Speckle Reduction In Ultrasound Images Of Atherosclerotic Carotid Plaque C. Loizou 1 , C. Christodoulou 2 , C. S. Pattichis 3 , R. Istepanian 4 , M. Pantziaris 2 , A. Nicolaides 2 1 Dep. of Computer Science, Intercollege Limassol Campus, P. O. Box 51604, CY-3507 Limassol-Cyprus, e-mail: christosl@lim.Intercollege.ac.cy, 2 Cyprus Institute of Neurology and Genetics, Nicosia-Cyprus, e-mail: cschr2@ucy.ac.cy, 3 Dep. of Computer Science, University of Cyprus, e-mail: pattichi@ucy.ac.cy, 4 Dept. of Electronic and Computer Engineering, Brunel University, e-mail: Robert.Istepanian@brunel.ac.uk ABSTRACT The objective of this work was to develop six speckle reduction-filtering techniques and evaluate them together with texture analysis in the assessment of 240 ultrasound images of the carotid artery. The de- speckled filters are based on anisotropic diffusion, local statistics with higher moments, and geometric filtering. Results showed that some improvement in class separation (between symptomatic and asymptomatic plaques) of the images was evident after de-speckle filtering. 1. INTRODUCTION Ultrasound (US) imaging being non-invasive is a powerful diagnostic tool in medicine [1]-[6]. Speckle, a form of multiplicative noise corrupts medical US imaging making visual observation difficult [2], [7]-[8]. Even radiologists with sufficient experience may not often draw useful conclusions from this texture. From an engineering point of view, speckle is most often considered a dominant source of noise in US and therefore should be filtered out [2], [7], [9]-[10]. For images that contain speckle, enhancing the image by removing the speckle without destroying important features is the goal. There are mainly three categories of speckle reduction techniques: 1) Techniques which operate in 3x3, 5x5 or larger pixel moving windows utilizing the statistical properties of the image neighbourhood. These can be separated into three broad categories [1]-[7]: a) Those utilizing the statistical properties, such as the first (mean) and the second moment (variance- 2 ) in a neighbourhood. b) Those utilizing the higher statistical properties such as the third moment ( 3 ) and/or the fourth moment ( 4 ) over a pixel neighbourhood. c) Geometric techniques, which are non-linear iterative algorithms. All the speckle filters discussed in this paper fall into this category. 2) Techniques utilizing the frequency spectrum of the image, which have been proved not to be very useful for speckle reduction or image enhancement and restoration [4], [7], [10]-[12]. 3) Averaging of uncorrelated images obtained from different spatial positions, a procedure that is computational costly. Also multiple images from the same object are required [1]-[6]. The majority of the techniques presented in the literature have certain limitations: a) They are sensitive to the size and shape of the window. b) They do not enhance edges-they only inhibit smoothing near edges. c) They are not directional in the sense that in the presence of an edge, all smoothing is precluded. Instead of inhibiting smoothing in directions perpendicular to the edge are encouraging smoothing in directions parallel to the edge. d) The thresholds used in the filtering process are insufficient in the window-based approaches [7], [13]-[14]. The objective of this study was to develop new speckle reduction techniques, investigate their performance on US images and evaluate them through a number of texture descriptors extracted form the original and filtered images. Our results show that the class separation between symptomatic and asymptomatic US images of the carotid artery are, slightly better after filtering. In the following section, theoretical concepts of the proposed de-speckle filters are presented. In section three, filter analysis and evaluation carried out using 17 different texture descriptors are discussed. Section four and five give the results, and concluding remarks respectively. 2. DE-SPECKLE FILTERS In this section, the following de-speckle filters are introduced: 2.1 speckle and amnoise using first order local statistics such as the mean and the variance, 2.2 rtd utilising anisotropic diffusion and the filter anisodiff utilising speckle anisotropic diffusion, 2.3 momente using local statistics with higher statistical moments such as the skewness and kurtosis of the histogram, and 2.4 the geometric filter speck. 2.1 Local statistic filters (speckle, amnoise) Most of the introduced techniques for speckle reduction in the literature use local statistics. Their working principle may be described by a weighted average calculation using sub region statistics to estimate statistical measures over pixel windows (typically 3x3, 5x5, 7x7 sliding pixel windows). They all assume that the speckle noise model has the following multiplicative form (x denotes multiplication) [2]: g i, j = f i, j x n i, j j, i N (2.1.1) where g i, j represents the noise pixel in the middle of the moving window, f i, j represents the noise-free pixel and n i, j is a Rayleigh distributed noise on pixel. Hence the algorithms in this class may be traced back to the following equation [2]: f i, j = j i g , + k i, j x [g i, j – j i g , ] (2.1.2) where f i, j is the new estimated pixel, g i, j is the old pixel in the middle of a moving window, j i g , is the local mean value of a N1xN2 region, k i, j is a weighting factor with k [0..1], and i, j the absolute pixel coordinates. The factor k i, j is a function of the local statistics in a moving window. It can be found in the literature [2], [7] and is derived as: IEEE 14th International IEEE Conference on Digital Signal Processing 1-3 July 2002 Santorini, Greece