Research On Multi-atlas Based Medical Image Segmentation Methods And Its Applications

Multi-atlas based label fusion, as a new research method, has attracted wide attentions in the field of pattern recognition and medical image segmentation gradually. It uses the expert prior knowledge, and propagates the atlas labels to the input image through the guidance of atlas images. How to make good use of the prior knowledge, and properly deal with the relationship between the various atlases to obtain the best segmentation results is the main purpose of this thesis.Firstly, we study the current existing multi-atlas based label fusion methods in detail, and propose a graph-based framework for label fusion, after that we integrate all the methods above into this framework, and we think that all the methods can be considered as a problem of combined classifiers. On the other hand, the unified framework can be in favor of the contrast and analysis in experiments.Secondly, we attempt to integrate the sparse representation method into our proposed framework. We discuss the notion of sparse representation, which will be introduced to our weights computing process. By sparsely selecting the over completed non-local information, we observe the impaction on the accuracy of various label fusion methods and we find the sparse method achieves good performance in our experiment.Finally, on the basis of sparse representation, we further put forward a semi-supervised learning method to improve the robustness brought from the sparse method. The sparse method can guarantee a good discrimination, but it can cause the un-smoothness of the final segmentation due to its sparse property. Therefore, we introduce a semi-supervised sparse method, by constructing graphs from atlas images and input image, this method study to learn the label of input image automatically. Our experiments show that semi-supervised method has effectively improved the smoothness of the segmentation and keep the structure information at the same time.