Improved Classification of Retinal Disease: An Ensemble Deep Learning Approach for Diabetic Retinopathy, Glaucoma, and Cataracts

Diabetic retinopathy, glaucoma, and cataracts are some of the most structurally significant, complicated, and common eye diseases in ophthalmology. This makes accurate and efficient detection methods desirable. In this study, different deep learning architectures are experimented with to categorize the Kaggle Eye Diseases Classification dataset with improved classification accuracy. We have used multiple state-of-the-art (SOTA) pre-trained Convolutional Neural Network (CNN) models—InceptionV3, ResNet50, MobileNetV3Large, and EfficientNetB3—to extract relevant features utilizing Transfer Learning (TL). Using the extracted features, we have discriminated among categories of eye diseases with multiple classifiers—Logistic Regression, Linear Support Vector Machine (LSVM), Perceptron, and a voting classifier ensembling the predictions of Logistic Regression, Linear Support Vector Machine (LSVM), Perceptron, quadratically penalized SVM, and SVM with quadratically smooth loss. The maximum classification accuracy of 93.02% is achieved. This is achieved by utilizing the ensemble approach on the features obtained from fusing the extracted features using ResNet50 and MobileNetV3Large. The findings highlight that ensemble learning methods could improve the classification accuracy of retinal diseases. This shows promise for automated diagnostics in ophthalmology.