Structure-adaptive sparse denoising for diffusion-tensor MRI

https://doi.org/10.1016/j.media.2013.01.006Get rights and content

Abstract

Diffusion tensor magnetic resonance imaging (DT-MRI) is becoming a prospective imaging technique in clinical applications because of its potential for in vivo and non-invasive characterization of tissue organization. However, the acquisition of diffusion-weighted images (DWIs) is often corrupted by noise and artifacts, and the intensity of diffusion-weighted signals is weaker than that of classical magnetic resonance signals. In this paper, we propose a new denoising method for DT-MRI, called structure-adaptive sparse denoising (SASD), which exploits self-similarity in DWIs. We define a similarity measure based on the local mean and on a modified structure-similarity index to find sets of similar patches that are arranged into three-dimensional arrays, and we propose a simple and efficient structure-adaptive window pursuit method to achieve sparse representation of these arrays. The noise component of the resulting structure-adaptive arrays is attenuated by Wiener shrinkage in a transform domain defined by two-dimensional principal component decomposition and Haar transformation. Experiments on both synthetic and real cardiac DT-MRI data show that the proposed SASD algorithm outperforms state-of-the-art methods for denoising images with structural redundancy. Moreover, SASD achieves a good trade-off between image contrast and image smoothness, and our experiments on synthetic data demonstrate that it produces more accurate tensor fields from which biologically relevant metrics can then be computed.

Highlights

► We propose a structure-adaptive sparse denoising method for diffusion tensor MRI. ► A similarity measure is defined to find similar patches to arrange into 3D arrays. ► A structure-adaptive window pursuit method is designed to achieve array sparsity. ► The proposed SASD performs well for images with sufficient structural redundancy. ► SASD facilitates subsequent tensor field and biological relevant metrics analysis.

Introduction

Magnetic resonance imaging (MRI) has benefited from many technological developments since its advent in the 1970s. However, the fundamental trade-offs between image resolution and signal-to-noise ratio (SNR) on the one hand, and between physiological and clinical constraints on acquisition speed on the other hand, often translate to spurious artifacts such as noise, partial volume, and bias field (Basser and Pajevic, 2000, Farrell et al., 2007). An eloquent example is diffusion-tensor MRI (DT-MRI), which has become quite popular over the last decade because of its potential for in vivo and non-invasive characterization of the three-dimensional (3D) fiber architecture of anatomical organs (Behrens et al., 2003, Bondiau et al., 2008, Clatz et al., 2005, Delingette et al., 2012, Deriche et al., 2009, Descoteaux et al., 2009, Durrleman et al., 2011, Fillard et al., 2011, Galanaud et al., 2010, Galban et al., 2005, Guevara et al., 2011, Le Bihan, 2003, Lenglet et al., 2009, Messe et al., 2011, Mori et al., 2009, Rohmer et al., 2007, Smith et al., 2006, Wakana et al., 2004, Wu and Tseng, 2006). The effects of Rician noise on DT-MRI are severe because of the inherent nature of the imaging process—the higher the tissue anisotropy, the lower the intensity in the diffusion-weighted images (DWIs), and hence the higher the sensitivity to noise (Awate and Whitaker, 2007).

Denoising magnetic resonance images is an important problem; the most popular approaches are Bayesian statistic approaches (Awate and Whitaker, 2007, Basser and Pajevic, 2003, Lu et al., 2006), PDE-based approaches (Chen and Edward, 2005, Fillard et al., 2007), wavelet-based methods (Pižurica et al., 2006, Yu et al., 2009), methods based on spatial correlation (Barash and Comaniciu, 2004, Kervrann and Boulanger, 2006, Manjón et al., 2008), and sparse representation denoising (Chatterjee and Milanfar, 2009, Donoho et al., 2006, Elad and Aharon, 2006, Varshney et al., 2008). No particular method shows good performance for all relevant aspects of MRI, including bias reduction, artifacts removal, structure- and edge-preservation, and good generality/reliability trade-off.

Buades et al. (2005) surveyed spatial correlation techniques and proposed the so-called non-local means (NL-means) method which is based on the assumption that natural images usually have structural redundancy. The underlying idea is that any pixel has similar pixels that are distributed not only in its local neighborhood but also in the whole spatial domain of the image, which allows to account for the global information associated with large structures, and hence to overcome the limitations of local search. The NL-means approach has been applied to denoising (Boulanger et al., 2010, Brox et al., 2008, Katkovnik et al., 2010, Katkovnik and Spokoiny, 2008), super-resolution restoration (Manjón et al., 2010, Rousseau, 2010), compressed sensing (Danielyan et al., 2010, Marim et al., 2010) and inpainting (Elad et al., 2005). NL-means detects pixel similarity by exploring patch similarity for better robustness to noise, and similar patches can be grouped together to achieve sparse representation in a suitable transform domain in which denoising reduces to a shrinkage operation. Therefore, non-local self-similarity and sparse representations are key elements of state-of-the-art filtering algorithms.

The similarity measure used in NL-means evaluates the correlation between patches globally via the standard Euclidean distance and neglects the local consistency between a patch and its center pixel. In particular, in the case where different patches are close in terms of Euclidean distance but have center pixels with quite different values, the standard similarity measure may cause a loss of fine structures and edge blurring – the same behavior is observed when the number of patches similar to the reference patch is too small. Alternatively, the block-matching algorithm (BM3D) proposed by Dabov et al. (2007), which exploits a specific non-local image model by grouping similar patches and by collaborative filtering, is another interesting denoising approach. The BM3D filter stacks similar two-dimensional (2D) image patches into 3D arrays which are processed in a 3D transform domain to attenuate the noise component. However, patches containing fine image details or singularities or sharp edges are examples where a non-adaptive transform is not able to produce good sparse representations (Foi et al., 2007, Hammond and Simoncelli, 2008, Sikora and Makai, 1995); for such patches, the filter may introduce artifacts around discontinuities where the visual attention is often mainly focused.

We propose a 3D structure-adaptive sparse denoising (SASD) algorithm for DT-MRI which we apply to human cardiac DWIs. Our method can be decomposed into four steps. First, we form groups of similar patches in the DWI to be denoised using a modified structure-similarity index. Second, each set of similar patches is arranged into a 3D array, and a structure-adaptive window pursuit method is used to adapt to the local image features. Third, the resulting structure-adaptive 3D arrays are denoised by Wiener shrinkage in the transform domain defined by 2D principal component analysis in the image domain and Haar transformation in the third dimension. Finally, the noise-free DWI estimate is obtained by weighted averaging of the denoised structure-adaptive patches.

The paper is organized as follows. Section 2 describes the SASD algorithm and its main components, including the definition of the proposed similarity measure, the design of the structure-adaptive window pursuit method, and the Wiener filtering operation in the transform domain. The experimental setup is discussed in Section 3, and our results are presented in Section 4, where our approach is compared to state-of-the-art denoising methods, namely, Bayesian least-squares Gaussian scale mixture (BLS-GSM) (Portilla et al., 2003) and Field of experts (FOE) (Roth and Black, 2005). Concluding remarks are given in Section 5.

Section snippets

Methodology

In a nutshell, the proposed approach is inspired by the NL-means method to filter sparse local representations of the image to be denoised. Therefore, we describe the standard NL-means method in Section 2.1 prior to our SASD algorithm in Section 2.2.

Simulated data

Since real noise-free cardiac DWIs cannot be obtained by real DT-MRI acquisitions, we simulated biologically informed DWIs based on physical measures from polarized light imaging (we used the simulation method proposed in Wang et al. (2012)). The simulated DWIs – an example of which is displayed in Fig. 5a – reveal the left and right ventricles of a human heart and contain various textures and fine structures. They form a set of realistic short-axis cardiac images (size 256 × 256, intensity range

Comparison of BLS-GSM, FOE and SASD

Fig. 5 shows the denoising results for the simulated DWI associated with the first diffusion direction. The image to be denoised is displayed in Fig. 5b; it was obtained by adding Rician noise with standard-deviation σ = 20 to the noise-free image shown in Fig. 5a (this noise component significantly affects the fine structures). The solution obtained by BLS-GSM is over-smoothed (Fig. 5c), and the output of FOE has chessboard artifacts pointed out by the arrows in Fig. 5d. By contrast, SASD shows

Conclusions

We proposed a new DWI denoising method, namely, structure-adaptive sparse denoising (SASD), that collects adaptive similar neighborhoods to increase redundancy and hence to facilitate the removal of the noise. To avoid the drawbacks of standard similarity measures, we have defined a new constrained similarity measure based on the local mean and the SSIM index. We also proposed to search for structure-adaptive neighborhoods in the reference patches, which allows local adaptation to image

Acknowledgments

This work was supported by the NNSF of China (61271092), the International S&T Cooperation Project of China (2007DFB30320), and the French ANR (ANR-09-BLAN-0372-01). The authors would like to thank L.H. Wang for supplying the simulated cardiac DWI data used in our experiments in Section 4.1.

References (66)

  • J.V. Manjón et al.

    MRI denoising using non-local means

    Med. Image Anal.

    (2008)
  • J.V. Manjón et al.

    Non-local MRI upsampling

    Med. Image Anal.

    (2010)
  • F. Rousseau

    A non-local approach for image super-resolution using intermodality priors

    Med. Image Anal.

    (2010)
  • S.M. Smith et al.

    Tract-based spatial statistics: voxelwise analysis of multi-subject diffusion data

    Neuroimage

    (2006)
  • L. Wang et al.

    Multiscale modeling and simulation of the cardiac fiber architecture for DMRI

    IEEE Trans. Biomed. Eng.

    (2012)
  • P.S. Awate et al.

    Feature-preserving MRI denoising: a nonparametric empirical bayes approach

    IEEE Trans. Med. Imaging

    (2007)
  • P.J. Basser et al.

    Statistical artifacts in diffusion tensor MRI caused by background noise

    Magn. Reson. Med.

    (2000)
  • P.J. Basser et al.

    A normal distribution for tensor-valued random variables: application to diffusion tensor MRI

    IEEE Trans. Med. Imaging

    (2003)
  • T.E. Behrens et al.

    Non-invasive mapping of connections between human thalamus and cortex using diffusion imaging

    Nat. Neurosci.

    (2003)
  • P.Y. Bondiau et al.

    Biocomputing: numerical simulation of glioblastoma growth using diffusion tensor imaging

    Phys. Med. Biol.

    (2008)
  • J. Boulanger et al.

    Patch-based nonlocal functional for denoising fluorescence microscopy image sequences

    IEEE Trans. Med. Imaging

    (2010)
  • T. Brox et al.

    Efficient nonlocal means for denoising of textural patterns

    IEEE Trans. Image Process.

    (2008)
  • A. Buades et al.

    A review of image denoising algorithms with a new one

    Multiscale Model. Simul.

    (2005)
  • P. Chatterjee et al.

    Clustering-based denoising with locally learned dictionaries

    IEEE Trans. Image Process.

    (2009)
  • B. Chen et al.

    Noise removal in magnetic resonance diffusion tensor imaging

    Magn. Reson. Med.

    (2005)
  • O. Clatz et al.

    Realistic simulation of the 3-D growth of brain tumors in MR images coupling diffusion with biomechanical deformation

    IEEE Trans. Med. Imaging

    (2005)
  • P. Coupé et al.

    An optimized blockwise nonlocal means denoising filter for 3-D magnetic resonance images

    IEEE Trans. Med. Imaging

    (2008)
  • K. Dabov et al.

    Image denoising by sparse 3D transform domain collaborative filtering

    IEEE Trans. Image Process.

    (2007)
  • A. Danielyan et al.

    Spatially adaptive filtering as regularization in inverse imaging: compressive sensing, upsampling and super-resolution

  • H. Delingette et al.

    Personalization of cardiac motion and contractility from images using variational data assimilation

    IEEE Trans. Biomed. Eng.

    (2012)
  • M. Descoteaux et al.

    Deterministic and probabilistic tractography based on complex fibre orientation distributions

    IEEE Trans. Med. Imaging

    (2009)
  • L.D. Donoho et al.

    Stable recovery of sparse overcomplete representations in the presence of noise

    IEEE Trans. Inf. Theory

    (2006)
  • M. Elad et al.

    Denoising via sparse and redundant representations over learned dictionaries

    IEEE Trans. Image Process.

    (2006)
  • Cited by (32)

    • Denoise magnitude diffusion magnetic resonance images via variance-stabilizing transformation and optimal singular-value manipulation

      2020, NeuroImage
      Citation Excerpt :

      Part of our future work would be to examine how the kernel size of each module can be optimized and how this module-specific kernel size optimization would further improve the denoising performance of the proposed method. Additionally, for the denoising module (in which a standard denoising method is applied), we will consider using more advanced patch definitions (such as those through non-local spatial (Buades et al., 2005; Dabov et al., 2007; Maggioni et al., 2013; Manjon et al., 2015) and angular (St-Jean et al., 2016) matching and through structure adaptation (Bao et al., 2013; Foi et al., 2007)) to form the data matrices to be denoised and will explore how the combined use of these more advanced patch definitions would leverage the denoising performance of the proposed method by maximizing the information redundancy (thereby minimizing the rank) of the data matrices. Incorporating Foi’s VST recipe (Foi, 2011), the proposed denoising method is specifically devised for denoising Rician images.

    • Atlas construction of cardiac fiber architecture using a multimodal registration approach

      2017, Neurocomputing
      Citation Excerpt :

      Up to now, in vivo cardiac DT-MRI is extremely challenging, mostly because of motion artifacts (caused by beating heart and breathing motion) [7] as well as limited image spatial resolution [4,8]. Alternatively, DTI on ex vivo hearts gives an insight view to qualify the myofiber micro-structure with more details [9–11]. As such, the availability of a 3D diffusion atlas of the myofibers makes the understanding of different myocardial mechanical functions more deep.

    • Denoise diffusion-weighted images using higher-order singular value decomposition

      2017, NeuroImage
      Citation Excerpt :

      Thus, the latter approach has been widely adopted to reduce the effect of noise on the diffusion parameter mapping in the past few decades. The current DW image denoising methods include algorithms based on wavelet transforms (Wirestam et al., 2006), partial differential equations (Ding et al., 2005; Parker et al., 2000), propagation-separation approaches (Becker et al., 2014, 2012), nonlocal means (Descoteaux et al., 2008; Wiest-Daessle et al., 2007), principal component analysis (Arfanakis et al., 2002; Bao et al., 2013; Manjon et al., 2013), low-rank approximations (Lam et al., 2014, 2016), and sparseness among grouped similar patches (Bao et al., 2013). Recently, we developed an MR image denoising algorithm based on the higher-order singular value decomposition (HOSVD) transform (Zhang et al., 2015a), which exhibits state-of-the-art performance in denoisingT1-, T2- and proton density-weighted MR images.

    View all citing articles on Scopus
    View full text