How eeeric-code/i3net works
Project type and purpose
• Type: Deep-learning research repository implementing the I³Net model for medical slice interpolation (CT & MR).
• Purpose: Train and evaluate an inter-intra-slice interpolation network that synthesizes missing 2D slices in 3D medical volumes.
High-level components
A. Data Acquisition & Preparation
B. Data Loading & Configuration
C. Model Definition (Generator & Discriminator)
D. Training & Optimization
E. Evaluation & Experiments Storage
Detailed breakdown & relationships
A. Data Acquisition & Preparation (green)
• Raw datasets: Medical Decathlon (Liver, Colon, Hepatic Vessel), IXI MR
• data_prepare/step1_split.py → splits volumes into train/test using test_set.pt
• data_prepare/step2_save_slice.py → writes individual slices to disk
→ Output: folder of 2D slice images (.pt or numpy)
B. Data Loading & Configuration (light blue)
• config.py: global hyperparameters (batch size, learning rate, paths)
• opt/i3net.json: model-specific options (number of slices to interpolate, DCT settings)
• data.py & util.py: PyTorch Dataset & DataLoader modules that read saved slices
• select_model.py: chooses I3Net implementation from model_zoo
• select_loss.py: picks loss functions from losses/
C. Model Definition (blue)
• model_zoo/i3net/basic_model.py: main Inter-Intra-slice U-Net architecture
• model_zoo/i3net/dct_util.py + utils_win.py: DCT transforms & sliding windows for intra-slice features
• losses/ (yellow): adversarial.py, contrastive_loss.py, edge_loss.py, focal_frequency_loss.py, vgg.py, discriminator.py
→ Generator network outputs synthesized slices
→ Discriminator network for adversarial training
D. Training & Optimization (orange)
• main.py: orchestrates training loop
• Loads config & opt JSON
• Instantiates DataLoader, model, losses, optimizer
• For each batch: forward pass → compute each loss → backward → optimizer.step()
• Writes checkpoints via train.sh wrapper and optim.py
• optim.py: builds Adam optimizers for generator & discriminator
• train.sh/test.sh: bash scripts that launch main.py/test.py with appropriate flags
• test.py: loads a checkpoint, runs inference on testset, computes PSNR/SSIM via util_evaluation.py
• experiments/i3net/...: folder hierarchy storing .pth checkpoints for different ×2/×4/×6 upsampling settings
• plots or metrics written to disk
Architectural patterns & design principles
• Modular separation: data_prepare, dataloading, model_zoo, losses
• Configurable via JSON + Python config
• Training loop decoupled from model/loss definitions
• Plugin-style selection: select_model.py & select_loss.py
[ Raw Data ] → [ Data Preparation ] → [ DataLoader / Config ] → [ Model + Loss Modules ] → [ Training Loop / Optimizer ] → [ Checkpoints & Evaluation ]
e.g., raw volumes → split → slices → DataLoader → model → loss → optimizer → checkpoint
• Data: green
• Config/Loading: cyan
• Model/Loss: blue/yellow
• Training: orange
• Storage: gray