Title : Deep learning convolutional neural network diabetic retinopathy detection using gaussian filter preprocessing
Abstract:
Background: Diabetic Retinopathy (DR) is a severe complication of diabetes that affects retinal blood vessels and can lead to blindness. The Middle East faces a significant public health challenge due to the high prevalence of DR among diabetics, with rates reaching up to 64% in Jordan. Early detection via fundus imaging is essential, but the variability and fatigue in human interpretation can impact diagnostic accuracy.
Methodology: 4,162 retinal images selected from the APTOS 2019 Blindness Detection dataset were used. Two models were trained: one with Gaussian filter as the primary preprocessing method to enhance feature clarity, and one without it. The CNN architecture included multiple convolutional and pooling layers, optimized to minimize the loss function.
Results: The CNN model with Gaussian filter preprocessing achieved 99.5% accuracy on the training set, 92.3% on the validation set, and 94.9% on the test set. In contrast, the non-filtered model exhibited overfitting, with higher training accuracy but lower validation and test accuracies. The filtered model also excelled in precision, recall, and F1-score across DR categories.
Conclusion: Our comparison showed that the CNN model with Gaussian filter preprocessing significantly outperforms the non-filtered model, demonstrating higher accuracy and better generalization. This underscores the value of advanced preprocessing in improving DR detection and its potential for enhancing diagnostic accuracy in the Middle East and beyond.
