Optimization Algorithms In Rigid Registration Of Medical Image

The process of medical images hogs a hot position in the study of inter-disciplines including applied mathematics. The development of visual technologies offers various modes of medical images and image registration accounts for a large part in medical image analysis. Since medical image registration enables doctors to synthesize information of image, it supports them technologically in searching lesions, correct diagnoses, and therapy. Medical image registration is meant to be seeking for a series of spatial transformation so as to enable the corresponding points on two pieces of medical images to accord with each other in special positions and anatomic structures. The rigid registration is a widely adopted technology in clinic application. Despite the ever-improvement of rigid registration, the present optimization algorithms remain to fail in avoiding the problems of local minimum. This paper devotes to optimization algorithms, which can search out the minimum of the whole by stably compressing solution space and without any loss in precision.Medical image registration devotes to a series of transformation parameters in essence, which assures the measure of two images’ similarity measure reaching the summit. The registration mainly contains two avenues:feature-based image registration and area-based image registration. Feature-based image registration costs such less calculation that it meets the clinic need on speed. Due to feature extraction, the precision of registration tends to be affected by the following errors. Area-based image registration has the virtue of no image reprocessing and high robustness yet requires more calculation. Among the image registration based on area, mutual information registration is a method that is applied in wide use. The paper mainly realizes two-dimension image rigid registration based on mutual information and analyzes theoretical bases and features of optimization of Powell, simplex and coordinate alternative method. As for the weakness of Powell which tends to collect local extreme, the paper compresses searching space by steps and at last, distributes in average within a small space determined. By simplex local searching, the way that selects the best point as final optimization result can effectively avoid incorrect minimum. Besides, to resolve the problem of low accuracy by space transformation within a large scale, the paper modifies this method and improves accuracy of registration as for excessively revolved images. It can be applied to products.