Evaluating Performance of Pre-trained Models for Diabetic Retinopathy Detection with a Minimal Dataset using Transfer Learning

Objectives: Evaluate the effectiveness of transfer learning with pre-trained CNNs for detecting diabetic retinopathy using a minimal dataset. Methods: Six pre-trained models (DenseNet201, ResNet152, VGG16, InceptionV3, MobileNet, and EfficientNetB0) were trained on 40 retinal fundus images each for training, validation, and testing. Models were evaluated over 5, 10, and 15 epochs, focusing on Test accuracy, Validation accuracy, and Validation loss. Findings: MobileNet achieved a Test accuracy of 70% at 15 epochs, consistently outperforming other models across all epochs. InceptionV3 initially performed well but was ultimately surpassed by MobileNet, which demonstrated stable and robust performance with limited data. These results suggest that MobileNet is highly effective in learning and generalizing features relevant to diabetic retinopathy detection, making it an optimal choice for use in resource-constrained settings. Novelty: This study highlights MobileNet's superior efficacy in diabetic retinopathy detection using minimal data, addressing an area that has been largely unexplored in existing literature. Keywords: Diabetic retinopathy, Convolutional neural network, Transfer learning, Pre­trained models, Minimal Dataset