Deep retino-network for automatic quantification of diabetic retinopathy

Abstract

Diabetic retinopathy (DR) is the ocular manifestation of the systemic disease. Since it is the most prevalent cause of blindness in the world, it demands a significant amount of therapeutic attention. As a result, a precise assessment of the DR condition as well as its evolution is very important for treatment. In this work, an automated quantification of diabetic retinopathy state (AQDRS) using fundus images is proposed. The state of DR is classified into 0 (low) to 3 (high) with the help of a deep retino-network (DRN). Before the classification by DRN, an image down-sampling scheme is employed. A DRN consists of convolution layer and max-pooling layers to extract the deep retina features and fully connected layer (FCL) for AQDRS where feed-forward neural network is employed for the classification. The performance of AQDRS by DRN for grading DR is evaluated using methods to evaluate segmentation and indexing techniques in the field of retinal ophthalmology (MESSIDOR) database. Results show that the AQDRS by DRN can able to extract the relevant discriminative information for grading the fundus image. The average accuracy on normal images in MESSIDOR database is 97.9% and it is 95.3% for DR images.