PaperlayerView exact revision on arXiv

Source-linked AI summary

OCTA-500: A Retinal Dataset for Optical Coherence Tomography Angiography Study

Mingchao Li, Kun Huang, Qiuzhuo Xu, Jiadong Yang, Yuhan Zhang, Zexuan Ji, Keren Xie, Songtao Yuan, Qinghuai Liu, Qiang Chen

arXiv:2012.07261v3eess.IVcs.CV

TL;DR

The paper addresses the need for reliable retinal vascular segmentation to support quantitative OCTA analysis. It introduces OCTA-500 and the CAVF task, optimizes IPN to IPN-V2, and studies dataset characteristics, finding that data scale and disease diversity strongly shape segmentation performance.

  • Problem

    Quantitative OCTA indicators require vascular segmentation, while existing methods perform poorly with artifacts, blur, weak contrast, low resolution, and disease.

  • Method

    The paper introduces OCTA-500, proposes the unified CAVF task, optimizes IPN into IPN-V2, and evaluates training size, inputs, baselines, and diseases.

  • Results

    Data are a driving factor for CAVF segmentation performance, while IPN-V2 improves segmentation quality and speed over IPN by a large margin and achieves competitive results.

  • Takeaways & Limitations

    OCTA-500 provides a broad benchmark whose disease diversity remains challenging, leaving room for improved methods and more data.

  • Takeaways & Limitations

    IPN training and testing require considerable GPU memory, so volume splitting can discard global information.

Abstract

from arXiv · show

Optical coherence tomography angiography (OCTA) is a novel imaging modality that has been widely utilized in ophthalmology and neuroscience studies to observe retinal vessels and microvascular systems. However, publicly available OCTA datasets remain scarce. In this paper, we introduce the largest and most comprehensive OCTA dataset dubbed OCTA-500, which contains OCTA imaging under two fields of view (FOVs) from 500 subjects. The dataset provides rich images and annotations including two modalities (OCT/OCTA volumes), six types of projections, four types of text labels (age / gender / eye / disease) and seven types of segmentation labels (large vessel/capillary/artery/vein/2D FAZ/3D FAZ/retinal layers). Then, we propose a multi-object segmentation task called CAVF, which integrates capillary segmentation, artery segmentation, vein segmentation, and FAZ segmentation under a unified framework. In addition, we optimize the 3D-to-2D image projection network (IPN) to IPN-V2 to serve as one of the segmentation baselines. Experimental results demonstrate that IPN-V2 achieves an ~10% mIoU improvement over IPN on CAVF task. Finally, we further study the impact of several dataset characteristics: the training set size, the model input (OCT/OCTA, 3D volume/2D projection), the baseline networks, and the diseases. The dataset and code are publicly available at: https://ieee-dataport.org/open-access/octa-500.

1 INTRODUCTION

OCTA provides functional views of retinal microvasculature, but public datasets remain scarce and limited in scale, diversity, modalities, and labels. OCTA-500 addresses these gaps with extensive multimodal data and annotations, while introducing the unified CAVF segmentation task and IPN-V2 baseline.

  • Motivation: OCTA complements OCT by providing functional information on retinal blood vessels and enabling projections from different retinal layers.OCT captures 3D retinal structure but cannot directly provide blood-flow information.
  • Motivation: Quantitative OCTA indicators such as vessel density and FAZ area require segmentation of vascular structures in OCTA images.These indicators support objective assessment of clinical outcomes.
  • Dataset gap: Public OCTA datasets are few and small, with limited disease diversity, single-modality projection maps, single-task labels, and scarce 3D or artery-vein annotations.Existing datasets lack publicly available 3D FAZ and artery-vein segmentation labels.
  • OCTA-500: OCTA-500 contains imaging from 500 subjects across two FOVs, OCT/OCTA volumes, six projections, four text-label types, and seven segmentation-label types.The dataset includes 361,600 scans and exceeds 80 GB in size.
  • Tasks and baselines: OCTA-500 introduces CAVF, which unifies capillary, artery, vein, and FAZ segmentation, and provides IPN-V2 as an optimized 3D-to-2D baseline.The experiments examine training-set size, input modality and structure, baseline networks, and disease effects.

2 RELATED WORK

Prior OCTA datasets and segmentation methods support vessel and FAZ analysis but remain narrow in task, modality, disease coverage, and annotation availability. OCTA-500 broadens this landscape with integrated labels and a unified multi-object segmentation direction.

  • OCTA datasets: Existing public OCTA datasets focus on single tasks, including vessel segmentation or FAZ segmentation, and no prior dataset provides multi-task OCTA annotations.Giarratano and ROSE target vessel segmentation, whereas OCTAGON and FAZID target FAZ segmentation.
  • OCTA datasets: Existing datasets generally contain only OCTA projection maps, with limited coverage of retinal layers, modalities, FOVs, and 3D volumes.ROSE-1 and OCTAGON include superficial and deep projections, while Giarratano and FAZID include only superficial projections.
  • OCTA datasets: Disease coverage is limited across prior datasets: FAZID includes two diseases, OCTAGON only diabetic retinopathy, and Giarratano only normal cases.Common diseases such as AMD, CNV, CSC, and RVO are absent from the cited existing datasets.
  • Vessel segmentation: Threshold-based and filtering-based vessel methods are vulnerable to noise, artifacts, blur, weak contrast, low resolution, and disease, while discriminating capillaries, arteries, and veins remains difficult.These methods can roughly support vessel-density and vessel-skeleton-density estimation but lack sufficient recognition capability for specific structures.
  • Vessel segmentation: Deep learning methods, including U-Net, CS-Net, OCTA-Net, and AV-Net, have advanced OCTA vessel segmentation, but artery-vein development remains constrained by limited public annotations.On one small capillary dataset, U-Net and CS-Net achieved Dice = 0.89, while OOF achieved Dice = 0.86.
  • FAZ segmentation: FAZ segmentation enables quantitative FAZ area and perimeter evaluation and foveal-center localization, motivating continued development of supervised and unsupervised methods.The FAZ is a nonperfused foveal region surrounded by interconnected retinal vessels, and its size is correlated with visual acuity and disease.
  • Main contributions: OCTA-500 addresses these gaps by providing comprehensive multimodal, multi-label data and a CAVF task integrating artery, vein, capillary, and FAZ segmentation.Its contributions also include optimizing IPN to IPN-V2 and studying training size, model input, baselines, and diseases.

3 OCTA-500 DATASET

OCTA-500 is organized into two field-of-view subsets and provides multimodal OCT/OCTA volumes, projections, clinical labels, and segmentation resources. Its projection maps use retinal-layer information with average or maximum axial projection, while disease and demographic distributions differ between subsets.

  • 3.1 Data Collection: OCTA-500 contains 500 subjects divided into OCTA_6mm (300 subjects, 6 mm × 6 mm) and OCTA_3mm (200 subjects, 3 mm × 3 mm).Both subsets were acquired using the same spectral-domain OCT system, with one eye included per subject.
  • 3.2 OCT and OCTA Volumes: The dataset includes OCT and OCTA volumes, with OCT providing retinal structure and OCTA providing blood-flow information.OCTA volumes were generated from OCT volumes using the SSADA algorithm and resized by bilinear interpolation.
  • Dataset Contents: The released volumes comprise 361,600 B-scans, and the dataset includes two FOVs, six projections, four text-label types, and seven segmentation-label types.The text labels cover age, gender, eye, and disease; the section’s distribution analysis reports differences in disease diversity between FOV subsets.
  • 3.3 Projection Maps: OCTA-500 provides six projection maps generated from ILM, OPL, and BM layer positions using average or maximum axial projection.Researchers can generate additional projection maps using the released layer segmentation annotations.
  • 3.4 Text Labels: OCTA_6mm has 69.7% subjects with ophthalmic diseases, whereas OCTA_3mm has 20%, with AMD, DR, and CNV represented in OCTA_3mm.OCTA_6mm includes AMD, DR, CNV, CSC, RVO, and additional low-count diseases.

3.5 Segmentation Labels

OCTA-500 supplies segmentation labels for vessels, capillaries, FAZ, and retinal layers through combinations of manual annotation, multimodal guidance, and model-assisted processing. The procedures address noise, incomplete visibility, disease-related layer errors, and the difficulty of dense capillary labeling.

  • 3.5 Segmentation Labels: The segmentation labels cover large vessels, arteries, veins, capillaries, 2D and 3D FAZ, and retinal layers.These labels support the dataset’s unified segmentation resources.
  • 3.5.1 Large Vessel: Large vessels were manually labeled on OCTA projection B5 using coarse annotation followed by fine-grained boundary correction.Manual correction was used because threshold binarization retained capillaries and noise.
  • 3.5.2 Artery and Vein: Artery-vein labels combine color fundus images for main-vessel classification with OCT/OCTA projections for crossover and branch-point information.Color fundus images provide wider fields and color information, while OCT/OCTA maps help determine vessel connectivity.
  • 3.5.3 Capillary: Capillary labels were produced from manually annotated B5 slices and model-assisted whole-map segmentation using IMN, topology optimization, and denoising.The process aimed to address the time cost, limited resolution, and noise that make exhaustive manual capillary labeling difficult.
  • 3.5.5 Retinal Layers: Six retinal layers were labeled and corrected, including ILM, IPL, OPL, ISOS, RPE, and BM.Disease cases such as AMD required corrections because fluid and pigment epithelial detachment altered layer structure.

4 CAVF TASK AND BASELINES

The CAVF task unifies capillary, artery, vein, and FAZ segmentation, while the baselines include 2D networks and optimized 3D-to-2D IPN-V2. IPN-V2 addresses IPN’s computational and semantic limitations through rapid 3D-to-2D projection followed by 2D segmentation.

  • 4.1 CAVF Task: CAVF unifies capillary, artery, vein, and FAZ segmentation under one framework for joint evaluation and quantitative analysis.The task groups interrelated structures whose separate segmentation would otherwise require one model per task.
  • 4.2 Baselines: The evaluated 2D-to-2D baselines include U-Net, UNet++, UNet 3+, Attention U-Net, CS-Net, and AV-Net.U-Net is described as a commonly used medical image segmentation architecture.
  • 4.2.2 3D-to-2D Baselines: IPN inputs 3D OCT/OCTA volumes and outputs 2D segmentation results end-to-end, avoiding dependence on layer segmentation for projection images.Using complete 3D information is described as reducing information loss and improving segmentation performance.
  • 4.2.2 3D-to-2D Baselines: IPN training and testing require considerable GPU memory, often forcing volume blocks that lose global information; its horizontal processing also lacks high-level semantic information.These limitations contributed to poor performance on the more challenging CAVF task.
  • 4.2.2 3D-to-2D Baselines: IPN-V2 uses a fast projection module to compress 3D information into 2D, then applies a U-Net-like 2D segmentation network end-to-end.The projection module combines multiscale convolutions and unidirectional pooling to condense useful features and stabilize training.
  • 4.3 Evaluation Metrics: mIoU averages IoU across the k segmentation objects, while Params, Memory, and test speed characterize computational complexity.The evaluation also defines Dice, accuracy, sensitivity, and specificity for object-level segmentation assessment.

5 EXPERIMENTS

Experiments evaluate training-set size, input modality, baseline architecture, and disease condition on OCTA-500. Performance generally improves with more data, IPN-V2 is the strongest baseline by mIoU, and disease-related vascular changes make segmentation harder.

  • 5 Experiments: The experiments study training-set size, model input, baseline networks, and disease conditions on OCTA-500.Experiments use defined training, validation, and test splits from the OCTA_6mm and OCTA_3mm subsets.
  • 5.2 Number of Training Data: mIoU increases with training samples, rising rapidly below 120 samples and becoming less pronounced above 120.The authors regard 120 samples as a basic requirement for this task.
  • 5.2 Number of Training Data: FAZ segmentation appears to require the most training samples, whereas capillary segmentation with U-Net requires relatively few.Artery, vein, and FAZ IoUs show trends broadly similar to mIoU.
  • 5.2 Number of Training Data: IPN-V2 surpasses U-Net when training samples exceed 40, indicating that the 3D-to-2D method requires more data than the 2D-to-2D method.For more than 20 samples, IPN-V2’s capillary IoU is significantly higher than U-Net’s because 2D projection noise can be misclassified as capillaries.
  • 5.3 Model Input: For 2D segmentation, all B1-B6 projections perform best while B4-B6 are only 0.1 mIoU lower; for 3D-to-2D segmentation, OCTA volume alone performs best.Using only OCT volume gives capillary IoU = 34.93%.
  • 5.4 Comparison of Different Baselines: IPN-V2 achieves the best mIoU on both OCTA_6mm and OCTA_3mm, with a large capillary-segmentation margin and competitive artery, vein, and FAZ results.Attention U-Net achieves the best IoU for all categories in OCTA_3mm.
  • 5.5 Performance in Different Diseases: Disease-condition mIoU ranks normal, CSC, CNV, AMD, DR, and RVO, with non-perfusion and vascular morphological changes increasing segmentation difficulty.Examples show FAZ over- and under-segmentation and some artery-vein mis-segmentation in disease cases.

6 DISCUSSION

OCTA-500 supports joint CAVF segmentation and additional retinal quantification tasks through its diverse annotations. Its paired modalities and disease labels also broaden possible OCT/OCTA research applications.

  • 6 Discussion: CAVF demonstrates joint segmentation of capillaries, arteries, veins, and the FAZ.The unified task is presented as a framework for these four segmentation targets.
  • 6 Discussion: OCTA-500’s labels support large-vessel, 3D FAZ, and retinal-layer segmentation alongside CAVF.These tasks enable retinal layer thickness, vessel density, and FAZ area quantification.
  • 6 Discussion: Paired OCT/OCTA volumes and disease labels support disease classification and broader multimodal OCTA studies.The paper identifies modality transformation as one possible application of the multimodal data.

7 CONCLUSION

The paper introduces OCTA-500, the CAVF task, and IPN-V2, then evaluates how dataset characteristics affect CAVF segmentation. Results identify data scale, model design, and disease diversity as important factors for future progress.

  • 7 Conclusion: OCTA-500 contains OCTA imaging from 500 subjects with rich images and annotations, supporting the proposed CAVF task.CAVF integrates artery, vein, capillary, and FAZ segmentation under a unified framework.
  • 7 Conclusion: The study evaluates training-set size, model input, baseline networks, and disease conditions for CAVF segmentation.IPN was optimized to IPN-V2 as one of the segmentation baselines.
  • 7 Conclusion: Data scale drives segmentation performance, while IPN-V2 improves segmentation quality and speed over IPN and achieves competitive results.The authors state that considered deep learning methods have not yet reached saturation on this task.
  • 7 Conclusion: OCTA-500 is intended to provide a systematic quantitative framework for retinal image analysis and stimulate further OCT/OCTA research.Future plans include collecting images, annotating ground truths, and optimizing methods.

RELATED WORK

Prior OCTA segmentation studies use varied architectures and datasets, but reported resources and task coverage remain fragmented across vessel and FAZ segmentation.

  • Vessel and related segmentation: Public OCTA studies include U-Net-like, attention-based, cascaded, semi-supervised, and 3D-to-2D segmentation approaches.The reviewed methods span vessel, artery-vein, FAZ, nonperfusion-area, and retinal-layer segmentation tasks.
  • Task organization: The related-work summary covers separate vessel and FAZ tasks rather than a unified capillary, artery, vein, and FAZ framework.The paper subsequently introduces CAVF as a unified multi-object segmentation task.
  • 3D-to-2D methods: IPN maps 3D OCTA data to 2D segmentation, while later variants add adaptive pooling, feature fusion, or dual-way projection learning.Reported OCTA-500 Dice values include IPN DSC = 0.8861, PAENet DSC (6 mm) = 0.8969, and RPS-Net DSC (6 mm) = 0.8989.
  • Reported OCTA-500 results: FARGO reports OCTA-500 DSC (6 mm) = 0.9272 and DSC (3 mm) = 0.9839 for FAZ segmentation, while RPS-Net reports DSC (6 mm) = 0.9155 and DSC (3 mm) = 0.9780.These results illustrate variation across methods and fields of view in the summarized OCTA literature.

ABLATION STUDY OF IPN-V2

The IPN-V2 ablation evaluates projection and segmentation components under memory-constrained OCTA inputs, then tests progressively larger input sizes.

  • Component ablations: The ablation compares basic IPN, an IPN-V2 CNN-backbone variant, and IPN-V2 without FPM on OCTA_6mm.The CNN variant removes down-sampling and concatenation operations to assess segmentation without high-level semantic information.
  • Input setting: IPN training uses randomly cropped OCTA patches of 128 × 100 × 100 because the full 128 × 400 × 400 volume exceeds available GPU memory.The same patch input is used across the ablation configurations.
  • Component ablations: IPN-V2 improves considerably over IPN, with gains attributed to FPM feature compression and U-Net high-level semantics.Artery, vein, and FAZ segmentation improve significantly after adding the semantic information from U-Net.
  • Input-size ablation: Performance improves as input size increases, suggesting that larger inputs provide more global features beneficial for segmentation.The study evaluates 128 × 200 × 200 and full-volume 128 × 400 × 400 inputs in addition to cropped patches.
  • Input-size ablation: With fewer parameters and lower GPU memory use than IPN, IPN-V2 processes the full OCTA volume and achieves more than 10% mIoU improvement.The improvement is reported for the full-volume IPN-V2 configuration relative to IPN.

SEGMENTATION TASKS ON OCTA-500

OCTA-500 evaluates multiple segmentation labels, including conventional vessel and FAZ tasks alongside the unified CAVF task and visualized task-specific outputs.

  • Segmentation task scope: OCTA-500 treats large vessel, capillary, artery, vein, FAZ, 3D FAZ, and retinal-layer labels as separate segmentation tasks.Results are reported for each task using the corresponding baselines and evaluation settings.
  • Task definitions: CAVF subtasks differ from the separate tasks: CAVF capillaries exclude arteries and veins, while its artery and vein labels intersect.Fig. S1 provides segmentation examples across the different task definitions.
  • Vessel segmentation: Tables S4–S6 report results for large vessel, capillary, and artery segmentation on OCTA-500.The supplied captions identify the reported metrics as percentages.
  • FAZ and layer segmentation: Tables S7–S10 report vein, FAZ, 3D FAZ, and retinal-layer segmentation results on OCTA-500.The tables distinguish 3D FAZ and layer segmentation from the 2D vessel and FAZ tasks.
  • Comparative visualizations: Supplementary figures provide PR curves for large vessel, capillary, artery, vein, and FAZ segmentation across different methods.These curves complement the tabulated task results.

EVALUATION OF COMPARATIVE EXPERIMENTS

Comparative experiments report several CAVF metrics and baseline speed-space behavior, showing that IPN-V2 retains computational advantages over typical 3D-to-2D processing.

  • CAVF metrics: CAVF evaluations report Dice, accuracy, sensitivity, and specificity across the compared methods.These metrics are provided in Tables S11–S14.
  • Speed and space: 2D-to-2D baselines use less training GPU memory and test faster than 3D-to-2D baselines.The difference is attributed to 3D volume reading and convolution costs in 3D-to-2D networks.
  • Speed and space: IPN-V2 reaches test speeds close to 2D baselines and completes testing with at least 2 subjects in 1 second.The reported speed is described as acceptable in the paper.
  • Speed and space: When projection-map generation and layer segmentation are included, 2D baselines can be slower than IPN-V2.The 2D baselines take projection maps as input, whose generation relies on layer segmentation.
Loading 2012.07261v3…