acceptodds
Under review as a conference paper at ICLR 2027

WaveJoint: Frequency-Guided Scale Transitions for OCT Angiography Segmentation

Abstract

Optical coherence tomography angiography (OCTA) supports two complementary biomarkers: the retinal vessel (RV) tree and the foveal avascular zone (FAZ). The capillaries of the RV tree span only 1–3 pixels and their decorrelation signals are weak, whereas the FAZ is a compact region whose accurate delineation depends on stable contextual information. Jointly segmenting both structures from a single en-face projection is desirable for clinical quantification, but the two targets place conflicting demands on the network. We present WaveJoint, a compact (12.4M-parameter) dual-task Transformer and, to our knowledge, the first OCTA joint-segmentation architecture built around an explicit frequency decomposition. At each encoder scale transition, a wavelet-guided downsampling module (WaveDown) applies fixed orthogonal Haar filters, which are parameter-free and satisfy perfect reconstruction, to compute directional detail through a 2-D discrete wavelet transform and to reweight the downsampled path before aliasing suppresses high-frequency detail. Two task-specific heads then predict the RV and FAZ masks under a balanced Dice–BCE objective. A 2×2 factorial ablation with interaction terms attributes the model's fine-vessel advantage to this downsampling path: this path shows a 2.1 percentage-point main effect on capillary-scale recall and is the only module whose individual removal incurs a significant structural cost. The same analysis explains why single-module ablations judged on aggregate Dice alone can suggest, incorrectly, that no module is important. We also contribute a structural evaluation protocol comprising skeleton recall stratified by vessel caliber, HD95 and ASSD boundary distances, a fragmentation ratio, and Holm-corrected paired Wilcoxon tests. Under this protocol, WaveJoint exceeds a protocol-matched U-Net by 3.25 percentage points of capillary recall with 48% lower fragmentation; under Dice alone, the two models are statistically indistinguishable. On the official OCTA-500 3 mm split, WaveJoint attains an RV Dice of 0.9220±0.0002 and an FAZ Dice of 0.9820±0.0013 over three seeds, matching the best published joint-segmentation systems while using a single 2-D input and no pretraining. The structural advantage persists under shifts in field of view, device, and pathology (OCTA-500 6 mm; FAROS), with capillary-recall gains between 3.3 and 12.5 percentage points. These results indicate that OCTA segmentation should report structural metrics and that the contribution of an architectural mechanism should be localized rather than presumed.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Optical Coherence Tomography Angiography, Medical Image Segmentation, Wavelet Transform, Multi-Task Learning, Vision Transformer

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

This paper presents WaveJoint, a dual-task Transformer for joint retinal-vessel and FAZ segmentation in OCTA. The method combines fixed Haar wavelet analysis with WaveDown and wavelet-aware attention to preserve fine vascular structures during scale transitions, and evaluates performance using both conventional segmentation metrics and structure-sensitive measures such as capillary recall and fragmentation. Experiments on OCTA-500 3 mm/6 mm and FAROS report improvements over several published and retrained baselines, together with factorial ablations and cross-dataset evaluations intended to analyze the contribution of the proposed frequency-guided design.

Critical strengths

The paper provides a useful structural evaluation protocol for OCTA that captures fine-vessel quality beyond aggregate Dice.

Critical weaknesses
  1. The architectural contribution appears too incremental relative to LWTformer (CVPR 2026 Findings), with highly similar modules sharing the same names, functions, and architectural designs, to meet the level of methodological novelty expected for ICLR.
  2. The experimental evaluation lacks sufficient consistency and rigor. The factorial main-effect magnitude is reported using an unconventional scaling, the same metrics are reported using different aggregation schemes across tables and therefore cannot be directly compared, and methods within the same table are evaluated using different checkpoint-selection and seed protocols. These issues reduce confidence in the fairness and interpretability of the reported results.
Questions for authors
  1. Please explicitly clarify the methodological differences between WaveJoint and LWTformer, particularly for WaveDown, WACGA, and SEA. Which operations are inherited, which are modified, and what is the motivation for each modification? If the main difference is the replacement of learnable wavelet filters with fixed Haar filters, please explain why this constitutes a sufficiently substantive methodological contribution.
  2. In Tables 1 and 2, why is WaveJoint evaluated using the last checkpoint while the retrained baselines use their best validation checkpoints? Was this evaluation policy determined a priori, before inspecting the test-set results? Please clarify how this asymmetric checkpoint-selection strategy ensures a fair comparison.
  3. In Table 3, the 2×2 factorial ablation is conducted using learnable wavelet filters, whereas the final WaveJoint adopts fixed Haar filters. Why was the factorial ablation not performed under the final fixed-Haar configuration? Please clarify whether the conclusions regarding the contributions of WaveDown and WACGA remain valid under the actual model setting.
  4. The protocol-matched baselines are relatively dated and may not adequately reflect the current state of medical image segmentation. Why were stronger and more widely adopted frameworks, such as nnU-Net, as well as more recent segmentation methods, not included for comparison?
Reviewer_K1WEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper combines wavelet frequency decomposition, scale transitions, and attention for joint segmentation of retinal vessels and the foveal avascular zone in OCT angiography. Besides Dice, it evaluates small-vessel recall, connected-component counts, external-data performance, and selected FAZ morphology errors. The central question is whether this combination preserves fine structures more reliably.

Critical strengths
  1. The evaluation addresses a meaningful concern in the task. In Table 2 on page 8, recall on the ≤2px ground-truth skeleton increases from 0.8029 to 0.8354 relative to U-Net, the strongest listed control on this metric. The three-seed factorial experiment also provides evidence that the scale-transition module affects fine-structure recovery. This adds useful information beyond Dice and should count toward the contribution.
  2. The authors report failures and provide a valid local control. The paper includes poor zero-shot small-vessel performance on FAROS. The fixed-Haar versus learnable-filter comparison under the same WACGA+WaveDown configuration and seed 42 is also valid. A higher result from another attention configuration does not invalidate that local comparison.
Critical weaknesses
  1. The general insight added beyond the existing architecture remains unclear. The cited LWTformer already includes related WaveDown, attention, and wavelet-processing components, as well as a fixed-Haar control. The paper therefore needs to clarify what generalizable knowledge is gained by adapting these designs to OCTA. The main three-seed factorial experiment uses learnable filters, while the final fixed-Haar choice mainly relies on a single-seed local experiment. That result is valid, but its stability across seeds and attention settings is not established. PixelUnshuffle controls are present. The remaining gap is a non-wavelet control matched to WaveDown in capacity and gating structure. The current evidence does not sufficiently separate the contribution of the wavelet representation from other architectural changes.
  2. A reduction in component count does not sufficiently establish better recovery of true vessel connectivity. In Table 2, the connected-component ratio decreases from 2.88 to 1.50, a reduction of about 48%. This cannot be directly interpreted as 48% fewer actual vessel breaks. Missing branches and false connections can both reduce component counts. Recall on the ground-truth skeleton also does not penalize all false predicted branches. The text reports a stronger U-Net control, but does not provide its structural metrics. The evidence therefore remains insufficient to establish fewer spatially matched breaks and false connections against that stronger control.
Questions for authors

1.Are multi-seed factorial results available for the final fixed-Haar configuration, together with a non-wavelet control matched in capacity and gating structure? Do they support the general explanation for the final design? 2.How does the stronger U-Net control perform on small-vessel recall and connectivity errors? Are spatially matched break/false-connection measurements or skeleton precision results available? 3.Please use a consistent main-effect scale and specify the small-vessel threshold definition and label version. Do these definitions affect the main comparison?

Reviewer_Rpk5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a method for segmentation of the retinal vessel (RV) tree and foveal avascular zone (FAZ) in Optical coherence tomography angiography (OCTA) images. The central struggle of this segmentation lies in the fine structures of RV (1-3 pixels) and the contextual information needed to segment FAZ. Towards that end, the authors propose a segmentation architecture that is designed to preserve the high frequency details through a Discrete Wavelet Transform (DWT) dependent attention and downsampling blocks. Quantitative and qualitative results show small but robust improvement over prior work.

Critical strengths
  1. The paper is well written and the method is well motivated

  2. Using DWT based inductive priors in attention and downsampling is a neat idea and as far as I know novel

  3. Quantitative and qualitative results show improvement over prior work

  4. Ablations are well conducted

Critical weaknesses
  1. Effect of the Wavelet-Aware Convolutional Gated Attention (WACGA) is not convincingly positive

  2. It would be helpful if some parts of Appendix E is in the main paper

Questions for authors
  1. I’m wondering if the authors explored using smaller kernels in the first few layers of the UNet to resolve the fine details instead of down a down sample after the first two conv layers?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.