The Application Of Joint Saliency Map Based Structural Similarity Measure In Nonrigid Image Registration

Nonrigid image registration has been widely used in image analysis, pattern recognition, remote sensing and clinical diagnosis. In recent years, the B-spline based nonrigid registration algorithm has attracted extensive attention due to its computational efficiency, general applicability and locally controllable property. However, the B-spline interpolation process in each resolution treats all the pixels in the same way while computing their displacement vectors, which will results in estimation inaccuracy and waste of time. Furthermore, this kind of algorithm will encounter many problems, and can hardly meet clinical requirement when dealing with brain tumor resection images with missing correspondence and complex local large distortions. In this thesis, we proposed an improved B-spline based nonrigid registration algorithm using joint saliency map (JSM) to constrain the displacement magnitude of the interpolated points, and employing local structure tensor as well as the weighted similarity measure to guarantee the topology preservation.We first detect the regions without correspondence by using the improved JSM, which indicates the local structural correspondence between the reference and floating images. In the JSM, the overlapping points in the mismatched regional structures which should be emphasized in the registration process will be assigned with high fractional joint saliency, while the effect of the background and the homogeneous pixel pairs will be suppressed by assigning with low joint saliency value. The new JSM is then refined iteratively to guide the B-spline deformation interpolation magnitude to emphasize on dramatically changed or unmatched anatomical structures, to avoid unnecessary calculations of the background and the homogeneous regions as well as to get more accurate results in the brain tumor resection regions.To guarantee that no folding, crossing or tearing problems appear during the registration, we integrate the local structure tensor information into the computation of the displacement vector field. Specifically, we adopt structure tensor as a description of local structure orientation to distinguish the different local structures in an image. The direction of the resulting displacement vector calculated by B-splines FFD can be simultaneously fine-tuned to preserve the topological structure and to ensure the smoothness of the transformation field by using the geometrical information obtained from the structure tensor.Experimental results confirmed that by using the new JSM as a measurement of the displacement amplitude and the corresponding eigenvectors of the local structure tensors to delicately adjust the deformation direction, our approach exhibits abilities of suppressing outlier effects and obtaining smoothed deformation field in nonrigid registration of images with unmatchable local structures and complex local deformations. Besides, our method yields higher accuracy with more efficiency in preserving the topological structures than the conventional B-splines based algorithm and other state-of-the-art algorithms.