A modern screening approach for detection of diabetic retinopathy

Diabetic retinopathy (DR) is a micro-vascular impediment of the diabetes which causes deformities in the retina. DR is the main source for the loss of vision and blindness. For the effective treatment of DR, early diagnosis of the disease is very important. The existing teleophthalmology screening models send all captured retinal images to the hospital via VSAT for evaluation by the expert ophthalmologists. These systems are very costlier and cause unnecessary data traffic in the internet as well as ophthalmologists has to evaluate all received images. We have proposed an automated fundus image analysis system for early stage detection of diabetic retinopathy to modify the conventional teleophthalmology. This modern teleophthalmology system captures retinal fundus images of patients by handheld fundus camera at the screening camp site. The captured images are accurately classified as normal or with DR using image processing techniques. From the camp site, only DR affected images will be sent to expert ophthalmologist through internet. The potential locations of different visual abnormalities associated with DR are highlighted on images. An automated prescreening system that determines whether or not any suspicious signs of DR are present in an image significantly reduces the workload of experts. The proposed system implements two stage classification, firstly at the screening camp it classifies images into DR and non DR, secondly to identify the potential lesions related to DR in the images which are sent to base hospital for expert review. In this paper we implemented holoentropy enabled decision tree classifier for the classification purpose. For both the stages the efficient features are extracted to form the feature vector. The experimental evaluation is performed on the publically available database DIARETDB1. The performance of the proposed teleophthalmology is evaluated using sensitivity, specificity and accuracy.