Catálogo de publicaciones - libros

We propose new techniques for unsupervised segmentation of multi-modal grayscale images such that each region-of-interest relates to a single dominant mode of the empirical marginal probability distribution of gray levels. We follow most conventional approaches such that initial images and desired maps of regions are described by a joint Markov–Gibbs random field (MGRF) model of independent image signals and interdependent region labels. But our focus is on more accurate model identification. To better specify region borders, each empirical distribution of image signals is precisely approximated by a linear combination of Gaussians (LCG) with positive and negative components. Initial segmentation based on the LCG-models is then iteratively refined by using the MGRF with analytically estimated potentials. The convergence of the overall segmentation algorithm at each stage is discussed. Experiments with medical images show that the proposed segmentation is more accurate than other known alternatives.

This paper presents an integrated algorithm for MRI (Magnetic Resonance Imaging) brain tissues segmentation. The method is composed of four stages. Noise in the MRI images is first reduced by a versatile wavelet-based filter. Then, the watershed algorithm is applied to brain tissues as an initial segmenting method. Because the result of classical watershed algorithm on grey-scale textured images such as tissue images is over-segmentation. The third stage is a merging process for the over-segmentation regions using fuzzy clustering algorithm (Fuzzy C-Means). But there are still some regions which are not divided completely due to the low contrast in them, particularly in the transitional regions of gray matter and white matter, or cerebrospinal fluid and gray matter. We exploited a method base on Minimum Covariance Determinant (MCD) estimator to detect the regions needed segmentation again, and then partition them by a supervised k-Nearest Neighbor (kNN) classifier. This integrated approach yields a robust and precise segmentation. The efficacy of the proposed algorithm is validated using extensive experiments.

Fluorescent confocal laser scanning microscope (CLSM) imaging has become popular in medical domain for the purpose of 3D information extraction. 3D information is extracted either by visual inspection or by automated techniques. Nonetheless, 3D information extraction from CLSM suffers from significant lateral intensity heterogeneity. We propose a novel lateral intensity heterogeneity correction technique to improve accurate image analysis, e.g., quantitative analysis, segmentation, or visualization. The proposed technique is novel in terms of its design (spatially adaptive mean-weight filtering) and application (CLSM), as well as its properties and full automation. The key properties of the intensity correction techniques include adjustment of intensity heterogeneity, preservation of fine structural details, and enhancement of image contrast. The full automation is achieved by data-driven parameter optimization and introduction of several evaluation metrics. We evaluated the performance by comparing with three other techniques, four quality metrics, and two realistic synthetic images and one real CLSM image.

In this paper we present a simple method for flattening of triangulated surfaces for mapping and imaging. The method is based on classical results of F. Gehring and Y. Väisälä regarding the existence of quasi-conformal and quasi-isometric mappings between Riemannian manifolds. A random starting triangle version of the algorithm is presented. A curvature based version is also applicable. In addition the algorithm enables the user to compute the maximal distortion and dilatation errors. Moreover, the algorithm makes no use to derivatives, hence it is robust and suitable for analysis of noisy data. The algorithm is tested on data obtained from real CT images of the human brain cortex and colon, as well as on a synthetic model of the human skull.

Many computer aided diagnosis schemes in chest radiography start with preprocessing steps that try to remove or suppress normal anatomical structures from the image. Examples of normal structures in posteroanterior chest radiographs are bony structures. Removing these kinds of structures can be done quite effectively if the right dual energy images—two radiographic images from the same patient taken with different energies—are available. Subtracting these two radiographs gives a soft-tissue image with most of the rib and other bony structures removed. In general, however, dual energy images are not readily available. We propose a supervised learning technique for inferring a soft-tissue image from a standard radiograph without explicitly determining the additional dual energy image. The procedure, called dual energy faking, is based on k -nearest neighbor regression, and incorporates knowledge obtained from a training set of dual energy radiographs with their corresponding subtraction images for the construction of a soft-tissue image from a previously unseen single standard chest image.

We introduce a novel minimally-interactive watershed algorithm that needs no initial parameterization, but lets the user refine the automatic segmentation close to real-time. In contrast to previous proposals, our algorithm encapsulates all time consuming calculation in a processing step executed only once. Thereby, a hierarchical subdivision of the incoming image data is generated. This subdivision serves as a basis for computing automatic segmentation results according to a given multi-dimensional classification scheme as well as for interactive refinement according to local markers. We have successfully applied our algorithm to efficiently removing bone structures from computed tomography angiography data, which is among the very challenging segmentation problems in medical image analysis.

The function of the human brain arises from computations that occur within and among billions of nerve cells known as neurons. A neuron is composed primarily of a cell body (soma) from which emanates a collection of branching structures (dendrites). How neuronal signals are processed is dependent on the dendrites’ specific morphology and distribution of voltage-gated ion channels. To understand this processing, it is necessary to acquire an accurate structural analysis of the cell. Toward this end, we present an automated reconstruction system which extracts the morphology of neurons imaged from confocal and multi-photon microscopes. As we place emphasis on this being a rapid (and therefore automated) process, we have developed several techniques that provide high-quality reconstructions with minimal human interaction. In addition to generating a tree of connected cylinders representing the reconstructed neuron, a computational model is also created for purposes of performing functional simulations. We present visual and statistical results from reconstructions performed both on real image volumes and on noised synthetic data from the Duke-Southampton archive.

A novel approach is presented for modeling the complex activity pattern of the heart in echocardiogram videos. In this approach, the heart is represented by the constellation of its chambers, where the constellation is modeled by pictorial structure at each instance in time. Pictorial structure is then extended to the temporal domain to simultaneously capture the evolution pattern of the appearance of each chamber, the evolving spatial relationships between them, and the topological transformations in their constellation due to phase transitions. Inference and learning algorithms are presented for the model. The problem of correspondence is solved at each stage of the inference process, by matching the evolving model of the complex activity pattern to the observed constellations. The model, which is trained using examples of normal echocardiogram videos is shown to be efficient in temporal segmentation of the content of echocardiogram videos into different phases during one cycle of heart activity.

This paper addresses two questions concerning JPEG2000 compression – firstly – how much has noise influence on compression performance – secondly – can compression performance be improved by applying a new complementary conception with introducing a denoising process before the application of compression Indeed, radiographic images are a combination between the relevant signal and noise, which is per definition not compressible. The noise behaves generally close to a mixture of Gaussian and/or Poisson statistics, which generally affects the compression performance. In this paper, the influence of noise on the compression performance of JPEG2000 images with investigating the parameters signal dynamic and spatial pattern frequency are considered; and the JPEG2000 compression scheme combined with a denoising process is analyzed on simulated and real dental ortho-pan-tomographic images. The test images are generated using Poisson statistics; the denoising utilizes a Monte Carlo noise modeling method. A hundred selected images are denoised and the compression ratio, using lossless and lossy JPEG 2000, is reported and evaluated.

This paper presents a semi-automatic method for tracking the mitral valve leaflet in transesophageal echocardiography. The algorithm requires a manual initialization and then segments an image sequence. The use of two constrained active contours and curve fitting techniques results in a fast segmentation algorithm. The active contours successfully track the inner cardiac muscle and the mitral valve leaflet axis. Three sequences have been processed and the generated muscle outline and leaflet axis have been visually assessed by an expert. This work is a part of a more general project which aims at providing real-time detection of the mitral valve leaflet in transesophageal echocardiography images.