Diabetic Retinopathy detection using geometrical techniques related to the underlying physiology

摘要

A computer algorithm is developed using a set of geometrical measures of Diabetic Retinopathy on fundus images. The algorithm is extended from previous work with a focus on accurately ruling out false exudates that are calculated from an initial screening. Sensitivities and specificities are used to measure the effectiveness of maintaining accurate detection of exudates while avoiding false positives with respect to an initial set of possible exudates. A very high specificity of 98.4% and good sensitivity of 88.6% on 12 images shows the potential of the method. These results justify future clinical trials involving Ophthalmologists. An eye tracking procedure that could characterize what regions of the retina that experienced Ophthalmologists focus on is discussed.