Research On Key Issues Of Medical Image Segmentation And Registration Methods

The rapid development of medical imaging techniques, as well as growing amount of imaging data, has been boosting the research on medical image processing and analysis. Aiming at assisting doctors in disease diagnosis, surgery planning and therapy assessment, much efforts are made in the filed of extracting and analyzing the structural, functional and pathological information containing in medical images. Medical image segmentation and medical image registration are the fundamental tasks in this field. But there are still many challenging problems existing in these two tasks, due to the facts such as the presence of multiple noise and artifact, the variety of tissues, organs and pathology, and inter-subject variance in medical images.In order to make the medical image segmentation and registration more robust and precise, we generally follow two roadmaps in this paper. First, we design and improve task-specific models based on different application backgrounds. Second, we formulate segmentation and registration tasks into optimization frameworks, in which we incorporate local statistical features or structural features using sparse representation and conduct the model optimization under the prior knowledge extracted from target objects. Some efficient algorithms have been proposed, such as improved active coutour model and graph cuts model based on local statistical similarity feature, an improved random walker model for texture segmentation based on LBP feature, a novel deformable model based on distributed discriminative dictionary learning, and a novel graph matching model under hierarchical and sparse framework. The primary work and contributions in the dissertation are as follows:1. To release the leakage issue under weak boundary in medical images, statistical image features from nonparametric estimation are measured with Bhattacharyya metric, which is further embedded into energy function construction in Geodesic Active Contour (GAC) and Graph Cuts (GC) models. The improved GAC and GC model benefit from the energy function based on the aforementioned metric, which introduces a pull-back strength into GAC to prevent from boundary leaking and helps the GC model in accurately estimating distribution from small samples as well as extracting objects in detail. Then the improved methods are applied to the medical image segmentation scenario which implements a bone and meniscus segmentation framework on knee MRI sequence. In the experimental section, quantitative and qualitative comparisons are conducted respectively. Experimental results indicate the increased precision of our method in segmenting medical images such as knee MRI sequences, which are affected by the noise and the partial volume effect.2. Observing that traditional random walker model for image segmantion only considers boundary information, we propose an improved version for texture image segmentation through solving a symmetric, semi-positive-definite system of linear equations equipped with the texture information. In the construction of the equations, we perform the feature extraction based on Local Binary Pattern (LBP) and map the original image into the space where textures are distinguished from each other (called as LBP map). The similarity between the pixels is then constructed by combining the LBP, gradient and geometric feature in a reciprocal fashion. These similarities are formed as the edge weights of the graph, which helps the labels of the seeds to be propagated to the unlabeled regions in the random walker process. Experiments on texture images, synthetic noise images and medical images shows that the proposed segmentation method extends the state-of-art random walker segmentation to texture images successfully and outperforms some other texture segmentation algorithms particularly on multi-label problem.3. To tackle the inherent limitation of active appearance/shape model, which assumes that both shape and appearance statistics of a target object follow Gaussian distributions, we propose an Distributed Discriminative Dictionary (DDD) learning model and integrate it into a deformable model to achieve the automatic3-D prostate segmentation from MR images. The DDD model describes image appearance features in a non-parametric and discriminative fashion and thus guides the segmenting evolution of active appearance model. In particular, three strategies are designed to boost the tissue discriminative power of DDD. First, mRMR feature selection is performed to constrain the dictionary learning in a discriminative feature space. Second, linear discriminant analysis (LDA) is employed to assemble residuals from different dictionaries for optimal separation between prostate and non-prostate tissues. Third, instead of learning the global dictionaries, we design a "divide-and-conquer" learning strategy and learn a set of local dictionaries for the local regions (each with small appearance variations) along prostate boundary, thus achieving better tissue differentiation locally. Besides, since Sparse Shape Composition (SSC) does not assume any parametric model of shape statistics, it can effectively model prostate shape priors, which may not follow a Gaussian distribution. Experiments on3D prostate MR images demonstrate the best performance of our method in terms of both visual and quantitative evaluation.4. The state-of-the art graph matching methods usually have limited correspondence accuracy especially under the situation of large inter-subject variation. In this paper, we present a novel, hierarchical sparse graph matching (HSGM) method to further augment the power of conventional graph matching methods in establishing anatomical correspondences, especially for the cases of large inter-subject variations in medical applications. Specifically, we first propose to measure the pairwise agreement between potential correspondences along a sequence of intensity profiles (called as line patch) which reduces the ambiguity in correspondence matching. By incorporating the line patch with geometric coherence, the robustness of measuring inter-pair agreement increases. We next introduce the concept of sparsity on the fuzziness of correspondences to suppress the distraction from misleading matches, which is very important for achieving the accurate, one-to-one correspondences. Finally, we integrate our graph matching method into a hierarchical correspondence matching framework, where we use multiple models to deal with the large inter-subject anatomical variations and gradually refine the correspondence matching results between the tentatively deformed model images and the underlying subject image. Evaluations on public hand X-ray images indicate that the proposed hierarchical sparse graph matching method yields the best correspondence matching performance in terms of both accuracy and robustness when compared with several conventional graph matching methods.