Deepretina: Deep Learning Approach To Detect Retinal Abnormality In Computer Vision

Abstract

Currently, almost 1.2 million people in our country are blind, while around 3.51 lakh people have low vision. The pattern of eye abnormalities is changing along with an increasing rate of dry eye, cornea-related problems and eye problems related to diabetes. Early identification of eye diseases especially retinal abnormality plays a vital role to prevent the blurry vision in patients. In my research, a hybrid deep learning model is proposed to detect retinal abnormality by scanning a single retinal image of a patient. First, a new multi-label retinal disease dataset, Retinal Fundus Multi-Disease Image Dataset (RFMiD) version 02 is collected from a renewed journal website “Multidisciplinary Digital Publishing Institute” (mdpi), where 46 retinal diseases labels are available with high resolution. Next, dataset is going through analysis and preprocessing techniques to deals with data imbalance and large size (8gb) problem. Numerous analysis and experiments are performed to evaluate the models for better results. In the model, Convolutional neural models – EfficientNet, VGG16, NesNetMobile are used to analysis comparative result as well. EffectiveNet gives the highest accuracy among them and that is 85%. Voting Ensemble method is used to increase model accuracy (88%) for better prediction than could be gained from any of the constituent learning algorithms. This model is used to detect normal or abnormal retinal conditions for early treatment.