3D-Fully Quantitative Flow

• 3D-FQFlow is a methodology that quantitatively reconstructs dense 3D fluid fields, ensuring physical validity across various imaging modalities.
• It formulates reconstruction as a constrained inverse problem by enforcing physical priors such as incompressibility and anatomical consistency.
• The open-source framework integrates vascular simulation, tissue motion, and optimized ultrasound imaging to derive robust, clinically relevant flow biomarkers.

3D-Fully Quantitative Flow (3D-FQFlow) denotes methodologies that recover quantitatively interpretable three-dimensional flow information from measurements or simulations, rather than only producing qualitative flow visualization. In the current literature, the expression appears in two closely related senses: as a generic descriptor for volumetric, physically grounded flow quantification across modalities, and as the formal name of an open-source framework for simulating 3D vascular flow with tissue motion for ultrafast power Doppler imaging (Fu et al., 5 Sep 2025). Across fluid mechanics, ultrasound, MRI, echocardiography, angiography, and patient-specific computational hemodynamics, the unifying objective is to obtain dense 3D fields or derived biomarkers that remain suitable for quantitative analysis, physical interpretation, and validation.

1. Terminology and scope

In adjacent research areas, “3D-FQFlow” is not confined to a single acquisition technology or reconstruction algorithm. In volumetric particle imaging, it refers to the direct recovery of a dense 3D fluid motion field in the entire domain while explicitly reconstructing tracer particles (Lasinger et al., 2018). In echocardiography, it refers to recovery of three-component velocity vector fields in a full intraventricular volume from triplane color Doppler (Vixège et al., 2021). In accelerated cardiovascular MRI, it refers to rapid reconstruction of undersampled 4D flow MRI data for time-resolved quantification of blood-flow dynamics (Vishnevskiy et al., 2020). In sparse-sensor reconstruction around bluff bodies, it refers to reconstruction of pressure, velocity components, Q-criterion fields, and lift and drag forces from limited measurements (Özbay et al., 2023).

This multiplicity of usage suggests that 3D-FQFlow is best understood as a methodological criterion rather than a single algorithmic lineage. The criterion is that the output must be volumetric, quantitatively validated, and sufficiently constrained to support downstream biomechanical or diagnostic interpretation.

2. Physical and inverse-problem foundations

A defining feature of 3D-FQFlow methods is that reconstruction is usually formulated as a constrained inverse problem. In the integrated particle-reconstruction framework for volumetric fluid estimation, explicit tracer particles and a dense motion field are estimated jointly through the variational objective

$$E(\mathcal{P}, \mathcal{C}, \mathcal{U}) = \frac{1}{2} E_D(\mathcal{P}, \mathcal{C}, \mathcal{U}) + \frac{\lambda}{2} E_S(\mathcal{U}) + \mu E_{Sp}(\mathcal{C}),$$

where the smoothness term enforces viscosity and a divergence-free prior, and the sparsity term suppresses ghost particles (Lasinger et al., 2018). The method uses inertial Proximal Alternating Linearized Minimization (iPALM), and its physical prior is expressed through penalization of velocity gradients together with incompressibility.

A closely related strategy appears in 3D intraventricular vector flow mapping, where triplane color Doppler supplies only radial velocity observations, and the missing components are inferred under mass conservation and free-slip boundary conditions on the endocardial wall. The inverse problem is written as a constrained least-squares system solved with Lagrange multipliers, with the divergence-free condition

$\nabla \cdot \vec{v} = 0$

and boundary kinematics imposed explicitly (Vixège et al., 2021).

Cardiac motion reconstruction has extended this logic from velocity estimation to kinematic consistency. A flow-compatible formulation couples anatomical observations with measured flow rates and, optionally, hemodynamic force, enforcing mass and momentum conservation through a variational optimization in which anatomical mismatch, mass mismatch, and momentum mismatch are jointly weighted (Capuano et al., 16 Apr 2025). This suggests that, in the strictest sense, “fully quantitative” denotes not only 3D output but also adherence to governing physical balances during reconstruction.

3. The open-source framework named 3D-FQFlow

The paper that explicitly names its framework “3D-Fully Quantitative Flow (3D-FQFlow)” introduces an open-source system for quantitative modeling of 3D vascular flow with tissue motion and ultrafast power Doppler imaging (Fu et al., 5 Sep 2025). Its architecture is modular and spans geometry generation, hemodynamics, motion, acoustic simulation, reconstruction, and image analysis.

Module	Function
L-system-based vascular generator	Generates anatomically plausible 3D vascular trees
SimVascular CFD	Computes vascular hemodynamics
Tissue motion simulator	Supports user-defined or clinical-data-driven condition
Optimized PFILED ultrasound simulator	Simulates 3D-uPDI data
Precomputed-matrix-based reconstructor	Reconstructs volumetric ultrasound images
Quantitative analyzer	Computes MSE, PSNR, and SSIM

The motion model updates scatterer positions according to

$$\vec{P}(t) = \vec{P}_0 + \int_0^t \vec{v}(\vec{P}(s)) \, ds,$$

while vascular flow is generated through SimVascular CFD and coupled to ultrasound image formation and reconstruction (Fu et al., 5 Sep 2025). The framework is reported as the first open-source framework for quantitative validation of uPDI under realistic vascular and motion conditions, and is released at the repository listed in the source paper.

Its reported quantitative results are centered on both fidelity and computational feasibility. The study demonstrated distinct influences of four motion patterns on SVD decomposition; successful 3D imaging of rabbit kidney with SSIM $= 0.951$, generated vasculature with SSIM $= 0.902$, and clinical pulmonary arteries with SSIM $= 0.850$; and GPU acceleration permitting 1-million-scatterer simulation in 4,117 seconds with 18.8* speedup for 100-frame 3D-uPDI generation (Fu et al., 5 Sep 2025). The same study also reports modality-specific degradation under stronger tissue motion, especially through clutter contamination of Doppler filtering.

4. Representative measurement and reconstruction systems

In multi-camera volumetric particle imaging, 3D-FQFlow is exemplified by an integrated hybrid Lagrangian/Eulerian method that jointly reconstructs tracer particles and a dense 3D motion field from only two time steps. At $0.1$ particle per pixel, the joint method reduced average endpoint error from $0.406$ to $0.136$ and average angular error from $6.74^\circ$ to $\nabla \cdot \vec{v} = 0$0, corresponding to an approximately $\nabla \cdot \vec{v} = 0$1 improvement over the prior two-step baseline; the authors further report performance comparable to state-of-the-art tracking-based methods such as Shake-the-Box while using only two frames (Lasinger et al., 2018).

In echocardiography, 3D-iVFM generalizes intraventricular vector flow mapping from 2-D to full-volume reconstruction using clinical triplane color Doppler. The reconstructed field is expressed in spherical coordinates as $\nabla \cdot \vec{v} = 0$2, with azimuthal behavior represented through a periodic trigonometric expansion. In silico validation against patient-specific CFD reported $\nabla \cdot \vec{v} = 0$3 nRMSE $\nabla \cdot \vec{v} = 0$4 with correlation $\nabla \cdot \vec{v} = 0$5, $\nabla \cdot \vec{v} = 0$6 nRMSE $\nabla \cdot \vec{v} = 0$7 with correlation $\nabla \cdot \vec{v} = 0$8, and $\nabla \cdot \vec{v} = 0$9 nRMSE $$\vec{P}(t) = \vec{P}_0 + \int_0^t \vec{v}(\vec{P}(s)) \, ds,$$0 with correlation $$\vec{P}(t) = \vec{P}_0 + \int_0^t \vec{v}(\vec{P}(s)) \, ds,$$1; all steps from acquisition to 3D velocity computation were reported to complete in under 20 seconds per volume (Vixège et al., 2021).

In accelerated 4D flow MRI, FlowVN applies model-driven unrolled reconstruction to highly undersampled phase-contrast data. The network uses 10 unrolled steps, groups 3D convolutions into $$\vec{P}(t) = \vec{P}_0 + \int_0^t \vec{v}(\vec{P}(s)) \, ds,$$2, $$\vec{P}(t) = \vec{P}_0 + \int_0^t \vec{v}(\vec{P}(s)) \, ds,$$3, $$\vec{P}(t) = \vec{P}_0 + \int_0^t \vec{v}(\vec{P}(s)) \, ds,$$4, and $$\vec{P}(t) = \vec{P}_0 + \int_0^t \vec{v}(\vec{P}(s)) \, ds,$$5 filter banks to avoid costly 4D convolutions, and was trained on 11 reference scans while generalizing to multiple acceleration factors and anatomies. A typical 4D flow MRI volume was reconstructed in about 21 seconds with 63,583 parameters; at acceleration $$\vec{P}(t) = \vec{P}_0 + \int_0^t \vec{v}(\vec{P}(s)) \, ds,$$6, reported performance was magnitude nRMSE $$\vec{P}(t) = \vec{P}_0 + \int_0^t \vec{v}(\vec{P}(s)) \, ds,$$7, velocity magnitude relative error $$\vec{P}(t) = \vec{P}_0 + \int_0^t \vec{v}(\vec{P}(s)) \, ds,$$8, and mean angular error $$\vec{P}(t) = \vec{P}_0 + \int_0^t \vec{v}(\vec{P}(s)) \, ds,$$9 (Vishnevskiy et al., 2020).

In neurovascular angiography, 4D angiograms have been reconstructed from two biplane DSA views using constrained back-projection and an a priori 3D vascular geometry. Patient-specific internal carotid aneurysm models supplied CFD-based ground truth, and the reconstructed 4D datasets achieved an average MSE of $= 0.951$0 across models and flow conditions. Intensity-derived angiographic parametric imaging metrics such as PH and AUC closely matched the CFD reference, whereas temporal metrics showed greater variability in regions with overlapping projections (Williams et al., 13 Feb 2025).

5. Geometry normalization, learned surrogates, and patient-specific digital twins

A major recent trend is the use of geometric normalization or anatomy-aware priors to make quantitative 3D reconstruction transferable across shapes. FR3D addresses unsteady flows around extruded bluff bodies by conformally mapping each 2D cross-section to an annulus, then reconstructing $= 0.951$1, $= 0.951$2, $= 0.951$3, and $= 0.951$4 on a $= 0.951$5 grid using a convolutional autoencoder and a sensor-to-latent embedder. The model was trained on 80 geometries and tested on 20 unseen geometries. In the sparse-sensor setup, reported MAPE values were $= 0.951$6 for $= 0.951$7, $= 0.951$8 for $= 0.951$9, $= 0.902$0 for $= 0.902$1, and $= 0.902$2 for $= 0.902$3, with min-max normalized errors below $= 0.902$4 for all velocities; average test-set force errors were $= 0.902$5 for $= 0.902$6 and $= 0.902$7 for $= 0.902$8, while plane-based sensing reduced these to $= 0.902$9 and $= 0.850$0 (Özbay et al., 2023). The same framework also reconstructs Q-criterion fields from the recovered velocity gradients.

At organ scale, image-based whole-heart cardiac flow simulation takes a complementary route. Rather than inferring fields directly from sparse observables, it reconstructs moving anatomies from time-resolved medical images using machine learning-based segmentation and mesh propagation, then solves CFD in deforming domains with all four valves represented by resistive immersed surfaces. The framework was applied to a healthy adult and a pediatric patient with congenital heart disease; in the healthy case it reproduced physiologic pressure-volume behavior, valve timing, and ventricular vortex formation, while in the congenital heart disease case simulated chamber and vessel pressures agreed with cardiac catheterization measurements and revealed altered diastolic flow organization and elevated normalized viscous dissipation (Kong et al., 10 May 2026).

Taken together, these studies indicate two complementary paths within 3D-FQFlow research. One path reconstructs dense quantitative fields from incomplete measurements; the other generates patient-specific or geometry-specific digital flow fields from anatomy plus physics. The distinction is methodological, but the target output is similar: a 3D time-resolved description of flow that can support quantitative interpretation.

6. Limitations, misconceptions, and research trajectory

A common misconception is that any 3D flow visualization is already “fully quantitative.” The surveyed work shows the opposite. Quantitative validity depends strongly on the observability structure of the modality, the explicit incorporation of priors, and independent validation. In biplane angiographic reconstruction, temporal API metrics degrade in overlapping and foreshortened regions because signal assignment becomes ambiguous (Williams et al., 13 Feb 2025). In ultrafast Doppler simulation, tissue motion can produce clutter signals $= 0.850$1 dB above flow signals, distort SVD decompositions, and generate visible flash noise (Fu et al., 5 Sep 2025). In triplane Doppler echocardiography, the azimuthal component remains the least directly observed because only six azimuthal samples are available from three planes (Vixège et al., 2021).

Another misconception is that machine learning makes physical constraints optional. The strongest results in this literature do not support that view. FlowVN remains model-based and enforces data consistency during unrolled reconstruction (Vishnevskiy et al., 2020). FR3D relies on conformal mapping so that sensor locations and surface integration remain geometrically comparable across unseen bluff-body shapes (Özbay et al., 2023). Flow-compatible motion reconstruction explicitly treats anatomical agreement alone as insufficient, introducing mass and momentum conservation as correction criteria for reconstructed cardiac kinematics (Capuano et al., 16 Apr 2025).

The current trajectory therefore points toward tighter coupling between imaging, mechanics, and flow physics. This is already visible in frameworks that combine anatomical imaging with measured flow rates or hemodynamic forces, and in patient-specific simulations that use image-derived moving geometries together with physiologically realistic valve dynamics (Capuano et al., 16 Apr 2025, Kong et al., 10 May 2026). A plausible implication is that future 3D-FQFlow systems will be judged less by visual plausibility alone and more by whether they satisfy dynamic conservation laws, reproduce independent physiologic measurements, and remain quantitatively stable under adverse acquisition conditions.