IMPROVING RETINAL DISEASE DETECTION USING AN ENSEMBLE OF CNN-BASED DEEP LEARNING MODELS

Abstract

Diabetic Retinopathy is a severe form of retinal diseases that cause vision impairment in most of the world. Timely diagnosis is the key to success in treatment and prevention of blindness. Nevertheless, the diagnosis of retinal fundus images manually involves the use of professional ophthalmologists and does not allow large-scale and timely screening, particularly in resource-limited settings. The present study presents an automated retinal disease detection model using DL methods. The model proposed combines three pre-trained CNN architectures and fine-tuning on retinal fundus image datasets with transfer learning. Resizing, normalization, and augmentation are employed to preprocess the input data to enhance the generalization and minimise overfitting. Focal loss and class weighting methods are used in the training to overcome the issue of imbalance class. The models are trained separately and their performance. The last prediction is derived under a soft voting ensemble strategy which involves the output of the individual models. The experimental results indicate that the proposed model is better than individual networks with an accuracy of 97.54 and AUC of 0.98 in the test data. The results indicate the usefulness of ensemble DL techniques to enhance the diagnostic results and minimize erroneous classification. The suggested system will be able to help medical staff with early screening and diagnosis, which will eventually lead to better patient outcomes and decrease the risk of vision loss.