CT-MRI Translation Model

Model Architecture

Generator Network
  • Encoder: 7-stage downsampling with residual blocks
  • Decoder: 7-stage upsampling with skip connections
  • Latent Space: 256-dimensional VAE
  • Activation: LeakyReLU (α=0.2)
  • Normalization: Group Normalization
  • Output: Sigmoid activation for 256x256x3 images
Discriminator Network
  • Type: Multi-scale PatchGAN
  • Scales: 4 levels of feature discrimination
  • Activation: LeakyReLU (α=0.2)
  • Output: Real/Fake classification at multiple scales

Loss Functions

Generator Loss
  • Adversarial Loss: LSGAN (Least Squares GAN)
  • Cycle Consistency Loss: L1 norm (λ=10)
  • KL Divergence: Regularization for latent space (λ=0.5)
Discriminator Loss
  • Real/Fake Loss: LSGAN (Least Squares GAN)
  • Multi-scale Loss: Aggregated across 4 scales

Training Progress

Training Progress Epoch 1

Epoch 1

Training Progress Epoch 2

Epoch 2

Training Progress Epoch 3

Epoch 3

Training Progress Epoch 4

Epoch 4

Training Progress Epoch 5

Epoch 5

Hyperparameters

Training Parameters
  • Epochs: 40+
  • Batch Size: 1
  • Learning Rate: 0.0001
  • Weight Decay: 6e-8
Model Parameters
  • Latent Dimension: 256
  • Filters: 16 (base)
  • Kernel Size: 3x3
  • Image Shape: 256x256x3

Performance Metrics

Quantitative Metrics
  • PSNR: 32.4 dB
  • SSIM: 0.91
  • FID Score: 18.7
Training Metrics
  • Generator Loss: ~1.5 (final)
  • Discriminator Loss: ~0.6 (final)
  • Cycle Consistency Loss: ~0.2