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

AI summary

Overview

Research area: Medical image processing / low-dose CT denoising, combining a data-alignment (purification) strategy with a generative Frequency-domain Flow Matching model.

Technical level: Advanced. The paper assumes familiarity with diffusion and flow-matching generative models, optimal transport, image registration, and CT image quality metrics.

Scope: The paper builds a real-world paired ultra-low-dose lung CT dataset, proposes an Image Purification (IP) strategy to fix anatomical misalignment between ultra-low-dose and normal-dose scans, and introduces a Frequency-domain Flow Matching (FFM) model evaluated against seven denoising networks on this dataset.

What This Paper Is About

Most low-dose CT denoising research trains and tests on synthetic or phantom data whose low-dose and normal-dose images are already anatomically aligned, whereas real patient scans are misaligned because of breathing and heartbeat between acquisitions. The authors collect a real paired lung dataset in which the low-dose scans use only 2% of the normal radiation dose, then propose an Image Purification strategy that rewrites anatomical contours in one image of each pair using contours from the other, producing aligned training and evaluation pairs. On top of this purified data they train a Frequency-domain Flow Matching model that is designed to preserve the patient's original anatomical structure while removing noise.

Key Contributions

1. A real clinical ultra-low-dose lung CT dataset (Patient-uLDCT). Collected from 6 patients with 4310 paired uLDCT and NDCT images, where the low-dose scans are approximately 2% of the normal dose, below the conventional 25% low-dose and 10% ultra-low-dose levels.

2. An Image Purification (IP) strategy for spatial misalignment. Operating at the image level, it uses Otsu-based binarization and a Common Mask to decompose the residual between uLDCT and NDCT into a structure-changing component and a texture-changing component, then recombines them to build structurally aligned pairs.

3. A Frequency-domain Flow Matching (FFM) model. Flow Matching's main network is migrated from the image domain to the frequency domain using a 2D FFT, separate 1x1 convolutions on magnitude and phase, and an inverse FFT, to strengthen anatomical contour preservation.

4. Demonstration that the IP strategy improves seven LDCT denoising networks (Flow Matching, CoreDiff, Cold Diffusion, REDCNN, EDCNN, DuGAN, MMCA) on the real patient dataset, with FFM plus IP reaching the best reported result.

Main Findings

• Misalignment is the core obstacle. Direct end-to-end training on misaligned uLDCT–NDCT pairs causes the model to learn incorrect mappings, producing anatomical distortion; even the advanced registration method uniGradICON still showed deficiencies in overall resolution and anatomical registration accuracy.

• The IP strategy straightens the mapping paths. Measured by a defined "crossing rate" (proportion of samples with intersecting paths in the dataset), IP reduced the crossing rate by a factor of 2 to 6 across similarity thresholds SSIM 0.95, 0.90, and 0.85.

• IP improves the data itself. Comparing uLDCT against NDCT versus IP-processed uLDCT against NDCT: FSIM rose from 0.5020 ± 0.0717 to 0.7889 ± 0.0871 (57%), GMSD fell from 0.7196 ± 0.0130 to 0.5870 ± 0.0982 (18%), and SSIM rose from 0.4788 ± 0.1197 to 0.7814 ± 0.1118 (63%).

• Every compared network improved on contour metrics with IP. Flow Matching + IP: FSIM 0.9262 ± 0.0359 (8%), GMSD 0.3691 ± 0.0627 (28%), SSIM 0.9160 ± 0.0639 (5%). Cold Diffusion + IP: FSIM 0.9308 ± 0.0302 (7%), GMSD 0.3788 ± 0.0551 (27%), SSIM 0.9120 ± 0.0577 (2%). DuGAN + IP: FSIM 0.9107 ± 0.0396 (14%), GMSD 0.3931 ± 0.0627 (25%), SSIM 0.8900 ± 0.0761 (6%). MMCA + IP: FSIM 0.8935 ± 0.0505 (38%), GMSD 0.4317 ± 0.0839 (28%), SSIM 0.8884 ± 0.0781 (11%). CoreDiff + IP showed the only negative cell reported: SSIM 0.9057 ± 0.0737, a 1% decrease from CoreDiff's 0.9134 ± 0.0408.

• Simpler models gained more from IP than strong generative models. REDCNN improved FSIM by 12% (0.8047 ± 0.0339 to 0.8985 ± 0.0538) and GMSD by 39% (0.6941 ± 0.0728 to 0.4201 ± 0.0896); EDCNN improved FSIM by 16% and GMSD by 30%. The authors interpret this as evidence that data alignment matters more than model architecture complexity for uLDCT.

• Self-supervised and latent-space methods underperformed. N2N (FSIM 0.5406 ± 0.0902, GMSD 0.7215 ± 0.0278, SSIM 0.5484 ± 0.1546) and IPDM (FSIM 0.5205 ± 0.0656, GMSD 0.7346 ± 0.0176, SSIM 0.5316 ± 0.1264) scored low for lack of paired supervision; Resshift (FSIM 0.8073 ± 0.0206, GMSD 0.6864 ± 0.0313, SSIM 0.7538 ± 0.0290) lost high-frequency detail through latent-space encoding and decoding, and its outputs contained visible horizontal stripe artifacts.

• FFM + IP gave the best result in Table 1. FSIM 0.9342 ± 0.0341, GMSD 0.3708 ± 0.0665, SSIM 0.9230 ± 0.0625.

• Visual comparisons supported the quantitative results. Without IP, Flow Matching, Cold Diffusion, and MMCA tended to fill in contours present in the uLDCT; CoreDiff and REDCNN deformed some contours; EDCNN and DuGAN blurred original contours. All models trained with IP-corrected pairs preserved the uLDCT anatomical contours. N2N showed almost no denoising capability and IPDM produced over-exposed results.

Methodology in Plain English

The authors first solve a data problem, then a modeling problem.

For the data problem, they note that in a real scan pair the two images come from two separate acquisitions, so the anatomy does not line up. Instead of registering the images or throwing away mismatched patches, they rewrite the anatomy:

1. Apply Otsu's Method to both the uLDCT and NDCT image to get binary masks separating anatomical contours from texture.
2. Combine the two masks into a Common Mask: a pixel is marked 1 if either mask marks it as contour.
3. Decompose the difference between the two images into a structure part and a texture part using the Common Mask and its complement (J − CM). Then build purified images: IP(uLDCT) = (J − CM) · uLDCT + CM · NDCT, and IP(NDCT) = (J − CM) · NDCT + CM · uLDCT. A tunable parameter T controls how much NDCT texture is blended into IP(uLDCT); T = 0 reduces to the base formula, and T = 1 makes IP(uLDCT) equal NDCT.
4. Choose the training pairing. Combination I (default): NDCT as label, IP(uLDCT) as input. Combination II: IP(NDCT) as label, uLDCT as input. Combination III: mix of the two.
5. Use IP(NDCT) as the evaluation label, since it carries the uLDCT's contours and the NDCT's texture, which matches what a good denoised output should look like.

For the modeling problem, they take Flow Matching as the baseline, which learns a velocity field along a straight interpolation path between a noisy image and a clean image, and train it with a mean squared error loss on the predicted versus true velocity. Their FFM variant moves the network's internal processing into the frequency domain at each U-Net layer: a 2D FFT produces magnitude and phase, two separate 1x1 convolutions process them, and an inverse FFT returns features to the spatial domain. Training runs for 200 epochs on an NVIDIA RTX 4090, and inference uses the Euler method with 10 time steps.

Why This Matters

Impact on research. The work argues that for ultra-low-dose CT, dataset alignment is a prerequisite that can outweigh architectural novelty: the IP strategy improved every one of the seven networks tested, and the simplest models (REDCNN, EDCNN) benefited most. It also supplies a real paired dataset at a 2% dose level, which is below the synthetic and phantom benchmarks the field commonly uses, such as MAYO2016 (about 25% for head and abdomen, 10% for lung) and Piglet (50%, 25%, 10%, 5% of normal dose).

Real-world applications:

• Lung cancer screening programs, where the National Lung Screening Trial reported a 20% reduction in lung cancer mortality with LDCT screening and where repeated imaging over years makes dose reduction clinically important.
• Pediatric and pregnant-patient imaging, named in the paper as groups vulnerable to ionizing radiation.
• COVID-19-era and other high-volume CT workflows, where the paper notes CT examination frequency increased significantly.
• Long-term or frequent-screening patients who accumulate cumulative radiation exposure.

Industry relevance. The framework is modular, with decoupled data preprocessing, model training, and evaluation stages, so the IP strategy can be plugged into existing denoising pipelines rather than replacing them. The code is released at https://github.com/MonkeyDadLufy/flow-matching, and the finding that data alignment unlocks gains across many architectures is directly relevant to scanner vendors and clinical AI vendors who already have deployed denoising models.

Future Directions

• Tuning the purification strength. The paper introduces parameter T to control how far IP(uLDCT) sits from the NDCT distribution (T = 0 leaves the purified image farthest from NDCT; T = 1 makes the two distributions overlap completely) and states that T is examined with all seven metrics in ablation studies, but the ablation results themselves are not reported in the provided text.
• Choosing among training combinations. Three pairings are defined (I, II, III), with Combination I as the default, and the paper states that exhaustive comparative analyses were conducted in ablation studies, but those results are not included in the provided content.
• Closing the gap for self-supervised methods. N2N and IPDM performed poorly here because they assume noise follows a known statistical distribution, while real uLDCT noise is highly coupled with anatomy; the paper leaves open how to remove such coupled noise without paired supervision.
• Recovering high-frequency detail in latent-space models. Resshift's latent encoding and decoding lost high-frequency information and produced stripe artifacts, pointing to a bottleneck for latent diffusion approaches in ultra-low-dose reconstruction.

Also open, and not reported in the provided text: the paper says Section 5 concludes the work, but the conclusion is absent from the supplied content, and no study of generalization beyond the 6 patients, nor any radiologist reader study, is reported.

Target Audience

Medical imaging and computer vision researchers working on low-dose CT denoising and generative image restoration; clinical translation researchers who need realistic paired datasets rather than synthetic benchmarks; and applied engineers at medical imaging or CT scanner companies who want a drop-in data-preprocessing strategy (IP) that improves existing denoising architectures. Readers without background in flow matching, optimal transport, or image quality metrics will find the modeling sections demanding, though the IP strategy itself is described step by step and can be understood on its own.

Authors’ 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.