Research On Supervoxel Generation Algorithm For Magnetic Resonance Images Of Brain

Superpixel or supervoxel is a kind of image preprocessing technology developed rapidly in recent years.This technique generates a sub-region that is consistent with the similar pixels or voxels in the local area of the image to maintain a certain local feature of the image.Compared with the basic unit pixel or voxel of traditional image processing,the superpixel or supervoxel has the following advantages:(1)The image redundancy information is acquired by grouping the pixels,which makes the size of the processing object greatly reduced,so that we can reduce the complexity of subsequent image processing operations;(2)Making the extraction of local features of the image more convenient,while more conducive to the expression of the image structure information.Brain magnetic image processing and analysis are a hot topic in the field of medical image processing.Magnetic resonance imageing(MRI)has the characteristics of clear image and multi-angle,and magnetic resonance images also contain rich human body soft tissue contrast information,so in medical imaging,especially on the brain image analysis,MRI is increasingly showing its superiority.And has been widely used in cognitive research,disease diagnosis and treatment and other fields.The supervoxel technique is to aggregate voxels with highly redundant properties into meaningful super voxels with uniform regions.Super voxels can provide local image features that can be applied to brain MRI images to reduce the complexity of subsequent processing and analysis tasks.However,due to the complex internal structure of the human brain and the partial volume effect of the brain MRI,there are still many limitations in the application of the existing supervoxel method to the brain MRI images.In order to solve this problem,this thesis presents a new iterative space fuzzy clustering supervoxel generation algorithm based on prior knowledge for 3D brain MRI images.First,since the human brain has the same topology,a set of seed templates can be obtained from the population-based brain MRI template and then projected onto the individual space to generate the initial seed.Secondly,in order to eliminate the influence of partial volume effect,we propose an efficient iterative spatial fuzzy clustering algorithm,which classifies voxels into each seed to generate supervoxels of brain MRI images.This thesis employs the widely utilized BrainWebl8 and IBSR18 datasets for assessments.The results are firstly visually evaluated,and then quantitatively assessed using three popular measurements,including the under-segmentation error,boundary recall and achievable segmentation accuracy.The algorithm is compared with three classical algorithms,including simple linear iterative clustering,graph-based,and regularity preserved supervoxel.Both the quantative and quantitative experiments demonstrate the effectiveness of the proposed algorithm.