Table 1 GAN architecture, hyperparameters, and training details.
From: Multimodal medical image fusion combining saliency perception and generative adversarial network
Component | Details |
|---|---|
Generator architecture | Consists of multiple convolutional layers with ReLU activation, followed by batch normalization. Upsampling layers refine image quality. |
Discriminator architecture | A series of convolutional layers with LeakyReLU activation, batch normalization, and a fully connected layer for classification. |
Input dimensions | Image patches of size 256 × 256 pixels for training. |
Loss function | Binary Cross-Entropy (BCE) loss for adversarial training. |
Optimizer | Adam optimizer with a learning rate of 0.0002 for stable convergence. |
Batch size | 32 images per batch to balance training stability and computational efficiency. |
Training epochs | 1200 epochs with early stopping based on validation loss. |
Regularization | Dropout (0.3) applied to prevent overfitting in both generator and discriminator. |
Evaluation metrics | F1-score, Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR) for quality assessment. |
Data augmentation | Random flipping, rotation, and contrast adjustments applied to improve generalization. |
