Research On Key Problems Of Medecail Image Registration

Medical image registration is a valuable tool for the diagnosis and treatment of disease. It has various clinical applications such as multi-modal brain image registration for diagnosis of disease and 4-dimensional (4D) computed tomography (CT) lung registration for tumor radiotherapy. In multi-modal image registration, there is no functional relation between the intensity mappings of aligned images since physical mechanism of different imaging modalities are different. In mono-modal image registration, non-uniform magnetic field and non-uniform noise leads to local intensity variations. Local intensity variations also lead to no functional relation between the intensity mappings of aligned images. This no functional relation hinders improving the accuracy of image registration, and becomes one of the key problems in medical image registration. This thesis proposes several image descriptors to reduce local intensity variations for multi-modal brain images and 4D CT scans.Lung cancer is the most common cause of cancer-related death all over the world. Image-guided radiation therapy (IGRT) plays an important role in treatment of lung cancer.4D CT lung image registration is the key step in localization of tumor target position. Combing image registration with radiotherapy techniques help to reduce radiation in normal tissue. Howerver, it is difficult to improve the accuracy of registration of lung 4D CT images due to local intensity variation, large displacement of lungs and discontinuity of lung motion caused by heartbeat. This thesis proposes image descriptors, coarse-to-fine strategy and bilateral filters for registration of lung 4D CT images. The proposed image descriptors copes well with local intensity variation. A coarse-to-fine strategy is adopted to deal with large motion. A bilateral filter is adopted to deal with discontinuity of lung motion. Combining these strategies with Lucus-Kanade method, we propose two high accurate image registration methods for lung 4D CT images.The main work can be summarized as follows:(1) Edge neighborhood descriptor for multi-modal medical image registrationLocal patch-based entropy images, manifold learning based descriptor and modality independent descriptor were recently proposed to deal with intensity distortions. However, this method suffers from the high complexity of calculation, large amount of computer memory and low accuracy of image registration. This thesis proposes a new dense descriptor for an image, which generates modality independent structural representation. The proposed descriptor assigns different values to different image regions, namely, image background, image edges, edge neighbourhoods and the remaining area. The proposed descriptor is named edge neighbourhood descriptor since it depends on edges and neighbours. The edge neighbourhood descriptor is 50 times faster than the entropy image. Experimental results show that this new descriptor can improve registration accuracy compared with the results of the entropy image and the mutual information method.(2) Image features-based manifold learning descriptor for multi-modal medical image registrationEntropy image descriptor, modality independent descriptor and edge neighbourhood descriptor employ local relation of local patch, while manifold learning based descriptor employ global relation of local patch. Due to high complexity of space and time, manifold learning based descriptor can hardly be applied to a 3D image. Therefore, it is necessary to research approximate methods of large-scale manifold learning. This thesis applies the classical random projection trees and k-means clusters to large-scale manifold based image descriptor. This thesis proposes two approximate large-scale manifold learning methods. The proposed approximate methods have the same idea, which creates a Laplacian manifold for patches of feature points and their neighbour points and then embeds the rest patches into the manifold. The first approximate method employs Canny detector to select points, while the second method employs Harris corner detection method to select points. The proposed methods are 45 to 51 times faster than the original manifold learning based method and are 3 to 5 times faster than the classical approximate manifold learning methods such as random projection trees and k-means clustering. Experimental results show that the proposed approximate methods can improve registration accuracy compared with the results of the classical approximate manifold learning methods and entropy image descriptor. (3) Hybrid local binary patterns for 4D CT lung registration.(3) Hybrid local binary patterns for 4D CT lung registrationDeformable image registration of lung Computed Tomography (CT) images is a valuable medical tool with many clinical applications. Howerver the accuracy of deformable image registration of lung CT suffers from the difficulties of dealing with high contrast intensity, local intensity variation and large motion. Census cost function has been proposed recently to deal with local intensity variation. Howerver, this method reduce the image contrast, which leads negative effects to image registration. This thesis generalizes one threshold of centre-symmetric local binary pattern to two thresholds, and combines the generalized centre-symmetric local binary pattern with median local binary pattern to generate a Hybrid Local Binary Pattern (HLBP). HLBP can deal with local intensity variation and preserve image contrast for using two thresholds. We propose an Accurate Inverse-consistent Symmetric Optical Flow (AISOF) method based on the classical Lucas-Kanade method. The AISOF applies a coarse-to-fine strategy, an inverse-consistent symmetric method and HLBP to the classical Lucas-Kanade optical flow method. The classical Lucas-Kanade optical flow is simple and efficient. The proposed HLBP can cope well with high contrast intensity and local intensity variation. Since the inverse-consistent symmetric method can reduce the inverse consistency errors in Markov random fields based registration methods, we adopt the inverse-consistent symmetric method to improve the accuracy of registration. A coarse-to-fine strategy is adopted to deal with large motion. Experimental results show that HLBP can improve image accuracy of modality independent descriptor, self-similarity context descriptor and the classical local binary pattern. The proposed AISOF method is evaluated on ten publicly available 4D CT lung datasets from DIR-Lab. The mean target registration error of the AISOF method is 1.16mm. This is 1.67mm less than the result of the classical Lucas-Kanade optical flow method. This result is also the most accurate result of all unmasked registration methods tested on this dataset.(4) Bilateral filters for 4D CT lung registrationThe early work of this thesis proposes a hybrid local binary pattern and AISOF method for 4D CT lung image registration. The hybrid local binary pattern generalizes one threshold of centre-symmetric local binary pattern to two thresholds. Only two thresholds can not deal with large local intensity variation. Therefore, we propose a new local binary pattern, named Multiple Thresholds Local Binary Pattern (MTLBP), which use three thresholds. The heart beat can lead to discontinuity of lung motion at the boundary. Bilateral filters can preserve image edge, which can deal with discontinuity of lung motion. Combing a MTLBP, bilateral filters, the inverse-consistent symmetrical method and the Lucas-Kanade method, we develop a novel accurate image registration method, which is named Bilateral filters based Inverse-consistency Symmetrical Optical Flow (BISOF). The experiments are performed on the publicly available 4D CT lung dataset from DIR-Lab. The mean target registration error of the BISOF method is 1.07mm. This result is 0.83mm less than the result of the bilateral filters based Demons. The computational speed of BISOF is 4 to 5 faster than that of bilateral filters based Demons.