TL;DR
This paper introduces OCTA-500, the largest publicly available dataset for Optical Coherence Tomography Angiography (OCTA) with 500 subjects and rich annotations. They also propose a novel multi-object segmentation task called CAVF (Capillary, Artery, Vein, FAZ) and introduce an improved 3D-to-2D segmentation network, IPN-V2, to tackle this challenge.
What Problem Does This Paper Address?
Optical Coherence Tomography Angiography (OCTA) is a powerful imaging technique for visualizing retinal vasculature, crucial for diagnosing and understanding various eye diseases. However, the field has been hampered by a lack of large, diverse, and well-annotated datasets. Existing datasets are often small, lack multi-modality data (OCT and OCTA volumes), and are limited in disease diversity. This scarcity makes it difficult to train robust deep learning models for comprehensive OCTA analysis, particularly for tasks like multi-object segmentation of different vascular components and the foveal avascular zone (FAZ).
Methods
The authors introduce OCTA-500, a comprehensive dataset featuring:
- 500 subjects with both OCT and OCTA volumes.
- Two fields of view (FOVs): 6mm x 6mm and 3mm x 3mm.
- Rich 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).
They propose the CAVF (Capillary, Artery, Vein, FAZ) task, a novel multi-object segmentation framework that unifies the segmentation of these critical vascular components. This unified approach aims to simplify quantitative analysis and model evaluation.
To address the 3D nature of OCTA data, they introduce IPN-V2, an optimized version of their previous Image Projection Network (IPN). IPN-V2 features a 3D-to-2D projection stage with a fast projection module (FPM) to efficiently compress 3D information into 2D features, followed by a 2D segmentation stage. This architecture aims to reduce computational cost and improve segmentation performance by leveraging both 3D context and high-level semantic information.
Datasets
The primary contribution is the introduction of the OCTA-500 dataset. This dataset is significantly larger and more diverse than existing OCTA datasets, offering:
- 500 subjects with both OCT and OCTA volumes.
- Two FOVs (6mm and 3mm).
- Multiple modalities (OCT and OCTA volumes).
- Extensive annotations: including large vessel, capillary, artery, vein, 2D FAZ, 3D FAZ, and retinal layers.
- Disease diversity: encompassing common retinal diseases like AMD, DR, CNV, CSC, RVO, and others.
The authors also mention using data from existing datasets like ROSE-1 and Giarratano for capillary annotation training.
Key Insights and Ideas
- The critical need for large, diverse datasets: The paper strongly emphasizes that the scarcity of such datasets is a major bottleneck for advancing OCTA analysis, particularly for deep learning models. OCTA-500 aims to fill this gap.
- Unified multi-object segmentation is beneficial: The CAVF task proposes that segmenting capillaries, arteries, veins, and FAZ together offers advantages over single-task segmentation, simplifying quantitative analysis and potentially improving overall performance.
- 3D context is crucial for robust segmentation: The development of IPN-V2 highlights the importance of leveraging 3D OCTA volumes, especially for tasks like capillary segmentation where noise in 2D projections can lead to misclassifications.
- Disease diversity significantly impacts segmentation performance: The experiments show that segmentation models perform less reliably on diseased retinas compared to healthy ones, underscoring the need for datasets with varied disease pathologies.
Key Results
- OCTA-500 is the largest OCTA dataset: It significantly surpasses previous datasets in terms of subject count and annotation richness.
- IPN-V2 outperforms IPN: The optimized IPN-V2 achieves an approximate 10% mIoU improvement over the original IPN on the CAVF task and significantly reduces GPU memory usage, allowing for larger input sizes.
- Impact of training data size: mIoU increases with more training data, with a rapid increase up to 120 samples, after which improvements become more gradual.
- Input modality matters: OCTA volumes provide the most dominant information for segmentation, outperforming OCT volumes and projection maps for certain tasks.
- Baseline performance: U-Net and its variants generally perform well on the CAVF task, with IPN-V2 showing superior performance, especially in capillary segmentation.
- Disease impact: Segmentation performance is notably lower for diseased retinas (AMD, DR, RVO) compared to normal retinas, with mIoU rankings generally following normal > CSC > CNV > AMD > DR > RVO.
Limitations and Future Work
- Computational cost of 3D models: While IPN-V2 improves efficiency, 3D-to-2D networks still require more computational resources than 2D baselines.
- Segmentation challenges in diseased retinas: The paper acknowledges that diseases like AMD, DR, and RVO, characterized by non-perfusion and morphological changes, pose significant challenges for segmentation accuracy.
- Artery-vein segmentation consistency: Differentiating arteries and veins, especially for smaller branches, remains a challenge, with IPN-V2 showing improvements but not complete resolution of inconsistencies.
- Future work: The authors plan to continue collecting images, annotating ground truths, and optimizing methods for more OCTA studies. They also suggest further research into improving segmentation performance under disease conditions and exploring the potential of OCTA-500 for other OCTA studies, such as modality transformation and artifact removal.
Summary updated