New public retinal image database for tortuosity estimation

Purpose : Tortuosity in retinal images can be used as a biomarker in the detection of several systemic diseases, including diabetes and hypertension. This work provides a new retinal image database to test and compare tortuosity metrics at both the vessel and image level, as well as a comparison of several of the popular methods for tortuosity estimation on the dataset.Methods : One macula-centered image was acquired for each eye in 37 patients (74 images) at the University of Edinburgh using a Canon non-mydriatic camera at 45° field of view. A total of 100 arteries and 100 veins were chosen and graded from the images. The tortuosity of these vessel segments was graded as either absent, low or high by two clinical specialists. Image-level tortuosity was also graded on the same scale by a total of 5 specialists. The database consists of the retinal images, vessel centerline points from chosen vessels (used to reproduce the vessel path), and the ground truth. Six previously developed tortuosity metrics were tested against a representative subset of the database (25 arteries and 25 veins). These metrics are the arch /chord ratio (DM), tortuosity density (TD), slope chain coding (SCC), and two integral curvature measures (Tau3, 5). Descriptions of algorithms have been reported previously (Lisowska et al., EMBC 2014). Full dataset and results will be made available at http://bioimlab.dei.unipd.it/Data%20Sets.htmResults : Intergrader variability was calculated for the subset of vessels. Cohen’s kappa for agreement was .73 for veins and .61 for arteries. Tortuosity metrics were calculated and agreement between metrics and the two graders can be seen in Table 1. Results show that no one metric had the highest agreement simultaneously across veins, arteries, and graders. The tortuosity density metric had the highest average agreement across all categories.Conclusions : This work provides a new public database for tortuosity estimation including images, vessel segments, and ground truth. Results on a subset of vessels suggests that a single tortuosity metric has difficulty capturing the qualitative grading of clinicians