Catálogo de publicaciones - libros

Malignant melanoma is the most deadly form of skin lesion. Early diagnosis is of critical importance to patient survival. Existent visual recognition algorithms for skin lesions classification focus mostly on segmentation and feature extraction. In this paper instead we put the emphasis on the learning process by using two kernel-based classifiers. We chose a discriminative approach using support vector machines, and a probabilistic approach using spin glass-Markov random fields. We benchmarked these algorithms against the (to our knowledge) state-of-the-art method on melanoma recognition, exploring how performance changes by using color or textural features, and how it is affected by the quality of the segmentation mask. We show with extensive experiments that the support vector machine approach outperforms the existing method and, on two classes out of three, it achieves performances comparable to those obtained by expert clinicians.

A computer-aided diagnosis (CAD) method is reported that allows the objective identification of subjects with connective tissue disorders from 3D aortic MR images using segmentation and independent component analysis (ICA). The first step to extend the model to 4D (3D + time) has also been taken. ICA is an effective tool for connective tissue disease detection in the presence of sparse data using prior knowledge to order the components, and the components can be inspected visually. 3D+time MR image data sets acquired from 31 normal and connective tissue disorder subjects at end-diastole (R-wave peak) and at 45% of the R-R interval were used to evaluate the performance of our method. The automated 3D segmentation result produced accurate aortic surfaces covering the aorta. The CAD method distinguished between normal and connective tissue disorder subjects with a classification accuracy of 93.5 %.

While learning ensembles have been widely used for various pattern recognition tasks, surprisingly, they have found limited application in problems related to medical image analysis and computer-aided diagnosis (CAD). In this paper we investigate the performance of several state-of-the-art machine-learning methods on a CAD method for detecting prostatic adenocarcinoma from high resolution (4 Tesla) ex vivo MRI studies. A total of 14 different feature ensemble methods from 4 different families of ensemble methods were compared: Bayesian learning, Boosting, Bagging, and the k -Nearest Neighbor ( k NN) classifier. Quantitative comparison of the methods was done on a total of 33 2D sections obtained from 5 different 3D MRI prostate studies. The tumor ground truth was determined on histologic sections and the regions manually mapped onto the corresponding individual MRI slices. All methods considered were found to be robust to changes in parameter settings and showed significantly less classification variability compared to inter-observer agreement among 5 experts. The k NN classifier was the best in terms of accuracy and ease of training, thus validating the principle of Occam’s Razor . The success of a simple non-parametric classifier requiring minimal training is significant for medical image analysis applications where large amounts of training data are usually unavailable.

We describe a method to accurately assess articular cartilage morphology using the three-dimensional laser scanning technology. Traditional methods to obtain ground truth for validating the assessment of cartilage morphology from MR images have relied on water displacement, anatomical sections obtained with a high precision band saw, stereophotogrammetry, manual segmentation, and phantoms of known geometry. However, these methods are either limited to overall measurements such as volume and area, require an extensive setup and a highly skilled operator, or are prone to artifacts due to tissue sectioning. Alternatively, 3D laser scanning is an established technology that can provide high resolution range images of cartilage and bone surfaces. We present a method to extract these surfaces from scanned range images, register them spatially, and combine them into a single surface representing the articular cartilage from which volume, area, and thickness can be computed. We validated the laser scanning approach using a knee model which was covered with a synthetic articular cartilage model and compared the computed volume against water displacement measurements. Using this method, the volume of articular cartilage in five sets of cadaver knees was compared to volume estimates obtained from segmentation of MR images.

Three-dimensional (3D) and four-dimensional (4D) imaging of dynamic structures is a rapidly developing area of research in medical imaging. Non-rigid registration plays an important role for the analysis of these datasets. In this paper we will show some of the work of our group using non-rigid registration techniques for the detection of temporal changes such as growth in brain MR images. We will also show how non-rigid registration can be used to analyze the motion of the heart from cardiac MR images.

We consider the problem of estimating the local accuracy of image registration when no ground truth data is available. The technique is based on a statistical resampling technique called bootstrap. Only the two input images are used, no other data are needed. The general bootstrap uncertainty estimation framework described here is in principle applicable to most of the existing pixel based registration techniques. In practice, a large computing power is required. We present experimental results for a block matching method on an ultrasound image sequence for elastography with both known and unknown deformation field.

Nonlinear image registration is a prerequisite for various medical image analysis applications. Many data acquisition protocols suffer from problems due to breathing motion which has to be taken into account for further analysis. Intensity based nonlinear registration is often used to align differing images, however this requires a large computational effort, is sensitive to intensity variations and has problems with matching small structures. In this work a feature-based image registration method is proposed that combines runtime efficiency with good registration accuracy by making use of a fully automatic feature matching and registration approach. The algorithm stages are 3D corner detection, calculation of local ( SIFT ) and global ( Shape Context ) 3D descriptors, robust feature matching and calculation of a dense displacement field. An evaluation of the algorithm on seven synthetic and four clinical data sets is presented. The quantitative and qualitative evaluations show lower runtime and superior results when compared to the Demons algorithm.

Here we present a new image registration algorithm for the alignment of histological sections that combines the ideas of B-spline based elastic registration and consistent image registration, to allow simultaneous registration of images in two directions (direct and inverse). In principle, deformations based on B-splines are not invertible. The consistency term overcomes this limitation and allows registration of two images in a completely symmetric way. This extension of the elastic registration method simplifies the search for the optimum deformation and allows registering with no information about landmarks or deformation regularization. This approach can also be used as the first step to solve the problem of group-wise registration.

Prior knowledge about shape may be quite important for image segmentation. In particular, a number of different methods have been proposed to compute the statistics on a set of training shapes, which are then used for a given image segmentation task to provide the shape prior. In this work, we perform a comparative analysis of shape learning techniques such as linear PCA, kernel PCA, locally linear embedding and propose a new method, kernelized locally linear embedding for doing shape analysis. The surfaces are represented as the zero level set of a signed distance function and shape learning is performed on the embeddings of these shapes. We carry out some experiments to see how well each of these methods can represent a shape, given the training set.

Dynamic emission tomography is a technique used for quantifying the biochemical and physiological processes within the body. For certain neuroimaging applications, like kinetic modelling in positron emission tomography (PET), segmenting the measured data into a fewer number of regions-of-interest (ROI) is an important procedure needed for calculation of regional time-activity curves (TACs). Conventional estimation of regional activities in image domain suffers from substantial errors due to the reconstruction artifacts and segmentation inaccuracies. In this study, we present an approach for separating the dynamic tomographic data directly in the projection space using the least-squares method. Sinogram ROIs are the fractional parts of different tissue types measured at each voxel. Regional TACs can be estimated from the segmented sinogram ROIs, thereby avoiding the image reconstruction step. The introduced technique was validated with the two dynamic synthetic phantoms simulated based on ^11C- and ^18F-labelled tracer distributions. From the quantitative point of view, TAC estimation from the segmented sinograms yielded more accurate results compared to the image-based segmentation.