(21) (22) (65)
(60) (51) (52)
Siemens Medical Solutions USA, Inc.,
Malvern, PA (US)
Subject to any disclaimer, the term of this patent is extended or adjusted under 35 U.S.C. 154(b) by 737 days.
(58) Field of Classification Search 382/190,
382/195, 201, 206, 220, 274, 276, 294; 345/589, 345/594,419, 427, 662, 639 See application file for complete search history.
(56) References Cited
U.S. PATENT DOCUMENTS
4,700,400 A * 10/1987 Ross 382/205
6,243,106 Bl* 6/2001 Rehgetal 345/473
6,430,307 Bl* 8/2002 Soumaetal 382/118
6,690,725 Bl* 2/2004 Abdeljaoud et al 375/240.08
7,177,486 B2* 2/2007 Stewart et al 382/294
* cited by examiner
Primary Examiner—Kanji Patel
(74) Attorney, Agent, or Firm—Donald B. Paschburg
A method of aligning a pair of images with a first image and a second image, wherein said images comprise a plurality of intensities corresponding to a domain of points in a D-dimensional space includes identifying feature points on both images using the same criteria, computing a feature vector for each feature point, measuring a feature dissimilarity for each pair of feature vectors, wherein a first feature vector of each pair is associated with a first feature point on the first image, and a second feature vector of each pair is associated with a second feature point on the second image. A correspondence mapping for each pair of feature points is determined using the feature dissimilarity associated with each feature point pair, and an image transformation is defined to align the second image with the first image using one or more pairs of feature points that are least dissimilar.
42 Claims, 2 Drawing Sheets
METHOD AND SYSTEM FOR HYBRID RIGID
REGISTRATION OF 2D/3D MEDICAL
IMAGES
CROSS REFERENCE TO RELATED UNITED 5
STATES APPLICATIONS
This application claims priority from "A Hybrid Rigid Registration Method for 2D/3D Medical Images", Provisional Patent Application No. 60/504,873 of Xu, et al., filed 10 Sept. 22, 2003, the contents of which are incorporated herein by reference.
BACKGROUND OF THE INVENTION
15
Image registration aims to spatially align one image to another. For that purpose, parameters of a global transformation model, such as rigid, affine or projective, are to be recovered to geometrically transform a moving image to achieve high spatial correspondence with a fixed image. The problem 20 has been studied in various contexts due to its significance in a wide range of areas, including medical image fusion, remote sensing, recognition, tracking, mosaicing, and so on.
Rigid registration of 2D/3D medical images is a vital com- 25 ponent of a large number of registration and fusion applications. In the areas of diagnosis, planning, evaluation of surgical and radio-therapeutical procedures, typically multiple single-modality, or multi-modality images are acquired in the clinical track of events. Since these images are complemen- 3Q tary to each other, the integration of useful data from separate images are often desired. Rigid registration, a first step in this integration process, aims to bring the multiple images involved into spatial alignment.
Existing methods for image registration can largely be 35 classified into three categories: feature-based methods, intensity-based methods, and hybrid methods that integrate the previous two. Feature-based methods use sparse geometric features such as points, curves, and/or surface patches, and their correspondences to compute an optimal transformation. 40 These methods are relatively fast. However, the main critiques of this type of methods in the literature are the robustness of feature extraction, the accuracy of feature correspondences, and the frequent need of user interaction. Intensitybased registration methods operate directly on the intensity 45 values from the full image content, without prior feature extraction. These methods have attracted much attention in recent years since they can be made fully automatic and can be used for multimodality image matching by utilizing appropriate similarity measures. However, these methods tend to 50 have high computational cost due to the need for optimization on complex, non-convex energy functions. In addition, they require the poses of two input images be close enough to converge to a local optimum. Furthermore, they often perform poorly when partial matching is required. 55
Recently, several hybrid methods are proposed that integrate the merits of both feature-based and intensity-based methods. Most focus on incorporating user provided or automatically extracted geometric feature constraints into the intensity-based energy functionals to achieve smoother and 60 faster optimization. Typically they are more flexible, and designed in such way that either intensity (gray values, gradients) information is incorporated into a feature-based algorithm, or feature (points, surfaces) information is introduced to a pixel/voxel intensity-based algorithm. The hybrid meth- 65 ods are expected to be more efficient and robust than the pure-feature, or pure-intensity based methods
A hybrid rigid registration method uses extracted features and feature correspondences. The method is expected to provide a fast (real time), robust, working registration system that can be used in practical applications on a regular basis. During registration, good salient anatomical features are treated differently from other pixels/voxels. They are used in the initial iterations and given much higher weights. By explicitly looking for salient anatomical feature correspondences, and using the correspondences to estimate a transformation, a registration result is provided having a satisfactory validation outcome.
In one aspect of the invention there is provided a method of aligning a pair of images with a first image and a second image, wherein the images comprise a plurality of intensities corresponding to a domain of points in a D-dimensional space. The method includes identifying feature points on both images using the same criteria, computing a feature vector for each feature point, measuring a feature dissimilarity for each pair of feature vectors, wherein a first feature vector of each pair is associated with a first feature point on the first image, and a second feature vector of each pair is associated with a second feature point on the second image, determining a correspondence mapping for each pair of feature points using the feature dissimilarity associated with each feature point pair; and defining an image transformation to align the second image with the first image using one or more pairs of feature points that are least dissimilar.
In further aspect of the invention, the steps of identifying feature points, computing a feature vector, measuring a feature dissimilarity, determining a correspondence mapping, and defining an image transformation are repeated for a predetermined number of iterations.
In a further aspect of the invention, the steps of identifying feature points, computing a feature vector, defining a feature dissimilarity, determining a correspondence mapping, and defining an image transformation are repeated until the transformed second image differs from the first image by a predetermined error.
In a further aspect of the invention, the criteria for identifying feature points include selecting points whose local intensity variance is a maximum.
In a further aspect of the invention, the criteria for identifying feature points include selecting points with a high local curvature.
In a further aspect of the invention, feature vector components are selected from a group that includes the coordinates of the feature point, the intensity associated with the feature point, the local curvature associated with the feature point, a histogram of local neighborhood intensity values, and intensity values of the neighborhood points.
In a further aspect of the invention, the dissimilarity of a pair of feature vectors can be measured by a distance between the respective feature vectors of the pair of vectors.
In a further aspect of the invention, the distance between the respective feature vectors is selected from a group that includes a Euclidean distance, a Mahalanobis distance, a normalized cross-correlation, and a mutual information.
In a further aspect of the invention, the method includes determining a correspondence mapping for each pair of feature points includes associating each feature vector on the first image with the closest feature vector on the second image, as determined by the smallest dissimilarity measure between the feature vector on the first image and each feature vector on the second image.
|
