acceptodds
Withdrawn from ICLR 2027

Frequency-Disentangled Multi-Phase Breast DCE-MRI Synthesis from Pre-Contrast MRI

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) captures spatially heterogeneous and temporally evolving tumor enhancement. Synthesizing multi-phase DCE-MRI from pre-contrast MRI therefore requires recovering patient-specific enhancement patterns rather than merely generating realistic images. However, existing methods often oversmooth intratumoral heterogeneity and introduce spurious enhancement outside lesions. We propose Freq-Syn, a frequency-disentangled framework that synthesizes multi-phase breast DCE-MRI from a pre-contrast 3D volume and target acquisition times. Unlike whole-image prediction, Freq-Syn represents enhancement as a progressive Gaussian scale-space hierarchy of global contrast, regional uptake, lesion morphology, and fine texture. Cross-phase interaction jointly models their temporal evolution, while synthesis-guided localization constrains where heterogeneous enhancement is generated. Lesion annotations provide targeted supervision during training but are not required at inference. On the internal and external test sets, Freq-Syn achieves SSIM scores of 0.83 and 0.81, respectively. Compared with the best-performing comparison methods, Freq-Syn reduces the false enhancement rate from 40.47% to 16.93% internally and from 37.49% to 19.59% externally. For downstream lesion segmentation, Freq-Syn consistently achieves higher Dice scores than all comparison synthesis methods on both test sets, demonstrating that the synthesized images preserve task-relevant tumor information. These results highlight the effectiveness of explicitly modeling spatial and temporal enhancement patterns for faithful multi-phase DCE-MRI synthesis. These results demonstrate the effectiveness of explicitly modeling spatial and temporal enhancement patterns for faithful multi-phase DCE-MRI synthesis.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Synthesis, Intratumoral heterogeneity

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

This paper tries to address the issues of excessive smoothing of intratumoral heterogeneity and the tendency for false-positive enhancement in normal tissues during multi-phase breast DCE-MRI synthesis. To tackle these challenges, a frequency-decoupling approach for DCE-MRI is proposed; specifically, a progressive 3D network architecture is constructed, and the method is validated through downstream clinical tasks.

Critical strengths

None

Critical weaknesses
  1. This paper is highly engineering-oriented and targets specific clinical problems, which is a strength; however, the specific clinical issues need to be clearly defined and elucidated using corresponding clinical data, and potential ethical risks should be addressed. I also have reservations regarding the reproducibility of the experimental results, as there are no evaluation experiments conducted specifically on public datasets. For a methodology-focused paper like those at ICLR, I would prefer to see a clearer emphasis on the advantages of the proposed method itself.

  2. The proposed method appears to involve a large number of hyperparameters and pipeline components—such as those listed in the appendix—whose selection process is unclear and seems to have been meticulously tuned. I question whether the baseline methods underwent similar careful tuning, which raises concerns about the fairness of the comparison.

Questions for authors
  1. Please clarify the rationale behind the parameter choices and the corresponding ablation studies; experiments on public datasets and detailed preprocessing procedures should also be provided.

  2. Suggestion: The font size in Figure 1(c) is too small; it is recommended to adjust it appropriately

1 comment

We sincerely thank the reviewer for the valuable feedback. We acknowledge that the current manuscript could better explain our methodological motivation and experimental details. Our key contribution lies in frequency-specific modeling of tumor enhancement, where each stage of the hierarchical architecture is designed to capture distinct scale-dependent features. We believe this provides a meaningful approach to preserving heterogeneous enhancement patterns. We appreciate the concerns regarding parameter selection, reproducibility, and experimental fairness, and will further clarify these aspects in the revised manuscript. We are grateful for the constructive suggestions, which will help us better communicate the contributions of our work.

Reviewer_DeypOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper faces the "oversmoothed intratumoral heterogeneity and spurious enhancement outside lesions" in multi-time-point breast DCE-MRI synthesis, and proposes Freq-Syn to generate multi-phase breast DCE-MRI via progressive Gaussian frequency decoding. The method achieves promising results on both internal and external datasets. However, the experimental fairness has notable limitations, and the proposed method adopts an incremental design. Last but not least, several identical numerical values appear in Table 1, which raises confusion.

Critical strengths
  1. The model performance is validated through evaluations on multiple downstream tasks as well as internal and external datasets.

  2. The approach supervises image generation separately using four progressive frequency components. This design is interesting and offers better interpretability compared with conventional generation methods.

  3. The paper investigates a valuable research topic for contrast-agent synthesis: heterogeneous enhancement dynamics.

Critical weaknesses
  1. The four modules in the Progressive Frequency Synthesis Network are designed in an incremental manner with excessive components. Furthermore, the motivation for this design is unclear. Why four layers instead of other numbers of layers? Why do these four layers correspond to global contrast, regional uptake, lesion morphology, and fine texture, respectively? Why does supervising these four components "make heterogeneous enhancement dynamics optimizable"? In Figure 2(b), the difference between B_L and B_M is not obvious. How can one represent regional uptake and the other represent lesion morphology? Additionally, what is the function of Frequency swapping in Figure 2(a)? Are (a) and (b) two distinct methods?

  2. There is insufficient experimentation and analysis to validate the claim that the model can “not only to determine when and how much a lesion should enhance, but also where heterogeneous enhancement should occur”. a. Comparisons of lesion-wide enhancement curves against the ground truth and competing baselines are missing. b. For the time-varying enhancement of intratumoral heterogeneity: does the synthesized image exhibit realistic heterogeneous enhancement? c. A reader study to evaluate the synthesized outputs is absent.

  3. Unconventional metrics in the experiments, such as tumor pattern correlation and FE-Rate, are not accompanied by their formulas to verify the rationality of the evaluation protocol.

  4. Several identical numerical values appear in Table 1, which raises confusion.

  5. Concerns regarding experimental fairness: Freq-Syn leverages real tumor masks during training. Although masks are only used in the training phase, some baseline methods do not support tumor mask inputs. How can a fair input setting be guaranteed?

  6. The figures are heavily blurred; small text in Figure 2 is illegible

Questions for authors
  1. Explanation for repeated identical values in Table 1
  2. Detailed motivation for Progressive Frequency Synthesis Network
  3. Add the essential model analysis to validate the claim.
1 comment

We thank the reviewer for the constructive comments and valuable suggestions. We address the concerns as follows.

  1. Motivation for four frequency levels. We adopt four levels to balance frequency decomposition granularity and model complexity. Too few levels may mix distinct enhancement patterns, while excessive levels may introduce redundancy. The interpretations of global contrast, regional uptake, lesion morphology, and fine texture are supported by the frequency-swapping analysis in Fig. 2, where exchanging corresponding frequency components between different time points reveals their distinct contributions. These represent dominant characteristics rather than strict one-to-one correspondences. Frequency-specific supervision encourages targeted feature extraction and learning at different scales, making heterogeneous enhancement patterns more explicitly optimizable.

  2. Clarification of Fig. 2. Figures 2(a) and 2(b) illustrate the same frequency-swapping analysis rather than two distinct methods. Fig. 2(a) exchanges the zero-frequency Gaussian-band components between two time points, while Fig. 2(b) progressively incorporates frequency-band components from another time point into the pre-contrast image. The distinction between (B_L) and (B_M) lies in their frequency characteristics, which may not be immediately apparent visually. We will clarify these relationships.

  3. Temporal enhancement analysis. We agree that lesion-wide enhancement curves and temporal heterogeneity analysis would provide stronger validation. While our current multi-phase evaluations assess synthesis quality, they do not fully establish the fidelity of temporal enhancement dynamics. We acknowledge this limitation and will clarify the scope of our conclusions.

  4. Evaluation metrics. FE-Rate measures the percentage of non-tumor voxels with ground-truth enhancement ≤ 0.02 where predicted enhancement exceeds the ground truth by more than 0.01, aggregated across target phases per patient and then averaged across patients. Tumor-PCorr computes the voxel-wise Pearson correlation between predicted and ground-truth enhancement within the tumor mask, averaged across phases and patients. These metrics evaluate spurious enhancement and intratumoral spatial heterogeneity, respectively. We will include their formulas.

  5. Identical values in Table 1. The identical reported values result from numerical rounding. Since tumors occupy relatively small regions, localized differences may have limited effects on aggregated metrics, leading to identical values at the reported precision. We will increase the displayed precision to avoid confusion.

  6. Experimental fairness. Ground-truth tumor masks are used only for training supervision, not as inference inputs. Thus, Freq-Syn does not require additional mask information at inference. We acknowledge that mask-guided supervision provides additional training information and may contribute to the performance gains. Accordingly, the comparisons evaluate the complete framework rather than isolating the architectural contribution. We will clarify this distinction.

  7. Figure quality. We will improve the figure resolution, enlarge the text, and enhance the visual clarity of Fig. 2.

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

This paper presents a model to predict post contrast enhancement images from pre-contrast MRI images in breast cancer diagnosis. The main claim of the paper is that decomposing the images in space-scale components and using these for supervising prediction is more effective than predicting the full image directly. The formulation of the method is presented, and experiments were conducted to evaluate performance and compare against relevant baselines.

Critical strengths
  • The formulation of the model is inspired by and follows recent work on frequency learning.
  • The adaptation to the MRI-DCE synthesis domain appears to be a novel application.
Critical weaknesses
  • A more intuitive description of the three levels of modeling and the four target components would greatly enhance the presentation of the paper. In its current state, it is very hard to follow, as it introduces terminology, symbols, and concepts without properly motivating or explaining where they come from.
  • The title and most of the paper indicate a frequency modeling or frequency learning approach. Previous work cited in the manuscript uses the Fourier transform to change representation to the frequency domain. However, this manuscript does not explicitly use the frequency domain; the connection is rather loose.
  • The performance of the proposed model is modestly better than the baselines according to most metrics, except False Enhancement Rate (FE-rate). The proposed model is better on this metric by a large margin. The main issue with this is that none of the baselines was trained with an FE-rate term as the proposed model is. This is biological supervision that other models do not seem to have. Ground-truth masks are used only for training supervision and while not required at inference time, it’s unclear which baselines use this guidance during training too to make the evaluation comparable.

Additional comments:

  • Why is the U-Net a deterministic reference? What does that mean?
  • Is the segmentation model trained on real or synthetic images?
  • No analysis of temporal performance was conducted.
Questions for authors
  • Can the authors frame the method as a space-scale decomposition of the image rather than a frequency decomposition? That is more accurate and less confusing.
  • Can the authors add more intuitive motivations and explanations for the formulation, decompositions, and modeling? This can be brief, and does not need to extend too much, but needs to go beyond the equations and technical variables.
  • Why the False Enhancement Rate metric is not explicitly presented? Were the baselines such as U-Net trained with this supervision?
  • Other questions presented above.
3 comments

We thank the reviewer for the suggestion. While our decomposition operates in the spatial domain, it has a clear frequency interpretation: Gaussian smoothing performs low-pass filtering, while inter-scale differences extract band-pass components. This is consistent with FreeScale (Qiu et al., ICCV 2025), which also employs Gaussian filtering to separate frequency components. Thus, explicit Fourier transformation is not necessary for frequency-aware modeling. We will clarify this connection and adopt the more precise term "spatial-frequency decomposition" in the revision.

We thank the reviewer for this constructive suggestion. Our three-level modeling strategy follows a coarse-to-fine refinement process, while the four target components decompose enhancement information into complementary spatial-frequency scales. This design aims to progressively capture global enhancement patterns, coarse structures, and fine-grained details, rather than predicting the full enhancement image directly. We agree that the motivation and relationships between these components were not sufficiently explained. We will add a more intuitive overview and clarify the terminology and formulation in the revised manuscript.

We acknowledge that the mask-guided supervision may contribute to the improvement in FE-rate. Our goal is to evaluate the effectiveness of the overall framework, including both the proposed modeling strategy and its supervision design, rather than attributing the improvement solely to the architecture. We will clarify the supervision settings of all baselines and provide an ablation study without the FE-related loss to better isolate its contribution. Importantly, the masks are only used during training and are not required at inference.

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

Related papers

Discussion (0)

Sign in to comment.