Use of deep learning to investigate the relationship between tortuosity of the vasculature of the retina and the bulbar conjunctiva

Abstract

Accurate segmentation of conjunctival vessels is critical for the non-invasive assessment of systemic microvascular disease. While diabetes is commonly evaluated using retinal imaging, alternative accessible biomarkers remain underexplored. Conjuncti val vessels are directly visible and may reflect systemic vascular changes; however, reliable quantification is challenging, as conventional deep learning models prioritize pixel-wise accuracy without preserving vascular topology or diameter-specific structure. In this thesis, we propose a transformer-augmented U-Net architecture designed for topology-preserving, diameter-sensitive segmentation of conjunctival vasculature from standard DSLR-captured images. The model integrates dilated convolutions, attention mechanisms, and transformer-based contextual modeling to balance local feature representation with global structural awareness, enabling improved detection of small-caliber vessels while maintaining vascular continuity. A curated dataset of conjunctival images with diameter-specific annotations was developed for training and evaluation. Through systematic ablation studies, we demonstrate that no single architectural component is sufficient in isolation; rather, the full hybrid design provides the most consistent preservation of vessel geometry. Compared with multiple state-of-the-art segmentation methods (U-Net, Attention U-Net, Tran sUNet, IterNet, and U2-Net, Swin-UNet), the proposed approach achieves comparable pixel-wise performance while significantly reducing tortuosity error and centerline deviation, highlighting the importance of topology-aware segmentation for downstream analysis. Clinically, the proposed framework enables robust estimation of vessel tortuosity, revealing significantly elevated conjunctival tortuosity in individuals with diabetes compared to healthy controls, consistent with established retinal findings. These re-suits support conjunctival imaging as a low-cost, non-invasive modality for microvas cular phenotyping and suggest its potential for scalable screening and longitudinal monitoring. Despite limitations related to dataset size, patch-based training, and cross-domain generalization, this work establishes that topology-preserving, diameter-aware segmentation is essential for clinically meaningful vascular analysis. Future work will focus on expanding multi-institutional datasets, improving spatial calibration, and validating the framework across diverse populations and vascular pathologies.