Research Of Brain Structure Segmentation Algorithms Of Infant Based On Improved Multi-atlas Fusion

Infancy is a critical period of brain development.Not only the brain is developed rapidly and has strong plasticity,but also the probability of suffering from a variety of encephalopathy is also higher than adult or in other periods.So it is significant to explore early diagnostic methods of infant brain disease.Modern medical imaging technology has become one of the most important methods in clinical diagnosis of brain diseases.Using nuclear magnetic resonance image to segment the brain structure of infants is the basis of the quantitative analysis of infant brain structure.It can also help to make the research on brain development and disease diagnosis.Deep brain structures include corpus callosum,hippocampus,cerebral ventricles,etc.With the development of the deep brain structure,the complex and irregular in gray and shape information compared to the adult,the segmentation is more difficult than the adults.Human hippocampus and corpus callosum are very important region of the brain,which are closely related with many diseases.Taking the hippocampus and corpus callosum as an example,the three-dimensional full automatic segmentation method of infant brain structure is deeply studied under the multi-atlas framework.The major work and contribution of this thesis include:(1)The algorithm model of multi-atlas segmentation based on non-local patch has been studied and implemented.The experimental data has been pre-processed.The algorithm of fusing target information using adaptive manner has been proposed based on the model of non-local patch.The prosed method fuse the information about themselves to the framework of label fusion.Experiments show that the improved algorithm can segment the MR image of the infant brain structure well and can modify imperfect place.(2)According to the low contrast and poor image quality of the infant brain MR image,the multi-atlas segmentation based on sparse representation has been studied and implemented.The brain structure segmentation algorithm fusing LBP feature and priori information of brain structure based on sparse representation has been prosed.The LBP feature has higher ability to discriminate local information of image blocks.The priori information of brain structure has the ability to guide and revise the segmentation of brain structure.Experiments show that the improved algorithm can segment MR image of the infant brain structure well and can improve the segmentation accuracy of local details.(3)According the blurred boundaries of brain structure in the infant brain MR image and the continual development of infant brain structure,the multi-atlas segmentation based on kernel dictionary learning has been prosed.In the process of segmentation,the training data and test data are then carried out to form a high dimensional virtual sample,and then combined with discriminative dictionary learning algorithm to segment brain structure.Qualitative and quantitative experiments show that the improved model can segment MR image of the infant brain structure well and can segment the edge of the brain structure more accurately.