Optideneffnet: Deep Learning Framework For Retinal Diseases Prediction

Abstract

The eye is the primary sensory organ involved in vision. Since some eye conditions might result in blindness, it is important to diagnose and treat them as soon as possible. In ophthalmic diagnostics, it is essential to identify various ocular illnesses from fundus images. On the other hand, subjectivity and mistake arise when ophthalmologists manually detect images. As a result, several automation techniques are developed, however they need to improve in terms of classification accuracy. This work proposes a novel multi-label automated Deep Learning (DL) prediction system called OptiDenEffNet, which can identify various retinal illnesses, as a solution for such challenges. The study begins with an extensive pre-processing step includes label encoding, data cleaning and normalization. The proposed OptiDenEffNet model, which combines many DL blocks, and the Attention Block, which act as feature descriptors, are then used to extract discriminative deep feature representations. Finally, a Discriminative Gated Recurrent Unit (DGRU), a highly developed prediction model is used. This DGRU can provide a probability distribution that precisely predict eye disorders on particular age by incorporating a Softmax layer to its architecture. Extensive tests are conducted on the challenging Ocular Disease Recognition Dataset using Python platform. The proposed model achieves the maximum accuracy of 98.269% and 99.691% in the 70:30 and 80:20 data splits respectively. The results specify that the proposed OptiDenEffNet model works better than state-of-the-art for predict retinal diseases.