GD-CSRNet

Overview

GD-CSRNet is a PyTorch-based multi-contrast MRI super-resolution training framework. The core model is the Gated Delta Cross-modal Structure Refinement Network (GD-CSRNet), designed for super-resolution reconstruction and quality enhancement of multi-modal MRI images.

---

Repository Structure

File / Directory	Description
train.py	Main entry point: loads configs, prepares data, builds model, runs train/val/test
engine.py	Training loop, validation & test logic; supports distributed training, model selection, and checkpoint management
utils.py	Utility functions: logging, optimizer, scheduler, and random seed setup
configs/config_setting.py	Default training config, model hyperparameters, data paths, and preprocessing transforms
datasets/dataset.py	BraTs_datasets class supporting BraTs2020 multi-contrast MRI loading and clinical degradation augmentation
model/gated_delta_sr.py	Main GD-CSRNet model definition
model/DGCFBlock.py	Key network module implementation
model/SDDFBlock.py	Key network module implementation

---

Requirements

Python 3.8+ is recommended. Install the main dependencies via:

pip install torch torchvision tensorboardX monai scipy SimpleITK medpy pillow opencv-python

For distributed training, ensure your PyTorch installation supports torch.distributed.

---

Dataset Format

The default configuration uses the BraTs2020_t1_t2 dataset. The expected directory structure is:

datasets/BraTs2020_t1_t2/
├── train/
│   ├── oriT2/          # High-resolution target images
│   ├── oriT1/          # Reference images
│   ├── orLRbicT1/x4/   # Low-resolution reference images
│   └── orLRbicT2/x4/   # Low-resolution target images
└── val/
├── oriT2/
├── oriT1/
├── orLRbicT1/x4/
└── orLRbicT2/x4/

The data loader in datasets/dataset.py expects four input types:

• oriT2 — High-resolution target image
• oriT1 — Reference image
• orLRbicT1/x4 — Low-resolution version of the reference image
• orLRbicT2/x4 — Low-resolution version of the target image

---

Getting Started

Run training with default settings:

python train.py

All configuration is read from configs/config_setting.py. To change the network architecture, dataset, or training hyperparameters, edit that file directly.

---

Default Configuration

Parameter	Default Value
network	'gdcsr'
datasets	'BraTs2020_t1_t2'
batch_size	4
epochs	150
num_gpus	2
enable_logging	True
work_dir	results/{network}_{datasets}_x{upscale}/

---

Model Architecture

The framework supports multiple network architectures selectable via config.network in train.py. The default is gdcsr.

Model files included in this repository:

• model/gated_delta_sr.py — Full GD-CSRNet definition
• model/DGCFBlock.py — Gated Delta Cross-modal Fusion Block
• model/SDDFBlock.py — Structure-guided Dual-domain Fusion Block

---

Training Outputs

All outputs are saved under work_dir/:

work_dir/
├── log/           # Training log files
├── checkpoints/   # Saved model weights
│   ├── latest.pth # Most recent checkpoint
│   └── best.pth   # Best checkpoint (by validation PSNR)
├── summary/       # TensorBoard logs
└── outputs/       # Optional output results

---

License

This project is released under the MIT License.