1. Introduction
Magnetic resonance imaging (MRI) is the primary non-invasive radiological method for evaluating a wide range of illnesses. However, it is still a time-consuming procedure that causes discomfort to patients, in particular to those suffering from neurological disorders and claustrophobia, as well as to paediatric patients who may not bear long time acquisition; moreover, the reduction in the acquisition time can also be considered an important goal in patients with MR-conditional implants.
Several MRI techniques have been developed to reduce acquisition times, ranging from accelerated sequences to reconstruction methods [
1]. The parallel imaging is a robust method, allowing to significantly reduce the scan time without affecting the image quality thanks to the properties of the new coil arrays [
2,
3,
4]; nowadays, the most common parallel imaging techniques are the SENSitivity Encoding-SENSE, Compressing Sense-CS, GeneRalized Autocalibrating Partial Parallel Acquisition—GRAPPA and the Iterative self-consistent parallel imaging reconstruction (SPIRiT) offered by the main MR vendors.
Although CS can reduce scan time by randomly under-sampling data, SENSE evenly under-samples data and uses coil sensitivity information to restore the entire image. The SENSE technique is a method for parallel imaging that uniformly under-samples data and uses coil sensitivity information to restore the full image. On the other hand, CS allows for scan time reduction by randomly under-sampling data compared to other parallel imaging techniques. The sampling in CS is described as “variable density incoherently under-sampled”, where in the centre of k-space is more densely sampled to account for the fact that most of the contrast information comes from this region [
5]. To improve the coherence of the sampling scheme, a sparsity-enforcing (iterative) reconstruction should be employed to reduce the pseudo-noise introduced by the unconventional sampling process. This can be achieved using a dedicated cost function for model-based reconstruction [
6].
Currently, SENSE and CS techniques can be combined to obtain images with acquisition acceleration factors far superior to those achievable with parallel or CS techniques alone [
7]. One such technique, called compressed sensing-sensitivity encoding (Compressed-SENSE), is provided by Philips Healthcare (Philips Healthcare, Amsterdam, The Netherlands).
However, to determine the extent to which acquisition can be accelerated without compromising image quality, quantitative and qualitative analyses are required. These analyses will determine if the diagnostic power would differ when using Compressed-SENSE and if any new types of artefacts would be observed.
The aim of this study was to compare qualitative and quantitative imaging techniques, specifically two-dimensional (2D) axial T1 Turbo Spin Echo-TSE, 2D axial T2-TSE and 3D Fluid Attenuated Inversion Recovery (FLAIR) with and without the Compressed-SENSE technique in healthy volunteers to avoid bias in the analysis related to specific neurological diseases. The contrast (C), contrast-to-noise ratio (CNR) and signal-to-noise ratio (SNR) were assessed.
2. Methods
2.1. Population
This retrospective study included brain MRI scans of 142 healthy subjects, 69 of which were acquired with Compressed-SENSE and 73 without Compressed-SENSE. The Compressed-SENSE group consisted of 31 men and 38 women (aged 19–86 years), while the No Compressed-SENSE group consisted of 36 men and 37 women (aged 13–84 years). The inclusion criteria were healthy volunteers, aged between 10 and 90 years without gender, race, language and educational predilection. The exclusion criteria for the enrolled subjects were history of head trauma, stroke, epilepsy, CNS infections, demyelinating disease, previous or active neurological diseases, tumours, coinfections, illicit substance abuse in five years and any kind of contraindication to MRI examinations.
The study was approved by the Ethics Committee of the University Hospital and Faculty of Medicine of Tor Vergata University of Rome. The authors confirm that all imaging methods were employed in accordance with national and institutional guidelines and regulations, in accordance with the 1964 Declaration of Helsinki and its subsequent amendments.
Informed written consent was obtained from all the patients (or their parents in the case of minors).
Sample size of 104 people was assessed according to Z formula and a confidence interval of 95% with 80% power to detect any significant difference between the compressed and non-compressed sequences with a significance level of 0.05.
2.2. Protocol Optimisation
Two neuroradiologists and an MRI physicist agreed to optimise the compressed sense factor (CS factor) in order to decrease examination time or improve image resolution, and thus included Compressed-SENSE in the clinical brain MR procedure. To guarantee that the final images were the same as those produced without Compressed-SENSE, the CS factor was carefully selected for each sequence. This method was applied to FLAIR and TSE sequences. The sequence was scanned with the CS factor while the denoising level was varied. This method was then used to figure out the denoising level.
2.3. MRI Protocol
MRI examinations were performed on a 3T scanner system (Achieva d-STREAM, Philips Medical Systems, Best, The Netherlands) with a dedicated 32-channel head coil.
The MRI protocol in both groups included the following: (a) axial T2-weighted (w) TSE sequence; (b) axial T1-w TSE sequence; (c) 3D T2-w FLAIR (T2 3D FLAIR) sequence; (d) echo planar diffusion weighted imaging (DWI); (e) axial T2 fast field echo (FFE) sequence. Compressed-SENSE was used to accelerate T2w TSE, T1w TSE, T2 3D FLAIR sequences only in Compressed-SENSE group.
Table 1 reports the specific acquisition parameters for the sequences that were evaluated.
Both quantitative and qualitative methods were used to analyse the Compressed-SENSE and conventional 3D FLAIR, 2D T2 TSE, and T1 IR TSE axial brain sequences obtained at 3T (
Figure 1).
2.4. Qualitative Image Analysis
Two double-blind radiologists, with more than 10 years and 3 years of neuroimaging experience, respectively, evaluated the qualitative analysis of Compressed-SENSE and No Compressed-SENSE images (T1w, T2w, and T2-FLAIR). The scoring system used for the analysis was as follows:
Score 5: Excellent; acceptable for diagnostic use, complete absence of artefacts;
Score 4: Good; acceptable for diagnostic use (only minor artefacts);
Score 3: Fair; acceptable for diagnostic use but with minor issues;
Score 2: Sufficient; acceptable for diagnostic use but severely mixed with the background;
Score 1: Insufficient; not acceptable for diagnostic use.
The qualitative score was obtained considering the following criteria: (i) demarcation of caudate head and sulci, (ii) grey-white matter differentiation (at the level of the lateral ventricles), (iii) artefacts in the posterior cranial fossa, (iv) motion artefacts, (v) low image resolution artefacts, (vi) low SNR, (vii) phase encoding, (viii) parallel imaging artefacts, (ix) flow pulsation artefacts in the ventricles in accordance with previous published papers [
8,
9,
10].
2.5. Quantitative Image Analysis
The Compressed-SENSE and No Compressed-SENSE T1w, T2w e 3D-FLAIR were visualised, contoured and opportunely spatially aligned and co-registered using (3D) Slicer Software (Release 4.6.2). Two radiologists, in consensus, manually drew twelve regions of interests (ROIs), with a diameter of 4 mm, on three different structures of the central nervous system (CNS): (i) white matter (WM), (ii) grey matter (GM), and (iii) cerebrospinal fluid (CSF) as reported by Di Giuliano et al. [
11,
12].
For the WM, the ROIs were positioned in the (i) left and right centrum semiovale and (ii) genu and splenium of corpus callosum; for the GM, the ROIs were positioned in the (i) left and right frontal and occipital cortex and (ii) thalami; for CSF, the ROIs were positioned in the left and right anterior horn of the lateral ventricles. The mean and the standard deviation of the signal intensity were calculated in each ROI.
The quantitative analysis of the images was performed by evaluating C, CNR and SNR, using the following equation:
where
Ii and
Ij are the mean signal intensity in two different tissue type;
is the median standard deviation of all 12 ROIs and
is the standard deviation for
ith ROI.
Patients were combined for C, CNR and SNR.
The ROIs were loaded in MATLAB Version 9.7.0, Release 2019b, where dedicated scripts for quantitative image analysis was developed.
2.6. Statistical Analysis
The weighted Kappa test was performed to evaluate the inter-rater agreement between the two radiologists for Compressed-SENSE and conventional T1, T2 and 3D-FLAIR sequences: a Kappa value > 0.81 was considered very good, a value between 0.61 and 0.80 was considered good, 0.41–0.60 was considered moderate, 0.21–0.40 was considered fair and a value < 0.20 was considered poor.
Because the sample size was fewer than 50, the Shapiro–Wilk normality test was used to assess the variables’ normal distribution. The variables in the study were non-normally distributed; therefore, the median and 95% confidence intervals (95% CIs) were used for all the analysed variables.
Data are expressed as median and 95% CIs. p-values refer to the Mann–Whitney U test.
4. Discussion
Compressed-SENSE is a strategy that accelerates MRI scans while preserving imaging quality for longer sequences, such as mDIXON, 3D BRAIN VIEW, and susceptibility-weighted imaging-phase (SWI). This is particularly useful for non-compliant and monitored patients. Additionally, Compressed-SENSE has been shown to decrease the Specific Absorption Rate (SAR) in long-term tests.
Few clinical studies have evaluated the performance of C-SENSE in body, brain and the cranial nerve imaging [
13,
14,
15,
16,
17,
18,
19,
20,
21,
22].
Qualitative analysis revealed a higher median value for the Compressed-SENSE T2w images compared to the No Compressed-SENSE T2w images with good agreement between observers. This finding is in line with Meister et al. [
19], who suggested that the slightly more homogeneous signal of GM and WM on Compressed-SENSE images, as opposed to standard T2-TSE images, may be due to the denoising capability of Compressed-SENSE [
4,
5,
22,
23]. However, Monch reported lower image quality for the T2-TSE sequence than for the Compressed-SENSE T2-TSE sequence, although the difference was not statistically significant [
24].
In contrast, the quality of Compressed-SENSE T1w and Compressed-SENSE 3D FLAIR was lower than that of non-Compressed-SENSE ones, with moderate inter-reader agreement. These results contradict the literature [
7,
17], which showed that accelerated Compressed-SENSE 3D FLAIR imaging provides equivalent image quality compared to corresponding conventional imaging.
Molnar et al. demonstrated a statistically significant negative correlation between the Compressed-SENSE acceleration factor and the evaluation score for corticomedullary differentiation, sulcus delineation, and noise in Compressed-SENSE T2 TSE axial scans. However, there was no statistically significant difference in the diagnostic quality of Compressed-SENSE T2-TSE images acquired with Compressed-SENSE factors ranging between 2 and 3 and those acquired without Compressed-SENSE [
25].
The Compressed-SENSE T1-TSE demonstrated significantly higher contrast values than the No Compressed-SENSE T1-TSE in the three different brain structures. The statistical analysis revealed that the No Compressed-SENSE T2-TSE and No Compressed-SENSE T2-3D-FLAIR sequences had significantly higher values than the Compressed-SENSE sequences, except for GM-WM, where no statistically significant differences were found. These results may be attributed to the intrinsic high tissue contrast, which could be reduced by the strong Compressed-SENSE denoising applied to the T2 weighted sequences. Additionally, the Compressed-SENSE T1-TSE sequence appears to be less sensitive to denoising.
Furthermore, the No Compressed-SENSE T2-TSE sequence exhibited significantly higher CNR values than the Compressed-SENSE T2-TSE sequence for the three brain structures, enabling better identification of brain lesions. The CNR did not differ significantly between the Compressed-SENSE and No Compressed-SENSE 3D FLAIR and T1-TSE sequences in GM-WM.
Even though TR and TE were the same for the T1-TSE No Compressed-SENSE and T1-TSE Compressed-SENSE, the TR values between Compressed-SENSE and No Compressed-SENSE T2-TSE and T2-3D-FLAIR were different: in particular, the preset TR values were higher for the T2 Compressed-SENSE sequences than the No Compressed-SENSE ones, and this seems to be related to the need to have more signal from the Compressed sequences due to the intrinsic under-sampling technique. Moreover, to gain signal, we increased the number of averages (NSA) for the Compressed-SENSE T2-TSE.
Although the contrast and CNR values were lower for the Compressed-SENSE T2-TSE and Compressed-SENSE T2-3D FLAIR sequences, as well as for the SNR values of the Compressed-SENSE T2-TSE, these did not affect the diagnostic ability of the Compressed-SENSE sequences.
Moreover, we are well aware that the higher default TR values for the Compressed-SENSE T2-TSE and Compressed-SENSE T2-3D-FLAIR than No-Compressed ones had a strong impact on time. In our opinion, the conjunction between the increase of NSA for the C-SENSE T2-TSE and TR values determined the increase in acquisition time and therefore a similar total scan time between the Compressed and No-Compressed T2 sequences. Nevertheless, in our opinion, a small reduction in the TR value for the C-SENSE T2 sequences could lead to time reduction without a significant impact on image quality.
The SNR values for the No Compressed-SENSE T2-TSE were higher than those for the Compressed-SENSE T2-TSE, except for the CSF. Our SNR values for the Compressed-SENSE T2-TSE were lower than the conventional values in all brain regions (
p < 0.05), including the basal ganglia. This may be related to the denoising level, which disagrees with Monch et al. [
24]. However, noise measurement might be difficult when using parallel imaging technology due to the spatial nonuniformity of the noise as well as the exacerbation of Rician bias at lower signal levels; even though we evaluated the SNR and CNR through the standard deviation of the contoured regions, further approaches are suggested in the literature [
26,
27,
28].
Our analysis revealed that higher acceleration factors were achieved for 3D scans compared to the 2D scans due to more room for aggressive under-sampling. This finding is consistent with Sartoretti et al. [
29], who observed a 32% reduction in sequence acquisition time while maintaining the same image quality using a C-SENSE factor of 7.8 for T2 3D FLAIR. We set the Compressed-SENSE factor at 9 with an acceleration of 17% for the same sequence, with no statistically significant difference in image quality.
Duan et al. [
30] evaluated the Compressed-SENSE method for 3D T1w turbo field echo (TFE) brain imaging. They compared it to the previous studies of 3D T1 [
18,
29,
31] and T2-FLAIR images. The authors evaluated more acceleration factors [
7,
29,
31,
32,
33] compared to both the traditional acceleration technique (SENSE) and the non-accelerated imaging [
18,
24]. The study demonstrated a significant decrease in SNR and CNR in all accelerated sequences, with CS factor 3 showing better performance in terms of image artefacts due to its shorter scan time, indicating that it can be recommended as the optimal acceleration factor [
30].
Sartoretti et al. [
34] demonstrated that the application of Compressed-SENSE led to an increase in motion artefacts with a higher spatial frequency. Additionally, images obtained with C-SENSE may exhibit specific artefacts, such as streaky-linear, wax-layer and starry-sky. In our study, we found ‘streaky-linear’ type A artefacts in 36% of patients in Compressed-SENSE 3D-FLAIR sequences. These artefacts were observed particularly in subjects with major motion artefacts, as reported by Sartoretti [
34]. Type A streaks, which were either long or short lines, appeared centrally and peripherally on the image in a horizontal or oblique arrangement. Sartoretti et al. [
34] state that type-A streaks can always be provoked on transverse, coronal and sagittal images if the reconstruction voxel is smaller than the acquisition voxel. Although our reconstruction voxel was the same as the acquisition voxel (both 1 mm), we also observed these types of artefacts.
Furthermore, in Compressed-SENSE 3D-FLAIR sequences, type B artefacts with a “streaky-linear” appearance were found in 16% of patients, while only one patient presented the ‘starry-sky’ artefacts, in which some structures appeared slightly pixelated and grainy. These results are in line with those observed by Sartoretti [
34]. The sequences with the highest Compressed-SENSE factor exhibited the most artefacts. The Compressed-SENSE factor used was 8.2 for 3D FLAIR, while ours was 9. In particular, we felt that the streaky artefacts had little impact on image quality and did not affect image evaluation, given their regular occurrence as described in the literature [
35,
36], although they may influence image interpretation if they remain undetected by image readers due to their unpredictable, often not easily recognisable and unfamiliar nature [
34].
Attention must be paid to the setting of the optimal denoising level for the Compressed-SENSE sequences; the aim of denoising is to minimise the amount of noise in the images due to unwanted signals that compromise the quality of a desired signal. The level of denoising determines how much noise is reduced or balanced: images with high denoising are smoother and less noisy, whereas images with weak denoising are sharper but noisier. Thereby strong denoising should be used with caution, as it can blur the white-grey matter interface.
This study had several limitations. A major limitation is that qualitative and quantitative image analysis of Compressed-SENSE and No Compressed-SENSE sequences were not acquired and compared in the same patients. Secondly, we did not assess the effect of Compressed-SENSE acceleration on selected sequences in the presence of different pathologies, or how it would affect the diagnostic task of their detection/characterisation: efforts should also be made to extend the evaluation of Compressed-SENSE sequences in different brain diseases with larger human datasets.