acceptodds
Withdrawn from ICLR 2027

HDRSNet: Learning Hemodynamic Representations via Residual Diffusion for Tumor Segmentation from Non-Contrast MRI

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Dynamic contrast-enhanced MRI (DCE-MRI) provides rich hemodynamic information for breast tumor segmentation, but its clinical use is often limited by contrast-agent risks, acquisition cost, and incomplete temporal sampling—sometimes with only a pre-contrast volume available. Existing generate-then-segment methods synthesize missing post-contrast phases before segmentation, but synthesis artifacts and over-smoothed enhancement dynamics can compromise downstream performance. To address this challenge, we propose HDRSNet, a Hemodynamic Diffusion Representation Network that learns task-aligned hemodynamic representations for accurate breast tumor segmentation from a single non-contrast MRI volume. During training, HDRSNet exploits complete DCE-MRI sequences to learn enhancement dynamics as residual variations relative to the pre-contrast baseline. Specifically, a residual diffusion module incorporates recurrent modeling and local decoupling to suppress background-driven noise accumulation and capture tumor-relevant enhancement patterns. The resulting hemodynamic representations are integrated into a segmentation backbone through multi-scale task-aligned conditioning, enabling the model to exploit learned dynamic cues when only a non-contrast volume is available at inference. Experiments on multiple DCE-MRI benchmarks demonstrate that HDRSNet outperforms state-of-the-art 4D models requiring complete temporal sequences and generate-then-segment pipelines. These results highlight task-aligned hemodynamic representation learning as an effective alternative to explicit image synthesis for tumor segmentation under missing contrast-enhanced observations. The source code is available at: https://anonymous.4open.science/r/HDRSNet-6A42.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Label-Efficient Segmentation, Active Learning, Foundation Models, Dynamic Medical Imaging

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_69QLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses breast tumor segmentation when only the pre-contrast volume of a DCE-MRI exam is available at test time. The full sequence is used only during training. The proposed HDRSNet has three parts. A recurrent conditional diffusion model (RHD) generates each post-contrast phase from the previous one; it is trained with a scheduled blend of real and generated conditioning (dynamic blending) and with a loss weighted by a soft gate computed from the enhancement magnitude and the tumor mask (local decoupling). Intermediate features of the denoising network, rather than the generated images, are injected into a 3D U-Net that takes the pre-contrast volume, through cross-attention at every scale (MTC). On DUKE (291 cases) and I-SPY2 (970 cases), the method is reported to reach 84.9 and 78.5 DSC from the pre-contrast volume alone, above the best full-sequence baseline (TSESNet, 83.7 and 76.5) and above three generate-then-segment pipelines.

Critical strengths
  1. The setting is clinically meaningful. Using the contrast-enhanced phases only as training-time supervision, and segmenting from a non-contrast scan at test time, would remove the need for contrast injection if it worked reliably.
  2. Taking intermediate features of a phase-synthesis diffusion model instead of its synthesized images is a reasonable idea, and it avoids passing image-level artifacts to the segmenter.
  3. The experimental section is broad: two main datasets, nine baselines, ablations of the blending schedule, loss weights, fusion levels and sampling steps, cross-center and CT experiments, and a limitations section.
Critical weaknesses
  1. The central result is hard to accept as presented. At inference the model sees only V0, so every hemodynamic feature is a function of V0 and random noise, yet HDRSNet is reported to beat all models that use the real post-contrast phases and to more than double the DSC of nnUNet and MedSegDiff given the same V0 input (78.5 versus 33.0 and 25.8 on I-SPY2, Table 12). Even the ablation whose generated phases reach an SSIM of only 0.30 scores 54.1 DSC (Table 3). The controls that could explain this are missing: the plain 3D U-Net on V0 trained with the same recipe, and HDRSNet with real phases at inference as an upper bound. A label leak also cannot be ruled out. The gate in Eq. (4) is built with the ground-truth mask Y and, by Eq. (7), multiplies the features passed to the segmentation network. The text says the gate at inference is computed "in the same form" from the predicted residual, Algorithm 1 applies no gate at inference, and Appendix B.2 gives two further definitions (Eq. 16 without Y, Eq. 19 with a different use of Y). The values of κ, τ and λ are not given.

  2. Table 5 reports p-values between 1e-7 and 1e-15 "from five runs", but the means and standard deviations in the same table do not support them. A two-sample t-test on those numbers gives p = 0.03 for DSC and 0.13 for HD against TSESNet on I-SPY2, and 0.27, 0.72 and 0.12 for the three metrics on DUKE. For I-SPY2, all nine reported p-values have the same leading digits as the recomputed ones, but exponents that are 2 to 10 orders of magnitude smaller. Standard deviations for the same configuration also differ between tables (on DUKE, 84.93 ± 1.67 and 6.87 ± 2.17 in Table 1 versus 84.93 ± 1.2 and 6.87 ± 1.5 in Table 11). With the reported variances, the advantage over TSESNet on DUKE is within run-to-run noise.

  3. In Table 2 the synthetic sequences appear to be fed to the TSESNet that was trained on real sequences ("the same TSESNet"), so these pipelines are tested under a domain shift that HDRSNet does not face. The direct test of the first contribution is missing, namely feeding HDRSNet's own generated phases to a segmenter and comparing this with using its features. On cost, the 216 G FLOPs in Table 1 cannot plausibly include the 125 denoising passes needed per window (five phases, 25 DDIM steps), and 2.2 s per case is hard to reconcile with sliding-window inference.

Questions for authors
  1. What exactly is the gate at inference? Is the mask Y used in any form at test time, and what are the values of κ, τ and λ? What is the result when the feature gate of Eq. (7) is removed during both training and inference?
  2. Could the authors report the segmentation backbone alone on V0 with the same training recipe, and HDRSNet with features extracted from the real phases at inference?
  3. How were the p-values and confidence intervals in Table 5 computed (which test, on which samples, with what n)? Why do the standard deviations of the same configuration differ between Table 1 and Table 11?
  4. In Table 2, was TSESNet retrained on the synthetic sequences? What is the result when HDRSNet's own generated phases are fed to TSESNet? How are FLOPs and time per case counted (number of windows, phases and DDIM steps)?
Reviewer_zE9DOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This a tumor segmentation paper (breast cancer) from DCE MRI sequences. Training uses complete DCE-MRI sequences. Test use pre contrast MRI. Interesting work. First of all, very limited settings we should accept as mammogram is the modality of breast cancer screening. When there is MRI, even DCE-MRI is questionable, but useful in certain settings. The problem is interesting and important. The method is called HDRSNet, and it has three parts. Experiments use I-SPY2 (970 cases) and DUKE (291 cases), five seeds, and 7:1:2 splits. Some liver cases are presented too (CT).

Critical strengths

-The clinical goal is important, and learning task-aligned representations instead of synthesizing pixels is a reasonable direction.

Critical weaknesses
  • The method as written routes the ground-truth tumor mask into the features consumed by the segmentation network (Eq. 4 and Eq. 7). So, not fully automated and its translation requires ground truth masks to be propagated.
  • The gate is applied at inference "consistently with training." That makes test-time leakage or a severe train/test mismatch unavoidable on the paper's own description.
  • The major result is a Vo-only model beating full-sequence 4D models and a Vo-only nnU-Net by 46 Dice points, which is the pattern such leakage would produce. -Independently, 17 of 18 reported p-values fall below the minimum achievable with five runs at the reported standard deviations. The DUKE confidence intervals follow no consistent formula.
  • Two tables report different values for what appears to be the same configuration, and those values share their decimal digits.
  • most im[protantlly, technical novelty is low. Really, I want to be positive but there is no technical innovation , things are combined engineering ways with limited exploration on ablations and potentially, highly likely a suboptimal design.
Questions for authors

-Why would features imagined from V₀ yield better segmentation than real post-contrast observations (TSESNet with V₀ to V₅)?

  • How does HDRSNet differ, conceptually and empirically, from modality hallucination (Hoffman et al., 2016) and from distilling a full-DCE teacher into a V₀ student?

  • Which publications support the claim that intermediate denoising features serve dense prediction? DiffBoost and DCE-Diff synthesize images.

-At inference, is Y used anywhere in computing Mᵢ, or in any other code path? Please provide the exact inference code for the gate.

-What happens to Dice if another patient's mask, or an empty mask, replaces Y in all code paths at inference?

-Which gate definition (Eq. 4, 16, or 19) produced the reported results, in training and at inference?

  • During training, features come from the denoiser processing a noised copy of the real Vᵢ at one random step. At inference, which DDIM steps' features are used, and how are they aggregated?

-How large is the performance gap between training-style features (from real phases) and inference-style features (from generated phases) on the test set?

Is RHD pretrained for 200 epochs (Section 4.1) or trained jointly from the start (Algorithm 1)?

  • Are the voxel masks from MAMA-MIA? Are splits at the patient level? Are phases co-registered before residuals are computed?

  • Are intensities normalized per phase or with V₀'s statistics?

What Dice does the same 3D U-Net achieve on V₀ alone with the same training budget?

Was TSESNet in Table 2 retrained on synthesized sequences?

How was SPHERE, a post-to-pre synthesizer, adapted to pre-to-post synthesis?

What are the total FLOPs and runtime per case for every method, including all denoiser passes?

Reviewer_Pq4ROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper asks whether a model can learn from contrast-enhanced MRI sequences during training and then segment tumors without contrast-enhanced scans at inference. It proposes HDRSNet for breast tumor segmentation from a single non-contrast MRI volume. A recurrent diffusion model learns enhancement changes relative to the pre-contrast scan, while spatial weighting emphasizes tumor-related changes during training. Multi-scale features from the diffusion model then guide the segmentation network. Experiments on two breast MRI datasets show that HDRSNet improves tumor segmentation using only pre-contrast MRI at inference, outperforming the compared generate-then-segment pipelines and full-sequence models.

Critical strengths
  1. The paper is clearly structured.
  2. Using a diffusion model to infer contrast-related features from non-contrast MRI is a novel aspect of HDRSNet.
Critical weaknesses

The model learns enhancement dynamics from fixed acquisition protocols, but it remains unclear how well these patterns transfer to new protocols or datasets with different phase counts and sampling intervals. This may limit how well it generalizes in practice.

Questions for authors
  1. How sensitive are the segmentation predictions to randomness in diffusion sampling? For the same trained model, do different initial noise samples produce consistent masks and segmentation performance?
  2. Could the authors clarify the training and inference compute budget, including GPU type, GPU hours, peak memory, and inference time per volume compared with the baselines? Do the reported FLOPs include all diffusion sampling steps across all generated phases?
  3. Which aspects of segmentation benefit most from the diffusion features: boundary delineation, tumor localization, or recovery of missed lesions? Could the authors clarify this using the existing predictions or feature analyses?
  4. How does the method compare with the same segmentation backbone without diffusion, with a deterministic feature predictor, or with simpler feature fusion? Can the benefits of loss weighting and feature gating in local decoupling be separated?
  5. The code link appears inaccessible. Could the authors provide a working link so that the implementation details and reproducibility can be checked?

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

Related papers

Discussion (0)

Sign in to comment.