Denoising of diffusion MRI improves peripheral nerve conspicuity and reproducibility.

Abstract

Quantitative diffusion MRI is a promising technique for evaluating peripheral nerve integrity but low signal-to-noise ratio (SNR) can impede measurement accuracy. To evaluate principal component analysis (PCA) and generalized spherical deconvolution (genSD) denoising techniques to improve within-subject reproducibility and peripheral nerve conspicuity. Prospective. Seven healthy volunteers and three peripheral neuropathy patients. 3T/multiband single-shot echo planar diffusion sequence using multishell 55-direction scheme. 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. 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. "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). 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. 2 Technical Efficacy Stage: 1 J. Magn. Reson. Imaging 2020;51:1128-1137.