Transformer-Based Diameter-Specific Segmentation of Conjunctival Vessels for Early Detection of Diabetic Vascular Changes

Early detection of systemic vascular disorders such as diabetes and arterial tortuosity syndrome relies on identifying changes in vessel morphology, including tortuosity and diameter. While retinal imaging is a clinical standard for microvascular assessment, its high cost and requirement for specialist expertise limit its accessibility for large-scale screening. The bulbar conjunctiva, by contrast, offers a non-invasive, cost-effective, and readily accessible alternative for vascular imaging. In this study, we propose a deep learning-based framework designed to extract diameter-specific conjunctival vessels using a hybrid architecture that combines dilated convolutions with transformer-based attention mechanisms. The model incorporates Multi-Head Self-Attention at the bottleneck to model long-range dependencies within the image, and Multi-Head Cross-Attention at each decoder level to enhance selective reconstruction of diameter-relevant features. Our architecture is evaluated on a custom dataset comprising healthy individuals and patients with diabetes with varying complications. Compared with state-of-the-art baselines, including IterNet and U$^2$-Net, the proposed method achieved superior segmentation performance, especially in isolating vessels within clinically relevant diameter ranges, enabling accurate quantification of vessel tortuosity. We demonstrate statistically significant differences in tortuosity across diabetic subgroups, demonstrating the translational potential of conjunctival imaging as a scalable tool for early screening and monitoring of diabetic and systemic vascular complications.