Smartphone-based decision support system for elimination of pathology-free images in diabetic retinopathy screening

Diabetic Retinopathy is a Diabetes complication and the leading cause of blindness in the United States. Early detection can be accomplished by analysis of images of the retina, generally obtained by expensive fundus cameras. Recent developments allow the use of mobile ophthalmoscopes that can be adapted to smartphones to acquire these images, but the low computational power of smartphones limits the use of Computer-Aided Diagnosis systems. In this paper, an approach for automatic retinal image analysis on a smartphone is proposed, with emphasis on high sensitivity and fast computation. A set of 1200 images from the Messidor database were analyzed for extraction of features related to vessel segmentation, presence of exudates and microaneurysms. SVM and k-NN classifier models were trained with these features, resulting in a sensitivity of 87% and a specificity of 66%. An analysis of the computational performance validates the feasibility of using this approach on quad-core smartphones.