The development of new vessels on the retina of people with diabetes is rare, but is likely to lead to severe visual impairment. This paper investigates the selection of suitable image features for the automatic detection of new vessels on the optic disc. The features are chosen based on their discrimination capability (tested using the non-parametric Wilcoxon rank sum and Ansari-Bradley dispersion tests) and absence of correlation with other features (tested using the Kendall Tau coefficient). Classification was performed using a support vector machine. The system was trained and tested by cross-validation using 38 images with new vessels and 71 normal images without new vessels. Fourteen features were selected, giving an area under the receiver operator characteristic curve of 0.911 for detecting images with new vessels on the disc. The method could have a useful role as part of an automated retinopathy analysis system.
