Classification of Fundus Images Using Custom CNN for Diabetic Retinopathy

Globally, retinal diseases represent a serious threat to vision, emphasizing the urgent need for efficient detection and intervention techniques. We present a novel deep learning classification system in this paper for the diagnosis of retinal diseases. Our approach uses sophisticated image processing techniques along with Convolutional neural networks (CNNs) to automate and improve the diagnosis accuracy of retinal diseases. We address the drawbacks of conventional diagnostic techniques and emphasize how deep learning has the potential to transform the detection of retinal diseases. In this paper, we have proposed a 5-layer CNN architecture CNNs which uses convolutional layers for feature extraction and classification using fully connected layers with softmax activation function. We validate the effectiveness of our approach in reliably detecting retinal abnormalities by comprehensive experiments on large-scale retinal imaging datasets of fundus images with Diabetic Retinopathy (DR) which is the main cause of vision impairment and blindness. Making use of custom Convolutional neural network architecture, it is demonstrated that an accuracy of 93.72% is achieved on APTOS fundus images dataset which shows significant improvement in classification accuracy over state-of-the art detection techniques.