Search

This paper theoretically analyses the propagation of measurement noise in Electrical Impedance Tomography (EIT) and its influence in image noise. The authors compare the behavior of three backprojection algorithms in front of noise. They present close-form approximated expressions for noise power in EIT. EIT image noise is argued to be gaussian distributed but with spatially variant statistics. There is a linear relation between measurement noise reduction and reduction in image noise. The study was validated with numerical simulation.

Active Shape Models (ASMs), a knowledge-based segmentation algorithm developed by Cootes and Taylor [1995, 1999], have become a standard and popular method for detecting structures in medical images. In ASMs-and various comparable approaches-the model of the object's shape and of its gray-level variations is based the assumption of linear distributions. In this work, we explore a new way to model the gray-level appearance of the objects, using a k-nearest-neighbors (kNN) classifier and a set of selected features for each location and resolution of the Active Shape Model. The construction of the kNN classifier and the selection of features from training images is fully automatic. We compare our approach with the standard ASMs on synthetic data and in four medical segmentation tasks. In all cases, the new method produces significantly better results (p

The decrease in the volume of viable tumor is an indicator for the effect preoperative chemotherapy has on bone tumors. We develop an approach for segmenting dynamic perfusion MR-images into viable tumor, nonviable tumor and healthy tissue. Two cascaded feedforward neural networks are trained to perform the pixel-based segmentation. As features, we use the parameters obtained from a pharmacokinetic model of the tissue perfusion (parametric images). Additional multiscale features that incorporate contextual information are included. Experiments indicate that multiscale blurred versions of the parametric images together with a multiscale formulation of the local image entropy are the most discriminative features.

A method to extract and track the position of a guide wire during endovascular interventions under X-ray fluoroscopy is presented and evaluated. The method can be used to improve guide wire visualization in the low quality fluoroscopic images and to estimate the position of the guide wire in world coordinates. A two-step procedure is utilized to track the guide wire in subsequent frames. First a rough estimate of the displacement is obtained using a template matching procedure. Subsequently, the position of the guide wire is determined by fitting a spline to a feature image in which line-like structures are enhanced. In the optimization step, the influence of the scale at which the feature is calculated is investigated. Also, the feature image is calculated both on the original image and on a preprocessed image in which coherent structures are enhanced. Finally, the influence of explicit endpoint detection is studied. The method is evaluated on 267 frames from 10 sequences. Using the automatic method, the guide wire could be tracked in 96% of the frames, with a greater accuracy than three observers. Endpoint detection improved the accuracy of the tip assessment, which was better than 1.3 mm.

In this paper we introduce the concept of statistical deformation models (SDM) which allow the construction of average models of the anatomy and their variability. SDMs are built by performing a statistical analysis of the deformations required to map anatomical features in one subject into the corresponding features of another subject. We demonstrate the applicability of this new framework to MR images of the brain and show results for the construction of anatomical models from 25 different subjects.

A novel method is introduced that allows for the generation of landmarks for three-dimensional shapes and the construction of the corresponding 3D Active Shape Models (ASM). Landmarking of a set of examples from a class of shapes is achieved by (i) construction of an atlas of the class, (ii) automatic extraction of the landmarks from the atlas, and (iii) subsequent propagation of these landmarks to each example shape via a volumetric elastic deformation procedure. This paper describes in detail the method to generate the atlas, and the landmark extraction and propagation procedures. This technique presents some advantages over previously published methods: it can treat multiple-part structures, and it requires less restrictive assumptions on the structure's topology. The applicability of the developed technique is demonstrated with two examples: CT bone data and MR brain data.

A 3D model-based approach for quantification of vascular morphology from several MRA acquisition protocols was evaluated. Accuracy, reproducibility, and influence of the image acquisition techniques were studied via in vitro experiments with ground truth diameters and the measurements of two expert readers as reference. The performance of the method was similar to or more accurate than the manual assessments and reproducibility was also improved. The methodology was applied to stenosis grading of carotid arteries from CE MRA data. In 11 patients, the approach was compared to manual scores (NASCET criterion) on CE MRA and DSA images, with the result that the model-based technique correlates better with DSA than the manual scores. Spearman’s correlation coefficient was 0.91 (P < 0.001) for the model-based technique and DSA vs. 0.80 and 0.84 (P < 0.001) between the manual scores and DSA. From the results it can be concluded that the approach is a promising objective technique to assess geometrical vascular parameters, including degree of stenosis. Magn Reson Med 45:311‐322, 2001. © 2001 Wiley-Liss, Inc.

In this paper we introduce the concept of statistical deformation models (SDM) which allow the construction of average models of the anatomy and their variability. SDMs are built by performing a statistical analysis of the deformations required to map anatomical features in one subject into the corresponding features in another subject. The concept of SDMs is similar to active shape models (ASM) which capture statistical information about shapes across a population but offers several new advantages over ASMs: Firstly, SDMs can be constructed directly from images such as MR or CT without the need for segmentation which is usually a prerequisite for the construction of active shape models. Instead a non-rigid registration algorithm is used to compute the deformations required to establish correspondences between the reference subject and the subjects in the population class under investigation. Secondly, SDMs allow the construction of an atlas of the average anatomy as well as its variability across a population of subjects. Finally, SDMs take the 3D nature of the underlying anatomy into account by analysing dense 3D deformation fields rather than only the 2D surface shape of anatomical structures. We demonstrate the applicability of this new framework to MR images of the brain and show results for the construction of anatomical models from 25 different subjects.

583 Added value of a statistical atlas-based quantification of motion abnormalities for the prediction of CRT response {#article-title-2} Purpose: The presence of septal flash (SF), a specific inward/outward motion pattern of the septum during the isovolumic contraction period, recently shown up as good predictor of cardiac resynchronization therapy (CRT) outcome. The ability of accurately identifying a SF fully conditions the predictive value of this mechanism. Current SF detection from US images is based on visual inspection only and suffers low reproducibility. Thus, we created an atlas-based statistical method for the automatic quantification of motion abnormalities that we apply to SF detection, and evaluated its predicting capability. Method: We built an atlas of normal motion from 21 healthy volunteers (age 30±5y, 14 male), to which we compared 75 CRT candidates (age 68±9y, 56 male) with left ventricle dyssynchrony. Abnormal motion is quantified locally in apical 4-chamber 2DUS sequences, from a p-value based on a statistical distance to normality. SF is identified using automatic recognition of the inward/outward events on spatiotemporal abnormality maps, as shown in Fig1. CRT response was defined at 6 months follow-up, as increase ≥10% in the 6' walking test or NYHA functional class reduction ≥1pt. Results: Comparison of atlas-based and visual assessment of SF leads to a Cohen's Kappa of 0.54 and an observed agreement of 0.81 (48 SF and 13 without SF). We reached a positive predictive value of CRT outcome of 85%, the visual assessment one being of 83%. Conclusion: We have demonstrated the potential of an atlas-based automatic quantification of motion abnormalities for the understanding of CRT outcome. In particular, we have demonstrated its performance for the prediction of clinical response when the methodology is applied to the identification of SF[⇓][1]. # 584 How reliably can myocardial blood flow be tracked? – a validation of commercial tracking software using computer-generated datasets {#article-title-3} Background: Recent advances in myocardial tracking technology have stimulated attempts to also track contrast enhanced intracavitary blood flow. Little is known, however, how basic imaging parameters (line density, frame rate, contrast bubble density) affect the quality of such tracking results. Our study aimed at investigating this by using simulated echo data sets. Methods: A 3D blood flow field of the entire left ventricle (LV) was calculated for a cardiac cycle based on mitral inflow and LV volume data using Fluent 12.1 (ANSYS Inc., USA). Then, the 3D motion of contrast microbubbles was simulated and 2D B-mode image loops were obtained (f = 4.5MHz; 50° sector angle). After conversion to DICOM, image loops were analyzed using flow tracking software (Qflow, Siemens, Mountainview, USA). Vorticity and amplitude of the resulting in-plane velocity vector field was calculated at different frame rates (227, 113 and 76 fps) and bubble densities (BD) (60, 35 and 18 B/ml) and compared to the ground truth known from the CFD model. Results: Quality of tracking was clearly related to BD and frame rate. Velocities, estimated by tracking, correlated best with the ground truth at 60 B/ml and 227fps (r = 0.61, p < 0.01) and deteriorated with lower settings (e.g. r = 0.25 at 60 B/ml and 76 fps). Estimated averaged vorticity also correlated best with the ground truth at 60 B/ml and 227fps. Interestingly, at low frame rates (76fps), tracking results improved with decreasing BD (r = 0.57 at 60 B/ml vs. r = 0.67 at 18 B/ml). Conclusions: Flow tracking by contrast enhanced echocardiography is feasible. Good tracking requires high frame rates and an optimized bubble density. Currently available tracking software shows acceptable quality of velocity estimates, facilitating the recognition of basic flow patterns, such as vortices[⇓][2]. # 585 Early detection of functional abnormalities in asymptomatic arrhythmogenic right ventricular cardiomyopathy gene carriers using echocardiographic deformation imaging {#article-title-4} Purpose: The first presentation of arrhythmogenic right ventricular cardiomyopathy (ARVC) is often potentially lethal ventricular arrhythmias originating from the right ventricle (RV), typically at a young age. This emphasizes the importance of an early recognition of this disease, for instance in ARVC- family members. The aim of this study is to evaluate the value of tissue deformation imaging to detect subclinical RV functional abnormalities in asymptomatic genotyped carriers of ARVC. Methods: A total of 43 asymptomatic first degree family members of ARVC probands (not fulfilling the diagnosis of ARVC according to the task-force criteria (TF-c)) were prospectively enrolled for echocardiographic examination. In a total of 14 (38.0±13.2 years), a genetic mutation (PKP2) could be identified (others had no mutation or genetic screening). All individuals were age-matched with 4 controls (n = 56, 38.2±12.7 years) undergoing the same echocardiographic evaluation (dimensions, global systolic parameters, visual assessment, and deformation imaging of the RV free wall). Echocardiographic evaluation was performed blinded. Deformation analysis was analyzed blinded to group and findings from the conventional echocardiogram. A peak systolic strain >−18% and/or post-systolic shortening (post-systolic index >15%) in any segment was considered abnormal. Results: No significant differences in baseline characteristics were seen between the groups. RV dimensions in the family group were similar to the controls (RVOT 15.4±2.9 vs. 14.4±1.9 mm/m2, RVIT 18.6±2.6 vs. 19.1±2.6 mm/m2, p=NS). Global systolic parameters were moderately reduced in the family group (TVI-syst 9.1±1.6 vs. 11.1±1.7 cm/s, TAPSE 20.0±3.2 vs. 23.9±2.8 mm, p −18% was seen in 6 family members (43%) and post systolic strain in 10 (71%). Either abnormality was observed in 11 (79%), almost exclusively in the basal segment, and in non of the controls. 2D-strain showed abnormal segments in 8 (57%) of family members and 5 (9%) controls. Conclusion: Echocardiographic deformation imaging detects functional abnormalities in the basal RV segment in almost 80% of asymptomatic ARVC gene carriers. Furthermore, false positive findings in visual assessment (28%) could be prevented since all showed normal deformation values and patterns. [1]: #F1 [2]: #F2

Integrative modeling of cardiac system is important for understanding the complex biophysical function of the heart]. To this end, multimodal cardiovascular imaging plays an important role in providing the computational domain, the boundary/initial conditions, and tissue function and properties. In particular, the incorporation of blood flow in the physiological models can help to simulate the hemodynamic properties and their effects on cardiac function. In this paper, we present a multimodal framework for quantitative and subject-specific analysis of blood flow in the cardiac chambers, including the left ventricle (LV). The 3D geometries of the LV at different time steps are extracted from medical images using an atlas of LV shape. The motion of the myocardium wall is used to extract the moving boundary data of the computational geometry. The data is used as a constraint for the computational fluid dynamics (CFD). An arbitrary Lagrangian–Eulerian (ALE) finite element method (FEM) formulation is used to derive a numerical solution of the transient dynamic equation of the fluid domain. With this method, simulation results describe detailed flow characteristics (such as velocity, pressure and wall shear stress) in the computational domain. The personalized hemodynamic characteristics obtained with the proposed approach can provide clinical value for diagnosis and treatment of abnormalities related to disturbed blood flow such as in myocardial remodeling and aortic sinus lesion formation.

Three dimensional magnetic resonance angiographic images (3D MRA) are routinely inspected using maximum intensity projections (MIP). However, accuracy of stenosis estimates based on projections is limited. Therefore, a method for quantitative 3D MRA is introduced. Linear vessel segments are modeled with a central vessel axis curve coupled to a vessel wall surface. First, the central vessel axis is determined. Subsequently, the vessel wall is segmented using knowledge of the acquisition process. The user interaction to initialize the model is performed in a 3D setting. The method is validated on a carotid bifurcation phantom and also illustrated on patient data.

The multiscale second order local structure of an image (Hessian) is examined with the purpose of developing a vessel enhancement filter. A vesselness measure is obtained on the basis of all eigenvalues of the Hessian. This measure is tested on two dimensional DSA and three dimensional aortoiliac and cerebral MRA data. Its clinical utility is shown by the simultaneous noise and background suppression and vessel enhancement in maximum intensity projections and volumetric displays.

Purpose: To generate accurate and realistic models of coronary artery bifurcations before and after percutaneous coronary intervention (PCI), using information from two image modalities. Because bifurcations are regions where atherosclerotic plaque appears frequently and intervention is more challenging, generation of such realistic models could be of high value to predict the risk of restenosis or thrombosis after stent implantation, and to study geometrical and hemodynamical changes. Methods: Two image modalities have been employed to generate the bifurcation models: computer tomography angiography (CTA) to obtain the 3D trajectory of vessels, and 2D conventional coronary angiography (CCA) to obtain radius information of the vessel lumen, due to its better contrast and image resolution. In addition, CCA can be acquired right before and after the intervention in the operation room; therefore, the combination of CTA and CCA allows the generation of realistic preprocedure and postprocedure models of coronary bifurcations. The method proposed is semiautomatic, based on landmarks manually placed on both image modalities. Results: A comparative study of the models obtained with the proposed method with models manually obtained using only CTA, shows more reliable results when both modalities are used together. The authors show that using preprocedure CTA and postprocedure CCA, realisticmore » postprocedure models can be obtained. Analysis carried out of the Murray's law in all patient bifurcations shows the geometric improvement of PCI in our models, better than using manual models from CTA alone. An experiment using a cardiac phantom also shows the feasibility of the proposed method. Conclusions: The authors have shown that fusion of CTA and CCA is feasible for realistic generation of coronary bifurcation models before and after PCI. The method proposed is efficient, and relies on minimal user interaction, and therefore is of high value to study geometric and hemodynamic changes of treated patients.« less

In surgical robotics, the ability to achieve high levels of autonomy is often limited by the complexity of the surgical scene. Autonomous interaction with soft tissues requires machines able to examine and understand the endoscopic video streams in real-time and identify the features of interest. In this work, we show the first example of spatio-temporal neural networks, based on the U-Net, aimed at segmenting soft tissues in endoscopic images. The networks, equipped with Long Short-Term Memory and Attention Gate cells, can extract the correlation between consecutive frames in an endoscopic video stream, thus enhancing the segmentation’s accuracy with respect to the standard U-Net. Initially, three configurations of the spatio-temporal layers are compared to select the best architecture. Afterwards, the parameters of the network are optimised and finally the results are compared with the standard U-Net. An accuracy of 83.77% ± 2.18% and a precision of 78.42% ± 7.38% are achieved by implementing both Long Short Term Memory (LSTM) convolutional layers and Attention Gate blocks. The results, although originated in the context of surgical tissue retraction, could benefit many autonomous tasks such as ablation, suturing and debridement.

This paper is framed in the field of statistical face analysis. In particular, the problem of accurate segmentation of prominent features of the face in frontal view images is addressed. Our method constitutes an extension of Cootes et al. [6] linear Active Shape Model (ASM) approach, which has already been used in this task [9]. The technique is built upon the development of a non-linear appearance model, incorporating a reduced set of differential invariant features as local image descriptors. These features are invariant to rigid transformations, and a subset of them is chosen by Sequential Feature Selection (SFS) for each landmark and resolution level. The new approach overcomes the unimodality and gaussianity assumptions of classical ASMs regarding the distribution of the intensity values across the training set. Validation of the method is presented against the linear ASM and its predecesor, the Optimal Features ASM (OF-ASM) [14] using the AR and XM2VTS databases as testbed.

Realistic models of the muscle fibers in the myocardium improve the understanding and simulation of the bio-mechanical behavior of the heart. Since Diffusion Tensor Imaging (DTI) allows to visualize the fiber structures in the tissues, this modality can be used to build fiber models. In this paper, we propose an automatic method for the analysis of the helix and transverse angles between the fibers and the myocardial wall. It computes automatically the theoretical value of these angles (according to a mathematical model described in the literature) at each voxel of the image as well as the real value based on the DTI acquisition. In addition, new parameters of the mathematical model can be estimated based on our approach to personalize the model for specific data-sets.

We present a numerical solver for the fast conduction system in the heart using both a CPU and a hybrid CPU/GPU implementation. To verify both implementations, we construct analytical solutions and show that the \(L^2\)-error is similar in both implementations and decreases linearly with the spatial step size. Finally, we test the performance of the implementations with networks of varying complexity, where the hybrid implementation is, on average, 5.8 times faster.

A model-based technique for quantitative analysis of three-dimensional magnetic resonance angiography (MRA) is presented. The model consists of a deformable B-spline representation of the central vessel axis and the vessel wall. An efficient interaction mechanism to initialize the models in a three-dimensional setting is devised. Two novel image features are introduced in order to deform the central vessel axis and the vessel wall model respectively. A segmental optimization scheme for deforming the B-spline vessel wall model is introduced. Finally, the paper presents results on clinical contrast enhanced (CE) MRA of 19 carotid arteries. Percentages of stenosis obtained with the authors' technique are compared to those of two experts using caliper measurements on Maximum Intensity Projection (MIP) CE MRA, and Digital Subtraction Angiography (DSA) as standard of reference.

With the exponential growth of clinical data, and the fast development of AI technologies, researchers are facing unprecedented challenges in managing data storage, scalable processing, and analysis capabilities for heterogeneous multisourced datasets. Beyond the complexity of executing data-intensive workflows over large-scale distributed data, the reproducibility of computed results is of paramount importance to validate scientific discoveries. In this paper, we present MULTIX, a cross-domain research-oriented platform, designed for collaborative and reproducible science. This cloud-based framework simplifies the logistical challenges of implementing data analytics and AI solutions by providing pre-configured environments with ad-hoc scalable computing resources and secure distributed storage, to efficiently build, test, share and reproduce scientific pipelines. An exemplary use-case in the area of cardiac image analysis will be presented together with the practical application of the platform for the analysis of ~20.000 subjects of the UK-Biobank database.

The increasing volume of data describing human/ndisease processes and the growing complexity of understanding, managing, and sharing such data presents a huge challenge for clinicians and medical researchers. This paper presents the/n@neurIST system, which provides an infrastructure for biomedical research while aiding clinical care, by bringing together heterogeneous data and complex processing and computing services. Although @neurIST targets the investigation and treatment of cerebral aneurysms, the system’s architecture is generic enough that it could be adapted to the treatment of other diseases./nInnovations in @neurIST include confining the patient data pertaining to aneurysms inside a single environment that offers clinicians/nthe tools to analyze and interpret patient data and make use of knowledge-based guidance in planning their treatment. Medical/nresearchers gain access to a critical mass of aneurysm related data due to the system’s ability to federate distributed information/nsources. A semantically mediated grid infrastructure ensures that both clinicians and researchers are able to seamlessly access and/nwork on data that is distributed across multiple sites in a secure way in addition to providing computing resources on demand for/nperforming computationally intensive simulations for treatment planning and research. This work was supported by the framework of the @neurIST Integrated Project, which is cofinanced by the European Commission under Contract IST-027703.

The article explores the application of machine learning approach to detect both single-vessel and multivessel coronary artery disease from X-ray angiography. Since the interpretation of coronary angiography images requires interventional cardiologists to have considerable training, our study is aimed at analysing, training, and assessing the potential of the existing object detectors for classifying and detecting coronary artery stenosis using angiographic imaging series. 100 patients who underwent coronary angiography at the Research Institute for Complex Issues of Cardiovascular Diseases were retrospectively enrolled in the study. To automate the medical data analysis, we examined and compared three models (SSD MobileNet V1, Faster-RCNN ResNet-50 V1, FasterRCNN NASNet) with various architecture, network complexity, and a number of weights. To compare developed deep learning models, we used the mean Average Precision (mAP) metric, training time, and inference time. Testing results show that the training/inference time is directly proportional to the model complexity. Thus, Faster-RCNN NASNet demonstrates the slowest inference time. Its mean inference time per one image made up 880 ms. In terms of accuracy, FasterRCNN ResNet-50 V1 demonstrates the highest prediction accuracy. This model has reached the mAP metric of 0.92 on the validation dataset. SSD MobileNet V1 has demonstrated the best inference time with the inference rate of 23 frames per second.

International audience; This paper presents a new method for classifying surface datavia spectral representations of shapes. Our approach benefits classificationproblems that involve data living on surfaces, such as in cortical parcellation.For instance, current methods for labeling cortical points into surface parcelsoften involve a slow mesh deformation toward pre-labeled atlases, requiringas much as 4 hours with the established FreeSurfer. This may burden neurosciencestudies involving region-specific measurements. Learning techniquesoffer an attractive computational advantage, however, their representation ofspatial information, typically defined in a Euclidean domain, may be inadequatefor cortical parcellation. Indeed, cortical data resides on surfaces thatare highly variable in space and shape. Consequently, Euclidean representationsof surface data may be inconsistent across individuals. We proposeto fundamentally change the spatial representation of surface data, by exploitingspectral coordinates derived from the Laplacian eigenfunctions ofshapes. They have the advantage over Euclidean coordinates, to be geometryaware and to parameterize surfaces explicitly. This change of paradigm,from Euclidean to spectral representations, enables a classifier to be applieddirectly on surface data via spectral coordinates. In this paper, we decide tobuild upon the successful Random Decision Forests algorithm and improve itsspatial representation with spectral features. Our method, Spectral Forests,is shown to significantly improve the accuracy of cortical parcellations overstandard Random Decision Forests (74% versus 28% Dice overlaps), and produceaccuracy equivalent to FreeSurfer in a fraction of its time (23 secondsversus 3 to 4 hours).

ISBN 978-3-319-24552-2; International audience; Functional Arterial Spin Labeling (fASL) MRI can provide a quantitative measurement of changes of cerebral blood flow induced by stimulation or task performance. fASL data is commonly analysed using a general linear model (GLM) with regressors based on the canonical hemodynamic response function. In this work, we consider instead a joint detection-estimation (JDE) framework which has the advantage of allowing the extraction of both task-related perfusion and hemodynamic responses not restricted to canonical shapes. Previous JDE attempts for ASL have been based on computer intensive sampling (MCMC) methods. Our contribution is to provide a comparison with an alternative variational expectation-maximization (VEM) algorithm on synthetic and real data.

International audience; Recent work on brain tumor growth modeling for glioblas-toma using reaction-diffusion equations suggests that the diffusion coefficient and the proliferation rate can be related to clinically relevant information. However, estimating these parameters is difficult due to the lack of identifiability of the parameters, the uncertainty in the tumor segmen-tations, and the model approximation, which cannot perfectly capture the dynamics of the tumor. Therefore, we propose a method for conducting the Bayesian personalization of the tumor growth model parameters. Our approach estimates the posterior probability of the parameters, and allows the analysis of the parameters correlations and uncertainty. Moreover , this method provides a way to compute the evidence of a model, which is a mathematically sound way of assessing the validity of different model hypotheses. Our approach is based on a highly parallelized implementation of the reaction-diffusion equation, and the Gaussian Process Hamiltonian Monte Carlo (GPHMC), a high acceptance rate Monte Carlo technique. We demonstrate our method on synthetic data, and four glioblastoma patients. This promising approach shows that the infiltration is better captured by the model compared to the speed of growth.

International audience; Medical image segmentation is often a prerequisite for clinical applications. As an ill-posed problem, it leads to uncertain estimations of the region of interest which may have a significant impact on downstream applications, such as therapy planning. To quantify the uncertainty related to image segmentations, a classical approach is to measure the effect of using various plausible segmentations. In this paper, a method for producing such image segmentation samples from a single expert segmentation is introduced. A probability distribution of image segmentation boundaries is defined as a Gaussian process, which leads to segmentations that are spatially coherent and consistent with the presence of salient borders in the image. The proposed approach outperforms previous generative segmentation approaches, and segmentation samples can be generated efficiently. The sample variability is governed by a parameter which is correlated with a simple DICE score. We show how this approach can have multiple useful applications in the field of uncertainty quantification, and an illustration is provided in radiotherapy planning.

International audience; Cardiac cine MRI with 3D isotropic resolution is challenging as it requires efficient data acquisition and motion management. It is proposed to use a 2D balanced SSFP (steady-state free precession) sequence rather than its 3D version as it provides better contrast between blood and tissue. In order to obtain 3D isotropic images, 2D multi-slice datasets are acquired in different orientations (short axis, horizontal long axis and vertical long axis) while the patient is breathing freely. Image reconstruction is performed in two steps: (i) a motion-compensated reconstruction of each image stack corrects for nonrigid cardiac and respiratory motion; (ii) a super-resolution (SR) algorithm combines the three motion-corrected volumes (with low resolution in the slice direction) into a single volume with isotropic resolution. The SR reconstruction was implemented with two regularization schemes including a conventional one (Tikhonov) and a feature-preserving one (Beltrami). The method was validated in 8 volunteers and 10 patients with breathing difficulties. Image sharpness, as assessed by intensity profiles and by objective metrics based on the structure tensor, was improved with both SR techniques. The Beltrami constraint provided efficient denoising without altering the effective resolution.