Research map / Image generation / Image restoration
Motion Artifact Removal in Pixel-Frequency Domain via Alternate Masks and Diffusion Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PFAD, an unsupervised MRI motion-artifact removal method that guides a pre-trained DDPM with low-frequency k-space information and alternating complementary masks over high-frequency and pixel components, balanced by a time-dependent weight.
Problem
Removing motion artifacts from MRI without paired clean/corrupted data, while accounting for artifacts that corrupt k-space (frequency domain) phase information.
Approach
Use a diffusion model pre-trained only on clean MRI, guided at inference by low-frequency k-space content of the corrupted image, with alternating complementary masks (flipped each reverse step) over high-frequency and pixel components, weighted by the DDPM noise term and fused across domains via a time-dependent balance parameter.
Key results
- On simulated brain images, PFAD reaches PSNR 27.60 vs 26.76 for the best baseline (Pix2pix), but Pix2pix has higher SSIM (0.841 vs 0.829).
- On real clinical abdominal images, two radiologists gave PFAD an overall Likert score of 4.63, versus 4.15 for DR2 (the next best).
- Ablation: frequency-only removal gives PSNR 27.60 while pixel-only gives 23.89 on brain, showing the frequency domain drives structure recovery.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PFAD instead of Pix2pix | Brain Random simulated · Brain Random simulated | PSNR: 26.76 -> 27.60 | Helped |
| Use PFAD instead of Pix2pix | Brain Random simulated · Brain Random simulated | SSIM: 0.841 -> 0.829 | Hurt |
| Combine frequency- and pixel-domain removal instead of pixel-domain removal only (ablation) | PFAD · HCP simulated | PSNR: 23.89 -> 27.60 | Helped |
| Add pixel-domain removal to frequency-domain removal (ablation) | PFAD · HCP simulated | Lpips: 0.154 -> 0.147 | Helped |
| Add masks to the model (ablation) | PFAD · HCP simulated | PSNR: 26.63 -> 26.01 | Hurt |
| Add frequency- and pixel-domain weights to masked guidance (ablation) | PFAD · HCP simulated | PSNR: 26.01 -> 27.60 | Helped |
| Set a to 0.9 rather than 0.7 (ablation) | PFAD · HCP simulated | Total: 30.513 -> 30.469 | Hurt |
| Set a to 0.9 rather than 0.7 (ablation) | PFAD · HCP simulated | Lpips: 0.147 -> 0.135 | Helped |
| Set the filter cutoff to π/10 rather than π/5 or π/20 (ablation) | PFAD · HCP simulated | PSNR: 25.89 / 26.34 -> 27.60 | Helped |
| Set checkerboard grid size to 16 rather than 64 or 128 (ablation) | PFAD · HCP simulated | PSNR: 27.42 / 27.24 -> 27.60 | Helped |
| Use PFAD instead of DR2 | Real clinical abdominal images · Real clinical abdominal images | Overall radiologist score: 4.15 -> 4.63 | Helped |
Surprising findings
- Pix2pix, a supervised method trained on paired data, achieves higher SSIM than PFAD on simulated brain images.
- Removing the mask-weight omega or using a fixed mask degrades results substantially relative to the full alternating scheme.
Limitations
- Simulated results lag supervised Pix2pix on some metrics (SSIM on brain).
- Cutoff frequency and mask grid size need careful selection, with performance sensitive to these choices.
- Real-artifact evaluation relies on qualitative radiologist scoring on 40 abdominal images, without quantitative ground truth.
Experimental setup
Datasets: HCP brain MRI fastMRI knee MRI private T2-weighted abdominal MRI (Xiangyang No. 1 People's Hospital)
Benchmarks: HCP brain simulated fastMRI knee simulated abdominal simulated real clinical abdominal T2-weighted images
Metrics: PSNR SSIM LPIPS FSIM VIF VSI GMSD Likert 5-scale radiologist score
Baselines: Pix2pix CycleGAN UDDN DR2 GDP
Builds on: DDPM guided diffusion DR2 GDP
Closest papers by meaning
- Autoregressive Image Diffusion: Generation of Image Sequence and Application in MRI (NeurIPS 2024 · similarity 0.86)
- Boosting ViT-based MRI Reconstruction from the Perspectives of Frequency Modulation, Spatial Purification, and Scale Diversification (AAAI 2025 · similarity 0.85)
- Accelerated Diffusion via High-Low Frequency Decomposition for Pan-Sharpening (AAAI 2025 · similarity 0.84)
- Score-based Self-supervised MRI Denoising (ICLR 2025 · similarity 0.83)
- Generating Content for HDR Deghosting from Frequency View (CVPR 2024 · similarity 0.83)
- LeFusion: Controllable Pathology Synthesis via Lesion-Focused Diffusion Models (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.