Table 1 Architectures and hyperparameters of 2D U-Net, 2.5Da U-Net, and 3D U-Net structures.
2D U-Net | 2.5Da U-Net | 3D U-Net | ||
|---|---|---|---|---|
Architecture | Convolution | Size = 3 × 3 Stride = 1 Zero-padding | Size = 3 × 3 Stride = 1 Zero-padding | Size = 3 × 3 × 3 Stride = 1 Zero-padding |
Down sampling maxpooling | Size = 2 × 2 Stride = 1 | Size = 2 × 2 Stride = 1 | Size = 2 × 2 × 2 Stride = 1 | |
Up sampling | Size = 2 × 2 Stride = 1 | Size = 2 × 2 Stride = 1 | Size = 2 × 2 × 2 Stride = 1 | |
Activation function | ReLu | ReLu | ReLu | |
U-Net layers | 4 | 4 | 4 | |
First layer features | 32 | 32 | 32 | |
Hyper parameter | Input data size | 512 × 512 × 1 | 512 × 512 × 3 | 64 × 64 × 128 |
Optimizer | Adam | Adam | Adam | |
Loss function | BCE | BCE | BCE | |
Initial learning rate | 0.0001 | 0.0001 | 0.0001 | |
Batch size | 12 | 12 | 6 | |
Epoch | 150 | 150 | 200 | |
Callback function | Reduce learning rate (newLR = LR × 0.95 when val_loss in 10 epochs are no better) Early stopping (training stop when val_loss in 50 epochs are no better) | |||
