Reduction of respiratory motion artifacts in gadoxetate-enhanced MR with a deep learning–based filter using convolutional neural network

Recent research demonstrated intravenous bolus injection of gadolinium-based contrast agents to be accompanied by motion-related image degradation in the arterial phase [1‐3]. This phenomenon is temporary and self-limited, wherefore the term “transient severe motion” (TSM) is often used. Although the underlying pathomechanism has not yet been conclusively solved, the incidence of these artifacts, ranging from 8 to 20% [1, 2, 4, 5], was found to be significantly higher after gadoxetate disodium administration compared to other contrast media [1, 2, 4]. TSM artifacts are of high clinical significance, as the arterial phase is essential for lesion characterization [6, 7]. Artifact reduction during image acquisition or afterwards, therefore, would contribute to the imaging-based diagnosis in the clinical setting.

In order to solve this problem, several strategies are conceivable. One potential solution can be patient-based: through informing patients concerning the problem of acute transient dyspnea or performing breath-hold practice before the scan, the rate of artifacts could be decreased [8]. Likewise, dilution of the contrast agent is able to reduce the incidence of artifacts [9, 10]. Another strategy can be a modulation of the data acquisition: fast scanning techniques leading to a shorter examination time consequently minimize the risk of patient motion. Compressed sensing is now widely used in clinical practice to shorten the acquisition time by the use of undersampled k-space data without compromising signal-to-noise ratio [11, 12]. Other studies focused on respiratory triggering, where arterial phase images are acquired with the help of respiratory tracings or navigator echoes [13]. The use of single-breath-hold multi-arterial phase acquisition provides adequate well-timed late hepatic arterial phase images in most patients even with transient severe motion [2, 14, 15]. Interestingly, Min et al [16] showed a similar incidence of TSM artifacts with multiple arterial phases using view sharing from two different vendors and conventional single arterial phase in a retrospective study.

All these approaches, however, have to be prepared before the scan at a time when we do not know if the individual patient will fail their breath-hold, and therefore turn out time-consuming and disrupt the clinical workflow. Postprocessing artifact removal allows for standardized MR protocols and does not require any change in sequence acquisition. On this field, deep learning methods show emerging applications in medical image reconstruction and related reduction of artifacts. In fact, a variety of studies have been published on the field of reconstruction [17‐19]. However, in terms of artifact reduction in MRI, there are only a few published papers addressing motion artifacts [20‐22]. Especially, the clinical utility of respiratory motion artifact reduction has not been studied so far.

Therefore, the purpose of our study was to reveal the utility of respiratory motion artifact reduction with convolutional neural network (MARC) in combination with advanced acquisition method of single-breath-hold multi-arterial phase acquisition (DISCO) in gadoxetate disodium–enhanced MRI of the liver.

The retrospective study was approved by the local institutional review board. The requirement to obtain patients’ informed consent was waived.

A study coordinator searched the electronic database for abdominal MR examinations under administration of gadoxetate disodium and found 192 consecutive examinations between March and June 2017. No examination had to be excluded. The study population consisted of 131 men and 61 women with a mean age of 68.7 ± 10.3 years (range 29–89 years).

With this retrospective analysis, we revealed that postprocessing using the MARC filter improved image quality in terms of respiratory motion artifacts in multiple arterial phase image acquisition with gadoxetate disodium.

Respiratory motion artifacts are a major problem in MR imaging of the abdomen. As described in previous studies, gadoxetate disodium–related transient severe motion degrades the quality of arterial phase images [1‐4, 14], resulting in decreased diagnostic accuracy, particularly for the detection and characterization of focal hepatic lesions [6, 7]. Among the multitude of strategies proposed, especially fast scanning techniques using compressed sensing offer a simple approach to minimize motion artifacts [11, 12]. Pietryga et al [2] showed that the use of multiple arterial phase acquisition within a rapid single breath-hold provides adequate well-timed late hepatic arterial phase images in most patients with TSM. In general, multiple arterial phase acquisition with view sharing technique can improve the time resolution. However, it often requires longer time for breath-hold. In our study, the acquisition time was relatively long, which probably led to more artifacts in the last 1–2 phases. In the meantime, most cases had phases without artifact in the front half, where appropriate late arterial phase images were included. These images without artifact can be used as a reference of anatomical details when convolutional neural network (CNN) works for artifact reduction. Taking these findings into account, in our study, we developed a deep learning filter based on CNN, which works in combination with multi-arterial phase acquisition using DISCO to achieve motion artifact removal in MRI of the liver.

Deep learning methods recently have shown promising performances on various vision tasks, which also include fields of importance for medical imaging. Deep learning is based on a neural network with a high number of hidden layers [25]. Central among the multitude of existing approaches, the feedforward technique of CNN shows emerging applications in image reconstruction, lesion/artifact detection, denoising, segmentation, or super resolution [19, 26‐28]. CNN employs trainable filters and convolution operations between input and output filters, reducing the amount of parameters necessary to be learned compared to classic multilayer perception methods [29, 30]. Few feasibility studies reported on motion artifact reduction. For instance, Hauptmann et al [20] generated a CNN for the suppression of spatiotemporal artifacts on cine images of cardiac MR studies. Pawar et al [22] successfully applied an encoder-decoder convolutional neural network on clinical brain scans for motion correction. However, clinically validated studies, especially focusing on motion artifacts in abdominal MR scans, remain scarce. In our study, the newly developed CNN filter MARC significantly improved image quality by reducing motion artifacts in dynamic liver MRI. We were able to apply this post hoc motion correction in a clinical setting. Furthermore, in almost all cases, observers preferred artifact-reduced MR images for evaluating focal liver lesions compared to the original ones or rated both datasets as equal. Comparing original and artifact-reduced images side-by-side, no loss of anatomical or pathological information after MARC application could be detected.

In our opinion, the use of the MARC filter offers a new perspective in artifact-reduced high-quality MRI. The application of MARC in combination with multi-arterial phase imaging technique is able to reduce artifacts in liver MRI. This can be of value, for instance, in cases where pseudolesions occur in original images due to artifacts, which might be clarified after artifact removal. The impact of artifact reduction in other regions, e.g., MRI of the lung, pancreas, kidney, and bowel, needs to be further explored. Nevertheless, an improvement in image quality does not automatically result in a better diagnosis.

Apart from that, advanced deep learning techniques, such as generative adversarial networks (GANs) [31], could improve our approach. Jiang et al [32] proposed a respiratory motion correction method for abdominal imaging using U-Net and GAN. Armanious et al [33] demonstrated that GAN enables rigid and non-rigid motion correction. These results show GAN as a promising approach for motion artifact correction, although there is a difficulty in determining the structure and hyperparameters due to simultaneous training of two networks of a generator and a discriminator [34].

Our study has several limitations. First, the study design was retrospective. Since the observers were informed about the study purpose, they might have been sensitized to TSM occurrence while examining MR images. This, however, accounts for original and artifact-reduced datasets alike. Furthermore, the introduced filter trained with the simulated datasets was not generalized for all possible cases since the simulation was performed based on the limited condition of respiratory motion. Likewise, the input of MARC filter was multiple arterial phase images acquired with a specific MRI pulse sequence/view sharing acquisition technique (DISCO). Additional validation in terms of diagnostic ability, i.e., sensitivity, specificity, and accuracy, with a larger patient cohort or different MRI sequence parameters might be beneficial for further improvement of the developed filter algorithm. There is still room for further improvement of motion artifact simulation. Since DISCO sequence uses a complicated scheme consisting of ARC reconstruction and irregular sub-sampling k-space ordering from low- to high-frequency components, phase corruption due to motion could lead to aliasing artifact in addition to motion artifact. Since the training datasets were generated from DICOM images in this study, it is difficult to simulate these complicated mechanisms. Therefore, we implemented the approximated simulation with straightforward centric k-space order. Further consideration will be needed to improve de-aliasing performance by generating datasets from complex raw data with exact acquisition ordering. Because the proposed CNN-based approach removes motion artifact by extracting deep features that are optimized with the training process, inappropriate training such as inadequate network structure and insufficient size of training datasets degrades the processing performance of the network. Therefore, a larger volume of training datasets or adopting a sophisticated network could yield further improvement.

In conclusion, the clinical evaluation of our newly developed CNN-based filter MARC significantly improved image quality of gadoxetate-enhanced MRI of the liver by reducing motion artifacts, especially in cases with substantial artifacts. This approach is of high clinical value in subjects who failed breath-hold in the scan.