Work

Data-Efficient Transformer Model for Retinal Artery-Vein Segmentation in Landmark-Deficient OCT Angiography

Year: 2026

Type: article

Abstract: Objective Accurately differentiating arteries from veins in highly magnified foveal 3 × 3-mm 2 macular optical coherence tomography angiography (OCTA) is critical for quantifying retinal vascular disease. However, existing models require macrovascular landmarks absent at this field of view, hindering vessel-specific analysis in clinical imaging. We developed and validated a data-efficient, two-stage transformer for AV segmentation in landmark-deficient macular OCTA and evaluated whether the resulting vessel maps could identify retinal vein occlusion (RVO) biomarkers without disease-specific retraining. Design Cross-sectional model development and validation study. Participants The development cohort comprised 17 eyes (29 scans) from 9 healthy adult volunteers (mean age, 32 years). A subsequent RVO pilot cohort included 9 foveal OCTA scans from patients with confirmed RVO, compared against 15 healthy control scans. Methods Independent two-stage Swin-Unet transformer trained on sparse annotations from 12 eyes, applied without retraining to five test eyes and the RVO pilot cohort. Main Outcome Measures Artery-vein segmentation Dice coefficients, area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. Clinical RVO endpoints included the major vein-to-artery sinuosity ratio and the 10th-percentile artery radius. Results The framework demonstrated exceptional internal validation and crucially, when applied to the external RVO cohort, the model successfully identified established pathophysiological hallmarks. Stage 1 global segmentation achieved AUC 0.973 and Dice 88.6%. Stage 2 AV classification achieved Dice 96.3% (arteries) and 96.1% (veins). End-to-end system Dice coefficients were 87.3% and 84.1% for arteries and veins, respectively. Applied to RVO without retraining, the model identified significantly elevated major vein/artery sinuosity (1.041 vs 1.010; P = .007; AUC, 0.837) and reduced 10th-percentile artery radius (13.2 vs 15.4 μm; P = .002; AUC, 0.885). Conclusions This data-efficient framework enables high fidelity AV segmentation in landmark deficient OCTA. By combining conventional methods with newer transformer models and decoupling structural segmentation from classification, the model identified biologically consistent biomarkers in RVO from small sets of scans. This supports highly scalable automated vessel specific quantification in clinical macular imaging; these pilot findings are hypothesis-generating and warrant prospective validation.

Source: Ophthalmology Science

DOI: 10.1016/j.xops.2026.101416