Research On The Cardiac Sequential Image Analysis

With the improvement of living condition and prolonged life expectancy, cardiovascular diseases (CVD) have been the number one cause of death in modem society. The early quantitative diagnosis and accurate evaluation of CVD are critical to improving quality of life and prolong life expectancy.Human heart is a chambered, muscular organ controlled by electrical rhythm that pumps blood into body. Many types of modem medical imaging equipment are capable of cardiovascular imaging. Recently the rapid development of imaging equipments such as magnetic resonance imaging (MRI), computer tomography (CT), ultrasonic (US) and nuclear medicine (NM), has led to dramatically shortening of imaging time and higher spatial and temporal resolution. These make it possible for dynamic 3D cardiovascular imaging.The merits of MRI include good soft tissue contrast, arbitrary imaging plane etc. MRI has natural high contrast between blood and other tissue because of the physical principle of imaging. It is capable of evaluating the anatomy, morphology, quantitative heart function indices, blood flow and viability of regional heart muscle at the same time. In recent years, 3D ultrasonic comes into use for clinical cardiac function evaluation. It is not only able to show the 2D cross section like traditional US, but also make real time volume imaging of heart structure. With the aid of post processing software, it can select arbitrary plane for diagnosis, which enable physician to see as many anatomy structures as possible and measure their properties quantitatively. After volume rending or surface rending, it is more intuitive to observe the structures than traditional 2D US and to some extent release the relying on medical expertise.The development of cardiac imaging equipments leads to enormous image data. Typically, cardiac imaging acquires 3D volume data which change with time during the cardiac cycle, so called 4D data. 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 hardly 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 only take use of some simple geometric models and calculate some global cardiac function indices. Therefore, new computer aided medical image analysis software for objective quantitative analysis and getting clinical significant information for diagnosis is indispensable. For cardiac image analysis the segmentation of anatomy structure such as left ventricle is the bottle-neck in clinical application.Image segmentation is one of the most fundamental and important research topics in the area of image processing and analysis. The researchers have proposed various segmentation methods based on different theory frameworks and image features. The successful implementation of image segmentation technique in medical images is closely related to the anatomy structure it deals with and the clinical environmental itbeing applied. The cardiac images have its unique properties. To achieve successfully segmentation of cardiac images, we must combine segmentation techniques, the specified image features and anatomy structure of interesting together.In this thesis, we propose several algorithms for cardiac MR and US image left ventricle segmentation, respectively. The main contribution of this thesis is list as following.1. A novel generalized fuzzy active contour model is proposed in this thesis. Combined with prior statistic shape model based on level set framework from expert manual outlines, the algorithm is capable of robust segmentation of endocardium borders in cardiac short axis MR images.Segmentation of left ventricle borders from MR dynamic 3D sequential images is the most important prerequisite for computer aided cardiac image analysis and quantitative measurement of heart function. It is because of complicate and non-rigid periodic motion of heart, respiratory motion and limit time resolution of MR that thecardiac MR images typically have low SNR and are full of weak edges. Further more, segmentation is an ill-posed problem, or say, the pose and reflection properties of the object and the noise from the acquisition devices are some of the factors that can interfere with it, there by, the traditional methods based on active contour model(ACM), regional growing often lead to failure in cardiac image segmentation.The introduction of gradient vector flow(GVF) greatly improves the dynamic capture range of curve and enabled the convergence to concave borders. The GVF is estimated from the continuous gradient space. The diffusion process of GVF leads to a measurement that is contextual. This is due to the fact that multiple boundary information contributes to the estimation of the GVF.The accurate image edge information is the key to calculate GVF. Considering the fuzzy and weak edges is common in cardiac MR images, the generalized fuzzy operator (GFO) is incorporated into the diffusion process of GVF in order to handle the weak edge detection. The resulting vector field is incorporating into the evolution equation of ACM. By designing the new external force based on general fuzzy gradient vector flow, the algorithm improves both dynamic capture range and robustness to fuzz3, object borders, and achieves successful application to the segmentation and tracking of LV borders in cardiac MR images. The advantages are the capability of locating the fuzzy object border in strong noisy image like cardiac MR and little human intervention.Further, past research indicated the incorporation of anatomy and human expertise as prior knowledge can improve the efficiency of computation and the robustness to noise and fuzzy edge. We construct a statistic shape model based on level set framework from human expert manual outlines of MR left ventricle. The prior shape model is used to guide the evolution of curve. Experiments on clinical ECG gated 4-D cardiac MR images show that the results are close to the manual outlines of medical experts.2. A wavelet multi-scale curve evolution algorithm based on level set framework is proposed and robust segmentation of cardiac ultrasonic left ventricle is achieved.Due to the existence of attenuation, shadows and speckle noises in ultrasonic images, the segmentation of ultrasonic images is a very difficult task. Traditional image segmentation methods based on Gaussian model often fail in echocardiographic images. Classic methods such as ACM are apt to leak due to the existence of weakand broken edges. In 2001, Tony proposed the model named Active Contour Without Edge (ACWE), which is based on regional homogeneity. ACWE has the merit of robustness and insensitivity to initial position. But the prerequisite of ACWE is the regional gray level distribution of image is Gaussian.Wavelets are a powerful mathematical tool for hierarchically decomposing functions and signals both in frequency and spatial domain. Using wavelets, a 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.Generally speaking, a typical echocardiographic image contains regular region with little change, texture region with similar intensity distribution pattern and sharp edges. Wavelet is successfully applied to model and remove noise in ultrasonic images. After wavelet decomposition, the different kind of information is divided into different scales. For example, the regular regions mainly exist in coarse scale and sharp edge in fine scale. In echocardiographic images, region inside the LV has the similar texture and intensity distribution which can be easily outlined by region based level set algorithm. Sharpe edge in fine scale can be better used by edge based level set evolution. These characteristics have motivated us to combine the region based and edge based level set based active contour model through the coarse to fine scale and use iner-scale constrain to improve the robustness of segmentation.After wavelet decomposition of ultrasonic images, the wavelet coefficient in high level low frequency sub-band is approximately Gaussian. Based on this characteristic, this paper proposes a novel wavelet multi-scale level set algorithm. Firstly, the echocardiographic image is wavelet transformed to get coarse scale approximation images. Then the algorithm begins with the highest level approximation image and outlines the left ventricle endocardium border with regional and edge constrained ACM. Then the result is interpolated into the next finer level of approximation image as a initial contour, and evolved with edge based and inter scales shape constrained ACM. The weighting parameter between inter scale constrain and image force is tuning during the intra-scale curve evolution.One of the important problems in curve evolution based on level set frame is the tuning of weighting parameters of different force. There exist three kinds of fore: regional force, edge force and inter-scale constrain force in our algorithm. How to choose the proper weighting parameters is difficult. We propose a dynamic adjusting scheme for these parameters. In the beginning, the curve should be faith to image information, but at end of evolution, inter-scale constrain should become more important to avoid the edge leakage and maintain the inte4-scale similarity.In sum, the algorithm can be descried by following procedures:a, Wavelet decomposition the original echocardiography images into multi-scale pyramid, use the edge and region compound model to get the initial endocardium contour in the coarsest level;b. Use the segmentation results in A_jf to interpolate the initial contour in finer level A_j-1f then evolve the contour and get the endocardium contour in A_j-1f;c. If reach the finest level (original image), stop the algorithm, elseifjump to step(b);Because of the inter-scale constrain and compound force, our algorithm is robust to local minimums and weak edges and suitable for the US image with high noise, poor contrast and small edge gradient. Comparison is made between the traditional ACM with edges and GACM methods. Experiments on clinical 3D echocardiographic images also show the algorithm's result is very close to the expert manual outlines.3. A new associate scheme for enpi- and endocardium segmentation and temporal tracking in cardiac MR image is proposedAccurate segmentation of endocardium and epicardium is the most important prerequisite for computer aid diagnosis of cardiac function. In this study, we use ^I_z,t(x,y) to donate the cardiac image in time t and slice position z. especially, ^I_z,1 stand for all cardiac images in slice position z of entire cardiac cycle and ^I(1-t) stand for images of all slice position in time frame t. The segmentation and tracking of endo- and epicardium steps is as follows:a. Segmentation of endocardiumLV is approximately the shape of truncated ellipse as illustrated by Fig. 5. In time^t₀, proper D and L are selected to get a rough initial shape of LV. Then the level set functions in all slices are initialized to embed the curves of truncated ellipse. The adapted geometric active contour model proposed in section 2 is then applied to evolve the curve to get the endocardium border in^I_:,to.b. Segmentation of epicardiumSegmentation of epicardium begins with the results of endocardium segmentation. In cardiac MR images, the cardiac muscles appear to have gray levels of Gaussian distribution. By an ad hoc designed regional balloon force controlled by local mean gray level and distance from endocardium, the curves evolve outward to get sensible epicardium borders. At each evolving steps, the local mean gray level ^m_p in a outward neighborhood is calculated and compare to the initial ^me₀ calculated in endocardium neighborhoods. If the ^m_p is different from ^m_e0, then the curve could have evolved out of the epicardium and the evolving speed should become slow. Further more, cording to the anatomy of human heart, the maximum distance of epicardium for endocardium should be around 2.5 cm, once the evolving curves are out of the this distance, the evolution should be stopped and pulled back. According to equation 7, when the curves are out of the cardiac muscle area (mean gray level changes), or the distance is bigger than 2.5 cm, the curves are slowed down or pulled back to ensure the results are robust and compliant with anatomy prior knowledge.c. Temporal propagationsEndocardium segmentation results in ⁽I_{zi, eo} are used as initial curve and propagated to the images in the same slice position z in other time frame and get the endocardium borders of^I_=,: in all time. Because the motion of ventricle border is regular and progressive (Fig.7), the time neighborhood GFGVF fields are weighted summed to get current frame GFGVF. After get the endocardium border in all slice position ^Z_i, the step b is executed to get the epicardium borders.4. Implement the quantitative cardiac function calculation according to the endoand epicardium segmentation from cardiac images.The ultimate goal of computer aided medical image analysis is some clinical significant quantitative cardiac indices. After segmentation of endo- and epicardium from cardiac MR image, 3D model is construct and various quantitative cardiac indices was calculated based on it.