Generation of Synthetic data using GANs with CLAHE processed images and comparative study of model performance on real and synthetic Dataset

The research involves generation of synthetic data images used for the prediction of disease such as glaucoma, age-related macular degeneration (ARMD), cataract, diabetic retinopathy and healthy eyes from fundus images. The research is focused on implementation of GAN model along with literature survey for the project". It starts with the careful gathering of different image datasets corresponding to these different eye diseases from various sources like HRF, FIVES, eyePACS, Papila etc. These images are resized which is needed for Generative Adversarial Networks (GANs) to fine-tune the aspect where generation of the best image is needed for validation. Contrast Limited Adaptive Histogram Equalization (CLAHE) is used to make these images clearer to improve model learnability for the dataset. A VAE – Variational auto encoders is used to fetch the latent vector from a input image and then train a GAN model usually a Deep Convolutional GAN (DCGAN) that has high-level optimization and training strategies (or) to create significant synthetic and eye fundus images. This synthetic data is important for the next step, which is retraining CNN models. Three different CNN architectures (various networks like resnet or mobilenet) are employed, one trained separately on real images, one separately on synthetic images, and one utilizing a mixed dataset of both real and synthetic images. All the models are evaluated robustly and their accuracies are calculated to know the efficiency of the models for disease prediction. The findings will add to the body of work on the use of GAN of data to improve deep learning models for the detection and diagnosis of eye diseases in medical imaging.