LVADNet3D: 3D Flow Reconstruction in LVADs

• The paper presents LVADNet3D, a deep 3D convolutional autoencoder that reconstructs full-resolution intraventricular blood flow from sparse measurements using anatomical and physiological priors.
• It employs a five-level encoder-decoder with hybrid downsampling and increased channel capacity, outperforming standard UNet3D in metrics like MSE and PSNR.
• Evaluations on a synthetic CFD dataset show that integrating Radial Distance Fields and inlet velocity significantly enhances reconstruction fidelity for clinical hemodynamic analysis.

Searching arXiv for the specified paper and closely related LVAD/hemodynamics reconstruction work. LVADNet3D is a deep 3D convolutional autoencoder designed to reconstruct full-resolution intraventricular blood-flow velocity fields from sparse hemodynamic measurements in patients supported by Left Ventricular Assist Devices (LVADs). The method targets a clinically important setting in which conventional imaging is either incompatible with LVADs or provides sparse, low-quality velocity information, while direct Computational Fluid Dynamics (CFD) simulation remains too computationally intensive for routine use. Within that setting, LVADNet3D reconstructs volumetric velocity fields on $128 \times 128 \times 128$ grids from sparse velocity vector inputs, using anatomical and physiological priors to improve fidelity and physical plausibility (Khan et al., 21 Sep 2025).

1. Clinical and computational setting

Accurate assessment of intraventricular blood flow is essential for evaluating hemodynamic conditions in LVAD-supported hearts. The central difficulty addressed by LVADNet3D is the mismatch between clinical measurability and the spatial resolution required for detailed flow analysis: clinical imaging is either incompatible with LVADs or yields sparse, low-quality velocity data, whereas CFD provides high-fidelity flow fields but is impractical for routine clinical use because of its computational cost (Khan et al., 21 Sep 2025).

The framework is therefore positioned between sparse observation and full numerical simulation. Its stated goal is the in silico recovery of fine-scale velocity fields from highly limited clinical measurements. Concretely, the target quantity is the full 3D velocity vector field

$\mathbf{V} = [V_x, V_y, V_z],$

with each component represented on a $128 \times 128 \times 128$ grid (Khan et al., 21 Sep 2025).

A common misconception is to equate such a model with direct image-based measurement. LVADNet3D instead operates as a reconstruction model trained on CFD-generated supervisory data. A plausible implication is that its performance depends not only on network design but also on how well the synthetic CFD dataset spans relevant ventricular anatomies and inlet conditions.

2. Reconstruction problem and conditioning variables

The reconstruction task is defined component-wise. For each velocity component $d \in \{x, y, z\}$, the model takes as input a sparse velocity field and an anatomical prior: $$(\mathbf{V}_d^s,\, \mathbf{R}) \in \mathbb{R}^{128 \times 128 \times 128 \times 2},$$ where $\mathbf{V}_d^s$ denotes the sparse component input and $\mathbf{R}$ is the Radial Distance Field (RDF), which encodes the distance of each voxel to the ventricular center (Khan et al., 21 Sep 2025).

The sparse input simulates clinical data availability by retaining only $5\%$ of voxels and zeroing the remainder, with a consistent mask across velocity components. The RDF is always dense. In addition to these spatial inputs, the model uses a physiological prior: the inlet velocity $v_{\text{in}}$, described as a scalar measurable in clinical practice (Khan et al., 21 Sep 2025).

The output for each component is a reconstructed full-resolution field,

$$\widehat{\mathbf{V}_d^f} \in \mathbb{R}^{128 \times 128 \times 128 \times 1}.$$

This formulation makes the model explicitly conditional on both chamber morphology and boundary-condition information. In the paper’s interpretation, the RDF guides reconstruction toward morphology-consistent flow, while the inlet velocity constrains the global flow regime (Khan et al., 21 Sep 2025).

The role of conditioning is central rather than auxiliary. The anatomical prior provides spatial context when sparse velocity samples cannot resolve boundaries, and the physiological prior directs the model toward the observed patient-specific flow state. This suggests that LVADNet3D should be understood not merely as a sparse-to-dense interpolator, but as a conditional generative reconstructor constrained by hemodynamic context.

3. Network architecture and design departures from UNet3D

LVADNet3D is described as a deep 3D convolutional autoencoder with a five-level encoder-decoder hierarchy, skip connections, hybrid downsampling, and increased channel capacity in deeper layers (Khan et al., 21 Sep 2025). Its design is explicitly contrasted with a standard UNet3D baseline.

In the encoder, each layer contains two 3D convolution blocks of the form

$\mathbf{V} = [V_x, V_y, V_z],$0

where $\mathbf{V} = [V_x, V_y, V_z],$1 is PReLU and the normalization is InstanceNorm (Khan et al., 21 Sep 2025). The encoder uses five layers, increasing channels from 16 and doubling at each stage until reaching 256 at the bottleneck. The downsampling strategy is hybrid: early layers ($\mathbf{V} = [V_x, V_y, V_z],$2) use max pooling, while later layers ($\mathbf{V} = [V_x, V_y, V_z],$3) use strided convolutions (Khan et al., 21 Sep 2025).

The latent representation is

$\mathbf{V} = [V_x, V_y, V_z],$4

The inlet velocity $\mathbf{V} = [V_x, V_y, V_z],$5 is broadcast and concatenated with this latent tensor to form

$\mathbf{V} = [V_x, V_y, V_z],$6

after which a $\mathbf{V} = [V_x, V_y, V_z],$7 convolution reduces the channel count back to 256 (Khan et al., 21 Sep 2025). The decoder then uses transpose convolutions for upsampling and concatenative skip connections from the encoder at each scale, progressively reducing channels back to 1 for the reconstructed component output.

The main architectural distinctions from UNet3D are summarized below.

Aspect	UNet3D	LVADNet3D
Depth	4	5
Downsampling	Max pooling only	Hybrid (max pool + strided conv)
Channels	Fewer	More at deeper layers

The paper attributes specific functional effects to these choices. The deeper hierarchy is said to enlarge the receptive field and improve integration of sparse input with anatomical context. Hybrid downsampling is intended to preserve fine detail early while enabling learnable abstraction later. Increased channel capacity at depth is presented as necessary for encoding complex spatial structure in intraventricular flow (Khan et al., 21 Sep 2025).

4. Training objective, synthetic CFD dataset, and experimental protocol

LVADNet3D is trained on a synthetic CFD dataset generated for LVAD-supported left ventricles. The dataset comprises 8 anatomically diverse left ventricles modeled in Fusion 360, meshed at 600k to 1.5M elements per mesh. CFD simulations were performed in ANSYS Fluent under laminar flow assumptions, with blood density $\mathbf{V} = [V_x, V_y, V_z],$8 and viscosity $\mathbf{V} = [V_x, V_y, V_z],$9, and inlet velocities ranging from $128 \times 128 \times 128$0 to $128 \times 128 \times 128$1. In total, 47 simulations were generated across multiple inlet conditions and geometries (Khan et al., 21 Sep 2025).

Ground-truth velocity fields were interpolated to regular $128 \times 128 \times 128$2 grids, producing tensors of shape $128 \times 128 \times 128$3. The RDF was computed as an additional anatomical channel. The dataset split used 5-fold cross-validation by geometry, so the hold-out condition was at the anatomy level rather than a random sample level (Khan et al., 21 Sep 2025). This matters because it evaluates generalization across ventricular geometries rather than only across flow realizations within the same anatomy.

The reconstruction loss is the Huber loss with $128 \times 128 \times 128$4: $128 \times 128 \times 128$5 with

$128 \times 128 \times 128$6

The paper describes this choice as robust to outliers (Khan et al., 21 Sep 2025).

The use of a CFD-derived training corpus is a defining methodological feature. It provides dense supervision that is difficult to obtain clinically, but it also indicates that the empirical scope of the model is tied to the simulated regime: eight ventricular geometries, laminar flow assumptions, and inlet velocities between $128 \times 128 \times 128$7 and $128 \times 128 \times 128$8. A plausible implication is that extrapolation outside those ranges would require additional evidence.

5. Evaluation metrics and empirical findings

The reported evaluation uses both per-component fidelity measures and aggregate flow measures. The metrics are Mean Squared Error (MSE), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Peak Signal-to-Noise Ratio (PSNR), with PSNR defined as

$128 \times 128 \times 128$9

For aggregate flow characterization, the paper also evaluates velocity magnitude: $d \in \{x, y, z\}$0 Qualitative assessment is performed with 2D slices showing local high- and low-flow recoveries (Khan et al., 21 Sep 2025).

Across all input settings, LVADNet3D is reported to outperform the baseline UNet3D, with lower MSE, MAE, and RMSE, and higher PSNR. The paper reports improvements of up to $d \in \{x, y, z\}$1 component-wise and $d \in \{x, y, z\}$2 on the most complete input configuration for velocity magnitude (Khan et al., 21 Sep 2025). For reconstruction of the $d \in \{x, y, z\}$3 component, the reported example is:

• MSE: 1.90 for LVADNet3D versus 5.36 for UNet3D
• PSNR: $d \in \{x, y, z\}$4 versus $d \in \{x, y, z\}$5 (Khan et al., 21 Sep 2025)

Ablation results further indicate that priors materially affect performance. Adding the RDF reduces error and increases PSNR by approximately $d \in \{x, y, z\}$6 for UNet3D and $d \in \{x, y, z\}$7 for LVADNet3D. Adding inlet velocity yields a further PSNR increase of approximately $d \in \{x, y, z\}$8, with the largest benefit reported for LVADNet3D, where the global flow pattern depends on the boundary condition. The best performance is obtained when all priors are included (Khan et al., 21 Sep 2025).

Qualitative evidence is described consistently with the quantitative findings: sliced images show that anatomical detail, including flow jets and recirculation, is best recovered when both anatomical and physiological priors are provided (Khan et al., 21 Sep 2025). This supports the interpretation that the model’s gains are not solely denoising gains, but improvements in anatomically and hemodynamically structured reconstruction.

6. Relation to broader LVAD hemodynamics modeling

LVADNet3D belongs to a broader effort to reduce the gap between high-fidelity CFD and clinically or operationally usable hemodynamic prediction. A related direction is non-intrusive reduced-order modeling (ROM) for LVAD-associated flow, exemplified by a data-driven PODI framework for aortic blood flow in the presence of an LVAD with varying pump flow rate (Girfoglio et al., 2020). That framework uses high-fidelity snapshots, proper orthogonal decomposition, and radial basis function interpolation to reconstruct pressure, velocity components, and wall shear stress for unseen parameter values, achieving approximately $d \in \{x, y, z\}$9 speedup relative to the full-order model within the trained range (Girfoglio et al., 2020).

The distinction between the two approaches is methodological. The ROM framework approximates parametric PDE solutions from a database of high-fidelity solutions and is applied to aortic blood flow under pump-flow variation (Girfoglio et al., 2020). LVADNet3D instead reconstructs intraventricular velocity fields from sparse volumetric inputs and explicitly conditions on RDF and inlet velocity (Khan et al., 21 Sep 2025). The former is a parametric surrogate for simulation; the latter is a sparse-data reconstruction network trained on synthetic CFD outputs.

This comparison helps clarify what LVADNet3D is and is not. It is not presented as a general reduced-order solver for arbitrary LVAD hemodynamics, nor as a direct substitute for all CFD workflows. Rather, it addresses a specific inverse problem: recovering high-resolution intraventricular flow from highly limited measurements under anatomically and physiologically informed constraints. This suggests a complementary relationship between sparse-data reconstruction and ROM-based surrogate modeling within the broader computational hemodynamics literature.

7. Significance, scope, and limitations

The principal contribution of LVADNet3D is the combination of three elements within a single framework: a CFD-based synthetic dataset for LVAD-supported intraventricular flow, a deeper 3D convolutional autoencoder with hybrid downsampling and increased channel capacity, and explicit conditioning on anatomical and physiological priors (Khan et al., 21 Sep 2025). In the reported experiments, these design choices consistently outperform a standard UNet3D baseline across all tested metrics and input settings.

Its significance lies in the specific problem formulation it operationalizes. By treating sparse velocity measurements, ventricular geometry encoding, and inlet velocity as jointly informative signals, the model shifts the reconstruction problem from purely local interpolation toward conditional volumetric estimation constrained by morphology and boundary conditions. The paper’s results indicate that this is especially important when the global flow pattern depends strongly on the inlet condition (Khan et al., 21 Sep 2025).

At the same time, the empirical basis is explicitly synthetic and CFD-derived. The training and evaluation data are generated from 47 simulations over 8 anatomically diverse left ventricles, under stated assumptions including laminar flow and inlet velocities in the range $$(\mathbf{V}_d^s,\, \mathbf{R}) \in \mathbb{R}^{128 \times 128 \times 128 \times 2},$$0–$$(\mathbf{V}_d^s,\, \mathbf{R}) \in \mathbb{R}^{128 \times 128 \times 128 \times 2},$$1 (Khan et al., 21 Sep 2025). A plausible implication is that claims of clinical deployment would require further validation on measured data and under conditions beyond the simulated regime. Within the scope reported, however, LVADNet3D is a specialized reconstruction architecture for high-resolution intraventricular flow recovery in LVAD-supported hearts, with demonstrated gains from deeper hierarchy, hybrid downsampling, increased channel capacity, and conditioning on RDF and inlet velocity (Khan et al., 21 Sep 2025).