Using blood speckle tracking (BST) based on high-frame-rate echocardiography (HFRE), we compared right ventricle (RV) flow dynamics in children with atrial septal defects (ASDs) and repaired tetralogy of Fallot (rTOF). Fifty-seven children with rTOF with severe pulmonary insufficiency (PI) (n = 21), large ASDs (n = 11) and healthy controls (CTL, n = 25) were included. Using a flow phantom, we studied the effects of imaging plane and smoothing parameters on 2-D energy loss (EL). RV diastolic EL was similar in ASD and rTOF, but both were greater than in CTL. Locations of high EL were similar in all groups in systole, occurring in the RV outflow tract and around the tricuspid valve leaflets in early diastole. An additional apical early diastolic area of EL was noted in rTOF, corresponding to colliding tricuspid inflow and PI. The flow phantom revealed that EL varied with imaging plane and smoothing settings but that the EL trend was preserved if kept consistent.

Background, Motivation, and Objective Ultrasound vector flow imaging (VFI) methods have shown promise for measuring intracardiac flow patterns, but are hampered by variance and clutter filter dropouts. Methods attempting to mediate often lead to feature blurring (smoothers) or scale poorly when moving to 4D imaging (model-regularizers). We propose a flexible reconstruction framework based on an efficient B-spline interpolation kernel and with model-based data regularization terms computed on the analytical spline gradients. Statement of Contribution/Methods A general purpose nD B-spline interpolator of arbitrary orders and differentials was developed in the open source TensorFlow framework. Sparse gradients supporting reverse mode automatic differentiation (AD) were implemented, enabling the use of stochastic gradient descent optimizers to minimize general differentiable cost functions even on memory constrained systems. This allows arbitrary models and data sources to be specified with a high level of implementation abstraction. Parallel forward pass and AD codes were written for CPU and GPU to increase performance across platforms. The framework was evaluated for vector flow reconstruction constrained by the incompressible Navier-Stokes (NS) equations. Results/Discussion Evaluation was done towards a computational fluid dynamics (CFD) phantom, subjected to semi-realistic artifacts and noise. Measurements were fitted to 4D spline grids, penalizing the deviation from the NS momentum and mass balance at each data point. This resulted in convincing reconstructions for moderately challenging scenarios, example seen in figure, where the (reconstructed) lateral and total RMSE were 3.5 mm/s and 3.0 cm/s respectively. The average 4D reconstruction time of the phantom on a NVIDIA Titan V was 3 minutes. An observed limitation with this model is the lack of inlet/outlet handling, leading to underestimation of the true velocities in these regions when momentum balance is strongly enforced. In vivo 4D data was acquired using a GE Vivid E95 system with a 4V probe, where VFI was done using 3D blood speckle tracking (BST), while the LV domain was extracted automatically using the open-source FAST library. We emphasize that the flexibility of the framework lies in the ease of specifying models and data sources, and its general purpose nature invites application to other regularization problems.

No abstract

No abstract

No abstract