Fast Shape-Based Nearest-Neighbor Search for Brain MRIs Using Hierarchical Feature Matching Peihong Zhu, Suyash P. Awate, Samuel Gerber, and Ross Whitaker Scientific Computing and Imaging Institute, University of Utah, USA Abstract. This paper presents a fast method for quantifying shape dif- ferences/similarities between pairs of magnetic resonance (MR) brain images. Most shape comparisons in the literature require some kind of deformable registration or identification of exact correspondences. The proposed approach relies on an optimal matching of a large collection of features, using a very fast, hierarchical method from the literature, called spatial pyramid matching (SPM). This paper shows that edge-based im- age features in combination with SPM results in a fast similarity mea- sure that captures relevant anatomical information in brain MRI. We present extensive comparisons against known methods for shape-based, k-nearest-neighbor lookup to evaluate the performance of the proposed method. Finally, we show that the method compares favorably with more computation-intensive methods in the construction of local atlases for use in brain MR image segmentation. 1 Introduction Large collections of medical images are becoming ubiquitous as public resources, and within specific clinical practices. Currently, large studies consist of thousands of images, but in the coming years databases of images of various types will grow to tens of thousands. The availability of such data demands new techniques for image analysis and associated algorithms that are able to efficiently take advantage of these large collections. We begin with a brief discussion of the kinds of algorithms that utilize these large sets of images and how these algorithms demand new technologies for fast image lookup, and end this section with a discussion of how the proposed method addresses this challenge. One use for a large collection of medical images is to aid in segmentation or tis- sue classification. Atlases, comprising voxel-wise tissue probabilities, incorporate information about spatial location of biological structures. Atlases with hard/soft tissue/object assignments can be used alone, or as “priors”, for segmentation and combined with voxel measurements of a specific image to generate label maps in previously unseen images. Atlases are typically constructed by summarizing information concerning image intensities and anatomical shapes from a training set that includes manual segmentations. The information is summarized in (i) an average image (template) and (ii) a tissue probability map. To segment a test image, the template is warped to the test image and the tissue probabilities in the atlas are then transferred to the test image. G. Fichtinger, A. Martel, and T. Peters (Eds.): MICCAI 2011, Part II, LNCS 6892, pp. 484–491, 2011. c  Springer-Verlag Berlin Heidelberg 2011