Speckle reduction in medical ultrasound images using an unbiased non-local means method
Introduction
Ultrasound (US) imaging is a prominent diagnostic imaging technique since it offers several benefits such as more economic, safe, real-time and ergonomically adaptable in practice. The clinical utility of ultrasound is useful in the visual inspection of internal organs, muscles and tissues or in quantitative analysis in order to obtain measures that can be used as biomarkers for diagnosis [1], [2], [3]. The impact of ultrasound is felt in several fields of medicine [4], [5], both in invasive and non-invasive applications. Non-invasive applications include the usage of US in outpatient clinics, while invasive applications include the assessment of many coronary arterial diseases [6], [7], [8]. B-mode US images suffer from quality degradation due to speckle noise [9]. Many post-processing algorithms such as image segmentation, registration or classification of tissue parenchyma are based on raw US images and hence can get affected by the presence of the speckle noise [10] if not removed. Hence, effective speckle reduction is vital for proper clinical interpretation and quantitative measurements.
A plethora of despeckling methods have been developed for improving the quality of US images that can be performed either in the transform domain or in the spatial domain [11]. Loizou et al. in [3] conducted a comparative study of different US despeckling methods for carotid artery, and provided a Matlab toolbox for US image despeckling. The details are available in [12]. The Lee's filter [13], Frost's filter [14], and Kuan's filter [15] are the most widely discussed spatial adaptive filters to attenuate the speckle noise. These classical filters consider the speckle as multiplicative noise and described them mathematically using a Gaussian distributed noise model. Lopes et al. [16] proposed improved versions of the Lee's and Frost's filters by organizing the pixels in different classes in which precise processing is defined. Squeeze box filter (SBF) designed in [17], [18] removes outliers at each iteration and smooths the random distributed pixel values to some confining value through adaptively computed mean. Recently, the Rayleigh-Maximum-Likelihood (R-ML) filter in [19] was employed with the Rayleigh density model and the ML method was adapted for solving the estimation problem.
Diffusion filters such as speckle reducing anisotropic diffusion filter (SRAD) [20] and its sophisticated variants such as the detail preserving anisotropic diffusion filter (DPAD) [21] and the oriented speckle reducing anisotropic diffusion filter (OSRAD) [22] eliminate the unwanted components from an image by finding solutions to a partial differential equation (PDE)[23]. However, these filters suffer from a significant loss of sharp-transition detail. The probability-driven OSRAD (POSRAD) [24] computes the statistical models of tissues and provides superior results compared to above mentioned diffusion filters.
The multi-scale techniques based on Wavelets are another class of filters introduced for speckle reduction in US imaging. These filters can be mainly classified into three categories-namely thresholding methods [25], [26], coefficients correlation methods [27] and Bayesian estimation methods. In [28], [29], [30], the Bayesian framework was investigated to perform wavelet thresholding adapted to the non-Gaussian statistics of the signal. Pizurica et al. [31] proposed a non-decimated wavelet-based denoising method, called GenLik, which uses a generalized likelihood ratio formulation and imposes no prior statistics for noise and data. Similar to wavelet based approaches, principle component analysis (PCA)-based method in [32] is an example of a transform domain despeckling filter used for denoising medical US images.
In the past, different hybrid filters have been proposed to integrate the benefits of the above mentioned paradigms. In [33], the image is decomposed into two components by an adaptive filter and each component, after performing Donoho's soft thresholding, is merged to suppress speckle. The hybrid filter in [34] combines PDE-based approaches and a wavelet transform. More recently, a patch-based non-local recovery paradigm such as optimized Bayesian NL-means with block selection (OBNLM) [35] has been applied to reduce speckle noise.
Since the speckle intensity has a signal dependent nature, filters based on the standard additive Gaussian noise model are inadequate. Therefore, specific filters are required to suppress speckle without compromising important image features. Many statistical models have been attempted in the literature; specifically, Rayleigh, K, homodyned K, Nakagami, generalized Nakagami and Rician inverse Gaussian (RiIG) distributions have been shown to achieve a perfect statistical characterisation of ultrasound signals and to address the restoration of speckled images [19]. When compressionless data with fully developed speckle is considered, it is well known that the interpolated output B-mode US images follow a Gamma distribution, which is a good approximation for the weighted sum of Rayleigh variables [36], [37].
In this article, an unbiased NLM speckle filter based on Gamma statistics has been presented. We used a three parameter Gamma distribution function to fit the real US image and the ML estimation was adapted to find the two key parameters that control the filter performance. The proposed filter was scientifically validated by taking into consideration both real US and synthetic standardized images.
The remainder of this paper is organized as follows: Section 2 gives an overview of noise characteristics in B-mode US images. Section 3 elaborates the ML approach to estimate the bias in terms of the parameters of the Gamma distribution. The proposed NLM filtering procedure is discussed in Section 4. Section 5 reports the quantitative and qualitative results, followed by the conclusions in Section 6.
Section snippets
Noise characteristics in B-mode ultrasound images
In [38], Goodman described the fundamentals and statistical properties of speckle noise and mathematically modelled it as a complex random walk, represented as a sum of a huge number of complex phasors. Even though these phasors can have either constructive or destructive relationship with each other, the destructive interferences cause the speckle formation. The severity of destructive interference depends on the relative phase between two overlapping reflected echoes produced by the
Maximum likelihood based parameter estimation
In this section, we first introduce our speckle model. Based on the assumption that the interpolated Cartesian image of the corrupted back-scattered signal fits Gamma distribution, a Gamma distributed synthetic image G can be produced as:where F and Ng denote the noiseless image and fading variable, respectively.
For the statistically independent random observations x = x1, x2, …, xn from a Gamma distribution, the PDF of the observations can be described by a three-parameter Gamma
Non-local means estimation of true underlying pixel value
The NLM algorithm compares neighbourhood around each pixel called blocks and performs a weighted average of pixels based on the similarity of neighbourhoods computed using Euclidean distance. Consider the discrete noisy image , where u(y) corresponds to the noisy image value at pixel location y. Let us assume that the search window size and similarity window size be ((2t1 + 1) × (2t1 + 1)) and ((2t2 + 1) × (2t2 + 1)), respectively. The NLM estimator for the filtered value at a location r is
Experiments and results
In this section, we report the experimental results obtained on synthetic and real US images. Results were evaluated through visual inspection and quantitative analysis and the proposed filter is compared with other state-of-the-art methods.
Various parameters used for the filters discussed in this section are as follows:
- 1.
SBF [17]: Averaging window size is 3 × 3; total iterations of 45 and Gaussian noise with standard deviation = 0.1 added at every 5th iteration.
- 2.
SRAD Filter [20]: The SRAD parameters
Conclusion
The proposed unbiased NLM speckle filter based on Gamma statistics has mainly two phases. In the first phase the shape and scale parameters, ρ and β, of the Gamma distribution are computed from the image by taking the mode of the ρ and β values estimated locally using an ML estimator. In the second phase, an unbiased NLM method was applied to estimate the true underlying intensity of each and every pixels in the image. Experiments were carried out on both synthetic and real B-mode US images to
Acknowledgements
We acknowledge Harman S. Suri, Mira Loma, Sacramento, CA, USA for proof reading the manuscript. We acknowledge our clinicians (co-authors of this paper) for supplying ultrasound image data and expert scorings.
References (51)
- et al.
Comparative study of different denoising filters for speckle noise reduction in ultrasonic B mode images
IJIGSP
(2013) - et al.
Despeckle filtering software toolbox for ultrasound imaging of the common carotid artery
Comput. Methods Progr. Biomed.
(2014) - et al.
Ultrasound echo envelope analysis using a homodyned K distribution signal model
Ultrason. Imaging
(1994) - et al.
Enhancement and bias removal of optical coherence tomography images: an iterative approach with adaptive bilateral filtering
Comput. Biol. Med.
(2016) - et al.
MRI denoising using non-local means
Med. Image Anal.
(2008) - et al.
A new method for IVUS-based coronary artery disease risk stratification: a link between coronary & carotid ultrasound plaque burdens
Comput. Methods Progr. Biomed.
(2015) - et al.
GyneScan: an improved online paradigm for screening of ovarian cancer via tissue characterization
Technol. Cancer Res. Treat.
(2014) - et al.
Comparative evaluation of despeckle filtering in ultrasound imaging of the carotid artery
IEEE Trans. Ultrason. Ferroelectr. Freq. Control
(2005) - et al.
Advances in Diagnostic and Therapeutic Ultrasound Imaging
(2008)