Using techniques like Image to Image translation for paired input input images, this work is successfully able to translate any MRI T1 input into a CT image in a supervised learning framework.
Experiment with hyperparameters, like weightage given to different loss functions, reconstruction loss, perceptual loss from VGG16, and adversarial loss. Performing a search over each parameter keeping others fixed.
Some fixed hyperparameters are:
- Learning rate: Here le = 2e-4 for 100 epochs, then linear learning rate decay till end ie. 200 epochs
- Discriminator loss reduction rate: Set to 0.5. We have to half GAN loss before backprop so that discriminator does not train much faster than generator
- Number of patches in discriminator: 16.
- Adam optimizer with beta_1 = 0.5 and beta_2 = 0.999
- Number of neighbouring slices used as input: 1. If we use more than 1 input slice, then contextual information can be used by the generator.
- Generator does not learn fast, maybe because of lots of loss applied to it
- G and D Loss fluctuate a lot in most cases
- Very small gain in accuracy is observed after 100 epochs, then it plateaus
Finding hyper-parameters is a hard problem using brute force. Some heuristic would help.