Validation of Highly Accelerated Wave–CAIPI SWI Compared with Conventional SWI and T2*-Weighted Gradient Recalled-Echo for Routine Clinical Brain MRI at 3T

Subjects and Study Design

A prospective comparative study was performed at a single institution (Massachusetts General Hospital, Boston, MA). Consecutive adult patients (older than 19 years of age; n = 246) undergoing clinical brain MR imaging were prospectively evaluated, including both inpatient and outpatient examinations. Imaging was performed on a 3T Magnetom Prisma scanner from March to June 2018 and on a 3T Magnetom Skyra MR imaging scanner from May to June 2018 (Siemens, Erlangen, Germany). There were no exclusion criteria beyond those for routine clinical MR imaging. The study was Health Insurance Portability and Accountability Act–compliant and approved by our institutional review board. Verbal consent was obtained before MR imaging. Written consent was waived by the institutional review board. Demographics of the study subjects and clinical indications for MR imaging are shown in On-line Table 1.

Wave-CAIPI SWI Pulse Sequence and Reconstruction

Wave-SWI was implemented using a work-in-progress dual-echo 3D gradient-echo pulse sequence⁶ (WIP1058C; Siemens). On-line reconstruction was performed using an autocalibrated procedure for simultaneous estimation of the parallel imaging reconstruction and true k-space trajectory (which accounts for potential gradient hardware imperfections),⁷ with a reconstruction time of approximately 60 seconds. This included phase unwrapping of the reconstructed multiecho data and a weighted combination that accounted for the TE phase evolution. The standard vendor SWI processing was then performed to produce high-pass-filtered phase images and associated SWI. Pulse sequence parameters could not be exactly matched between the wave-SWI and standard SWI sequences due to vendor constraints on the available parameter options, but they were approximated as closely as possible within the allowable range of parameter values (On-line Table 2).

MR Imaging Protocol

Each patient underwent brain MR imaging on 1 of two 3T MR imaging scanners using commercially available 20- and 32-channel receiver coil arrays (Siemens). Standard institutional brain MR imaging protocols were selected by the radiologist on the basis of the provided clinical indication. Each scan included a conventional magnetic susceptibility–weighted sequence (either standard SWI or T2*W GRE, decided at the discretion of the protocoling radiologist based on the clinical indication) and a highly accelerated wave-SWI sequence performed immediately before or after the conventional susceptibility-weighted sequence. Acceleration factors of R = 3 × 2 and R = 3 × 3 were chosen on the 20-channel and 32-channel coils, respectively, to balance scan time with an acceptable SNR for each coil configuration.¹⁰ The acquisition order alternated on a weekly basis to control for the possible effect of increased patient motion on the later sequence. A summary of the magnetic susceptibility-weighted sequence parameters is provided in On-line Table 2.

Image Evaluation

A semiquantitative grading system based on predetermined criteria was used to compare wave-SWI with the conventional susceptibility-weighted sequence acquired for each patient (standard SWI or T2*W GRE). The DICOM datasets were anonymized and transferred to an independent workstation. Blinded to patient information and protocol type, 2 neuroradiologists (O.R. and S.Y.H.) with 17 and 7 years of experience, respectively, independently reviewed all images in randomized order. To obtain optimal visualization, adjustments of window widths and levels were allowed. Only the magnetic susceptibility-weighted sequences were evaluated.

To provide a comprehensive assessment of the diagnostic performance of the wave-SWI and conventional susceptibility-weighted sequences, we divided the review sessions into an individual image series analysis (“individual analysis”), in which images obtained using each sequence were presented in isolation, and a head-to-head image series analysis (“head-to-head analysis”), in which images obtained using the wave-SWI and standard magnetic susceptibility-weighted sequence were presented side-by-side with randomly selected left and right screen positions. The individual analysis was conducted in 2 sessions so that only 1 sequence per patient was presented to the reviewer in each session. The individual analysis sessions were separated by at least 2 weeks to minimize memory bias.

In the individual analysis, reviewers evaluated the following: the presence of hemorrhage, the number and location of the cerebral microhemorrhages (if present), the degree of motion, and whether the images were of diagnostic quality (yes or no). Motion was scored on a predefined 4-point scale. To minimize subjectivity, representative images of each motion score were available to reviewers during the evaluation (On-line Figure). Microhemorrhages were defined according to the Microbleed Anatomical Rating Scale (MARS)¹¹ and were counted by a single reviewer, excluding cases with gross structural lesions (eg, large parenchymal hematoma, infarct, or surgical cavity) that may limit the reliability of the microhemorrhage counts. A subset of 20 randomly selected cases was evaluated by a second reviewer to determine interrater reliability of the microhemorrhage counts.

In the head-to-head analysis, all cases in which either reviewer identified an abnormality on either of the 2 conventional susceptibility sequences in the individual analysis were presented to the reviewers in a separate blinded review, with the wave-SWI and conventional susceptibility-weighted sequence simultaneously displayed on the left and right halves of the screen, labeled image A and image B. The order of the cases and the screen position of the sequences were randomized. The raters compared and scored the 2 sequences for each of the following variables: visualization of pathology (primary outcome), the presence of artifacts (including motion, signal drop-out, and parallel imaging artifacts), and overall diagnostic quality. A predefined 5-point scale was used, in which positive numbers favored the sequence on the right and negative numbers favored the sequence on the left side of the screen (On-line Table 3). Disagreements between readers were adjudicated by a third neuroradiologist (P.W.S.) with >20 years of experience.

Statistical Analysis

In the individual analysis, we used the McNemar test for comparison of dichotomous variables and the Wilcoxon signed rank test for comparison of ordinal variables between sequences. For motion grading, scores of 0 and 1 (ie, “no motion” and “mild motion that is not clinically relevant”) were combined in a single category, because this distinction was not clinically meaningful. In the head-to-head analysis, we tested for superiority of wave-SWI compared with T2*WI GRE and for noninferiority of wave-SWI compared with standard SWI. This approach was selected a priori on the basis of our hypothesis that the wave-SWI was comparable with the standard SWI but superior to the T2*-weighted GRE images. For superiority testing, we compared the ordinal radiologists’ scores using the Wilcoxon signed rank test with the null hypothesis (H₀) of no difference between sequences. For noninferiority testing,¹² a noninferiority margin (Δ) of 15% was chosen, with the null hypothesis (H₀) that the proportion of cases in which standard SWI was preferred over wave-SWI was >15%. We used the z statistic to calculate the probability of the standard sequence being preferred over the wave-SWI sequence in >15% of cases (H₀ > Δ), with a type I error rate (α) of 0.05. We also calculated the upper bound of the 95% confidence interval for the proportion of cases in which the standard SWI was preferred over wave-SWI (ie, the critical value, P_critical).¹³ The required sample size was estimated as described by Cohen¹⁴ for a single proportion (the proportion of cases in which visualization of pathology was preferred on standard SWI over wave-SWI), for an effect size of 0.15, a type I error rate (α) of 0.05, and a power (1–β) of 0.90. According to this calculation, a minimum of 63 cases with abnormal findings was required. For dichotomous variables, interrater agreement was evaluated using the unweighted Cohen κ coefficient.

For ordinal variables, interrater agreement was reported using the quadratically weighted Cohen κ, to disproportionately penalize larger disagreements. For the numeric microhemorrhage counts, interrater agreement was reported using the intraclass correlation coefficient. Agreement was interpreted according to Landis and Koch.¹⁵ We applied a Bonferroni correction for 5 comparisons (the presence of hemorrhage, motion artifacts, visualization of pathology, artifacts, and overall diagnostic quality), with a corrected threshold for a statistical significance of .05/5 = .01. We also performed exploratory univariate testing evaluation for the possible effect of age and study indication on motion scores. For age, we calculated the Pearson correlation coefficient between patient age and the motion score. For indication, we performed a multinomial logistic regression with the indication as the independent variable and motion score as the dependent variable. All statistical calculations were performed using R statistical and computing software, Version 3.4.3 (http://www.r-project.org/).