Compressed Sensing–Sensitivity Encoding (CS-SENSE) Accelerated Brain Imaging: Reduced Scan Time without Reduced Image Quality

BACKGROUND AND PURPOSE: Compressed sensing–sensitivity encoding is a promising MR imaging acceleration technique. This study compares the image quality of compressed sensing–sensitivity encoding accelerated imaging with conventional MR imaging sequences. MATERIALS AND METHODS: Patients with known, treated, or suspected brain tumors underwent compressed sensing–sensitivity encoding accelerated 3D T1-echo-spoiled gradient echo or 3D T2-FLAIR sequences in addition to the corresponding conventional acquisition as part of their clinical brain MR imaging. Two neuroradiologists blinded to sequence and patient information independently evaluated both the accelerated and corresponding conventional acquisitions. The sequences were evaluated on 4- or 5-point Likert scales for overall image quality, SNR, extent/severity of artifacts, and gray-white junction and lesion boundary sharpness. SNR and contrast-to-noise ratio values were compared. RESULTS: Sixty-six patients were included in the study. For T1-echo-spoiled gradient echo, image quality in all 5 metrics was slightly better for compressed sensing–sensitivity encoding than conventional images on average, though it was not statistically significant, and the lower bounds of the 95% confidence intervals indicated that compressed sensing–sensitivity encoding image quality was within 10% of conventional imaging. For T2-FLAIR, image quality of the compressed sensing–sensitivity encoding images was within 10% of the conventional images on average for 3 of 5 metrics. The compressed sensing–sensitivity encoding images had somewhat more artifacts (P = .068) and less gray-white matter sharpness (P = .36) than the conventional images, though neither difference was significant. There was no significant difference in the SNR and contrast-to-noise ratio. There was 25% and 35% scan-time reduction with compressed sensing–sensitivity encoding for FLAIR and echo-spoiled gradient echo sequences, respectively. CONCLUSIONS: Compressed sensing–sensitivity encoding accelerated 3D T1-echo-spoiled gradient echo and T2-FLAIR sequences of the brain show image quality similar to that of standard acquisitions with reduced scan time. Compressed sensing–sensitivity encoding may reduce scan time without sacrificing image quality.

T he excellent soft-tissue contrast resolution and specialized sequences targeting different aspects of pathophysiology make MR imaging the optimal technique for studying the brain. Despite the many advantages of brain MR imaging, MR imaging acquisition is a time-consuming endeavor compared with CT. Long im-age-acquisition times limit both the clinical application and practicality of MR imaging, particularly in medically unstable and pediatric patients.
MR imaging acquisition time is largely influenced by the number of data points sampled from k-space, the way these data points are sampled, and the way in which image reconstruction is performed. Several image-acquisition and postprocessing techniques have been developed to reduce image-acquisition time while still preserving image quality. 1,2 These include parallel MR imaging and compressed sensing (CS) MR imaging techniques, which rely on different reconstruction constraints to accelerate image production. 3 Combining these techniques can lead to image-acquisition acceleration factors that far exceed what is achievable by either parallel or CS MR imaging alone. [3][4][5][6] This combined image-acceleration technique is referred to as CS-sensitivity encoding (SENSE) MR imaging, and it has the potential to dramatically decrease overall imaging times while still preserving image quality.
Despite the many technical advancements that have been made in accelerating MR imaging acquisition and image reconstruction, robust evaluation of these acceleration techniques in clinical practice is still warranted. Clinical verification of the ability of these accelerated image-acquisition techniques to produce diagnostic-quality images of the central nervous system is essential before broader implementation of these imaging techniques into clinical practice can occur. Only a small number of studies have investigated the performance of CS-SENSE MR imaging in limited patient populations as it relates to body imaging. 4,5 Very few studies have evaluated CS in brain MR imaging, with the studies performed focusing on the evaluation of multiple sclerosis lesions on T2-FLAIR, 7 brain MR imaging quality assessment in healthy controls, 8 and evaluating achievable acceleration, reconstruction schemes, and artifacts generated from retrospective CS. 9 To date, however, no one has critically evaluated the clinical performance of the integrated CS-SENSE algorithm for MR imaging applied to imaging of the central nervous system, to our knowledge. In addition, we present the first work to apply CS acceleration in a brain tumor clinical population. The purpose of the current study was to compare the image quality of CS-SENSE accelerated 3D T1echo-spoiled gradient echo (SPGR) (CS-SENSE SPGR) and T2-FLAIR (CS-SENSE FLAIR) sequences with the corresponding conventional acquisitions. We hypothesized that CS-SENSE accelerated sequences will have image quality equivalent to that of conventional acquisitions while accelerating imaging.

Patient Selection
With our institutional review board approval and after obtaining informed written consent, adult patients (18 years of age or older) were prospectively scanned between February 8, 2017, and January 19, 2018, for assessment of the MR imaging brain tumor protocols of our institution with inclusion of a conventional sequence and a corresponding CS-SENSE accelerated acquisition when the clinical schedule permitted. CS-SENSE accelerated acquisition was performed before or after the corresponding conventional acquisition in alternating order (to mitigate potential bias from ordering effects), with both sequences performed after gadolinium administration for both T2-FLAIR and SPGR.

Image Acquisition
All imaging was performed on a 3T Ingenia MR imaging scanner (Philips Healthcare, Best, the Netherlands) using a 16-channel head coil (In Vivo, Gainesville, Florida). Each patient underwent the brain tumor imaging protocol of our institution. This included the following sequences: axial DWI, axial T1-spin-echo, sagittal 3D T2-FLAIR with gadolinium, axial 3D T1-SPGR with gadolinium, and coronal and axial T1-spin-echo with gadolinium. In addition to these conventional acquisitions, each patient underwent either a CS-SENSE accelerated 3D T2-FLAIR (Fig 1) or a CS-SENSE accelerated gadolinium-enhanced 3D T1-SPGR sequence (Fig 2), which was performed during their routine MR imaging examination. The sequence scan parameters for both the conventional and CS-SENSE MR imaging sequences are listed in Table 1. The CS-SENSE FLAIR and CS-SENSE SPGR sequences had acceleration factors of 1.3 and 1.7 with scan time reduc- Conventional and CS-SENSE accelerated sagittal 3D T2-FLAIR images from the same patient demonstrate a treated primary brain tumor within the left frontal lobe. Note the sharp borders of the brain parenchymal lesion detected in both images, while CS-SENSE 3D FLAIR (right) was acquired with a 25% scan time reduction.

FIG 2.
Conventional and CS-SENSE accelerated axial T1-SPGR images are from the same patient. The arrow demonstrates a small metastasis within the left cerebellar hemisphere that was detected by both sequences equally well. Acquisition of the CS-SENSE SPGR (right) was 35% faster than the conventional SPGR (left).
tions of 25% and 35% compared with the conventional acquisition counterparts, respectively. These CS-SENSE accelerated acquisitions used a balanced variable density incoherent undersampling acquisition scheme and iterative reconstruction to solve an inverse problem with a sparsity constraint. Specifically, the images were acquired using a random undersampling pattern with the Poisson disc style distribution. Image reconstruction was performed using a wavelet transform for the sparsity term, according to the common CS and parallel imaging problem definitions. Prior knowledge of noise decorrelation, regularization, and coil sensitivities was used to provide an optimal SNR as a starting point, allowing additional acceleration capabilities via sparsity constraining. The reconstruction algorithm was based on a modified fast iterative shrinkage/ soft thresholding algorithm (FISTA) scheme, 10 which entails iterative reconstruction. Conventional clinical acquisition T2-FLAIR and T1-SPGR sequences served as imaging control sequences against which the CS-SENSE FLAIR and SPGR sequences could be compared.

Image Evaluation
Two experienced board-certified neuroradiologists (M.M.-B. and N.M.C.) blinded to the imaging technique and patient clinical information independently evaluated all CS-SENSE and corresponding conventional sequences. All imaging studies were deidentified and randomized so that each rater was unaware of whether they were reviewing a CS-SENSE or conventional acquisition. Raters evaluated overall imaging quality on the following 4-point scale: 1, nondiagnostic; 2, limited but interpretable; 3, minimally limited; and 4, optimal quality. Image SNR was rated on the following 5-point scale: 1, markedly diminished SNR that renders the images uninterpretable; 2, moderately diminished SNR that affects interpretation; 3, diminished SNR that only mildly limits interpretation; 4, mildly diminished SNR that does not affect image interpretation; and 5, optimal SNR. Image artifacts were evaluated on the following 5-point scale: 1, severe image artifacts; 2, moderate artifacts; 3, mild artifacts; 4, trace artifacts; and 5, no artifacts. Brain parenchymal lesion border sharpness was evaluated on the following 3-point scale: 1, a lesion whose borders were indistinguishable from background brain; 2, a lesion with blurry margins; and 3, sharp lesion margins. Finally, gray-white matter sharpness was assessed using the following 3-point scale: 1, indistinguishable gray-white sharpness; 2, blurry gray-white sharpness; and 3, well-defined gray-white sharpness.
In addition to qualitative image evaluation, quantitative evaluation of SNR and the contrast-to-noise ratio (CNR) was also performed for each CS-SENSE and conventional acquisition, respectively. For each patient, ROIs were drawn on representative images from each CS-SENSE and conventional acquisition. ROIs were positioned within the normal subcortical white matter, within a focal brain lesion, and outside the patient, in what was classified as image background. All ROIs were the same size and had nearly identical positioning between sequences. As reported elsewhere in the literature, 7,11 SNR and CNR were calculated as follows: SNR ϭ SI/SD noise and CNR ϭ (SI lesion Ϫ SI WM )/SD noise , where SI is the average signal intensity of the lesion or white matter and SD noise is the SD of noise.

Statistical Analysis
While raters evaluated images using 4-or 5-point scales, they ultimately only used 2-3 levels of each scale, with the middle level being the most common. To improve interpretability, we dichotomized all scales, mainly to get the best possible balance of ratings above and below the threshold. Specifically, image quality was dichotomized as optimal image quality (5 versus 1-4), optimal SNR (5 versus 1-4), no or trace artifacts (4 -5 versus 1-3), sharp gray-white matter boundaries (3 versus 1-2), and sharp lesion boundaries (3 versus 1-2).
The percentages for each image-quality metric were compared between CS-SENSE and the corresponding conventional images (FLAIR versus CS-FLAIR, SPGR versus CS-SPGR) using the nonparametric bootstrap to calculate 95% CI and P values for the differences. The widths of the 95% CIs were used to help assess a plausible range of differences in image quality between CS-SENSE and the corresponding conventional images. Ratings from both raters were analyzed together for the primary analysis and separately as a sensitivity analysis. Bootstrap resampling was performed by patient to account for the nonindependence of ratings by both raters of the same images and for multiple scans acquired from some patients.
Interrater agreement was assessed using the Cohen and by counting how often both raters, 1 rater, and neither rater rated CS-SENSE images at least as highly as conventional images. All statistical calculations were conducted with R statistical and computing software (Version 3.1.1; http://www.r-project.org/). Throughout, 2-tailed tests were used with statistical significance defined as P Ͻ .05.

Patient Data
Sixty-nine patients were reviewed. Three patients were scanned with 1 of the 2 accelerated image-acquisition sequences but were not scanned with the corresponding conventional sequence and were excluded from analysis.  Table 1). There were a total of 89 and 56 lesions on FLAIR and SPGR, respectively. There were lesions in 34 patients (34 scans) in the FLAIR cohort, 22 of whom had multiple lesions. Lesions were also present in 29 patients (32 scans) in the SPGR cohort, 10 of whom had multiple lesions. Lesion sizes are summarized in On-line Table 2.

Qualitative Image Comparison
Pooled image-quality ratings are summarized in Table 2. For FLAIR, there were no statistically significant differences in overall image quality, SNR, gray-white matter boundary sharpness, or lesion-border sharpness between CS-SENSE and conventional sequences, with the lower bound of the 95% CIs indicating that image quality of the CS-SENSE images was within approximately 10% of the conventional images by these metrics. However, there was a trend toward more artifacts on CS-SENSE compared with conventional images (11.4%, P ϭ .068). For SPGR, there were no significant differences in any imagequality metric between CS-SENSE and conventional SPGR, though CS-SENSE images had slightly higher image-quality ratings on average than the conventional images. By each metric, image-quality ratings of CS-SENSE were within 10% of the conventional SPGR ratings based on the lower bound of the 95% CI. Differences in image quality were most noticeable between the CS-SENSE SPGR and the standard SPGR; 50% of accelerated SPGR studies demonstrated optimal image quality compared with 37% of the standard SPGR acquisitions.

Interrater Agreement
Interrater agreement scores for the CS-SENSE and conventional sequences are listed in Table 3. Interrater agreement for each image quality was mostly poor to fair for FLAIR ( Ͻ 0.4) but fair to moderate for SPGR ( ϭ 0.2-0.6). Despite some differences in absolute ratings, raters both agreed 77%-91% of the time that the image-quality metrics of the CS-SENSE FLAIR were at least as good as the those of conventional images and rarely agreed that the CS-SENSE FLAIR images were worse than conventional images (Table 4). Similarly, raters both agreed 67%-87% of the time that the image-quality metrics of the CS-SENSE SPGR were at least as good as those of the conventional images, while only agreeing 0%-6% of the time that the conventional images were better. In terms of disagreement, there was Ͼ1 disagreement on the Likert scale only for artifact severity (On-line Figure), which occurred in 7/69 comparisons. This level of disagreement did not occur for any other qualitative metrics.

Quantitative Assessment
The white matter SNR or lesion CNR measurements were high for both conventional and CS-SENSE FLAIR acquisitions (Ն44 in all cases), though there was a trend toward higher SNR and CNR values on average for the CS-SENSE acquisition (Table 5). White matter SNR measurements were also relatively high for conventional and CS-SENSE SPGR acquisitions (Ն28 in all cases) with little numeric difference between them on average (difference in medians, 0.3; P ϭ .75). The lesion CNR measurements from the SPGR acquisitions tended to be lower and ranged from 9 to 24 overall. The CNR was slightly higher on average in CS-SENSE than on the conventional acquisitions (difference in medians, 3.3; P ϭ .31), but the difference was not statistically significant (Table 5).

DISCUSSION
Long MR imaging acquisition times represent a significant limitation to widespread use of MR imaging. This is especially true for MR imaging in the evaluation of both clinically unstable and pediatric patients: Increased MR imaging scan time may expose these individuals to an increased need for sedation or result in limited diagnostic quality due to motion. Long image-acquisition times also negatively impact radiology workflow, leading to scheduling bottlenecks. Finally, long acquisition times contribute to the high cost of MR imaging. Given the significance of imageacquisition time for patient safety, clinical efficiency, image quality, and cost, technical effort has been made to decrease imageacquisition and reconstruction times. CS techniques show promise in providing imaging acceleration without significant image-quality degradation. Despite the promise these acceleration techniques hold for improving patient throughput and decreasing imaging cost, rigorous evaluation of the performance of these acceleration techniques in a clinical imaging population has yet to be undertaken. To our knowledge, this is the first study to translate CS-SENSE, which combines and integrates CS and SENSE parallel imaging, to a clinical brain tumor patient popula-tion to evaluate image quality relative to corresponding conventional MR imaging sequences.
In the current study, we hypothesized that CS-SENSE accelerated sequences would have image quality equivalent to that of standard acquisitions while accelerating imaging. To evaluate this hypothesis, we compared the clinical performance of 2 CS-SENSE accelerated MR imaging sequences with their corresponding conventional sequences in a clinical cohort undergoing brain tumor MR imaging scans. On the basis of blinded multirater evaluations of multiple clinically pertinent imaging variables, these accelerated acquisitions largely performed as well as their conventional counterparts across several image-quality metrics, including overall image quality, SNR, image artifacts, gray-white matter boundary sharpness, and parenchymal lesion border sharpness. In particular, the lower bounds of 95% CIs of the differences in image quality between CS-SENSE and conventional images indicated that the CS-SENSE images were within 10% of the conventional images for all metrics for the SPGR sequence and for 3 of 5 metrics for the FLAIR sequence. The CS-SENSE acquisitions had no significant differences in white matter SNR and lesion CNR relative to their corresponding conventional acquisitions, and in fact, there was a trend toward higher values for the CS-SENSE FLAIR relative to the conventional FLAIR. CS-SENSE showed at least comparable SNR and CNR measures relative to their conventional counterparts. While this finding is somewhat counterintuitive considering the undersampling algorithm used by CS-SENSE, increased/similar SNR is thought be a result of the denoising algorithm incorporated into CS-SENSE. CS-SENSE FLAIR and SPGR sequences decreased imaging time by 25% and 35% relative to conventional sequences, respectively. These results confirm that CS-SENSE sequences produce diagnostic-quality MR images of the brain specifically for brain tumor protocols while reducing overall image-acquisition time compared with conventional acquisitions.
To date, multiple MR imaging techniques have been developed with the goal of accelerating image-acquisition and recon-