Research On Technology For Left Ventricle Computer Aided Segmentation In Cardiac Cine MR Images

According to the new report of the World Health Organization,cardiovascular diseases(CVDs)have become the number one cause of death globally,and its incidence increases year by year.The prevention,diagnosis and treatment of CVDs have become worldwide problems which need to be addressed with urgency.Because Cardiac Magnetic Resonance Imaging possesses some advantages such as the non-invasive property,the excellent contrast between soft tissues,the ability of multi-directional tomography imaging,it has been widely used for the early detection,the risk assessment and the operation guidance of CVDs.Physicians can obtain some important parameters like ventricle volumes,masses and ejection fraction(EF)by segmenting left ventricle(LV)based on CMR images,which are used to evaluate LV systolic function.However,in the current clinical practice,LV myocardium is manually segmented by experienced doctors.This work is tedious,labor-intensive,time-consuming and subjective.Therefore,the research of LV computer-aided segmentation based on CMR images has been a hot research direction in the field of medical image processing.This thesis focuses on the study of the automatic and semi-automatic LV segmentation methods based on the short-axis CMR cine images,which involves three main contents: the automatic extraction of the region of interest(ROI),the automatic segmentation of LV endocardium and the automatic delineation of LV epicardium.We aim to solve the difficulties existed in the LV automatic segmentation and to conquer some disadvantages of the published LV segmentation methods.The main work and innovations of the thesis are as follows:1.In order to reduce the requirements of the computation and space,at the same time,to improve the accuracy of the subsequent LV segmentation,we research the problem of ROI automatic extraction and contribute two new ideas.One is an ROI automatic extraction algorithm based on the time-domain dynamic characteristics of CMR cine images,the other is an LV cavity automatic location algorithm based on the space-domain gray distribution characteristics of CMR cine images.Because heart changes most along time axis and its location will not change too much on different slices,we calculate a rectangular ROI including both LV and RV(right ventricle)by processing the middle slice images.Firstly,we compute the absolute difference between each frame and the mean image along time axis.Then,we process the difference image with some simple procedures such as image accumulation,pixel attenuation,contrast stretching and Gaussian blur to highlight the heart.Finally,the ROI is determined by using Otsu segmentation and morphological operation.It is worth noticing that we define an attenuation factor based on the pixel position to reduce the interference from imaging artifacts and make the ROI extraction robust.After extracting ROI,we detect the LV cavity point based on the 3-Dimension data of end diastolic by applying the image polar transform,Gaussian blur,the min-max normalization and the maxima detection.The automatic ROI extraction algorithm and the automatic LV location method are respectively tested on both the MICCAI 2009 database and the CAP database,and experimental results show that both the two methods are accurate and fast.2.For the automatic segmentation of LV endocardium,the accuracy and stability of the traditional image-driven based methods are low.To solve this problem,we propose a novel method based on the simplified pulse coupled neural network(SPCNN).Firstly,SPCNN model is iteratively applied to produce a series of binary images.Secondly,an optimal LV cavity is detected from all SPCNN outputs based on some prior information including LV size,LV location,the intensity homogeneity of LV myocardium and the obvious gray contrast between cavity and myocardium.Finally,the fine endocardium is delineated by using the convex hull calculation,B-spline smoothing and the maximum gradient searching.For the SPCNN parameter setting,we iteratively increase C in equal increments while setting other parameters as fixed values.For the optimal LV cavity determination,we not only integrate the constraints of LV size and location between slices but also define a comprehensive evaluation index reflecting the LV characteristics.When tested on the MICCAI 2009 database and the CAP database,the proposed approach achieves encouraging results which indicates its effectiveness and competitiveness.3.In order to solve the edge disclosure usually occurred in the conventional Snake models,we put forward an improved Snake model for the LV epicardium segmentation.In the modified Snake model,the gradient vector flow(GVF)force drives curve to move towards boundaries,and the balloon and shape forces control curve to converge at weak edges.The shape force with an iteratively increasing strength is defined based on the shape similarity of endocardium and epicardium.In addition,an edge map modification is employed to avoid the interference from endocardium.The epicardium segmentation based on the modified Snake model is evaluated on MICCAI 2009 database and CAP database,and it performs better than some methods published in recent years.4.In order to improve the precision of the endo-/epi-cardium segmentation,at the same time,to keep a good anatomic structure for LV myocardium,we study the LV segmentation techniques based on statistic shape models and propose an automatic LV segmentation method by combining SPCNN and ASM(Active Shape Model).For the modeling,we divide the 4-Dimension data into nine sets and create their 2D(two-dimensional)statistical models respectively;for the ASM initialization,an algorithm based on SPCNN segmentation and geometric transformation is designed.Experimental results based on MICCAI 2009 database show that our method not only can automatically realize ASM initialization,but also can achieve good performance for both the endocardium and the epicardium segmentation.