TL;DR
This paper introduces MedGAN, a novel Generative Adversarial Network (GAN) framework for retrospectively correcting both rigid and non-rigid motion artifacts in MRI scans. Unlike previous methods, MedGAN doesn't require external tracking systems or reference images, and it achieves state-of-the-art results in producing artifact-free images across different body regions.
What Problem Does This Paper Address?
Magnetic Resonance Imaging (MRI) is a vital tool in medical diagnostics, but its image quality is often compromised by patient motion during scans. This motion can manifest as blurring, aliasing, and stretching of anatomical structures, hindering accurate diagnosis. Existing solutions include prospective correction (intercepting motion during acquisition, which has safety and time limitations) and retrospective correction. While retrospective methods exist, they often rely on computationally expensive iterative algorithms or struggle with complex, non-rigid motion. This paper aims to address the limitations of current retrospective MRI motion correction techniques by developing a more efficient and effective deep learning-based approach.
Methods
The core of MedGAN is a conditional Generative Adversarial Network (cGAN) architecture. This framework consists of two main components:
- Generator (G): Takes a motion-corrupted MR image as input and outputs a denoised, artifact-free image. MedGAN employs a novel generator architecture, an extension of the U-Net, which concatenates three U-nets in an end-to-end manner (referred to as CasNet). This multi-stage approach allows for progressive enhancement of the corrected image.
- Discriminator (D): Acts as a binary classifier, distinguishing between real, artifact-free MR images and the synthetic images generated by G.
MedGAN goes beyond standard cGANs by incorporating non-adversarial losses to improve image quality:
- Perceptual Loss: This loss component uses the discriminator itself as a feature extractor. It calculates the mean absolute error (MAE) between intermediate features of the generated image and the ground-truth image, ensuring global consistency and perceptual quality.
- Style Reconstruction Loss: Inspired by neural style transfer, this loss uses a pre-trained VGG-19 network to capture and reconstruct the texture and fine details of the target images. It achieves this by minimizing the difference between the Gram matrices of the generated and target images.
The paper also introduces a joint correction model (MedGAN-joint) that is trained on data from all body regions simultaneously, aiming to leverage learned correlations between different anatomical areas.
Datasets
The study utilized a dataset from 17 anonymized volunteers, covering head, pelvis, and abdomen regions. For each volunteer, scans were acquired under resting (head, pelvis) and breath-hold (abdomen) conditions, as well as with induced rigid motion (head/hips) and non-rigid motion (respiratory). 2D slices were extracted and scaled to 256x256 pixels.
The training dataset comprised:
- Rigid motion artifacts (head/pelvis): 980 paired images (motion-free and corrupted) from 14 patients, with 105 validation images from 3 patients.
- Non-rigid motion artifacts (abdomen): 420 paired trained images from 14 patients, with 90 validation images from 3 patients.
The authors compared their method against pix2pix and ID-cGAN, which were trained for single-body region correction.
Key Insights and Ideas
- GANs for Retrospective Motion Correction: The paper demonstrates the efficacy of GANs, particularly with the addition of perceptual and style losses, for tackling complex MRI motion artifacts retrospectively.
- Joint Correction Advantage: Training a single model (MedGAN-joint) on data from multiple body regions simultaneously leads to improved results, suggesting that learned correlations between different anatomical areas enhance the correction process.
- Beyond Adversarial Loss: The inclusion of non-adversarial losses (perceptual and style) is crucial for preserving fine details and global consistency, which are often lost in standard GAN-based image translation tasks.
- No Reference Image Needed: A significant advantage is the ability to correct motion artifacts without requiring a corresponding motion-free reference scan, simplifying the workflow.
Key Results
- Quantitative Performance: MedGAN and MedGAN-joint consistently outperformed pix2pix and ID-CGAN across various quantitative metrics, including SSIM, VIF, UQI, and LPIPS, for head, abdomen, and pelvis motion correction. For instance, in head motion correction, MedGAN-joint achieved the highest SSIM (0.8314) and UQI (0.6571).
- Qualitative Performance: Visual comparisons showed that MedGAN produced sharper, more homogeneous, and globally consistent images compared to pix2pix (which lacked consistency) and ID-CGAN (which resulted in blurred images). MedGAN-joint further enhanced these qualities, producing images that closely resembled the target motion-free scans.
- Comparison to State-of-the-Art: MedGAN demonstrated superior performance compared to established GAN-based methods like pix2pix and ID-cGAN.
Limitations and Future Work
- Diagnostic Quality: The authors acknowledge that while MedGAN produces near-realistic images, it is not yet intended for direct diagnostic use. Further investigation is needed to ensure that relevant diagnostic information is not lost during the correction process.
- Complex-Valued Data: The current framework primarily uses magnitude information from MR images. Future work will explore extending the model to handle complex-valued data (including phase information) for potentially improved results.
- 3D Correction: The current study focuses on 2D slices. Extending the model to 3D space is a planned future direction.
- Subjective Evaluation: The authors plan to conduct further subjective evaluations with experienced radiologists to assess the diagnostic utility of the corrected images.
- Applications: The authors suggest that MedGAN could be valuable for post-processing tasks such as segmentation, organ volume estimation, and attenuation correction for PET/MR scans.
Summary updated