Analysis of Eye Disease classification by Fundus Images using Different Machine/Deep/Transfer Learning Techniques

The fundus image is a crucial diagnostic tool for diseases like diabetic retinopathy (DR) because it catches a retina, a back of an eye. This issue is most prevalent among diabetics and a leading contributor to blindness, particularly in young adults and those engaged in the workforce. Detecting DR in its early stages is crucial in reducing the likelihood of blindness in patients. An ophthalmologist performed DR screening and looked for lesions on the fundus imaging. Intelligent systems have the ability to make early identification of DR more efficient, especially with the recent advancements in machine learning in healthcare. This study compares and contrasts many methods for detecting DR in fundus images. This study lays out methods for improving upon pre-trained architectures such that they can correctly classify fundus images, allowing for early detection and prediction of diabetes. A concatenate model is presented in this comparative study as a method for classifying fundus images using ML and deep/transfer learning. Various model architectures are used, including SVM, Gradient Boosting, Efficient Net, Efficient-Net B4 and B7, MLP, CNN, ResNet, VGG-19, and DNN, among others. The models' accuracy, F1-score, recall, and precision are measured in experimental findings that compare them. Our strategies outperform a single model with results ranging from 84% to 96% accuracy, respectively. They found that our ML- DL-TL methods work well for automated DR classification employing fundus image data in this experiment.