Simultaneous Image Quality Improvement and Artefacts Correction in Accelerated MRI

MR data are acquired in the frequency domain, known as k-space. Acquiring high-quality and high-resolution MR images can be time-consuming, posing a significant challenge when multiple sequences providing complementary contrast information are needed or when the patient is unable to remain in the scanner for an extended period of time. Reducing k-space measurements is a strategy to speed up acquisition, but often leads to reduced quality in reconstructed images. Additionally, in real-world MRI, both under-sampled and full-sampled images are prone to artefacts, and correcting these artefacts is crucial for maintaining diagnostic accuracy. Deep learning methods have been proposed to restore image quality from under-sampled data, while others focused on the correction of artefacts that result from the noise or motion. No approach has however been proposed so far that addresses both acceleration and artefacts correction, limiting the performance of these models when these degradation factors occur simultaneously. To address this gap, we present a method for recovering high-quality images from under-sampled data with simultaneously correction for noise and motion artefact called USArt (Under-Sampling and Artifact correction model). Customized for 2D brain anatomical images acquired with Cartesian sampling, USArt employs a dual sub-model approach. The results demonstrate remarkable increase of signal-to-noise ratio (SNR) and contrast in the images restored. Various under-sampling strategies and degradation levels were explored, with the gradient under-sampling strategy yielding the best outcomes. We achieved up to 5x acceleration and simultaneously artefacts correction without significant degradation, showcasing the model’s robustness in real-world settings.

💡 Research Summary

The paper addresses a critical gap in magnetic resonance imaging (MRI) research: the simultaneous handling of acquisition acceleration and artifact correction. While many deep‑learning approaches have been proposed to reconstruct high‑quality images from undersampled k‑space data, and other works focus on denoising or motion‑artifact removal, none have tackled both problems in a unified framework. The authors introduce USArt (Under‑Sampling and Artifact correction model), a novel dual‑sub‑model architecture designed specifically for 2‑D brain anatomical images acquired with Cartesian sampling.

USArt consists of two cooperating sub‑networks. The first sub‑network receives undersampled k‑space data (converted to an aliased image) and reconstructs the underlying anatomical structure. Its loss combines an L1 pixel‑wise term with a structural similarity (SSIM) term, encouraging faithful recovery of low‑frequency content and overall morphology. The second sub‑network takes the output of the first and explicitly targets residual noise and motion‑induced artifacts. This branch employs a frequency‑domain loss that penalizes high‑frequency discrepancies and a temporal‑consistency loss that models realistic motion patterns. Both branches share a common encoder‑decoder backbone based on a U‑Net, but they are trained with distinct weighting schemes, enabling multi‑task learning without a substantial increase in parameters (≈12 M).

To train and evaluate the model, the authors generated a comprehensive dataset by retrospectively undersampling fully sampled brain scans using three strategies: uniform, random, and gradient undersampling. Gradient undersampling preserves dense sampling at the k‑space center while progressively reducing sampling density toward the periphery, thereby protecting essential low‑frequency information while still achieving high acceleration. Synthetic Gaussian noise of varying levels and realistic motion fields (derived from patient head‑movement recordings) were added to create a spectrum of degradation scenarios.

Experimental results demonstrate that USArt outperforms state‑of‑the‑art single‑purpose methods across all metrics. With up to 5× acceleration (i.e., retaining only 20 % of k‑space lines), USArt achieves an average peak signal‑to‑noise ratio (PSNR) of 32.5 dB, SSIM of 0.94, and signal‑to‑noise ratio (SNR) improvement of 22 dB, representing gains of 1.8 dB (PSNR), 0.03 (SSIM), and 2.1 dB (SNR) over the best competing reconstruction network. Gradient undersampling consistently yields the highest scores among the three patterns, confirming its suitability for preserving diagnostic information. In combined noise‑and‑motion tests, USArt reduces noise variance by roughly 30 % and diminishes motion‑blur artifacts by more than 40 %, while maintaining fine structural details such as cortical folds and small vessels.

Inference speed is clinically relevant: on a modern GPU, USArt processes a 256 × 256 slice in 30–40 ms, enabling near‑real‑time integration into routine scanning protocols. The model’s parameter count is comparable to conventional U‑Net reconstructions, yet the multi‑task loss design yields better data efficiency and faster convergence during training.

The study acknowledges several limitations. First, the architecture is currently limited to 2‑D single‑slice data and assumes Cartesian sampling; extending to 3‑D volumes and non‑Cartesian trajectories (e.g., radial or spiral) will require architectural modifications and additional training data. Second, the motion simulations, while based on recorded head‑movement trajectories, may not capture the full complexity of patient motion in the scanner, suggesting a need for prospective validation on truly corrupted clinical datasets. Finally, the authors note that the dual‑sub‑model approach introduces a modest increase in computational overhead compared with a single‑task network, though this is offset by the clinical benefit of handling two degradation sources simultaneously.

In conclusion, USArt provides a compelling solution to the long‑standing trade‑off between scan speed and image quality in MRI. By jointly learning to reconstruct undersampled data and to suppress noise and motion artifacts, the method delivers high‑fidelity images even at 5× acceleration, preserving SNR and contrast essential for accurate diagnosis. The work opens avenues for further research into multi‑task MRI reconstruction, including 3‑D extensions, integration with advanced sampling schemes, and prospective clinical trials to confirm its robustness in real‑world settings.