Font Size: a A A

Research On Medical Image Registration And Segmentation Algorithms In Cancer Diagnosis And Treatment

Posted on:2015-02-03Degree:DoctorType:Dissertation
Country:ChinaCandidate:P SuFull Text:PDF
GTID:1228330452965516Subject:Detection Technology and Automation
Abstract/Summary:PDF Full Text Request
At present, more and more attention has been drawn to medical image processing andanalysis which plays an important role in cancer diagnosis and treatment. As the two keytechniques of medical image processing and analysis, medical image registration andsegmentation are the foundation and prerequisite for image fusion, image reconstruction,image visualization, image understanding and target recognition, and have been widelyapplied in the field of diagnosis, radiotherapy planning, surgical navigation, responseevaluation, anatomy teaching and research. However, there are still many problems existingin these two tasks, due to the facts that image registration and segmentation are always ill-posed, plus the complexity and diversity of medical image. Especially, the advent of somenew imaging technologies such as4D CT, CTF(CT Fluoroscopy), multimodal MRI etc.,makes medical image registration and segmentation more challenging. In this dissertation, wefocus on the algorithms of medical image registration and segmentation, and mainly researchon registration and segmentation problems of some new imaging technologies involved incancer diagnosis and treatment.The main work and contributions of this dissertation are listed as follows:(1)Accurate estimation of lung motion is of great significance to design optimalradiotherapy planning. In previous studies, respiratory sensor or X-ray, CT are used toestimate and measure the lung motion. However, limited by methods themselves or imagingmodes, the estimation results are always unsatisfactory. In this paper,4D CT is treated asresearch object and a B-Spline based deformable registration algorithm is proposed. Accurateestimation of lung motion is achieved by registering the different respiratory phase images in4D CT using proposed algorithm. Considering that the respiratory motion is non-uniform andit contains both big deformation and small local deformation, while B-Spline has the propertyof local control, it is suitable to utilize B-Spline based deformation model to describe therespiratory motion. To address registration optimization problem, the finite differentialmethod is used, which takes advantage of local controllability of B-Spline and can acceleratethe registration procedure. Meanwhile, the multiresolution optimization strategy is adopted toimprove the performance of registration. Then, a region growing based segmentationalgorithm is utilized to extract lung field coarsely. Finally, our registration algorithm isevaluated by experiments. The results show that our algorithm is accurate enough to estimate lung motion. As an application, the lung motion of eight patients is estimated using proposedalgorithm and the statistical results are given, too.(2)As a guidance tool for image guided intervention such as percutaneous lung biopsy,CTF can provide near real-time feedback of patients’ anatomy. However, CTF only capturesfour to ten2D CT images and cannot provide sufficient3D anatomical information. Inaddition, the intervention procedure requires frequent CTF scans, which may causeunnecessary radiation exposure to physicians and patients. To better utilize CTF guidance,we first propose a fast CT-CTF deformable registration algorithm, that register the inhalepreprocedural CT onto the intraprocedural CTF for3D lung intervention guidance. BecauseCT-CTF registration is a special registration problem: CT is a3D image, and covers the wholelung, while CTF can be considered as a2.5D image, for which there is a limited number ofslices. So traditional B-Spline based deformable registration algorithms do not fit such a noveltask. To address this problem, we modified the B-Spline deformation model so that tehdeformation in the x-y transverse plane is modeled using2D B-Spline, and the deformationalong z-direction is regularized using general smoothness criterion, the whole deformationfield is still defined in3D. To improve registration accuracy, a respiratory motioncompensation (MC) model is incorporated into the registration. In order to improvecomputation speed and accomplish the registration in few seconds, a partitioning parallelstrategy is adopted. Similarly, the finite differential method and the multiresolutionoptimization is used to improve the performance in registration optimization. Experiments aredesigned to evaluate the performance of proposed registration using simulated data and realdata. Experimental results demonstrate that our algorithm can yield satisfactory registrationresults.(3)Compared with single modality magnetic resonance (MR) images, multimodal MRimages can provide more sufficient tissue information of lesion from different angles and havebeen widely used in diagnosis and treatment of glioblastoma multiforme (GBM). Due to thecomplex imaging characteristics such as large diversity in shapes and appearance combiningwith deformation of surrounding tissues, it is a challenging task to segment GBM frommultimodal MR images. The most existing segmentation methods are not efficient or cannotachieve accurate GBM segmentation. To overcome this problem, we first propose asuperpixel-based GBM segmentation algorithm which combining superpixel approach withspectral clustering and dynamic region merging algorithm. First, after comprehensivelyinvestigated the superpixel generation algorithm, a local k-means clustering with weighteddistance algorithm is proposed to over-segment the multimodal MR images into a number of superpixels which are homogeneous, compact and adhere well to image boundaries. Then,spectral clustering and dynamic region merging algorithm are utilized separately to processsuperpixles, so that the different GBM tissues including necrosis, enhanced tumor and edemaare extracted. Qualitative and quantitative experiments based on image data collected from15GBM patients are carried out. Experiment results demonstrate better performance of theproposed algorithm by comparing with some classic segmentation methods such as fuzzy C-means clustering(FCM) based method and normalized cut (Ncut).
Keywords/Search Tags:medical image registration, medical image segmentation, respiratory motionestimation, respiratory motion compensation, superpixel
PDF Full Text Request
Related items