FlowOCT: Wavelet Flow Matching for OCT-to-OCTA Translation

1 Introduction

Retinal diseases frequently present with both structural and vascular alterations. Structural optical coherence tomography (OCT) is routinely used to evaluate retinal morphology, while optical coherence tomography angiography (OCTA) provides complementary depth-resolved visualization of retinal blood flow and vascular organization [Huang et al., 1991]. The combined interpretation of OCT and OCTA supports diagnosis, staging, and monitoring of retinal diseases by providing complementary structural and vascular perspectives [De Carlo et al., 2015].

Despite its clinical value, OCTA requires more advanced instrumentation and remains slower to acquire, more sensitive to motion artifacts, and less widely available than OCT

∗Corresponding author.

[Jia et al., 2012]. In clinical practice, many diagnostic decisions are therefore made based solely on structural OCT [Pan et al., 2025], without direct access to angiographic information. This limitation motivates a fundamental question: Can clinically meaningful angiographic information be recovered from structural OCT?

Recent learning-based methods attempt to synthesize OCTA from OCT using generative models [Ran and Cheung, 2021]. While these approaches can produce visually plausible angiograms, most advanced methods rely on stochastic sampling conditioned on OCT, which neglects key properties of retinal vasculature [Zhu et al., 2025a; Zhu et al., 2025b]. Angiographic signals are highly sparse, and clinically meaningful structures are dominated by low-amplitude, high-frequency components. Treating OCT merely as a conditioning signal makes such information easy to ignore, leading models to emphasize large vessels while failing to preserve fine capillary networks. Visual comparison in Figure 1 highlights distinct vascular artifacts in synthesized OCTA images from baseline methods. The image generated by the Diffusion model [Ho et al., 2020] (a) exhibits vessel discontinuity and false vessel growth. Meanwhile, the result from the MuTri model [Chen et al., 2025b] (b) suffers from false vessel growth and vessel adhesion. In contrast, our proposed FlowOCT method (c) effectively mitigates these artifacts, producing vascular structures that are morphologically closer to the ground truth (d). These systematic error patterns highlight the limitations of existing approaches in achieving anatomically plausible OCTA synthesis. As a result, existing generative methods often produce anatomically implausible outputs[Ran and Cheung, 2021; Courtie et al., 2024]. Therefore, the OCTA images synthesized by existing methods often suffer from anatomical inaccuracies. Although visually subtle, such errors can introduce risks in clinical interpretation, as the integrity, connectivity, and morphological fidelity of vascular structures directly impact diagnostic conclusions[Coscas et al., 2016].

Instead of treating OCT as a conditioning signal, we formulate OCT-to-OCTA synthesis as a noise-free flowmatching problem that starts directly from structural OCT and learns a deterministic velocity field to transport it toward the OCTA distribution. This formulation preserves patientspecific anatomy throughout generation and avoids artifacts introduced by stochastic sampling [Lipman et al., 2023]. To faithfully recover vascular structure across spatial scales, we perform flow matching in a wavelet representation rather than in pixel space. This representation separates coarse anatomical layout from fine vascular detail and aligns naturally with the multi-scale organization of retinal vasculature.

Accurate angiographic synthesis further requires correct localization of vascular signals, which are confined to narrow axial bands and exhibit structured spatial organization [Garrity et al., 2017]. To preserve this anatomy, we impose two complementary constraints. First, to prevent diffuse textures and vessel displacement across depth, we introduce retinaldepth contrastive alignment, which enforces depth-aware discrimination by aligning vascular signals to the correct retinal slice while remaining distinct from other depths. Second, to reflect how retinal imaging is interpreted clinically, we constrain synthesis through spatial projections to twodimensional (2D) retinal layer views, ensuring that vascular spatial organization aligns with standard clinical assessment.

Contributions. We make the following contributions:

• We propose FlowOCT, a wavelet-based flow-matching framework that reformulates OCT-to-OCTA synthesis as a noise-free, anatomically faithful transport process from structural OCT wavelet representations to OCTA signals.
• To ensure anatomically and clinically meaningful vascular localization, we introduce depth- and spatial-aware constraints, including retinal-depth contrastive alignment (RDCA) and spatial retinal layer projection con-

straint.

• We introduce a Structure-Focused Objective that mitigates signal sparsity by emphasizing vascular regions, thereby improving capillary quality in training and establishing a correlated foreground evaluation metric.
• We demonstrate through image quality metrics and downstream diagnostic tasks that our method improves structural fidelity and clinical reliability.

2 RelatedWork

2.1 OCT-to-OCTA Image Translation

Cross-modality OCT-to-OCTA synthesis is challenging because of the intrinsic modality gap between static tissue structures and dynamic blood flow. Early 2D methods evolved from general frameworks like Pix2Pix [Isola et al., 2017] to specialized architectures, including Texture-UNet [Zhang et al., 2021] and AdjacentGAN [Li et al., 2020b] for slicewise B-scans, and MultiGAN [Pan et al., 2022] for en-face projections. Despite their utility in quantifying retinal features [Badhon et al., 2024; Wang et al., 2025], 2D approaches inherently lose volumetric information and suffer from interslice discontinuity, leading to fragmented reconstruction.

With the availability of large-scale datasets like OCTA500 [Li et al., 2024a], recent research has shifted toward 3D volumetric translation. State-of-the-art frameworks, such as TransPro [Li et al., 2024b] and XOCT [Khosravi et al., 2025], utilize projection consistency and cross-dimension supervision, while MuTri [Chen et al., 2025b] employs multi-view semantic alignment via VQ-VAE. However, these methods often overlook the inherent sparsity and multi-scale organization of vascular signals. Consequently, they struggle to recover fine capillary networks and frequently produce anatomical distortions, such as vessel adhesion, fragmentation, and spurious vessels, which compromise clinical reliability.

2.2 Diffusion and FlowMatchingModels

Generative modeling has transitioned from GANs toward Diffusion Models (DMs) [Ho et al., 2020; Song et al., 2021]. Despite their success in medical reconstruction [Wu et al., 2023; M¨uller-Franzes et al., 2023], DMs face high computational costs due to the iterative nature of reversing stochastic paths. To improve efficiency, recent studies have conducted diffusion in compressed domains, such as latent spaces [Rombach et al., 2022] or the wavelet domain [Phung et al., 2023; Huang et al., 2024], which leverages frequency decomposition to preserve structural sparsity and fine vascular details while reducing signal complexity.

To further accelerate inference, FlowMatching (FM) [Lipman et al., 2023; Albergo and Vanden-Eijnden, 2023; Liu et al., 2023; Tong et al., 2023] has emerged as a simulation-free framework based on Continuous Normalizing Flows (CNFs). Recent progress has been done on medical imaging [Chen et al., 2025a; Chen et al., 2026; Jedynak, 2025] . By integrating wavelet representations with Flow Matching, our method unifies frequency-domain spatial efficiency with straight-line inference to ensure high-fidelity and rapid 3D OCT-to-OCTA translation.

3 Methodology

As shown in Figure 2, our proposed FlowOCT integrates four complementary modules that jointly address the challenges of signal sparsity, depth specificity, and clinical consistency in OCTA synthesis, enabling anatomically faithful and clinically viable vascular generation.

3.1 Problem Setup

Let π0 and π1 denote the distributions of structural OCT and OCTA volumes, respectively. Both modalities are defined on a shared spatial domain Ω ⊂ R3 with coordinates (h, w, z), where z indexes retinal depth. Given paired samples

our goal is to learn a conditional generative model that maps an observed OCT volume x0 ∼ π0 to a synthesized OCTA volume ˆx1 ∼ π1. N denotes the number of paired training samples.

3.2 Wavelet Representation

Angiographic signals exhibit strong multi-scale structure, with fine vessels encoded at high frequencies and global vascular organization at coarse scales. To explicitly represent this structure, we operate in the wavelet domain.

Let W denote 3-Dimension discrete wavelet transform (DWT). For a volume x ∈ RH×W×Z , we compute

with s indexing wavelet scales and B denoting orientation subbands. The inverse transform recovers the original volume:

3.3 Noise-FreeWavelet FlowMatching

We then define the generative process directly in wavelet space. Let

denote the wavelet representations of OCT and OCTA volumes. We model a continuous flow governed by

where vθ is a learnable velocity field conditioned on the source OCT wavelet coefficients w0.

The flow is initialized asw(0) = w0,

and integrated to t = 1, yielding synthesized coefficients

The final OCTA volume is obtained via

Flowmatching objective. Given paired samples (w0,w1), we define a linear interpolation path

with target velocity

The flow-matching loss is

where

3.4 Retinal-Depth Contrastive Alignment

Retinal vasculature is organized into depth-specific plexuses occupying narrow axial bands. To preserve this organization, we introduce Retinal-Depth Contrastive Alignment (RDCA) to enhance depth-specific vascular delineation and prevent axial diffusion and misalignment.

Given real and synthesized OCTA volumes x1 and ˆx1, we extract axial slices

Each slice is encoded using a shared depth encoder fd, producing normalized embeddings

We enforce depth-wise alignment using a contrastive objective

3.5 Spatial Retinal Layer Projection Constraint

While RDCA enforces correct depth-wise localization, clinically meaningful interpretation of OCTA is performed at two complementary levels: voxel-level analysis for fine vessel details and layer-projection assessment for global vascular patterns. To ensure high-quality reconstruction that meets both clinical needs, we introduce a spatial retinal layer projection constraint that incorporates these two aspects.

First, to preserve fine vascular structures at the original spatial resolution, we introduce a voxel-wise L1 loss:

Second, to reflect clinical practice and enforce spatial consistency, we introduce a retinal layer projection constraint, following a setup similar to [Li et al., 2024b; Khosravi et al., 2025]. Given a volumetric OCTA image x1, we project it along the depth axis within predefined retinal layer ranges {Zk}Kk=1 and enforce consistency between real and synthesized views:

Here, A ∈ {mean, max} denotes the mean or maximum projection operator.

Combining these two complementary constraints, we obtain the spatial retinal layer projection constraint:

where β balances the contribution of the layer-projection term. Both terms employ the L1 norm for its robustness to outliers and sparse signal characteristics.

3.6 Structure-Focused Objective

Since angiographic signals in OCTA volumes are highly sparse, with informative vascular structures occupying only a small fraction of the spatial domain, standard reconstruction losses are dominated by large background regions and therefore tend to underemphasize fine vascular details. To address this imbalance, we reformulate existing losses as structurefocused losses that emphasize informative vascular regions.

Specifically, we construct a 3D binary mask M from the ground-truth OCTA by adaptively detecting content regions with significant angiographic signals via thresholding, thereby identifying regions with significant angiographic signal. The loss is then computed betweenM⊙ˆx1 andM⊙x1, rather than between ˆx1 and x1. Similarly, for LFM, we compute frequency masks and are applied to feature maps w instead of the input x. We denote the structure-focused loss as Ls .

The full training objective then is

where the λProj and λRDCA are loss weighting term.

4 Experiments

4.1 Experiments Setup

Dataset. We evaluate our method on the OCTA-500 dataset [Li et al., 2024a; Li et al., 2020a], which comprises two subsets with different fields of view: OCTA-3M (200 pairs, 304 × 304 × 640 voxels) and OCTA-6M (300 pairs, 400 × 400 × 640 voxels). Adopting the protocol from [Li et al., 2024b], we partitioned OCTA-3M into 140/10/50 and OCTA-6M into 180/20/100 for training, validation, and testing, respectively. All volumes underwent standardized preprocessing, including intensity normalization and retinal region cropping. Furthermore, all data pairs are annotated with seven diagnostic categories to support the downstream clinical evaluation detailed in Section 4.3.

Implementation Details. For the proposed method, we parameterize the flow matching process using the DiffusionModelUnet framework from the Monai library. Models are trained with the Adam optimizer at a learning rate of 2×10-4 . Unless otherwise stated, we set β = 0.2, λProj = 0.5 and λRDCA = 0.005 (calibrated to balance loss scales: RDCA ∼ 4 vs. other losses ∼ 0.001). Through comparative testing, we have chosen A = max as the projection operator (Eq.(17)). We train the model with a batch size of 1 for 200 epochs. All experiments are conducted on a virtual GPU (vGPU) with 48 GB of memory.

Inference Configuration. For all generation tasks on the validation and test sets, we employ a fixed 3-step ODE solver to numerically solve the learned probability flow ODE [Lipman et al., 2023]. This configuration balances optimal quantitative performance with high visual clarity while maintaining efficient inference. Further increasing the number of steps yields marginal or diminishing returns with substantially higher computational costs.

Evaluation Metrics. We employ a comprehensive set of metrics, including Mean Absolute Error (MAE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM) [Hor´e and Ziou, 2010], Learned Perceptual Image Patch Similarity (LPIPS) [Zhang et al., 2018], and Fr´echet Inception Distance (FID) [Heusel et al., 2017], to thoroughly evaluate the fidelity of the reconstructed results. Specifically, LPIPS uses AlexNet as the feature extraction network, while FID is computed based on Inception-v3.

We observed that calculating these metrics on the full image volume yields artificially inflated scores due to interference from extensive background regions. This directly motivates our structure-focused objective function. To concentrate evaluation on meaningful content, we propose Foreground Metrics, which apply the evaluation only to the foreground region extracted via the structure-guided objective, in contrast to traditional Full Volume Metrics.

4.2 Comparison with Baseline Methods

We compare the proposed method with several representative baseline methods for OCT-to-OCTA synthesis. To ensure a fair comparison, all methods were trained and evaluated using the same data splits and preprocessing pipelines. The baseline methods include: Pix2pix 2D and Pix2pix 3D [Isola et al., 2017], Adjacent GAN [Li et al., 2020b], Diffusion Model [Ho et al., 2020], Flow Matching [Moschetto et al., 2025], Transpro [Li et al., 2024b], and MuTri [Chen et al., 2025b]. All baseline implementations followed their respective original frameworks or public repositories.

Experimental results on OCTA-500 dataset

Table 1 and Table 2 summarize the quantitative comparison results between the proposed method, FlowOCT, and various baseline methods on the OCTA-3M and OCTA-6M datasets.

On the OCTA-3M dataset, our method achieves optimal performance in perceptual quality (PSNR, SSIM) and perceptual similarity (LPIPS), outperforming state-of-the-art methods with a 1.7% improvement in Full PSNR, a 2.1% improvement in Foreground SSIM, and a 16.60% reduction in Foreground FID. On the OCTA-6M dataset, the advantages of our method in structural preservation and perceptual quality are further demonstrated. Compared to the second-best method, our approach achieves a 6.25% improvement in Foreground MAE, a 4.8% increase in Foreground SSIM, along with an 8.1% reduction in Full LPIPS and an 8.7% reduction in Foreground FID.

Overall, FlowOCT consistently outperforms all baselines across both datasets and all metrics, demonstrating its superior capability in preserving vascular structures and achieving high-fidelity OCTA synthesis.

Visualization results on OCTA-500 dataset

For qualitative evaluation of the results, we selected visual comparison examples across three projection dimensions (axial, coronal, and sagittal) generated by different methods, as illustrated in Figure 3. Additionally, we generated difference maps (see Figure 4) by performing pixel-wise subtraction between the synthesized OCTA images and the ground truth. Overall, GAN-based methods (Pix2pix 2D/3D [Isola et al., 2017] and Adjacent GAN [Li et al., 2020b]) tended to exhibit checkerboard artifacts or stripe-like noise, along with lower brightness. The Diffusion Model [Ho et al., 2020] produced overly blurry outputs with erroneous vessel connections and breaks. Transpro [Li et al., 2024b] showed improved structural reconstruction and brightness, yet still suffered from vessel over-generation and adhesion. In contrast, our proposed FlowOCT method consistently achieved superior performance across all views: the generated OCTA images exhibited the smallest discrepancy in the difference maps, producing clearer, more continuous vessel structures with finer details, and delivering visual results closest to the ground truth.

4.3 DownstreamDiagnostic Evaluation

This study employs a downstream diagnostic task on the OCTA-6M dataset to evaluate generative models. A ResNet classifier under the MONAI framework is trained on a combination of 180 real training cases and the Seen Set data. The real test set is randomly split into a Seen Set (40 cases) for authenticity assessment and an Unseen Set (60 cases) for generalizability evaluation. Each generative model synthesizes corresponding data under identical conditions. Classification accuracy on the Synthetic-Seen Set measures the model's ability to imitate known real-world patterns (authenticity), while accuracy on the Synthetic-Unseen Set gauges its generalization to unknown distributions (generalizability). This dual-dimensional strategy avoids complex classifier tuning and provides a comparable quantitative benchmark.

All models are evaluated under uniform conditions. The results are ranked primarily based on the Synthetic-Unseen Set performance, supplemented by a consistency analysis of the Synthetic-Seen Set results, with a focus on the deviation from real-data performance.

As summarized in Table 3, the evaluation reveals clear distinctions. In the authenticity dimension, where real OCTA data serves as an upper bound at 82.50%, most models achieve between 47.50% and 55.00%. Notably, our method FlowOCT reaches 65.00%, performing the best among all synthetic models and demonstrating effective imitation of known patterns. Regarding generalizability, the real data itself attains only 51.67%, underscoring the task's difficulty. Here, FlowOCT again leads at 46.67%, outperforming the second-best model (FlowMatching [Moschetto et al., 2025]) by 1.67% , which confirms its strong out-of-distribution generalization. In contrast, most other models show limited generalizability (23.33%-40.00%), indicating their generated data fail to transfer effectively to unseen real distributions.

4.4 Ablation and Sensitivity Analysis

We conducted ablation studies to validate the effectiveness of each component and performed sensitivity analysis on lossweight configurations, thereby confirming the robustness of the overall framework.

Table 4 demonstrates that FlowOCT achieves optimal performance across all metrics. Specifically, removing the wavelet representation degrades overall image quality, while omitting the RDCA module significantly impairs vascular structural integrity. Furthermore, ablating the projection constraint and structure-focused loss reduces cross-view consistency and reconstruction accuracy, respectively.

In summary, the wavelet-domain and RDCA module drive the primary performance gains by enhancing frequencydomain modeling and depth-wise structure preservation. Coupled with projection constraint and structure-focused loss, these components ensure clinical utility and detail precision, validating their collective necessity.

To further assess robustness, we perform a joint sensitivity analysis over the two loss weights, λRDCA and λProj, which jointly govern anatomical alignment and projection consistency. As shown in Table 5, performance remains stable across a wide range of loss-weight combinations, with the default setting λRDCA = 0.005 and λProj = 0.5 achieving the best overall results. This joint sensitivity analysis demonstrates that the proposed framework is not overly sensitive to precise hyperparameter tuning, supporting its robustness and practical applicability.

5 Conclusion

In this paper, we present FlowOCT, a novel framework for OCT-to-OCTA synthesis that incorporates distributional modeling and anatomical constraints. By formulating angiography generation as conditional flow matching, our method directly maps structural OCT to the OCTA distribution via deterministic paths. Operating in the wavelet domain enables the explicit modeling of vascular structures across multiple scales, while the retinal-depth contrastive alignment (RDCA) ensures anatomically consistent vessel organization. Experimental results demonstrate that FlowOCT consistently outperforms state-of-the-art methods in both reconstruction fidelity and downstream diagnostic accuracy. This work provides a principled framework for anatomy-aware angiographic synthesis and suggests a path toward leveraging routine OCT for vascular imaging.

Acknowledgments

This work was supported in part by the National Natural Science Foundation of China under Grants 62541207, W2511070, 32241027 and 62472034, and in part by the Key Research and Development and Achievement Transformation Project of Inner Mongolia Autonomous Region in the Social Welfare Field under Grant 2026YFSH0177. The corresponding author is Dehui Qiu (qiudehui@cnu.edu.cn).