DP-NSL: Dual-Prior Guided Null-Space Learning with Mixture-of-Splines for Arbitrary Medical Slice Super-Resolution (ECCV2026)

Haofei Song, Siyuan Xu, Xintian Mao, Shaojie Guo, Qingli Li, Yan Wang*

Paper Link: arxiv

Environment

Install PyTorch following the official instructions for your CUDA version, then install the remaining dependencies:

pip install -r requirements.txt

Data Preparation

The data preprocessing protocol follows the slice-based setting used by I3Net. For the original preprocessing reference, please see the data_prepare/ folder in the I3Net repository.

This repository already includes a small debug set under data/, so users can run the training and testing scripts without preparing the full dataset first.

data/
  data_slice/
    Task10_Colon/
      imagesTr/
        colon_001/
          colon_001_slice_000.pt
          colon_001_slice_001.pt
          ...
  data_volume/
    Task10_Colon/
      imagesTs/
        colon_018.pt

For full experiments, prepare your dataset with the same directory structure, then update the paths in config.py, yaml/dpnsl/dpnsl.yaml, or the command line arguments as needed.

Checkpoint

An official checkpoint is provided at:

ckpt/1000.pth

Configuration

The default model configuration is:

yaml/dpnsl/dpnsl.yaml

Quick Run

Evaluate the provided checkpoint on the included debug test volume:

bash test.sh

Run training on the included debug training slices:

bash train.sh

For a shorter smoke test, you can run one epoch directly:

python -m torch.distributed.launch \
    --nproc_per_node=1 \
    --master_port 12345 main.py \
    --config_yaml=yaml/dpnsl/dpnsl.yaml \
    --upscale 2,3,4 \
    --upscale_test 2,3,4 \
    --ckpt_dir debug_train \
    --batch_size 1 \
    --num_workers 4 \
    --lr 1e-4 \
    --test_period=1 \
    --save_ckpt_period=1 \
    --amp \
    --kwargs max_epoch=1

Training logs and checkpoints are written under:

log/
experiments/

Repository Structure

.
├── main.py                  # training entry point
├── test.py                  # evaluation entry point
├── config.py                # CLI and default config
├── data.py                  # dataset loader
├── val.py                   # validation / evaluation loop
├── model_zoo/
│   ├── dpnsl/               # DP-NSL model
│   └── encoder/             # encoder backbones
├── data_prepare/            # preprocessing notes, adapted from I3Net
├── yaml/dpnsl/dpnsl.yaml    # model config
├── train.sh                 # training launcher
└── test.sh                  # evaluation launcher

Citation

If this repository is useful for your research, please cite:

@article{
  coming soon
}

Contact

If you have any question, please contact hfsong@stu.ecnu.edu.cn

---

More Open-Source Projects

Tip

Interested in more open-source methods for medical image reconstruction? Visit our Medical Image Reconstruction project hub for a collection of papers, source code, and available resources from DeepMed Lab.

View all open-source medical image reconstruction projects

​​