Research On Optimized Multi-Atlas Medical Image Segmentation

The hippocampus,located at the edge of the medial temporal lobe beneath the cerebral cortex,is an important part of human and other vertebrate brains.Its function is related to short-term memory,cognition and emotion,so studying the hippocampus is of great importance to psychology and neurology.Nuclear magnetic resonance images can avoid such hazards as X-rays and ionizing radiation,have safety and high resolution characteristics,and the brain's MRI can provide three-dimensional brain structure information to facilitate the study of complex tissue structures in the brain similar to the hippocampus.Therefore,this paper studies how to separate the hippocampus from brain tissue in MR images,which provides a visual basis for medical research on the volume morphological changes of hippocampus.For medical image segmentation,the complexity of its processing is caused by the particularity of its image type.On the other hand,because the tissues in the brain have irregular shape,small volume,the edge is connected to other surrounding tissues,the gray level is similar to each other,making the segmentation more difficult than the general image,and the traditional segmentation method is difficult to obtain satisfactory results.The segmentation method based on Atlas registration is a hotspot in recent years,and it has a good effect in separating the specific tissues within brain medical image,the principle of which is:using expert pre-segmented atlas to register with target image,and then guiding the target segmentation through deformation field mapping.Among them,the segmentation method based on single atlas can easily lead to serious segmentation error for the large difference between the atlas and the target image.The multiple atlas segmentation method can effectively avoid the above error by synthetically considering the priori information of multiple atlas.In order to achieve the accurate and efficient segmentation of the hippocampus in three-dimensional human brain MRI images,this paper improves the algorithms of registration and label fusion in the multi-atlas segmentation,proposes an improved registration method based on resampling,and realized the goal of dividing the hippocampus with the optimized label fusion algorithms.The experimental results show that the improved algorithms can shorten the segmentation time under the precondition of guaranteeing the segmentation precision.At present,the multi-atlas segmentation of three-dimensional human brain MR images has two difficulties:the first is the improvement of precision,the second is the reduction of computational complexity.In this paper,the necessary pre-processing for the target image and atlas is done,such as skull culling,extraction of interest region and gray-level pretreatment.In the phase of image registration,an improved registration scheme based on resampling is proposed in this paper.Unlike the traditional "coarse-fine" hybrid registration,resampling can simplify the "coarse" registration process,make the image to be sampled have the same isotropic sampling rate as the reference image,and make them have the consistent image size and center.Therefore,resampling makes the atlas to be prepared for the Diffeomorphic Demons registration.And then the combination of diffeomorphic demons algorithm and multiresolution registration ensures the accuracy of registration,and improves the registration speed.In the label fusion stage,this paper tries to improve two fusion algorithms:the first is the weighted selection label fusion algorithm based on K nearest neighbor search,the second is a discriminative model-constrained graph cuts approach.The weighted selection fusion algorithm based on K nearest neighbor searching method is based on the algorithm of nonlocal patch weighted fusion.K-nearest neighbor is used to find the K-patches which is the most similar to the target,to avoid weighting the local patches with lower similarity in the neighborhood of the voxel.This improved algorithm increased the computational speed and improved the accuracy of weight calculation.In many cases,segmentation is regarded as the process of classifying objects in a scene,so this paper studied another improved algorithm which is discriminative model-constrained graph cuts.This algorithm combined machine learning method with graph cut algorithm,so that the segmentation result is optimized when the prior knowledge of the atlas is effectively utilized.