Diabetic Retinopathy Classification Using Hybrid Optimized U-Net and Improved ResNet-18 with MISH Activation

Diabetic Retinopathy (DR) refers to damage to the retina caused by diabetes, which can cause visual impairments or potentially lead to blindness. The process of manually identifying diabetic retinopathy is slow and can easily be affected by human mistakes because of the eye's complex anatomy. This paper aims to determine the optimal model for accurately staging diabetic retinopathy (DR) across five DR categories. The proposed system involves image pre-processing, segmentation and classification. A well-structured preprocessing framework was implemented, integrating Median Filtering technique to decrease noise and Gamma Correction for improved image quality. Data augmentation is performed through methods like flipping, cropping rotating, and translation of fundus images. In this study, a hybrid optimization strategy is proposed by combining the Adam optimizer with Simulated Annealing (SA) and Cosine Annealing for training a U-Net model for diabetic retinopathy (DR) lesion segmentation. Canny Edge Detection is employed to discover edges correctly within the image. For DR categorisation in this work, enhanced Resnet 18 combined with MISH activation function is used. The suggested model makes use of residual blocks in conjunction with identity and convolutional blocks. Experiments on the APTOS dataset show that the proposed model outperforms state-of-the-art techniques. Accuracy, precision, recall, and F1 score are all improved by the suggested Resnet18 model, which receives scores of 99.45%, 98.64%, 98.63%, and 98.63%, respectively.