Study On Automatic And Interactive Segmentation Of Left Ventricle Contour In MR Images

Recently, Magnetic resonance imaging (MRI) has become an important assistant measure in the clinical diagnosis of heart diseases. It has been widely used, because it is no invasive and harmful and especially efficient to parenchyma. Through the cardiac MR images, the physicians not only can observe the structure of the heart, but also can estimate the global function and local myocardium function of the ventricles. It can be used to diagnose many heart diseases. Especially, Left ventricle (LV) is the pump of the blood circulation of the whole body. It plays an important role in cardiac function. So LV is the focus in the current research. The parameters of global function and local myocardium function of LV are significant in the clinical diagnosis. In order to estimate these parameters, we need to reconstruct the left ventricle surface and motion. Segmentation of left ventricle MR images is the basis of LV surface and motion reconstruction. Therefore, it is very important to segement LV precisely.MRI is divided into tow category: tagged MRI and untagged MRI. Untagged MRI can provide the high quality images of myocardium, but can not give the myocardium motion information. By applying a special radio-frequency pulse to alter the magnetic property of selective tissues, Tagged MRI creates dark tags on generated images. Since tag stripes are magnetically embedded, they move with the underlying tissue during heart deformation, thus providing myocardium motion information. Tagged MRI has provided a powerful tool to study the lefte ventricle motion. We analyse the characteristic of tagged MR and untagged MR images and propose some novel models and algorithms. Our work mainly includes the following parts:(1) It is proposed that a new method to segment the epicardium and endocardium in left ventricle MR images automatically. Based on the prior shape of epicardium and endocardium which are like circles, it uses Hough transform to detect circles for initial contours of LV. This method modifies the geodesic active contour model by integrating K-means clustering information and anatomical constraints. K-means can provide regional information and anatomical constraints can provide shape constraints for segmentation. Taking the initial contour detected by Hough transform, it can segment both the epicardium and endocardium automatically and accurately. Experimental results demonstrate the effectiveness of this method.(2) It is proposed that a method based on graph cuts active contours and shape statistics to segment left ventricle MR images interactively. This method uses Graph cuts based active contours to convert the image segmentation into the globally optimal partition after we transform the image into a graph. Then, we introduce shape statistics into Graph cuts based active contours. The introduction of shape statistics can prevent the deformation curve from leaking out of actual boundaries. This method has a large capture range and is insensitive to initial contour and occlusion. Only one parameter need to tune manually. Its speed is fast and results are not infected by transformation and rotation and scale. It also provides an interactive modification of segmentation results.(3) It is proposed that an improved ASM method integrateing feature classification for the automatic segmentation of left ventricle. Image classification based on features can provide regional information. The classifaction accuray can be improved by dimensionality reduction. We improve PLS taking spatial label context into account and also use a new way of encoding the class label of CCP and PLS using Gaussian weighting function. We call them CPLS, GCCP and GPLS respectly. These new feature dimensionality reduction methods have better performance. We use these methods to reduce the gray level features of image and to low the classification error as well as consuming time. And KNN classifier was trained with dimensionality reduced features. Instead of sampling the normalized derivative profiles scheme of original ASM, we improve it with a new way that the feature at each position along the profile perpendicular to the object contour is fed into a trained classifier to determine the edge point. The experiments show that the new method can get accurate result robustly.(4) It is proposed that an improved ASM method for automatic segmentation of left ventricle tagged MR images. Based on the idea of feature fusion, we used canonical correlation analysis (CCA) to combine the features extracted from tagged MR images by LM filter bank. Then, a classifier was constructed to determine edge point using SVM. CCA can decrease the classification error and improve the classification performance. Instead of sampling the normalized derivative profiles, the feature at each position along the profile perpendicular to the object contour is fed into a trained classifier to determine the edge point. Experimental results show that our method can achieve a highly accurate and robust performance.(5) It is proposed that an improved texture classification and shape statistics variational approach for the automatic segmentation of the epicardium and endocardium of left ventricle. We introduce texture classification information and shape statistical knowledge into the Mumford-Shah model, and then use the output of support vector machine (SVM) classifier relying on S filter banks to construct a new region-based image energy term. The introduction of shape statistics can improve the segmentation with broken boundaries. Segmentation results demonstrate that our method can achieve a higher segmentation precision and provide a promising way to clinical application.