Quantification and classification of retinal vessel tortuosity

The clinical recognition of abnormal retinal tortuosity enables the diagnosis of many diseases. Tortuosity is often interpreted as points of high curvature of the blood vessel along certain segments. Quantitative measures proposed so far depend on or are functions of the curvature of the vessel axis. In this paper, we propose a parallel algorithm to quantify retinal vessel tortuosity using a robust metric based on the curvature calculated from an improved chain code algorithm. We suggest that the tortuosity evaluation depends not only on the accuracy of curvature determination, but primarily on the precise determination of the region of support. The region of support, and hence the corresponding scale, was optimally selected from a quantitative experiment where it was varied from a vessel contour of two to ten pixels, before computing the curvature for each proposed metric. Scale factor optimization was based on the classification accuracy of the classifiers used, which was calculated by comparing the estimated results with ground truths from expert ophthalmologists for the integrated proposed index. We demonstrate the authenticity of the proposed metric as an indicator of changes in morphology using both simulated curves and actual vessels. The performance of each classifier is evaluated based on sensitivity, specificity, accuracy, positive predictive value, negative predictive value, and positive likelihood ratio. Our method is effective at evaluating the range of clinically relevant patterns of abnormality such as those in retinopathy of prematurity. While all the proposed metrics are sensitive to curved or kinked vessels, the integrated proposed index achieves the best sensitivity and classification rate of 97.8% and 93.6%, respectively, on 45 infant retinal images.