EVALUATION OF FEATURE-BASED REGISTRATION ALGORITHMS: University of Pennsylvania
Project Overview | Current Work & ObjectivesPROJECT OVERVIEW
RELATED RESEARCHThe objective of the project is to systematically evaluate and compare the performance of several fully-automated feature-based registration algorithms. Therefore, expertly-segmented brain data will be generated for the evaluation of intersubject registration, and the accuracy of registration algorithms will be evaluated. Moreover the performance of different features used for registration will also be studied based on developing a test bed which can utilize different features and image similarity measure.
Specifically, our tasks include:
- Create a general construction model for different simulations. Develop a landmark/label-based simulator of deformations, which combines interpolation schemes with our topology-correcting transformation method and provides less-biased deformations.
- Evaluate the performance of different features (attribute vectors) for defining correspondences among images and for performing image registration.
1. HAMMER.
HAMMER is an volumetric image registration algorithm, which uses moment-based features as voxel attributes and hierarchically performs image warping.
The following pictures show an example of warping our Jacob model image and its label image onto an MNI image (icbmfun_02026). The image icbmfun_02026 is thus labeled using the warped Jacob label image.
Figure 1. Apply HAMMER to register our model image (called "Jakob image") to the MNI data. This figure gives an example of the registration results, where the the Jakob model was warped onto the input MNI image, "icbmfun_02026".
Segmented Jacob Model Image Jacob Label Segmented Icbmfun_02026 Warped Jacob Label Overlay
2. The RAVENS function.
The RAVENs (Regional Analysis of Volumes Examined in Normalized Space) function was implemented, which calculates the local volumetric changes from the transformation field between two brain images. RAVENS map is one of the more powerful methods used to quantitatively capture the morphometric changes between medial images. It can also be used to evaluate the performance of registration algorithms by analyzing their ability to capturing known or simulated deformations. The following pictures show an example of the RAVENS map for White matter, Gray matter and Ventricle respectively.
Figure 2. RAVENS map on VN, WM, and GM. The colormaps show the degree of local volume contraction/expansion when the subject is registered onto the model. Green means there is no change, yellow/red/white mean there is a contraction, and blue/purple/dark mean an expansion.
3. Generate Simulated Atrophy.
We have generated images that have systematic contraction of certain gyri in order to test the ability of different methods to detect localized atrophy.
Atrophy is simulated by reducing the volume of the brain tissue (gray matter and white matter) in a designated region, via a method that imposes certain values on the Jacobian of the transformation on every single voxel. This way, an atrophy of a desired percentage level can be prescribed in a region of interest, with the rest of the Jacobians prescribed the value of 1 (no atrophy). The solution to the prescribed constraints is found via a computationally complex iterative procedure, involving projections on convex sets.
Figure 3. An example of simulated atrophy -- shrunken superior temporal gyrus and its neighboring region.4. Surface Extraction and Correction
The software package to extract the cortical surface from the MR brain image and to detect and correct the handles in a surface is also available.
A Summary of Current work and short term objectives from Penn.
1) Deformation Simulation.
We will use the existing sets of landmarks to develop a landmark-based simulator of deformations. The realistic deformations will be obtained by combining interpolation schemes with our topology-correcting transformation method. In particular, Gaussian, spline or other interpolants will be used to interpolate the deformation based on the landmarks, and to line up the gross boundaries in the source and target images in order to give plausible validity to the measurement. However, this deformation will introduce topological folds, generate unrealistic deformations due either to the sparsity of landmarks, to the interpolant itself (e.g. splines), or to sensitivity to errors in landmark placement. Therefore, we will add constraints on the Jacobian that can be as simple as positivity, which maintains a correct topology, to lower and upper bounds that limit the deformation effect and can therefore eliminate unrealistic deformations introduced during interpolation.
2) Attribute investigation
We will investigate the merits and limitations of different attributes in terms of accurate correspondence detection, based on the sets of landmarks. In particular, it includes defining figures of merit for correspondence detection based on attribute vectors (e.g. uniqueness and inverse consistency), and testing various attribute vectors. Thus we will establish the role of different scales/frequencies in determining correspondence (e.g. are local features, such as edges, useful or global features are more informative? Which combinations of features are best?).
The following figures show some ideas of evaluating attribute vectors (the inverse consistency and uniqueness of correspondence detection).
Figure 4. Inverse consistency. Given a point x in image 1, denoting its corresponding point in image 2 as y*, and the corresponding point of y* in image 1 as x*, the mapping x-->y*-->x* is said consistent if the distance between x and x* is small. The inverse consistency can be used as a factor to evaluate the performance of attribute vector in defining correspondences.
Figure 5. Uniqueness of correspondence. The colormap on the right shows the similarity between the attirbute vector of a point in image 1 and the attribute vector of every point in image 2, where the brown/red indicate higher similarities. Therefore, the number of peaks, and the width of the peak will indicate the uniqueness of correspondence matching.
Figure 6. A comparison of correspondence uniqueness using different attribute vectors. In the first row, the wavelet-based attribute vector is used, while moment-based attribute vector is used in the second row. From the similarity maps, it can be seen that the former attribute vector may produce more distinctive or more unique correspondence matching than the latter.
In this project, we will compare the performance of various kinds of attribute vectors in defining correspondences of landmarks, through defining measurement of merit for correspondence detection.
3) Evaluation of Warping Algorithms
Working with other groups to evaluate the warping algorithms, including their abilities in determining the correspondences of the landmarks, their abilities in registering images (by comparing with the simulated deformation from landmarks, and other GoWs), their abilities in capturing simulated atrophies, and others.
Back to Top
Section of Biomedical Image Analysis (SBIA), University of Pennsylvania.
URL: http://www.rad.upenn.edu/sbia/
Last Updated: Sept 30, 2004.