Figure 2.

Scheme of our 3D CNN for training (A) and using for external dataset (B). In (A), the input is a T1w brain MRI previously registered to the MNI space, the subject’s chronological age, and the output is the predicted brain age of the subject. Training and developing (dev) datasets are also shown, with data augmentation being performed in 70% of the training set. In (B), the input is only a T1w MRI previously registered to the MNI space before using the trained network (model’s weight) to predict brain age in the new and unseen data of an external dataset. Conv3D, 3D convolution; BatchNorm3d, 3D batch normalisation; Dev, development dataset (for testing model’s performance while training); (Leaky)ReLu, (leaky) rectified linear unit; MaxPool3D, 3D max pooling; AvgPool3D, 3D average pooling; FC, fully connected layer; ResNet, residual network block.