Research On Object Segmentation In Cardiac CT Images By Using Random Forests

Cardiovascular diseases has seriously threaten human life and health. Early diagnosis and quantitative assessment of the risk of cardiovascular disease plays an important role in prolonging human life and health. With the rapid development of science and technology, functionality and image quality of diagnostic imaging equipment has been greatly improved. The rapid development of the computed tomography has great influence on the diagnostic mode of human disease and has gradually become an important way of cardiac diagnosis. The left ventricular is an important content in studies of heart diseases because it is the core region of heart. Therefore, it will be very meaningful to conduct research on the left ventricle of heart with the help of cardiac CT images.Random forests is regarded as a flexible frame model, and shows a good prospect in the field of medical image processing. The main work of the paper is based on the random forests model. Firstly, we study the location problem of anatomical landmark in cardiac CT images. Then, we make a further research on image segmentation of left ventricular myocardium. First of all, random forests regression model and classification model are jointed to detect and localize three anatomic landmarks in cardiac CT images which contain aortic valve center, mitral valve center and left ventricle apex. In the process of detection of the three landmarks, random forests regression and classification model which has combined with 3D Haar-like features can produce the distance maps and probability density maps of the anatomical points, respectively. Exploiting mean-shift algorithm unites distance maps with probability density maps, that make prediction for locations of anatomical points. Then, in the process of segmentation of the myocardium, the approximate region of left ventricular can be obtained by locations of three landmarks. More importantly, we can build the heart coordinate system based on the locations of the three landmarks. Then, cardiac CT images are transformed to a similar coordinate space. On this basis, we extract the image grayscale characteristics and geometrical characteristics training random forests classifier which perform classification for transformed voxels within specific region, segmenting the myocardial from CT images. In the experiment of myocardial segmentation, the paper uses two different way to get positon of landmarks, respectively are manual annotation and automatic detection by proposed localization algorithm. Thus, we can get two different myocardial segmentation algorithms, and the results of two segmentation methods are compared by us. Macroscopically speaking, our approach is based on supervised machine learning for location and segmentation of anatomical structures in cardiac CT images. By learning the relationship between anatomical structure and image information, we can build the random forest model, then uses it to analyze the unknown image.