TL;DR
Multishot MRI accelerates data acquisition but is highly sensitive to patient movement, which introduces severe motion artifacts into the final images. This paper introduces an end-to-end Generative Adversarial Network (GAN) framework that bypasses slow, traditional motion-estimation algorithms, correcting artifacts up to two orders of magnitude faster while delivering superior image fidelity.
What Problem Does This Paper Address?
While classical MRI can be slow, multishot Magnetic Resonance Imaging (MRI) speeds up the acquisition process by filling the "k-space" (the raw spatial frequency data) fractionally across multiple separate "shots." However, this makes the technique incredibly sensitive to patient movement. Even minor displacements from respiration or slight head shifts dislocate the k-space data between shots, creating disruptive motion artifacts that can lead to medical misdiagnosis.
Existing state-of-the-art motion correction methods rely on iterative algorithms to estimate and correct these motion states. Unfortunately, these methods are computationally expensive and painfully slow, making them impractical for real-time clinical workflows. The authors aim to eliminate this computational bottleneck using deep learning.
Methods
The authors propose a hybrid framework that pairs a standard parallel MRI reconstruction algorithm with a deep generative model:
- Initial Reconstruction: The motion-corrupted k-space data is first reconstructed into the image domain using the standard Conjugate Gradient Sensitivity Encoding (CG SENSE) algorithm.
- GAN-based Correction: The resulting artifact-heavy image is then passed through a Generative Adversarial Network (GAN) tasked entirely with removing the motion artifacts.
Network Architecture
- Generator ($G$): A U-Net-like convolutional architecture consisting of encoding and decoding blocks. Each encoder block contains 5 convolutional layers. The decoder blocks mirror this structure but replace standard convolutions with transposed convolutions to upsample the image back to its original resolution.
- Discriminator ($D$): Matches the exact architecture of the U-Net's encoder stage.
The network is trained using a standard minimax adversarial loss function, optimized via the RMSProp optimizer. The generator is also pre-trained to ensure faster, more stable convergence.
Datasets
Because no public datasets of motion-corrupted MR images exist, the authors synthetically simulated data using the BRATS 2015 dataset (a collection of motion-free brain tumor images).
To create realistic multishot artifacts, they transformed the image-domain data back into simulated k-space data, introduced varying degrees of sudden rotational motion (5°, 8°, 10°, 12°, and 14°) for individual shots, and then recombined the segments. The final dataset was split into:
- Training set: 13,019 images (70%)
- Testing set: 5,580 images (30%)
Key Insights and Ideas
- Implicit vs. Explicit Correction: Traditional methods try to explicitly calculate the exact mathematical parameters of the patient's physical movement. The GAN bypasses this entirely by implicitly learning the statistical mapping from "corrupted image space" to "clean image space."
- Graceful Degradation: While traditional iterative methods completely break down or fail to converge when head rotation exceeds 10°, the GAN's deep priors allow it to recover highly acceptable, clinically useful images even at extreme 14° rotations.
Key Results
The quantitative and qualitative improvements over the baseline iterative method (Cordero et al.) are stark:
- Massive Speedups: For a mild 5° motion artifact, the GAN approach corrects the image in just 0.22 seconds, compared to 25.23 seconds for the baseline—a 114x speedup. At a severe 14° rotation, the GAN takes 0.34 seconds while the baseline balloons to 152.03 seconds.
- Image Quality: The framework achieves a Peak Signal to Noise Ratio (PSNR) of 31.82 and a Structural Similarity Index (SSIM) of 0.95 for 5° rotations, maintaining a highly acceptable PSNR of 26.82 even at 14°.
As seen in the qualitative results, the GAN approach yields significantly sharper brain structures and fewer residual ghosting artifacts than previous approaches.
Limitations and Future Work
Despite the impressive speed and structural recovery, the authors note that at severe levels of motion degradation (e.g., 14°), the network tends to generate overly smooth images, losing some fine-grained textures due to the intensity of the corruption.
Additionally, this initial study was limited strictly to 2-shot MRI acquisitions. The authors plan to scale their GAN approach to handle a higher number of shots and conduct more extensive clinical validation in future work.
Summary updated