跳转至

物理引导的流匹配用于CT图像重建

文章背景与核心概要

在计算机断层扫描(CT)中,深度生成模型正越来越多地被用作强大的先验,以解决病态逆问题。尽管基于扩散的模型已经取得了最先进的成果,但它们通常依赖于随机采样、漫长的推理轨迹和复杂的噪声调度,这限制了其在高分辨率下的数值稳定性和计算效率。

本文引入了流匹配(Flow Matching)作为CT重建的强大替代生成先验:通过在梅奥诊所(Mayo Clinic)低剂量CT数据集的 \(256 \times 256\) 胸部图像上训练高分辨率的纠正流匹配(Rectified Flow Matching)模型,并采用两阶段策略来应对过拟合与解剖结构方差有限的问题。研究评估了多种流匹配重建技术,并将其与主流扩散算法进行对比,结果表明流匹配方法在峰值信噪比(PSNR)、结构相似性指数(SSIM)和感知指标上均可靠地优于扩散方法,同时需要更少的采样步数。


概览 (Overview)

作者: Davide Evangelista
提交时间: 2026年8月28日
主要学科: 人工智能 (cs.AI)
次要学科: 计算机视觉与模式识别 (cs.CV)
arXiv 标识符: arXiv:2608.28256
DOI: 10.48550/arXiv.2608.28256

Overview

Authors: Davide Evangelista
Submitted: August 28, 2026
Primary Subject: Artificial Intelligence (cs.AI)
Secondary Subject: Computer Vision and Pattern Recognition (cs.CV)
arXiv Identifier: arXiv:2608.28256
DOI: 10.48550/arXiv.2608.28256


摘要总结 (Abstract Summary)

深度生成模型正越来越多地被用作强大的先验,以求解计算机断层扫描(CT)中的病态逆问题。尽管基于扩散的模型已经取得了最先进的结果,但它们通常依赖于随机采样、漫长的推理轨迹和复杂的噪声调度,这限制了其在高分辨率下的数值稳定性和计算效率。

这项工作引入了流匹配(Flow Matching)作为CT重建的一种稳健的替代生成先验: * 模型训练: 在来自梅奥诊所低剂量CT数据集的 \(256 \times 256\) 胸部图像上训练了一个高分辨率的纠正流匹配(Rectified Flow Matching)模型。 * 两阶段策略: 为了解决过拟合和解剖学方差有限的问题,训练结合了第一阶段的强解剖学信息数据增强,随后是进行极小或无增强的微调阶段,以最大化结构保真度。 * 重建评估: 本文评估了多种流匹配重建技术——包括 Plug-and-Play FlowFlowDPSFlowerFlow-Priors (ICTM)——并将其与诸如 DDRMDPSDiffPIR 的著名扩散算法进行了对比。 * 主要发现: 流匹配方法在峰值信噪比(PSNR)、结构相似性指数(SSIM)和感知指标上均稳定地优于扩散方法,同时所需的采样步数更少。

Abstract Summary

Deep generative models are increasingly used as powerful priors to solve ill-posed inverse problems in Computed Tomography (CT). While diffusion-based models have achieved state-of-the-art results, they often rely on stochastic sampling, lengthy inference trajectories, and complex noise schedules, limiting their numerical stability and computational efficiency at high resolutions.

This work introduces Flow Matching as a robust alternative generative prior for CT reconstruction: * Model Training: A high-resolution Rectified Flow Matching model is trained on \(256 \times 256\) chest images from the Mayo Clinic Low-Dose CT dataset. * Two-Stage Strategy: To tackle overfitting and limited anatomical variance, the training incorporates an initial phase of strong anatomically informed data augmentation, followed by a fine-tuning phase with minimal or no augmentation to maximize structural fidelity. * Reconstruction Evaluation: The paper evaluates multiple Flow Matching reconstruction techniques—including Plug-and-Play Flow, FlowDPS, Flower, and Flow-Priors (ICTM)—and contrasts them with prominent diffusion algorithms like DDRM, DPS, and DiffPIR. * Key Findings: Flow Matching approaches reliably outperform diffusion methods across Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and perceptual metrics while demanding fewer sampling steps.


访问与资源 (Access & Resources)

Access & Resources