of children under 2 years old with a MAE of 28.9 days, a RMSE of 37.0 days, a MRE of 7.8%, a R of 0.983, and a R ² of 0.967 is much better than that over 2 with a MAE of 110.0 days, a RMSE of 133.5 days, a MRE of 8.2%, a R of 0.883, and a R ² of 0.780.

Keywords: artificial intelligence; brain age; convolutional neural network; deep learning; magnetic resonance imaging.

Figure 1

Illustration of data augmentation to a 3D image.

Figure 2

The hierarchical architecture of proposed 3D CNN. The “32” and “128 × 116 × 12” in “32@128 × 116 × 12” denote the number and size of feature maps.

Figure 3

The hierarchical architecture of the 2D CNN.

Figure 4

Distribution of participant ages.

Figure 5

Training performance of one run. (A) Loss against training epoch, and (B) MAE against training epoch. The loss and MAE are the average of all iterations in one epoch.

Figure 6

Prediction results of the proposed 3D CNN. The error bar represents the average and standard deviation of the prediction results over 10-run.

Figure 7

Bland-Altman plots for the proposed 3D CNN. Plot (A–C) denote 0–2, 2–5, 0–5 age groups, respectively.

Figure 8

Prediction results of the proposed method without data augmentation. The error bar represents the average and standard deviation of the prediction results over 10-run.

Figure 9

Comparison of different network depths. “6, 2” (“No. of convolution layers, No. of fully connected layers”) denotes 6 convolution layers and 2 fully connected layers.

Figure 10

Training performance of one run without batch normalization. (A) Loss against training epoch, and (B) MAE against training epoch. The loss and MAE are the average of all iterations in one epoch.