Deformable neuroimage registration is an active and challenging research
area. It forms a crucial component of many
computational and clinical neuroscience applications, including computer
aided diagnosis, statistical quantification of human brain,
and atlas-based neuroimage segmentation.
Maximizing the number of correctly estimated voxel correspondences enhances the accuracy of a deformable registration algorithm. Most existing feature-based deformable registration algorithms use a pre-defined set of image features to estimate correspondences for all voxels. These methods have two main weaknesses. First, the feature vector is constructed by the authors of the algorithms rather than automatically selected to minimize registration error. Second, the same feature vector is used for all the voxels in the whole brain image, without consideration given to the inhomogeneity of the anatomical structures and their corresponding voxels.
We propose a new learning-based deformable registration algorithm that performs feature selection for every voxel. Our algorithm can be trained to accurately register specific anatomical structures as well as the entire neuroimages of specific patient groups. The main novelty of our approach is that it automatically learns feature vectors for distinguished individual image voxels, thus increasing correspondence estimation accuracy. Our method utilizes a decision theoretic approach to systematically calculate the expected correspondence estimation error for a voxel in many different feature spaces, and then select the space with the smallest error. Our feasibility study on the 2D midsagittal slices shows that learning feature subspace increases number of correctly estimated correspondences by 20%.
We will quantitatively evaluate the performance of our deformable registration algorithm and apply it to several medical image analysis problems.
Back to the Main Page
Pradeep Ravikumar Last modified: Mon Sep 25 09:21:54 EDT 2006