Figure 4.

Illustration of the CNN architecture. Arrays of size ${96}^3$ are used as input. First, three convolutional blocks having convolutional layers with 32, 48, and 64 filters and average pooling layers are used to extract microstructural features at different scales. Then, four fully-connected layers with 256, 192, 128, and 64 nodes are used to produce the logit-transformed diffusivity prediction as output. Elu activation is applied after all convolutional and fully-connected layers.