Denoising of diffusion MRI in the cervical spinal cord – effects of denoising strategy and acquisition on intra-cord contrast, signal modeling, and feature conspicuity

Abstract

Quantitative diffusion MRI (dMRI) is a promising technique for evaluating the spinal cord in health and disease. However, low signal-to-noise ratio (SNR) can impede interpretation and quantification of these images. The purpose of this study is to evaluate several dMRI denoising approaches on their ability to improve the quality, reliability, and accuracy of quantitative diffusion MRI of the spinal cord. We evaluate three denoising approaches (Non-Local Means, Marchenko-Pastur PCA, and a newly proposed Patch2Self algorithm) and conduct five experiments to validate the denoising performance on clinical-quality and commonly-acquired dMRI acquisitions: 1) a phantom experiment to assess denoising error and bias; 2) a multi-vendor, multi-acquisition open experiment for both qualitative and quantitative evaluation of noise residuals; 3) a bootstrapping experiment to estimate uncertainty of parametric maps; 4) an assessment of spinal cord lesion conspicuity in a multiple sclerosis group; and 5) an evaluation of denoising for advanced parametric multi-compartment modeling. We find that all methods improve signal-to-noise ratio and conspicuity of MS lesions in individual diffusion weighted images (DWIs), but MPPCA and Patch2Self excel at improving the quality and intra-cord contrast of diffusion weighted images – removing signal fluctuations due to thermal noise while improving precision of estimation of diffusion parameters even with very few DWIs (i.e., 16–32) typical of clinical acquisitions. These denoising approaches hold promise for facilitating reliable diffusion observations and measurements in the spinal cord to investigate biological and pathological processes.

1. Introduction

Quantitative diffusion MRI (dMRI) is a promising tool to study the tissue microstructure of the spinal cord in health and disease. To date, the most commonly utilized dMRI technique is diffusion tensor imaging (DTI). DTI provides quantitative indices including fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD), and radial diffusivity (RD), which have been shown to be sensitive to tissue properties such as axon density, axonal injury, and degree of myelination (Beaulieu, 2002). In addition to DTI, a growing number of more advanced microstructure models, or multi-compartment models, of diffusion have been applied to the spinal cord (Duval et al., 2015; Grussu et al., 2019; Grussu et al., 2017; Grussu et al., 2016; Moccia et al., 2019; Saliani et al., 2017; Schilling et al., 2019), improving pathological specificity to tissue damage. For example, in multiple sclerosis (MS), a condition featured by a complex interplay among inflammation, demyelination and axonal loss, DTI indices have been shown to correlate well with clinical measures of disability (Moccia et al., 2019). Further, advanced measures of neurite density, compartment diffusivities, and axonal disorganization have shown sensitivity in identifying abnormal changes in MS spinal cord lesions, as well as in normal-appearing white matter (WM) (By et al., 2017, 2018; Cohen-Adad, 2018; Grussu et al., 2016; Grussu et al., 2015; Schilling et al., 2019), and suggest an expanding use for diffusion-derived metrics in MS clinical practice and trials.

Yet, spinal cord dMRI on clinical scanners is challenging. The spinal cord is a thin, complex structure requiring relatively high spatial resolution for adequate anatomic depiction. Spinal cord imaging is also complicated by artifacts from motion and local susceptibility, resulting in lower SNR and precision in quantitative analysis. At the same time, in vivo imaging has scan time constraints, which limits the number of diffusion weighted images (DWIs) that can be acquired and requires fundamental tradeoffs with image resolution and image quality. Thus, improving image quality is imperative to facilitate both research investigations and clinical application of spinal cord dMRI.

Towards this end, several denoising approaches have been applied to dMRI data, typically in the brain, in order to improve signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR), as well as reproducibility and precision of diffusion-derived features (Buades et al., 2005; Coupe et al., 2008; Elad and Aharon, 2006; Knoll et al., 2011; Manjon et al., 2013; Rudin et al., 1992). For example, some may assume that the signal is intrinsically low-rank and can be denoised through dimensionality reduction, as in the Marchenko-Pastur PCA, or MPPCA, denoising algorithm (Veraart et al., 2016b). Others assume that the signal patterns may be replicated throughout the tissue and can be denoised through averaging of similar signals as performed in non-local means (NLM) (Coupé et al., 2012; Coupe et al., 2008). Finally, rather than assumptions on the signal, a recently proposed approach, called Patch2Self (Schilling et al., 2021), assumes that the noise is random and uncorrelated across DWIs in order to separate signal from noise.

Despite proven enhancements of signal and contrast in the brain, the effectiveness of these methods in the spinal cord has not been fully investigated. Of particular interest, the MPPCA algorithm has indeed been shown to significantly enhance SNR and improve parameters maps in the spinal cord (Grussu et al., 2020). However, this technique exploits data redundancy (i.e. multiple DWIs or multi-contrast datasets) which may not exist with time-limited datasets and low numbers of DWIs. In fact, clinical investigations using diffusion datasets are often limited to less than 10–20 DWIs (Rutman et al., 2018; Zaninovich et al., 2019), and state-of-the-art spinal cord research protocols to 30–32 directions (Cohen-Adad et al., 2021a; Cohen-Adad et al., 2021b), with few studies considering multiple diffusion weightings. Thus, validating and investigating this algorithm in clinical-quality datasets is needed for a vast majority of applications of diffusion MRI in the spinal cord. Moreover, to the best of our knowledge, other denoising algorithms, i.e., Patch2Self or NLM, have not been tested in this structure.

Thus, the aim of this study is to evaluate the efficacy of denoising algorithms on clinical-quality dMRI data in the cervical spinal cord in healthy volunteers and in patients with MS. We do so by evaluating three algorithms, NLM, MPPCA, and Patch2Self across a range of image acquisition settings, image qualities, and scan vendors. We first utilize a numerical phantom dataset to evaluate the bias and errors introduced in the denoising process. Next, we assess denoising on empirically acquired datasets from open-source databases to evaluate visual quality and noise statistics. Third, we utilize a multi-repeat dataset to evaluate precision in the estimation of diffusion parameters. Fourth, we apply the algorithm to data from an MS cohort and assess lesion conspicuity before and after denoising. Finally, we apply the algorithms to a multi-shell high angular resolution dataset to evaluate advanced biophysical modeling. We hypothesize that denoising will lead to improved intra-cord contrast, model-fitting, and SNR, and improve lesion conspicuity in spinal cord diffusion images. Further, we hypothesize that there will be tradeoffs in algorithm choice, depending on image acquisition and processing choices.

2. Materials and methods

This study included 5 datasets that were used to evaluate different effects of the denoising process (Figure 1). In all cases NLM, MPPCA, and P2S denoising strategies were employed using the DIPY toolbox (Garyfallidis et al., 2014). All denoising was applied on raw, unprocessed images in native space.

Fig. 1.

Experimental datasets and designs. This study includes 5 datasets, including (1) a digital phantom to investigate error and bias in denoising, (2) empirical datasets from multiple vendors and acquisitions to assess noise statistics, (3) a multi-repeat dataset to study precision and bias with bootstrapping, (4) MS datasets to assess lesion conspicuity, and (5) a multi-shell high angular resolution dataset for multi-compartment biophysical modeling.

NLM is an algorithm that is based on the redundancy of periodic or textured images. In contrast to a local-means filtering which takes the mean value of a patch of voxels surrounding the target voxel we intend to denoise, the ‘non-local’ algorithm adapts the mean based on the similarity in spatial and frequency information in blocks that may not be near (i.e. may not be local to) the patch under investigation. In this way, signal is filtered locally, but edges are preserved. In all studies, NLM is implemented using DIPYs non_local_means algorithm using default parameters, patch_radius = 2 (5 × 5 × 5 voxels), block_radius = 5 (11 × 11 × 11 voxels).

MPPCA exploits intrinsic redundancy in dMRI data to remove components of the data that are classified as noise (Veraart et al., 2016b). These noise-only principal components can be identified, and removed, from the signal through automatic detection based on properties of the eigenspectrum of random covariance matrices. Because noise may be spatially varying, noise level estimation and denoising again operates in patches around each voxel. In all cases, patch size was chosen to ensure that the number of voxels in a patch is greater than or equal to the number of diffusion weighted images in the dataset based on recommendations from Cordero-Grande et al., 2019, which was typically patch_radius = 1 (3 × 3 × 3) or patch_radius = 2 (5 × 5 × 5).

Unlike the previous algorithms, Patch2Self does not make assumptions about the dMRI signal, rather it assumes that noise across different 3D volumes of the DWI signal is random and uncorrelated. Patch2Self begins with a self-supervised training by extracting 3D patches of signal from all DWI volumes, except for one held out target volume to denoise. A regressor is trained to predict the center voxel of each patch in the target volume. Once the training is done, the same trained model is now used to predict the held-out volume using the trained function. The prediction obtained is the denoised volume and this procedure is applied iteratively on all 3D volumes of the 4D data. This procedure can be seen as a q-space in-painting method where each gradient direction is represented as a linear combination of the remaining gradient directions. In all studies, Patch2Self was implemented in Dipy, with patch_radius = 2 (5 × 5 × 5) for all experiments, ordinary least squares for regression, shift_intensity = True (see Discussion), and a mask of the spinal cord so that background is not used in training. We used the DIPY toolbox (Garyfallidis et al., 2014) to implement Patch2Self, but note that it has recently been made available in the Spinal Cord Toolbox (De Leener et al., 2017) as well, which was utilized to perform additional processing steps described below.

In addition to these settings, anisotropic patches of all methods were tested due to the anisotropic nature of the acquired datasets, and no significant differences were observed in the results of any experiments described below.

2.1. Digital phantom experiment

The first dataset was a digital phantom of the cervical spinal cord. This digital phantom was created as the population-average diffusion tensors derived from 260 healthy subjects from 42 centers as described in (Cohen-Adad et al., 2021a) and publicly available via (https://doi.org/10.6084/m9.figshare.14052269). Local ethics committees of the participating institutions approved the study protocol and signed informed consent was obtained from all participants. All data was acquired using the ‘spinal cord generic protocol’ (Cohen-Adad et al., 2021b) designed to improve inter-site reproducibility of spinal cord imaging. Data was transformed to PAM50 template space (De Leener et al., 2018) using the Spinal Cord Toolbox (De Leener et al., 2017) and simply averaged across subjects at the C3/C4 vertebral levels. Simple averaging is adequate for the purposes of this study, where we only required a ‘ground truth’ tensor (and subsequent ground truth signal) which covered an expected range of biophysical values which will be used to compare against denoising strategies. From this phantom, datasets of 16, 32, and 96 uniformly sampled diffusion directions were simulated, and corrupted with noise in quadrature at SNR = 10, 20, and 50. Denoising was performed for all combinations of number of directions and SNR, and bias and error in reconstructed signal were assessed by calculating the difference in denoised signal from the ground truth, and mean squared error (analysis within the spinal cord only, not including cerebrospinal fluid nor background).

2.2. Empirical data from spinal cord generic protocol

For this experiment, we aimed to assess denoising strategies on empirically acquired datasets, with a range of acquisition strategies and from multiple vendors. To do this, we selected example datasets from the previously described spinal cord generic protocol study (Cohen-Adad et al., 2021a) publicly available via (https://doi.org/10.6084/m9.figshare.14052269) as well two datasets acquired at Vanderbilt University Institute of Imaging Science (VUIIS). These datasets gave us variation in vendors, variation in SNR, in direction sets, in b-values, in image resolution, and in possible preprocessing done on the scanner including interpolation and averaging. Most datasets chosen are representative of data that may typically be acquired clinically – a single b-value with relatively low number of gradient directions.

The open sourced datasets included one acquired on a Philips Achieva 3T scanner with 32 DWIs at a b-value of 800s/mm² at 0.875 × 0.875 × 5 mm³ resolution (designated in Figures as d32b800-Achieva for simplicity), one acquired on a Siemens Prisma 3T scanner with 30 DWIs at a b-value of 800s/mm² at 0.9 × 0.9 × 5 mm³ resolution (d30b800-Prisma), one acquired on a Siemens Magnetom 3T scanner with 30 DWIs at a b-value of 800s/mm² at 0.9 × 0.9 × 5 mm³ resolution (d30b800-Magnetom), and one acquired on a GE Discovery 3T scanner with 30 DWIs at a b-value of 800s/mm² at 0.9 × 0.9 × 5 mm³ resolution but resampled to 0.33 × 0.33 × 5 mm³ resolution on the scanner (d30b800-GE-Interp). Imaging at VUIIS was performed using a 3T Philips Elition MR scanner and included three different datasets. First was a single-shell acquisition with 15 diffusion-weighted directions at b = 750s/mm² (TR/TE = 5 beats (~5000 ms)/65 ms, resolution = 1.1 × 1.1mm², slice thickness = 5 mm, FOV = 80 × 57.5 × 70 mm, SENSE (RL) = 1.8, partial Fourier = 0.693). Importantly, this dataset included 3 averages (NEX = 3) performed on the scanner (d15b750-NEX3). Second, this same dataset was used with interpolation to 0.375 × 0.375 × 5mm³ resolution, a ‘feature’ that is commonly done in many clinical protocols (d15b750-NEX3-interp). Finally, a multi-shell acquisition was performed with 48 DWIs at b-values of both 700 and 2800s/mm², an acquisition that may be typical for multicompartment modeling. All datasets were acquired with axial orientation. This study at VUIIS involving human participants was reviewed and approved by the Vanderbilt University Institutional Review Board, and the participants provided their signed informed consent prior to examination.

For this experiment, we first assess denoising through qualitative visualization of individual acquired and denoised images. Second, we calculate and show estimated error, or residual maps, of each denoising approach in order to visually verify signal-preservation (confirmed by lack of anatomy in error maps) and properties of noise removed (by comparing variance of residuals to standard normal distributions) (Veraart et al., 2016b). Finally, we visualize DTI-derived maps of FA, MD, AD, and RD on all datasets before and after denoising.

2.3. Bootstrap experiment

We used a bootstrapping experiment to estimate uncertainty in DTI-derived parameters after denoising. To do this, a healthy volunteer underwent imaging on a 3T Philips Elition MR scanner at VUIIS. Data was acquired with five repetitions using the Spinal Cord Generic protocol which included acquisition of 32 DWIs at a b-value of 800s/mm² at 0.875 × 0.875 × 5 mm³ resolution. 100 bootstrap realizations were selected by randomly selecting individual measurements from the five repetitions of the same measurement. DTI parameters of interest from each of the many bootstrap realizations, with and without denoising, were computed, and variability of these parameters between the different realizations were evaluated to study the effect of denoising on the precision of diffusion parameters. In all cases, denoising was performed on the raw, bootstrapped data, followed by motion correction and DTI fitting using the Spinal Cord Toolbox (De Leener et al., 2017).

2.4. Multiple sclerosis datasets

The next experiment aimed to assess the effects of denoising on image quality, contrast, and lesion conspicuity in MS lesions. The cohort for this experiment consisted of N = 16 people with relapsing-remitting MS (pwRRMS) (20–42 years old, 9F/7M, Expanded Disability Status Scale scores 0–1.5) with one session per patient. Imaging was performed using the same 3T Philips Elition MR scanner as previously described, using a single-shell acquisition with 15 diffusion-weighted directions at b = 750s/mm² (TR/TE = 5 beats (~5000 ms)/65 ms, resolution = 1.1 × 1.1mm², slice thickness = 5 mm, FOV = 80 × 57.5 × 70 mm, SENSE (RL) = 1.8, partial Fourier = 0.693, NEX = 3). Denoising quality was assessed on both scans reconstructed at native resolution, and also up-sampled during reconstruction by the scanner console (0.375 × 0.375 × 5mm³ resolution). A high-resolution (0.65 × 0.65 × 5 mm³) multi-slice, multi-echo gradient echo (mFFE) anatomical image (Held et al., 2003) was acquired (TR/TE/ΔTE = 700/8.0/9.2 ms, α = 28 degrees, number of slices = 14, 6:12 minutes) for co-registration and to serve as a structural reference image.

In this experiment, lesion conspicuity was assessed before and after denoising on both individual DWIs as well as the mean DWI, both of which may assist in highlighting either increased diffusion restrictions in lesions (visible as hyperintense diffusion signal) or vasogenic edema in acute demyelinating lesions with increased diffusion coefficients (hypointense diffusion signal).

2.5. Biophysical modeling

Finally, in contrast to many of the clinical-quality, single-shell experiments above, we aimed to investigate denoising approaches on multi-shell diffusion datasets and subsequent multi-compartment modeling. Towards this end, a 3-shell dataset was acquired, with b-values of 700, 1400, and 2800s/mm² with 40 directions each. Images with and without denoising were fit to the Neurite Orientation and Dispersion Density Imaging (NODDI) model (Zhang et al., 2012), which results in indices of orientation dispersion index (ODI), intracellular volume fraction (ICVF), and isotropic volume fraction (ISO).

3. Results

3.1. Phantom experiments

The ground truth phantom data with simulated acquisitions ranging from 16 to 96 DWIs was corrupted with various SNR levels and denoised using three denoising approaches. The bias and error are shown for a selected DWI image chosen with gradient directions nearly parallel to the spinal cord (i.e., expected highest attenuation and lowest signal) in Figure 2. Here, we see that all denoising methods remove noise, reducing error between the noisy and noiseless data, even at very low SNR. However, NLM overly smooths the image, reducing white/gray matter contrast, and results in higher mean-squared error than MPPCA and P2S. MPPCA and P2S excel at removing noise at low SNR and with remarkably few diffusion directions, with little to no bias in results at SNR = 20 and above. These results generalize to all DWIs, with results shown nearly perpendicular to the cord (i.e., lower attenuation, higher signal) in Supplementary Figure 1.

Fig. 2.

Denoising reduces noise levels, even with low SNR and with few diffusion directions. For SNR = 10 and SNR = 20, DWIs are shown after denoising for NLM, MPPCA, and P2S, along with error from ground truth phantom, and a plot of denoised versus true signal with mean-squared error (MSE) given in plots. These results are shown for a DWI most aligned with cord (greatest signal attenuation), and an example DWI most perpendicular to cord is shown in Supplementary Figure 1.

3.2. Empirical data

Denoising performance on empirical datasets is shown in Figure 3 for datasets from different vendors and with different acquisition conditions. A few observations are apparent. First, this implementation of NLM overly smoothes the structure in the cord. Second, MPPCA and P2S excel at improving intra-cord contrast, particularly at low SNR. Third, P2S is additionally able to recover signal when data is interpolated to high resolution on the scanner. Finally, MPPCA and P2S excel even with low numbers of DWIs.

Fig. 3.

Denoising improves image quality from different vendors and with different acquisition parameters. Results are shown for seven datasets, from 3 vendors, with/without interpolation and signal averaging on the scanner, for diffusion directions against (left) and along (right) the cord.

Residual maps from each denoising technique are shown in Figure 4, along with the distribution of normalized residuals. As described in (Veraart et al., 2016b), the zero-centered residuals and lack of anatomical structure indicate good preservation of signal. Here, MPPCA and P2S perform well, with no anatomical structure visible within the cord and zero-centered residuals when averaged over all DWIs. Plots of the distribution (or here, logarithm of the distribution) of residuals allow comparisons to a normal distribution (black line), where normalized residuals are ideally described by this distribution when all noise is accumulated in the residuals (i.e., denoising removed all noise) (Veraart et al., 2016b). Here, all algorithms are well approximated by a normal distribution, where NLM occasionally has a higher variance than explained by the noise and is removing true signal.

Fig. 4.

Noise properties estimated from denoising algorithms suggest MPPCA and P2S preserve signal while removing noise. Sigma-normalized residual maps between denoised DWIs and the original data for a single image (left) and averaged over all DWIs (right) are shown, where the presence of anatomical structure indicates interference of algorithms with the signal. Plots at the right show the distribution of normalized residuals with a standard normal distribution (solid black line) as reference. All algorithms have similar magnitude of denoising (standard deviation ~1) although NLM has negative shift of signal.

Finally, parametric maps of derived FA are shown in Figure 5. MP-PCA and P2S show increased contrast within the cord but have very small overall effects on these derived parameters, particularly for acquisitions with large sampling schemes. Effects of denoising on derived diffusivities are much less visible (Supplementary Figure 2).

Fig. 5.

FA maps from acquired and denoised data show small, but visible, improvements in intra-cord quantitative contrast.

3.3. Bootstrap experiment

Results of the bootstrap experiment are shown in Figure 6 for derived FA and MD. Here, typical voxel-wise standard variation of this dataset without denoising is on the order of 0.1 for FA and 0.8E-4mm²/s for MD. All investigated denoising algorithms reduce variability and increase precision of estimated parameters.

Fig. 6.

All denoising techniques improve precision in estimated FA (top, unitless) and MD (bottom, units mm²/s). Bootstrapping experiments enable quantitative estimates of variability shown as maps and as histograms (including voxels in the cord only).

3.4. Multiple sclerosis datasets

The effects of denoising on datasets with MS lesions are shown in Figures 7 and 8. Here, Figure 7 shows the effects of denoising on the mean diffusion weighted image while Figure 8 shows the effects of denoising on a single DWI, chosen most parallel to the cord which is expected to best highlight diffusion abnormalities due to lesions. In both mean and individual DWIs, lesions are clearly distinguished through an increased signal visible in both raw and denoised images. However, denoising has very little effect on the mean DWI, as there is already adequate SNR when averaging many DWIs. Optimistically, denoising does increase lesion conspicuity (and white/gray matter contrast) on individual DWIs, with hyperintense signal clearly aligned with lesions visible in a structural image, particularly with MPPCA and P2S denoising. Similar results are obtained with images interpolated on the scanner (Supplementary Figure 3).

Fig. 7.

Effects of denoising on lesion conspicuity in mean DWIs. Lesions are visible as hyperintense regions on the mean DWI, however, denoising does not improve visibility on these images. Anatomical mFFE images are shown in the first column for reference.

Fig. 8.

Effects of denoising on lesion conspicuity in individual DWIs. Lesions are visible as hyperintense regions in DWIs, and denoising improves conspicuity of lesions as well as white/gray matter contrast. Anatomical mFFE images are shown in the first column for reference.

3.5. Biophysical modeling

Results of denoising on multi-shell high angular resolution diffusion datasets, and subsequent multi-compartment modeling, are shown in Supplementary Figure 4. While individual DWIs show improved contrast, derived microstructure features are not visibly changed, in line with observations for the DTI model.

4. Discussion

This work demonstrates advantages of denoising techniques for clinically-feasible in vivo diffusion imaging of the spinal cord. The key findings of this work are that (1) all denoising techniques reduce thermal noise and increase intra-cord contrast; (2) MPPCA and P2S do this while also not removing anatomical signal or compromising accuracy while (3) also working well on datasets of varying vendors, acquisitions, and image quality including very few diffusion directions; (4) precision of DTI-estimated parameters is marginally improved with all denoising; and (5) MPPCA and P2S increase conspicuity of image features and MS lesions in diffusion weighted images. Finally, (6) despite these advantages to individual diffusion weighted volumes, derived parameters from DTI and multi-compartment models, and lesion conspicuity on direction-averaged DWIs, are not visibly improved.

Advanced MRI of the cervical spinal cord is challenging due to its small size and mobility, its proximity to tissue interfaces, and its susceptibility to physiological noise from respiratory, cardiac, and pulsatile CSF flow sources. Optimization of acquisition and processing techniques is therefore necessary, as are strategies to increase image quality in post-processing (Rutman et al., 2018). The reliability and sensitivity to pathology of DTI as a quantitative MRI technique make it valuable for various neuroimaging applications.

Spinal cord involvement is a central feature of all MS subtypes and is partly responsible for the accumulation of clinical disability. As a research tool, new and exciting diffusion MRI endeavors aim to potentially provide quantitative measures to assess demyelination, edema, fiber integrity and axonal loss in normal-appearing and lesional tissue, and are applied in the spinal cord with increasing frequency. DTI metrics have also shown correlations with disability, and with upper and lower limb motor function (Moccia et al., 2019). Reproducible scans with higher SNR will further facilitate such investigations. While MRI is the modality of choice for the management and study of MS, DWI currently has a limited role in the diagnostic work-up (Wattjes et al., 2021). Abnormal signal hyperintensity in DWIs, such as that observed in the MS lesions in our study, may be clinically useful in increasing confidence in lesion identification and potentially help differentiate active from nonactive lesions in conjunction with anatomical images and apparent diffusion coefficient (ADC) maps. Although DWIs for a single direction are not typically used clinically, improving feature conspicuity in these images with denoising may reveal abnormalities that are obscured in a mean DWI derived from the average of many diffusion directions. The investigation of spinal cord features in health and development, as well as research into acute and chronic injury, degenerative conditions such as neuromyelitis optica spectrum disorder and amyotrophic lateral sclerosis also stand to benefit from improvements in image quality.

Despite improvements in contrast of individual diffusion weighted images, the contrast and conspicuity of DTI-derived metrics as well as direction-averaged DWIs did not significantly increase after denoising. This suggests that, with this acquisition protocol, there is already adequate SNR with the DWIs utilized (as low as 15) to reliably compute the tensor or direction-averaged signal. Reassuringly, MPPCA and P2S did not significantly alter the quantified indices, result in lower contrast, nor blur the biological inter-subject differences.

Overall, the findings in this study support the use of MPPCA and P2S denoising for spinal cord diffusion imaging. While MPPCA algorithms have become popular in brain imaging (Veraart et al., 2016b), it has not been widely used in the cord. One exception is work from Grussu et al. which showed substantial improvements when used with high angular resolution diffusion in combination with multiple contrasts utilizing the same image readout (Grussu et al., 2020). In this study, application of MPPCA showed noise reduction and improved precision in estimated parameters without compromising accuracy in datasets with as few as 15 DWIs. Thus, this denoising technique which excels with oversampling and high-resolution datasets also does well with standard, or generic, spinal cord diffusion protocols.

Patch2Self makes a very weak assumption on the noise (Fadnavis et al., 2020). By assuming that the noise is randomly fluctuating across different gradient directions, it only relies on the assumption of statistical independence. This property enables the application of Patch2Self at any point of the pre-processing pipeline, for example after interpolation or averaging on the scanner. While most other denoisers such as the MPPCA need to be applied to the raw data (Grussu et al., 2020; Veraart et al., 2016a; Veraart et al., 2016b), Patch2Self does not impose any such restriction and is unaffected by pre-volume (3D) interpolations by construction. Since the upsampling or downsampling is performed on a 3D volume and does not correlate the noise across volumes, the statistical independence assumption of Patch2Self is unaffected. Even after the interpolation is applied the noise across volumes is still randomly fluctuating and uncorrelated. This allows Patch2Self to still suppress the residual noise where other denoisers cannot be applied.

Conclusions

Here, we have shown that the application of MPPCA and P2S denoising in clinical-quality spinal cord diffusion data improves intra-cord contrast, signal fitting, SNR, and lesion conspicuity. These denoising approaches are freely available in the DIPY software package and P2S in Spinal Cord Toolbox, and they can be implemented on any in vivo dMRI spinal cord acquisitions. These algorithms and approaches hold promise for facilitating reliable diffusion measures in the spinal cord to investigate biological and pathological processes.

Data and code availability

The datasets supporting the conclusions of this article will be made available by the authors upon request.

Data Availability

I have shared the link to open-sourced data in article