A Denoising Framework for Real-World Ultra-Low-Dose Lung CT Images Based on an Image Purification Strategy

TL;DR

This paper tackles the challenge of denoising real-world ultra-low-dose lung CT (uLDCT) by addressing spatial misalignment between uLDCT and normal-dose CT (NDCT) data. It introduces an Image Purification (IP) strategy to construct structurally aligned training pairs and a Frequency-domain Flow Matching (FFM) model to preserve anatomical contours during reconstruction. The approach is validated on a real-world paired uLDCT/NDCT dataset, showing consistent improvements in contour-preservation metrics and achieving state-of-the-art performance for structure-sensitive denoising. The work offers a practical path toward clinically viable uLDCT denoising by combining data purification with frequency-domain diffusion modeling.

Abstract

Computed Tomography (CT) is a vital diagnostic tool in clinical practice, yet the health risks associated with ionizing radiation cannot be overlooked. Low-dose CT (LDCT) helps mitigate radiation exposure but simultaneously leads to reduced image quality. Consequently, researchers have sought to reconstruct clear images from LDCT scans using artificial intelligence-based image enhancement techniques. However, these studies typically rely on synthetic LDCT images for algorithm training, which introduces significant domain-shift issues and limits the practical effectiveness of these algorithms in real-world scenarios. To address this challenge, we constructed a real-world paired lung dataset, referred to as Patient-uLDCT (ultra-low-dose CT), by performing multiple scans on volunteers. The radiation dose for the low-dose images in this dataset is only 2% of the normal dose, substantially lower than the conventional 25% low-dose and 10% ultra-low-dose levels. Furthermore, to resolve the anatomical misalignment between normal-dose and uLDCT images caused by respiratory motion during acquisition, we propose a novel purification strategy to construct corresponding aligned image pairs. Finally, we introduce a Frequency-domain Flow Matching model (FFM) that achieves excellent image reconstruction performance. Code is available at https://github.com/MonkeyDadLufy/flow-matching.

Figures (11)

• Figure 1: Motivation for the Image Purification (IP) strategy. (a) Ultra-low-dose CT (uLDCT) image with severe noise affecting structural clarity. (b) Corresponding normal-dose CT (NDCT) image for (a). Zoomed-in black areas show anatomical structure morphology. (c) NDCT image after IP processing, used as the label during training. (d) NDCT image after image registration. (e) Denoising result of end-to-end trained Flow Matching (FM) FM on (a). (f) Denoising result of Flow Matching FM trained with the PSPpsp strategy on (a). (g) Denoising result of our Flow Matching FM trained with the IP strategy on (a). Red arrows in (c)-(g) indicate inconsistent structures; green arrows indicate consistent structures.
• Figure 2: Concept of the Image Purification (IP) Strategy. (a)Mapping relationship of input data($X_0$,$X_1$) $\sim$ (NDCT,uLDCT).Structural misalignment between $X_0$ and $X_1$ leads to intersecting trajectories. (b)By establishing a new distribution $\mathrm{X}_1{ }^{\prime} \sim IP(uLDCT)$,decompose $v$ into structure-consistent $v_1$ and texture-consistent $v_2$.Use $v_1$ as the training data for the model. (c)During inference, $X_1$ is implicitly mapped from uLDCT to IP(uLDCT), and then reaches NDCT via the trained velocity field. (d)Parameter T influences the magnitude of $v_1$.
• Figure 3: Overall Framework.
• Figure 4: Effect of the Image Purification (IP) Strategy on Straightening Network Paths. We define the crossover rate as the proportion of samples with intersecting paths in the dataset. SSIM0.95, SSIM0.90, SSIM0.85 represent the cases where the similarity at the crossover point is $\geq$ 0.95, 0.90, 0.85, respectively, among the samples with crossovers. Details are in the \ref{['sec:补充材料']}.
• Figure 5: Proposed Image Purification (IP) strategy workflow.