An Initialization Method of B-Spline Transformation for Medical Atlases Alignment

Atlases are widely used in medical image processing, especially in the case of image analysis based on machine learning. Spatial alignment between atlases is an essential step for these image processing tasks. The target of spatial alignment of atlases is to find an optimal spatial transformation that best aligns the structures of interest in the input images. For the case of medical image processing, due to strong ability of simulating non-rigid deformation, B-splines transformation model has been widely used. The goal of finding optimal spatial transformation is usually achieved by optimisation method. Before the optimisation starts to search for the optimal spatial transformation, an initialization of the spatial transformation model must be specified. This initialization has important influence to the searching procedure and result. A bad initialization will greatly increase search time and even more make searching algorithm converge to local minimum and then get an incorrect result. In this paper, a novel initialization method of B-spline transformation is proposed to improve the alignment of medical atlases. Specifically, the information in label images, which is neglected in conventional atlases alignment method, is used to generate an initial displacement field and then the initial displacement field is convert to parameters of B-splines transformation model. Experimental results show that our initialization approach can greatly enhance the robustness of atlases alignment method.