acceptodds
Under review as a conference paper at ICLR 2027

DuoPrior-CFFlow: Dual-Prior Coarse-to-Fine Flow Matching for 3D OCT-to-OCTA Translation

Abstract

Optical coherence tomography angiography (OCTA) provides clinically valuable retinal vascular information but remains more difficult to acquire than structural optical coherence tomography (OCT), motivating reliable OCT-to-OCTA transla- tion. Three-dimensional OCT-to-OCTA translation is challenging because it must simultaneously preserve continuous global retinal structure and recover sparse, fine vascular details. We propose DuoPrior-CFFlow, a dual-prior coarse-to-fine flow-matching framework for paired three-dimensional OCT-to-OCTA transla- tion. Its CF-Flow path decomposes volumetric representations into low- and high- frequency components, modeling global structural transport with Coarse Flow and vascular-detail transport with Fine Flow. The multi-scale Z-aware (MSZ) prior provides global depth, local depth, and depth-gradient cues for structural guid- ance, while VesselAux promotes three-dimensional vascular consistency through auxiliary vessel supervision. DuoPrior-CFFlow outperforms the directly compa- rable baselines on most quantitative metrics and better preserves retinal vascular structures. Structural preservation and downstream utility evaluations show that our translated 3D OCTA volumes achieve the highest 3D vessel-mask consistency, the lowest vascular occupancy in the foveal avascular zone, and the highest disease classification accuracy. Code is available at https://anonymous.4open.science/r/duoprior-cfflow-code-6D9F/.

Primary area:
generative models
Keywords:
flow matching, 3D image-to-image translation, volumetric medical image synthesis, retinal OCT, OCT angiography, vessel-aware generation

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_Ci36OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on the OCT-to-OCTA translation task and proposes a dual-prior coarse-to-fine flow-matching framework. The key lies in the fact that the CF-Flow path decomposes volumetric representations into low- and high-frequency components, modeling global structural transport with Coarse Flow and vascular-detail transport with Fine Flow. The results show the improvement of the proposed method.

Critical strengths
  1. The idea of Coarse Flow first modeling low-frequency structural transport, while Fine Flow uses the predicted coarse coefficients as context to recover high-frequency vascular detail, is reasonable.
  2. Comprehensive quantitative and qualitative comparisons with existing methods.
Critical weaknesses
  1. The contribution of the proposed method over existing works is incremental, as flow matching and wavelet transformation are existing works. Coarse and fine flow matching also appear in existing work. Also, only a vessel head connects the proposed method to the OCT domain. This work seems to be a lot of a combination of existing works.

  2. The improvement over FlowOCT is not significant. The need for coarse and fine branches remains questionable.

  3. The reliance on a pre-trained vessel labeler limits its generalizability to new domains.

Questions for authors
  1. The contribution of the proposed method over existing work is incremental, as flow matching and wavelet transformation are existing methods. Coarse and fine flow matching also appear in existing work. Also, only a vessel head connects the proposed method to the OCT domain.

  2. The improvement over FlowOCT is not significant. The need for coarse and fine branches remains questionable.

  3. The reliance on a pre-trained vessel labeler limits its generalizability to new domains.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 3

Q1: The method appears to combine existing techniques, including flow matching, wavelet transformation, and coarse-to-fine flow matching. The OCT-specific contribution also appears limited to a vessel head.

R1: We clarify that our contribution is not the general invention of flow matching, wavelet transforms, or coarse-to-fine modeling. It is their task-specific design for paired 3D OCT-to-OCTA translation. CF-Flow assigns distinct roles to low-frequency structural transport and high-frequency vascular-detail refinement, with the predicted coarse representation conditioning Fine Flow. In addition, the method incorporates two OCTA-oriented priors: MSZ Prior derives global-depth, local-depth, and depth-gradient cues from the source OCT to guide structural emphasis, while VesselAux supervises the high-frequency branch using 3D vessel targets. Thus, the OCT-specific guidance is provided by both the source-conditioned MSZ Prior and the vessel-level auxiliary supervision, rather than by the Vessel Head alone.

Q2: The improvement over FlowOCT is not significant, and the need for coarse and fine branches remains questionable.

R2: We agree that the results should not be interpreted as showing that the dual-path design is universally necessary or statistically superior. In Table 5, the full DuoPrior-CFFlow configuration improves over the Single-Path Flow (FlowOCT) result from 1.22 to 1.13 in MAE, from 28.80 to 29.38 dB in PSNR, from 86.82 to 87.65 in SSIM, from 5.36 to 5.12 in LPIPS, and from 18.76 to 17.48 in FID. The ablation also compares Coarse Flow with CF-Flow to examine the effect of adding Fine Flow. These results support the selected design under our evaluation protocol, but we do not claim statistical significance or that every branch is necessary in all settings.

Q3: Reliance on a pretrained vessel labeler may limit generalizability to new domains.

R3: We clarify that SAM-OCTA is used to construct vessel supervision and to compute the reported segmentation-derived vessel metrics; it is not required by the generator at inference. Therefore, generation does not require running the pretrained labeler on a new-domain input. However, VesselAux training and the reported vessel evaluations do depend on suitable vessel masks, and we have not evaluated cross-domain transfer. We therefore do not claim that the current vessel-supervision protocol establishes generalizability to new domains.

Reviewer_XqdhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes deterministic OCT-to-OCTA translation through separate low- and high-frequency wavelet flows, with the fine branch conditioned on the coarse prediction. An OCT-derived depth prior and auxiliary vessel supervision guide training. At inference, three Euler steps transform OCT wavelet coefficients into predicted OCTA coefficients, followed by an inverse wavelet transform to reconstruct the OCTA volume.

Critical strengths
  1. The frequency decomposition matches the anatomy and gives vessel supervision a natural attachment point. The low-frequency LLL coefficients carry the continuous retinal structure, while the high-frequency subbands carry the sparse, fine vascular details. Attaching the VesselAux head to the high-frequency detail volume follows this structure, and it produces the clearest independent gain in the ablation—on a distributional metric that does not depend on the SAM-OCTA labeler.

  2. The method is efficient without sacrificing quality. It matches or exceeds the baselines on the reconstruction metrics and LPIPS while using a substantially smaller generator and faster inference. For 3D OCT-to-OCTA translation, this is a practical advantage: the gains are not bought with model capacity.

Critical weaknesses

Independent validation of vascular fidelity is missing, and the reported vessel metrics are confounded by the training objective. VesselAux trains a vessel head on the high-frequency-only volume against SAM-OCTA masks of real OCTA, with gradients passing through the head and inverse DWT into Fine Flow. The generator is therefore optimized so that a vessel head reading its high-frequency output reproduces the labeler's masks; at test time, the same labeler produces the vessel maps used for Dice/IoU and FAZ occupancy. The FAZ region is the official annotation, but the vessels inside it are labeler-derived; because the training mask is near-empty there, FAZ emptiness is encouraged, not merely measured. Clinician review of the targets is a sanity check, not a quantitative independent validation, and the remaining metrics are global volume/slice scores that do not localize errors to vessels. The paper's central claim therefore has no independent vessel-specific measurement.

Questions for authors
  1. What are the Dice/IoU and FAZ numbers for CF-Flow without VesselAux? If they fall to baseline levels, the vessel metrics are measuring the auxiliary loss rather than vascular fidelity.

  2. On a human-annotated subset (20–30 volumes), does the ranking between your method and FlowOCT survive?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 2

Q1: Independent validation of vascular fidelity is missing, and the vessel metrics may be confounded by using SAM-OCTA-derived masks for both VesselAux supervision and evaluation.

R1: We clarify that the Dice/IoU and FAZ metrics measure agreement with 3D vessel masks generated by the fixed SAM-OCTA model, which is applied consistently to translated and real OCTA volumes. They are not comparisons against manually annotated 3D vessel ground truth. Although VesselAux uses SAM-OCTA-derived targets, the reported metrics are computed by applying the segmenter to the generated volumes and comparing the resulting masks; they are not the VesselAux loss itself. The model is also trained with flow, reconstruction, and structural objectives. We agree, however, that using the same label source for supervision and evaluation does not constitute independent validation and may favor patterns recognized by that labeler. Clinician review of the masks serves as a qualitative quality check, not as quantitative independent validation.

Q2: What are the Dice/IoU and FAZ results for CF-Flow without VesselAux?

R2: These vessel and FAZ metrics were not computed for the CF-Flow-only ablation. The existing ablation results therefore do not establish how removing VesselAux affects these measurements, and we cannot provide numerical results or infer that comparison from the reported reconstruction metrics.

Q3: Does the ranking between DuoPrior-CFFlow and FlowOCT persist on an independently human-annotated subset?

R3: We did not evaluate the two methods on an independently human-annotated subset. The clinician-reviewed masks were generated from SAM-OCTA predictions and therefore do not provide an independent annotation set for this comparison. Accordingly, the current results do not establish whether the ranking persists against manual vessel annotations.

Thank you for your frank reply. I will keep my score.

Reviewer_CXj6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DuoPrior-CFFlow

Critical strengths

The paper presents a well-motivated coarse-to-fine flow matching framework for 3D OCT-to-OCTA translation, supported by comprehensive experiments demonstrating improvements in reconstruction quality, vascular preservation, and downstream utility.

Critical weaknesses

The experimental comparisons are relatively limited, with several baselines being outdated, such as Pix2Pix. More recent open-source image translation methods should be included to provide a more comprehensive evaluation of the proposed framework. Figure 2 appears to contain AI-generated visual elements, which detract from its clarity and academic presentation. The authors are strongly encouraged to redesign the framework diagram using conventional scientific illustration tools. The authors should provide a clearer justification for adopting flow matching instead of diffusion models, which are more widely used in image generation. Additional comparisons with representative flow-matching and diffusion-based methods would help establish the advantages of the proposed approach. "FlowOCT: Wavelet Flow Matching for OCT-to-OCTA Translation in IJCAI" is similar to your work

Questions for authors

N/A

1 comment
AuthorsOpenReview ↗

Response to Reviewer 1

Q1:The experimental comparisons are relatively limited, with several baselines being outdated, such as Pix2Pix. More recent open-source image translation methods should be included to provide a more comprehensive evaluation of the proposed framework.

R1:Our evaluation covers recent methods developed specifically for OCT-to-OCTA translation from 2023 to 2026, including the latest FlowOCT (2026), together with representative 2D and 3D baselines. We also examined the publicly available XOCT code. However, the required 3D retinal-layer segmentation labels were not available in a usable form in the released materials we could access, preventing us from reproducing XOCT under the same data partitions and preprocessing protocol. We therefore compare against recent methods that can be fairly reproduced under our unified evaluation protocol.

Q2: The authors should provide a clearer justification for adopting flow matching instead of diffusion models, which are more widely used in image generation.

R2:We chose flow matching because our task is paired OCT-to-OCTA translation: each structural OCT volume has a spatially corresponding OCTA target. Flow matching provides a continuous transport formulation that can model this paired mapping directly, while allowing us to define a deterministic path between the paired representations. This is a natural fit for learning anatomy-preserving translation. Diffusion models are also applicable, and we include BBDM as a diffusion-based baseline; our choice of flow matching is motivated by its suitability for paired volumetric translation, not by a claim that it is universally superior to diffusion.

Q3:Additional comparisons with representative flow-matching and diffusion-based methods would help establish the advantages of the proposed approach.

R3:We have already included representative flow-matching and diffusion-based comparisons. Specifically, FlowOCT serves as the flow-matching baseline, while BBDM 2D and BBDM 3D represent diffusion-based baselines.

Q4:"FlowOCT: Wavelet Flow Matching for OCT-to-OCTA Translation in IJCAI" is similar to your work.

R4:We acknowledge the close relationship between FlowOCT and our work. We discuss its 3D wavelet flow-matching approach and the limitations of existing structural constraints in Related Work (Section 2.1). The Introduction identifies the separation of low-frequency retinal-structure transport from high-frequency vessel-detail transport, together with richer depth-aware guidance and volumetric vessel supervision, as the gaps addressed by our method. Beyond the dual-path CF-Flow design, DuoPrior-CFFlow introduces two complementary priors: the MSZ Prior provides multi-scale depth-aware structural guidance, while VesselAux supplies 3D vessel supervision to the high-frequency branch. In Section 4.4, we compare the Single-Path Flow (FlowOCT) baseline with CF-Flow, and Coarse Flow with CF-Flow, to examine the dual-path design and the role of high-frequency transport.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.