Unsupervised MRI motion artifact disentanglement: introducing MAUDGAN

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Motion artifacts are a significant challenge in Magnetic Resonance Imaging (MRI), potentially compromising diagnostic accuracy, especially in patients with brain tumors.
• Existing artifact reduction methods often require supervision or fail to adequately preserve crucial image details in multi-parametric MRI sequences.

Purpose of the Study:

• To develop and validate an unsupervised deep learning framework for reducing motion artifacts in contrast-enhanced T1-weighted (ceT1W) and T2-FLAIR MRI images of brain tumor patients.
• To compare the proposed method against established techniques like Pix2pix and CycleGAN in terms of quantitative metrics and qualitative assessment.

Main Methods:

• A novel framework comprising two generators, two discriminators, and two feature extractor networks was designed.
• The model was trained using 3-fold cross-validation on 230 brain tumor MRI datasets and tested on 148 in-vivo patient datasets.
• Performance was evaluated using metrics such as NMSE, SSIM, PSNR, VIF, and MS-GMSD, alongside qualitative assessments by expert evaluators.

Main Results:

• The proposed unsupervised method demonstrated superior performance in reducing heavy motion artifacts, achieving the lowest NMSE and MS-GMSD compared to Pix2pix and CycleGAN.
• Generated images exhibited enhanced quality with the highest SSIM, PSNR, and VIF values, indicating effective artifact removal and preservation of image integrity.
• Qualitative evaluations confirmed the method's effectiveness, with significantly higher Likert scale scores for artifact reduction and image quality compared to other models.

Conclusions:

• The developed unsupervised method effectively reduces motion artifacts in multi-parametric brain MRI scans of tumor patients.
• This approach offers a promising solution for improving the diagnostic quality of MRI scans affected by patient motion, particularly in challenging clinical scenarios.
• The framework's ability to handle various artifact levels and its consistent performance highlight its potential clinical utility.