Deep Learning for Diabetic Retinopathy Detection: An Elaborative Study Using Fundus Imaging and Generative Approaches

Diabetic Retinopathy (DR) remains one of the highest contributors to blindness globally, making the need for early Automated screenings a prioritised focus within the field of medical imaging. The standard diagnostic approach consists of the labour-intensive manual analysis of retinal fundus images by an ophthalmologist, a process laden with errors and variability due to the subjective nature of the assessment. The development of deep learning technologies, especially convolutional neural networks (CNNs) and attention-based methods, permits rapid and reliable analyses of retinal images, allowing for the accurate classification of images and staging of the disease. This work outlines the construction and assessment of a deep learning pipeline for the detection of DR using Kaggle’s Diabetic Retinopathy dataset. This work describes how image resizing, histogram equalisation, and contrast enhancement as preprocessing steps set the standard for work performed with leading CNNs, including Inception V3 and U-Net-based models. From the experiments, we are able to report a state-of-the-art benchmark score of 88.7% accuracy and a 0.82 quadratic weighted kappa score while surpassing foundational methods implemented within Kaggle. To increase the understanding of the model, we use LayerCAM to provide visual evidence and interpret the working of the model. This identifies and segments microaneurysms and exudates, which forms the basis of disease to improve the confidence of medical practitioners. Finally, we introduce the use of Generative Adversarial Networks (GANs) for image colourisation and augmentation to demonstrate the use of generative methods on dataset diversification and class imbalance. The developed approach that integrates generative synthesis, explainable AI, and discriminative classification provides a holistic approach to fully automate the detection of diabetic retinopathy (DR). Our research demonstrates the effectiveness of deep learning in diabetic retinopathy screening and suggests that using interpretability and augmentation techniques, deep learning will almost certainly help to diminish the burden on ophthalmologists, allowing for earlier interventions to avert vision loss.