Table 2.
Architecture of the CNN for brain morphometry.
| Layer | Kernel | Stride | Filters | Output size | Activation function |
|---|---|---|---|---|---|
| Input | – | – | – | 256 × 256 × 256 | – |
| Conv3D | 11 × 11 × 11 | 3 | 144 | 86 × 86 × 86 × 144 | ReLU |
| MaxPool | 3 × 3 × 3 | 2 | – | 42 × 42 × 42 × 144 | – |
| Conv3D | 5 × 5 × 5 | 2 | 192 | 21 × 21 × 21 × 192 | ReLU |
| MaxPool | 3 × 3 × 3 | 2 | – | 10 × 10 × 10 × 192 | – |
| Conv3D | 5 × 5 × 5 | 1 | 192 | 10 × 10 × 10 × 192 | ReLU |
| MaxPool | 3 × 3 × 3 | 2 | – | 4 × 4 × 4 × 192 | – |
| FC | – | – | – | 374 | ReLU |
| FC | – | – | – | 192 | – |
| FC | – | – | – | 165 | – |
Dropout (0.4) is applied after the last MaxPool layer and after first FC layer.
A bias is added to the first convolutional and all fully connected layers. Conv3D, 3D convolution; FC, fully connected layer; ReLU, rectified linear unit.