Classification of Eye Diseases using Hybrid CNN-RNN Models

Early diagnosis of eye diseases is essential to prevent irreversible vision loss. The traditional method used by ophthalmologists is to manually screen, images captured of back of an eye called as fundus images. Patient care is being harmed by an increase in the number of patients and a scarcity of qualified ophthalmologists. This study focuses on classifying fundus images into different types of eye illnesses: cataracts, glaucoma, and retinal diseases. The Convolution Neural Network-Recurrent Neural Networks (CNN-RNN) model has proven to be effective in a variety of disease classification challenges. The advantages of Transfer Learning and Long Short-Term Memory (LSTM) are combined in the proposed hybrid CNN-RNN models. Multiple different features are extracted from fundus images by Transfer Learning models, InceptionV3, InceptionResNetV2, and DenseNet169. The extracted features are classified using LSTM. The research is based on the Kaggle dataset, which comprises an imbalance number of images in each category. Therefore, augmentation approaches are used to balance the dataset. The performance of the models improved when trained and tested on increased and balanced data. The hybrid DenseNet169-LSTM model achieved the highest accuracy of 69.50% (87.40% specificity and 69.50% sensitivity)