Training a neural network for Gibbs and noise removal in diffusion MRI

Abstract

To develop and evaluate a neural network-based method for Gibbs artifact and noise removal. Methods: A convolutional neural network (CNN) was designed for artifact removal in diffusion-weighted imaging data. Two implementations were considered: one for magnitude images and one for complex images. Both models were based on the same encoder-decoder structure and were trained by simulating MRI acquisitions on synthetic non-MRI images. Results: Both machine learning methods were able to mitigate artifacts in diffusionweighted images and diffusion parameter maps. The CNN for complex images was also able to reduce artifacts in partial Fourier acquisitions. Conclusions: The proposed CNNs extend the ability of artifact correction in diffusion MRI. The machine learning method described here can be applied on each imaging slice independently, allowing it to be used flexibly in clinical applications.

Fast qualitY conTrol meThod foR derIved diffUsion Metrics (YTTRIUM) in big data analysis: U.K. Biobank 18,608 example

“…Nevertheless, derived diffusion metrics from DTI or DKI may still deviate from expected range, for example, due to remaining artefacts and numerical misestimations (see Supporting Information for examples of the distorted diffusion maps). Despite improved post‐processing algorithms (Ades‐Aron et al, 2018) for raw diffusion data, there is no consensus yet about a unified pipeline for diffusion data, for example, noise correction methods are regularly revised (Muckley et al, 2021), Gibbs ringing artefacts can remain in the images due to different origins such as a partial Fourier (Muckley et al, 2021), frequency drift effect (Vos et al, 2016) can bias the estimations, in particular in the case of advanced dMRI protocols, and diffusion gradient non‐linearity correction (Rudrapatna, Parker, Roberts, & Jones, 2020) might be important as well. Notably, a number of artefacts in the scalar diffusion maps could be minimised by applying a state‐of‐the‐art algorithms such as, for example, eddy_gpu, if a computational facility allows that.…”

Same Brain, Different Look? – the Impact of Scanner, Sequence and Preprocessing on Diffusion Imaging Outcome Parameters

“…Neither the Siemens default window filtering nor the denoising tool did attenuate the GR -from the latter one, we did not expect any artefact reduction but the filter method by Siemens was expected to reduce the GR artefact through global smoothing and did not meet these expectations. Promising future steps towards automatic GR artefact detection and reduction besides the Kellner tool might be the application of convolutional neural networks as suggested and experimentally verified by Zhang et al (2019), Zhao et al (2020) and Muckley et al (2021).…”

“…While several a posteriori correction methods have been implemented in commonly used preprocessing software to mitigate such imaging artefacts (Smith et al, 2004;Woolrich et al, 2009;Andersson & Sotiropoulos, 2016), one of the most ubiquitous artefacts, the Gibbs ringing (GR), received less attention. Only in recent years, attempts addressing the removal of this artefact have been published (Perrone et al, 2015;Kellner et al, 2016;Veraart et al, 2016a;Zhang et al, 2019;Zhao et al 2020;Muckley et al, 2021). GR appears due to a k-space truncation along finite image sampling and presents as signal oscillations at sharp intensity transitions leading to physically implausible signals (PIS) and erroneous FA values (e.g.…”

“…Physically implausible FA values show as bright voxels in the DW images. Even though this adverse impact of GR on DTI-derived metrics has been recognised for almost three decades (Constable & Henkelmann, 1991;Pan, 2001;Gelb & Archibald, 2002;Perrone et al, 2015;Veraart et al, 2016a;Muckley et al, 2021), until recently, no state-of-the-art preprocessing pipeline such as FSL (Smith et al, 2004;Woolrich et al, 2009)) had a GR artefact removal tool included yet. Whether such preprocessing would indeed lessen acquisition-related differences in DTI-derived outcome measures has not been evaluated yet.…”

Training Data Distribution Significantly Impacts the Estimation of Tissue Microstructure with Machine Learning

“…Recent works that leverage supervised ML for model parameter estimation typically employ one of two training strategies: (1) parameter combinations obtained from traditional model fitting and the corresponding measured qMRI signals [4] [6] [9] [14] [15] [11] [16] [17], or (2) parameters sampled uniformly from the entire plausible parameter space with simulated qMRI signals [5] [18] [19] [20] [21] [22] [23] [24]. While both of these approaches are limited by the model used to estimate parameters or simulate signals, simulations allow considerably more freedom in choosing training data [25] [26] [27]. However, it is not clear how best to utilise this freedom, as the impact of training data distribution on parameter estimation has yet to be examined.…”