Each data point in k-space corresponds to the frequency content of the image, and the alteration of a single point in k-space influences the whole image. MR data acquisition proceeds in its Fourier domain, i.e., k-space. We then propose a new motion-simulation tool, called view2Dmotion, which can generate massive training datasets for motion simulation. In particular, we pay attention to the methods used for network training. In this paper, we review previous studies for motion-correction methods based on deep learning.
Thus, network training requires new data-generation methods to consider the motion artifacts generated for various reasons. However, since these previous studies considered only simple and fixed motion patterns, applications to various motion artifacts from patients whose motions are unexpected to have limitations ( 31). To generate motion-corrupted images, motion simulations in the image domain ( 29, 30, 31, 32) or k-space domain ( 33, 34, 35, 36) are essential. In this process, a neural network is trained with an enormous training dataset, which typically consists of motion-corrupted images as an input and motion-clean images as a label. Also, it has been suggested to use deep learning for the motion correction as an alternative method, since it does not require any additional scan time, motion-tracking systems, or alteration of sequence. Recently, deep learning has shown remarkable acceptability in MR image processing ( 28). Overall, these conventional methods incur additional cost, such as from prolonged scan time ( 30, 31) and modification of sequences. However, the computational limits are evident because of the complex and unpredictable patient motions in this case ( 28, 29). In addition, these retrospective methods could be executed with algorithms estimating motion without acquiring motion information. Using trajectory information, these prospective methods compensate for or reacquire the k-space partially during data acquisition ( 3, 12, 15, 16, 17, 18, 19, 20, 21, 22, 23), although retrospective methods, which process the data after the MR acquisition ( 4, 24, 25, 26, 27), could also be considered for the motion correction. In terms of MR systems, MR navigators ( 9, 10, 11, 12, 13), which are additional pulses to track motion or motion-tracking systems, such as an in-bore camera and markers ( 14, 15, 16, 17, 18), are generally used to measure the patient's motion. This method can reduce scan times with fewer phase encodings during MR scanning, but it is still affected by artifacts from patients who cannot control their behavior. However, since this method asks the patient to actively control motions, there was also an effort to use the parallel imaging technique in order to reduce the burden on the patient ( 7, 8). The basic straightforward method is to make restrain patients' motions by means of sedation or breath-holding ( 5, 6). To address motion-related problems, there have been many methods to prevent motion or to correct artifacts. Thus, the subject motion is an important factor that has been continuously considered in MR imaging. Most importantly, in clinical settings, these motion artifacts may affect the diagnosis for patients who cannot control their movements. Additionally, technological developments in MR hardware have improved spatial resolution and signal-to-noise ratio (SNR), but can be more susceptible to motion artifacts because of extended scan times ( 1, 3, 4). The subject's motion creates incorrect allocation values of the k-space signal from MRI machines and results in blurring or ghosting artifacts ( 1, 2), which can be typically observed in the reconstructed images. However, MRI is sensitive to movement of the subject because of the long acquisition time, which causes motion artifacts ( 1).
Magnetic resonance imaging (MRI) is a non-invasive soft-tissue mapping technique, which is widely used to diagnose various diseases without radiation exposure.
0 kommentar(er)
