Catálogo de publicaciones - libros

Non-rigid image registration is an ill-posed yet challenging problem due to its supernormal high degree of freedoms and inherent requirement of smoothness. Graph-cuts method is a powerful combinatorial optimization tool which has been successfully applied into image segmentation and stereo matching. Under some specific constraints, graph-cuts method yields either a global minimum or a local minimum in a strong sense. Thus, it is interesting to see the effects of using graph-cuts in non-rigid image registration. In this paper, we formulate non-rigid image registration as a discrete labeling problem. Each pixel in the source image is assigned a displacement label (which is a vector) indicating which position in the floating image it is spatially corresponding to. A smoothness constraint based on first derivative is used to penalize sharp changes in displacement labels across pixels. The whole system can be optimized by using the graph-cuts method via alpha-expansions. We compare 2D and 3D registration results of our method with two state-of-the-art approaches. It is found that our method is more robust to different challenging non-rigid registration cases with higher registration accuracy.

Recent developments in freehand 3D ultrasound (US) have shown how image registration and speckle decorrelation methods can be used for 3D reconstruction instead of relying on a tracking device. Estimating elevational separation between untracked US images using speckle decorrelation is error prone due to the uncertainty that plagues the correlation measurements. In this paper, using maximum entropy estimation methods, the uncertainty is directly modeled from the calibration data normally used to estimate an average decorrelation curve. Multiple correlation measurements can then be fused within a maximum likelihood estimation framework in order to reduce the drift in elevational pose estimation over large image sequences. The approach is shown to be effective through empirical results on simulated and phantom US data.

Dynamic enhancement causes serious problems for registration of contrast enhanced breast MRI, due to variable uptakes of agent on different tissues or even same tissues in the breast. We present an iterative optimization algorithm to de-enhance the dynamic contrast-enhanced breast MRI and then register them for avoiding the effects of enhancement on image registration. In particular, the spatially varying enhancements are modeled by a Markov Random Field, and estimated by a locally smooth function with boundaries using a graph cut algorithm. The de-enhanced images are then registered by conventional B-spline based registration algorithm. These two steps benefit from each other and are repeated until the results converge. Experimental results show that our two-step registration algorithm performs much better than conventional mutual information based registration algorithm. Also, the effects of tumor shrinking in the conventional registration algorithms can be effectively avoided by our registration algorithm.

There exists a large body of literature on shape matching and registration in medical image analysis. However, most of the previous work is focused on matching particular sets of features—point-sets, lines, curves and surfaces. In this work, we forsake specific geometric shape representations and instead seek probabilistic representations—specifically Gaussian mixture models—of shapes. We evaluate a closed-form distance between two probabilistic shape representations for the general case where the mixture models differ in variance and the number of components. We then cast non-rigid registration as a deformable density matching problem. In our approach, we take one mixture density onto another by deforming the component centroids via a thin-plate spline (TPS) and also minimizing the distance with respect to the variance parameters. We validate our approach on synthetic and 3D arterial tree data and evaluate it on 3D hippocampal shapes.

We present a new method for the fast and robust computation of information theoretic similarity measures for alignment of multi-modality medical images. The proposed method defines a non-uniform, adaptive sampling scheme for estimating the entropies of the images, which is less vulnerable to local maxima as compared to uniform and random sampling. The sampling is defined using an octree partition of the template image, and is preferable over other proposed methods of non-uniform sampling since it respects the underlying data distribution. It also extends naturally to a multi-resolution registration approach, which is commonly employed in the alignment of medical images. The effectiveness of the proposed method is demonstrated using both simulated MR images obtained from the BrainWeb database and clinical CT and SPECT images.

A fundamental problem when computing statistical shape models is the determination of correspondences between the instances of the associated data set. Often, homologies between points that represent the surfaces are assumed which might lead to imprecise mean shape and variability results. We propose an approach where exact correspondences are replaced by evolving correspondence probabilities. These are the basis for a novel algorithm that computes a generative statistical shape model. We developed an unified MAP framework to compute the model parameters (’mean shape’ and ’modes of variation’) and the nuisance parameters which leads to an optimal adaption of the model to the set of observations. The registration of the model on the instances is solved using the Expectation Maximization - Iterative Closest Point algorithm which is based on probabilistic correspondences and proved to be robust and fast. The alternated optimization of the MAP explanation with respect to the observation and the generative model parameters leads to very efficient and closed-form solutions for (almost) all parameters. Experimental results on brain structure data sets demonstrate the efficiency and well-posedness of the approach. The algorithm is then extended to an automatic classification method using the k-means clustering and applied to synthetic data as well as brain structure classification problems.

In this paper we propose a weakly supervised learning algorithm for appearance models based on the minimum description length (MDL) principle. From a set of training images or volumes depicting examples of an anatomical structure, correspondences for a set of landmarks are established by group-wise registration. The approach does not require any annotation. In contrast to existing methods no assumptions about the topology of the data are made, and the topology can change throughout the data set. Instead of a continuous representation of the volumes or images, only sparse finite sets of interest points are used to represent the examples during optimization. This enables the algorithm to efficiently use distinctive points, and to handle texture variations robustly. In contrast to standard elasticity based deformation constraints the MDL criterion accounts for systematic deformations typical for training sets stemming from medical image data. Experimental results are reported for five different 2D and 3D data sets.

In this paper, we present a new segmentation evaluation method that can simulate radiologist’s subjective assessment of 3D tumor segmentation in CT images. The method uses a new metric defined as a linear combination of a set of commonly used objective metrics. The weighing parameters of the linear combination are determined by maximizing the rank correlation between radiologist’s subjective rating and objective measurements. Experimental results on 93 lesions demonstrate that the new composite metric shows better performance in segmentation evaluation than each individual objective metric. Also, segmentation rating using the composite metric compares well with radiologist’s subjective evaluation. Our method has the potential to facilitate the development of new tumor segmentation algorithms and assist large scale segmentation evaluation studies.

We propose an extended graph-shifts algorithm for image segmentation and labeling. This algorithm performs energy minimization by manipulating a dynamic hierarchical representation of the image. It consists of a set of moves occurring at different levels of the hierarchy where the types of move, and the level of the hierarchy, are chosen automatically so as to maximally decrease the energy. Extended graph-shifts can be applied to a broad range of problems in medical imaging. In this paper, we apply extended graph-shifts to the detection of pathological brain structures: (i) segmentation of brain tumors, and (ii) detection of multiple sclerosis lesions. The energy terms in these tasks are learned from training data by statistical learning algorithms. We demonstrate accurate results, precision and recall in the order of 93%, and also show that the algorithm is computationally efficient, segmenting a full 3D volume in about one minute.

Minimally invasive surgery in orthopedic field is considered to be a challenging problem with a milling robot. One objective of this study is to minimize collision of the cutting tool with soft tissue. The authors have developed a robot with redundant axis to avoid the collision so far. Some important components are modeled based on physical requirements, and a geometric optimization approach based on the model has been also proposed to improve performance. In this paper, a protective mechanism to cover the non-working part of the cutting edge is proposed to avoid soft tissue damage. Hardware and software have been developed for this application and the effectiveness of this technique was evaluated with urethane bone.