Explainable AI Based Diabetic Retinopathy Detection Using Transfer Learning and Grad CAM

Abstract

- Diabetic retinopathy is a diabetes related retinal disorder that can lead to progressive vision impairment and blindness when it is not detected and managed at an early stage. Automated analysis of retinal fundus images using deep learning has shown considerable potential for assisting diabetic retinopathy screening and severity classification. However, the complex decision making process of deep learning models can make their predictions difficult to interpret, which limits transparency in medical applications. This study investigates an explainable artificial intelligence based approach for diabetic retinopathy detection using transfer learning and Grad CAM. Pretrained convolutional neural network architectures, including ResNet50 and EfficientNet, are considered for classifying retinal fundus images into different stages of diabetic retinopathy. Image preprocessing and augmentation techniques are applied to improve the suitability of retinal images for model training, followed by model fine tuning on the selected dataset. The models are evaluated using accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve. Grad CAM is integrated to generate visual explanations that highlight the retinal regions contributing to the model predictions. The study evaluates both predictive performance and visual interpretability, providing a comparative analysis of transfer learning models and an understanding of the visual evidence associated with their classifications. The experimental results will be used to determine the effectiveness of the proposed approach and its potential for developing more transparent computer aided diabetic retinopathy screening systems.