DENOISING AND SEGMENTATION OF ECHOCARDIOGRAPHIC IMAGES IN MULTILPE VIEWS

Abstract

Echocardiography is the most commonly used first-line imaging modality in the assessment of cardiac chamber and valvular abnormalities. The valvular abnormalities can be classified as regurgitation and stenosis. Aortic regurgitation (AR) is a valvular disorder due to the retrograde flow of blood from the aortic valve into the left ventricle during the diastole. The aetiologies and the consequences of AR are diagnosed using transthoracic echocardiographic (TTE) images acquired in parastrenal long axis (PLAX), parasternal short axis (PSAX), apical four chamber (A4C), apical two chamber (A2C) and apical five chamber (A5C) views. The Doppler imaging modalities such as continuous wave Doppler (CWD) and color Doppler echocardiography along with conventional B-Mode (brightness mode) and MMode (motion mode) images are used hand-in-hand to detect the prevalence of regurgitation, better understanding of the mechanism of regurgitation and quantification of severity along with its repercussions. But the technical research is more concentrated on despeckling, and segmentation of good quality B-Mode images acquired in a particular view from healthy adults. Hence, it is necessary to study the applications of despeckling and segmentation techniques for TTE images acquired in multiple views from patients diagnosed with AR. Current research work attempts to overcome the existing caveats by integrated processing of B-Mode, CWD, and color Doppler echocardiography images. The work looks for the best denoising and segmentation techniques suitable for different modalities of echocardiographic imaging in multiple views. Based on the exhaustive technical and clinical literature review, the following research objectives have been framed: 1) To propose despeckling methods for the B-Mode TTE images of aortic valve and cardiac chambers acquired in multiple views using different windows. 2) Comparative analysis of state-of-the-art despeckling techniques and texture features for the B-Mode and CWD images. 3) To propose delineation techniques for tracing the outer spectrum of CWD images. 4) Comparative analysis of segmentation techniques using the TTE images acquired in multiple views and windows. These research objectives are accomplished in the following manner. Speckle noise present in cross sectional TTE images makes it difficult to consistently perform delineation of the cardiac structure. It is necessary to suppress noise and enhance contrast without altering the fine details present in the images. To address the issue of speckle noise, six despeckling techniques are proposed in this thesis. The first proposed technique is based on multiscale techniques. The multiscale techniques are employed for speckle noise in the logarithmic domain, considering the approximated additive noise model of despeckling. The transformation of images into the logarithmic domain, application of shrinkage techniques and bringing the image back into the original space are the three common steps for all multiscale techniques. Eight shrinkage techniques are analyzed for ii despeckling of TTE images in multiple views. The performances of logarithmic multiscale techniques are compared with adaptive and diffusion based filtering techniques. The denoised images are enhanced using Butterworth filter. The integrated effects of denoising and enhancement are tested on active contour, region, watershed and edge based segmentation techniques. Further, to make use of the advantages of M-band wavelets, it is proposed to combine M-band ridgelet with neighborhood coefficient thresholding for the despeckling of TTE images. The thresholds of coefficients are computed using neighborhood coefficient thresholding technique and reconstructed to obtain the denoised images. The second proposed despeckling technique is known as the hybrid triangulation moving average (TMAV) fuzzy filter. The performance of TMAV filter is fine tuned for despeckling of TTE images by combining it with adaptive Wiener filter. Four fuzzy filters have been analyzed in the logarithmic domain for speckle noise reduction. The despeckling performances of all the four fuzzy filters are fine tuned by combining them individually with adaptive Wiener filter. The integrated fuzzy filter is the third proposed despeckling technique which is the improved version of hybrid TMAV fuzzy filter. It is based on the integration of geometric, Wiener and fuzzy filters. The hybrid homomorphic fuzzy (HHF) is the fourth proposed technique which is the combination of logarithmic fuzzy filter and anisotropic diffusion filter. The diffusion based methods are known for speckle noise suppression, and edge preservation capability but they do not perform well when the noise contamination in the images is high. The fuzzy filters have superior figure of merit in comparison to anisotropic diffusion filters. The advantages of anisotropic diffusion and fuzzy filters are integrated in the proposed HHF filter. The hybrid posterior sampling based Bayesian estimation (PSBE) is the fifth proposed technique for despeckling of TTE images. The performance of logarithmic PSBE technique degrades considerably for images contaminated with high amount of noise. To address this issue, an additional adaptive filter is embedded into logarithmic PSBE and is known as the hybrid PSBE technique. The contrast of output images are enhanced using Butterworth filter. The effects of denoising and enhancement on segmentation are studied using three basic techniques namely the edge, region and multistage watershed. The extreme total variation bilateral (ETVB) is the sixth proposed technique for denoising of TTE images of the cardiac structures. The regularizer term of the total variation (TV) filter is replaced with the bilateral (BL) term in the proposed ETVB filter. The true information is incorporated in the algorithm using Bayesian inference and probability density function. Applications of gradient projection based restoration methods have been analyzed for speckle noise reduction of TTE images. In an effort to define the best despeckling filter for the B-Mode, CWD and color Doppler images in multiple views, a comparative study is under taken in this thesis. The applications of 48 filters are analyzed for the B-Mode, CWD, MM, and color Doppler images where the performance analysis is in terms of sixteen image quality metrics, visual quality assessment iii and clinical validation. Both, traditional and blind assessment parameters are computed for assessment of noise suppression, edge and structure preservation. The despeckling filters are grouped into eight type’s namely local statistics, fuzzy, Fourier, multiscale, nonlinear iterative, total variation, nonlocal mean and hybrid filters. It has been observed that the median filter and Gaussian low pass filter are more commonly used for reduction of noise in the CWD images. The applications of state-of-theart despeckling filters have not been extensively analyzed for the CWD images. Therefore, applications of despeckling filters analyzed for B-Mode images have also been analyzed for the CWD, pulse wave Doppler (PWD), M-Mode and color Doppler images. The despeckling performance of filters have also been compared in terms of 65 texture features computed from the denoised B-Mode, M-Mode, CWD, and color Doppler echocardiographic images. The set of features include five first order statistical (FOS), 26 spatial gray level dependence matrix (SGLDM), four gray level difference statistics (GLDS), four statistical feature matrix (SFM), six Laws texture energy measures (LTEM), four fractal dimension, two Fourier power spectrum and five neighborhood gray tone difference matrix features. All the texture features have been computed before and after the application of despeckling filters. The segmentation of original and the pre-processed B-Mode, CWD and color Doppler images is taken up as the next objective. This objective looks for the best delineation technique for each modality image in multiple views. Initially, synthetic images with different amount of intensity in-homogeneity have been delineated using segmentation techniques based on various variants of the edge, region, watershed, fuzzy, active contour and level set techniques. The applications of these techniques are analyzed for the contouring of synthetic, color Doppler, B-Mode and CWD images. The objective of segmenting the B-Mode images is to trace the inner boundaries of left ventricle (LV) and aortic valve (AV) along with the leaflets of AV and mitral valve (MV). The B-Mode images acquired in two parasternal and three apical windows are used for analysis of segmentation techniques. The color Doppler images in PLAX and A5C are segmented to trace the outer boundary of regurgitant jet area. The outer spectrum of the CWD images is traced using multiple segmentation techniques. To begin with, the edge, region and multi-stage watershed segmentation techniques are analysed. The performances of these techniques are improved by combining them with filters. The existing techniques such as wavelet based scale multiplication edge detection (SMED) approach, intuitionistic fuzzy divergence (IFD) based edge detection, soft thresholding, topological derivative based delineation, Magagnin and Kiruthika method have been employed for tracing the outer boundaries of the images. Further, the techniques based on active contour and level set have been employed for tracing the boundaries in the presence of intensity in-homogeneity. This set of techniques includes methods such as reaction diffusion, region scalable fitting (RSF), global minimization of active contour iv (GMAC), Laplacian fitting energy, statistical and variational multiphase level set (SVMLS) approach, active contour without edges, selective binary and Gaussian filtering regularized level set. The manually segmented B-Mode images are compared with results obtained on application of local region based active contour segmentation technique. The estimated parameters on manual segmentation are compared with those obtained on application of semi-automated segmentation. The analysis of segmentation techniques for CWD images is carried out using filtered as well original noisy images. The Gaussian and median filters used in Kiruthika method, Magagnin method and reaction diffusion (RD) based active contour method are replaced by the despeckling techniques in the proposed modifications for these three delineation techniques. These basic filters are replaced by ten despeckling filters such as DsFlsmv, DsFmedian, DsFhmedain, DsFad, DsFsrad, DsFlsminsc, DsFhomog, DsFWiener, DsFhomog and DsFgf4d. Further, performances of the modified RD method with various despeckling filters are tested using low contrast images with higher intensity in-homogeneity. The boundaries traced show that embedding of despeckling filter as replacement for the Gaussian filter in the RD method can be employed in the delineation of CWD images even in the presence of intensity in-homogeneity.