Efficient Flow Matching for Sparse-View CT Reconstruction

This paper proposes FMCT and its efficient variant EFMCT, which leverage the deterministic nature of Flow Matching and a velocity field reuse strategy to achieve high-quality, computationally efficient sparse-view CT reconstruction with bounded error and significantly fewer neural network function evaluations compared to diffusion-based methods.

The Gist

Imagine you are trying to reconstruct a blurry, incomplete photograph of a person, but you only have a few scattered puzzle pieces (the "sparse-view" CT scans). Your goal is to fill in the missing parts to create a clear, complete picture.

In the medical world, this is a huge challenge. Doctors need these pictures fast, especially in emergencies.

Here is the story of how this paper solves that problem, using simple analogies.

1. The Old Way: The "Drunk Navigator" (Diffusion Models)

Previously, the best AI tools for this job were called Diffusion Models. You can think of them as a drunk navigator trying to find their way home.

• How it works: The navigator starts far away from the house (a random mess of pixels) and tries to walk toward the house (the clear image).
• The Problem: Because the navigator is "drunk" (stochastic/random), they stumble and sway left and right. Every time they take a step, they have to check a map (the "data consistency" check) to make sure they are still on the right path.
• The Conflict: The map says "Go straight," but the drunk navigator keeps swaying. They have to fight against their own stumbling. This "push-and-pull" makes the journey slow and unstable. To get a good result, the navigator has to take thousands of tiny, shaky steps. In a hospital, waiting for thousands of steps is too slow.

2. The New Way: The "GPS Driver" (Flow Matching)

The authors of this paper introduced a new method called Flow Matching (FM). Think of this as a high-tech GPS driver.

• How it works: Instead of stumbling, the GPS calculates a perfectly smooth, straight line from the starting point to the destination. There is no randomness, no swaying. It's a deterministic (predictable) path.
• The Benefit: Because the path is smooth and straight, the "map check" (data consistency) works perfectly with the driver's instructions. They don't fight each other; they work together. This makes the journey much more stable.

3. The Secret Sauce: "Skipping Steps" (Velocity Reuse)

Even with a GPS driver, calculating the exact direction for every single step takes time. The computer has to ask the AI, "Which way should I go?" thousands of times.

The authors noticed something clever: The GPS driver doesn't need to recalculate the direction every single second.

• The Observation: If you are driving down a straight highway, your steering wheel doesn't change much from second to second. You can keep going in the same direction for a while before you need to check the map again.
• The Innovation (EFMCT): The authors created a strategy called Velocity Reuse.
  • The AI calculates the direction once.
  • Then, it says, "Okay, I'll keep driving in this exact same direction for the next 5 or 10 steps without asking the AI again."
  • It only asks the AI for a new direction if it starts to drift too far off course.

4. Why This Matters

• Speed: By reusing the "direction" (velocity) multiple times, the computer doesn't have to do the heavy math as often. It's like skipping the "calculate route" button on your GPS and just cruising for a bit.
• Safety: The authors proved mathematically that skipping these calculations doesn't ruin the picture. The "drift" is so small that the "map check" (the physics of the CT scan) easily corrects it.
• Result: They achieved the same high-quality image as the old methods but in half the time (or even less).

The Bottom Line

Imagine you are painting a masterpiece, but you only have a few clues.

• Old Method: You take a step, stumble, check your clues, stumble again, check your clues... it takes forever.
• New Method: You draw a smooth line, and then you realize, "I can keep sliding along this line for a while without stopping to check my notes."

This paper gives doctors a faster, smarter way to see inside the human body, potentially saving lives by getting clear images to the surgeon's screen much quicker.

Technical Summary

1. Problem Statement

Sparse-View CT Reconstruction is an ill-posed inverse problem where the number of projection measurements ($n$) is significantly lower than the number of image pixels ($m$). This leads to severe artifacts and information loss.

• Current State-of-the-Art (SOTA): Generative models, particularly Diffusion Models (DM), have shown promise by learning expressive image priors. However, they rely on Stochastic Differential Equations (SDEs) for sampling.
• Key Limitations of DMs:
  1. Stochasticity vs. Data Consistency: CT reconstruction requires repeated "data consistency" corrections (ensuring the image matches the physical measurements). The stochastic noise injection in DMs interferes with these corrections, causing a "push-and-pull" effect that leads to instability and requires thousands of iterations (Neural Network Function Evaluations or NFEs) to converge.
  2. Computational Cost: In clinical settings (emergency/interventional), reconstruction speed is critical. The high NFE count of diffusion methods makes them impractical for time-critical workflows.

2. Methodology

The authors propose FMCT (Flow Matching for CT) and its efficient variant EFMCT (Efficient FMCT).

A. Flow Matching (FM) Framework

Unlike Diffusion Models, Flow Matching models the generation process as a deterministic Ordinary Differential Equation (ODE) rather than an SDE.

• Mechanism: It learns a time-dependent velocity field $v_t(x_t)$ that transports samples from a source distribution (noise) to a target distribution (CT images) along straight-line trajectories (Rectified Flow).
• Advantage: The deterministic nature allows for smooth trajectories without stochastic noise, making it naturally compatible with repeated data consistency operations without the instability seen in DMs.

B. Integration with CT Reconstruction

The sampling process integrates physics-based data consistency at each step:

1. Flow Transport: An explicit Euler step moves the image along the learned velocity field.
2. Data Consistency (DC): The algorithm estimates the image at $t=0$, applies a data consistency operator (implemented via Conjugate Gradient) to align the image with the measured projections $y = Ax$, and calculates a correction vector.
3. Update: The final step combines the flow transport and the data consistency correction.

C. The EFMCT Innovation: Velocity Reuse

The core contribution of this paper is the Velocity Reuse Strategy to drastically reduce NFEs.

• Observation: The predicted velocity fields in FM exhibit strong correlation across consecutive time steps, resulting in near-straight trajectories.
• Strategy: Instead of evaluating the neural network at every integration step, the method reuses the velocity vector predicted at step $t$ for up to $M$ subsequent steps.
• Adaptive Control: An adaptive check ensures that the data fidelity ($\|Ax - y\|^2$) does not degrade beyond a relaxation factor ($\eta$). If the condition is violated, the velocity is recomputed.
• Theoretical Guarantee: The authors prove that the error introduced by reusing a velocity is of the same order as the Euler discretization error ($O(\Delta t^2)$). When combined with non-expansive data consistency operators, the accumulated error remains bounded, preserving the convergence behavior.

3. Key Contributions

1. First FM-based CT Framework: Proposes the first application of Flow Matching to CT reconstruction, establishing a deterministic alternative to Diffusion Models.
2. Efficient Velocity Reuse: Introduces a strategy to reuse predicted velocity fields, reducing the number of Neural Network Function Evaluations (NFEs) by up to 89% without significant quality loss.
3. Theoretical Analysis: Provides a proof showing that velocity reuse introduces errors comparable to standard Euler discretization and that these errors remain bounded when coupled with data consistency corrections.
4. Comprehensive Evaluation: Demonstrates competitive reconstruction quality against SOTA diffusion methods (DPS, MCG, PGDM, DDS) and classical methods (FBP, ADMM-TV) on multiple datasets.

4. Experimental Results

Experiments were conducted on the AAPM 2016 Low Dose CT and Medical Segmentation Decathlon datasets with 20 and 40 views (sparse-view scenarios).

• Reconstruction Quality:
  • FMCT/EFMCT achieve PSNR and SSIM scores comparable to or better than diffusion-based baselines.
  • Visual Quality: Generative methods (FM/DM) produce sharper images with finer details compared to the overly smooth results of ADMM-TV, though they may show slight deviations in fine-scale details due to the ill-posed nature of the problem.
• Efficiency:
  • NFE Reduction: EFMCT reduces NFEs from ~50 (FMCT) to 7–11 (depending on the dataset), a 75–89% reduction.
  • Time Savings: Reconstruction time is reduced by 78% compared to the most efficient diffusion baseline (DDS) and significantly outperforms standard diffusion samplers (which take ~140s vs. ~0.8s for EFMCT on RTX 4090).
• Ablation Studies:
  • Reusing velocities too early in the sampling process degrades performance.
  • Reusing velocities in the mid-to-late stages (after the first iteration) is optimal.
  • A limit of ~10 consecutive reuse steps offers the best trade-off between speed and accuracy.

5. Significance

This work addresses the critical bottleneck of computational efficiency in generative medical imaging.

• Clinical Viability: By reducing reconstruction time from minutes to seconds while maintaining high quality, the proposed method (EFMCT) makes advanced generative priors viable for time-critical clinical scenarios (e.g., emergency rooms, interventional radiology).
• Paradigm Shift: It demonstrates that deterministic Flow Matching is a superior foundation for inverse problems requiring strict data consistency compared to stochastic Diffusion Models.
• Open Source: The authors have open-sourced the codebase, facilitating further research and adoption in the medical imaging community.