Diabetic Retinopathy Classification Using PSOSVM Based Deep Learning Model

Diabetic retinopathy (DR) is caused by diabetes. DR is required to be identified in its earlier stage otherwise long-term uncontrolled DR can result in complete blindness of the eyes. DR diagnosis can be carried out manually using retinal fundus images of human eye. However, automated methods are more cost-effective and reliable. This paper presents a deep learning methodology for DR image classification by classifying retinal fundus images in two separate classes named as ‘No DR’ and ‘DR’. The pre-trained VGG-19 model is utilized to extract the data’s deep features followed by features optimization using particle swarm optimization (PSO) technique. Support vector machine (SVM) classifier, which functions as a binary classifier for the classification, is then used to classify the selected features. APTOS-2019 binary dataset consists of total 5590 retinal fundus images out of which 3662 are used for training and 1928 are used for testing the proposed model and five other existing models. The classification performance of the suggested model, as well as five existing methods, was assessed through extensive simulation. The proposed model achieved an accuracy of 93%, sensitivity of 99.21%, specificity of 98.07%, and F1-score value of 94.32% which is the best among the models used for comparison. Consequently, the presented method exhibits great potential for effectively classifying retinal fundus images.