Research On Three-dimensional Hippocampus Segmentation Algorithm Of Brain MRI

The hippocampus(HC) is an important part of brain limbic nervous system, while which plays an important role in learning and memory, especially in spatial localization. MRI images are significant data to study the HC morphology, which provide three-dimensional brain information with rich contrast and high-resolution. Therefore, for mental neurological diseases such as Alzheimer’s syndrome diagnosis, in the study of characterizing the HC volume and morphology, accurately segmentation HC of brain MRI has important medical implications.Atlas-based registration methods make the use of expert annotations in the form of prior information to achieve automatic segmentation. However, due to the small size and complex anatomy of the HC, single-atlas cannot be well adapted to the complexity of the differences between individuals, prone to error segmentation. Thus, in this paper we focus on 3D hippocampus automatic segmentation method based on multi-atlas registration, a two-stage local registration method and an improved evolution method based on point distribution model(PDM) are proposed and its main ideas are as follows:Firstly, we focus on 3D HC segmentation method based on multi-atlas registration. Analysis commonly used methods of registration and the label fusion, which are two key issues for multi-atlas segmentation method. Particularly expounds the multi-atlas segmentation method based on point distribution model.Secondly, a two-stage multi-atlas local registration method based on the standard template is developed. In the 3D segmentation of HC, global linear registration method is commonly used in registration process, but due to the small size and irregular shape of HC, the shape and position differences among the HCs are still large after registration. We use a two-stage local registration method to gain a more accurate corresponds. First, all images are global linearly registered to the standard template with affine transformation; Secondary, local registration with a HC mask. The multiple initial segmentation results are gained form label deformation. Lastly, Resulted label is gained from label fusion based on 3D point distribution model. This two-stage local registration method improves the accuracy of result.Thirdly, For the HC segmentation algorithm based on 3D point distribution model, except for two-stage local registration, an improved local neighborhood point spread evolution method is proposed. Due to the point by point evolution method cannot meet the requirements of surface smoothness, prone topology abnormal deformation and easy to fall into local minimum. In this paper, a local neighborhood point spread evolution is proposed, which not only effectively improves the continuity of the shape surface, but also improves the evolution efficiency.Finally,choose 50 T1 MR images as experimental data. Do experiments with improved algorithm and unimproved algorithm and compare the results with expert manual segmentation(commonly known as the gold standard) for quantitative analysis. Experimental results show that the proposed segmentation approach improve the results.