Denoising of diffusion MRI improves peripheral nerve conspicuity and reproducibility

Abstract

Background Quantitative diffusion MRI is a promising technique for evaluating peripheral nerve integrity but low signal‐to‐noise ratio (SNR) can impede measurement accuracy. Purpose To evaluate principal component analysis (PCA) and generalized spherical deconvolution (genSD) denoising techniques to improve within‐subject reproducibility and peripheral nerve conspicuity. Study Type Prospective. Subjects Seven healthy volunteers and three peripheral neuropathy patients. Field Strength/Sequence 3T/multiband single‐shot echo planar diffusion sequence using multishell 55‐direction scheme. Assessment Images were processed using four methods: "original" (no denoising), "average" (10 repetitions), "PCA‐only," and "PCA + genSD." Tibial and common peroneal nerve segmentations and masks were generated from volunteer diffusion data. Quantitative (SNR and contrast‐to‐noise ratio [CNR]) values were calculated. Three radiologists qualitatively evaluated nerve conspicuity for each method. The two denoising methods were also performed in three patients with peripheral neuropathies. Statistical Tests For healthy volunteers, calculations included SNR and CNRFA (computed using FA values). Coefficient of variation (CV%) of CNRFA quantified within‐subject reproducibility. Groups were compared with two‐sample t‐tests (significance P < 0.05; two‐tailed, Bonferroni‐corrected). Odds ratios (ORs) quantified the relative rates of each of three radiologists confidently identifying a nerve, per slice, for the four methods. Results "PCA + genSD" yielded the highest SNR (meanoverall = 14.83 ± 1.99) and tibial and common peroneal nerve CNRFA (meantibial = 3.45, meanperoneal = 2.34) compared to "original" (P SNR < 0.001; P CNR = 0.011) and "PCA‐only" (P SNR < 0.001, P CNR < 0.001). "PCA + genSD" had higher within‐subject reproducibility (low CV%) for tibial (6.04 ± 1.98) and common peroneal nerves (8.27 ± 2.75) compared to "original" and "PCA‐only." The mean FA was higher for "original" than "average" (P < 0.001), but did not differ significantly between "average" and "PCA + genSD" (P = 0.14). "PCA + genSD" had higher tibial and common peroneal nerve conspicuity than "PCA‐only" (ORtibial = 2.50, P < 0.001; ORperoneal = 1.86, P < 0.001) and "original" (ORtibial = 2.73, P < 0.001; ORperoneal = 2.43, P < 0.001). Data Conclusion PCA + genSD denoising method improved SNR, CNRFA, and within‐subject reproducibility (CV%) without biasing FA and nerve conspicuity. This technique holds promise for facilitating more reliable, unbiased diffusion measurements of peripheral nerves. Level of Evidence: 2 Technical Efficacy Stage: 1 J. Magn. Reson. Imaging 2020;51:1128–1137.

Denoising and Multiple Tissue Compartment Visualization of Multi‐b‐Valued Breast Diffusion MRI

“…The denoising scheme was shown to be highly effective at suppressing erroneously high FA without compromising spatial resolution, similar to that observed in multi-shell diffusion denoising in the peripheral nerves. 27 In this breast DWI work, the FA reductions were particularly conspicuous in the fatty tissue regions with low signal due to fat-suppression and in tissue boundary regions (fat-air, fat-NFT interfaces, for instance). Erroneously high FA bias is a common DTI postprocessing issue, 35 especially in tissue boundary regions where misregistration due to either motion or eddy currents can severely bias FA calculation.…”

“…With the 12-directions data, there were further increases in the spread of the distributions, along with a pronounced bias in the vascular FA to about 0.4-0.6. Figure 3 shows the quantitative differences between denoised and undenoised metrics for each of the compartments, and at all of the tested number of diffusion directions (12)(13)(14)(15)(16)(17)(18)(19)(20)(21)(22)(23)(24)(25)(26)(27)(28)(29)(30). Denoising on the "fully-sampled" resulted in varying levels of bias in the diffusivity metrics.…”

“…It has shown promising results in removing outliers in diffusion kurtosis imaging in the brain, as well as in improving nerve conspicuity in peripheral nerve imaging. 27 The goal of this work was to adapt the model-based diffusion denoising technique, 26 for breast DWI, where the diffusivity, anisotropy and type of signal compartments are inherently different compared to the brain. Multi-shell diffusion (with more than six directions in a given shell) breast MRI acquisition was used to 1) more broadly determine the effects of denoising on both mean diffusivity and fractional anisotropy.…”

“…This approach utilizes multiple compartments and is applicable primarily to multi‐b‐valued diffusion acquisition. It has shown promising results in removing outliers in diffusion kurtosis imaging in the brain, as well as in improving nerve conspicuity in peripheral nerve imaging 27 …”

Getting Quantitative Diffusion‐Weighted MR Neurography and Tractography Ready for Clinical Practice

“…The article "Denoising of Diffusion MRI Improves Peripheral Nerve Conspicuity and Within‐Subject Reproducibility" describes the use of a combined principal component analysis (PCA) and generalized spherical devolution (genSD) denoising technique to improve signal and contrast, nerve visibility, and quantitative accuracy of diffusion tensor imaging data that were acquired with an acceleration scheme combining parallel imaging and simultaneous multislice techniques . A 6‐minute prototype spin echo echo‐planar pulse sequence with a multishell 55‐direction scheme was used to obtain anisotropic diffusion data.…”