ANALYSIS OF MITRAL REGURGITATION USING ECHOCARDIOGRAPHIC IMAGES

Abstract

The valvular heart disease (VHD) is one of the major cause of morbidity and mortality throughout the world. It affects any of the four valves of the heart (the aortic and mitral valves on the left and the pulmonary and tricuspid valves on the right). The valvular abnormality can be of two types (i) regurgitation and (ii) stenosis. A thorough understanding of the valvular abnormalities are very important to aid in the management of patients with valvular abnormality. Valvular regurgitation is defined as the backward or retrograde flow of blood from the valves into the cardiac chambers when the leaflets do not close completely. This backward flow is referred as “regurgitant flow”. In valvular stenosis, the valvular leaflet becomes stiffer which narrows the valve opening and reduces the blood flow through it. Each of the four valves of the heart may exhibit these abnormalities. The presented research work carried out with an aim to enhance the diagnostic potential of conventional B-mode ultrasound imaging modality for the diagnosis and severity analysis of mitral regurgitation (MR). The MR is the most common valvular disorder in modern clinical practice. It is known as the reverse blood flow from the left ventricle (LV) into the left atrium (LA) during the systole process. The ultrasound of the heart is known as echocardiography and is most commonly used in the assessment of cardiac chamber and valvular abnormalities. The aetiologies and the consequences of valvular abnormalities are diagnosed using transthoracic echocardiographic (TTE) images acquired in apical two chamber (A2C), apical four chamber (A4C) and parasternal long axis (PLAX) views. The ultrasound imaging modalities such as conventional B-Mode (brightness mode), M-Mode (motion mode), continuous wave Doppler (CWD), and color Doppler echocardiography are used hand-in-hand to detect the prevalence of regurgitation, a better understanding of the mechanism of regurgitation and quantification of severity along with its repercussions. A fine-grained textural pattern known as speckle noise is observed in B-mode TTE images which reduce the contrast resolution and masks the texture details. It often makes quantitative measurements and the automatic analysis of ultrasound images difficult. The contemporary research demonstrate the importance and superiority of despeckling techniques. Individual techniques have their merits and demerits. The performance of the despeckling filter is i measured utilizing different performance parameters and visual quality parameters. The automatic estimation of cardiac structures and shape and size become very difficult because of the speckle noise. The speckle noise also affects location of edge features, therefore it is very important to preserve the edges in medical images while doing despeckling. In order to address this issue, as a first objective of this work, a comparative study of despeckling filters of five categories (i) local adaptive, (ii) synthetic aperture radar (SAR), (iii) anisotropic diffusion, (iv) non-local mean, and (v) fuzzy filters have been implemented on test images and TTE images. The despeckling capabilities of these filters have been evaluated in terms of traditional image quality metrics as well as blind image quality metrics. A hybrid homomorphic fuzzy (HHF) filter, combining the advantages of NLM and fuzzy filters, has been proposed here. The denoising performance parameters of the HHF filter are compared with despeckling techniques in the homomorphic and non-homomorphic domain. The HFF filter performed better in terms of edge preservation compared to other fuzzy filters. The commonly used methods for the quantification of MR still has many limitations, such as the uncertainties in orifice location, multiple jets, and a hemispheric convergence assumption that often results in over or underestimation of flow rate and regurgitation orifice area. These techniques are operator-dependent and often significant training is required to acquire good quality and correct data. It is difficult to analyze manually the acquired data due to the poor quality of echocardiographic images during the diagnosis at various stages. It is quite challenging to derive the necessary information from the acquired data by the technicians. Also, the manual analysis is a subjective methodology; it compromises on the accuracy of diagnosis and severity estimation. It is very difficult to reproduce quantitative measurements, manually. To overcome the issues associated with the severity analysis of MR, machine vision technology has been employed. Keeping the above facts in view, the second objective of this work has been planned as to develop the texture feature extraction techniques to acquire substantial information from echocardiographic images and to improve the classification accuracy of the computer-aided diagnosis (CAD) system for the severity analysis of MR. In this approach, two steps have been involved. The first step has been to extract the features of the echocardiographic images and in the second step the relevant features among the above extracted features on the basis ii of certain criteria have been considered as input to the various classifier techniques. An optimized classifier has been selected on the basis of experimentation. Here, the texture analysis of the MR images has been accomplished through proposed three approaches. In first approach, eight texture features, (i) first order statistics (FOS), (ii) spatial gray level dependence matrices (SGLDM), (iii) gray level difference statistics (GLDS), (iv) neighborhood gray tone difference matrix (NGTDM), (v)statistical feature matrix (SFM), (vi) Laws textures energy measure (Laws TEM), (vii) fractal dimension texture analysis (FDTA), and (viii) Fourier power spectrum (FPS) have been extracted from the MR image database in three views with four color spaces. The minimum Redundancy Maximum Relevance (mRMR) feature selection techniques have been chosen to eliminate the less relevant features. Finally, two supervised classifiers support vector machine (SVM) with three kernel (linear, polynomial, radial basis function (RBF)) and random forest (RF) has been used along with 10-fold and leave-one-out cross validation technique to reduce biasness. The classification accuracy of the proposed CAD system in red green blue (RGB) color space has been found slightly better than gray-scale color space, however the computational time of the scheme was thrice compared to gray-scale model. Subsequently, the performance of the CAD system has been evaluated in terms of the classification accuracy, severity, and specificity. In the second approach, the multiresolution based texture feature extraction techniques have been utilized in proposed CAD system. The images are decomposed at several levels of different resolution, where each of the sub-images contain varied and valuable information about the original image. In order to enrich the quality of the texture feature extraction technique the Gaussian pyramid has been used because of less computational requirement. Moreover, to extract the texture features from the decomposed images, variants of the local binary pattern (LBP) such as uniform local binary pattern (LBPu2), rotation invariant local binary pattern (LBPri), rotation invariant uniform local binary pattern (LBPriu2), center-symmetric local binary pattern (CSLBP), local binary pattern histogram Fourier features (LBP-HF), and completed local binary pattern (CLBP) have been applied. Furthermore, the performance of the extracted features has been evaluated using SVM and random forest (RF) classifier with 10-fold cross validation. The Gaussian pyramid based completed local binary pattern (GPiii CLBP) technique was found to be produced the best classification accuracy amongst the all proposed features. In the third approach, the discriminatory capability of Daubechies wavelet-based texture modeling has been assessed for the severity analysis of MR. The transform domain techniques have been opted due to their multiresolution capability for analyzing images at different frequencies of several levels of resolutions. The different frequency sub-band images provide substantial information about the various objects of the images compared to the information obtained in spatial domain grayscale images. In the present work, discrete wavelet transform (DWT) technique has been utilized to get transformed domain features as it has the property to emphasize the directional information of the images. The Daubechies wavelet family has been utilized for the image decomposition because of its approximate shift invariance property. This multiresolution based texture feature extraction techniques produces a large number of complex features and many of the features may not be significant. Keeping this aspect in mind, PCA has been used as feature reduction technique in order to reduce the dimension of feature vector. Additionally, the mRMR has been chosen as feature selection techniques to eliminate the less relevant features. At the end, both the utilized techniques, improve the classification performance of the CAD system and reduce the computational time. The db4 offered best characteristics among the Daubechies wavelet family considered for precise severity investigation of the MR images. The significant contribution of this thesis work can be summed up as the development of a CAD system for severity analysis of mitral regurgitation using echocardiographic images. The CAD system developed for the severity analysis of MR is capable to classify the different categories of the MRstages with reasonable accuracy. The achieved classification accuracy of the proposed CAD systems helped to enhance the productivity of clinicians while supporting them with some useful information.