[KongchangAI]
· 2 min read· 1,400 words

Geometric Bridge Flow Matching: A New Breakthrough in Blind Rigid Motion Correction

Geometric Bridge Flow Matching: A New Breakthrough in Blind Rigid Motion Correction

A geometric bridge flow matching prior enables blind rigid motion correction in head CBCT, achieving state-of-the-art results on simulated data.

This paper proposes a blind rigid motion correction method based on a 3D flow matching prior for head cone-beam CT. Its core is a "geometric bridge" that uses forward projector geometric derivatives to provide analytic velocity targets, connecting motion-corrupted and motion-free FDK reconstructions through gradually decreasing simulated motion to learn motion-induced image changes. Inference uses a plug-and-play iterative framework alternating flow matching prior updates, spline-based pose estimation, conjugate gradient data consistency, and total variation regularization. On simulated data from 30 CQ500 patients, average PSNR reaches 37.1 dB and RPE is 0.27 mm, outperforming learned autofocus and diffusion methods; the geometric bridge improves PSNR by 1.6 dB and reduces RPE by 33% over image-linear training. Processing takes about 9.1 minutes per patient with no extra motion monitoring equipment.

Geometric Bridge Flow Matching: A New Breakthrough in Blind Rigid Motion Correction

During head cone-beam CT (CBCT) scans, rigid patient head motion degrades image quality and makes it difficult to recover accurate anatomical structures from projection data. To address this challenge, researchers proposed a method based on a 3D flow matching prior that learns motion-induced image changes to achieve blind rigid motion correction. The method achieves state-of-the-art reconstruction quality and motion estimation accuracy on simulated data.

Overview of the geometric bridge flow matching method

Core Method: Geometric Bridge Flow Matching

The key innovation of this work is the construction of a "geometric bridge" that connects motion-corrupted and motion-free FDK reconstructions of the same source volume through gradually decreasing simulated motion. Specifically, the method learns a 3D flow matching prior that captures the patterns of motion-induced image changes.

The construction of the geometric bridge leverages the geometric derivatives of the forward projector to provide analytic velocity targets for the transformation along the path. This design enables the model to understand how motion affects the reconstructed image, thereby better correcting motion artifacts during inference.

According to the original text, the geometric bridge connects motion-corrupted and motion-free FDK reconstructions of the same source volume by gradually reducing simulated motion, with the geometric derivatives of the forward projector providing analytic velocity targets along the path.

Flow Matching is a class of generative modeling methods whose core idea is to learn a continuous-time velocity field that gradually transforms a simple prior distribution (such as Gaussian noise) into the target data distribution. Unlike diffusion models, which require simulating a stochastic forward noising process, flow matching directly constructs a deterministic probability path from source to target (commonly an optimal transport path) and regresses the conditional velocity field along the path as the training objective. At inference time, generation is completed simply by numerically integrating an ordinary differential equation. Because the velocity field is analytically definable, the training objective is clear and stable, and sampling efficiency is typically superior to diffusion models. In this paper, it is used as an image prior, learning the transformation direction from "motion-corrupted image → motion-free image" rather than stochastic generation.

FDK reconstruction is an analytic reconstruction algorithm widely used in cone-beam CT, proposed by Feldkamp, Davis, and Kress. It is essentially a generalization of filtered backprojection to cone-beam geometry: projection data are first distance-weighted and ramp-filtered, then backprojected along the cone-beam ray directions onto the 3D voxel grid. FDK is approximately exact under circular trajectories and small cone angles, is computationally fast, and is the standard algorithm for clinical CBCT. However, when the patient moves, projections from different views correspond to different anatomical positions, and backprojection superposition produces obvious artifacts (such as streaks, blurring, and ghosting). Therefore, FDK is often used as the initial reconstruction or data consistency reference in motion correction iterations.

Plug-and-Play Optimization Framework

The method combines the learned image updates with spline-based pose estimation, conjugate gradient data consistency, and image total variation regularization to form a plug-and-play iterative optimization loop. This design allows flexible integration of multiple regularization approaches, improving reconstruction robustness.

In each iteration, the image is first updated using the flow matching prior, and then further optimized through pose estimation and data consistency steps. Total variation regularization is used to suppress noise and artifacts. The entire process alternates between the image domain and the projection domain, ensuring that the final result both fits the measured data and satisfies prior constraints.

Conjugate Gradient (CG) data consistency refers to the step of minimizing the projection-domain residual ‖A x − y‖² in optimization, where A is the forward projection operator, x is the image to be reconstructed, and y is the measured projection. The normal equation AᵀA x = Aᵀy corresponding to this quadratic objective is symmetric positive definite, and the conjugate gradient method can efficiently solve it iteratively without explicitly constructing large-scale matrices, with faster convergence than steepest descent. In plug-and-play frameworks, CG typically handles the "data fidelity" subproblem, alternating with the subproblem responsible for "prior regularization," which is the typical structure of splitting algorithms such as ADMM and Plug-and-Play.

Total Variation (TV) regularization was proposed by Rudin, Osher, and Fatemi. It suppresses noise and artifacts by penalizing the L1 norm of the image gradient while preserving edges. Its advantage is that it allows piecewise-constant-like images (such as reconstructions with clear anatomical boundaries) to be denoised without blurring boundaries. In CT reconstruction, TV is often written as λ‖∇x‖₁ and used together with a data consistency term; however, it can also cause staircase effects and over-smooth texture-rich regions, so in practice it is often combined with other priors or weighted variants.

Experimental Setup and Results

The researchers conducted simulated 360-view scan experiments on 30 held-out CQ500 patient datasets, with motion magnitudes for each component nominally within 10 mm/10°. Compared with methods such as learned autofocus and diffusion-based joint reconstruction, the framework achieves state-of-the-art performance in both image reconstruction quality and motion accuracy.

In terms of specific metrics, the average peak signal-to-noise ratio (PSNR) reaches 37.1 dB, and the average detector-plane reprojection error (RPE) is 0.27 mm. When the inference setting matches the endpoints, the geometric bridge improves PSNR by 1.6 dB and reduces RPE by 33% relative to image-linear training.

Original data: average PSNR 37.1 dB, average RPE 0.27 mm; the geometric bridge improves PSNR by 1.6 dB and reduces RPE by 33% compared with image-linear training.

Peak Signal-to-Noise Ratio (PSNR) is a commonly used metric for measuring pixel-level error between a reconstructed image and a reference image, defined as PSNR = 10·log₁₀(MAX²/MSE), where MAX is the maximum possible pixel value and MSE is the mean squared error. It is expressed in dB, and higher values indicate closer similarity to the reference. It is simple to compute and comparable across methods, but it does not fully align with human visual perception, so it is often used together with structural similarity metrics such as SSIM.

Detector-plane reprojection error (RPE) measures the average geometric deviation (often in millimeters) between the forward projection of the reconstructed volume onto the detector plane under estimated motion parameters and the measured projection. RPE directly reflects the accuracy of motion estimation: the smaller the RPE, the more geometrically consistent the reconstructed image and projection data are. Unlike image-domain metrics, RPE is evaluated in the projection domain and is more sensitive to motion parameter errors, so it is often used as a geometric accuracy metric for motion correction methods.

Computational Efficiency and Practical Applications

On an RTX A6000 GPU, with 50 prior updates and coarse-to-fine motion estimation, the average processing time is approximately 9.1 minutes per patient. This runtime is potentially feasible for clinical workflows, especially in scenarios requiring high-precision reconstruction.

A notable advantage of this method is that it requires no additional motion monitoring equipment; rigid motion can be estimated and corrected solely from projection data, falling into the category of "blind" correction. This reduces hardware cost and operational complexity and is expected to be useful in applications such as head CBCT-guided radiotherapy.

Summary and Outlook

The geometric bridge flow matching method provides a novel solution for blind rigid motion correction. By combining a flow matching prior with geometric information, it demonstrates superior performance on simulated data. Future work may include validation on real clinical data, further reduction of processing time, and extension to more complex motion patterns.

One-line Summary

Proposes a geometric bridge flow matching prior for blind rigid motion correction in head CBCT, achieving state-of-the-art results on simulated data.

Paragraph Summary

This paper proposes a blind rigid motion correction method based on a 3D flow matching prior for head cone-beam CT. Its core is the construction of a "geometric bridge" that uses the geometric derivatives of the forward projector to provide analytic velocity targets, connecting motion-corrupted FDK reconstructions and motion-free reconstructions of the same source volume through gradually decreasing simulated motion, thereby learning patterns of motion-induced image changes. During inference, a plug-and-play iterative framework alternates flow matching prior updates, spline-based pose estimation, conjugate gradient data consistency, and total variation regularization. On simulated data from 30 CQ500 patients, the average PSNR reaches 37.1 dB and RPE is 0.27 mm, outperforming learned autofocus and diffusion methods; the geometric bridge improves PSNR by 1.6 dB and reduces RPE by 33% relative to image-linear training. Processing takes about 9.1 minutes per patient and requires no additional motion monitoring equipment.

#geometric bridge flow matching#blind rigid motion correction#cone-beam CT#flow matching prior#FDK reconstruction#motion artifact correction#CBCT reconstruction
Share:

Related articles