Research Of Similarity Measure In Medical Image Registration

Medical image registration is an important branch in the field of computer vision and image processing. It is an essential step for medical image analysis, medical image understanding, clinical diagnosis and clinical treatment. So far, medical image registration has been studied for many years in the word. It is widely used in medical domain. Image registration is the process of finding a spatial transformation (or a set of transformations) to align two (or many) images, which makes the pixels have the same spatial location in two (or many) images.Medical image shows information about patient, such as the shape, size, structure and spatial relationships of tissue or lesion. It is important for clinician to diagnose and treat disease. Generally, medical image acquired at different time but same medical imaging equipment is called single modality image. Medical image acquired at different medical imaging equipment but same time is called multi-modality image. Single modality image registration is mainly used for monitoring tissue growth, subtraction imaging and so on. In general, the technique of single modality image registration is mature to meet the needs of clinical practice for aligning the common medical images. However, the single modality registration of non-uniform medical image is a challenge task in the word because of its violation on the assumption of the pixel-wise independence or stationarity. Here, the intensity distortion is mainly caused by intensity bias field and noise in MR. In addition, single modality medical image can only provide a kind of information about patient. It is concluded that the information provided by single modality image is simplex and incomplete. However, the information provided by multi-modality medical image is complementary. The integration of medical images in different modality can correct the misdiagnosis caused by the incomplete information produced by single modality image. So, clinicians need to fuse several multi-modality images to obtain more information of patient, which can help them make an accurate diagnosis and logical treatment.Medical image registration is a key prerequisite for medical image fusion. Hence, the intensive study on the image registration would be of great clinical interest. In a word, medical image registration is a key step for image contrast, change analysis, object tracking and object recognition. It is the baseline for image process, such as, medical image segmentation, medical image reconstruction. In addition, medical image registration has a wide range of applications in the comparison of medical image with atlas, surgical navigation, estimation of cardiac motion, estimation of tumor parameters, creation of average atlas, surgery location, radiation therapy, and so on.According to the benchmark of the characteristics of registration, medical image registration can be classified into two categories. One is feature-based image registration. Another is intensity-based image registration. In most cases, the complexity of medical image feature makes the extraction difficult, which will result in undesirable registration result. The mage registration based on intensity does not need to extract image feature. It just adopts image intensity to make the accurate registration result easier.Medical image registration based on maxing mutual information is an intensity-based method in recent years. Originating from informatics theory, mutual information describes the similarity between two random variables. It does not need to select marking point, extract image feature, or assume the linear relationship among the gray values in medical images. Hence, it is an effective method for multi-modality medical image registration. However, mutual information is sensitive to noise, changes in overlap of two images and the number of sampling points. It will generate mis-registration results if the image is defect or with low resolution. To tackle this problem, lots of methods have been proposed. Studholme proposed normalized mutual information (NMI), which refers to the ratio of marginal entropy and joint entropy. The marginal entropy is limited by joint entropy. NMI is less sensitive to changes in overlap. However, it ignores the spatial information and orientation information. Pluim proposed to combine image gradient and mutual information. It takes the spatial information into account. In general, it can get good registration result. However, it is sensitive to noise. So, the registration result will be not good if the image contains some noise. Also, computing image gradient is time consuming. Russakoff proposed regional mutual information (RMI), which is an extension to MI. It can restrain noise in a way because of computing MI in the region. Here, we proposed a new similarity measure, phase congruency and regional mutual information (PCRMI), which combines RMI and phase congruency (PC) describing image feature. The significant advantage of this approach is that not only intensity information, but also spatial information is taken into account. We will describe this method in section3.The aforementioned methods are restricted to the assumption that the intensity relationship of the corresponding pixels is independent and stationary. However, images with noise, intensity distortion or intensity bias field don’t satisfy the assumption. These traditional similarity measures do not produce satisfied results. This problem was solved by physical method early. That is, change imaging parameters to suppress the intensity distortion in medical equipment. This method is not universal. Therefore, it could not be widely used in clinic. Today, people always use medical image processing to tackle this problem. Many scholars used some models to simulate the intensity bias field. But the function index affects the accuracy of simulation. The function with high index will result in complex computation. Then, local measure defined on a small region was utilized. The intuitive idea of such approach is that a spatially-varying intensity distortion is constant within a small pixel neighborhood. In general, methods based on local similarity measures performed better than those using global similarity measures. However, such local approaches are much more sensitive to noise and outliers than global measures. Moreover, it is also an expensive computation. Markov-based technique was used to register image with intensity distortion, but this method heavily relies on the definition of local intensity interactions and initial parameter. Khader and Hamz proposed a generalized information-theoretic similarity measure. This method optimized the Jensen-Tsallis (JT) entropic similarity measure using the Quasi-Newton as optimization scheme. Cubic B-splines was used to model the non-rigid deformation field. Then, the analytical gradient of JT measure was derived so that an efficient and accurate image registration can be achieved. However, it must use a similarity measure with suitable optimization technique to improve the image registration. Andriy Myronenko performed residual complexity measure to solve the intensity correction field. The problem of such approach is that it is sensitive to parameter. Relatively minor modification of the parameter can generate dramatically different results. In addition, it is sensitive to noise and outliers. We mainly focus our research on two image registration topics, as follows:1, Medical image registration based on phase congruency and regional mutual informationRegional mutual information allows neighborhood information to be taken into account to avoid noise and local extremum. It is a smooth and accurate similarity measure. That’s because the mutual information is computed in region. Phase congruency, depicting the structure information, is a novel image feature to measure the phase for each pixel in frequency domain. It defines the max phase as feature point. In addition, it is invariant to image brightness and image contrast because the computation of phase congruency is nothing to do with image gradient. Traditional edge detection operator, such as Marr, Sobel and Canny, is computed by the image gradient. Hence, the detection relies heavily on image brightness and image contrast, which will make the result undesirable. However, phase congruency could overcome above problem due to the invariant to image contrast and noise. So a new approach of multi-modality image registration is presented with not only image intensity, but also features describing image structure. Two aspects of the registration are addressed to make this method feasible. Firstly, instead of standard MI (Mutual Information) based on joint intensity histogram, RMI (Regional Mutual Information) is employed, which allows neighborhood information to be taken into account. Secondly, obtain the phase congruency image in different orientation. Finally, by incorporating these feature images and original images into RMI to computer the new similarity PCRMI. Then, we can combine aspects of both structural and intensity information together, which offers a more robust and a high level of registration accuracy. The experiment demonstrated that our method is smooth and robust.2, Intensity based image registration by exponential function weighted residual complexityAs we all know, Magnetic Resonance Image (MRI), a non-invasive technique, do not use the harmful substance, such as X ray and contrast agent, to collect clinical data of patient. It describes the information of soft tissue with high resolution. In clinic, physician always use it to diagnose and treat kinds of diseases. MR image often contains noise, motion artifacts and intensity non-uniformity (INU) caused by bias field (BF) generated by inhomogeneous magnetic field. The intensity non-uniformity refers to the slow change of image intensity for the same tissue, which makes the registration of images with intensity distortion difficult. In addition, compared to pre-contrast enhanced image, post-contrast enhanced image of the same patient could be supposed to possess the property of intensity non-uniformity.In this paper, we propose a novel intensity-based similarity measure for medical image registration. Traditional intensity-based methods do not consider the relationships of pixels in medical images. It is sensitive to intensity distortions, contrast agent and noise. Although residual complexity can solve this problem in certain situations, relative modification of the parameter can generate dramatically different results. The new similarity measure, Exponential function Weighted Residual Complexity (EWRC) utilizes local variance of the reference image to model the exponential weighting function. It assigns low weight to the pixel with large value in residual image, and assigns high weight to the pixel with small value in residual image. Then, the residual complex of the constrained residual image is computed as the new similarity measure. EWRC considers the relationships of pixels in medical images, and incorporates the image spatial information into registration process to make the results much smoother and better. The proposed technique was applied to simulated data and patient data. The experimental results clearly indicated that the proposed approach has achieved more accurate and robust performance to bias field, noise and parameters.