Analysis of Activation Functions and Hyper-Parameters in Transfer Learning Models for Classifying Diabetic Retinopathy

Abstract

A comparative study of transfer learning models with hyperparameter optimization for classifying diabetic retinopathy has been derived in the present work. This thesis focuses on the classification of diabetic retinopathy using transfer learning models with an optimization of hyperparameters. Chapter 1, deals with providing information about diabetic retinopathy, which is diabetesrelated retinal disorder that ranks as one of the most common causes of blindness. We elaborate on the importance of early diagnosis and possible difficulties with traditional approaches to screening. The possibility of deep learning AI technology and its application in the identification of diabetic retinopathy is discussed. It is crucial to focus on the design of accurate and pragmatic automated detection systems to support the healthcare staff in early identification and early action. The chapter also describes the purpose of this research and the role this research can play in enhancing the quality of patient care and easing the burden of healthcare systems. The current understanding and gaps of literature in existing approaches are discussed in Chapter 2, which is a literature review. As mentioned in Chapter 3, the transfer learning-based solution applies three types of pre-trained CNN and examines different activation functions, optimizers as well as padding modes in a manner of systematic analysis. Hyperparameters tuning is used in the optimization of each of the developed diabetic retinopathy detection model. Chapter 4 covers the results of the experiments carried out, where the various evaluation measures including accuracy, precision, recall, F1 score, specificity, and AUC-ROC are used to compare the models. ROC curves and plots of accuracy vs the epoch and loss vs the epoch is used to understand the performance of the model and characteristics of training. The following are the results that we have got that supports our findings. This work will supplement the existing scholarship on automating diabetic retinopathy screening and could positively impact early detection and intervention