Research On Level Set And Multi-Atlas Method For Medical Image Segmentation

Medical image segmentation is one of the important researches in medical image processing.To extract an interested organ or focus from an image,the accurate segmentation of medical image is an essential step in computer assisted diagnosis or other medical imaging applications.Since medical images are often the characteristics of low resolutions,weak edges,serious noise and intensity inhomogeneity,achieving accurate segmentation has greater difficulty.At present,there are no universal segmentation algorithms which can perfectly solve medical image segmentation problems.Therefore,researches on medical image segmentation technology have very important theoretical significance and practical value.The most existing image segmentation algorithms use images' characteristics,such as gray-scale,texture,such that the segmentation of the interested area is achieved.Among these algorithms,level set method and multi-atlas method have obtained broad applying prospect on the field of medical image segmentation.This method is called geometric active contour model.Its basic idea is that the evolving curve is implicitly expressed for zero level of higher-dimensional level set functions to better handle curves' topological changes.Region scalable fitting level set model is to use local area information as external energy term which can drive curve motion,and the common gray-scale inhomoaeneity segmentation problem in medical images can be effectively solved.Multi-atlas based segmentation method is different from the above methods,it is a prior knowledge of clinicians guided segmentation method.The core idea is to transfer image segmentation problem into image registration and obtain the final segmentation results by using the prior knowledge of label image.Label fusion is the key step of multi-atlas based segmentation method and can reduce segmented errors which are caused by registration method.Additionally,the effect of label fusion algorithm also directly affects the accuracy of the final segmentation results.Since label images from atlas can include prior anatomical information of multi-objectives,multi-atlas method can solve multi-objective segmentation of medical images.However,the traditional multi-atlas based segmentation method easily occurs the segmented problems of multi-objectives' overlap and incorrect gap when multi-atlas is segmented.The above two segmented algorithms have different their own advantages and disadvantages and are often used to solve medical image segmentation problems of different characteristics.Therefore,the dissertation lays emphasis on the research of the two medical image segmentation methods,i.e.,level set and multi-atlas.Their state-of-the-art and algorithmic principles are deeply explored and the improved algorithms which are respectively based on level set and multi-atlas algorithms are proposed.Finally,by combining of level set and multi-atlas algorithms' advantages.a more robust combined algorithm is also proposed.For the above researches,The innovation points of the dissertation:Since level set based medical image segmentation method needs to select manually initial contour and proceed parameters' settings of level set functions,an automatic initial level set segmentation method is proposed.By using mean shift(MS)clustering algorithm,RSF level set model' initial contour can be automatically obtained and an adaptive window width selection algorithm is also proposed in order to improve mean shift.Finally,selected velocity and accuracy of the initial contour are improved.Meanwhile,clustering results and gray-scale features of images to be segmented based parameter estimation equation is proposed,which can automatically generate parameters of level set function.Additionally,the research also introduces a new initial method to avoid reinitialization.In this way,the evolved velocity and numerical stability of level set functions are enhanced in terms of RSF level set model,and segmentation efficiency and accuracy of algorithm are further improved.Considering overlap and error gap problems of multi-objects medical image segmentation,a new label fusion algorithm is proposed,i.e.,combine voting(CV)algorithm which is used to segment multi-atlas based medical images.CV algorithm can proceed label fusion for all objectives to be segmented by only one computation,while the traditional multi-atlas segmentation method needs respectively to proceed label fusion for each objective to be segmented.Therefore,comparing with the traditional multi-atlas segmentation method,the proposed algorithm markedly reduces computation cost.Meanwhile,the proposed CV based algorithm uses one label fusion and also effectively eliminates objective area overlap and error gap problems by causing the traditional multiple label fusion.Furthermore,each objective's segmentation accuracy is improved.This advantage can make the proposed improved multi-atlas segmentation method use to partial volume correction algorithm of medical images.In this way,manual segmentation method can be replaced and the computation efficiency of partial volume correction is dramatically enhanced.When medical images to be segmented exist significant anatomic specificity,multi-atlas based medical image segmentation method is often difficult to obtain the satisfied segmentation results.In order to solve this problem,level set model based boundary correlation multi-atlas segmentation method is proposed.Before the label fusion step of multi-atlas based segmentation method,this algorithm adds level set model based objective boundary correlation,which makes multi-atlas algorithm not only use prior anatomical information from multi-atlas,but also use gray-scale features from images themselves to be segmented.Segmentation accuracy and robust of this algorithm are improved.Additionally,since this algorithm uses CV algorithm which is proposed by the dissertation as label fusion method,when multi-objects medical image segmentation problems are handled,among multi-objects overlap and error gap problems are also effectively avoided.