Research On Left Ventricle Segmentation Algorithm In Cardiac MR Images

In recent years, cardiovascular diseases(CVD) have become the most dangerous killer to human health. With the improvement of living condition and development of modern medical techniques , the early quantitative diagnosis and accurate evaluation of CVD are critical to improving quality of life and prolong life expectancy.Under the control of electrocardio signal, the heart contracts and relaxes periodically in order to pump blood to the whole body to maintain the metabolism of the tissues. With the rapid development of medical imaging techniques, there are the modern medical imaging equipments which are capable of cardiovascular imaging such as magnetic resonance imaging (MRI), computer tomography (CT) and ultrasonic (US) . It is the dramatically shortening of imaging time and rapid improving in spatial and temporal resolution that make these equipments feasible for dynamic 3D cardiovascular imaging.MR is an ideal modality for cardiac imaging for it's good soft tissue contrast, nonradioactive, no need of tracer and arbitrary imaging plane. Additionally, it is feasible to evaluate the velocity and volume of blood flow through cardiac MR because of its good contrast for blood. As a whole, cardiac MR can show the cardiac anatomy, morphology and function, the properties of blood flow and viability of heart muscle.The development of cardiac imaging equipments leads to enormous image data. Typically, cardiac imaging acquires 4D data which include 3D volume data changed with time during the cardiac cycle. Traditional method of 2D diagnosis by clinical experts' subjective observation is problematic. Firstly, it is hard to show 4D data in 2D form, the traditional way of observing projection image or tomography imaging can hardly pick up any clinical significant information from 4D data. Secondly, it is hard to get quantitative information merely by subjective analysis because manual delineation of chamber boundaries is time-consuming and prone to intra observer and inter observer variability. These facts limit the full exploitation of the imaging ability. Although some of the imaging equipments provide some simple computer aided image analysis software, they work based on some simple geometric models to calculate some global cardiac function indices. Therefore, new computer aided medical image analysis software aimed at getting clinical significant information for objective quantitative analysis is indispensable for diagnosis of cardiac diseases. It is the segmentation of anatomy structure such as left ventricle that has become the bottle-neck of cardiac image analysis which limit it's clinical application.Image segmentation is one of the most fundamental,the most important and the most widespread research topics in the research areas of image processing and computer vision. Many segmentation methods which are based on different theory frameworks and image features have been proposed in recent years. The successful implementation of image segmentation technique in medical images is closely related to the objects which it deals with and the practical environment. Because the cardiac images have unique properties, it is necessary to combine segmentation techniques, the specified image features and anatomy structure of interesting together to achieve successful segmentation.In this thesis, we have proposed several algorithms for segmentation of left ventricle in cardiac MR images. The main contributions of this thesis are listed as following.一. A novel dynamic directional gradient vector flow active contour model in thewavelet multiscale framework is proposed in this thesis. The proposed algorithmis capable of delineating the endocardium borders in cardiac short axis MR imagesrobustly.The classical active contour model is a powerful tool for image segmentation. However, when delieating left ventricle contour in cardiac MR images with classical active contour, the interference of right ventricle blood pool and the tissues outside the epicardium often lead to leakage of the contour which evolve solely based on gradient information. In 2006, Chen Jierong et al. proposed dynamic directional gradient vector active contour model which could seek edge with different direction and force the contour evolve into deeply concave region with low sensitivity to initial position of the contour, but the prerequisite gaussian smoothing would blur the edge and produce erroneous results.Wavelets are a powerful mathematical tool for hierarchically decomposing signals both in frequency and spatial domain. The wavelets multiscale analysis is a milestone in signal analysis and is applied widly in the reseach area of signal analysis,image analysis,pattern recognitiona and computer vision. Because of the smoothing and downsampling of wavelet basis, the adjacent low frequency coefficients in higher scale have weak correlations which is similar to gaussian distribution. Additionally, in higher scale of wavelet decomposition, the gray information which include the main energy of the image are preserved and a lot of noise are suppressed. Thus, it is feasible to apply the same segmentation model in the low frequency images on different scales and restrain the evolution of the contour on high scales with the result on lower scales in order to get accurate and robust segmentation.In order to get robust segmentation of left ventricle, the rude contour is delineated on the higher scale with DDGVF active contour firstly. On the current scale, the result of higher scale is used as initial position and the evolution of the contour is driven by image information. The process is repeated until the last scale is reached and the final segmentation result is left ventricle contour.signal can be described in terms of a coarse approximation, plus details that range from broad to narrow. Regardless of whether the function of interest is an image, a curve, or a surface, wavelets provide an elegant technique for representing the levels of detail present. Wavelet theory uses a two-dimensional expansion set to characterize and give a time-frequency localization of a one-dimensional signal. Since this is a linear system, the signal can be reconstructed by a weighted sum of the basis functions. In contrast to the one-dimensional Fourier basis localized in only frequency, the wavelet basis is two-dimensional - localized in both frequency and time. A signal's energy, therefore, is usually well represented by just a few wavelet expansion coefficients. The algorithm can be descried by following procedures:(一) The cardiac MR images are transformed using dyadic wavelet whose wavelet basis is the first derivative of gaussian function to obtain the low frequency images on each scale. From the highest scale to the lowest scale, the modulus and angle on each low frequency image are calculated to obtain the maximum of modulus which is used to calculate DDGVF;(二) On the higher scale, the contour of left ventricle is delineated with corresponding DDGVF active contour model;(三) On the current scale, the result of previous scale is used as initial contour andthe DDGVF is also calculated with the modulus maximum;(四) The process is repeated until the last scale is reached and the final result is left ventricle.Because the edge information on each scale are utilized together, the active contour can avoid leakage and has low sensitivity to the initial position. Encouraging experimental results are provided using real cardiac MR data.二. A novel algorithm based on geometric active contour model which integrates the edge preserved adaptive anisotropic diffusion filtering and K means clustering is proposed in this paper for robust segmentation of left ventricle in cardiac MR images.Because of the non-rigid rapid motion of heart and the physical principle of fast imaging sequence, the boundary among myocardium and ventricles is often blured. In addition, the cardiac MR images contain not only rich texture information, but also much noise and artifacts. In the process of delineating the contour of left ventricle, the noise and irrelevant texture will have great effect upon the final segmentation. The accuracy and robustness could not be guaranteed without appropriate preprocessing.The classical preprocessing is smoothing based on gaussian kernal. Through the gaussian smoothing would remove a portion of noise, it also would blur the edge of the object, which would result in the loss of edge information, moreover, the gaussian smoothing has no effect on irrelevant texture. The edge preserved adaptive anisotropic diffusion filtering based upon local discontinuity and texture discontinuity can not only remove much noise,artifacts and unrelated texture, but also perserve the edge information. In cardiac MR images, the pixels in the ventricle blood pool and myocardium have low local discontinuity and texture discontinuity for the similarity in gray level, thus the homogeneity of these region will be improved after smoothing. On the interface of myocardium and ventricle, the local discontinuity and texture discontinuity are high, therefore the discontinuity is perserved after smoothing. The pixels produced by noise and unrelated texture with high local discontinuity and low texture discontinuity will be removed after the smoothing.K means algorithm is a dynamic clustering method based on global region information. Through the K means clustering, the homogeneity in the ventricle blood pool and myocardium and the contrast between them will be improved. The endocardium is delineated with a geometric active contour model which integrates the boundary information and global region intensity information. In this model, deviation of the contour with signed distance function is used as internal energy so that the reinitialization in the classical level set method become needless and the rate of convergence and robustness will be increased. The proposed algorithm has been applied to real images with promising results.三. An novel approach integrating global region intensity, boundary and prior thatused for associate segmentation of endocardium and epicardium is proposed.In order to calculate the left ventricle mass,ejection fraction and stroke volume, the endocarium and epicardium should all be delineated and the inhomogeneity of gray level,ventricle blood pool and the loss and incomplete boundary information will interfere with the segmentation. A variational framework integrating visual information with prior knowledge is proposed so as to delineated the endocardium and epicardium robustly. The framework can be separate to three parts:(一) The boundary driven termThe DDGVF active contour model is a parametric active contour model which can't handle the topology changes of the contour. In this paper, a novel boundary attraction term is designed to embody the merits of DDGVF and geometric active contour model. The endocardium and epicardium can be differentiated by DDGVF and the topology changes can be handled through level set representations. The DDGVF is an dynamic external force field defined on the interface and should be extended to the whole image domain.(二) The region driven termBecause of the noise and complicated edge information, the segmentation methods which primarily rely on boundary information have great dependency of initial conditions and bad robustness, however, the approaches relied on global intensity properties have low sensitivities to initial conditions and good robustness and anti-interference. In this paper, a statistical region component based on bayesian maximum a posterior estimation is proposed on the assumption that the gray level of the pixels in the cardiac MR images agree with gaussian distribution. In order to increase to velocity of convergence, the deviation between the contour and signed distance function is used as an constraint to omit the procedure of reinitiation in the level set methods.(三) The prior knowledgeBecause of the inhomogeneity of radio-frequency field and the interference of noise, the visual information related with considered application can be misleading, physically corrupted and some time incomplete. Therefore, it can lead to leakage of the active contour model without taking into account specific application constraints. In order to deal with these limitations and physically corrupted visual information, the constraints based on anatomical structures is proposed. Firstly, the epicardium is partitioned into four regions and through anatomical knowledge, the region with great probability to contain physically corrupted visual information can be found. Secondly, because the thickness of myocardium can't change much during one cardiac cycle, the maximum euclidean distance between the epicardium and the centroid of left ventricle can't exceed a specific threshold. If the distance is close to the threshold, the evolving velocity of the contour should be slowed down, if it exceed the threshold, the evolution should be changed or stopped.Experiments on real cardiac MR images with excouraging results show that the endocardium and epicardium can be delineated with good anti-interference and robustness, the dependency can be exempted from initial conditions and leakage and intervention of papillary can't be prevented.