Efficient 3D Whole-Body PET Image Denoising via Conditional Rectified Flow With Optimized Sampling Strategy
Jiale Shen, Guolin Wang, Chenhao Wang, Xinhui Su, Wei Luo, Feng Yu
cs.CV
Quick verdict
A one-pass conditional 3D rectified flow with a tuned exponential sampling schedule denoises ultra-low-dose whole-body PET in ~30s with PSNR/SSIM matching a 1000-step DDPM, though claims rest on surrogate metrics rather than reader or quantitative validation.
Summary
The paper trains a one-pass conditional 3D rectified flow (207M-parameter 3D U-Net velocity field) for whole-body PET denoising, learning a linear-interpolant velocity objective conditioned on low-dose input, and pairs it with an empirically tuned exponential non-uniform time schedule. Training used the public UDPET cohort (300 train/16 validation/50 internal test) at a single 1/20 dose-reduction factor; evaluation covered all six UDPET dose levels (1/2 to 1/100) and a 50-patient external FAHZU cohort (30-120 s/bed, 519 physician-annotated lesions) without fine-tuning. Metrics were PSNR, SSIM, and lesion-conspicuity CNR, with Wilcoxon signed-rank tests and Benjamini-Hochberg FDR correction. With two Euler integration steps under the optimized schedule, inference takes ~31 s per volume versus 4.27 h for a 1000-step 3D DDPM, while PSNR/SSIM match or slightly exceed it (40.70 dB/0.9790 vs 40.06 dB/0.9741 on FAHZU 30 s), with the largest gains at ultra-low doses; the schedule ablation shows modest but significant gains over uniform sampling. Key limitations: all conclusions rest on surrogate metrics without reader studies, SUV bias analysis, or formal lesion-detection analysis; the sampling 'strategy' is effectively a two-hyperparameter grid search with post-hoc interpretation; a few reported numbers (peak step count, baseline timing accounting) are unexplained.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A one-pass conditional 3D rectified flow with linear-interpolant velocity matching enables whole-body PET denoising in approximately 30 seconds per volume using only two integration steps, versus multi-hour inference for a 1000-step 3D DDPM.
- The empirically selected non-uniform exponential schedule (N=2, k=-15), which concentrates steps near the trajectory endpoint where measured errors surge, yields statistically significant PSNR/SSIM gains of +0.29 to +0.62 dB over uniform scheduling across dose levels.
- The proposed model outperforms evaluated 3D DDPM and DDIM baselines on PSNR/SSIM and lesion-conspicuity CNR, with the advantage most pronounced at ultra-low doses (1/100, 1/50).
- A model trained on a single dose level (UDPET DRF=20) transfers zero-shot to unseen dose levels and to an independent clinical cohort (FAHZU) without fine-tuning.
Strengths
Orders-of-magnitude inference speedup at no quality cost on reported metrics
Table I: 2-step 3D Flow reconstructs a full volume in 30.74 s versus 4.27 h for 1000-step 3D DDPM, while PSNR/SSIM are slightly better (40.70 dB/0.9790 vs 40.06 dB/0.9741 on FAHZU 30 s); Section IV-B argues this enables online denoising.
Genuinely strict zero-shot evaluation design
Section III-A: training used only UDPET DRF=20; the same model was tested across DRF 2-100 (Fig. 4) and on an independent center (different scanner, real short-acquisition protocols) with no fine-tuning, with per-dose significance markers in Figs. 4-5.
Ablation links schedule choice to observed trajectory behavior
Evaluation goes beyond global similarity metrics
Weaknesses
All headline evidence is surrogate-metric based, favoring smoothness
The 'optimized sampling strategy' is a grid search with post-hoc rationale
Internal inconsistencies and opaque baseline accounting
Incompletely characterized baseline comparisons
Uncertainty and CNR variability are shown but not quantified
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is rectified flow and how is it used for PET image denoising in this paper?
Rectified flow is a generative modeling technique that learns a velocity field transporting a simple noise distribution straight to the target data distribution along linear paths, so it needs very few numerical integration steps at inference. Here the authors condition that velocity field on a low-dose whole-body PET volume using a 3D U-Net, so the model converts an ultra-low-dose scan into a normal-dose-quality image with only two Euler steps.
How fast is the proposed 3D Flow PET denoising method?
The paper reports about 30 seconds (30.74 s) to denoise one full 3D whole-body PET volume on a single RTX 4090 GPU, versus about 4.27 hours for a 1000-step 3D diffusion model and about 51 minutes for a 200-step DDIM, using the same patch-based inference protocol and similar image-quality metrics.
How low a PET dose can the method handle, and was it retrained for each dose?
No retraining was used. A single model trained only on 1/20-dose data was applied unchanged to dose levels from 1/2 down to 1/100 of the standard dose, and also to an external hospital dataset acquired for 30-120 seconds per bed position, with the largest reported quality advantages at the lowest doses (1/100 and 1/50).
Has the method been clinically validated?
Not yet. Evidence is limited to image-similarity metrics (PSNR, SSIM), a lesion-conspicuity measure (CNR) on 519 annotated lesions, and visual comparisons; the authors themselves state that reader studies, SUV quantification bias analysis, and formal lesion-detection evaluation are needed before any diagnostic use.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The extracted text may be truncated and figures (bar plots, heatmaps, uncertainty maps, qualitative comparisons) cannot be analyzed from text, so figure-only results such as the N/k ablation grid and significance markers could not be verified.
- Reference [26] (Sun & Mawlawi, PET denoising with rectified flow) is cited without venue or year detail; the incremental novelty over it could not be assessed from the provided text.
- Baseline implementations ('reproduced' DDIM/DDPM/DPM-Solver++) cannot be audited from the text; the apparent inference-time accounting discrepancy may reflect undocumented protocol details.
- Reviewer judgment is based solely on the provided extraction; page images and supplementary material were not available.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.