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.
Magnetic Resonance Imaging (MRI) provides detailed anatomical and functional information on soft tissues by collecting raw data in k-space [1]. The MRI scan time is influenced by the number of phase encoding steps required to reconstruct an image. Increasing the resolution or quality of an image typically requires more phase encoding steps, leading to longer scan times. This poses challenges, especially for patients who struggle to remain still, such as children and people with claustrophobia or uncontrolled movements disorder. Reducing scan time improves patient comfort and lowers medical costs by increasing throughput, but it can also reduce image quality [2][3][4].
Accelerating MRI acquisition has been a major focus in the field [5][6][7][8][9][10][11][12], leveraging both physics-driven and data-driven strategies to reduce scan times and improve image quality. Physics-driven methods like parallel imaging (SENSE, GRAPPA), compressed sensing (CS), and Echo Planar Imaging (EPI) exploit physical principles to decrease acquisition time but can introduce artefacts like noise amplification, residual aliasing, and Nyquist ghosting. Data-driven approaches, particularly deep learning with convolutional neural networks [13] (CNNs) or autoencoders, such as U-net [14], have also been proposed to predict missing data resulting from various under-sampled k-space data acquisition strategies, offering robust image reconstruction.
In real-world MRI, signal acquisition is also subject to various degradation factors that cause artefacts, which are undesired and unreal information that appear in the reconstructed image. Although full sampling may theoretically provide the most complete data, it is still susceptible to motion, noise and other imperfections inherent in the imaging process. These issues distort anatomical structures, introduce false information or cause signal loss, compromising diagnostic accuracy. Therefore, in the pursuit of improving image quality through acceleration techniques, it’s crucial to not only address artefacts arising from under-sampling but also to carefully manage and mitigate additional artefacts inherent to imperfect acquisitions settings that may be further amplified by the under-sampling process.
In the present paper, we present a new method to restore quality in images reconstructed from under-sampled MRI data acquisitions. We propose a neural network model called USArt (Under-Sampling and Artifact correction model) that restores missing k-space data and simultaneously corrects for motion and noise related artefacts. By addressing both under-sampling and artefacts correction, we aim to enhance the overall quality and accuracy of real-life fast MRI methods, thus contributing to more reliable and effective image reconstruction techniques. Our approach involves a dual-model framework, featuring one model operating in the k-space domain and another in the image domain, inspired by prior work that exploits the different characteristics of these domains [7,8,12]. We examined different under-sampling strategies, acceleration factors and artefacts. This project focuses on single-channel coils and Cartesian sampling for simplicity; hence, parallel MRI is not discussed.
In vivo 2D T2w anatomical images of mouse brain with tumors were acquired according to established protocols [15,16]. Images acquisitions used a Cartesian sampling trajectory and were performed on a preclinical 3T MRI system (MRSolutions, UK) equipped with a quadrature head coil. The dataset consists of 5649 complex-valued images from 224 different mice. The train/validation dataset includes 5000/449 images from 204/30 subjects. The remaining 200 images from 10 different subjects were used to test the performance of the trained networks. There was no overlapping of the same subjects or images in the different datasets.
Under-sampling was achieved by retrospectively dropping lines in fully acquired Cartesian k-spaces, corresponding to phase-encoding steps. Masks were for this purpose applied to selectively zero out lines in the k-space using one of three strategies: gradient, random, and uniform under-sampling. The gradient undersampling mask progressively reduces the number of lines acquired as the trajectory moves away from the k-space center. This favors low frequency k-space information that determines contrast, brightness, and general shapes, over high frequencies that pertain to edges and sharp transitions. Random under-sampling selects the retained lines randomly. Uniform under-sampling uniformly discard lines in the k-space, without targeting specific regions or frequencies. Three under-sampling acceleration factors were used: 2×, 5×, and 10×. For all undersampling strategies, we additionally retained some low-frequency lines in order to reduce the aliasing artefact; 25%, 10%, and 4% of the k-space’s lines for 2×, 5×, and 10× accelerations respectively.
Artefacts in real-world acquisitions were simulated in the k-space
This content is AI-processed based on open access ArXiv data.
