Iterative Denoising Accelerated 3D FLAIR Sequence for Hydrops MR Imaging at 3T

DISCUSSION

In this study, we demonstrated that an accelerated 3D FLAIR sequence combined with an ID algorithm enabled reducing the scan time by 33% without compromising image quality and diagnostic performance for EH. As expected, the SNR was significantly increased with acc-FLAIR-ID compared with conv-FLAIR because a noise-reduction algorithm was used.

An MR imaging evaluation of the endolymphatic space relies on the selective enhancement of the perilymphatic space after administration of IV contrast media, which enables distinguishing the endolymphatic and perilymphatic spaces.^1,3 The main limiting factor is the low concentration of gadolinium obtained in the perilymphatic space. To overcome this, we optimized several parameters to increase the signal intensity. First, the administration of gadolinium-based contrast agents with higher longitudinal relaxivity and concentration has been recommended.¹⁵ Second, a constant flip angle instead of a variable flip angle provides higher signal and contrast in the perilymphatic space, by shortening the longitudinal relaxation induced by gadolinium.⁹ However, the use of a constant flip angle with a high echo-train length compromises the SNR. Third, the signal intensity of the perilymphatic space increases using a long TR (16,000 ms), which allows sufficient longitudinal magnetization regrowth to detect minor T1-shortening related to low gadolinium concentration.^19,20 Yet, a long TR contributes to the long acquisition time, which is an important limitation to the wide spread of this protocol.

PI acceleration techniques, based on phased array coils, might be used to significantly decrease scan time to overcome the long acquisition time of these sequences. However, by increasing the acceleration factor, the reduction of the acquisition time is limited by a significant decrease in SNR (by a factor of the square root of the acceleration factor) because fewer data points are acquired and averaged.^5,21 In PI-reconstructed images, the SNR also depends on the spatially varying noise characteristics and amplification in the final images, quantified by the g-factor, which originates from the coil sensitivities. Because inner ear imaging requires reduced section acquisition coverage with activation of a few coil elements, the use of a higher acceleration factor is limited. Moreover, PI techniques are particularly sensitive to motion artifacts that might occur between the time of the calibration scan and image acquisition. Thus, we have decided to increase the turbo factor and the receiver bandwidth instead of increasing the PI acceleration factor. 3D FLAIR sequences for EH exploration are less susceptible to signal loss due to the weaker later echoes affected by T2 decay. Indeed, inner ear imaging is particularly suitable for the use of a high turbo factor with a long echo-train because the T2 values of the labyrinthine fluid are high (similar to CSF, which is around 2000 ms at 3T). Despite the high turbo factor used, blurring is avoided because the echo-train duration remains inferior to 2–3 times the T2 values of the primary interest area.⁸ By increasing the bandwidth, we were able to reduce interecho spacing to maintain the same readout time while increasing the number of echoes, decreasing the total scan time. Nevertheless, the inherent result of an increased bandwidth is a 25% SNR loss because of the amount of noise that is sampled due to the larger frequency range.

The ID algorithm compensated for the signal loss caused by using a constant flip angle with a long echo-train length and the increased bandwidth. Recently, Naganawa et al¹⁰ achieved a 5-minute HYDROPS-Mi2 sequence with DLR. The reduction in acquisition time was mainly obtained by decreasing the number of slices (224 to 60), while the SNR loss was compensated for with the DLR. The DLR tool incorporates a deep convolutional neural network restoration process into the reconstruction flow and enables noise reduction. DLR is a nonlinear processing with behaviors potentially difficult to predict.²² Thus, in our experience, ID allows better control of the denoising process, parameters, and strength, ensuring a precise estimation of the heterogeneous noise distribution.

Other acceleration approaches such as compressed sensing (CS) have been introduced to reduce the scan time. CS is based on incoherent subsampling of the Fourier space, transformation of the image into a sparse representation, and nonlinear iterative reconstruction.²³ It is used in various applications and is particularly suitable for indications in which images are sparse, such as MRA.²⁴ However, other applications with low sparsity, such as 3D morphologic sequences with high spatial resolution, offer little acceleration potential with CS. In addition, artifacts such as image blurring and global ringing have been described with CS, notably for MR neuroimaging,²⁵ which limits the acceleration rates achievable. CS is also limited by its extended reconstruction time, which can be reduced with the use of a graphic processing unit, though it is not available on all clinical MR imaging scanners.

Conversely, an ID algorithm can be performed on conventional computers without a significant increase of reconstruction time. The use of a quantitative noise map in ID is particularly suited to limit the g-factor penalty associated with high acceleration rates, as well as the SNR loss related to the increased bandwidth.

Our study has several limitations. EH was reported as present or absent by an anatomic system, but we did not use grading or volumetric assessment. However, our hydrops assessment was based on a previously reported anatomically-based grading system,¹⁷ and the grading used should not have impacted diagnostic relevancy.

The use of a 3T system and a 64-channel phased array head coil contributed to the high image quality. The SNR improvement with ID could improve the image quality of 1.5T scanners, which are more available, and further studies should be performed at 1.5T or with lower head coil density.

Our study has several clinical implications. By reducing the scan time, patient comfort and satisfaction are increased, reducing the risk of motion artifacts. Shortening the imaging time will also allow a higher patient throughput and is expected to promote wider use of MR imaging for the evaluation of EH. In our institution, about 40 patients per week undergo inner ear MR imaging. A 33% scan time reduction of 8 minutes 15 seconds would allow 120 minutes of additional machine time.

Along with the scan time, the 4-hour delay after gadolinium injection contributes to the logistical strains of EH imaging. Our group showed in a recent work that with optimized 3D FLAIR parameters, the postinjection delay could be shortened to 2 hours with sufficient contrast for EH evaluation, which should further shorten the imaging time and promote a wider use of EH MR imaging.²⁶