OCTA Vessel Segmentation via UNet Transformers and Deformable Receptive Fields: A Motion-Artefact-Resilient Deep Learning Framework.
Authors
Affiliations (8)
Affiliations (8)
- Department of Data Science and Digital Innovation, The Hong Kong Shue Yan University, Hong Kong SAR, China. [email protected].
- Department of Biomedical Engineering, The Chinese University of Hong Kong, Hong Kong SAR, China.
- Department of Data Science and Digital Innovation, The Hong Kong Shue Yan University, Hong Kong SAR, China.
- Division of Science, Engineering, and Health Studies (SEHS), School of Professional Education and Executive Development, The Hong Kong Polytechnic University, Hong Kong SAR, China.
- English Schools Foundation College, Sha Tin, Hong Kong SAR, China.
- Department of Agriculture, Bahauddin Zakariya University, Multan, Pakistan.
- Academy of Wellness and Human Development, Hong Kong Baptist University, Hong Kong SAR, China.
- Department of Biomedical Engineering, The Chinese University of Hong Kong, Hong Kong SAR, China. [email protected].
Abstract
Optical coherence tomography angiography (OCTA) is a non-invasive medical imaging technique that helps doctors see the blood vessels of the retina and choroid in fine detail. However, OCTA images often suffer from motion artefacts, which lower the accuracy of vessel segmentation in the retina. In this paper, we propose V2Vnet, a robust volume-to-volume deep learning framework that combines a UNet transformer with deformable receptive fields (DRF) to overcome this problem. The transformer encoder captures the overall structure of the input volume and gathers multi-scale information about retinal blood vessels, while the DRF block makes the convolutional sampling grid more flexible by adding learnable two-dimensional offsets to the standard grid. Together, these components allow the network to detect and segment blood vessels accurately in motion-insensitive 3D OCT images from a single scan, removing the need for repeated acquisitions. We trained and evaluated V2Vnet on the public OCTA-500 dataset, and we further tested it on an independent retinal-layer parcellation dataset to check generalisation. Our proposed technique achieved an accuracy of <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>99.56</mn> <mo>%</mo></mrow> </math> , an mIoU of <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>69.64</mn> <mo>%</mo></mrow> </math> , an mDice of <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>79.14</mn> <mo>%</mo></mrow> </math> , and an MSE of 0.0151 on 3D OCTA segmentation, outperforming UNet3D, UNETR, and IPNV2 across all metrics. Even when synthetic motion artefacts were introduced, V2Vnet still delivered the best mIoU and mDice scores. These results show that our framework is reliable, cost-effective, and well suited to clinical use, especially in low- and middle-income settings where multi-scan OCTA hardware may not be available.