Medical Image Registration Based On Intensity Similarity Measures

Medical image registration and fusion is an active research hotspot of medical image process and analysis.It plays a key role in clinic diagnosis,radiation therapy, image guided surgical planning and functional neuroanatomy.In monomodal applications,the images acquired from different time are registered to inspect the growth of bone or tumour and evaluate curative effect.Multimodal image registration has applications in treatment planning system,computer assisted neurosurgery navigation and so on.Among all the registration methods,intensity similarity based registration methods are more suitable for clinical application and represent the trend of registration method due to their properties of noninvasive,automatic and accurate. Mutual information(MI)is a popular measure based on intensity similarity.In order to reduce local maximum and misregistration,mutual information is extended to general entropy,finformation and mean divergence measures in this dissertation.The content of this dissertation main includes:1)The registration process of mutual information based method is comprehensively induced,include:the definition and properties of Shannon entropy and mutual information,gray probability distribution estimation,interpolation method and several improved forms of mutual information.In order to reduce local maximum and misregistration of mutual information in medical image registration, three information measures based on generalized entropy instead of the Shannon entropy,named as FRI-alpha,SRI-alpha and GMI-t information measures,are proposed.The convergence width and radius are used for evaluating the measure convergence.The computing time,convergence and accuracy are studied by applying these measures to rigid registration of medical images.The results of tests show that the generalized entropy measures outperform normalized mutual information in convergence performance,without compromising computational speed and registration accuracy. 2)According to the relation of mutual information and f information,several f information measures are introduced,they can also measure the distance between two distributions.Then,two improved I-alpha-information measures,named as FNI-alpha information measure and SNI-alpha information measure,are proposed.The function curves,computing time and convergence are studied by applying the following measures to rigid registration of multimodal images.The measures include V-information,mutual information,I-alpha information,FNI-alpha information and SNI-alpha information.The results of tests show that the FNI-alpha information outperforms mutual information and other measures in convergence performance,and the computational speed of SNI-alpha information is faster than mutual information.3)In order to improve registration speed of mutual information measure,the mean divergence measures are used as the similarity measure of medical image registration.The square root arithmetic mean divergence(SAM),square root geometric mean divergence(SGM),square root harmonic mean divergence(SHM), arithmetic geometric mean divergence(AGM),and arithmetic harmonic mean divergence(AHM)measures are applied to rigid registration of multimodal images. The function curves,registration accuracy,registration time and convergence of these measures are studied in comparison with that of mutual information.The results show that the proposed registration measures have similar function curves and accuracy with mutual information,and the AHM and SAM measures have significant improvements in registration speed.4)A novel image segmentation method based on mean divergence is proposed to automatically determine the number of classes in image segmentation.The AGM and AHM are applied as measures to measure the similarity between the original image and its segmentation result.This method uses AGM and AHM as optimization object and simulated annealing as optimization strategy to find the optimal threshold. Experimental results show that this method solved the problems which fuzzy c-means (FCM)clustering algorithm has,such as determining the number of classes and getting local extremum.In addition,compared with the method based on mutual information,the speed has significant improvement. 5)For registration of multiple images,high dimensional mutual information has large computational cost.A new measure for multiple medical image registration is proposed based on mutual information matrix.The method first calculates the mutual information matrix,and then calculates the entropy of the matrix.When substitute AGM for mutual information,the AGM matrix is obtained.The maximal entropy corresponds to the optimal registration solution.The obtained results show that the proposed method can dramatically decrease registration time with acceptable accuracy.6)Nonrigid image registration is a popular and difficult research field of medical image registration.An automatic method to extract landmark points for elastic medical image registration is proposed.The reference image and floating image are registered rigidly by AHM measure.Then,the two images are divided into uniform sub-blocks.The landmarks in object are extracted by regional rigid registration by maximizing AHM measure between the corresponding sub-images. According to corresponding landmarks,the global elastic registration is achieved by bicubic B-spline functions.Experiments on registration of brain magnetic resonance (MR)images and computed tomography(CT)images show that the proposed method is accurate and fast.