A Fine-Tuned ResNet50 Model for Diabetic Retinopathy Detection Using Fundus Images

Diabetic Retinopathy (DR) is a progressive eye relate disorder and a leading cause of blindness worldwide. Early detection and accurate classification of DR stages is crucial for on time treatment and avoidance of vision loss. In this study a Deep Learning (DL) based approach of ResNet50 model for automated DR categorization using images of retinal fundus. Proposed ResNet50 based model on deep Convolutional Neural Network (CNN) architecture derived from Residual Network model comprises 50 layers, trained with the use of pre-trained ImageNet weights and random weights separately, to compare performance. Prior to training, extensive data preprocessing steps were applied, including circular cropping, grayscale enhancement, and Gaussian blurring to improve image quality and model focus. Experimental results demonstrate that Transfer Learning (TL) with ImageNet weights significantly enhances performance, achieving a Cohen Kappa training score of 0.991 and training accuracy of 0.982, outperforming training from scratch. Effectiveness of the model is further analyzed according to the comparing of predicted values and actual values which highlighting strong classification for DR stages and also providing common misclassifications for DR stages. This work showcases the impact of TL, preprocessing, and type of augmenting techniques in improving DR classification accuracy, making it an optimistic solution for real-world scenario.