Integrated Computational Retinal Vascular Morphometry Using Fractal, Skeleton-Based, and Phi-Related Features for Diabetic Retinopathy and Glaucoma.
Keywords:
Retinal Fundus Imaging; Fractal Dimension; Vascular Skeleton; Diabetic Retinopathy; Glaucoma; Golden Ratio; Morphometry.Abstract
Background: Fundus images can be used to quantify retinal vascular morphology using density, fractal and graph-based metrics. Exploratory work with phi-related proportional indices is most commonly incomplete or there are no real-world reports.
Objective: To assess whether the global vascular morphology, fractal complexity, skeleton topology, and phi-derived proportional features of healthy, diabetic retinopathy, and glaucoma eyes are different in the HRF dataset.
Methods: Out of 45 HRF fundus images analyzed, 15 each were healthy, diabetic retinopathy, and glaucoma with expert matched vessel and field-of-view masks. Global features included vessel density, edge density, fractal dimension, lacunarity, ratios of regional densities, and phi-deviation indices. The skeletons in downsampled vessels utilized at branch level so that the subsequent features are extracted from the skeleton length of the vessel, endpoint of the vessel, cluster of the branches, ratio of the segments, geo-metric tortuosity proxies, branch phi-deviation features. Differences in group characteristics using permutation ANOVA with 5,000 permutations were investigated. Moreover, the use of a leave-one-out nearest-centroid model was explored.
Results: Healthy retinas showed higher vessel density and fractal dimension, whereas diabetic retinopathy and glaucoma showed higher lacunarity. Skeleton analysis showed higher skeleton length and branch-cluster density in healthy eyes, while disease groups showed much higher endpoint density and endpoint/branch-cluster ratios. Integrated global morphology plus skeleton features achieved 84.4% exploratory classification accuracy.
Conclusion: Morphometric differences linked to diseases were captured using fractal and skeleton techniques in the HRF dataset. Indices related to phi must be considered exploratory descriptors, rather than bona fide biomarkers.
