Algorithm 1
From: High-resolution conditional MR image synthesis through the PACGAN framework
PACGAN training procedure. The training process consists of multiple phases, during which the algorithm operates on images of increasing resolution, ranging from 4 × 4 to 256 × 256. For each resolution, training is performed for a fixed number of epochs, denoted as #epochs. During each epoch, the discriminator is trained ncritic times (with ncritic varying depending on the resolution), while the generator is trained once. After completing 60% of the total training epochs, the validation_loss is computed for each epoch to identify the best_epoch, which corresponds to the model at the current resolution that exhibits superior generalization performance on the validation set. The model obtained at the best_epoch is considered the best model and serves as the starting point for training the subsequent resolution.
