Adaptive Medical Image Registration Based On Joint Salient Region

Image registration plays an important role in many applications in medical imaging, computer vision, pattern recognition and remote sensing. In this thesis, based on joint salient region detection and some robust computation methods, adaptive rigid and non-rigid medical image registration is proposed to address the problem of the gross outliers and local large deformation resulted from space occupying tumor-like gross outlier in multi-modality and multi-temporal medical images.One contributions of this thesis is rigid registration of medical images with gross outlier using joint saliency map. First, based on local energy function in Markov random field signifying the spatial interaction between the local neighboring voxels, a scale-invariant saliency operator is used to construct saliency map for two images. Second, the saliency maps for two images to be registered are compared to construct the normalized joint saliency map. Each normalized joint saliency value at every voxel of joint saliency map indicates the extent to which the underlying two overlapped voxels are mostly from corresponding salient regions of floating image and reference image and simultaneously indicates to what extent that these corresponding voxels potentially don't belong to gross outliers. Furthermore, in calculating joint histogram of two images, the contribution that the intensity of a sampled voxel at floating image is given to the joint histogram is weighted using the normalized joint saliency value, which can produce histogram blurring to improve registration robustness. At last, the original two images are then matched using normalized mutual information computation guided by joint saliency map to automatically exclude gross outliers in iterative registration procedure. The results show that the method is sufficiently accurate and robust to gross outliers for multi-temporal and multi-modal medical image registration.The second part of the thesis is devoted to novel adaptive non-rigid medical image registration based on control point sets of joint salient regions. Using the scale-invariant salient point detector, the control points are extracted for the two images to be registered. The joint saliency map combined with cluster analysis is computed for detecting the outlier control points. After gradually screening out of control points, some stable sub-class of control points that well represent the structures within joint salient region at two image could be obtained to construct the deformation field between the two images in the subsequent matching procedure. By iteratively finding correspondences of each sub-class of control points, the smooth deformation field is approximated simultaneously using radial basis functions with compact support until the convergence to the steady-state solution is achieved. Experimental results show that the proposed method is able to recover local deformation caused by tumor-like gross outlier, and ensure a point-to-point correspondence in multi-modal and multi-temporal non-rigid medical image registration.