3013
Parallelized Patch2Self & other denoising methods on 7T diffusion weighted imaging: comparisons of quality and effects on tractometry1Beckman Institute for Advanced Science & Technology, University of Illinois at Urbana Champaign, Urbana, IL, United States, 2Johnson & Johnson, Cambridge, MA, United States, 3Carle-Illinois Advanced Imaging Center, Carle Health, Urbana, IL, United States, 4Intelligent Systems Engineering, Indiana University Bloomington, Bloomington, IN, United States, 5Stephens Family Clinical Research Institute, Carle Health, Urbana, IL, United States, 6Bioengineering, University of Illinois at Urbana Champaign, Urbana, IL, United States
Synopsis
Keywords: Data Processing, Data Processing, Denoising, Diffusion Preprocessing
Motivation: Performance of denoising methods and effects on downstream analyses for diffusion-weighted imaging at 7 Tesla are understudied.
Goal(s): Determine which denoising methods provide better performance and whether any skew tractometry outcomes.
Approach: Preliminary data acquired using two different diffusion sequences - 30 or 64 directions/shell multi-shell - were separately denoised using either Patch2Self, oversampled local-PCA, non-local means, or Marchenko-Pastur PCA before QSIPrep pre-processing and DSI Studio AutoTrack GQI.
Results: Contrast-to-noise ratios were best for Patch2Self and oversampled local-PCA, agreeing with visual assessment. Fractional anisotropy distributions were higher and mean diffusivity lower in several major bundles for non-local means than Patch2Self, especially with lower angular resolution.
Impact: The computationally-efficient parallel Patch2Self improves 7T diffusion data quality and produces lower fractional anisotropy values in tractometry of common bundles for biomarker searches than non-local means. Denoising methods should be considered in literature comparisons and image preprocessing in clinical trials.
Introduction
White matter microstructure abnormalities - attributed to inflammation, edema, demyelination, axonal degeneration, etc.- are quantified by a growing number of metrics in the search for robust biomarkers. The variety of options available for preprocessing and analyses of MRI data are a major issue in reproducible neuroscience 1,2; especially important as we address artifacts in ultra-high field 7 Tesla MRI. In this preliminary work, we explore the differences in data quality metrics and tractometry results between four common denoising methods when applied to 7 Tesla diffusion MRI:- Non-local means (NLMEANS) 3, 4: similarity-weighted means of all voxels to each target voxel.
- Overcomplete local PCA (LPCA) 5,6: applying principal component analysis (PCA) around each voxel, thresholding the eigenvalues based on estimating the original local noise variance, and performing PCA reconstruction.
- Marchenko-Pastur PCA (MP-PCA) 7,8: removing components classified as noise automatically based on the Marcenko-Pastur distribution.
- Parallelization-accelerated version of Patch2Self 9 (https://github.com/ShreyasFadnavis/p2s_parallel): a self-supervised learning method that uses the entire volume to learn a full-rank locally linear
denoiser.
Methods
The data included here are a subset of the Champaign-Urbana Population Study (https://cupopulationstudy.illinois.edu/) - an observational longitudinal study recruited based on the criteria in Table 1 - with 10 participants each from an early protocol and from the first cohort with the updated protocol. Preprocessing begins with DICOM to Brain Imaging Data Structure (BIDS) format NIFTI conversion using HeuDiConv 13 and MP2RAGE UNI images denoising using the LN_MP2RAGE_DNOISE tool from LAYNII 14,15, followed by preprocessing using QSIPrep 0.19.1 16-26. For more details of the pipeline, see the section corresponding to workflows in QSIPrep’s documentation (https://qsiprep.readthedocs.io/en/latest/preprocessing.html). Diffusion orientation distribution functions (ODFs) were reconstructed using generalized q-sampling imaging 27 with a ratio of mean diffusion distance of 1.250000 in DSI Studio). Automatic Tractography was run in DSI Studio and bundle shape statistics were calculated 28. The following quality control metrics were compared across denoising methods: contrast to noise ratio (CNR), framewise displacement 29, and neighbor correlation 30. Code for performing these analyses is available at https://tinyurl.com/2n68fymc.Results
From visual comparison, Patch2Self outperformed NLMEANS and MP-PCA in both the 64d and 30d data (see Figure 2). We note that the residuals from NLMEANS and LPCA appear to include more visually identifiable structure than those of Patch2Self, especially in the 30 directions per shell data. For the 64d data, the mean contrast-to-noise ratio in the high b-value shell was highest for both Patch2Self and LPCA; producing better neighbor correlations than NLMEANS or MP-PCA (see Figure 3) . The distribution of mean fractional anisotropy for the corpus callosum body was highest for data denoised with NLMEANS, while Patch2Self showed the lowest distribution for both acquisitions (see Figure 4). Mean diffusivity showed a higher distribution for Patch2Self than LPCA or NLMEANS. This general pattern - though varying in magnitude - appears to be common across other tracts, suggesting a skew that is stronger in lower angular resolution data. Mean length shows some differences in distributions, with more pronounced differences in the 64d data.Discussion
Quality control metrics suggest that Patch2Self performs on par with the less computationally efficient LPCA in 7T data with 64 directions per shell, especially in the higher b = 2000s/mm2 shell. Considering the 6- to 7- fold increase in speed achieved with Patch2Self Parallel (versus the original Patch2Self as shown in Figure 1), the data quality improvement in less than three minutes of real time is an appreciable benefit for large 7T diffusion datasets.As biomarker discovery continues to incorporate the more spatially informative tract profiles from tractometry, our findings show the visible effects of denoising method choice on both the shape and diffusion metric measures of key white matter bundles. Note that not all tracts are identified for all denoised data by DSI Studio AutoTrack, precluding paired differences testing in this preliminary analysis. Future work with the full CUPS dataset and 7T DWI from other sites will determine which tracts are robustly identified, include paired differences statistics, and investigate how age and site affect the magnitude of differences in data quality, motion estimates, and downstream tractometry outcomes produced by using the denoising methods compared here.
Conclusion
Compared to MP-PCA and NLMEANS, there are significant benefits for using Patch2Self Parallel for denoising in both 30 direction and 64 direction multi-shell 7T diffusion-weighted imaging data. Similar results can be achieved with LPCA, though the computational efficiency of Patch2Self Parallel is an important consideration for large datasets. Our results suggest that choice of denoising methods has a significant impact on tractometry outcomes at 7T and should be considered in as tractometry is incorporated in biomarker research.Acknowledgements
We extend our gratitude to all participants in the Champaign-Urbana Population Study, the research team, and technicians at the Carle-Illinois Advanced Imaging Center. This research was supported in part by the Illinois Computes project which is supported by the University of Illinois Urbana-Champaign and the University of Illinois System.References
1. R. A. Poldrack, C. I. Baker, J. Durnez, K. J. Gorgolewski, P. M. Matthews, M. R. Munaf`o, T. E. Nichols, J.-B. Poline, E. Vul, and T. Yarkoni, “Scanning the horizon: towards transparent and reproducible neuroimaging research,” Nature reviews neuroscience, vol. 18, no. 2, pp. 115–126, 2017.
2. R. Botvinik-Nezer, F. Holzmeister, C. F. Camerer, A. Dreber, J. Huber, M. Johannesson, M. Kirchler, R. Iwanir, J. A. Mumford, R. A. Adcock et al., “Variability in the analysis of a single neuroimaging dataset by many teams,” Nature, vol. 582, no. 7810, pp. 84–88, 2020.
3. P. Coup ́e, P. Yger, S. Prima, P. Hellier, C. Kervrann, and C. Barillot, “An optimized blockwise nonlocal means denoising filter for 3-d magnetic resonance images,” IEEE transactions on medical imaging, vol. 27, no. 4, pp. 425–441, 2008.
4. P. Coup ́e, J. V. Manj ́on, M. Robles, and D. L. Collins, “Adaptive multiresolution non-local means filter for three- dimensional magnetic resonance image denoising,” IET image Processing, vol. 6, no. 5, pp. 558–568, 2012.
5. J. V. Manj ́on, P. Coup ́e, L. Concha, A. Buades, D. L. Collins, and M. Robles, “Diffusion weighted image denoising using overcomplete local pca,” PloS one, vol. 8, no. 9, p. e73021, 2013.
6. C. G. Koay and P. J. Basser, “Analytically exact correction scheme for signal extraction from noisy magnitude mr signals,” Journal of magnetic resonance, vol. 179, no. 2, pp. 317–322, 2006.
7. J. Veraart, D. S. Novikov, D. Christiaens, B. Ades-Aron, J. Sijbers, and E. Fieremans, “Denoising of diffusion mri using random matrix theory,” NeuroImage, vol. 142, pp. 394–406, 2016.
8. J. Veraart, E. Fieremans, and D. S. Novikov, “Diffusion mri noise mapping using random matrix theory,” Magnetic resonance in medicine, vol. 76, no. 5, pp. 1582–1593, 2016.
9. S. Fadnavis, J. Batson, and E. Garyfallidis, “Patch2self: Denoising diffusion mri with self-supervised learning,” Advances in Neural Information Processing Systems, vol. 33, 2020.
10. J. P. Marques, T. Kober, G. Krueger, W. van der Zwaag, P.-F. Van de Moortele, and R. Gruetter, “Mp2rage, a self bias-field corrected sequence for improved segmentation and t1-mapping at high field,” Neuroimage, vol. 49, no. 2, pp. 1271–1281, 2010.
11. E. J. Auerbach, J. Xu, E. Yacoub, S. Moeller, and K. U ̆gurbil, “Multiband accelerated spin-echo echo planar imaging
with reduced peak rf power using time-shifted rf pulses,” Magnetic resonance in medicine, vol. 69, no. 5, pp. 1261–1267,
2013.
12. K. Setsompop, J. Cohen-Adad, B. A. Gagoski, T. Raij, A. Yendiki, B. Keil, V. J. Wedeen, and L. L. Wald, “Improving diffusion mri using simultaneous multi-slice echo planar imaging,” Neuroimage, vol. 63, no. 1, pp. 569–580, 2012.
13. X. Li, P. S. Morgan, J. Ashburner, J. Smith, and C. Rorden, “The first step for neuroimaging data analysis: Dicom to nifti conversion,” Journal of neuroscience methods, vol. 264, pp. 47–56, 2016.
14. L. Huber, B. A. Poser, P. A. Bandettini, K. Arora, K. Wagstyl, S. Cho, J. Goense, N. Nothnagel, A. T. Morgan, J. van den Hurk, A. K. M ̈uller, R. C. Reynolds, D. R. Glen, R. Goebel, and O. F. Gulban, “Laynii: A software suite for layer-fmri,” NeuroImage, vol. 237, p. 118091, 2021. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S1053811921003682
15. K. O’Brien, G. Kr ̈uger, F. Lazeyras, R. Gruetter, and A. Roche, “A simple method to denoise mp2rage,” 2013.
16. K. Gorgolewski, C. D. Burns, C. Madison, D. Clark, Y. O. Halchenko, M. L. Waskom, and S. Ghosh, “Nipype: a flexible, lightweight and extensible neuroimaging data processing framework in python,” Frontiers in Neuroinformatics, vol. 5, p. 13, 2011.
17. K. J. Gorgolewski, O. Esteban, C. J. Markiewicz, E. Ziegler, D. G. Ellis, M. P. Notter, D. Jarecka, H. Johnson, C. Burns, A. Manh ̃aes-Savio, C. Hamalainen, B. Yvernault, T. Salo, K. Jordan, M. Goncalves, M. Waskom, D. Clark, J. Wong, F. Loney, M. Modat, B. E. Dewey, C. Madison, M. Visconti di Oleggio Castello, M. G. Clark, M. Dayan, D. Clark, A. Keshavan, B. Pinsard, A. Gramfort, S. Berleant, D. M. Nielson, S. Bougacha, G. Varoquaux, B. Cipollini, R. Markello, A. Rokem, B. Moloney, Y. O. Halchenko, D. Wassermann, M. Hanke, C. Horea, J. Kaczmarzyk, G. de Hollander, E. DuPre, A. Gillman, D. Mordom, C. Buchanan, R. Tungaraza, W. M. Pauli, S. Iqbal, S. Sikka, M. Mancini, Y. Schwartz, I. B. Malone, M. Dubois, C. Frohlich, D. Welch, J. Forbes, J. Kent, A. Watanabe, C. Cumba, J. M. Huntenburg, E. Kastman, B. N. Nichols, A. Eshaghi, D. Ginsburg, A. Schaefer, B. Acland, S. Giavasis, J. Kleesiek, D. Erickson, R. K ̈uttner, C. Haselgrove, C. Correa, A. Ghayoor, F. Liem, J. Millman, D. Haehn, J. Lai, D. Zhou, R. Blair, T. Glatard, M. Renfro, S. Liu, A. E. Kahn, F. P ́erez-Garc ́ıa, W. Triplett, L. Lampe, J. Stadler, X.-Z. Kong, M. Hallquist, A. Chetverikov, J. Salvatore, A. Park, R. Poldrack, R. C. Craddock, S. Inati, O. Hinds, G. Cooper, L. N. Perkins, A. Marina, A. Mattfeld, M. Noel, L. Snoek, K. Matsubara, B. Cheung, S. Rothmei, S. Urchs, J. Durnez, F. Mertz, D. Geisler, A. Floren, S. Gerhard, P. Sharp, M. Molina-Romero, A. Weinstein, W. Broderick, V. Saase, S. K. Andberg, R. Harms, K. Schlamp, J. Arias, D. Papadopoulos Orfanos, C. Tarbert, A. Tambini, A. De La Vega, T. Nickson, M. Brett, M. Falkiewicz, K. Podranski, J. Linkersd ̈orfer, G. Flandin, E. Ort, D. Shachnev, D. McNamee, A. Davison, J. Varada, I. Schwabacher, J. Pellman, M. Perez-Guevara, R. Khanuja, N. Pannetier, C. McDermottroe, and S. Ghosh, “Nipype,” Software, 2018.
18. A. Abraham, F. Pedregosa, M. Eickenberg, P. Gervais, A. Mueller, J. Kossaifi, A. Gramfort, B. Thirion, and G. Varoquaux, “Machine learning for neuroimaging with scikit-learn,” Frontiers in Neuroinformatics, vol. 8, 2014. [Online]. Available: https://www.frontiersin.org/articles/10.3389/fninf.2014.00014/full
19. E. Garyfallidis, M. Brett, B. Amirbekian, A. Rokem, S. Van Der Walt, M. Descoteaux, and I. Nimmo-Smith, “Dipy, a library for the analysis of diffusion mri data,” Frontiers in neuroinformatics, vol. 8, p. 8, 2014.
20. N. J. Tustison, B. B. Avants, P. A. Cook, Y. Zheng, A. Egan, P. A. Yushkevich, and J. C. Gee, “N4itk: Improved n3 bias correction,” IEEE Transactions on Medical Imaging, vol. 29, no. 6, pp. 1310–1320, 2010.
21. A. Hoopes, J. S. Mora, A. V. Dalca, B. Fischl, and M. Hoffmann, “Synthstrip: skull-stripping for any brain image,” NeuroImage, vol. 260, p. 119474, 2022.
22. B. Billot, C. Magdamo, Y. Cheng, S. E. Arnold, S. Das, and J. E. Iglesias, “Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain mri datasets,” Proceedings of the National Academy of Sciences, vol. 120, no. 9, p. e2216399120, 2023.
23. B. Billot, D. N. Greve, O. Puonti, A. Thielscher, K. Van Leemput, B. Fischl, A. V. Dalca, J. E. Iglesias et al., “Synthseg: Segmentation of brain mri scans of any contrast and resolution without retraining,” Medical image analysis, vol. 86, p. 102789, 2023.
24. J. L. Andersson and S. N. Sotiropoulos, “An integrated approach to correction for off-resonance effects and subject movement in diffusion mr imaging,” Neuroimage, vol. 125, pp. 1063–1078, 2016.
25. J. L. Andersson, M. S. Graham, E. Zsoldos, and S. N. Sotiropoulos, “Incorporating outlier detection and replacement into a non-parametric framework for movement and distortion correction of diffusion mr images,” Neuroimage, vol. 141, pp. 556–572, 2016.
26. J. L. Andersson, S. Skare, and J. Ashburner, “How to correct susceptibility distortions in spin-echo echo-planar images: application to diffusion tensor imaging,” Neuroimage, vol. 20, no. 2, pp. 870–888, 2003.
27. F.-C. Yeh, V. J. Wedeen, and W.-Y. I. Tseng, “Generalized q-sampling imaging,” IEEE transactions on medical imaging, vol. 29, no. 9, pp. 1626–1635, 2010.
28. F.-C. Yeh, “Shape analysis of the human association pathways,” Neuroimage, vol. 223, p. 117329, 2020.
29. J. D. Power, K. A. Barnes, A. Z. Snyder, B. L. Schlaggar, and S. E. Petersen, “Spurious but systematic correlations in functional connectivity mri networks arise from subject motion,” Neuroimage, vol. 59, no. 3, pp. 2142–2154, 2012.
30. F.-C. Yeh, I. M. Zaydan, V. R. Suski, D. Lacomis, R. M. Richardson, J. C. Maroon, and J. Barrios-Martinez, “Differential tractography as a track-based biomarker for neuronal injury,” Neuroimage, vol. 202, p. 116131,2019.
Figures
