[Paper Review] OCTA-500: A Retinal Dataset for Optical Coherence Tomography Angiography Study

Mingchao Li, Kun Huang+8 more|arXiv (Cornell University)|Dec 14, 2020

Retinal Imaging and Analysis21 citations

TL;DR

OCTA-500 introduces the largest publicly available retinal dataset for optical coherence tomography angiography, comprising 500 subjects with dual-field-of-view OCTA and OCT volumes, extensive annotations (7 segmentation types, 4 labels), and a novel CAVF multi-object segmentation task. The study proposes IPN-V2, an optimized 3D-to-2D projection network achieving ~10% mIoU improvement over baseline on CAVF, and evaluates dataset impact on model performance across input types, training size, and disease conditions.

ABSTRACT

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.

Motivation & Objective

• To address the scarcity of publicly available, large-scale OCTA datasets in retinal imaging and neuroscience research.
• To provide a richly annotated dataset with dual-field-of-view OCTA and OCT volumes, multi-modal projections, and detailed clinical and anatomical labels.
• To introduce a unified multi-object segmentation task (CAVF) integrating capillary, artery, vein, and FAZ segmentation.
• To develop and evaluate IPN-V2, an improved 3D-to-2D image projection network for enhanced segmentation performance.
• To investigate the impact of dataset characteristics—such as training set size, input modality (OCT/OCTA), model architecture, and disease presence—on segmentation performance.

Proposed method

• The OCTA-500 dataset includes 500 subjects with two fields of view (FOVs), capturing both OCT and OCTA volumes for each subject.
• The dataset provides six types of 2D projections (e.g., en face) from 3D volumes, enabling multi-scale analysis of retinal microvasculature.
• A novel multi-object segmentation task, CAVF, unifies capillary, artery, vein, and foveal avascular zone (FAZ) segmentation into a single framework.
• The IPN-V2 network is developed by optimizing the original IPN architecture to improve feature representation and spatial alignment in 3D-to-2D projection.
• The study evaluates model performance across varying training set sizes, input modalities (OCT vs. OCTA, 3D vs. 2D), and baseline networks.
• Extensive ablation studies assess the influence of disease status and data diversity on segmentation accuracy.

Experimental results

Research questions

• RQ1 How does the size of the training set affect segmentation performance on the CAVF task in OCTA imaging?
• RQ2 What is the comparative performance of 3D volume inputs versus 2D projection inputs in retinal vessel segmentation tasks?
• RQ3 How do different baseline networks influence segmentation accuracy on the OCTA-500 dataset?
• RQ4 To what extent does disease status in the training data affect model generalization and performance?
• RQ5 What improvements does the IPN-V2 architecture achieve over the original IPN in multi-object retinal segmentation?

Key findings

• IPN-V2 achieves approximately a 10% improvement in mean Intersection over Union (mIoU) over the original IPN on the CAVF multi-object segmentation task.
• The dataset demonstrates strong generalization potential, with performance gains observed when training on larger subsets of the 500 subjects.
• 2D projections from 3D OCTA volumes yield superior segmentation performance compared to 3D volume inputs in the CAVF benchmark.
• Incorporating disease-affected cases into training data improves model robustness, particularly for FAZ and capillary segmentation.
• The inclusion of both OCT and OCTA modalities in the dataset enables better feature learning, especially for vessel boundary delineation.
• The dataset’s rich annotations, including 7 segmentation types and 4 clinical labels, support diverse downstream applications in retinal disease analysis.