Accurate Segmentation For Cardiac Right Ventricles Based On Multi-Atlas

Nowadays, cardiovascular and cerebrovascular disease increased year by year, topped in the death list, is a great threat to human health. How fast and accurate diagnosis and prevention of them has become the hot issues. Studies have shown that the right ventricle, as important as left ventricle, plays an important role to predict and diagnose heart disease, such as heart failure, ischemic cardiomyopathy, pulmonary hypertension and so on. MRI, with its unique advantages, has become an important clinical diagnosis mean for cardiac image. Meanwhile, segmentation of cardiac MRI is also a hot topic in medical image processing. However, due to its high variability, thin wall, blur boundary and low contrast with the surrounding tissue, right ventricle segmentation is still a difficult problem.In this thesis, some traditional segmentation methods such as Threshold, Region Growing, and Level Set segmentation are implemented to segment right ventricles based on the study of cardiac anatomy on ITK platform. It was found that there are some limits for these methods, for example, region growing could lead to border disclosure easily, while the level set method depends on the initial evolution curve much. Therefore, some new segmentation methods should be considered for right ventricles.Multi-Atlas based segmentation transfers the problem of image segmentation into that of image registration taking full advantage of the prior knowledge of anatomy, and seeks the transformation parameters from the registration of different segmented Atlas intensity images to target image, then applies these parameters on Atlas label images through a procedure called label propagation, obtaining multiple coarse segmentation results, then combines these results together through a label fusion strategy to get the final result. This method has always been a popular way to segment brain tissues and liver tissues, but rarely to right ventricle. While the premise of success for this method is that the Atlas images are similar to target image on some level, so Atlas selection seems quite important to the whole segmentation. A new Atlas selection method called affinity propagation clustering atlas selection is proposed in this thesis, it treats all Atlas images as a series of data nodes and clusters them through massage propagation to obtain different exemplars, then uses these exemplars to participate in the normal atlas segmentation, orders them based on the dice similar coefficient, iterate, gets the final accurate segmentation. In terms of registration, three-step registration are adopted: first rigid registration, then affine registration and B spline registration. Besides, two fusion strategy called STAPLE and Joint Label Fusion are used, where joint label fusion is first used for right ventricle.At last, the right ventricle data form MICCAI held in October2012are used with the above methods, and dice coefficient is selected as the evaluation criteria to analyze the segmentation results. It was found that the result from the proposed affinity propagation clustering Atlas selection combine with STAPLE fusion are more accurate than other methods, and joint label fusion is also better than STAPLE in some case, indicating the effectiveness of the method proposed in this paper.