Research On Key Algorithms For Segmentation And Visualization Of Cardiac Tissues

With the improvement of living standard and lengthening of human life expectancy, cardiovascular disease (CVD) has become to be known as the number one killer of human health. Suffered from population aging, gender imbalance, air pollution, and other factors, Chinese are faced with a higher morbidity of CVD. Therefore, it is urgently needed for China to develop early CVD diagnosis and assessment techniques to reduce tremendous pressure on the limited medical resources brought by CVD.Although the mushroom growth of medical imaing brings new hope to CVD disgnosis, it also increases the burden of reading for radiologists in traditional diagnostic methods. It is significant to reduce death ratio caused by CVD and even improve life quality of CVD patients that making researches on key algorithms for segmentation and visualization of cardiac tissues and then quantitatively analyzing cardiac function and morphological changes of human being heart based on medical image processing technologies.In consideration of the research significance as given above, status of current research, and accumulation of current technologies, this thesis focuses on resolving difficulties and meeting challenges in this field. Improved methods are proposed for imaging human heart fast and clearly, eliminating noises and correcting the bias field existing in cardiac images, segmenting cardiac ventricles and analyzing ventricular functions, segmenting coronary artery, detecting and quantitatively analyzing coronary artery stenosis, and visualizing the heart. The target of this thesis is making great efforts on proposing new methods and trying to obtain better results in one or two areas so as to provide technical accumulation for the development of relative fields in China. Hence, the following contents are concerned in this thesis:(1) In order to improve time performance of Cine MR PROPELLER imaging, CUD A based methods are proposed to accelerate artifact correction and uniformly re-sampling algorithms. Experiments show the presence of write-write conflicts in the accelerated data-driven uniformly resampling algorithm. Therefore, a grid-driven uniformly resampling algorithm is proposed and also accelerated by CUDA. Experiments demonstrate that the accelerated algothrims are correct and the speedup ratio reaches 6.5 for artifact correction algorithms and 10 for uniformly resampling algorithm.(2) As equivalent treatment of conventional image denoising algorithms in tiny details and noises, noise elimination might simultaneously distort the struction information of medical images, thereby introcuce new artifacts. A 3D+T Non-Local denoising algorithm is proposed in this thesis and a CUDA accelerated method is proposed to improve its time performance. Experimental results show that the proposed denoising algorithm can ensure weak boundaries and original distribution of image intensities. The results show that algorithm performance accelerated by nearly 150 times.(3) In order to achieve accurate segmentation of the left ventricle and then calculate the exact parameters of left ventricular function, a two-layer level set method is poposed to segment the left ventricle. In this method, endocardium and epicardium are presented by two level contours of one level set function. Distance between these two contours is regularized by a soft constraint as a distance regularization term in the final energy functional. Experiments show that the proposed method does not require registration or training to segment the left ventricle by ensuring its anatomical structure. Comparative analysis demonstrates that the proposed method achieves the highest segmentation accuracy compared with algorithms attended in MICCAI challenge. In order to improve performance of the proposed method, we use CUDA to accelerate it and obtain a 15 times improvement using Tesla 1050C.(4) Taking into account the weak boundaries and intensity inhomogeneity existing in CTA images, a level set method is proposed to segement coronary artery. This method is a hybrid model of DRLSE and RSF, which means that it is a boundary- and region-based level set method. Experimental results show that the proposed method uses advantages of these two models to achieve the purpose of segmenting the coronary tree. A CUDA based method is proposed to accelerate its time performance, which wins a 12-fold speed increase in Tesla 1050C. Taking advantage of the nature of the level set function from the proposed method when coronary artery segmentation is achieved, a centerline extraction method is proposed by considering local minimum of level set function as point in the central path. A coronary artery stenosis dectection and stenoses staging method is proposed by calculating area, the shortest path, and radius of intersection along the central path, then estimating the healthy values of these parameters using regression, and finally computing difference percentage to the healthy values. Experimental results show that the proposed method can be used to detect coronary artery stenosis and achieve the purpose of quantitative analysis of the degree of stenosis.(5) In order to reduce the occupancy of video memory in GPU based texture mapping visualization, a 2D texture mapping method is proposed. The proposed method binds the volumetic data into three 2D texture objects along three axes, which can greatly reduce the required memory space. In process of interactive operations, the proposed method switches texture object according to the gazing direction. This method requires less to hardware and is downward compatibility. Experiments show that the proposed 2D texture mapping method is correct and visualization quality and frames per second rendered by this method both meet clinical needs. In order to visualize the beating human heart, a dynamic texture mapping method is proposed. The proposed method binds all the volumetric data into one texture object and switchs the volumetric data according to the texture object when it renders. Experiments show the proposed method is correct and the reconstruction quality and and frames per second both fundamentally satisfy clinical requirements.In summary, this thesis focuses on following topics which include cardiac imaging, cardiac image pre-processing, cardiac tissue segmentation and function analysis, and cardiac visualization to carry out the study work. Improved medical image processing and analysis methods are proposed and leading research results in cardiac left ventricular segmentation are achieved, which provides technical accumulation for the development of related fileds in China.