Functional Imaging and Modeling of the Heart

We propose a strategy to perform population-based personalization of a model, to overcome the limits of case-based personalization for generating virtual populations from models that include randomness. We formulate the problem as matching the synthetic and real populations by minimizing the Kullback-Leibler divergence between their distributions. As an analytical formulation of the models is complex or even impossible, the personalization is addressed by a gradient-free method: the CMA-ES algorithm, whose relevance was demonstrated for the case-based personalization of complex biomechanical cardiac models. The algorithm iteratively adapts the covariance matrix which in our problem encodes the distribution of the synthetic data.We demonstrate the feasibility of this approach on two simple geometrical models of myocardial infarction, in 2D, to better focus on the relevance of the personalization process. Our strategy is able to reproduce the distribution of 2D myocardial infarcts from the segmented late Gadolinium images of 123 subjects with acute myocardial infarction.

Cardiac magnetic resonance (CMR) is an important diagnostic imaging modality in cardiovascular medicine. Estimation of myocardium motion derived from CMR scans is routinely used to measure the cardiac mechanics. However, tracking a rapidly moving organ can be compromised by artefacts or impaired image quality. To assess how in-plane image resolution and slice sampling in short and long axis scans impact errors in motion tracking, we utilised retrospective gated cardiac computed tomography (CCT) imaging as a surrogate groundtruth for motion estimation across 10 clinical datasets, since these scans have a higher isotropic resolution than CMR and the ability to capture full 3D motion. In our work, the left atrial (LA) and ventricular (LV) cavities were first delineated, and then reconstructed short and long axis images were used in a non-rigid registration method with optimised hyperparameters to track endocardial motion. Finally, global and regional functions in the form of area, circumferential, and longitudinal strains were computed. Our findings showed that tracking LA was more sensitive than LV to changes in the in-plane resolution and magnitude of strain was robust to resolution changes in short axis images, when correlated with the groundtruth (r: 0.87–0.99, $$R^2$$ R 2 : 0.75–0.98). We also found that 9 short axis slices could capture the motion of LV almost as accurately as 36 slices captured in long axis (r: 0.89 vs. 0.90, $$R^2$$ R 2 : 0.80 vs. 0.80), illustrating that the cardiac mechanics measured by short axis scans are more likely to be robust to image artefacts and reconstruction parameters.

2D Echocardiography is a popular and cost-efficient tool for cardiac dysfunction diagnosis. Automatic solutions that could effectively and efficiently analyse cardiac functions are highly desired in clinical situations. Segmentation and motion tracking are two important techniques to extract useful cardiac indexes, such as left ventricle ejection fraction (LVEF), global longitudinal strain (GLS), etc. However, these tasks are non-trivial since ultrasound images usually suffer from poor signal-to-noise ratio, boundary ambiguity and out of view problem. In this paper, we explore how to introduce shape constraints from global, regional and pixel level into a baseline U-Net model for better segmentation and landmark tracking. Our experiments show that all the three propositions perform similarly as the baseline model in terms of geometrical scores, while our pixel-level model, which uses a multi-class contour loss, reduces segmentation outliers and improves the tracking accuracy of 3 landmarks used for GLS computation. With appropriate augmentation techniques, our models also show a good generalisation performance when testing on a larger unseen cohort.

Abdominal aortic aneurysm (AAA) growth is correlated with rupture risk, but predicting either AAA growth or rupture remains challenging. Global aneurysm geometric properties have been linked with elevated peak AAA wall stress when using finite element analysis (FEA) and may predict AAA growth. We used a machine learning model to evaluate whether image-derived geometric parameters, calculated both globally and locally over the surface of the aneurysm can predict local AAA wall growth, avoiding material property assumptions used in FEA. Sequential CTAs one year apart were collected from 10 patients with AAAs. The luminal and aortic wall were segmented in patient’s baseline CTA. In order to calculate local geometric properties, each baseline AAA was divided into 64 regions to define regional geometric aneurysm characteristics from vertices in that region, and into 1,500 sub-regions in order to define sub-regional geometric characteristics. The global and local (regional and sub-regional) aortic geometric properties were all derived from the images and determined from the aortic segmentation and surface mesh. Local AAA growth between CTAs was determined at the sub-regional level using deformable image registration and was the outcome variable for the model. Patient demographics, as well as the global and local geometric aneurysm properties were used to predict local AAA growth using an eXtreme gradient boosted regression tree using a performance metric of root-mean-square error (RMSE) with 80/20 training to testing split. Mean relative error in predicting maximum AAA growth was 10.5% in the testing set. The most impactful predictors were AAA volume, regional maximum diameter, regional maximum Gaussian surface curvature, regional median aneurysm thickness, and patient age. Removal of local geometric properties from the model increased RMSE from 0.5 to 1.1 and decreased model performance by likelihood test (P = 0.01). Utilizing both global and local aneurysm geometric characteristics better predicts local aortic wall growth in AAAs, avoiding assumptions required using FEA.

Deep learning has demonstrated promise for cardiac magnetic resonance image (MRI) segmentation. However, the performance is degraded when a trained model is applied to previously unseen datasets. In this work, we developed a way to employ a pre-trained model to segment the left ventricle (LV) and quantify LV indices in a new dataset. We trained a U-net with Monte-Carlo dropout on 45 cine MR images and applied the model to 10 subjects from the ACDC dataset. The initial segmentation was refined using a continuous kernel-cut algorithm and the refined segmentation was used to fine-tune the pre-trained U-net for 10 min. This process was iterated several times until convergence and the updated model was used to segment the remaining 90 patients in the ACDC dataset. For the test dataset, we achieved Dice-similarity-coefficient of 0.81 ± 0.12 for LV myocardium and 0.90 ± 0.09 for LV cavity. Algorithm LV indices were strongly correlated with manual results (r = 0.86–0.99, p < 0.0001) with marginal biases of –8.8 g for LV myocardial mass, –0.9 ml for LV end-diastolic volume, –0.2 ml for LV end-systolic volume, –0.7 ml for LV stroke volume, and –0.6% for LV ejection fraction. The proposed approach required 12 min for fine-tuning without requiring manual annotations of the new datasets and 1 s to segment a new image. These results suggest that the developed approach is effective in segmenting a previously unseen cardiac MRI dataset and quantifying LV indices without requiring manual segmentation of the new dataset.

Myocardial Velocity Mapping Cardiac MR (MVM-CMR) can be used to measure global and regional myocardial velocities with proved reproducibility. Accurate left ventricle delineation is a prerequisite for robust and reproducible myocardial velocity estimation. Conventional manual segmentation on this dataset can be time-consuming and subjective, and an effective fully automated delineation method is highly in demand. By leveraging recently proposed deep learning-based semantic segmentation approaches, in this study, we propose a novel fully automated framework incorporating a 3D-UNet backbone architecture with Embedded multichannel Attention mechanism and LSTM based Recurrent neural networks (RNN) for the MVM-CMR datasets (dubbed 3D-EAR segmentor). The proposed method also utilises the amalgamation of magnitude and phase images as input to realise an information fusion of this multichannel dataset and exploring the correlations of temporal frames via the embedded RNN. By comparing the baseline model of 3D-UNet and ablation studies with and without embedded attentive LSTM modules and various loss functions, we can demonstrate that the proposed model has outperformed the state-of-the-art baseline models with significant improvement.

Coronary computed tomography angiography (CCTA) provides detailed anatomical information on all chambers of the heart. Existing segmentation tools can label the gross anatomy, but addition of application-specific labels can require detailed and often manual refinement. We developed a U-Net based framework to i) extrapolate a new label from existing labels, and ii) parcellate one label into multiple labels, both using label-to-label mapping, to create a desired segmentation that could then be learnt directly from the image (image- to-label mapping). This approach only required manual correction in a small subset of cases (80 for extrapolation, 50 for parcellation, compared with 260 for initial labels). An initial 6-label segmentation (left ventricle, left ventricular myocardium, right ventricle, left atrium, right atrium and aorta) was refined to a 10-label segmentation that added a label for the pulmonary artery and divided the left atrium label into body, left and right veins and appendage components. The final method was tested using 30 cases, 10 each from Philips, Siemens and Toshiba scanners. In addition to the new labels, the median Dice scores were improved for all the initial 6 labels to be above 95% in the 10-label segmentation, e.g. from 91% to 97% for the left atrium body and from 92% to 96% for the right ventricle. This method provides a simple framework for flexible refinement of anatomical labels. The code and executables are available at cemrg.com .

Retrospective gated cardiac computed tomography (CCT) images can provide high contrast and resolution images of the heart throughout the cardiac cycle. Feature tracking in retrospective CCT images using the temporal sparse free-form deformations (TSFFDs) registration method has previously been optimised for the left ventricle (LV). However, there is limited work on optimising nonrigid registration methods for feature tracking in the left atria (LA). This paper systematically optimises the sparsity weight (SW) and bending energy (BE) as two hyperparameters of the TSFFD method to track the LA endocardium from end-diastole (ED) to end-systole (ES) using 10-frame retrospective gated CCT images. The effect of two different control point (CP) grid resolutions was also investigated. TSFFD optimisation was achieved using the average surface distance (ASD), directed Hausdorff distance (DHD) and Dice score between the registered and ground truth surface meshes and segmentations at ES. For baseline comparison, the configuration optimised for LV feature tracking gave errors across the cohort of 0.826 ± 0.172 mm ASD, 5.882 ± 1.524 mm DHD, and 0.912 ± 0.033 Dice score. Optimising the SW and BE hyperparameters improved the TSFFD performance in tracking LA features, with case specific optimisations giving errors across the cohort of 0.750 ± 0.144 mm ASD, 5.096 ± 1.246 mm DHD, and 0.919 ± 0.029 Dice score. Increasing the CP resolution and optimising the SW and BE further improved tracking performance, with case specific optimisation errors of 0.372 ± 0.051 mm ASD, 2.739 ± 0.843 mm DHD and 0.949 ± 0.018 Dice score across the cohort. We therefore show LA feature tracking using TSFFDs is improved through a chamber-specific optimised configuration.

The ellipse can be an appropriate geometry for aorta cross-section fitting on the lumen contour. However, in some regions of the aorta, such as the Sinuses of Valsalva, this approximation can suffer of a relatively high error. Thus, some authors use closed polynomial curves for a better representation of the cross section. This paper presents a detailed comparison between the use of an elliptic cross section model and a spline based model with different number of knots. We use a cohort of 32 thoracic aorta geometries (segmented triangle meshes), obtained using CT scan in the mesosystole phase of the cardiac cycle, for the assessment of both methods. We use the root mean squared error of the fitting of the studied methods to quantify their accuracy. As expected, the spline based model improves the fitting accuracy of the elliptic one and specially in complex aorta cross-sections. However, we have observed that with a high number of knots some cross sections may show high error values due to the adaption of the function to noise.

High frame rate (HFR) speckle tracking echocardiography (STE) assesses myocardial function by quantifying motion and deformation at high temporal resolution. Our lab recently proposed a two-step HFR STE methodology based on 1-D cross-correlation [1]. Even if it was proved to be accurate for a global assessment of the mid-wall myocardial motion, an impaired sensitivity to motion and lower feasibilities for the apical regions gave higher errors. Thus, the aim of this study was to improve the speckle tracking algorithm by improving the tracking quality in the apical region specifically while preserving tracking quality on the other segments as well as picking up movement transmurally, i.e. the endocardial, mid-wall and epicardial motion. Hereto, the original algorithm was modified by improving robustness and accuracy of the lateral motion estimation. Simulation results showed that the proposed changes resulted in a significantly lower error in the estimation of the global longitudinal strain (GLS). Moreover, these improvements were mostly visible in the apical region (e.g. strain error 4.64 ± 2.65% vs 1.19 ± 0.71% for the septum) and prioritized the local movements resulting in lower error ranges between contours ([0.2–0.46]% vs [0.33–1.66]%). Finally, for a qualitative comparison, a preliminary in vivo acquisition was performed.

Deep learning methods have reached state-of-the-art performance in cardiac image segmentation. Currently, the main bottleneck towards their effective translation into clinics requires assuring continuous high model performance and segmentation results. In this work, we present a novel learning framework to monitor the performance of heart segmentation models in the absence of ground truth. Formulated as an anomaly detection problem, the monitoring framework allows deriving surrogate quality measures for a segmentation and allows flagging suspicious results. We propose two different types of quality measures, a global score and a pixel-wise map. We demonstrate their use by reproducing the final rankings of a cardiac segmentation challenge in the absence of ground truth. Results show that our framework is accurate, fast, and scalable, confirming it is a viable option for quality control monitoring in clinical practice and large population studies.

The lack of standardized pipelines for image processing has prevented the application of deep learning (DL) techniques for the segmentation of the aorta in phase-contrast enhanced magnetic resonance angiography (PC-MRA). Furthermore, large, well-curated and annotated datasets, which are needed to create DL-based models able to generalize, are rare. We present the adaptation of the popular nnU-net DL framework to automatically segment the aorta in 4D flow MRI-derived angiograms. The resulting segmentations in a large database ( $$> 300$$ > 300 cases) with normal cases and examples of different pathologies of the aorta provided from a single centre were excellent after post-processing (Dice score of 0.944). Subsequently, we explored the generalisation of the trained network in a small dataset of images (around 20 cases) acquired in a different hospital with another scanner. Without domain adaptation, only with a model trained with the large dataset, the obtained results were substantially worst than with adding a few cases of the small dataset (Dice scores of 0.61 vs 0.86, respectively). The obtained results created good quality segmentations of the aorta in 4D flow MRI, which can later be post-processed to assess blood flow patterns, similarly than with manual annotations. However, advanced domain adaptation schemes are very important in 4D flow MRI due to the large differences in image characteristics between different vendor scanners available in multiple centers.

Segmentation of cardiac magnetic resonance (cMR) images is often the first step necessary to compute common diagnostic biomarkers, such as myocardial mass and left ventricle (LV) ejection fraction. Often image segmentation and analysis require significant, time-consuming user input. Machine learning has been increasingly adopted to automatically segment medical images to lessen the burden on image segmentation and image analysis for model construction and validation. In this work we present a multi-step machine learning approach to segment short axis cMR images based on a heart locator and the weighted average of 2D and 2D++ UNets. The presence of a heart locator led to more accurate results and allowed to increase the neural network training batch size. Finally, the obtained segmentations are post-processed using spline interpolation and the Loop scheme to generate left ventricular endocardial and epicardial surfaces at the end of diastole and end of systole.

Cardiac diffusion tensor magnetic resonance imaging (cDTI) allows estimating the aggregate cardiomyocyte architecture in healthy subjects and its remodeling as a result of cardiac disease. In this study, cDTI was used to quantify microstructural changes occurring in swine (N = 7) six to ten weeks after myocardial infarction. Each heart was extracted and imaged ex vivo with $$1\,\mathrm {mm}$$ 1 mm isotropic spatial resolution. Microstructural changes were quantified in the border zone and infarct region by comparing diffusion tensor invariants – fractional anisotropy (FA), mode, and mean diffusivity (MD) – radial diffusivity, and diffusion tensor eigenvalues with the corresponding values in the remote myocardium. MD and radial diffusivity increased in the infarct and border regions with respect to the remote myocardium ( $$p< 0.01$$ p < 0.01 ). In contrast, FA and mode decreased in the infarct and border regions ( $$p< 0.01$$ p < 0.01 ). Diffusion tensor eigenvalues also increased in the infarct and border regions, with a larger increase in the secondary and tertiary eigenvalues.

Parametric cardiac magnetic resonance techniques, such as T1 mapping with MOLLI sequences, enable quantitative imaging of tissue properties, which can be a powerful tool in the diagnosis and prognosis of different cardiovascular conditions. Conventional parameter estimation methods are often based on pixel-wise curve fitting, ignoring spatial information. In this study, an automatic pipeline based on a spatially constrained deep learning algorithm is presented, to compute the myocardial T1 values from MOLLI sequences, within clinically acceptable computation times. The proposed algorithm is based on the DeepBLESS architecture, modified to incorporate local spatial information and regularization. The model was trained on a large database of clinical MOLLI cases (from 186 patients), showing promising preliminary results, obtaining T1 maps faster and more robust to noise.

The sheetlet angle is a biomarker that can describe the local arrangement of the sheetlets making the heart wall tissue. It could be of major interest for the analysis of cardiac function. In this preliminary study, we use the skeleton method to measure the local sheetlet angle in two human left ventricular transparietal tissue samples. The samples were imaged using synchrotron X-rays phase contrast micro-tomography and reconstructed in 3D with isotropic voxels of 3.5 $$\upmu $$ μ m edges. We extract the skeleton from each sheetlet. Next, we scan the skeleton, voxel by voxel, performing principal component analysis (PCA) in a sliding cubic working window (WW) of 112 $$\upmu $$ μ m edges. The tertiary eigenvector of the PCA provides the sheetlet angle, step by step, along the skeleton. We show that the proposed methodology is able to provide the transmural distribution of sheetlets angle at a scale of 112 $$\upmu $$ μ m and locally to investigate the spatial evolution of the angle along the sheetlets. We compare the results obtained at a larger scale with the Fourier-based method.

Pulmonary arterial hypertension (PAH) imposes a pressure overload on the right ventricle (RV), leading to myofiber hypertrophy and remodeling of the extracellular collagen fiber network. While the macroscopic behavior of healthy and post-PAH RV free wall (RVFW) tissue has been studied previously, the mechanical microenvironment that drives remodeling events in the myofibers and the extracellular matrix (ECM) remains largely unexplored. We hypothesize that multiscale computational modeling of the heart, linking cellular-scale events to tissue-scale behavior, can improve our understanding of cardiac remodeling and better identify therapeutic targets. We have developed a high-fidelity microanatomically realistic model of ventricular myocardium, combining confocal microscopy techniques, soft tissue mechanics, and finite element modeling. We match our microanatomical model to the tissue-scale mechanical response of previous studies on biaxial properties of RVFW and examine the local myofiber-ECM interactions to study fiber-specific mechanics at the scale of individual myofibers. Through this approach, we determine that the interactions occurring at the tissue scale can be accounted for by accurately representing the geometry of the myofiber-collagen arrangement at the micro scale. Ultimately, models such as these can be used to link cellular-level adaptations with organ-level adaptations to lead to the development of patient-specific treatments for PAH.

In this study an equine heart of a 13-year-old Belgian Warmblood is scanned using diffusion tensor magnetic resonance imaging (DT-CMR) to perform quantitative analysis on myocyte orientation in the left and right ventricular free wall (LVFW and RVFW) as well as in the interventricular septum (IVS). Transmural helical angle profiles are obtained at basal, mid-ventricular and apical height. This equine heart with high transmural DT-CMR resolution revealed that the transmural helix angle range in the RVFW is approximately twice as large as the one in the LVFW. Moreover, the transmural helix angle distributions in the IVS could be separated into an RV and LV side with gradients in fiber angle orientation that were comparable to the RVFW and LVFW, respectively.

Cardiomyocytes branch and interconnect with one another, providing important redundancy for propagation of electro-chemical signals. Despite this, cardiomyocyte branching structure remains poorly understood. Herein, myocardium from spontaneously hypertensive rats (SHR) was imaged using extended volume confocal microscopy. Samples from untreated SHRs ( $$n = 2$$ n = 2 ) were compared with SHRs undergoing ACE inhibitor treatment ( $$n = 2$$ n = 2 ). From these image-stacks the cardiomyocyte network (center-lines and branches) were manually tracked. The frequency of cardiomyocyte branching was calculated and these branching frequencies were compared according to spatial position within a myocardial sheetlet. ACE inhibitor treatment resulted in significantly reduced total cardiomyocyte branching compared with untreated SHR myocardium at 24 mo (0.49 ± 0.04 vs 1.07 ± 0.15 branches per 100 $$\upmu $$ μ m, $$P = 0.020$$ P = 0.020 ). Cardiomyocytes on the sheetlet-surface branched more frequently within their respective cell-layer (0.59 ± 0.07 branches per 100 $$\upmu $$ μ m) compared with cardiomyocytes in the sheetlet interior (0.29 ± 0.12 branches per 100 $$\upmu $$ μ m). The cardiomyocytes in the sheetlet interior exhibited more frequent between-layer branching (0.56 ± 0.06 branches per 100 $$\upmu $$ μ m) compared to sheetlet-surface cardiomyocytes (0.17 ± 0.03 branches per 100 $$\upmu $$ μ m). The ratio of within-layer to between-layer branching was significantly greater at the surface layer compared with the interior layer (3.93 ± 1.06 vs 0.47 ± 0.16, $$P=0.018$$ P = 0.018 ). This proof-of-concept study demonstrates an approach to measuring branching cardiomyocyte networks and shows the spatial heterogeneity of cardiomyocyte branching.

Diffusion tensor imaging (DTI) is a promising imaging technique to non-invasively study diffusion properties and fiber structures of myocardial tissues. Previous studies have investigated the influence of noise or angular resolution independently on the estimation of diffusion tensors in DTI. However, the joint influence of these two factors in DTI remains unclear. In this paper, we propose to systematically study the joint influence of angular resolutions and noise levels on the estimation of diffusion tensors and tensor-derived fractional anisotropy (FA) and mean diffusivity (MD). The results showed that, as expected, given a certain noise level and sufficient acquisition time, the accuracy of diffusion tensor, FA and MD all increase as the angular resolution. Moreover, when the angular resolution reached a certain value, further increasing the number of angular resolutions has little effect on the estimation of diffusion tensor, FA and MD. Also, both the mean and variance of FA or MD decrease as the angular resolution increases. For an imposed acquisition time, increasing the angular resolution reduces SNR of DW images. When fixing SNR, higher angular resolution can be obtained at the expense of longer acquisition time. These findings suggest the necessity of an optimized trade-off when designing DTI protocols.

Cardiac tagged MR images allow for deformation fields to be measured in the heart by tracking the motion of tag lines throughout the cardiac cycle. Machine learning (ML) algorithms enable accurate and robust tracking of tag lines. Herein, the use of a massive synthetic physics-driven training dataset with known ground truth was used to train an ML network to enable tracking any number of points at arbitrary positions rather than anchored to the tag lines themselves. The tag tracking and strain calculation methods were investigated in a computational deforming cardiac phantom with known (ground truth) strain values. This enabled both tag tracking and strain accuracy to be characterized for a range of image acquisition and tag tracking parameters. The methods were also tested on in vivo volunteer data. Median tracking error was< 0.26 mm in the computational phantom, and strain measurements were improved in vivo when using the arbitrary point tracking for a standard clinical protocol.

Myocardial shape and deformation are two relevant descriptors for the study of cardiac function and can undergo strong interactions depending on diseases. Manifold learning provides low dimensional representations of these high-dimensional descriptors, but the choice of normalization can strongly affect the analysis. Besides, whether the shape normalization should include a scale factor is still an open question.In this paper, we investigate the influence of normalization choices on the study of the interactions between cardiac shape and deformation using Multiple Manifold Learning, a dimensionality reduction method that considers inter- and intra-descriptors link between samples. By studying the main variations of two different shape normalizations (one including scaling, the other one not) we observed that the scaled normalization concentrates variations of a given physiological characteristic on only one mode. The influence of the associated choice of the deformation normalization was evaluated by quantifying differences between the estimated low-dimensional spaces (one for each choice against a combination of both), revealing potential analysis biases that may arise depending on such choices.

Cardiomyopathy is the leading cause of mortality in boys with Duchenne muscular dystrophy (DMD). Left ventricular (LV) peak mid-wall circumferential strain (Ecc) is a sensitive early biomarker for evaluating both the subtle and variable onset and the progression of cardiomyopathy in pediatric subjects with DMD. Cine Displacement Encoding with Stimulated Echoes (DENSE) has proven sensitive to changes in Ecc, but its reproducibility has not been reported in a pediatric cohort or a DMD cohort. The objective was to quantify the intra-observer repeatability, and intra-exam and inter-observer reproducibility of global and regional Ecc derived from cine DENSE in DMD patients (N = 10) and age- and sex-matched controls (N = 10). Global and regional Ecc measures were considered reproducible in the intra-exam, intra-observer, and inter-observer comparisons. Intra-observer repeatability was highest, followed by intra-exam reproducibility and then inter-observer reproducibility. The smallest detectable change in Ecc was 0.01 for the intra-observer comparison, which is below the previously reported yearly decrease of 0.013 ± 0.015 in Ecc in DMD patients.

Quantification of regional cardiac function is a central goal of cardiology. Multiple methods, such as Coherent Point Drift (CPD) and Simultaneous Subdivision Surface Registration (SiSSR), have been used to register meshes to the endocardial surface. However, these methods do not distinguish between cardiac chambers during registration, and consequently the mesh may “slip” across the interface between two structures during contraction, resulting in inaccurate regional functional measurements. Here, we present Multilabel-SiSSR (M-SiSSR), a novel method for registering a “labeled” cardiac mesh (with each triangle assigned to a cardiac structure). We compare our results to the original, label-agnostic version of SiSSR and find both a visual and quantitative improvement in tracking of the mitral valve plane.

Patient-specific left ventricle (LV) myocardial models have the potential to be used in a variety of clinical scenarios for improved diagnosis and treatment plans. Cine cardiac magnetic resonance (MR) imaging provides high resolution images to reconstruct patient-specific geometric models of the LV myocardium. With the advent of deep learning, accurate segmentation of cardiac chambers from cine cardiac MR images and unsupervised learning for image registration for cardiac motion estimation on a large number of image datasets is attainable. Here, we propose a deep leaning-based framework for the development of patient-specific geometric models of LV myocardium from cine cardiac MR images, using the Automated Cardiac Diagnosis Challenge (ACDC) dataset. We use the deformation field estimated from the VoxelMorph-based convolutional neural network (CNN) to propagate the isosurface mesh and volume mesh of the end-diastole (ED) frame to the subsequent frames of the cardiac cycle. We assess the CNN-based propagated models against segmented models at each cardiac phase, as well as models propagated using another traditional nonrigid image registration technique. Additionally, we generate dynamic LV myocardial volume meshes at all phases of the cardiac cycle using the log barrier-based mesh warping (LBWARP) method and compare them with the CNN-propagated volume meshes.

We propose a multiscale spatio-temporal graph convolutional network (MST-GCN) approach to learn the left ventricular (LV) motion patterns from cardiac MR image sequences. The MST-GCN follows an encoder-decoder framework. The encoder uses a sequence of multiscale graph computation units (MGCUs). The myocardial geometry is represented as a graph. The network models the internal relations of the graph nodes via feature extraction at different scales and fuses the feature across scales to form a global representation of the input cardiac motion. Based on this, the decoder employs a graph-based gated recurrent unit (G-GRU) to predict future cardiac motion. We show that the MST-GCN can automatically quantify the spatio-temporal patterns in cardiac MR that characterise cardiac motion. Experiments are performed on mid-ventricular short-axis view cardiac MR image sequence from the UK Biobank dataset. We compare the performance of cardiac motion prediction of the proposed method with ten different architectures and parameter settings. Experiments show that the proposed method inputting node positions and node velocities with multiscale graphs achieves the best performance with a mean squared error of 0.25 pixel between the ground truth node locations and our prediction. We also show that the proposed method can estimate a number of motion-related metrics, including endocardial radii, thickness and strain which are useful for regional LV function assessment.

Accurate and efficient quantification of cardiac motion offers promising biomarkers for non-invasive diagnosis and prognosis of structural heart diseases. Cine cardiac magnetic resonance imaging remains one of the most advanced imaging tools to provide image acquisitions needed to assess and quantify in-vivo heart kinematics. The majority of cardiac motion studies are focused on human data, and there remains a need to develop and implement an image-registration pipeline to quantify full three-dimensional (3D) cardiac motion in mice where ideal image acquisition is challenged by the subject size and heart rate and the possibility of traditional tagged imaging is hampered. In this study, we used diffeomorphic image registration to estimate strains in the left ventricular wall in two wild-type mice and one diabetic mouse. Our pipeline resulted in a continuous and fully 3D strain map over one cardiac cycle. The estimation of 3D regional and transmural variations of strains is a critical step towards identifying mechanistic biomarkers for improved diagnosis and phenotyping of structural left heart diseases including heart failure with reduced or preserved ejection fraction.

Heart failure is known to be associated with substantial changes in the mechanical properties of the heart muscle. Biomechanical parameters, such as myocardial stiffness, have the potential to help clinicians diagnose and determine and monitor treatment options. The impact of inter-observer variability of geometric modelling on the estimation of passive myocardial stiffness has not yet been systematically investigated. We aimed to examine the sensitivity of myocardial stiffness estimates with respect to inter-observer geometric model variability. Twenty-four subjects (5 controls, 19 patients with heart failure) underwent left heart catheterisation and cardiovascular magnetic resonance (CMR) imaging. Three expert analysts independently constructed three-dimensional geometric models of the left ventricle (LV), which were used to estimate myocardial stiffness using finite element simulations that combined cine CMR data and LV pressure measurements. Bland-Altman analysis was used to assess the inter-observer effects on the reproducibility of myocardial stiffness. The inter-observer variations were ±7.69 mL/m $$^2$$ 2 and ±9.78 g/m $$^2$$ 2 for the LV end-diastolic volume and mass indices, respectively. Meanwhile, the variability ranged by up to ±0.51 kPa for inter-observer analysis on the estimated intrinsic myocardial stiffness values. Findings from this pilot study highlight the importance of the accuracy of image-based geometric modelling when estimating myocardial stiffness.

In this work, we use a Finite Element model of the left ventricular (LV) mechanics to assess the sensitivity of strains to fiber orientation, modeled using both helix and transverse angles using a 5-parameter rule-based model. The ranges for the five parameters represent the variability within the human population, as inferred from DTI measurements found in literature. End-systolic Green-Lagrange strains with respect of the end-diastolic state were expressed according to the local circumferential (c), longitudinal (l) and transmural (t) direction. The results show that the transverse angle is more influential than the helix angle and that the strain components most affected by fiber orientation are $$E_{ct}$$ E ct ( $$38\pm 18\%$$ 38 ± 18 % ), and $$E_{cl}$$ E cl ( $$25\pm 8\%$$ 25 ± 8 % ). The most influential parameters are found to be $$t_u^0$$ t u 0 , describing the longitudinal offset of the transverse angle, and $$h_v^0$$ h v 0 , describing the transmural offset of the helix angle. Instead, $$h_v^1$$ h v 1 , describing the transmural slope of the helix angle, resulted to be the least influential parameter.

MRI-driven computational modeling is increasingly used to simulate in vivo cardiac mechanical behavior and estimate subject-specific myocardial stiffness. However, in vivo validation of these estimates is exceedingly difficult due to the lack of a known ground-truth in vivo myocardial stiffness. We have developed 3D-printed heart phantoms of known myocardium-mimicking stiffness and MRI relaxation properties and incorporated the heart phantoms within a highly controlled MRI-compatible setup to simulate in vivo diastolic filling. The setup enables the acquisition of experimental data needed to evaluate myocardial stiffness using computational constitutive modeling: phantom geometry, loading pressures, boundary conditions, and filling strains. The pressure-volume relationship obtained from the phantom setup was used to calibrate an in silico model of the heart phantom undergoing simulated diastolic filling. The model estimated stiffness was compared to a ground-truth stiffness obtained from uniaxial tensile testing. Ultimately, the setup is designed to enable extensive validation of MRI and FEM-based myocardial stiffness estimation frameworks.

Perivascular support and tethering are likely relevant factors in vascular mechanics and one of the possible causes of deformational heterogeneity of the aortic wall. Besides these effects, the thoracic aorta interacts with the heart and other moving tissues and organs. We propose a generalized approach to model the effect of aortic interactions with static and dynamic perivascular structures. Periaortic interactions are modeled as a heterogeneous Elastic Foundation Boundary Condition (EFBC). This is implemented in the Finite Element model as a collection of unidimensional springs attached to the adventitial surface and a movable opposite end. An optimization algorithm iterates over the material constants and EFBC parameters to fit the simulated nodal displacements or the aortic wall to patient-specific DENSE MRI-derived displacements. We hypothesize that the adventitial load distribution that replicates the in vivo motion and deformation of the aorta is representative of the actual periaortic interactions. We study 3 aortic locations: the distal aortic arch, the descending thoracic aorta, and the infrarenal abdominal aorta. Our method reproduced the in vivo DENSE MRI-derived displacements with a median error below 30% of the pixel-size resolution (1.2–1.6 mm). The resulting average adventitial load is circumferentially and axially heterogeneous and ranged between 30 and 60% of the luminal pressure-pulse depending on the local nature of the periaortic interaction. Adequate modeling of periaortic interactions may bring a better understanding of its role in the normal and pathological function of the aorta in vivo.

Growth and Remodelling (G&R) processes are typical responses to changes in the heart’s loading conditions. The most frequent types of growth in the left ventricle (LV) are thought to involve growth parallel to (eccentric) or perpendicular to (concentric) the fiber direction. However, hypertrophic cardiomyopathy (HCM), a genetic mutation of the sarcomeric proteins, exhibits heterogeneous patterns of growth and fiber disarray despite the absence of clear changes in loading conditions. Previous studies have predicted cardiac growth due to increased overload in the heart [7, 12, 23] as well as modelled inverse G&R post-treatment [1, 14]. Since observed growth patterns in HCM are more complex than standard models of hypertrophy in the heart, fewer studies focus on the geometric changes in this pathological case. By adapting established kinematic growth tensors for the standard types of hypertrophy in an isotropic and orthotropic material model, the paper aims to identify different factors which contribute to the heterogeneous growth patterns observed in HCM. Consequently, it was possible to distinguish that fiber disarray alone does not appear to induce the typical phenotypes of HCM. Instead, it appears that an underlying trigger for growth in HCM might be a consequence of factors stimulating isotropic growth (e.g., inflammation). Additionally, morphological changes in the septal region resulted in higher amounts of incompatibility, evidenced by increased residual stresses in the grown region.

Fiber orientation is a major factor in the determination of end-systolic strains within models of cardiac mechanics. Unfortunately, direct patient-specific acquisition of fiber orientation is not readily available nowadays in the clinic. As an alternative, we propose to use the Reduced Order Unscented Kalman Filter to estimate rule-based fiber orientation parameters from end-systolic wall strains that can be obtained using more traditional imaging methodologies. We address the estimation of fiber orientation in the physiological left ventricle, where end-systolic strains were generated in-silico using a 12-parameter rule-based fiber model. The estimation process focused on the determination of the three most influential parameters of an imperfect 5-parameter rule-based fiber model. Our results show that these three fiber parameters can be estimated within an average deviation of 6 $$^{\circ }$$ ∘ from a combination of three end-systolic strains even when the initial guess for each estimated parameter was set 10 $$^{\circ }$$ ∘ away from the ground truth value.

The helix orientated fibres in the ventricular wall modulate the cardiac electromechanical functions. Experimental data of the helix angle through the ventricular wall have been reported from histological and image-based methods, exhibiting large variability. It is, however, still unclear how this variability influences electrocardiographic characteristics and mechanical functions of human hearts, as characterized through computer simulations. This paper investigates the effects of the range and transmural gradient of the helix angle on electrocardiogram, pressure-volume loops, circumferential contraction, wall thickening, longitudinal shortening and twist, by using state-of-the-art computational human biventricular modelling and simulation. Five models of the helix angle are considered based on in vivo diffusion tensor magnetic resonance imaging data. We found that both electrocardiographic and mechanical biomarkers are influenced by these two factors, through the mechanism of regulating the proportion of circumferentially-orientated fibres. With the increase in this proportion, the T-wave amplitude decreases, circumferential contraction and twist increase while longitudinal shortening decreases.

When combining cardiovascular magnetic resonance imaging (CMR) with pressure catheter measurements, the acquired image and pressure data need to be synchronized in time. The time offset between the image and pressure data depends on a number of factors, such as the type and settings of the MR sequence, duration and shape of QRS complex or the type of catheter, and cannot be typically estimated beforehand. In the present work we propose using a biophysical heart model to synchronize the left ventricular (LV) pressure and volume (P-V) data. Ten patients, who underwent CMR and LV catheterization, were included. A biophysical model of reduced geometrical complexity with physiologically substantiated timing of each phase of the cardiac cycle was first adjusted to individual patients using basic morphological and functional indicators. The pressure and volume waveforms simulated by the patient-specific models were then used as templates to detect the time offset between the acquired ventricular pressure and volume waveforms. Time-varying ventricular elastance was derived from clinical data both as originally acquired as well as when time-synchronized, and normalized with respect to end-systolic time and maximum elastance value ( $$E^N_\text {orig}(t)$$ E orig N ( t ) , $$E^N_\text {t-syn}(t)$$ E t-syn N ( t ) , respectively). $$E^N_\text {t-syn}(t)$$ E t-syn N ( t ) was significantly closer to the experimentally obtained $$E^N_\text {exp}(t)$$ E exp N ( t ) published in the literature (p < 0.05, $$L^2$$ L 2 norm). The work concludes that the model-driven time-synchronization of P-V data obtained by catheter measurement and CMR allows to generate high quality P-V loops, which can then be used for clinical interpretation.

Aortic stenosis (AS) is the most common acquired heart valve disease in the developed world. Traditional methods of grading AS have relied on the measurement of aortic valve area and transvalvular pressure gradient. Recent research has highlighted the existence of AS variants that do not meet classic criteria for severe AS such as low-flow, low-gradient AS. With the development of sophisticated multi-scale computational models, investigation into the left ventricular (LV) biomechanics of AS offers new insights into the pathophysiology that may guide treatment decisions surrounding AS. Building upon our prior study entailing LV-aortic coupling where AS conditions were applied to the idealized geometry of the Living Heart Human Model, we now describe the first patient-specific adaptation of the model to a case of low flow, low gradient AS. EKG-gated cardiac computed tomography images were segmented to provide surfaces to which the generic Living Heart model was adapted. The model was coupled to a lumped-parameter circulatory system; it was then calibrated to patient clinical data from echocardiography/cardiac catheterization with strong correlation (simulation versus clinical measurement): ascending aorta systolic pressure: 109 mmHg vs 116 mmHg, ascending aorta diastolic pressure 50 mmHg vs 45 mmHg, LV systolic pressure: 118 mmHg vs 128 mmHg, peak transvalvular gradient: 9 mmHg vs 12 mmHg, LV ejection fraction: 23% vs 25%. This work illustrates how the Living Heart Human Model geometry can be efficiently adapted to patient-specific parameters, enabling future biomechanics investigations into the LV dysfunction of AS.

A well-established treatment option for advanced heart failure is the implantation of a ventricular assist device (VAD) in the left heart. In over one quarter of patients, however, failure of the right ventricle (RV) occurs shortly after implantation, with a paucity of options for RV failure management in this clinical context. A possible treatment for RV failure is the application of regional mechanical support to the free surface of the RV. Here, we investigate the effect of this treatment using a multiscale finite element model. We discuss a trade-off between hemodynamic benefits and biomechanical effects of simulated interventions with respect to the complex dynamics of RV contraction. Specifically, we report on timing of support with respect to the cardiac cycle, duration of applied force, and force profile distribution. Insights from these preliminary studies can be informative in the rational design of RV-specific mechanical support solutions.

In vivo pressure-volume loops (PVLs) are the gold standard measurement to assess ventricular function. We developed a pipeline to integrate hemodynamic measurements with real-time three-dimensional (3D) echocardiographic data to construct in vivo PVLs for 25 post-heart transplant patients. We then evaluated left ventricular diastolic function for these patients by calculating chamber stiffness from a cubic polynomial fit of the diastolic pressure-volume relationships (PVR). We examined the ability of a well-established mathematical (Klotz) model to predict the patient-specific diastolic PVRs. We found that beat-to-beat variation in hemodynamic measurement was typical for this group of patients, which resulted in mean ± standard deviation end-diastolic chamber stiffness estimates of 0.75 ± 0.40 mmHg/ml. The cubic polynomial fits of the individual diastolic PVRs resulted in much smaller errors (0.25 ± 0.01 mmHg) compared to those associated with the Klotz predicted diastolic PVRs (4.0 ± 0.27 mmHg), which provided a poor representation of the in vivo diastolic PVRs. The proposed framework enables the temporal alignment between hemodynamic and 3D imaging data to produce in vivo PVLs that can be used not only to quantify global ventricular function, but also to estimate mechanical properties of the myocardium.

Omecamtiv mecarbil (OM) is a cardiac myosin activator developed as a treatment of heart failure. OM acts on cross-bridge formation without disrupting intracellular calcium homeostasis. OM effects are extensively characterised both in vitro and in vivo yet how these mechanistically translate from the sarcomere to whole-heart function is not fully understood. We employed a 3D biventricular contraction model of a healthy rat heart that was fitted to anatomic, structural, and hemodynamic and volumetric functional data. The model incorporates pre-load, after-load, fibre orientation, passive material properties, anatomy, calcium transients, and thin and thick filament dynamics. We identified 4 sarcomere model parameters that reflect cross-bridge behavior. Gaussian process emulators (GPEs) were trained to map these parameters to pressure- and volume-based indexes of left ventricular (LV) function. We constrained the 4-parameter space using preclinical OM data, either (1) in vivo whole-heart hemodynamics data (using the Bayesian history matching technique), or (2) in vitro force-pCa measurements. The OM-compatible sarcomere parameter space from case (1) was used to directly calculate force-pCa curves, while the one resulting from case (2) was mapped to the LV indexes using the trained GPEs. We found that our mapping from LV features to force-pCa and vice versa was in agreement with experimental data. In addition, our simulations supported the latest evidence that OM indirectly alters thin filament calcium sensitivity. Our work demonstrates how quantitative mapping from cellular to whole-organ level can be used to improve our understanding of drug action mechanisms.

The full characterization of three-dimensional (3D) mechanical behaviour of myocardium is essential in understanding their function in health and disease. The hierarchical structure of myocardium results in their highly anisotropic mechanical behaviors, with the spatial variations in fiber structure giving rise to heterogeneity. The optimal set of loading paths has been used to estimate the constitutive parameters of myocardium using a novel numerical-experimental approach with full 3D kinematically controlled (triaxial) experiments [1, 2]. Due to the natural variations in soft tissue structures, the mechanical behaviors of myocardium can vary dramatically within the same organ. To alleviate the associated computational costs for obtaining responses of myocardium under a range of loading conditions with a given realization of structure, we developed a neural network-based method integrated with finite elements. The boundary conditions were parameterized. The neural network generated a corresponding trial solution of the underling hyperelasticity problem for each boundary condition. Thus, the neural network approximated the parameter-to-state map. A physics-informed approach was used to train the neural network. Due to their learnability characteristics, the neural network was able to predict solutions for a range of boundary conditions with given individual specimen fiber structures. The neural network was validated with finite element solutions. This method will provide efficient and robust computational models for clinical evaluation to improve patient outcomes.

Ischemic mitral regurgitation (IMR) is manifested by the inability of the mitral valve (MV) to form a completed sealed shape, which is induced by rapidly impairing contractile function of acute myocardial infarction (MI). Mitral valve repair with undersized ring annuloplasty is currently the preferred treatment strategy for IMR. However, the overall persistence and recurrence rate of moderate or severe IMR within 12 months of surgery has been consistently reported as high, which is a direct consequence of adverse left ventricle (LV) remodeling after MI. In this study, we developed a detailed finite element model with coupled left ventricle-mitral valve structure including mitral valve leaflets, chordae tendineae (CT), papillary muscles, and myocardium. In addition, this model was consisted of high fidelity structure segmented from image data, a novel structural constitutive model of MV leaflets and mechanical properties of CT measured using an in-vitro mechanical testing in an integrated computational modeling framework. Discrepancy of strain mapping has been found between in-silico model and in-vivo strain analysis and including pre-strain of mitral valve leaflets in in-silico model was necessary to have more agreement with in-vivo data. Our findings suggests our LV-MV model is capable of predicting IMR results by shutting down regional contractility and pre-strain should be incorporated into future LV-MV model for more accuracy.

Cardiovascular data of 8 patients with Multisystem Inflammatory Syndrome in Children (MIS-C), caused by an aberrant reaction of immune system to the SARS-CoV-2 coronarovirus, were retrospectively analyzed by using patient-specific biomechanical modeling. The first goal was to increase the understanding of the pathophysiology of the cardiovascular involvement in MIS-C, during the inpatient stay at Intensive Care Unit and after discharge from the hospital. Secondly, hypothetical action of various types of pharmacological therapy was tested in silico using the created patient-specific models, aiming to contribute into the optimal pharmacological management during the acute stage of MIS-C.

In this work we present a coupled electromechanical model of the heart for patient-specific simulations, and in particular cardiac resynchronisation therapy. To this end, we propose a fast fully autonomous and flexible pipeline to generate and optimise the data required to run the mechanical simulation. After the meshing of the biventricular segmentation image and the construction of the associated fibres arrangement, we compute the electrical potential propagation in the myocardial tissue from selected onset points on the endocardium. We generate a 12-lead electrocardiogram corresponding to the latter activation map by extrapolating the electrical potential on a virtual torso. This electrical activation is coupled to a mechanical model, featuring a small set of interpretable parameters. We also propose an efficient algorithm to optimise the model parameters, based on patient data. The whole pipeline including a cardiac cycle is computed in 30 min, enabling to use this digital twin for diagnosis and therapy planning.

The aim of this study is to create an automatic framework for sustained ventricular arrhythmia (VA) prediction using cardiac computed tomography (CT) images. We built an image processing pipeline and a deep learning network to explore the relation between post-infarct left ventricular myocardium thickness and previous occurrence of VA. Our pipeline generated a 2D myocardium thickness map (TM) from the 3D imaging input. Our network consisted of a conditional variational autoencoder (CVAE) and a classifier model. The CVAE was used to compress the TM into a low dimensional latent space, then the classifier utilised the latent variables to predict between healthy and VA patient. We studied the network on a large clinical database of 504 healthy and 182 VA patients. Using our method, we achieved a mean classification accuracy of $$75\% \pm 4$$ 75 % ± 4 on the testing dataset, compared to $$71\% \pm 4$$ 71 % ± 4 from the classification using the classical left ventricular ejection fraction (LVEF).

Accuracy of estimating the heart’s electrical activity with Electrocardiographic Imaging (ECGI) is challenging due to using an error-prone physics-based model (forward model). While getting better results than the traditional numerical methods following the underlying physics, modern deep learning approaches ignore the physics behind the electrical propagation in the body and do not allow the use of patient-specific geometry. We introduce a deep-learning-based ECGI framework capable of understanding the underlying physics, aware of geometry, and adjustable to patient-specific data. Using a variational autoencoder (VAE), we uncover the forward model’s parameter space, and when solving the inverse problem, these parameters will be optimized to reduce the errors in the forward model. In both simulation and real data experiments, we demonstrated the ability of the presented framework to provide accurate reconstruction of the heart’s electrical potentials and localization of the earliest activation sites.

Cardiac electrophysiology models achieved good progress in simulating cardiac electrical activity. However, it is still challenging to leverage clinical measurements due to the discrepancy between idealised models and patient-specific conditions. In the last few years, data-driven machine learning methods have been actively used to learn dynamics and physical model parameters from data. In this paper, we propose a principled deep learning approach to learn the cardiac electrophysiology dynamics from data in the presence of scars in the cardiac tissue slab. We demonstrate that this technique is indeed able to reproduce the transmembrane potential dynamics in situations close to the training context. We then focus on evaluating the ability of the trained networks to generalize outside their training domain. We show experimentally that our model is able to generalize to new conditions including more complex scar geometries, multiple signal onsets and various conduction velocities.

Electrocardiographic imaging (ECGI) is an effective tool for noninvasive diagnosis of a range of cardiac dysfunctions. ECGI leverages a model of how cardiac bioelectric sources appear on the torso surface (the forward problem) and uses recorded body surface potential signals to reconstruct the bioelectric source (the inverse problem). Solutions to the inverse problem are sensitive to noise and variations in the body surface potential (BSP) recordings such as those caused by changes or errors in cardiac position. Techniques such as signal averaging seek to improve ECGI solutions by incorporating BSP signals from multiple heartbeats into an averaged BSP with a higher SNR to use when estimating the cardiac bioelectric source. However, signal averaging is limited when it comes to addressing sources of BSP variability such as beat to beat differences in the forward solution. We present a novel joint inverse formulation to solve for the cardiac source given multiple BSP recordings and known changes in the forward solution, here changes in the heart position. We report improved ECGI accuracy over signal averaging and averaged individual inverse solutions using this joint inverse formulation across multiple activation sequence types and regularization techniques with measured canine data and simulated heart motion. Our joint inverse formulation builds upon established techniques and consequently can easily be applied with many existing regularization techniques, source models, and forward problem formulations.

Computing a physiologically accurate electrocardiogram (ECG) is one of the key outcomes of cardiac electrophysiology (EP) simulations. Indeed, the simulated ECG serves as a validation, may be the target for optimization in inverse EP problems, and in general allows to link simulation results to clinical ECG data. Several approaches are available to compute the ECG corresponding to an EP simulation. Lead field approaches are commonly used to compute ECGs from cardiac EP simulations using the Monodomain or Eikonal models. A coupled passive conductor model is instead common when the full Bidomain model is adopted. An approach based on solving an auxiliary Poisson problem propagating the activation field from the heart surface to the torso surface is also possible, although not commonly described in the literature. In this work, through a series of numerical experiments, we investigate the limits of validity of the different approaches to compute the ECG from simulations based on the Monodomain and Bidomain models. Significant discrepancies are observed between the common lead field and direct ECG approaches in most realistic cases – e.g., when conduction anisotropy is included – while the ECG computed via solution of an auxiliary Poisson problem is similar to the direct ECG approach. We conclude that either the direct ECG or Poisson approach should be adopted to improve the accuracy of the computed ECG.

Despite advances in many of the techniques used in Electrocardiographic Imaging (ECGI), uncertainty remains insufficiently quantified for many aspects of the pipeline. The effect of geometric uncertainty, particularly due to segmentation variability, may be the least explored to date. We use statistical shape modeling and uncertainty quantification (UQ) to compute the effect of segmentation variability on ECGI solutions. The shape model was made with Shapeworks from nine segmentations of the same patient and incorporated into an ECGI pipeline. We computed uncertainty of the pericardial potentials and local activation times (LATs) using polynomial chaos expansion (PCE) implemented in UncertainSCI. Uncertainty in pericardial potentials from segmentation variation mirrored areas of high variability in the shape model, near the base of the heart and the right ventricular outflow tract, and that ECGI was less sensitive to uncertainty in the posterior region of the heart. Subsequently LAT calculations could vary dramatically due to segmentation variability, with a standard deviation as high as 126ms, yet mainly in regions with low conduction velocity. Our shape modeling and UQ pipeline presented possible uncertainty in ECGI due to segmentation variability and can be used by researchers to reduce said uncertainty or mitigate its effects. The demonstrated use of statistical shape modeling and UQ can also be extended to other types of modeling pipelines.

Aim: Computer models enabled the study of the fundamental mechanisms responsible for arrhythmias and have the potential of optimizing the clinical procedure for an individual patients pathology. The model complexity and the computational costs affecting computer models hamper their application on a routinely performed procedure. In this work, we aim to design a computer model suitable for clinical time scales. Methods: We adopt a (multi-front) eikonal model that adapts the conduction velocity to the underlying electrophysiology; we describe the diffusion current using a parametrised form, fitted to reproduce the monodomain profile. Results: We simulated spiral waves on a 3D tissue slab and bi-atrial anatomy. We compared the numerical results obtained with a monodomain formulation with those obtained with the new method. Both models provided the same pattern of the spiral waves. While the monodomain model presented slower propagation fronts, the eikonal model captured the correct value of the conduction velocity CV even using a coarse resolution. Conclusion: The eikonal model has the potential of enabling computer-guided procedures when adapts the conduction velocity to the underlying electrophysiology and characterises the diffusion current with a parametrised form.

Patients that have suffered a myocardial infarction are at lifetime high risk for sudden cardiac death (SCD). Personalized 3D computational modeling and simulation can help to find non-invasively arrhythmogenic features of patients’ infarcts, and to provide additional information for stratification and planning of radiofrequency ablation (RFA). Currently, multiscale biophysical models require high computational resources and long simulations times, which make them impractical for clinical environments. In this paper, we develop a phenomenological solver based on cellular automata to simulate cardiac electrophysiology, with results comparable to those of biophysical models. The solver can run simulations in the order of seconds and reproduce rotor dynamics, and ventricular tachycardia in infarcted patients, using a virtual pacing protocol. This model could be use to plan RFA intervention without the time constrains of complex models.

Cardiac smooth muscle cell mathematical models are increasingly being used in clinical decision making and drug testing. The cell models also have the potential to assist interpretation and extending of our multi-scale experimental findings. Components of the models interact with each other to regulate model behavior in a non-linear manner. To permit meaningful deployment of the models, it is therefore a necessity to understand the regulatory significance of model components’ parameters on the model’s behavior. In this study, the regulation of mean intra-cellular calcium and mean membrane potential (model behavior) by underlying model parameters (regulators) in a smooth muscle cell mathematical model was quantified using two sensitivity analysis methods. It was found that extracellular electrolytes and gating kinetics are prime model behavior regulators. A representative case relevant to widespread hypertension focusing on the L-type channel’s parameters is presented. This sensitivity analysis will guide our future data driven modelling efforts.

Electrocardiographic Imaging (ECGI) is a non-invasive procedure that allows to reconstruct the electrical activity of the heart from body surface potential map (BSPM). In this paper, we present a volume model to solve the electrocardiography inverse problem capable to take into account structural informations obtained by imaging techniques. Thedirect problem maps a volume current in the cardiac muscle (ventricles) to the body surface electrical measures. The model is based on coupling bidomain heart model with torso conduction. The corresponding inverse problem is solved with the Tikhonov regularization. Simulated database are used for the evaluation of this method and we compared them to standard method of fundamental solutions (MFS). The sensitivity to noise is also assessed. The correlation coefficients (CC) and the relative error (RE) of activation times were computed. Results show that the CC (respectively RE) median is respectively 0.75 for the volume model and 0.4 for the MFS (respectively 0.31 vs 0.35) on the epicardium. On the endocardium, the CC and the RE median are 0.65 and 0.33 for the volume method. In conclusion, the volume method performs better than the method of fundamental solutions (MFS) for any noise level, and reconstruct in addition endocardial information.

The standard electrocardiogram (ECG) is a point-wise evaluation of the body potential at certain given locations. These locations are subject to uncertainty and may vary from patient to patient or even for a single patient. In this work, we estimate the uncertainty in the ECG induced by uncertain electrode positions when the ECG is derived from the forward bidomain model. In order to avoid the high computational cost associated to the solution of the bidomain model in the entire torso, we propose a low-rank approach to solve the uncertainty quantification (UQ) problem. More precisely, we exploit the sparsity of the ECG and the lead field theory to translate it into a set of deterministic, time-independent problems, whose solution is eventually used to evaluate expectation and covariance of the ECG. We assess the approach with numerical experiments in a simple geometry.

The analysis of quantitative hemodynamics and luminal pressure may add valuable information to aid treatment strategies and prognosis for aortic dissections. This work directly compared in vitro 4D-flow magnetic resonance imaging (MRI), catheter-based pressure measurements, and computational fluid dynamics that integrated fluid-structure interaction (CFD FSI). Experimental data was acquired with a compliant 3D-printed model of a type-B aortic dissection (TBAD) that was embedded into a flow circuit with tunable boundary conditions. In vitro flow and relative pressure information were used to tune the CFD FSI Windkessel boundary conditions. Results showed overall agreement of complex flow patterns, true to false lumen flow splits, and pressure distribution. This work demonstrates feasibility of a tunable experimental setup that integrates a patient-specific compliant model and provides a test bed for exploring critical imaging and modeling parameters that ultimately may improve the prognosis for patients with aortic dissections.

We generalized and improved our clinical technique of two-dimensional intraventricular vector flow mapping (2D-iVFM) for a full-volume three-component analysis of the intraventricular blood flow (3D-iVFM). While 2D-iVFM uses three-chamber color Doppler images, 3D-iVFM is based on the clinical mode of triplane color Doppler echocardiography. As in the previous two-dimensional version, 3D-iVFM relies on mass conservation and free-slip endocardial boundary conditions. For sake of robustness, the optimization problem was written as a constrained least-squares problem. We tested and validated 3D-iVFM in silico through a patient-specific heart-flow CFD (computational fluid dynamics) model, as well as in vivo in one healthy volunteer. The intraventricular vortex that forms during left ventricular filling was deciphered. After further validation, 3D-iVFM could offer clinically compatible 3-D echocardiographic insights into left intraventricular hemodynamics.

The treatment of coronary stenosis relies on invasive high risk surgical assessment to generate the fractional flow reserve, a ratio of distal to proximal pressures in respect of the stenosis. Non-invasive methods are therefore desirable. Non-invasive imaging-computational methodologies call for robust and calibrated mathematical descriptions of the coronary vasculature that can be personalized. In addition, it is important to understand extra-coronary co-morbidities that may affect fractional flow estimates. In this preliminary theoretical work, a 0D human coronary vasculature model was implemented, and used to demonstrate the distinct roles of focal and extended stenosis (intra-coronary), as well as microvascular disease and atrial fibrillation (extra-coronary) on fractional flow reserve estimation. It was found that the right coronary artery is maximally affected by diffuse stenosis and microvascular disease. The model predicts that the presence, rather than severity, of both microvascular disease and atrial fibrillation affect coronary flow deleteriously. The model provides a computationally inexpensive instrument for future in silico coronary blood flow investigations as well as clinical-imaging decision making. The framework provided is extensible as well as can be personalized. Furthermore, it provides a starting point and crucial boundary conditions for future 3D computational hemodynamics flow estimation.

Atrial fibrillation (AF) is considered the most common human arrhythmia. Around 99% of thrombi in non-valvular AF are formed in the left atrial appendage (LAA). Studies suggest that abnormal LAA haemodynamics and the subsequently stagnated flow are the factors triggering clot formation. However, the relation between LAA morphology, the blood pattern and the triggering is not fully understood. Moreover, the impact of structures such as the pulmonary veins (PVs) on LA haemodynamics has not been thoroughly studied due to the difficulties of acquiring appropriate data. On the other hand, in-silico studies and flow simulations allow a thorough analysis of haemodynamics, analysing the 4D nature of blood flow patterns under different boundary conditions. However, the reduced number of cases reported on the literature of these studies has been a limitation. The main goal of this work was to study the influence of PVs on left atrium (LA) and LAA haemodynamics. Computational fluid dynamics simulations were run on 52 patients, the largest cohort so far in the literature, where different parameters were individually studied: pulmonary veins orientation and configuration; LAA and LA volumes and its ratio; and flow velocities. Our computational analysis showed how the right pulmonary vein height and angulation have a great influence on LA haemodynamics. Additionally, we found that LAA with great bending with its tip pointing towards the mitral valve could contribute to favour flow stagnation.

In humans, there is a well-established relationship between atrial fibrillation (AF), blood flow abnormalities and thrombus formation, even if there is no clear consensus on the role of left atrial appendage (LAA) morphologies. Cats can also suffer heart diseases, often leading to an enlargement of the left atrium that promotes stagnant blood flow, activating the clotting process and promoting feline aortic thromboembolism. The majority of pathological feline hearts have echocardiographic evidence of abnormal left ventricular filling, usually assessed with 2D and Doppler echocardiography and standard imaging tools. Actually, veterinary professionals have limited access to advanced computational techniques that would enable a better understanding of feline heart pathologies with improved morphological and haemodynamic descriptors. In this work, we applied state-of-the-art image processing and computational fluid simulations based on micro-computed tomography images acquired in 24 cases, including normal cats and cats with varying severity of cardiomyopathy. The main goal of the study was to identify differences in the LA/LAA morphologies and blood flow patterns in the analysed cohorts with respect to thrombus formation and cardiac pathology. The obtained results show significant differences between normal and pathological feline hearts, as well as in thrombus vs non-thrombus cases and asymptomatic vs symptomatic cases, while it was not possible to discern in congestive heart failure with thrombus and from non-thrombus cases. Additionally, in-silico fluid simulations demonstrated lower LAA blood flow velocities and higher thrombotic risk in the thrombus cases.

Current biophysical atrial models for investigating atrial fibrillation (AF) mechanisms and treatment approaches use imaging data to define patient-specific anatomy. Electrophysiology of the models can be calibrated using invasive electrical data collected using electroanatomic mapping (EAM) systems. However, these EAM data are typically only available after the catheter ablation procedure has begun, which makes it challenging to use personalised biophysical simulations for informing procedures. In this study, we first aimed to derive a mapping between LGE-MRI intensity and EAM conduction velocity (CV) for calibrating patient-specific left atrial electrophysiology models. Second, we investigated the functional effects of this calibration on simulated arrhythmia properties. To achieve this, we used the Universal Atrial Coordinate (UAC) system to register LGE-MRI and EAM meshes for ten patients. We then post-processed these data to investigate the relationship between LGE-MRI intensities and EAM CV. Mean atrial CV decreased from 0.81 ± 0.31 m/s to 0.58 ± 0.18 m/s as LGE-MRI image intensity ratio (IIR) increased from IIR < 0.9 to 1.6 ≤ IIR. The relationship between IIR and CV was used to calibrate conductivity for a cohort of 50 patient-specific models constructed from LGE-MRI data. This calibration increased the mean number of phase singularities during simulated arrhythmia from 2.67 ± 0.94 to 5.15 ± 2.60.

The assessment of left atrial appendage (LAA) thrombogenesis has experienced major advances with the adoption of patient-specific computational fluid dynamics (CFD) simulations. Nonetheless, due to the vast computational resources and long execution times required by fluid dynamics solvers, there is an ever-growing body of work aiming to develop surrogate models of fluid flow simulations based on neural networks. The present study builds on this foundation by developing a deep learning (DL) framework capable of predicting the endothelial cell activation potential (ECAP), linked to the risk of thrombosis, solely from the patient-specific LAA geometry. To this end, we leveraged recent advancements in Geometric DL, which seamlessly extend the unparalleled potential of convolutional neural networks (CNN), to non-Euclidean data such as meshes. The model was trained with a dataset combining 202 synthetic and 54 real LAA, predicting the ECAP distributions instantaneously, with an average mean absolute error of 0.563. Moreover, the resulting framework manages to predict the anatomical features related to higher ECAP values even when trained exclusively on synthetic cases.

Electroanatomical maps are a key tool in the diagnosis and treatment of atrial fibrillation. Current approaches focus on the activation times recorded. However, more information can be extracted from the available data. The fibers in cardiac tissue conduct the electrical wave faster, and their direction could be inferred from activation times. In this work, we employ a recently developed approach, called physics informed neural networks, to learn the fiber orientations from electroanatomical maps, taking into account the physics of the electrical wave propagation. In particular, we train the neural network to weakly satisfy the anisotropic eikonal equation and to predict the measured activation times. We use a local basis for the anisotropic conductivity tensor, which encodes the fiber orientation. The methodology is tested both in a synthetic example and for patient data. Our approach shows good agreement in both cases, with an RMSE of 2.2 ms on the in-silico data and outperforming a state of the art method on the patient data. The results show a first step towards learning the fiber orientations from electroanatomical maps with physics-informed neural networks.

Cardiac output is dependent on the tight coupling between atrial and ventricular function. The study of such interaction mechanisms is hindered by their complexity, and therefore requires a systematic approach. We have developed a four-chamber closed-loop cardiac electromechanics model which, through the coupling of the chambers with a closed-loop cardiovascular system model and the effect of the pericardium, is able to capture atrioventricular interaction. Our model simulates electrical activation and contraction of the atria and the ventricles coupled with a closed-loop model based on the CircAdapt framework. We include the effect of the pericardium on the heart using normal springs, scaling the local spring stiffness based on image-derived motion. The coupled model was used to study the impact of ventricular myofibre orientation on atrial dynamics by varying ventricular fibre orientation from –40 $$^\circ $$ ∘ /+40 $$^\circ $$ ∘ to –70 $$^\circ $$ ∘ /+70 $$^\circ $$ ∘ . We found that steeper fibres increase atrioventricular valve plane motion from 1.0 mm to 14.0 mm, leading to a lower minimum left atrial (LA) pressure (–0.4 mmHg vs –1.1 mmHg) and greater venous return (LA maximum volume: 168 mL vs 182 mL), and that fibres angles –50 $$^\circ $$ ∘ /+50 $$^\circ $$ ∘ were consistent with a physiological atrial contraction and filling pattern. Our framework is capable of capturing complex interaction dynamics between the atria, the ventricles and the circulatory system accounting for the effect of the pericardium. Such simulation platform represents a useful tool to study both systolic and filling phases of all cardiac chambers, and how these get altered in diseased states and in response to treatment.

Intracardiac devices separate atrial arrhythmias (AA) from sinus rhythm (SR) using electrogram (EGM) features such as rate, that are imperfect. We hypothesized that machine learning could improve this classification.In 71 persistent AF patients (50 male, 65 ± 11 years) we recorded unipolar and bipolar intracardiac EGMs for 1 min prior to ablation, providing 50,190 unipolar and 44,490 bipolar non-overlapping 4 s segments. We developed custom deep learning models to detect SR or AA, with 10-fold cross-validation, compared to classical analyses of cycle length (CL), Dominant Frequency (DF) and autocorrelation.Classical analyses of single features were modestly effective with AUC ranging from 0.91 (DF) to 0.70 for other rate metrics. Performance increased by combining features linearly (AUC 0.991/0.987 for unipolar/bipolar), by Bagged Trees (0.995/0.991) or K-Nearest Neighbors (0.985/0.991). Convolutional deep learning of raw EGMs with no feature engineering provided improved AUC of 0.998/0.995 to separate AA from SR.Deep learning of raw EGMs outperforms classic rule-based classifiers of SR or AA. This could improve device diagnosis, and the logic developed by deep learning could shed novel insights into EGM analyses beyond current classification based on EGM features and rules.

Atrial fibrillation is a prevalent cardiac arrhythmia, and may reduce cerebral blood perfusion augmenting the risk of dementia. It is thought that geometric variations in the cerebral arterial structure called the Circle of Willis play an important role influencing cerebral perfusion. The objective of this work is to use computational modelling to investigate the role of variations in cerebral vascular structure on cerebral blood flow dynamics during atrial fibrillation.A computational blood flow model was developed by coupling whole-body and detailed cerebral circulation models, modified to represent the most common variations of the Circle of Willis. Cerebral blood flow dynamics were simulated in common Circle of Willis variants, with imposed atrial fibrillation conditions. Perfusion and its heterogeneity were quantified using segment-wise hypoperfusion events and mean perfusion at terminals.It was found that cerebral perfusion and the rate of hypoperfusion events strongly depend on Circle of Willis geometry as well as atrial fibrillation induced stochastic heart rates. The missing ACA1 variant had a 25% decrease in hypoperfusion events compared to normal, while the missing PCA1 and PCoA variant had a 550% increase. A similar trend was observed in flow heterogeneity. The hypoperfusion events were specific to particular arteries within each variant. Our results, based on biophysical principles, suggest that cerebral vascular geometry plays an important role influencing cerebral hemodynamics during atrial fibrillation. Additionally, our findings suggest potential clinical assessment sites. Further work will be conducted using spatially resolved 1D modelling.