1888
Patch2self denoising of magnitude- and complex-averaged peripheral nerve DTI data: Challenges and opportunities1Imaging Research, Barrow Neurological Institute, Phoenix, AZ, United States, 2Barrow Neurological Institute, Phoenix, AZ, United States, 3Barrow Neurological Insitute, Phoenix, AZ, United States
Synopsis
Keywords: Peripheral Nerves, Nerves, denoising, DTI, FA
Motivation: DTI is effective in characterizing nerve pathology but requires a high-resolution scans, resulting in low SNR data.
Goal(s): To analyze the effectiveness of a self-supervised Patch2Self denoising technique on magnitude-averaged and complex-averaged DTI data of human median and ulnar nerve.
Approach: The magnitude-averaged and complex-averaged DTI data was denoised with the Patch2Self algorithm; and DTI fitting performed was performed to obtained estimates of FA, AD, and RD in nerves.
Results: Patch2Self denoising reduced the variability in the data by 20% at the cost of systematic bias, while complex averaging improved the contrast between muscle/nerve and suppressed fat.
Impact: FA estimates obtained from DTI helps monitor nerve regeneration following catastrophic nerve injuries. Improving the quality of the DTI data may improves the reliability of FA estimates so that more subtle treatment effects can be detected.
Introduction
Peripheral nerve damage following trauma results in a catastrophic loss of sensorimotor function. In severe cases, surgical repair is required to regain function, but outcomes remain suboptimal[1][2]. After surgery, it can take months for electrodiagnostics to indicate whether axons are sprouting across the repair site. This results in delayed clinical decision-making and increases the likelihood of permanent functional deficits. We demonstrated that i) fractional anisotropy (FA) values from ex vivo rat nerves relate to axon densities and behavioral outcomes following trauma/repair [3] and ii) FA values from human nerves report on failed surgeries, successful reoperations, and injury severity[4].Despite this promise, peripheral nerve DTI requires high spatial resolution, resulting in low signal-to-noise ratio (SNR) scans that can affect the reliability of FA estimates and our ability to detect changes related to de/regeneration. Patch2self [5] is a self-supervised learning algorithm, which separates signal structure from noise and has been successfully applied in spinal cord DTI[6]. However, its performance with respect to precision and bias has yet to be tested in clinical peripheral nerve DTI data, where fewer diffusion directions/shells are often available.
Therefore, we evaluated the effectiveness of Patch2self denoising with respect to both precision and bias in human DTI data collected from median and ulnar nerves of the forearm. Furthermore, conventional magnitude-averaging was compared to a complex-averaging scheme[7], which has shown improved performance in low SNR regimes.
Methods
Data were collected on a Philips 3.0-T Ingenia MRI system and 8-ch extremity coil. Heathy subjects (N=3, 1/2 M/F, 20-50y.o.) were scanned in the prone position with the arm extended above the head. A fat-suppressed DTI sequence was collected from the mid-forearm to wrist (Figure 1) using: TE/TR=75/3000 ms, resolution= 1x1x5 mm3, averages=10, slices=20, multiband factor=2, SENSE factor=1.5, 20 diffusion directions (b=0,800 s/mm2), and scan time≈10 minutes. The resulting data were reconstructed using a vendor-provided magnitude-averaging and a custom complex-averaging scheme, the latter of which performed phase corrections for each average prior to averaging[7]. The resulting data were denoised using Patch2self in DIPY (https://dipy.org) (patch radius= (2,2,0)). Finally, DTI tensor estimation was performed before and after denoising to estimate FA, AD, and RD. The median and the ulnar nerves were then manually segmented using 3D slicer (https://www.slicer.org), and i) mean region-of-interest (ROI) tensor estimates and ii) SNR values were calculated.Results and Discussion
The magnitude- and complex-averaged data before and after Patch2Self denoising can be seen in Figure 2. Visually, the data after denoising appeared less noisy for both the magnitude- and complex-averaged data, with an SNR increase of approximately 20% in the Patch2Self denoised data. Furthermore, while the complex-averaged data demonstrated better suppression of fat signal (red arrow) than the magnitude-averaged data, the SNR values of the complex-data were 15% lower than the magnitude-averaged data in muscle (yellow arrow) and nerves (blue arrows). This is consistent with the idea that complex-averaging reduces signal bias related to the Rician nature of magnitude MRI data [8] in low SNR regimes (the partially suppressed superficial fat signal) but has little impact on the signal in higher SNR regimes (muscles and nerves). However, complex-averaging did improve the contrast between muscle/nerve and fat, which may be useful in clinical DWI scans. Furthermore, this may prove useful at higher b-values in multishell diffusion acquisitions.Corresponding FA maps for raw and denoised magnitude- and complex-averaged data can be seen in Figure 3. Of note, FA maps showed reduce variability after denoising. However, systematic differences in FA/AD/RD values were observed after denoising for both the magnitude- and complex-averaged data (Tables 1 and 2). This suggests that while Patch2Self model removes noise, it is not capable of fully separating signal contributions. While previous spinal cord DTI data showed promising results using Patch2Self with a small number of diffusion directions [6], the effect of diffusion sampling on Patcth2Self performance has yet to be systematically studied and may be affecting the accuracy of the denoising algorithm in our clinical nerve data.
Conclusion
Given the low SNR of high-resolution nerve DTI, denoising methods may play a key role in improving the reliability of DTI parameter estimates. Patch2self was found to effectively denoise DTI data but may introduce systematic bias in clinical scenarios where a limited number of diffusion direction are available. The Patch2Self model used herein was based on ordinary least-squares; however, future work will test whether regularized and/or nonlinear models offer improvements. In addition, the impact of denoising on test-retest repeatability will be evaluated.Acknowledgements
We thank NIH/NINDS R61 NS127268 and DOD/PRMRP PR211292 for support. We would also like to thank our MRI Tech Sharmeen Maze.References
[1] Brown, J. M., Tung, T. H. H. & Mackinnon, S. E. Median to radial nerve transfer to restore wrist and finger extension: technical nuances. Neurosurgery (2010). doi:10.1227/01.neu.0000354364.13224.c8
[2] P, M., S, H., M, D. & M, S. Advances of Peripheral Nerve Repair Techniques to Improve Hand Function: A Systematic Review of Literature. Open Orthop. J. (2012). doi:10.2174/1874325001206010060
[3] Farinas, A. F. et al. Diffusion Magnetic Resonance Imaging Predicts Peripheral Nerve Recovery in a Rat Sciatic Nerve Injury Model. Plast. Reconstr. Surg. (2020). doi:10.1097/PRS.0000000000006638
[4]. Pridmore, M. D. et al. Initial findings in traumatic peripheral nerve injury and repair with diffusion tensor imaging. Ann. Clin. Transl. Neurol. (2021). doi:10.1002/acn3.51270
[5] Fadnavis, Shreyas, et al. Patch2Self: Denoising Diffusion MRI with Self-Supervised Learning.”Advances in Neural Information Processing Systems. (2023)
[6] Schilling KG, Fadnavis S, Batson J, Visagie M, Combes AJE, By S, McKnight CD, Bagnato F, Garyfallidis E, Landman BA, Smith SA, O'Grady KP. Denoising of diffusion MRI in the cervical spinal cord - effects of denoising strategy and acquisition on intra-cord contrast, signal modeling, and feature conspicuity. Neuroimage. (2023). doi: 10.1016/j.neuroimage.2022.119826.
[7] Werner S, Zinsser D, Esser M, Nickel D, Nikolaou K, Othman AE. Enhanced Image Processing Using Complex Averaging in Diffusion-Weighted Imaging of the Prostate: The Impact on Image Quality and Lesion Detectability. Diagnostics (Basel). (2023) doi: 10.3390/diagnostics13142325.
[8] Gudbjartsson H, Patz S. The Rician distribution of noisy MRI data. Magn Reson Med. 1995 Dec;34(6):910-4. doi: 10.1002/mrm.1910340618. Erratum in: Magn Reson Med 1996 Aug;36(2):332.
Figures
Figure 2: Magnitude-averaged DW image (with the diffusion gradient applied parallel to the plane of the image, or approximately perpendicular to the course of the nerves) before (a) and after (b) Patch2Self denoising. Similar results are shown for complex-averaged data before (c) and after (d) denoising in the bottom row. Red arrow: suppressed fat signal; yellow arrow: muscle; blue arrows: median and ulnar nerves.
