Abstract
Objective
To demonstrate that a T2 periodically rotated overlapping parallel lines with enhanced reconstruction (PROPELLER) technique using deep learning reconstruction (DLR) will provide better image quality and decrease image noise.
Methods
From December 2020 to March 2021, 35 patients examined cervical spine MRI were included in this study. Four sets of images including fast spin echo (FSE), original PROPELLER, PROPELLER DLR50%, and DLR75% were quantitatively and qualitatively reviewed. We calculated the signal-to-noise ratio (SNR) of the spinal cord and sternocleidomastoid (SCM) muscle and the contrast-to-noise ratio (CNR) of the spinal cord by applying region-of-interest at the spinal cord, SCM muscle, and background air. We evaluated image noise with regard to the spinal cord, SCM, and back muscles at each level from C2-3 to C6-7 in the 4 sets.
Results
At all disc levels, the mean SNR values for the spinal cord and SCM muscles were significantly higher in PROPELLER DLR50% and DLR75% compared to FSE and original PROPELLER images (P < .0083). The mean CNR values of the spinal cord were significantly higher in PROPELLER DLR50% and DLR75% compared to FSE at the C3-4 and 4-5 levels and PROPELLER DLR75% compared to FSE at the C6-7 level (P < .0083). Qualitative analysis of image noise on the spinal cord, SCM, and back muscles showed that PROPELLER DLR50% and PROPELLER DLR75% images showed a significant denoising effect compared to the FSE and original PROPELLER images.
Conclusion
The combination of PROPELLER and DLR improved image quality with a high SNR and CNR and reduced noise.
Advances in knowledge
Motion-insensitive imaging technique (PROPELLER) increased the image quality compared to conventional FSE images. PROPELLER technique with a DLR reduced image noise and improved image quality.
Keywords: deep learning reconstruction, noise reduction, magnetic resonance imaging, PROPELLER
Introduction
High-resolution MRIs with high signal-to-noise ratio (SNR) enable better visualization of precise anatomical structures, improving diagnostic accuracy.1 An axial T2 fast spin echo (FSE) image in cervical spine MRI is essential for evaluating the spinal cord and discs. In many cases, it is difficult to evaluate the spinal cord because the signal intensity (SI) of the spinal cord is inhomogeneous due to motion artefacts, especially by swallowing or breathing. Motion artefacts may have an important influence on image quality and diagnostic value.
The periodically rotated overlapping parallel lines with enhanced reconstruction (PROPELLER) technique reduces motion artefacts in MRI of the abdomen, chest, spine, and brain.2–5 This sequence is a radial k-space sampling concept with parallel data lines rotating around the centre of the k-space, which allows correction of spatial inconsistencies. Motion artefacts are further reduced through averaging in low spatial frequencies.2 Although PROPELLER image can reduce noise, it is often associated with image blurring caused by motion artefacts.
Several denoising techniques have been used to improve the image quality of low SNR images. Deep learning (DL) has recently been used in medical research and reported for improved image quality and noise reduction. DL methods have also been applied to MR artefact detection such as the automated reference-free detection of patient motion artefacts in MRI.6
The deep learning reconstruction (DLR) pipeline uses raw k-space data as its input and generates high fidelity images as its output. The goal is to produce images that are consistent with the acquired data, have no ringing artefacts and have reduced noise power, ultimately improving diagnostic confidence compared to conventional methods.7
The purpose of this study was to demonstrate that a T2 PROPELLER technique using a DLR algorithm will provide better image quality and decrease image noise compared to the conventional T2 FSE or original PROPELLER images of cervical spine MRI.
Methods
Our institutional review board approved this study with a prospective and retrospective design that involves performing MRI examinations prospectively and imaging analysis retrospectively. The requirement for informed consent was waived due to the routine examination and retrospective nature of the study. Consent for contrast material was not required because of non-enhanced examination.
Patient group
From December 2020 to March 2021, 79 patients underwent non-enhanced cervical spine MRI using the same 3 T scanner (SIGNA Architect, GE Healthcare, Waukesha, United States).
After excluding cases with metal fixation or surgery (n = 12) and severe compression of the spinal cord by variable pathological conditions (n = 32), a total of 35 patients (age range, 19-76 years; mean, 53.4 years; 12 females and 23 males) were included.
MR imaging technique
The sequences and parameters used in this study are listed in Table 1.
Table 1.
Scan parameters of cervical spine MR image.
| Axial FSE T2WI | Axial PROPELLER T2WI | Axial T1WI | Sagittal T1WI | Sagittal T2WI | |
|---|---|---|---|---|---|
| TR (ms) | 4021-6601 | 4594-6486 | 645-728 | 607-736 | 2697-3328 |
| TE (ms) | 94.9-102.9 | 95.6-110 | 12-13.6 | 12.1-12.7 | 104.3-106.3 |
| Echo train lengths | 16-18 | 28 | 3 | 6-8 | 20 |
| Flip angle | 160 | 150-160 | 150-160 | 142 | 160 |
| Matrix | 300 × 250 | 300 | 256 × 200 | 352 × 224 | 416 × 256 |
| FOV (mm) | 140-150 | 140-150 | 140-150 | 250-260 | 250-260 |
| Thickness/gap | 3.0/0.3 | 3.0/0.3 | 3.0/0.3 | 3.0/0.3 | 3.0/0.3 |
| No. of slices | 18 | 15 | 30 | 13 | 13 |
| Bandwidth (Hz) | 41.67 | 50 | 41.67 | 62.50 | 62.50 |
| Parallel imaging factor | 2 | 2 | 2 | 1 | 1 |
| Total scan time (min: sec) | 2:37 | 2:15 | 2:56 | 2:30 | 2:20 |
Abbreviations: FOV = field of view; FSE = fast spin echo; No. = number; PROPELLER = periodically rotated overlapping parallel lines with enhanced reconstruction; T1WI = T1-weighted image; T2WI = T2-weighted image; TE = echo time; TR = repetition time.
In addition to routine protocols including axial and sagittal FSE T1- and T2-weighted images (WIs), axial T2WIs using the PROPELLER technique were obtained. Each sequence was obtained within ∼3 min.
While adding the axial PROPELLER technique, the axial gradient echo image was excluded from the routine protocol. Therefore, the overall examination duration was not affected.
Imaging process
Two sets of images, the original PROPELLER and DLR images were reconstructed for each acquisition using conventional reconstruction and DLR algorithms. The vendor provided DLR comprised a deep convolutional residual encoder network trained to remove noise and truncation artefacts, improving SNR, image sharpness, and resolution.8,9 The DLR algorithm also provided a tunable denoising factor to accommodate the user preferences. To evaluate the denoising effect of the DLR, we obtained images with 2 different noise reduction factors (DLR 50% and 75%).
Imaging analysis
We evaluated the performance of DLR not only for denoising but also for delineation of spinal cord structures on T2WI.
Four types of images including FSE, original PROPELLER, PROPELLER DLR50%, and DLR75% were quantitatively and qualitatively reviewed by 2 radiologists with 23 and 13 years of experience in musculoskeletal MRI, respectively.
Quantitative image analysis
To calculate the contrast-to-noise ratio (CNR) of spinal cord and SNR of the spinal cord and sternocleidomastoid (SCM) muscle, positioning and drawing of the regions of interest (ROIs) were performed by a musculoskeletal radiologist with 23 years of experience. To ensure reproducibility of the ROI location and size, 1 radiologist divided the patients into 10 unspecified patients and repeated the measurements several times, and it was confirmed that similar results were obtained through statistical analysis. In the middle slice of each cervical disc level (from C2-3 to C6-7) on each axial FSE T2WI, PROPELLER T2WI, PROPELLER DLR50%, and PROPELLER DLR75% image, round- or ovoid-shaped ROIs were placed on the spinal cord, SCM muscle, and proximal and distal background air. The size of the ROI was adjusted to include as much of the entire spinal cord as possible. This ROI was then copied and located in the SCM muscle and background air. The ROI of the SCM muscle was located on the right side in most patients; however, if the artefact or distortion was severe, the ROI was located on the left side. The ROIs were placed in the background air closest to (proximal) and furthest (distal) from the outermost border of the imaged structure. The location of the ROI was arbitrarily determined; however, in one patient, they were located at a constant location for each level. For each patient, the ROI of the background air was placed in the same position as much as possible. The window level was adjusted to determine the appropriate ROI location in the background air (Figure 1).
Figure 1.
Drawing and placement of the ROI for measurements of SNR of the spinal cord and SCM muscle and CNR of the spinal cord. Round- or ovoid-shaped ROIs were placed on the spinal cord, SCM muscle, and proximal and distal background air. The size of the ROI was adjusted to include as much of the entire spinal cord as possible. This ROI was then copied and located in the SCM muscle and background air. The ROIs of the background air were located adjacent to the skin (proximal) and at the corner furthest from it (distal). Abbreviations: CNR = contrast-to-noise ratio; ROI = region of interest; SCM = sternocleidomastoid; SNR = signal-to-noise ratio.
We calculated the SNR and CNR using the following formulas from the literature.: SNR = mean SIcord or SCMmuscle/SDbackground, CNR = SIcord−SISCM muscle/SISCM muscle. The SI was taken as the mean SI of the ROIs and SD as the standard deviation of the ROIs. We used the average (SD) value of proximal and distal as the SD of background air to calculate SNR.
Qualitative image analysis
Two independent musculoskeletal radiologists (L.I.S. and S.Y.S.) evaluated the image noise of the spinal cord, SCM, and back muscles at each level from C2-3 to C6-7 in the 4 sets of axial T2WIs including FSE, original PROPELLER, PROPELLER DLR50%, and PROEPLLER DLR75% sequences. Before analysing the images independently, they randomly analysed the images together for 10 to achieve a somewhat similar level of understanding. And then, 3 weeks later, the 2 radiologists independently performed qualitative analysis.
After the MR images were presented in a random order, they were assessed independently without knowledge of the sequence parameters.
Image noise were scored qualitatively as follows: none (0) = no visible noise and clear anatomic structures; mild noise (1) = visible noise, but no problem in decision; moderate noise (2) = relatively indistinct decision by noise; severe noise (3) = not diagnostic by prominent noise.
The average scores of the 2 radiologists were used for statistical analysis.
Statistical analysis
Using the values of SIs obtained through ROIs of the spinal cord, SCM muscle, and background, the SNR of the spinal cord and SCM muscle and the CNR of the spinal cord were obtained. For quantitative analysis, all numerical values reported as mean ± SD.
A paired t-test was used to determine whether the overall SNR and CNR were significantly different between the original PROPELLER and PROPELLER DLR75% images using all ROI values measured in various anatomical structures of all patients and all disc levels. To obtain the coefficient of variation of the foreground, MRQy, an open-source tool for quality control of MR imaging data, was used.10
Analysis of variance and post-hoc t-tests were used to determine whether there was a difference in the SNR and CNR between each sequence. In ANOVA analysis, P < .05 was considered to be significant, but in post-hoc t-test, P < .0083 (0.05/6, applied Bonferroni correction for multiple comparison correction) was considered to be significant.
For qualitative analysis, the mean values of subjective scores measured by 2 radiologists for the spinal cord, SCM muscle, and back muscles were used. Additionally, ANOVA and post-hoc t-tests were used to investigate whether there was a difference in image noise for the spinal cord, SCM, and back muscles between each image. Inter-reader agreements for qualitative scores of the spinal cord, SCM, and back muscles were assessed by using a linear-weighted kappa (κ) statistic for qualitative analysis with the following scale: 0-0.20, poor; 0.21-0.40, fair; 0.41-0.60, moderate; 0.61-0.80, good; and 0.81-1.00, excellent.
All statistical analyses were performed using SPSS Statistics for Windows (version 25, IBM).
Results
The overall SNR and CNR obtained by using ROIs measured in various anatomic structures of all patients and all disc levels showed a significant difference between the original PROPELLER image and the PROPELLER DLR75% image (P < .0001); the overall SNR value = original PROPELLER (34.98 ± 13.9) & PROPELLER DLR75% (44.32 ± 19.5), the overall CNR value = original PROPELLER (9.87 ± 4.0) & PROPELLER DLR75% (11.17 ± 4.4). The coefficient of variation of the foreground for shadowing and inhomogeneous artefacts showed no significant difference between these 2 images (P = .79).
Quantitative analysis
The overall SNR values are as follows; FSE (47.74 ± 18.24), original PROPELLER (54.59 ± 21.06), PROPELLER DLR50% (85.64 ± 34.84), and PROPELLER DLR75% (135.36 ± 71.69) for spinal cord, and FSE (24.33 ± 9.10), original PROPELLER (27.79 ± 10.36), PROPELLER DLR50% (43.57 ± 17.09), and PROPELLER DLR75% (68.91 ± 35.54) for SCM muscle. The overall CNR values of spinal cord are as follows; FSE (2.94 ± 0.84), original PROPELLER (3.21 ± 0.75), PROPELLER DLR50% (3.23 ± 0.78), and PROPELLER DLR75% (3.25 ± 0.79).
The mean and SD of the SNR values of the spinal cord and SCM muscle at each cervical disc level are shown in Table 2.
Table 2.
Mean and standard deviation of signal-to-noise (SNR) values of the spinal cord and sternocleidomastoid (SCM) muscle at each cervical disc level according to each sequence of T2-weighted images.
| Level | No.a | Sequences | Spinal cord | P-valueb | SCM muscle | P-valueb |
|---|---|---|---|---|---|---|
| C2-3 | 1 | FSE | 41.63 ± 15.98 | 1&3, 1&4, 2&3, 2&4, <.0001 | 13.36 ± 3.84 | 1&3, 1&4, 2&3, 2&4, <.0001 |
| 2 | PROPELLER | 45.99 ± 12.14 | 13.81 ± 3.62 | |||
| 3 | PROPELLER DLR50% | 60.81 ± 22.56 | 18.27 ± 7.14 | |||
| 4 | PROPELLER DLR75% | 79.20 ± 36.57 | 23.84 ± 11.33 | |||
| C3-4 | 1 | FSE | 38.12 ± 9.24 | 1&3, 1&4, 2&3, 2&4, 3&4, <.0001 | 13.00 ± 3.93 | 1&4, 2&3, 2&4, 3&4, <.0001 |
| 2 | PROPELLER | 40.76 ± 10.30 | 12.56 ± 4.88 | |||
| 3 | PROPELLER DLR50% | 53.72 ± 15.97 | 16.90 ± 9.58 | |||
| 4 | PROPELLER DLR75% | 71.48 ± 29.08 | 22.73 ± 17.30 | |||
| C4-5 | 1 | FSE | 35.10 ± 7.06 | 1&3, 1&4, 2&3, 2&4, 3&4, <.0001 | 12.60 ± 3.45 | 1&4, 2&3, 2&4, 3&4, <.0001 |
| 2 | PROPELLER | 36.97 ± 8.27 | 11.86 ± 3.99 | |||
| 3 | PROPELLER DLR50% | 47.86 ± 13.27 | 15.47 ± 7.96 | |||
| 4 | PROPELLER DLR75% | 64.05 ± 27.46 | 20.72 ± 13.14 | |||
| C5-6 | 1 | FSE | 36.46 ± 8.46 | 1&3, 1&4, 2&3, 3&4, <.0001 | 13.53 ± 4.45 | 1&4, 2&3, 2&4, 3&4, <.0001 |
| 2 | PROPELLER | 35.94 ± 9.47 | 12.29 ± 4.35 | |||
| 3 | PROPELLER DLR50% | 47.80 ± 16.91 | 16.62 ± 8.46 | |||
| 4 | PROPELLER DLR75% | 68.13 ± 34.86 | 23.70 ± 16.19 | |||
| C6-7 | 1 | FSE | 33.37 ± 8.78 | 1&3, 1&4, 2&3, 2&4, 3&4, <.0001 | 13.08 ± 4.00 | 1&4, 2&4, 3&4, <.0001 |
| 2 | PROPELLER | 33.04 ± 10.35 | 11.51 ± 4.08 | |||
| 3 | PROPELLER DLR50% | 43.77 ± 16.71 | 15.56 ± 7.46 | |||
| 4 | PROPELLER DLR75% | 65.60 ± 31.68 | 23.44 ± 14.48 |
By assigning unique numbers to the 4 sequences, we tried to express them more simply when presenting statistical results.
Post-hoc t-test (P < .0083, significant).
Abbreviations: DLR = deep learning reconstruction; FSE = fast spin echo image; PROPELLER = periodically rotated overlapping parallel lines with enhanced reconstruction.
The mean SNR values for the spinal cord of all disc levels and SCM muscles from C2-3 to C5-6 level were significantly higher in PROPELLER DLR50% and DLR75% compared to FSE and original PROPELLER images P < .0083). From C3-4 to C6-7 level, the mean SNR values of the spinal cord and SCM muscles were also significantly different between PROPELLER DLR50% and PROPELLER DLR75% (P < .0083). Therefore, the mean SNR values increased significantly from the FSE to the PROPELLER DLR 75% images (Figure 2).
Figure 2.
Mean signal-to-noise ratio (SNR) values of the spinal cord and sternocleidomastoid (SCM) muscle according to each disc level and imaging sequence. The mean SNR values for the spinal cord of all disc levels and SCM muscles from C2-3 to C5-6 level were significantly higher in PROPELLER DLR50% and DLR75% compared to fast spin echo (FSE) and original PROPELLER images (P < .0083). The line with * indicates significance in post-hoc t-tests with a threshold of P<.0083 (adjusted for multiple comparisons using Bonferroni correction). Abbreviations: DLR = deep learning reconstruction; PROPELLER = periodically rotated overlapping parallel lines with enhanced reconstruction.
The mean and SD values of the CNR of the spinal cord at each cervical disc level are presented in Table 3. The mean CNR values of the spinal cord were significantly higher in PROPELLER DLR50% and DLR75 compared to FSE images at the C3-4 and 4-5 levels and PROPELLER DLR75% compared to FSE at the C6-7 level (P < .0083) (Figure 3).
Table 3.
Mean and standard deviation of the contrast-to-noise (CNR) values of the spinal cord at each cervical disc level.
| Level | No.a | sequences | Spinal cord | P-valueb |
|---|---|---|---|---|
| C2-3 | 1 | FSE | 3.13±0.68 | All >.0083 |
| 2 | PROPELLER | 3.38±0.59 | ||
| 3 | PROPELLER DLR50% | 3.41±0.65 | ||
| 4 | PROPELLER DLR75% | 3.42±0.72 | ||
| C3-4 | 1 | FSE | 3.06±0.81 | 1&3, .007/1&4, .007 |
| 2 | PROPELLER | 3.41±0.75 | ||
| 3 | PROPELLER DLR50% | 3.40±0.63 | ||
| 4 | PROPELLER DLR75% | 3.44±0.68 | ||
| C4-5 | 1 | FSE | 2.92±0.73 | 1&2, .008/1&3, .004/1&4, .003 |
| 2 | PROPELLER | 3.23±0.57 | ||
| 3 | PROPELLER DLR50% | 3.28±0.58 | ||
| 4 | PROPELLER DLR75% | 3.28±0.56 | ||
| C5-6 | 1 | FSE | 2.89±0.88 | All >.0083 |
| 2 | PROPELLER | 3.06±0.74 | ||
| 3 | PROPELLER DLR50% | 3.06±0.82 | ||
| 4 | PROPELLER DLR75% | 3.09±0.79 | ||
| C6-7 | 1 | FSE | 2.71±1.04 | 1&4, .006 |
| 2 | PROPELLER | 2.99±0.96 | ||
| 3 | PROPELLER DLR50% | 2.99±1.06 | ||
| 4 | PROPELLER DLR75% | 3.02±1.06 |
By assigning unique numbers to the 4 sequences, we tried to express them more simply when presenting statistical results.
Post-hoc t-test (P < .0083, significant).
Abbreviations: DLR = deep learning reconstruction; FSE = fast spin echo image; PROPELLER = periodically rotated overlapping parallel lines with enhanced reconstruction.
Figure 3.
Mean contrast-to-noise ratio (CNR) values of the spinal cord at each cervical disc level. The mean CNR values of the spinal cord were significantly higher in PROPELLER, PROPELLER with DLR50%, and PROPELLER with DLR75% images than in FSE at C3-4 and C4-5 levels and PROPELLER with DLR75% compared to FSE at C6-7 levels (P < .0083). The line with * indicates significance in post-hoc t-tests with a threshold of P<.0083 (adjusted for multiple comparisons using Bonferroni correction). Abbreviations: DLR = deep learning reconstruction; FSE = fast spin echo; PROPELLER = periodically rotated overlapping parallel lines with enhanced reconstruction.
Qualitative analysis
The kappa values of inter-reader agreement were all 1 (excellent agreement) for the image noise scores of the spinal cord, SCM, and back muscles (P < .0001).
In the qualitative analysis of image noise on the spinal cord, SCM, and back muscles, PROPELLER DLR50% and PROPELLER DLR75% images showed a significant denoising effect compared to the FSE and original PROPELLER images. However, there were no significant differences between the FSE and original PROPELLER images and between the PROPELLER DLR50% and PROPELLER DLR75% images (Figures 4 and 5).
Figure 4.
Qualitative analysis of image noise in the spinal cord (A), sternocleidomastoid (SCM) (B), and back muscles (C). The PROPELLER DLR50% and PROPELLER DLR75% images showed a significant denoising effect compared to the fast spin echo (FSE) and original PROPELLER images. However, there was no significant difference between the PROPELLER DLR50% and PROPELLER DLR75% images. The line with * indicates significance in post-hoc t-tests with a threshold of P <.0083 (adjusted for multiple comparisons using Bonferroni correction). Abbreviations: DLR = deep learning reconstruction; PROPELLER = periodically rotated overlapping parallel lines with enhanced reconstruction.
Figure 5.
Qualitative or visual analysis of axial T2 fast spin echo (FSE), original PROPELLER, PROPELLER DLR 50%, and PROPELLER DLR75%. On the axial T2 FSE image (A), signal intensities of the spinal cord and muscles including the SCM and back muscles are heterogeneous and coarse, and muscular structures appear to be blurred. Although the blurring or motion artefacts of the SCM and back muscles are decreased in the original PROPELLER image (B) compared to the FSE image, the signal intensity of the spinal cord still appears to be heterogeneous. The prevertebral soft tissues showed heterogeneous signal intensities and blurring due to motion. PROPELLER DLR75% (D) shows more homogenous signal intensities of the spinal cord and muscular structures compared to PROPELLER DLR50% (C), although both DLR 75% and 50% images demonstrate better image quality compared to the FSE and original PROPELLER images. Abbreviations: DLR = deep learning reconstruction; PROPELLER = periodically rotated overlapping parallel lines with enhanced reconstruction; SCM = sternocleidomastoid.
Discussion
Many clinical applications require in greater image quality improvement in terms of SNR and spatial resolution rather than additional acceleration.7 Spine MR images usually have more inhomogeneous signal strength because spinal MR images use a phased-array surface coil, which has a lower coil sensitivity toward the deeper portions from the surface.11 In particular, cervical spine MRI has a more pronounced image blurring effect due to a patient’s unconscious swallowing than other levels. Thus, spinal cord evaluation is often limited because of inhomogeneous SI.
To reduce motion artefacts associated with a long acquisition time, the most effective solution is shorten the image acquisition time. Several techniques for reducing the image acquisition time have been used, including parallel imaging and compressed sensing.12 Pipe2 stated that the PROPELLER technique is well suited for imaging moving objects because of its inherent ability to reject some of the in-plane and through-plane motions and its inherent averaging of the remaining data inconsistencies. Fellner et al5 evaluated sagittal T2-weighted PROPELLER MRI in the cervical spine and found a significantly increased SNR in the vertebral body and spinal cord for the PROPELLER technique compared with the standard FSE sequence. In our study, the SNR and CNR of the original PROPELLER images were higher than those of the FSE images; however, the difference was not statistically significant. Qualitative analysis also showed differences in image quality between the FSE and PROPELLER images at some levels, but most were not statistically significant. Therefore, there seems to be a limit to improving the quality of noisy images obtained by motion using only the simple PROPELLER technique.
Recent advances in DL have improved the computational cost, training time, and amount of data required, which have led to its application in medical imaging denoising.13 DL is a set of techniques and algorithms that enable computers to discover complicated patterns in large datasets.14 The DLR method used in our study includes deep convolutional neural networks (CNNs) that operate on raw, complex-valued imaging data to produce a clean output image. Specifically, the CNN is designed to provide a user-tunable reduction in image noise, reduce truncation artefacts, and improve edge sharpness.7
The DLR was designed to consistently suppress ringing artefacts without reducing the resolution, while allowing the user to adjust the denoising level based on individual preferences. Thus, the DLR is an image reconstruction pipeline leveraging artificial intelligence that improves image SNR and sharpness while reducing truncation artefacts.7 Kidoh et al15 reported that DLR significantly reduced image noise while preserving image quality for brain MR images obtained in a relatively short acquisition time. Moreover, they suggested that a denoising method using DL works effectively even in images with small anatomical structures. In our study, additional DLR achieved a significant elevation of the SNR without decreasing tissue contrast. In addition, with the use of DLR in addition to PROPELLER for MRI of the cervical spine, motion artefacts were significantly reduced, and the overall image quality and delineation for all investigated anatomic structures could be improved compared with conventional FSE images.
Optimization of the pre-DLR image quality as well as the level of DLR denoising according to the type of imaging sequence used and the anatomical targets of interest are required when using DLR, especially in the evaluation of precise anatomical regions.15 We also obtained images with 2 different noise-reduction factors (DLR 50% and 75%) to assess the denoising effect of the DLR. In our study, the DLR50% image showed no significant difference in SNR and CNR compared to the DLR75% image in some cases, but the DLR75% image had an overall higher SNR and CNR.
Limitations
This study had some limitations. First, we evaluated the performance of DLR when applied only to axial T2WIs because it is a key sequence for the diagnosis of various spinal disorders. Further studies are warranted to evaluate the DLR performance in other sequences, such as gradient echo images or diffusion-WIs. Second, we did not evaluate pathological lesions in this study. It is usually difficult to evaluate signal alterations in the spinal cord. Therefore, pathologic conditions were excluded because we wanted to determine how homogeneously the normal spinal cord could be seen by applying DL reconstruction. Third, the number of patients included in this study was small and the study period was also short. However, the amount of data used for analysis was not small because 4 sets of different images were analysed from the C2-3 to C6-7 level.
Conclusions
Axial T2 PROPELLER imaging of the cervical spine demonstrated improved image quality with higher CNR and SNR and reduced image noise by adding a DLR. Therefore, to better evaluate the spinal cord and surrounding soft tissue structures in the cervical spine, PROPELLER with DLR is recommended, which increases CNR and SNR while reducing motion artefacts compared to the conventional FSE sequence.
Contributor Information
You Seon Song, Pusan National University School of Medicine, Busan, Korea; Department of Radiology, Pusan National University Hospital, Biomedical Research Institute, Busan 49241, Korea.
In Sook Lee, Pusan National University School of Medicine, Busan, Korea; Department of Radiology, Pusan National University Hospital, Biomedical Research Institute, Busan 49241, Korea.
Moonjung Hwang, GE Healthcare, 15F Seoul Square 416, Seoul, Seoul 04367, Korea.
Kyoungeun Jang, AIRS Medical, 13-14F, Keungil Tower, Seoul, Seoul 06142, Korea.
Xinzeng Wang, GE Healthcare, MR Clinical Solutions & Research Collaborations, Houston, Texas 77081, United States.
Maggie Fung, GE Healthcare, MR Clinical Solutions & Research Collaborations, New York, NY 10032, United States.
Funding
None declared.
Conflicts of interest
None declared.
References
- 1. Jonkman LE, Klaver R, Fleysher L, Inglese M, Geurts JJ.. Ultra-high-field MRI visualization of cortical multiple sclerosis lesions with T2 and T2: a postmortem MRI and histopathology study. AJNR Am J Neuroradiol. 2015;36(11):2062-2067. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. Pipe JG. Motion correction with PROPELLER MRI: application to head motion and free-breathing cardiac imaging. Magn Reson Med. 1999;42(5):963-969. [DOI] [PubMed] [Google Scholar]
- 3. Alibek S, Adamietz B, Cavallaro A, et al. Contrast-enhanced T1-weighted fluid-attenuated inversion-recovery BLADE magnetic resonance imaging of the brain: an alternative to spin-echo technique for detection of brain lesions in the unsedated pediatric patient? Acad Radiol. 2008;15(8):986-995. [DOI] [PubMed] [Google Scholar]
- 4. Hirokawa Y, Isoda H, Maetani Y, Arizono S, Shimada K, Togashi K.. Evaluation of motion correction effect and image quality with the periodically rotated overlapping parallel lines with enhanced reconstruction (PROPELLER) (BLADE) and parallel imaging acquisition technique in the upper abdomen. J Magn Reson Imaging. 2008;28(4):957-962. [DOI] [PubMed] [Google Scholar]
- 5. Fellner C, Menzel C, Fellner F, et al. BLADE in sagittal T2-weighted MR imaging of the cervical spine. AJNR Am J Neuroradiol. 2010;31(4):674-681. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Kustner T, Liebgott A, Mauch L, et al. Automated reference-free detection of motion artifacts in magnetic resonance images. MAGMA. 2018;31(2):243-256. [DOI] [PubMed] [Google Scholar]
- 7. Lebel RM. 2020. Performance characterization of a novel deep learning-based MR image reconstruction pipeline. arXiv, arxiv.org/abs/2008.06559, preprint: not peer reviewed.
- 8. Wang X, Litwiller D, Lebel M, et al. High resolution T2W imaging using deep learning reconstruction and reduced field-of-view PROPELLER. In: Proceedings of the ISMRM. 2020 ISMRM & SMRT conference & exhibition; 2020.
- 9. Wang X, Litwiller D, Ersoz A, et al. Diffusion weighted imaging using PROPELLER acquisition and a deep learning based reconstruction. In: Proceedings of the ISMRM. 2020 ISMRM & SMRT conference & exhibition; 2020.
- 10. Sadri AR, Janowczyk A, Zhou R, et al. Technical note: MRQy-an open-source tool for quality control of MR imaging data. Med Phys. 2020;47(12):6029-6038. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Dietrich O, Raya JG, Reeder SB, et al. Measurement of signal-to-noise ratios in MR images: influence of multichannel coils, parallel imaging, and reconstruction filters. J Magn Reson Imaging. 2007;26(2):375-385. [DOI] [PubMed] [Google Scholar]
- 12. Haji-Valizadeh H, Rahsepar AA, Collins JD, et al. ; CKD Optimal Management with Binders and NicotinamidE (COMBINE) Study Group. Validation of highly accelerated real-time cardiac cine MRI with radial k-space sampling and compressed sensing in patients at 1.5T and 3T. Magn Reson Med. 2018;79(5):2745-2751. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Zhang K, Zuo W, Chen Y, et al. Beyond a Gaussian denoiser: residual learning of deep CNN for image denoising. IEEE Trans Image Process. 2017;26(7):3142-3155. [DOI] [PubMed] [Google Scholar]
- 14. Lundervold AS, Lundervold A.. An overview of deep learning in medical imaging focusing on MRI. Z Med Phys. 2019;29(2):102-127. [DOI] [PubMed] [Google Scholar]
- 15. Kidoh M, Shinoda K, Kitajima M, et al. Deep learning based noise reduction for brain MR imaging: Tests on phantoms and healthy volunteers. Magn Reson Med Sci. 2020;19(3):195-206. [DOI] [PMC free article] [PubMed] [Google Scholar]