CTFlow: Continuous-Time Flow Models

• CTFlow is a suite of continuous-time flow models that parameterize data synthesis and inversion using ODEs/SDEs with learned velocity fields, enhancing CT reconstruction accuracy.
• It integrates differentiable physical models, such as the Radon transform and PDE-based constraints, to enable efficient sparse-view CT reconstruction and synthetic imaging.
• Notable applications include CT/MRI synthesis, cardiovascular simulation, and dynamic snapshot reconstruction, achieving significant speed-ups and improved robustness over traditional methods.

CTFlow refers to a set of closely related methodologies and models that employ flow-based and flow-matching frameworks for continuous-time modeling, generation, or inference in computed tomography (CT), cardiovascular physics, and broader medical and scientific imaging contexts. These approaches unify data-driven machine learning with physical modeling via ordinary (ODE) or stochastic (SDE) differential equations and differentiable projections, often outperforming traditional methods in efficiency, accuracy, or robustness in ill-posed inverse problems, synthetic data generation, and continuous-time dynamics reconstruction (Wang et al., 18 Aug 2025, Nannini et al., 2024, Hadzic et al., 6 Oct 2025, Shi et al., 27 Feb 2026, Thombre et al., 3 Jun 2026).

1. Mathematical and Algorithmic Foundations

CTFlow encompasses a family of continuous-time flow models. The core is the parameterization of data transport and synthesis as the solution to ODEs or SDEs, in which a learned velocity (drift) field defines the temporal evolution from noise (prior) to data (posterior or observation), or between domain pairs (e.g., MRI → CT):

• Basic ODE Formulation (Flow Matching):

$$\frac{dx_t}{dt} = v_\theta(x_t, t \mid c), \; t \in [0, 1]$$

where $x_0$ is sampled from an isotropic Gaussian or prior, and $x_1$ is the target data. The vector field $v_\theta$ is trained to match an analytically known or empirically estimated “bridge” velocity based on the task (Shi et al., 27 Feb 2026, Hadzic et al., 6 Oct 2025, Ye et al., 2024).

• Training Loss (Conditional Flow Matching):

$$\mathcal{L}_{\mathrm{CFM}} = \mathbb{E}_{t, z, x \sim p_t(\cdot \mid z)} \left\| v_\theta(x, t \mid c) - u_t(x \mid z) \right\|^2$$

with the ground-truth velocity field $u_t(x \mid z) = x_1 - x_0$ along the path connecting pairs of prior/true samples (Ye et al., 2024, Hadzic et al., 6 Oct 2025).

• Topology Optimization and Physics-Based Extensions:

In vascular reconstruction, CTFlow models simultaneously reconstruct geometry (via a density field $\chi(x)$), steady-state velocities, and contrast transport by solving PDE-constrained optimization, integrating differentiable tomography operators for projection matching in sinogram space (Thombre et al., 3 Jun 2026).

• Integration with CT Physics:

CTFlow algorithms may embed the Radon transform or other forward-imaging models directly into the loss function—for both image synthesis and physical parameter inference—allowing end-to-end, data-consistent inversion from projected measurement data (Guo et al., 5 Nov 2025, Hadzic et al., 6 Oct 2025).

2. Key CTFlow Architectures and Implementations

Multiple architectural realizations of CTFlow appear across tasks:

• Image and Volume Synthesis:

E.g., for MRI→CT or CBCT→CT, CTFlow leverages a lightweight 3D encoder for conditioning, concatenated with the noisy input, and processed through a 3D U-Net, with ODE integration via an RK4 solver. Deep latent flow matching with transformers and clinical-report conditioning enables text-to-3D-CT volume synthesis in the latent space, utilizing cross-attention to radiology text embeddings (Hadzic et al., 6 Oct 2025, Wang et al., 18 Aug 2025).

• Reconstruction from Sparse-View CT:

FMCT and its efficient variant EFMCT use a rectified flow-matching U-Net with strong sharing of predicted velocity fields across consecutive ODE timesteps, with adaptive data-consistency (CG-based) corrections at every step (Shi et al., 27 Feb 2026).

• Automated Cardiovascular Simulation:

The pipeline for non-invasive coronary flow simulation constructs patient-specific geometry and mesh from CCTA, infers patient-specific boundary conditions from clinical signals, and runs steady-state CFD with self-tuning inlet flows, with an option for subsequent transient simulation. All stages are fully automated (Nannini et al., 2024).

• Joint Geometry and Physics Reconstruction:

In the blood flow topology optimization setting, the pipeline iteratively solves for both the geometry field and boundary conditions using adjoint methods and implicit function theorem-based differentiation, embedding physical constraints for both flow and contrast transport (Thombre et al., 3 Jun 2026).

3. Applications and Use Cases

CTFlow methodologies have been demonstrated in several major domains:

Domain	Core Task/Problem	Exemplary CTFlow Realization
CT/MRI Synthesis	MRI→synthetic CT; CBCT→synthetic CT	Conditional flow-matching, 3D U-Net (Hadzic et al., 6 Oct 2025)
Sparse-View CT Reconstruction	Inverse imaging, rapid reconstruction	Deterministic ODE flow-matching (FMCT/EFMCT) (Shi et al., 27 Feb 2026)
Synthetic 3D CT Generation	Text/clinical report → CT volume generation	Latent flow-matching transformer with report conditioning (Wang et al., 18 Aug 2025)
Cardiovascular Simulation	CT-based FFR estimation, blood flow simulation	Automated pipeline, steady CFD (Nannini et al., 2024)
Hemodynamic Inversion	Geometry, flow from direct sinogram	PDE-constrained topology optimization (Thombre et al., 3 Jun 2026)
Physics-Informed Neural Inverse Problems	Flow recovery from CT data, PINNs	Data-fidelity in sinogram space, direct PINN inversion (Guo et al., 5 Nov 2025)
Continuous-Time Snapshot Dynamics	Dynamics from noisy/irregular snapshots	Partial OT time alignment, kernel smoothing, ODE/SDE vector field learning (Kawano et al., 23 May 2025)

In every case, CTFlow advances a physically or statistically principled approach to data-driven inversion, generation, or synthesis, leveraging continuous-time modeling for efficiency, stability, or increased physical fidelity.

4. Evaluation, Validation, and Comparative Benchmarks

Extensive evaluation highlights key strengths and issues:

• Accuracy:

CTFlow-based methods achieve state-of-the-art or competitive performance. For CT synthesis, global anatomical reconstruction is accurate, though fine structural details may be limited by training resolution (Hadzic et al., 6 Oct 2025). In sparse-view CT, EFMCT matches or slightly exceeds diffusion baselines in PSNR and SSIM, with substantially reduced neural function evaluations (Shi et al., 27 Feb 2026). For FFR-CT, predicted values correlate strongly with invasive gold-standards (r = 0.797, AUC = 0.923) (Nannini et al., 2024).

• Efficiency:

Flow-matching-based approaches allow order-of-magnitude speed-ups over diffusion-based models. FMCT/EFMCT achieves high-quality reconstructions in ≤2 s with as little as 7 network evaluations per image (Shi et al., 27 Feb 2026). In trajectory generation, up to 100× speed-up over diffusion is reported (Ye et al., 2024). Automated FFR-CT delivers 30× shorter runtime than transient CFD, enabling same-session clinical use (Nannini et al., 2024).

• Robustness and Data-Fidelity:

Embedding data consistency—via direct matching in sinogram space or differentiable projection—improves robustness to projection noise, motion artifacts, or data sparsity, and circumvents errors arising from intermediate (filtered backprojection) reconstructions (Guo et al., 5 Nov 2025, Thombre et al., 3 Jun 2026).

• Limitations:

Limitations include loss of fine detail due to low spatial resolution (image synthesis) (Hadzic et al., 6 Oct 2025), dependency on idealized scanner models (e.g., parallel-beam), and current restriction to 2D/steady-state in some physics-based pipelines (Thombre et al., 3 Jun 2026). Automation of boundary condition inference and mesh generation has reduced, but not eliminated, uncertainties in clinical simulations (Nannini et al., 2024).

5. Extensions and Future Directions

Several future directions and extensions are explicitly proposed:

• High-Resolution and Latent-Space Flows:

Patch-based training and flow-matching in learned latent spaces are suggested to overcome limits of GPU memory or to improve anatomical detail (Hadzic et al., 6 Oct 2025, Wang et al., 18 Aug 2025).

• 3D, Transient, and Compliance Models:

Extending 2D/steady geometry-physics frameworks to 3D, temporally resolved, and wall-compliant (FSI) regimes is a major focus for hemodynamics inversion (Thombre et al., 3 Jun 2026, Guo et al., 5 Nov 2025).

• Probabilistic, Multi-Modal, and Uncertainty Quantification:

Probabilistic extensions for capturing uncertainty (e.g., multiple plausible geometries or flows) are identified as important future work (Thombre et al., 3 Jun 2026, Kawano et al., 23 May 2025).

• Self-Supervised and Snapshot Learning:

New flow-matching frameworks for continuous dynamics from discrete, temporally noisy snapshot data advance ODE/SDE learning where fine trajectory data are unavailable, with rigorous alignment of timestamps and robust kernel smoothing (Kawano et al., 23 May 2025).

• Integration with Clinical Workflow:

The pipeline automation, rapid turnaround, and minimal intervention in clinical CTFlow enable new translational applications in immediate diagnosis, intervention planning, and non-invasive physiological assessment (Nannini et al., 2024).

6. Theoretical Insights and Guarantees

CTFlow is supported by several foundational theoretical results:

• Convergence:

Discretized flow-matching schemes converge to the true target distribution under mild regularity and proper scaling of step size and number of steps (Chen et al., 2017).

• Partial Optimal Transport Relaxation:

For time-label inference in snapshot datasets, the partial OT relaxation is shown to be a tight, tractable surrogate for mixed-integer boundary-labeling MILPs, with theoretical correctness in the infinite-resolution, low-noise regime (Kawano et al., 23 May 2025).

• Error Bounds in ODE Integration:

Velocity reuse and adaptive data consistency in FMCT/EFMCT introduce no greater error than ODE discretization itself, with global errors scaling as $O(\Delta t)$ over integration intervals (Shi et al., 27 Feb 2026).

• Adjoint-Based Gradients:

For PDE-constrained and topology optimization settings, exact gradients via adjoint systems and implicit differentiation allow efficient optimization even in large-scale, non-linear settings (Thombre et al., 3 Jun 2026).

• Unified Generative–Inference Frameworks:

Continuous-time flows unify generative modeling and variational inference in a principled, flexible architecture, with maximum-likelihood, adversarial, and Wasserstein-theoretic underpinnings (Chen et al., 2017).

In summary, CTFlow designates a set of models and pipelines at the intersection of deep generative modeling, physical simulation, and inverse problem theory, centered on deterministic or stochastic flow-matching in continuous time. By combining differentiable physical models and efficient ODE-based sample transport, CTFlow methods enable scalable, accurate, and automated solutions to a range of synthesis, reconstruction, and simulation tasks in medical imaging and dynamic systems inference (Shi et al., 27 Feb 2026, Hadzic et al., 6 Oct 2025, Wang et al., 18 Aug 2025, Thombre et al., 3 Jun 2026, Nannini et al., 2024, Guo et al., 5 Nov 2025, Kawano et al., 23 May 2025, Ye et al., 2024, Chen et al., 2017).