Maria Lorenzo-Valdés

Atlas-Based Segmentation and Tracking of 3D Cardiac MR Images Using Non-rigid Registration

We propose a novel method for fully automated segmentation and tracking of the myocardium and left and right ventricles (LV and RV) using 4D MRimages. The method uses non-rigid registration to elastically deform a cardiac atlas built automatically from 14 normal subjects. The registration yields robust performance and is particularly suitable for processing a sequence of 3D images in a cardiac cycle. Transformations are calculated to obtain the deformations between images in a sequence. The registration algorithm aligns the cardiac atlas to a subject specific atlas of the sequence generated with the transformations. The method relates images spatially and temporally and is suitable for measuring regional motion and deformation, as well as for labelling and tracking specific regions of the heart. In this work experiments for the registration, segmentation and tracking of a cardiac cycle are presented on nine MRI data sets. Validation against manual segmentations and computation of the correlation between manual and automatic tracking and segmentation on 141 3D volumes were calculated. Results show that the procedure can accurately track the left ventricle (r=0.99), myocardium (r=0.98) and right ventricle (r=0.96). Results for segmentation are also obtained for left ventricle (r=0.92), myocardium (r=0.82) and right ventricle (r=0.90).

Segmentation of 4D Cardiac MR Images Using a Probabilistic Atlas and the EM Algorithm

In this paper an automatic atlas-based segmentation algorithm for 4D cardiac MR images is proposed. The algorithm is based on the 4D extension of the expectation maximisation (EM) algorithm. The EM algorithm uses a 4D probabilistic cardiac atlas to estimate the initial model parameters and to integrate a-priori information into the classification process. The probabilistic cardiac atlas has been constructed from the manual segmentations of 3D cardiac image sequences of 14 subjects. It provides space and time-varying probability maps for the left and right ventricle, the myocardium, and background structures such as the liver, stomach, lungs and skin. In addition to the probabilistic cardiac atlas, the segmentation algorithm incorporates spatial and temporal contextual information by using 4D Markov Random Fields (MRF). Validation against manual segmentations and computation of the correlation between manual and automatic segmentation on 249 3D volumes were calculated. Results show that the procedure can successfully segment the left ventricle (LV) (r=0.95), myocardium (r=0.83) and right ventricle (RV) (r=0.91).

Comparison of Cardiac Motion Across Subjects Using Non-rigid Registration

We present a novel technique that enables a direct quantitative comparison of cardiac motion derived from 4D MR image sequences to be made either within or across patients. This is achieved by registering the images that describe the anatomy of both subjects and then using the computed transformation to map the motion fields of each subject into the same coordinate system. The motion fields are calculated by registering each of the frames in a sequence of tagged short-axis MRI images to the end-diastolic frame using a non-rigid registration technique based on multi-level free-form deformations. The end-diastolic untagged short-axis images acquired shortly after the tagged images were obtained are registered using non-rigid registration to determine an inter-subject mapping, which is used to transform the motion fields of one of the subjects into the coordinate system of the other, which is thus our reference coordinate system. The results show the transformed myocardial motion fields of a series of volunteers, and clearly demonstrate the potential of the proposed technique.

3D/4D Cardiac Segmentation Using Active Appearance Models, Non-rigid Registration, and the Insight Toolkit

We describe the design of a statistical atlas-based 3D/4D cardiac segmentation system using a combination of active appearance models (AAM) and statistical deformation models with the Insight Toolkit as an underlying implementation framework. Since the original AAM approach was developed for 2D applications and makes use of manually set landmarks its extension to higher dimensional data sets cannot be easily achieved. We therefore apply the idea of statistical deformation models to AAMs and use a deformable registration step for establishing point-to-point correspondences. An evaluation of the implemented system was performed by segmenting the left ventricle cavity, myocardium and right ventricle of ten cardiac MRI and ten CT datasets. The comparison of automatic and manual segmentations showed encouraging results with a mean segmentation error of 2.2±1.1 mm. We conclude that the combination of a non-rigid registration step with the statistical analysis concepts of the AAM is both feasible and useful and allows for its application to 3D and 4D data.

Construction of a cardiac motion atlas from MR using non-rigid registration

In this paper we present a technique for constructing a cardiac motion atlas using the myocardial motion fields derived from 4D MR image sequences of a series of subjects. This is achieved by transforming the motion field of each subject into a the coordinate system of a reference subject, and then averaging the transformed fields to give a vector field representing the mean motion of the heart. The motion fields of each subject are calculated by registering each of the frames in the sequence of tagged short-axis and long-axis MRI images to the end-diastolic frame using a non-rigid registration technique based on multi-level free-form deformations. The end-diastolic untagged short-axis images of each subject, which are acquired shortly after the tagged images, are registered to the corresponding image of a designated reference subject using non-rigid registration to determine reference-subject mappings, which are then used to transform the corresponding motion fields into that of the reference subject. Finally, the mean transformed motion field is calculated to give the cardiac motion atlas.

Building a 4D atlas of the cardiac anatomy and motion using MR imaging

In this paper we describe the construction of 4D atlas of human heart using cardiac MR imaging. This probabilistic atlas captures the cardiac anatomy and function of a healthy heart. In order to build the atlas we have acquired tagged as well as untagged MR image sequences from 11 healthy volunteers. The untagged MR image sequences for each subject are segmented and then mapped into a common reference coordinate system using a novel spatio-temporal registration algorithm to produce a 4D probabilistic model of the cardiac anatomy. In addition, the tagged MR image sequences are used to derive motion fields between the end-diastolic and the end-systolic frames which describe myocardial contraction patterns in each subject. These motion fields are also mapped into our spatio-temporal reference coordinate system to produce a 4D statistical model of cardiac function.

Spatio-temporal alignment of 4D cardiac MR images

A 4D registration method for the spatio-temporal alignment of cardiac MR image sequences has been developed. The registration algorithm has the ability not only to correct any spatial misalignment between the image sequences but also any temporal misalignment which maybe the result of differences in the cardiac cycle between subjects and differences in the temporal acquisition parameters. The algorithm uses a 4D transformation model which is separated into a spatial and a temporal component: the spatial component is a 3D affine transformation which corrects for any misalignment between the two image sequences. The temporal component uses an affine transformation which corrects the temporal misalignment caused by differences in the initial acquisition offset and length of the two cardiac cycles. The method was applied to seven cardiac MR image sequences from healthy volunteers. The registration was qualitatively evaluated by visual inspection and quantitatively by measuring the volume difference and overlap of anatomical regions between the sequences. The results indicated a significant improvement in the spatio-temporal alignment of the sequences.

Non-rigid registration of cardiac MR: application to motion modelling and atlas-based segmentation

Three-dimensional (3D) and four-dimensional (4D) imaging of the heart is a rapidly developing area of research in medical imaging. Recent advances in development of MR imaging for fast spatio-temporal cardiac imaging have led to an increased interest in the use of MR imaging for functional analysis of the cardiovascular system. Segmentation of left and right ventricle and modelling of myocardial motion provide important information for quantitative functional analysis. In this paper we show how non-rigid registration techniques can be used to solve these problems in cardiac MR images.

Determination of Aortic Distensibility Using Non-rigid Registration of Cine MR Images

A novel method for the estimation of areas in 2D MR images of the aorta is presented. The method uses spatio-temporal non-rigid registration in order to obtain the 2D deformation fields of the vessels during the cardiac cycle. This is accomplished by aligning all time frames in the image sequence simultaneously to the first one. The determinant of the Jacobian of the 2D deformation fields are then computed to obtain the expansion (or contraction) at each time frame, with respect to the first time frame. By using 3D splines, the method exploits the relation between time frames in order to obtain continuous and smooth distensibility measurements throughout the cardiac cycle. Validation was carried out with MR images of the aorta. Experiments for the registration and estimation of areas in the aorta are presented in 60 data sets corresponding to three different sections of the aorta (proximal, mid and distal) in 20 different subjects, where each set consisted of 17 to 38 time frames. Manually estimated areas are compared to the areas estimated automatically in 8 data sets where the average error is 2.3% of the area manually obtained.

Segmentation of cardiac MR images using - the EM algorithm with a 4D probabilistic atlas and a global connectivity filter

In this paper an automatic atlas-based segmentation algorithm for 4D cardiac MR images is described. The algorithm is based on the 4D extension of the expectation maximisation (EM) algorithm. The EM algorithm uses a 4D probabilistic cardiac atlas to estimate the initial model parameters and to integrate spatially-varying a-priori information into the classification process. It provides space and time-varying probability maps for the left and right ventricle, the myocardium, and background structures such as the liver, stomach, lungs and skin. The segmentation algorithm also incorporates spatial and temporal contextual information by using 4D Markov Random Fields (MRF). After the classification, the largest connected component of each structure is used as a global connectivity filter that improves the results significantly, especially for the myocardium. Validation against manual segmentations and computation of the correlation between manual and automatic segmentation on 249 3D volumes were calculated. Results show that the procedure can successfully segment the left ventricle (LV) (r=0.96), myocardium (r=0.92) and right ventricle (RV) (r=0.92).