Detecting Eye Diseases and Intraocular Lesions from Fundus Images Using Deep Learning Approaches

Abstract

In this study, the focus begins with addressing the critical issue of diabetic retinopathy (DR) detection, a leading cause of blindness globally, by using a combination of SHapley Additive exPlanations (SHAP) analysis and transfer learning ResNet50 model. Achieving impressive accuracy rates of 97% for binary and 81% for multi-class DR classification, the study demonstrates the potential of SHAP analysis to enhance interpretability and contextual understanding of prediction outcomes. Shifting the study to uveal melanoma (UM), an intraocular cancer with significant risks, the research used similar methodologies to predict UM, achieving a high binary classification accuracy of 82.5% in InceptionV3 model. The application of SHAP analysis once again highlights its value in shedding light on prediction rationales and improving result comprehension. The study further extends into the use of four distinct convolutional neural network (CNN)-based architectures for UM detection, emphasizing the manual collection and preprocessing of 854 RGB fundus images. Through transfer learning, DenseNet169 appears as the most accurate model, achieving 89% accuracy in binary classification of choroidal nevus (CN). Essentially, SHAP analysis continues to play an essential role in enhancing interpretability, offering detailed insights into the significant image regions influencing CN predictions. In conclusion, this study emphasises the power of combining deep transfer learning CNN-based models, and SHAP analysis to not only achieve robust predictive performance but also to address the critical challenge of interpretability in deep learning models, contributing significantly to the fields of medical image analysis and diagnostic decision-making.