Research Projects

1 A Framework for the Interactive Retrieval of Multimedia Objects

In this project we are developing an efficient and effective framework to interactively perform image segmentation and find “regions of interest” in a user input multimedia object. Similar multimedia objects may then be retrieved from a repository. The repository stores trained feature data of multimedia objects, which are obtained by applying feature extraction and dimension reduction analysis to the regions of interest. An advantage of our framework is that we divide the task into two independent modules: objects are processed in the image domain and the feature domain respectively. This two-module-based framework makes the independent development and implementation of the software for each domain possible. The framework has been successfully applied to image databases and it can be easily modified to accommodate multimedia objects.

2 Wavelet Markov Models for Image Segmentation and Denoising Applications

For statistical modeling of images, an image is treated as a realization of a spatial stochastic process defined on some domain in R². One advantage of statistical approaches for image modeling and processing is that they can provide a unified view of learning, classification and generation. In this project, several wavelet domain multiresolution hidden Markov models are studied and proposed in terms of the applications to image restoration and image segmentation. We also investigate the shift-variance problem caused by real wavelet transforms. A new local hidden Markov model is proposed based on the dual-tree complex wavelet transform that is approximately shift-invariant. This algorithm is applied to denoising problems to remove additive white Gaussian noise in an image. Our scheme outperforms those approaches based on real wavelet transforms and provides state-of-the-art image denoising performance.

3 Medical Image Registration

There are a number of medical imaging technologies and one technology may image one component of the anatomy, for example bone or soft tissue, better than another technology. Registration is the alignment of images of the same area produced by different technologies (multimodal images). The resulting aligned images give medical professionals a much better view of the imaged region. Medical image registration has been utilized in the diagnosis of breast cancer, cardiac studies, and a variety of neurological disorders including brain tumors.

A new approach for registration called maximization of mutual information, the quantifying of the similarity between corresponding voxel (a small 3-D region) intensities of two images, has been demonstrated to be a very powerful criterion for three-dimensional medical image registration. We proposed new implementation approaches for multimodal brain image registration based on mutual information. Each of these approaches gives improved accuracy or/and speed. We illustrate these approaches by presenting the results of tests from the data of a group of patients. Registration accuracy is evaluated as part of the Retrospective Registration Evaluation Project. We also compare the performance of each approach to results available in the literature. Our results have shown that each approach can reach sub-voxel accuracy with no loss of speed.

4 Retinal Image Registration

The retinal image registration work develops an object-oriented software system for automatic retinal image registration by mutual information maximization. For maximum portability the software is written in Java, using MVC (model-view-control) framework. We use the simplex downhill method as the optimization algorithm which is easy to implement, and is quick in practice. We demonstrate that this algorithm registers temporal and stereo retinal image pairs with a very high success rate, a satisfactory registration accuracy compared to point matching results, and within a clinically acceptable time.

5 Non-Rigid Medical Image Registration

The clinical need for PET-CT registration (Positron Emission Tomography-Cat Scan, a multimodal image registration) is becoming a necessity. This registration gives an interpretation of functional and structural images. PET provides information regarding functional abnormalities that in many cases are not detectable with CT imaging, the details depicted on CT scans can help pinpoint the location of the disease with a level of reliability that PET images alone can not match. In multimodal registration images may not be the same size so non-rigid registration approaches are necessary. In order to optimize the quality of CT-PET fused images, non-rigid image registration methods are needed to find local deformation between the two types of images. We propose a hybrid non-rigid medical image registration method using both feature and intensity based approaches.

6 Fast Region Growing Approaches for CT Angiography Applications

Much research on region growing has focused on the definition of the homogeneity criterion or the growing and merging criterion. However, one disadvantage of conventional region growing is redundancy. The redundancy causes a large memory usage, and the computation-efficiency is very low, especially for 3D images. To overcome this problem, a non-recursive single-pass 3D region growing was implemented and successfully applied to 3D CT angiography applications for vessel segmentation and bone removal. Clinical testings of this algorithm on brain CT angiography show this technique could effectively remove from the image the whole skull and most of the bones on the skull base, and thus reveal the cerebral vascular structures clearly.

7 Decomposition of signals with a G-reductive sum of Gaussians

For any given signal, there is generally more than one, perhaps many, Gaussian solutions within a given tolerance. A decomposition algorithm which tends to find a smaller Gaussian solution than another decomposition algorithm is said to be a more G-reductive algorithm. Our work here focuses on speeding up the process of G-reductive Gaussian decomposition without significantly worsening G-reduction. An effective parallel model for Gaussian decomposition is discussed. Also a novel method is presented in which each iteration has a low time complexity.

8 Edge Oriented and Region Oriented Segmentation Methods

Segmentation techniques are used to automate the process of isolating the relevant objects within a digital image from the extraneous background. In an edge-oriented scheme, the relevant object is identified by locating its outer edges. Such methods may be used, for example, for the segmentation of license plate numbers from an image of an automobile. In a region-oriented scheme, the relevant object is identified by accumulating neighboring pixels of similar intensities. This sort of method was developed to segment the soft tissue, lung and external regions in a tomographic image obtained from nuclear medicine imaging.