Improved Nerve Conspicuity with Water-Weighting and Denoising in Two-Point Dixon Magnetic Resonance Neurography

Abstract

Background:

T₂-weighted, two-point Dixon fast-spin-echo (FSE) is an effective technique for magnetic resonance neurography (MRN) that can provide quantitative assessment of muscle denervation. Low signal-to-noise ratio and inadequate fat suppression, however, can impede accurate interpretation.

Purpose:

To quantify effects of principal component analysis (PCA) denoising on tissue signal intensities and fat fraction (FF) and to determine qualitative image quality improvements from both denoising and water-weighting (WW) algorithms to improve nerve conspicuity and fat suppression.

Study Type:

Prospective.

Subjects:

Twenty-one subjects undergoing MR neurography evaluation (11/10 male/female, mean age=46.3+/−13.7 years) with 60 image volumes. Twelve subjects (23 image volumes) were determined to have muscle denervation based on diffusely elevated T₂ signal intensity.

Field Strength/Sequence:

3T, 2D, two-point Dixon FSE.

Assessment:

Qualitative assessment included overall image quality, nerve conspicuity, fat suppression, pulsation and ringing artifacts by 3 radiologists separately on a three-point scale (1=poor, 2=average, 3=excellent). Quantitative measurements for FF and signal intensity relative to normal muscle were made for nerve, abnormal muscle and subcutaneous fat.

Statistical Tests:

Linear and ordinal regression models were used for quantitative and qualitative comparisons, respectively; 95% confidence intervals (CIs) and p-values for pairwise comparisons were adjusted using the Holm-Bonferroni method. Inter-rater agreement was assessed using Gwet’s agreement coefficient (AC₂).

Results:

Simulations showed PCA-denoising reduced FF error from 2.0% to 1.0%, and from 7.6% to 3.1% at noise levels of 10% and 30%, respectively. In human subjects, PCA-denoising did not change signal levels and FF quantitatively. WW decreased fat signal significantly (−83.6%, p<0.001). Nerve conspicuity was improved by WW (odds ratio, OR=5.8, p<0.001). Fat suppression was improved by both PCA (OR=3.6, p<0.001) and WW (OR=2.2, p<0.001). Overall image quality was improved by PCA+WW (OR=1.7, p=0.04).

Conclusions:

WW and PCA-denoising improved nerve conspicuity and fat suppression in MR neurography. Denoising can potentially provide improved accuracy of FF maps for assessing fat-infiltrated muscle.

1. Introduction

Peripheral nerve MRI (MR neurography, MRN) is effectively performed with T₂-weighted imaging[1] using a fast spin echo (FSE) readout to exploit the contrast between the higher T₂ of nerves and the lower T₂ of surrounding soft tissue. T₂-weighted imaging also highlights muscle edema-like patterns in actively denervated muscle[2], which may be secondary to extracellular fluid accumulation from increased muscle perfusion and capillary permeability [3]. T₂-weighted FSE requires effective fat suppression, as fat has a high T₂ value and therefore appears hyperintense relative to other soft tissues. Fat suppression is especially important in MRN, where hyperintense perineural fat obscures adjacent, small nerves. Although inversion recovery[4] and chemical saturation[5] are used routinely for fat suppression in musculoskeletal imaging[6], Dixon-based FSE provides high signal-to-noise ratio (SNR) relative to inversion recovery-based techniques and superior robustness to B₀-inhomogeneity relative to chemical fat saturation [7]. Dixon-based FSE[8,9] is therefore particularly effective for fat suppression in MRN [10], including of the brachial plexus[11].

In Dixon-based fat suppression, images are acquired at different echo times (TE) that have different phases between spins at “water” resonance and “fat” resonance. “Water” and “fat” images are resolved by either subtracting echoes with opposite phase shifts or by solving a system of equations from three or more echoes acquired at varying phase shifts[12]. Prior to solving these equations, Dixon-processing requires removal of the confounding phase due to B₀-inhomogeneity for which several effective methods, including phase-fitting[13] and region-growing[14], have become routinely used. In Dixon-based FSE, each additional echo increases scan time proportionately to maintain the same echo train length (ETL) [9]. MRN requires high spatial resolution (~0.3–0.5 mm in-plane) to visualize small nerves and their fascicular architecture, and to delineate these from the surrounding soft tissues; consequently, scan times can be long (4–7 minutes). Since two-point or two-echo Dixon is effective at providing spatially uniform fat suppression, Dixon T₂-weighted FSE is often acquired with just two echoes to minimize scan time.

Dixon-based T₂-weighted FSE currently has several limitations. First, the contrast between fat and nerves can be sub-optimal due to the wide spectrum of frequency shift and T₂ in fatty tissue[15]. Incomplete fat suppression can result in reduced conspicuity of small nerves in regions directly adjacent to perineural fat. Neither adding echoes nor changing TE improves fat suppression due to the need for contrast against both low-T₂ muscle and high-T₂ fat. Second, the need for high spatial resolution in MRN often results in sub-optimal SNR, which motivates performing image denoising. A promising denoising approach utilizes principal components analysis (PCA) across multiple images by reorganizing multi-echo data as matrices and attributing noise to the smaller principal components[16]. While PCA-denoising may be readily applied to T₂-mapping by reorganizing matrices in the echo- and spatial-dimensions[17], having only two echoes in MRN makes it practically infeasible to perform PCA with a matrix size of 2. However, if data from multi-channel phased-array coils are considered, the matrix dimensionality of Dixon-FSE can be increased to make it suitable for PCA-denoising.

An additional motivation for denoising is fat fraction (FF) quantification, which requires high SNR[18]. During the chronic phase of muscle denervation, the presence and extent of fat infiltration may be assessed qualitatively from the “water” image without relying on the “fat” image. However, quantitative estimation of the MRI-based fat content[18], while different from actual fat content[19], may help characterize the extent and stage of muscle denervation or muscle function loss[20]. While FF quantification is more efficiently performed using gradient recalled echo readout[12], the ability to quantify FF using an FSE readout[21] that can be performed in combination with T₂-mapping[22,23] would add to the utility of this MRN pulse sequence.

This study’s objective was to improve key image quality metrics in two-point Dixon FSE MRN by exploiting the “fat” image. We hypothesized that the proposed water-weighting and PCA-denoising would improve overall image quality, nerve conspicuity and fat suppression in two-point Dixon FSE MRN.

2. Material and Methods

Fig. 1a shows the overall processing steps for the PCA-denoising and water-weighting methods introduced in this work. These steps are described in more detail below.

Figure 1.

Schematics of methodology, including (a) a flowchart of processing steps in this work, including principal components analysis (PCA)-denoising and water-weighting processing, for generating water (W), fat (F) and fat fraction (FF) images; (b) organization of image kernel data into matrix S for PCA.

2.1. Denoising with Principal Components Analysis

In PCA-denoising[16], the image data is first organized into an M-by-N matrix S and undergoes eigenvalue decomposition into its eigenvector matrices U and V, and eigenvalue matrix Λ, which is a diagonal matrix of values with scalar values Λ_i,i sorted in decreasing order where i ∈ [1,N]. When M=N:

$$\mathbf{S}=\mathbf{U\Lambda }{\mathbf{V}}^T,\mathbf{\Lambda }=\left[\begin{array}{ccc}{\Lambda }_{1,1}& \cdots & 0\\ ⋮& \ddots & ⋮\\ 0& \cdots & {\Lambda }_{N,N}\end{array}\right].$$ (1)

The denoised eigenvalue matrix Λ′ is obtained by nulling all Λ_i,i except the P largest eigenvalues, where 1 ≤ P ≤ {M, N}. When M ≠ N, singular value decomposition may instead be used without loss of generality. The denoised image data S′ is obtained with S′ = UΛ′V^T. PCA-denoising may be applied ascribing the number of echoes to one dimension (M), and pixels to the other (N) as suggested previously[17]. However, the choice of these dimensions could be sub-optimal for denoising two-point Dixon data as the rank of Λ is low at 2 (M=2).

To increase dimensions and rank of S (and Λ), we propose utilizing the channel dimension from multi-channel phased-array coils, similar to that proposed elsewhere [24]. This requires denoising to be performed just prior to complex coil combination, which means the data would be complex, containing real and imaginary components. In addition, the image can be separated into real/imaginary channels to double the dimensionality. In this work, the first dimension (M) consisted of echoes, the image-kernel (of 3×3 to 7×7 pixels), concatenated by real/imaginary channels; the second dimension (N) was for coil channels. The matrix organization is illustrated in Fig. 1b. For a 3×3 kernel, two real/imaginary channels and two echoes for two-point Dixon, and 32-channel phased-array coil, one obtains MxN=36×32, which would have sufficiently high matrix rank.

2.2. Water Weighting (WW)

To exploit the fat image for further suppression of fat signal in the water image, we propose water-weighting (WW) processing following PCA-denoising. WW exploits the weighting from FF, which is defined by the ratio between fat (F) and water (W) signals, FF=F/(W+F) and is sensitive to noise due to normalization. With PCA-denoising, the noise in FF will be reduced, which incidentally also benefits WW processing. To apply WW, we propose a simple weighting of:

$$W_{\text{WW}}=W(1-\text{FF})=\frac{W^2}{(W+F)}.$$ (2)

The rationale for this weighting is to further suppress fat signal while maintaining the contrast of most tissue in the WW in a similar range as the original water image. Signal intensities in image regions with high FF will be significantly attenuated, while those with low FF will be minimally attenuated, creating increased contrast of water signal. It is expected that nerve conspicuity will be improved, as fat suppression will be improved. The contrast of non-denervated and denervated but non-fat-infiltrated muscle (low FF) would also be expected to increase compared to fat-infiltrated denervated muscle, but this was not evaluated in this work.

2.4. Simulations - PCA

To determine the effects from PCA-denoising, signal simulations were performed at different levels of noise. The signals were processed using mostly the same steps as the in vivo MRI data (Fig. 1a). The simulated signals were generated using coil-sensitivities ρ_c obtained from 400 representative pixels acquired from a 16-channel phased-array coil on human subject data. The impact of additive Gaussian noise, η was tested at 7 different noise levels ranging from 0.1% to 100%. To exclude possible effects on the phase from PCA-denoising that could impact Dixon-processing, the B₀ differences between the in-phase (s_in) and out-of-phase signals (s_out) were simulated with a small amount of B₀ inhomogeneity, leading to phase accrual θ (−30° to +30 °) between the in-phase (s_in) and out-of-phase (s_out) signals from the c-th coil:

$$s_{in,c}={\rho }_c(W+F)+\eta ,\text{and}$$ (3)

$$s_{out,c}={\rho }_c{e}^{i\theta }(W+F)+\eta .$$ (4)

The signals were combined in the coil-dimension using optimal reconstruction[25] to obtain the coil-combined signals s_in and s_out. To perform B₀-correction, the measured phase accrual, $\varphi =\measuredangle \left({\overline{\mathbf{s}}}_{out}{\mathbf{s}}_{in}\right)$, was used to correct the out-of-phase signal under the premise that ϕ ≈ −θ if there were no deleterious effects from PCA:

$${s^{\prime }}_{out,c}=e^{i\varphi }{s}_{out,c}.$$ (5)

The simulated water and fat signals were then obtained using the standard Dixon-based addition and subtraction of s_in and s’_out and sum-of-squares coil combination. These steps were repeated without and with PCA-denoising and at different PCA thresholds, P to determine the optimum threshold.

2.5. MR Imaging

For in vivo imaging, 21 subjects (11 male/10 female; age 46.3+/−13.7 years old) undergoing MRN of varying anatomies were prospectively included in this IRB-approved study. The clinical indications for MRI included suspected brachial plexopathy (n= 5), Parsonage-Turner syndrome (n=4), pain and weakness following motor vehicle or bicycle accident (n=3), lower lumbosacral plexopathy or proximal sciatic neuropathy (n=2), concern for mass (n=2), finger pain, paresthesia or weakness (n=2), foot drop (n= 1), and pudendal neuralgia (n=1). 2D, two-point-Dixon, T₂-weighted FSE images, acquired as part of standard-of-care, used parameters that varied depending on the imaged anatomy (TR/TE=3077–7520/78.7–88.5 ms, FOV=10–36 cm, matrix=320×224, 30–100 slices, slice thickness=1.2–4.0 mm (no slice gap), ETL=11–18, bandwidth-per-pixel=244–391 Hz). Acquisitions were performed at 3T MRI (Signa Premier, and MR750, GE Healthcare, Waukesha, WI, USA) using either one or two 16-channel flexible coils or a 32-channel torso coil, depending on the anatomy imaged. Raw data was reconstructed offline using in-house code developed on Matlab (Mathworks Inc., Natick, USA); image reconstruction was performed without and with PCA. A 3×3 kernel was used in PCA (Fig. 1b), with real and imaginary components separated for PCA matrix sizes of either MxN=36×16 or 36×32 depending on the coils used. The optimum number of eigenvalues being retained, P, was determined from the simulations. Phase-fitting was applied for B₀-correction. The same P was used for the entire image, so as not to introduce any spatially varying artifacts. WW-processing was applied to the PCA images. The processing times for PCA and water-weighting for each image series were 26–58 minutes and <1 minute, respectively (on a 2.8GHz Pentium IV 16GB RAM workstation).

2.6. Hypotheses and Statistical Analysis

We hypothesized that PCA-denoising would reduce FF in normal muscle and increase FF in subcutaneous fat, without significantly changing water and fat signals in nerve, abnormal muscle, or subcutaneous fat ROIs. To test this hypothesis, the mean values of water signal, fat signal, and FF from the standard vs. PCA images for each ROI were measured. In order to facilitate averaging of signal intensities (water and fat) across subjects, the signals for nerve, subcutaneous fat, and abnormal muscle ROIs were normalized to that of the normal muscle ROI. As FF is a normalized quantity, normalization of FF was not performed. The FF was adjusted for relaxation by assuming a normal muscle T₂* of 21.6 ms. We hypothesized that WW would improve fat suppression by reducing fat signal without reducing water signal. The water and fat signals from the PCA and WW+PCA images were also measured and normalized to normal muscle ROIs. For quantitative outcomes, marginal linear regression models estimated with small sample bias-corrected empirical standard errors were specified to evaluate for mean differences between image type by ROI [26]. Subjects were treated as the repeated factor to account for within-patient correlations between image type and ROIs.

We also hypothesized that, qualitatively, PCA-denoising and WW would improve overall image quality, nerve conspicuity, and fat suppression without increasing pulsation and ringing artifacts. With reduced noise and a new intensity-weighting scheme, it was conceivable that such artifacts common to MRN could become more pronounced. Three radiologists with 7 to 10 years of dedicated musculoskeletal MR experience (YE, AJB, DBS) graded the water images from the Dixon-FSE acquisition. These were assessed for overall image quality, nerve conspicuity (i.e. ability to confidently visualize the nerve, separate from the surrounding soft tissue), fat suppression (1=poor, 2=average, 3=excellent), and pulsation and ringing artifacts (1=none, 2=mild, 3=severe). For each qualitative outcome, an ordinal logistic regression model with random subject effects was specified to compare standard, PCA, and WW+PCA images, accounting for the three readers. Odds ratios (OR) were estimated from the three pairwise image comparisons (PCA vs standard, WW+PCA vs PCA, WW+PCA vs Standard) with 95% confidence intervals (CIs) and p-values subsequently adjusted using the Holm-Bonferroni method. Statistical significance was set at p<0.05. To assess interrater agreement of the readers’ assessments, ordinal weighted Gwet’s agreement coefficients (AC₂) were estimated, with AC₂ ≥ 0.7 considered substantial. All statistical analyses were performed using SAS v9.4 (SAS Institute, Cary, NC, USA).

The ROIs were manually drawn by one radiologist with 7 years of musculoskeletal subspecialty experience (DBS) on the uncorrected water image – one nerve, subcutaneous fat, normal muscle (expected signal intensity), and abnormal muscle (diffusely increased T₂ signal intensity). The qualitative diagnosis of ‘normal’ versus ‘abnormal’ muscle was agreed upon independently by all three readers.

2.7. Muscle and nerve ROIs

Table 1 provides a summary of the locations of normal and abnormal muscle ROIs. Muscle abnormality was determined qualitatively by the presence of diffuse signal hyperintensity of the muscle on the T₂-weighted Dixon water image (without denoising). In total, 60 image volumes covering a broad range of anatomy (27 brachial plexus, 15 upper-arm/elbow, 18 pelvis/lower extremities) were evaluated in 21 subjects. All subjects had normal muscle as determined on MRI, with normal muscle ROIs drawn in 59 image volumes (1 image volume was excluded due to excessive motion). Twelve of the 21 subjects had abnormal muscle determined on MRI, with abnormal muscle ROIs drawn in 23 image volumes. There were no cases of fatty infiltrated muscle. EMGs were available in 14 of the 21 subjects. EMG demonstrated abnormal motor unit recruitment in 16 of the 23 abnormal muscles; the remaining 7 abnormal muscles were not tested on EMG.

Table 1.

Patient demographics and body parts imaged and analyzed

Age (Years)		Mean	SD	Minimum	Maximum
		46.3	13.7	19	65
Sex		Normal Muscle		Abnormal Muscle
		Subjects	Image Series	Subjects	Image Series
Male		3	30	8	17
Female		6	29	4	6
Total:		9	59	12	23
Body Part by MRI Study		Normal Muscle Image		Abnormal Muscle Image
Description		Series		Series
Brachial Plexus (Total):		15		-
Nerves:	C5 nerve root	1		-
	C6 nerve root	3		-
	C7 nerve root	2		-
	Anterior division of upper trunk	1		-
	Posterior division of upper trunk	2		-
	Medial cord	1		-
	Posterior cord	5		-
Trunk/shoulder girdle/neck (Total):		26		16
Muscles:	Supraspinatus	2		8
	Infraspinatus	2		2
	Subscapularis	0		1
	Teres minor	1		0
	Trapezius	10		1
	Deltoid	5		3
	Anterior Scalene	1		0
	Middle Scalene	4		0
	Posterior paraspinals	0		1
	Pectoralis major	1		0
Upper Extremities Nerves (Total):		12		-
Nerves:	Axillary	1		-
	Ulnar	7		-
	Median	3		-
	Radial	1		-
Upper Extremities Muscles (Total):		15		6
Muscles:	Biceps	4		0
	Brachialis	2		1
	Pronator Teres	2		0
	Flexor Carpi Ulnaris	2		2
	Flexor Carpi Radialis	1		0
	Flexor Digitorum Profundus	2		1
	Pronator Quadratus	1		0
	Flexor Pollicis Longus	0		2
	1st Dorsal Interosseous	1		0
Lumbar Plexus (Total):		18		-
Nerves:	L5 nerve root	3		-
	S1 nerve root	4		-
	S2 nerve root	1		-
	Sciatic nerve	10		-
Lower Extremities Muscles (Total):		18		1
Muscles:	Gluteus maximus	10		0
	Gluteus medius	2		0
	Gluteus minimus	3		0
	Piriformis	1		0
	Obturator internus	1		0
	Vastus medialis	1		0
	Long head bicep femoris	0		1

In total, 45 nerve ROIs were drawn, undifferentiated by the presence of a nerve abnormality. Fifteen nerve ROIs could not be drawn reliably because the nerves were either too small or the acquired images were not orthogonal to the longitudinal course of the nerve.

3. Results

3.1. Simulations

Fig. 2a–b show that the error in FF (measured FF minus true FF) was low at low noise and was not improved significantly with PCA-denoising with P=1. Increased noise levels (Fig. 2c) increased FF error, which showed positive bias at low FF and negative bias at high FF. The FF error was reduced significantly with PCA-denoising (Fig. 2d). Also, there were no increased errors due to B₀ inhomogeneity, either with or without PCA-denoising.

Figure 2.

Simulation results of normalized mean error in computed fat fraction (FF) as a function of the simulated FF and off-resonance. At low noise of (a) 3% (b) PCA (P=1) did not result in significant improvement to FF error. When noise was increased to (c) 30%, bias in FF increased, with positive bias at low FF and negative bias at high FF. Applying (d) PCA-denoising to the 30% noise case reduced the bias to levels similar to that in 3%. The level of bias was invariant to the degree of off-resonance.

Fig. 3 summarizes the simulation results for different noise levels and values of P. At noise levels of 1% and below, the root-mean-square error (RMSE) was less than 0.1%. The error increased with increased noise levels. PCA-denoising with P=1 resulted in the lowest error at all noise levels compared to higher P and with no PCA-denoising. The largest improvement in error was seen with the highest noise level of 100% (RMSE reduced from 20.2% to 12.6%). However, when normalized to its non-PCA-denoised RMSE, the RMSE reduction was more substantial at noise levels of 30% (RMSE reduced from 7.6% to 3.1%) and 10% (RMSE reduced from 2.0% to 1.0%).

Figure 3.

Summary of root-mean-square error (RMSE) of fat fraction results simulated at various noise levels and without (none) and with PCA-denoising at different threshold P levels.

3.2. Quantitative Analysis

Using the previous simulation results, the lowest P=1 was used to process the in vivo images. PCA-denoising did not reduce the relative nerve, subcutaneous fat, and abnormal muscle signal (Table 2). PCA-denoising also did not significantly alter FF, but for subcutaneous fat the average change was positive (+0.1%) and for other ROIs the average change was negative (−0.9 to −0.3%). WW decreased the relative signal intensity of subcutaneous fat significantly (−83.6%, p<0.001), without significantly altering the relative signal intensities of nerve (−4.0%, p=0.66) and abnormal muscle (+6.0%, p=0.62).

Table 2.

Summary of quantitative results. The results indicate mean percentage changes in water image signal (normalized to normal muscle signal) and fat fraction (FF) from application of principal components analysis (PCA)-denoising and water-weighting (WW) in nerve, normal muscle, subcutaneous fat, and abnormal muscle. WW was applied on PCA images.

		Change in Water Image Signal Intensity (% Relative to Normal Muscle Signal)		Change in FF (%)
Quantitative		PCA vs Standard	WW+PCA vs PCA	PCA vs Standard
	Mean (95% CI) p-value	−2.2 (−19.0, 14.5)	−4.0 (−21.8, 13.8)	−0.3 (−2.2, 1.6)
Nerve	p-value	0.796	0.657	0.751
	Mean (95% CI)	-	-	−0.9 ( −2.7, 0.8)
Normal Muscle	p-value	-	-	0.301
	Mean (95% CI)	−2.1 (−16.6, 12.3)	−83.6 (−100.3, −66.8)	0.1 (−1.5, 1.8)
Subcutaneous Fat	p-value	0.770	<.001*	0.864
	Mean (95% CI)	−0.3 (−23.7, 23.0)	6.0 (−17.9, 30)	−0.3 (−2.9, 2.4)
Abnormal Muscle	p-value	0.977	0.621	0.845

3.3. Qualitative Analysis

Fig. 4 shows the distribution of qualitative results averaged across all readers. The fraction of scans rated as excellent for overall image quality was increased by PCA (from 19.0% to 21.0%), and further by WW (to 30.3%). For nerve conspicuity, scans rated as excellent were increased by PCA (from 25.7% to 27.0%) and further by WW (to 67.0%). For fat suppression, scans rated as excellent were increased by PCA (from 22.0% to 53.7%) and further by WW (to 69.7%). Artifact scores were mostly unchanged.

Figure 4.

Distribution of qualitative scoring, showing averaged fraction of responses from all readers with ratings (1=poor, 2=average, 3=excellent) for (a) overall image quality, (b) nerve conspicuity, and (c) fat suppression, and artifact ratings (1=none, 2=mild, 3=severe) for (d) pulsatility and (e) ringing.

Table 3 summarizes results of the paired comparisons. PCA resulted in significantly improved fat suppression by an OR=3.6 (p<0.001); the improvements in overall image quality (OR=1.08) and nerve conspicuity (OR=1.05) were, however, not statistically significant (p=0.72 to 0.81). WW resulted in significantly improved nerve conspicuity by OR=5.7 (p<0.001), and fat suppression by OR=2.2 (p<0.001). WW with PCA resulted in improved overall image quality by OR=1.7 (p=0.04), along with improved nerve conspicuity by OR=6.1 (p<0.001) and fat suppression by OR=8.1 (p<0.001). Pulsatility and ringing artifacts were unchanged by PCA with or without WW.

Table 3.

Summary of qualitative results comparing standard images vs. principal components analysis (PCA) denoising, vs. water-weighted (WW) images. WW was applied on PCA-denoised images. Results of the odds ratio (OR) were based on a 3-point assessment of image quality (IQ) or artifacts.

Qualitative		PCA vs Standard	WW vs PCA	WW vs Standard
	Odds Ratio (95% CI)	1.08 (0.65,1.78)	1.56 (0.94,2.57)	1.68 (1.02, 2.76)
Overall IQ	p-value	0.724	0.067	0.040*
	Odds Ratio (95% CI)	1.05 (0.64, 1.73)	5.77 (3.41, 9.76)	6.08 (3.59, 10.28)
Nerve Conspicuity	p-value	0.805	<.001*	<.001*
	Odds Ratio (95% CI)	3.62 (2.17, 6.02)	2.24 (1.28, 3.9)	8.09 (4.65, 14.08)
Fat Suppression	p-value	<.001*	<.001*	<.001*
	Odds Ratio (95% CI) p-value	1 (0.54, 1.86)	1.04 (0.56, 1.94)	1.04 (0.56, 1.94)
Pulsatility Artifact	p-value	0.999	0.999	0.999
	Odds Ratio (95% CI)	1 (0.61, 1.64)	1.18 (0.72, 1.95)	1.18 (0.72, 1.94)
Ringing Artifact	p-value	0.999	0.999	0.999

Interpretation:

OR > 1 for A vs B implies superior scores in A compared to B

OR = 1 for A vs B implies no difference in scores between A and B

OR < 1 for A vs B implies superior scores in B compared to A

Table 4 summarizes the inter-rater agreement between the three readers. There was substantial agreement (≥0.7) in 10 out of 15 of the assessed image metrics. The range of the agreement coefficient was between 0.59 and 0.89.

Table 4.

Inter-rater agreement of qualitative measures assessed using ordinal weighted Gwet’s agreement coefficient (standard error in parentheses), where a coefficient greater than or equal to 0.7 was considered substantial agreement.

	Standard	PCA	WW+PCA
Overall Image Quality	0.77 ( 0.04)	0.74 ( 0.04)	0.69 ( 0.05)
Nerve Conspicuity	0.62 ( 0.06)	0.59 ( 0.06)	0.76 ( 0.04)
Fat Suppression	0.73 ( 0.06)	0.62 ( 0.06)	0.77 ( 0.05)
Pulsatility Artifact	0.89 ( 0.03)	0.89 ( 0.03)	0.63 ( 0.08)
Ringing Artifact	0.76 ( 0.04)	0.76 ( 0.04)	0.75 ( 0.04)

Fig. 5 shows water images and FF maps in a patient with polyneuropathy of unknown etiology and denervated muscle. PCA improved fat suppression and conspicuously decreased FF in the musculature, especially in regions that were more central. These regions corresponded to reduced noise in the PCA+WW images. The nerve conspicuity was unaltered by PCA. The image contrasts for denervated vs. normal muscle were similar, but the subcutaneous fat signal was significantly suppressed.

Figure 5.

Images without PCA (Std, top row) and with PCA-denoising (bottom row) from a 65-year-old man with denervation edema pattern of multiple forearm flexor and extensor muscles, showing (from left to right) water, fat, fat fraction (FF) and water-weighted (WW) images. PCA provided similar levels of proximal forearm ulnar nerve conspicuity (yellow arrows) and abnormal muscle edema (red dashed arrows). PCA improved homogeneity of subcutaneous fat suppression (green arrows) relative to the adjacent muscle. PCA also reduced the FF levels and noisy appearance of FF in the normal muscle (blue dashed arrows). In the WW images, the regions where FF were reduced in normal muscle corresponded to improved muscle signal homogeneity as well.

Fig. 6 compares standard water images, PCA water images, and WW images from two different subjects and anatomies. The WW images show significantly improved subcutaneous fat signal in both subjects (Fig. 6d, f), as well as suppression of perineural fat (Fig. 6d, f, insets), which improved the conspicuity of the sciatic (Fig. 6d) and radial nerves (Fig. 6f). The contrast of the denervated muscle vs. normal muscle was altered neither by PCA nor WW.

Figure 6.

Comparisons in (a-c) pelvic MR neurography in a 41-year-old man with right-sided sciatica but normal appearance of the sciatic nerve. The right sciatic nerve was not conspicuous with (a) standard and (b) PCA (yellow inset and yellow arrows), but was conspicuous on the (c) water-weighted (WW) PCA image. WW markedly improved suppression of subcutaneous fat signal (white arrow). Aliasing and subcutaneous fat (green arrow) were less conspicuous with PCA and WW. Similar comparison of (d-f) elbow MR neurography in a 65-year-old man with mild denervation edema pattern of multiple forearm flexor and extensor muscles (red arrows). The branches of the radial nerve were more conspicuous on (f) WW than on (d) standard and (e) PCA only (yellow inset and yellow arrows).

4. Discussion

WW qualitatively and quantitatively improved fat suppression and qualitatively improved nerve conspicuity in T₂-weighted, two-point Dixon FSE MRN. PCA-denoising demonstrated quantitative reduction in FF error in simulations, with significantly improved fat suppression qualitatively. PCA-denoising did not alter image signal significantly, suggesting no strong bias was introduced to any tissue contrast; the improvement in qualitative fat suppression might in part be due to the denoised muscle signal. Together, WW and PCA-denoising significantly improved the overall image quality of MRN.

The in vivo results from PCA-denoising were consistent with simulations between 3–10% noise level, i.e. resulting in about −1% FF in normal muscle. While not statistically significant, PCA-denoising increased FF in subcutaneous fat and decreased FF in other tissues, consistent with the expected physiology; variation from the wide sampling parameters and range of anatomical regions may have contributed to a non-statistically significant result. There was also variation in effects from motion, especially respiratory-induced motion about the brachial plexus, despite the use of respiratory-gating[27].

The denoising and WW processing techniques may also be applicable to other Dixon-based acquisitions. For example, Dixon-based fat suppression has been applied in reversed fast imaging steady-state free precession (PSIF) in MRN [28] and dual-echo steady-state (DESS) in musculoskeletal imaging[29]. Denoising may also be applicable to quantitative T2-mapping sequences that use the Dixon method for separating water and fat components [22].

4.1. Limitations

In this study, the choice of two-point Dixon imaging was based on protocols optimized for high in-plane spatial resolution and scan times commonly used in clinical practice. This necessarily limited the potential size of the PCA matrix and capability for denoising that generally increases with the matrix size. For example, matrix size can be increased by 50% with a three-point Dixon technique and doubled with four-point Dixon, albeit at the cost of scan time. Furthermore, more echoes could be useful for solving for T₂* simultaneously[12] and could provide a more accurate B₀-map for reducing FF calculation bias[18]. Also, we did not utilize the Marchenko-Pastur theory previously proposed for determining the optimum threshold[16]; instead, a uniform threshold was chosen to provide spatial uniformity in the extent of denoising. The lowest possible threshold of P=1was found to be optimal, which suggests a high level of informational redundancy in coil-sensitivity and is likely a reasonable assumption for small kernel sizes. In this work, we did not evaluate WW without denoising, which did not allow evaluation of interaction between PCA-denoising and WW. We speculate that WW would also provide improved nerve conspicuity and fat suppression in non-denoised images, given it works complementarily to PCA-denoising. Given the significant contrast improvement from WW processing observed in this study, we anticipate that WW effects would dominate those of PCA-denoising. Another limitation was that this study did not include normal subject controls; muscle was deemed to be ‘normal’ by qualitative assessment by the radiologists rather than by electrophysiological or physical exam testing.

In this work, we focused on improvements in visualization of nerve conspicuity and potential improvements in FF accuracy. However, we did not evaluate potential improvements in visualizing edema patterns in denervated muscle due to the smaller subset of denervated muscle. In addition, we did not evaluate the potential for using FF to quantify fat infiltration in denervated muscle. In this cohort, only 11 subjects were diagnosed with abnormal muscle and none of these muscles demonstrated fat infiltration qualitatively, as seen in states of chronic denervation. We speculate that PCA-denoising and WW together would provide improved depiction of denervated muscle, especially in discerning patterns of fat infiltration due to the increased fat suppression. Given the simplicity of WW as compared to PCA, WW would be more feasibly implemented in real-time. Future work could include comparing Dixon FSE-based FF measurements with gradient-echo-based methods, as well as correlating Dixon FSE-based FF measurements with electrophysiology results in a larger cohort with varying stages and severities of muscle denervation.

5. Conclusions

This study demonstrated that PCA-denoising and water-weighting improved nerve conspicuity and fat suppression on T₂-weighted, Dixon FSE MRN sequences. PCA-denoising also improved quantitative accuracy of FF maps, which might be useful in assessing chronic muscle denervation.