RISE-MAR: Radiologist-in-the-Loop Self-Training for Generalizable CT Metal Artifact Reduction

This is the official repository for the paper "Radiologist-in-the-Loop Self-Training for Generalizable CT Metal Artifact Reduction". [arxiv, tmi]

Highlights

• RISE-MAR Framework: A semi-supervised learning framework that ensures high-quality pseudo groundtruths for clinical CT metal artifact reduction (MAR).
• Clinical Quality Assessor (CQA): A pretrained model that evaluates the quality of CT images potentially affected by metal artifacts

Project Status

• Preprint released
• Core model implementation
• Training code available
• CQA model weights

Implementation Guide

To implement the full version of RISE-MAR, follow these steps:

1. Data Preparation

1. Download relevant CT datasets
2. Simulate metal artifact-affected data to construct paired data
3. Organize data using JSON metadata files

Paired Data Example (please see ./data/meta/example_paired_deepl.json):

{
    "case_001": {
        "metal_img": "path/to/artifact_affected_image",
        "li_img": "path/to/LI_corrected_image",
        "gt_img": "path/to/artifact_free_ground_truth",
        "metal_mask": "path/to/metal_mask",
        "root_dir": "optional/path/prefix" 
    },
    // Additional cases...
}

UnPaired Data Example (please see ./data/meta/example_unpaired_deepl.json):

{
    "ma": {  // Metal artifact-affected images
        "case_001": {
            "img": "path/to/artifact_affected_image",
            "quality": 5,  // For CQA model training, optional
            "root_dir": "optional/path/prefix"
        },
        // Additional cases...
    },
    "mf": {  // Metal artifact-free images
        "case_001": {
            "img": "path/to/artifact_free_image",
            "quality": 9,
            "root_dir": "optional/path/prefix"
        },
        // Additional cases...
    }
}

After preparation, browse ./configs/__init__.py and update the paths for the metadata.

2. Baseline Model Training

We refer to "baseline models" as models that produce imperfect prediction for clinical MAR. These models can be obtained from:

• Some conventional methods (e.g., LI, NMAR, ...)
• Insufficient training (e.g., early stop before model convergence)
• Models with limited domain transferability (e.g., can handle simulated metal artifacts very well but perform poor MAR on real data)
• ...

Turns out that they are still useful! In our work, they:

• serve as undertrained MAR models that provide CT images with low-to-moderate quality, greatly enhancing the diversity of our clinical quality assessment dataset.
• provide better network initialization, which often leads to a more stable training process for unsupervised or semisupervised methods.
• serve as the prior model, i.e., teacher network in our paper. A stronger prior model certainly perform better MAR, but with our RISE-MAR framework, we can still employ an imperfect one while expecting satisfactory MAR results ;)
• ...

As an example of training script, please run:

bash scripts/train_supervised_mar.sh

Once the baseline models are prepared, configure related paths (UNDETRAINED_WEIGHTS and PRETRAINED_*) in ./configs/__init__.py.

3. Clinical Quality Assessor (CQA) Training

Adjust the arguments in scripts/train_cqa.sh and run the training script:

bash scripts/train_cqa.sh

Update the path to the CQA model weights (PRETRAINED_CQA_PATH) in ./configs/__init__.py after training.

4. MAR Model Training

Adjust the arguments in scripts/train_risemar.sh and run the training script:

bash scripts/train_risemar.sh

Citation

If you find our work and code helpful, please kindly cite our paper 💙

@ARTICLE{ma2025risemar,
  author={Ma, Chenglong and Li, Zilong and Li, Yuanlin and Han, Jing and Zhang, Junping and Zhang, Yi and Liu, Jiannan and Shan, Hongming},
  journal={IEEE Transactions on Medical Imaging}, 
  title={Radiologist-in-the-Loop Self-Training for Generalizable {CT} Metal Artifact Reduction}, 
  year={2025},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TMI.2025.3535906}
}