Catálogo de publicaciones - libros

The presence of microcalcification clusters in digital mammograms is a primary indicator of early stages of malignant types of breast cancer and its detection is important to prevent the disease. This paper uses a procedure for the classification of microcalcification clusters in mammograms using sequential Difference of Gaussian filters (DoG) and feedforward Neural Networks (NN). Three methods using class separability, forward sequential search and genetic algorithms for feature selection are compared. We found that the use of Genetic Algorithms (GAs) for selecting the features from microcalcifications and microcalcification clusters that will be the inputs of a feedforward Neural Network (NN) results mainly in improvements in overall accuracy, sensitivity and specificity of the classification.

The research presented in this paper is aimed at the development of an automatic mass classification of mammograms. This paper focuses on using contourlet-based multi-resolution texture analysis. The contourlet transform is a new two-dimensional extension of the wavelet transform using multi-scale framework as well as directional filter banks. The proposed method consists of three steps: removing pectoral muscle and segmenting regions of interest, extracting the most discriminative texture features based on the contourlet coefficients, and finally creating a classifier, which identifies various tissues. In this research classification is performed based on the idea of Successive Enhancement Learning (SEL) weighted Support Vector Machine (SVM). The main contribution of this work is exploiting the superiority of the contourlets to the-state-of-the-art multi-scale techniques. Experimental results show that contourlet-based feature extraction in conjunction with the SEL weighted SVM classifier significantly improves breast mass detection.

This paper introduces a fully automated segmentation scheme for carotid artery ultrasound images. The proposed scheme is based on fuzzy c-means clustering. It consists of four major stages. These stages are pre-processing, feature extraction, fuzzy c-means clustering, and finally boundary extraction. Experimental results demonstrated the efficiency of the proposed scheme in segmenting carotid artery ultrasound images.

In this paper a novel method that automatically detects the lumen-intima border on an intravascular ultrasound sequence (IVUS) is presented. First, a 3D co-occurrence matrix was used to efficiently extract the texture information of the IVUS images through the temporal sequence. By extracting several co-occurrence matrices a complete characterization feature space was determined. Secondly, using a -means algorithm, all the pixels in the IVUS images were classified by determining if they belong to either the lumen or the other vessel tissues. This enables automatic clustering and therefore no learning step was required. The classification of the pixels within the feature space was obtained using 3 clusters: two clusters for the vessel tissues, one cluster for the lumen, while the remaining pixels are labeled as unclassified. Experimental results show that the proposed method is robust to noisy images and yields segmented lumen-intima contours validated by an expert in more than 80% of a total of 300 IVUS images.

Ultrasound provides a non-invasive means for visualizing various tissues within the human body. However, these visualizations tend to be filled with speckle noise and other artifacts, due to the sporadic nature of high frequency sound waves. Many techniques for segmenting ultrasound images have been introduced in order to deal with these problems. One such technique is the active contouring.

In this paper, two proposed alterations to the dynamic programming parametric active contour model (or snake) are introduced. The first alteration allows the snake to converge to the one-response result of a modified Canny edge detector. The second provides a function that allows a user to preset knowledge about a given object being detected, by means of curve fitting and energy modification. The results yield accurate segmentations of cross-sectional transverse carotid artery ultrasound images that are validated by an independent clinical radiologist. Utilizing the proposed alterations leads to a reduction of clinician interaction time while maintaining an acceptable level of accuracy for varying measures such as percent stenosis.

Transmission scans are performed to provide attenuation correction factors (ACFs) information for positron emission tomography(PET). Long acquisition or scan times for transmission tomography, although alleviating the noise effect of the count-limited scans, are blamed for patient uncomfortableness and movements. So, the quality of transmission tomography from short scan time often suffers heavily from noise effect and limited counts. Bayesian approaches, or maximum (MAP) methods, have been accepted as an effective solution to overcome the ill-posed problem of count-limited transmission tomography. Based on Bayesian and Markov Random Fields(MRF)theories, prior information of the objective image can be incorporated to improve the reconstructions from count-limited and noise-contaminating transmission scans. However, information of traditional priors comes from a simply weighted differences between the pixel densities within local neighborhoods, so only limited prior information can be provided for reconstructions. In this paper, a novel MRF prior, which is able to exploit global information of image by choosing large neighborhoods and a new weighting method, is proposed.Two-step monotonical reconstruction algorithm is also given for PET transmission tomography. Experimentations show that the reconstructions using the nonlocal prior can reconstruct better transmission images and overcome the ill-posed problem even when the scan time is relatively short.

Abstract. This paper presents a new 3-D image registration method based on the equivalent meridian plane and mutual information.Cnamed as EMP-MI algorithm. Compared with MI based registration method using the whole volume intensity information, our approach utilizes the equivalent meridian plane information. First, principal component analysis (PCA) is applied to determine the equivalent meridian plane. The volume is roughly aligned by taking advantage of the EMP. Second, registration is refined via maximizing the MI of the EMP. We evaluate the effectiveness of the EMP-MI approach by applying it to the simulated and real brain image data (CT, MR, PET, and SPECT). The experimental results indicate that the algorithm is effective in reducing computation time as well as in helping to avoid local minima.

In this paper, a wavelet based approach is proposed for the detection and diagnosis of prostate cancer in Trans Rectal UltraSound (TRUS) images. A texture feature extraction filter was implemented to extract textural features from TRUS images that characterize malignant and benign tissues. The filter is based on the wavelet decomposition. It is demonstrated that the wavelet decomposition reveals details in the malignant and benign regions in TRUS images of the prostate which correlate with their pathological representations. The wavelet decomposition is applied to enhance the visual distinction between the malignant and benign regions, which could be used by radiologists as a supplementary tool for making manual classification decisions. The proposed filter could be used to extract texture features which linearly separate the malignant and benign regions in the feature domain. The extracted feature could be used as an input to a complex classifier for automated malignancy region classification.

Transmission electron microscope images of biological membranes are difficult to segment because they are low-contrasted images with heterogeneous gray levels. Added to that are the many possible types of membranes, the variable degree of aggregation, and the negative staining of the sample. We therefore develop a multi-scale approach to detect the edges at the appropriate scales. For these images, the study of the amplitude of the first derivative through the scales simplifies the feature tracking and the scale selection. A scale-adapted threshold is automatically applied to gradient images to progressively segment edges through the scales. The edges found at the different scales are then combined into a gradient-like image. The watershed algorithm is finally applied to segment the image into homogeneous regions, automatically selecting the edges found at the finest resolution.

A robust method to segment intervertebral disks and spinal canal in magnetic resonance images is required as part of a precise 3D reconstruction for computer assistance during diskectomy procedure with minimally invasive surgery approach. In this paper, an unsupervised segmentation technique for intervertebral disks and spinal canal from MRI data is presented. The proposed scheme uses a watershed transform and morphological operations to locate regions containing structures of interest. Results show that the method is robust enough to cope with variability of shapes and topologies characterizing MRI images of scoliotic patients.