The SIFT detector has been successfully applied to various applications such as face recognition [33] and medical image registration [24�C27]. In image registration, the SIFT keypoints [24, 27] localized at specific anatomical structures could be automatically selected for the adaptive setup of the irregular control point grids for the local deformation of specific anatomical structures. Without tedious manual selection of control points, the adaptive setup of irregular control point grids could alleviate the computational cost and the registration inaccuracy that are related to the regular grids [34] of control points arranged for the local large deformation at the tumor resection region.

In general, it is often difficult to correctly match local keypoints [26, 28] by using only the similarity between SIFT descriptors when complex local deformation and outliers exist in brain images.

The deformation invariant local feature descriptor was presented in [29], however this topic is beyond the scope of this paper.Although the results of these above methods clearly demonstrate the power of local invariant feature-based nonrigid deformations, the desired landmark-based registration algorithm should establish robust control point correspondence to accurately model the complex local deformation around the tumor resection region. To find robust point correspondence, some approaches are proposed including soft correspondence detections [20, 35], joint clustering-matching strategy [36] and modeling point sets by kernel density function [37].

Compared with the classical template matching, the iterative closest point [38] and the correspondence by sensitivity to movement [39], the self-organizing map [40] algorithm was considered in [41] to be the most effective method in 2D feature point correspondence detection. However, these methods do not consider the complexity of correspondence detection in the context of local large structure distortion combined with the outliers. In this work, we first compute the global correspondence between the contiguous matching areas of normal tissues and the tumor regions in the two images by using MI-based rigid registration method.

The global correspondence is then used to introduce Dacomitinib the cluster correspondence between Cilengitide the paired pools of DoG keypoints [23] that are detected and clustered at those contiguous matching areas in the two images. An important contribution is that we have proposed the cluster-to-cluster correspondence can be introduced as a useful constraint for the local point correspondence detection within the paired pools of keypoints.