Skip to content

README.md

Latest commit

 

History

History
106 lines (74 loc) · 3.27 KB

README.md

File metadata and controls

106 lines (74 loc) · 3.27 KB
Raw

EDA — CT Metal Artifact Reduction (MAR)

Official implementation of metal artifact reduction for the paper:

Elucidating the Design Space of Arbitrary-Noise-Based Diffusion Models (EDA)
Xingyu Qiu, Mengying Yang, Xinghua Ma, Dong Liang, Fanding Li, Gongning Luo, Wei Wang, Kuanquan Wang, Shuo Li
CVPR 2026

MAR is Task 2 in EDA: it removes metal streak artifacts from CT by diffusing sample-dependent basis noise h = X_ma − x_0 (EDA Case 2, η = 10), with a metal-aware weighted loss.

Acknowledgement

This module is built on top of the Denoising Diffusion Implicit Models (DDIM) PyTorch codebase. We thank Jiaming Song, Chenlin Meng, and Stefano Ermon for releasing their implementation.

Original DDIM repository: https://github.com/ermongroup/ddim

Method Summary (Paper Settings)

Item Setting
Basis Sample-dependent h = X_ma − x_0
EDA case Case 2 — sample-dependent basis
η (training noise) 10: N = (10 + ε) / 11 · h, ε ~ N(0,1)
Forward process s(t) = 1, σ(t) = √(1 − ᾱ_t)
Loss Weighted MSE with metal mask m (Appendix E.3)
Input size 416 × 416, grayscale
Optimizer Adam, lr = 1×10⁻⁵, EMA = 0.9999
Training 1000 epochs, batch size 1
Sampling 5-step deterministic Euler, η = 0

Repository Layout

MAR/
├── main.py                 # Entry point (train / sample)
├── configs/mydataset.yml   # MAR hyperparameters
├── runners/diffusion.py    # EDA MAR training & sampling
├── functions/
│   ├── losses.py           # Weighted noise-prediction loss
│   └── denoising.py        # Deterministic sampling steps
├── datasets/
│   └── data_loading.py     # MARTrainDataset (pre-simulated h5 pairs)
└── models/diffusion.py     # UNet denoiser (from DDIM)

Requirements

This module is built on DDIM. Install Python, PyTorch, and other dependencies according to the baseline repository:

Baseline GitHub: https://github.com/ermongroup/ddim

Data Preparation

Training and testing use pre-simulated SynDeepLesion pairs stored as HDF5 files:

data/SynDeepLesion/
├── trainmask.npy
├── train_640geo_dir.txt
├── test_640geo_dir.txt
├── train_640geo/          # GT *.h5 and ma_CT *.h5
└── test_640geo/

Each volume has paired image (GT) and ma_CT (metal-affected) arrays. Set data.data_root in configs/mydataset.yml or pass --data_root at runtime.

Usage

Train

python main.py --config mydataset.yml --exp exp --doc mar_eda --ni

Sample

python main.py --config mydataset.yml --exp exp --doc mar_eda --sample --sequence --timesteps 5 --ni

See arg.txt for additional command examples.

Citation

@inproceedings{qiu2026eda,
  title={Elucidating the Design Space of Arbitrary-Noise-Based Diffusion Models},
  author={Qiu, Xingyu and Yang, Mengying and Ma, Xinghua and others},
  booktitle={CVPR},
  year={2026}
}