Enhanced DE-GAN with Dynamic Parameters and Global Positional Encoding

This implementation has two key improvements:

• Uses learned weight combinations to adaptively process features
• In DecoderDynamic: 4 different convolution outputs are weighted by learned parameters
• Includes MixStyle: style augmentation for domain robustness
• Helps the model learn adaptive feature combinations for different image characteristics

• Concatenates normalized (x, y) coordinate maps to feature tensors
• Provides explicit spatial information to the decoder
• Helps preserve spatial structure in generated images
• Particularly useful for medical image enhancement

models/
├── degan_generator_dynamic.py   # Generator
├── degan_discriminator.py       # Discriminator
networks/
├── unet.py                       # Contains ConvDDynamic, CoordConv2d

DEGANGeneratorDynamic
├── CalibrationBlock 
│   └── Channel-wise attention
└── DynamicUNetBlock 
    ├── ConvDDynamic encoders (layers 1-3)
    │   ├── ConvDDynamic: Dynamic conv + MixStyle
    │   ├── MixStyle: Adaptive instance normalization
    │   └── Learned weight combinations
    ├── Standard Conv2d encoders (layers 4-5)
    └── Decoder with CoordConv2d
        └── CoordConv2d: Global positional encoding

Create "data_adr.txt" file and determine requirement as bellow:

#############################################

[data]

data_root = datapath

data_names = config/train_name_all.txt

modality_postfix = [flair]

file_postfix = nii.gz

Put name of each Subject ID on "train_name_all.txt"

Test mode (CPU, 1 epoch):

python train_degan.py --device cpu --test

Full training (GPU):

python train_degan.py --epochs 400 --batch-size 8 --device cuda

After training, use the model for enhancement:

from models.degan_generator_dynamic import DEGANGeneratorDynamic
import torch

# Load model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = DEGANGeneratorDynamic().to(device)
model.load_state_dict(torch.load('path/to/data'))
model.eval()

# Inference
with torch.no_grad():
    input_tensor = torch.randn(1, 1, 128, 128).to(device)
    output = model(input_tensor)

Advantages:

• ✓ Adaptive feature processing
• ✓ Better spatial information awareness
• ✓ More robust to style variations
• ✓ Preserves original CalibrationBlock benefits

Computational Cost:

• ~10-15% slower than original (due to CoordConv2d + dynamic operations)
• Memory usage: slightly higher (~5-10%)

1. Start with test mode:

   python train_degan.py --device cpu --test

2. Monitor losses:

   • Generator loss should decrease
   • Discriminator loss should stay stable
   • If D_loss → 0 too quickly, generator is winning (reduce G learning rate)

3. Save checkpoints:

   • Models saved every 40 epochs
   • Load best checkpoint for inference

The dynamic components are based on:

• ConvDDynamic: Dynamic convolution with MixStyle augmentation
• CoordConv2d: Coordinate-based convolutions for spatial awareness
• MixStyle: Domain generalisation technique for medical imaging