A three-dimensional wavelet-based algorithm for nonlinear registration of an elastic body model of the brain is developed. Surfaces of external and internal anatomic brain structures are used to guide alignment. The deformation field is represented with a multiresolution wavelet expansion and is modeled by the partial differential equations of linear elasticity. A progressive estimation of the registration parameters and the usage of an adaptive distance map reduce algorithm complexity, thereby providing computational flexibility that allows mapping of large, high resolution datasets. The performance of the algorithm was evaluated on rat brains. The wavelet-based registration method yielded a twofold improvement over affine registration.

Breast cancer diagnosis through ultrasound tissue characterization was studied using receiver operating characteristic (ROC) analysis of combinations of acoustic features, patient age, and radiological findings. A feature fusion method was devised that operates even if only partial diagnostic data are available. The ROC methodology uses ordinal dominance theory and bootstrap resampling to evaluate A(z) and confidence intervals in simple as well as paired data analyses. The combined diagnostic feature had an A(z) of 0.96 with a confidence interval of at a significance level of 0.05. The combined features show statistically significant improvement over prebiopsy radiological findings. These results indicate that ultrasound tissue characterization, in combination with patient record and clinical findings, may greatly reduce the need to perform biopsies of benign breast lesions.

Classical interpolation theory, proposes to find a function f(x) which satisfies a set of interpolative constraints, f(x/sub i/)=y/sub i/. This approach has been used in the alignment of images and volumetric data sets. However, in many applications, it is difficult if not impossible to find corresponding fiducial points. In these applications however, it is possible to identify homologous geometrical structures, such as surfaces. We have explored algorithms for using these sorts of data for image alignment, which requires a novel formulation of the interpolation problem. This paper describes the principle of surface-based alignment, and examines its performance and accuracy.

Discusses spatial transformation techniques for multidimensional image registration. Examples from brain mapping are given. The authors illustrate the problem by reviewing point matching, which is the 'gold standard' in this area. The point matching formulation is not adequate for all applications, and they describe various extensions and generalizations. One important issue is multi-set registration, others are more general nonlinear models, as well as the problems of registration on the basis of intensity data and when the geometrical objects to be registered are curves, surfaces, or volumes. This leads to many interesting problems that deserve further study. Finally, the authors discuss the accuracy in registration.

The problem of adaptive segmentation of images of objects with smooth surfaces is addressed. The images are composed of regions of slowly varying intensities that may be corrupted by additive noise. The underlying field is modeled by Markov random field that consists of both a label process which contains the classification of each pixel in the image and intensity functions which contain the possible grey levels that each pixel may take. The algorithm iteratively repeats two steps: the parameter estimation step, in which the maximum-likelihood (ML) estimates of the associated parameters are obtained; and the restoration step, in which the underlying field is estimated through the maximum-a-posteriori (MAP) method. The concept of allowing the pixel grey values to vary across the image regions is discussed. These values are estimated by using windows on the observed data. As the algorithm progresses, the window size is decreased so that the algorithm adapts to the characteristics of each region.>

A new algorithm for unsupervised textured image segmentation is presented. The image comprises M textured regions, each of which is modeled by a stationary Gaussian Markov random field. A feature vector is computed for each pixel in the original image where these vectors are normally distributed and cluster about some vector means. Thus, the problem is reduced to one of restoring a vector valued underlying field embedded in additive Gaussian noise. The vector means corresponding to the different regions are estimated by using the expectation-maximization (EM) algorithm. An iterative algorithm is used with the underlying field modeled as a multilevel logistic Markov random field. The results obtained on two-region and four-region textured images are impressive, and the classification error is less than 3%. The algorithm is not limited to textured images but can also be applied to any vector-valued signals.>

Addresses the problem of detecting small regions of interest embedded in larger areas of interest as imaged by a fluorescence imaging system. The application exhibits the variation observed in biological specimens, which are primarily shape and intensity uncertainties. The presence of these variability together with uneven illumination and the lack of a global model, implies that a reliable nonparametric technique should be used to detect the objects of interest. The detection task is formulated as a two step hierarchical approach which integrates both parametric and nonparametric techniques. The image as a whole is considered as a slowly varying multi-modal Gaussian field. The classification of which is obtained through the expectation maximisation algorithm, and a spatially smoother segmentation is accomplished by using a Gibbsian segmenter. Shape deformation constraints retain only the so-called valid objects. A similar approach is employed in the second step, where objects within the already detected objects are identified.>

A robust and efficient technique for detection of region boundaries in a sequence of images is the foundation for a hierarchical framework for visual target tracking. This robust technique is a multi-hypothesis test that searches for a boundary that maximizes a certain performance criterion. The framework also consists of a technique to restore noisy images when the parameters are not known and also filtering to predict positions and orientations of successive images.>

The problem of region recognition in a sequence of images is addressed, and a recognition system that finds and tracks region-of-interest boundaries in those images is presented. These regions are not stationary: parts of the boundary may be missing or completely blurred and outliers are likely to exist. Thus, the emphasis is on robustification and efficiency. The region segmentation problem was formulated as a multihypothesis test that seeks the boundary that maximizes a performance criterion which is general in terms of blur and noise. Efficiency is obtained by restricting outline candidates to an adaptive search area near the optimal boundary from the previous section. The search for the maximum is cast into a fast first-order dynamic programming procedure. Robust statistical techniques are used in the multihypothesis test to reduce the sensitivity to outliers and unexpected noise. The inconsistent parts of the optimal boundary are then detected by using a robust expectation maximization algorithm and are interpolated from higher-quality parts. The boundary obtained by this method is used as the reference boundary for the next image.>

The problem of restoring noisy images when the model parameters are not known is discussed. The underlying field, x, is modeled as a noncausal Markov random field (MRF), namely, either a multilevel logistic (MLL) or a Gaussian MRF, and is corrupted by additive independently identically distributed (i.i.d.) Gaussian noise. The application is a restoration/segmentation of regions of interest in an image obtained from histologies of brain sections, which suggests an MLL modeling since the regions are spatially smooth. The presented algorithm maximizes the joint likelihood of the observations, y, and x given the unknown parameters. The parameters of the noise and the random field are estimated separately through a maximum likelihood technique given the current estimate of x, and the underlying field is estimated through a maximum a posteriori method. In the case of images modeled by MLL MRFs, the result of the restoration is actually a segmentation since the collection of all pixels with the same level defines a region. The results show that the algorithm successfully segments the region of interest even when the signal-to-noise ratio is low.>

A theory is proposed for the structure of edges in a two dimensional visual field. This theory consists of models of ideal edges, and of distortion criteria for the evaluation of the accuracy with which a visual field fits this model. In this formulation the edge is specified by a curve described by a pair of parametric equations, and the line or edge detection problem is solved by minimizing a cost function over the set of curves. The substance of the paper consists of mathema tical transformations of the above problem which lead to efficient algorithms and novel edge detection operators.