Diabetic retinopathy detection by fundus images using fine tuned deep learning model

Abstract

This study employs transfer learning using a fine-tuned pretrained EfficientNetB0 convolutional neural network (CNN) model to accurately detect the various stages of Diabetic Retinopathy. The training process involved utilizing three datasets: Messidor, IDRiD (Indian Diabetic Retinopathy Detection), and APTOS 2019 Blindness Detection, which collectively encompassed 5,379 fundus images. Different types of processed fundus images were fed into the model to determine the optimal pre-processing approach for stage detection in Diabetic Retinopathy. The model was assessed on the original dataset with some augmentation techniques applied. According to the training data, the model achieved a maximum accuracy of 72%. However, converting the dataset to grayscale yielded an improved accuracy of 80%. Similarly, extracting the green, red, and blue channels individually resulted in accuracies of 72%, 76%, and 73% respectively. Notably, when the green channel extracted images underwent histogram equalization, the model achieved its highest accuracy of 83%. Furthermore, the application of a Sobel filter to the red channel images led to a maximum accuracy of 51%. Finally, to determine the effectiveness of each processed image type, sensitivity and specificity measures were compared. Among all the variations, the green channel extracted images with histogram equalization demonstrated superior performance in correctly identifying the respective classes, outperforming the other approaches.