Abstract
Systemic diseases such as diabetes, hypertension, and atherosclerosis are among the leading causes of annual human mortality. Retinal and conjunctival vascular tortuosity are potential biomarkers for these conditions. Crucially, tortuosity depends on vessel thickness, necessitating selective calculation based on the specific disease under analysis. This paper proposes a thickness-sensitive vessel extraction framework applicable to retinal and conjunctival studies. The framework utilizes a Convolutional Neural Network (CNN) based on the IterNet architecture to generate vascular probability maps, followed by a multi-scale vessel enhancement technique to extract vessels of specified thicknesses. Evaluated on four datasets (including DRIVE and SBVPI), the framework achieved Matthew's Correlation Coefficient (MCC) values exceeding 0.71. Clinical application showed that diabetic patients had significantly higher retinal vascular tortuosity but significantly lower conjunctival vascular tortuosity (p < .05) compared to non-diabetics, aligning with existing literature and validating the framework's utility.
Introduction
The microvasculature of the eye provides a unique window into systemic health. Changes in vascular tortuosity—the degree of "twistiness" in blood vessels—often precede more severe clinical symptoms of diabetes and hypertension. However, accurately measuring these changes is challenging because they are often scale-dependent; certain diseases affect thicker vessels differently than thinner ones.
Most existing segmentation tools lack the precision to isolate vessels by width. Additionally, while retinal fundus imaging is the gold standard, its high cost limits accessibility. This research addresses these gaps by proposing a framework that handles both retinal and conjunctival images, the latter of which can be captured using standard DSLR cameras with macro lenses.
Methods
The proposed framework follows a multi-stage pipeline designed to maximize both structural accuracy and scale selectivity. For external eye images, a specialized U-Net (ScleraUNet) first segments the sclera to eliminate noise from surrounding tissues like eyelids. The core of the vessel detection relies on IterNet, a cascaded CNN architecture that iteratively refines probability maps to capture the full vascular tree, including fine details that single-pass networks might miss.
To achieve thickness sensitivity, the framework applies a multi-scale Frangi filter to the IterNet output. By combining structural maps (from the CNN) with redness maps (from the original green channel), the system can selectively extract vessels of specific pixel widths. Finally, these vessels are skeletonized to calculate mathematical tortuosity indices, such as the Eccentricity-based Tortuosity Index.
Results and Discussion
The framework's performance was validated across four datasets: DRIVE, SBVPI, and the authors' REIDA (Retinal and Eye Images for Diabetic Analysis) dataset. The system achieved reliable segmentation with MCC values above 0.71 across all tests, indicating high reliability despite the inherent class imbalance in vessel segmentation.
A key clinical finding was the divergent behavior of the retina and conjunctiva in diabetic patients. While retinal vessels showed significantly increased tortuosity, conjunctival vessels in the same group appeared straighter (lower tortuosity) than those in healthy controls. This suggests that while both are biomarkers for diabetes, they reflect different physiological responses. The study also confirmed that combining CNN-derived structural data with raw image contrast (redness maps) yields superior results compared to using either alone.
Conclusion
This thickness-sensitive framework provides a robust tool for microvascular analysis. By enabling precise vessel extraction across different scales and modalities, it facilitates the use of accessible conjunctival imaging for systemic disease screening. Future work will focus on improving the validation of extremely thin vessels and developing specialized indices tailored to the unique geometry of conjunctival vasculature, as current indices may not fully capture the structural differences between retinal and external eye vessels.
Summary updated