acceptodds
Withdrawn from ICLR 2027

SAFE-Diff: Where Anatomy Meets Frequency in Virtual Contrast Enhancement

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

SAFE-Diff: Segmentation-Conditioned and Frequency-Aware Diffusion for Anatomy-Preserving Contrast Synthesis Download PDF Yuankai Wang, Jianyang Xie, Feixiang Zhou, Zhongli Wu, Tim Fairbairn, Gregory Yoke Hong Lip, John K Field, Yalin Zheng 20 Jul 2026 (modified: 20 Aug 2026) AAAI 2027 Conference Submission Conference, Area Chairs, Senior Program Committee, Program Committee, Authors Revisions CC BY 4.0 TL;DR: SAFE-Diff is a segmentation-conditioned, frequency-aware diffusion framework for synthesizing contrast-enhanced medical images from non-contrast scans while preserving fine anatomical structures. Abstract: Angiography is a contrast-enhanced imaging technique that improves the visualization of vascular structures, thereby facilitating the detection and assessment of vascular dis- eases. but many angiographic examinations rely on con- trast agents that may pose risks to susceptible patients. Vir- tual angiography synthesis from non-contrast scans may re- duce this reliance; however, existing medical image trans- lation methods can still have difficulty preserving fine vas- cular structures, occasionally producing blurred, displaced, or spurious details despite visually plausible contrast pat- terns.To address these challenges, we propose SAFE-Diff, a segmentation-conditioned and frequency-aware diffusion framework that combines global cross-modal alignment with local anatomical guidance. First, modality-specific VAE en- coders and a contrastive projection objective align paired source and target representations at the case level, provid- ing a global cross-modal condition for generation. Second, source-derived segmentation masks and outline maps are in- corporated into each denoising step to encourage consis- tency with the source anatomy. A multi-band spectral at- tention module further modulates bottleneck features using frequency-domain statistics to improve the representation of fine vessels and anatomical boundaries. We evaluate SAFE- Diff on 1,600 paired non-contrast CT and CTA volumes and on brain T1-to-T1ce synthesis, comparing it with represen- tative paired and unpaired translation methods. SAFE-Diff achieves 25.1 dB PSNR, 0.77 SSIM, and 0.26 LPIPS on CT- to-CTA synthesis, and 29.7 dB PSNR, 0.87 SSIM, and 0.19 LPIPS on T1-to-T1ce synthesis, obtaining the best or tied- best performance across the reported metrics. Component ab- lations, downstream structural evaluation, and robustness ex- periments further support the effectiveness of cross-modal alignment, structural conditioning, and spectral modulation. These findings provide evidence that SAFE-Diff improves global fidelity and proxy structural consistency on the evalu- ated paired datasets, although external and clinical validation remains necessary.

Primary area:
generative models
Keywords:
image-to-image translation, misaligned supervision, diffusion models, medical image synthesis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_mMN2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SAFE-Diff, a segmentation-conditioned and frequency-aware diffusion framework for paired virtual contrast-enhanced medical image synthesis. It follows a two-stage process. In Stage 1, modality-specific VAE encoders are pretrained and a contrastive projection head is used to align source and target representations at the case level, reducing the cross-modal semantic gap. In Stage 2, a latent diffusion model synthesizes the contrast-enhanced target while being conditioned on source semantic features, segmentation masks, and anatomical boundary maps. The key module, Multi-Band Spectral Attention (MBSA), adaptively reweights bottleneck features according to their frequency responses, emphasizing anatomically meaningful high-frequency structures while suppressing noise-like fluctuations. Extensive experiments on a large-scale cardiac CT dataset and the BraTS dataset demonstrate that SAFE-Diff outperforms representative UNet-, GAN-, and diffusion-based baselines in both image fidelity metrics and downstream anatomical segmentation tasks.

Critical strengths

The overall method is coherent and aligned with the goal of anatomy-preserving virtual contrast enhancement. The experiments are comprehensive, including an evaluation of the model's robustness to imperfect anatomical priors.

Critical weaknesses
  • The compared baselines are relatively limited for supporting a state-of-the-art claim. The methods listed in Table 1 are general-purpose and relatively older image-to-image translation methods. It is advisable to include comparisons with recent works focused on anatomy preservation or contrast synthesis, such as APT and VASTSD.
  • The downstream segmentation-based anatomical evaluation and the structural priors used by SAFE-Diff are not entirely independent, as the processes for prior generation and evaluation may rely on the same segmentation software framework.
  • Minor issues. There are some formatting and layout issues in the paper. For instance, the subplot label "(a)" appears repeatedly in Fig. 2; additionally, the same paper by Konz et al. (2024) is listed twice in the references as 2024a and 2024b. While these issues do not affect the technical content, they detract from the paper's overall professionalism and polish.
Questions for authors

Table 2 states that anatomical evaluation is conducted on 55 expert-labelled CTA cases, whereas Supplementary Section S2.4 refers to 20 expert-annotated CTA subjects. Do these numbers correspond to different subsets or annotation protocols? If so, could the authors specify the number of cases used for each metric in Table 2?

Reviewer_yYEXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SAFE-Diff studies paired virtual contrast enhancement for cardiac CT-to-CTA and brain T1-to-T1ce synthesis. It combines case-level contrastive alignment, source-predicted segmentation and outline conditions, and a multi-band spectral attention module in a 3D latent diffusion model. Experiments on an in-house cardiac cohort and BraTS show improved image similarity and downstream segmentation scores. The main claimed contribution is anatomy-conditioned frequency modulation for preserving patient-specific structures.

Critical strengths

(1) The problem is relevant, and the paired cardiac cohort of 1,600 patients is a useful scale for studying this task. The source-only inference setting is explicitly described in the supplement, including how the anatomical priors are obtained.

(2) The evaluation goes beyond global image similarity. Table 2 includes coronary Dice, clDice and surface distance against expert annotations, with real CTA as a reference. The improvement in coronary mDice from 15.3 for RegGAN to 33.2 for SAFE-Diff is encouraging. The component ablations and prior-corruption experiment also provide useful evidence, although their interpretation needs more care.

Critical weaknesses

(1) In Eq. (7), the spectral descriptor consists of three globally pooled band energies per sample. Eq. (8) combines this descriptor with a globally pooled outline representation to produce channel-wise gates shared across spatial locations. This can still improve local anatomy through the learned feature channels. The unresolved question is what outline-conditioned gating adds beyond general channel attention. Table 4 removes MBSA as a whole, while Table 3(a) changes the structural priors; neither clearly isolates the outline term inside MBSA with the other spatial conditions kept fixed. Fig. 1 also includes paths labelled “Channel” and “Spatial”, but their correspondence to Eqs. (7)–(8) needs clarification.

(2) Fan et al. already combine outline conditioning and multi-frequency attention for pre-to-post contrast synthesis. FcaNet establishes DCT-based channel recalibration, and CoLa-Diff contains both anatomical guidance and wavelet-domain processing. SAFE-Diff may have a narrower contribution through learnable 3D radial bands and direct outline-conditioned gating, but this distinction is not isolated experimentally. VasTSD is also a close 3D angiography-synthesis method and is cited but not compared. The results currently support a promising integration of existing components more clearly than a distinct anatomy–frequency mechanism.

(3) Table 2 and its accompanying text report 55 expert-labelled CTA cases, whereas Sec. S2.4 states that the structural evaluation uses 20 held-out subjects. The manuscript does not make the relationship between these subsets clear. This matters because Table 2 carries much of the evidence for anatomical preservation. The actual sample size, selection procedure and correspondence between the reported evaluations need to be unambiguous.

(4) The downstream results are meaningful, but they measure segmentation performance on synthesized images. Sec. S2.4 appropriately describes them as a proxy evaluation. Coronary mDice/clDice are 33.2/21.5, compared with 64.2/55.0 on real CTA, and ASSD is 3.8 versus 1.2 mm. These results leave a substantial gap. There is no direct analysis of omitted or false branches, caliber errors, or preservation of abnormal enhancement. The BraTS experiment also lacks a lesion-focused evaluation.The fine-anatomical-fidelity claims consequently go beyond what these measurements directly verify.

(5)SAFE-Diff uses pretrained anatomical predictors and, for BraTS, a prior network trained with tumor annotations. The paired baselines are not reported to receive equivalent masks or outlines. This is a legitimate source-only pipeline, but it leaves the architectural benefit mixed with the benefit of additional anatomical supervision. The source conditioning retained in the “Baseline” and “w/o CAP” variants is also unclear. Identical optimizer settings across methods do not by themselves establish comparable tuning.

(6) The private cardiac cohort makes a precise public reproduction particularly valuable, but “BraTS 2016/2017 official splits” is not enough to identify the MRI experiment. The provided configuration omits important choices such as patch size, training budget, intensity normalization, prior/evaluator checkpoints, and the full sampling and stitching procedure. CT/CTA registration is not described. It is also unclear whether 256×144×144 is obtained by cropping/padding or further resizing after resampling to 1.1 mm. These choices directly affect vascular geometry and the interpretation of distances in millimetres.

(7)Under erosion plus shift, coronary mDice falls from 33.2 to 17.5, a reduction of 15.7 points, or 47.3% relative to the clean-prior score. The discussion instead emphasizes the smaller cardiac and global-image changes when describing the degradation as graceful. Also, the reported mask Dice measures agreement with the original prediction, not accuracy against true anatomy. The perturbation magnitudes and real prior-prediction errors are needed to judge how representative this stress test is of difficult pipeline inputs.

(8) No patient-level uncertainty or number of runs is reported. This makes it difficult to assess differences such as 0.5 dB PSNR and 0.01 LPIPS against RegGAN. Brain SSIM is tied with cGAN at the reported precision. The ablation interpretation also needs correction: the text describes CAP as having a smaller effect, while Table 4 shows that removing CAP causes the largest PSNR and SSIM losses. Different metrics may favor different components, but the current broad interpretation is not supported consistently.

Questions for authors

(1) Which MBSA implementation produced the results: the architecture in Fig. 1 or the operator in Eqs. (7)–(8)? With the backbone and structural inputs fixed, how does it compare with a spectral-only gate and a parameter-matched standard channel gate? A fixed-band control would also help identify the benefit of learning the bands. In “w/o Struct.”, is the outline term inside MBSA removed as well?

(2) What is the precise mechanism-level difference from Fan et al. and FcaNet? Which original CoLa-Diff components are retained or replaced? Can the authors provide a comparison with a close synthesis method such as Fan et al. or VasTSD, or explain a concrete obstacle to making that comparison? These comparisons would make the incremental contribution easier to judge.

(3) Do the anatomical results use 55 patients, 20 patients, or different subsets for different structures? Please specify the patient count and selection criteria for each endpoint, the relationship between the subsets, and whether every method was evaluated on the same fixed cases.

(4) Can the authors provide an independent analysis of the claimed anatomical benefit on an annotated held-out subset, for example branch precision/recall, centerline or caliber errors, or agreement on abnormal vascular regions? For BraTS, enhancing-tumor sensitivity and false enhancement would be informative. Which evidence supports preservation of patient-specific abnormality beyond easier downstream segmentation?

(5) What is the performance of a strong paired diffusion baseline given the same source features, predicted masks, outlines and annotation resources? Please clarify the conditions retained in each ablation, especially whether “w/o CAP” keeps the same global conditioning pathway without contrastive training. How were the paired baselines tuned on validation data?

(6) Can the authors provide exact BraTS subject partitions and a runnable configuration covering preprocessing, prior generation, training and volume inference? Please also explain CT/CTA registration, the fixed-size conversion, and how physical spacing is handled during anatomical evaluation. What artifacts or access procedure can support verification of the private cardiac results?

(7) What erosion radii and translation ranges were used, and were the outlines regenerated from the perturbed masks? How do coronary metrics change with actual prior-prediction errors against annotations? Please compare the difficult-prior cases with the model without structural priors and discuss the coronary degradation explicitly.

(8) Can the main improvements be reported with patient-level paired confidence intervals, together with the number of training runs and sampling seeds? Please also reconcile the component-importance discussion with Table 4 and distinguish strict improvements from ties at the reported precision.

Reviewer_bpJsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces "SAFE-Diff", a segmentation-conditioned and frequency-aware diffusion model for paired data that predicts contrast-enhanced CT and MRI images from non-contrast scans such that susceptible patients can be spared the negative reaction to the contrast agent. The authors tackle limitations of previous image translation approaches, such as preserving local anatomical details and structurally faithful synthesis with regards to each patient. The authors evaluate their method on 1600 paired CT and T1-weighted MRI volumes, yielding best performance across common image quality and fidelity metrics, and best performance in downstream anatomical tasks. In their ablations, the authors evaluate different structural priors and robustness of SAFE-Diff to prior degradation.

Critical strengths
  1. The paper is well-motivated and addresses relevant limitations in the field in their approach: The method section describes this in an easy-to-follow, modular way: Cross-modal correspondence is reached via pooling and InfoNCE, local anatomy is preserved via region and boundary-level anatomic guidance, and distinction between high-frequency noise and anatomy to be preserved is established through the MBSA module.
  2. The visual comparison in Fig. 2 shows that SAFE-Diff retains local anatomy very well compared to the baselines, which supports one of the method's main claims, while producing clear contrast.
Critical weaknesses
  1. Improvement over baselines is marginal in many cases. For instance, in the quantitative comparison (Table 1), SAFE-Diff yields only marginal improvement regarding PSNR, SSIM and LPIPS of CT to CTA synthesis to the second-best baseline RegGAN, which is a comparatively older method, as well as marginal LPIPS gains in T1 to T1ce synthesis compared to RegGAN and ResViT, with SSIM parity to cGAN. In the downstream anatomical evaluation (Table 2), the second-best entries are not underlined, which makes comparison less easily readable. Regarding mDice of Cardiac, Aorta Dice and PA Dice of the Great vessels, and ASSD of the Coronary arteries, gains are marginal again compared to the second-best scores. Overall, the performance can be read as competitive, not marginally outperforming compared baselines. Also, no standard deviations of SAFE-Diff's experiments or a performance of significance tests is reported. Another point I'd like to make here is that more baselines, such as SegGuidedDiff or DCE-Diff could be included, as they are also mentioned in the Related Work chapter. The manuscript could also be strengthened by considering other recent work on T1->T1ce synthesis, such as [1].
  2. The method section is difficult to follow as Fig. 1 sometimes does not correspond to the descriptions in Sec. 3 "Method", making it more difficult to follow the approach: For example, the notation of the pre-trained VAE's is not in the Figure, that an uuUnet predicts the segmentation mask and edge outlines is not clear, FSF^S and FTF^T is represented differently, the FFN is not in the text and the normalization nor in the plot. Also, only "global average pooling" is mentioned in the text, whereas the Figure shows two kinds of pooling operations, which the authors probably summarized as "Pool" in the manuscript. Z~m\mathbf{\tilde{Z}}^m becomes Zm\mathbf{Z}^m in the text. P\mathcal{P} becomes "Projection P" in the plot, etc.

[1] Brandstötter, B., & Kobler, E. (2025). Synthesizing Accurate and Realistic T1-Weighted Contrast-Enhanced MR Images Using Posterior-Mean Rectified Flow. In International Workshop on Simulation and Synthesis in Medical Imaging (pp. 159-169). Cham: Springer Nature Switzerland.

Questions for authors

Questions to the authors whose answers could improve my assessment:

  • Why do the authors list SegGuidedDiff, FGDM, and DCE-Diff in their Related Work section, but do not include them in their Tab. 1? Are there comparative findings available?
  • Can the authors provide standard deviation/error bars, seeded results or significance tests to show the robustness of their method?

Clarification requests:

  • Do the authors plan on making the cardiac cohort dataset available to others such that the work becomes fully reproducible? And why is an in-house dataset used over a publicly-available one which would show a more transparent comparison?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.