Medical Image Registration And Application

Medical image registration is one hot topic in the field of medical image processing. It is widely used in clinical diagnosis and treatment. However, there are still some draw backs within the current registration methods. In this paper, the key problems such as accuracy and speed for image registration were thoroughly researched and some special registration methods were developed for special clinical applications.Mutual information (MI) criterion is widely used in various clinical applications because it is highly data-independent and allows for robust and completely automatic registration without any prior preprocessing steps. A new bilateral image registration method based on combining normalization mutual information with gradient information (BGNMI) was developed in this paper. BGNMI considered geometric relationship between corresponding points in different images and asymmetry of MI criterion. Experimental results on 3D medical images including MR T1/T2 images and PET/CT images indicated the new bilateral similarity was more robust and accurate than conventional mutual information by improving the success rate more than 20%.Real-time registration is very important in medical application, while registration based on MI is a voxel-based method whose main disadvantage is time-consuming. There are two reasons for time-consuming, one reason is each time evaluation for criterion involves in all voxels in images overlap, the other is the number of evaluations is very large. Therefore, in the paper, we speeded up the voxel-based registration method in two processes, one was to reduce time for one evaluation of the MI criterion by using incremental coordinate, the other was to reduce the iteration times by using multi-resolution and a new coarse registration method named principal axis section MI method (PAMI). Experimental results showed the hybrid optimization method was faster than traditional maximum mutual information method.Three special registration methods were developed for different clinical applications in the paper. An automated mask segmentation method was proposed to assist PET & CT image registration. Furthermore, we proposed a hybrid automatic registration method which combined automated mask segmentation method with BNMI and the hybrid optimization method mentioned above.25 patients with PET/CT scans were involved in the study to test this hybrid registration method for its robustness, stability, accuracy and speed. Experimental results showed the registration method was robust and stable whose accuracy was less than one pixel size. It was about 20 seconds for whole-body image registration. This method was also effective for MR and PET image registration. In order to calculate glomerular filtration rate (GFR) based on 3D MR images, precise segmentation of kidney in MR images was needed. We proposed an automatic kidney position and segmentation method by conbining registration and level set segmentation performed at different times. We developed an automatic visual tool based on coherent point drift (CPD) method which could show lesions distribution in the standard mammographic image. We tested this visual tool by using 4 large mass data sets. This tool could assist doctors to analyze lesion distribution conveniently.BNMI method was more robust and accurate than conventional NMI and gradient NMI method for medical rigid registration. The hybrid method was not only faster but also reduced failure rate caused by local extreme value.During the study, concerning the requirement from doctors' clinical diagnosis, we developed different automatic hybrid registration methods for different applications in order to simplify the operation of doctors. These methods were integrated in different application products such as PET&CT registration workstation, UroCARE and mammo lesion distrabution tool.