Retinal fundus image enhancement using an ensemble framework for accurate glaucoma detection

Retinal fundus imaging plays a crucial role in the diagnosis of ophthalmic diseases such as glaucoma, a significant cause of vision loss worldwide. Accurate detection of glaucoma using image processing, machine learning, and deep learning approaches depends on the effectiveness with which the retinal fundus images are captured. Poor-quality images with artifacts, including uneven illumination, blur, and color distortion, can lead to incorrect diagnoses. In this work, we propose an end-to-end glaucoma detection model based on the ensemble of image enhancement networks, segmentation networks, and image classification networks. The proposed approach consists of an improved version of generative adversarial network (GAN) called the cycle consistency GAN (cycle-GAN) for image quality enhancement, U-Net for optic cup and optic disc segmentation, and support vector machine for image classification. The cycle-GAN model uses autoencoders as generators and a deep convolutional neural network (CNN) as discriminators to generate high-quality fundus images. The cup-to-disc ratio, a popular feature, is utilized to categorize fundus images as either glaucomatous or non-glaucomatous. We use six imbalanced datasets for experimental analysis of the proposed ensemble model, including ORIGA, ACRIMA, DRISTI-GS, REFUGE, Messidor, and Mendeley. The experimental findings demonstrate that the proposed ensemble model works better than individual models such as GAN, Autoencoder, deep CNN, and also from existing methods. The proposed method not only reduces the artifacts from fundus images but also solves the problem of imbalanced datasets for accurate glaucoma detection. The experimental results show maximum accuracy, precision, recall, and F-measure values of 0.968, 0.821, 0.974, and 0.891, respectively.