Leveraging DenseNet121 and Data Augmentation for High-Accuracy Diabetic Retinopathy Screening

This study aims to classify retinal fundus images to detect diabetic retinopathy using convolutional neural networks (CNNs). The research methodology involved developing a CNN based on the DenseNet121 architecture, augmented with a dense layer of 256 neurons and a 5-neuron SoftMax output layer. Two experiments were conducted: the first applied Deep Learning (DL) to an unbalanced dataset, while the second utilized Transfer Learning (TL) combined with Data Augmentation on a balanced dataset. Results showed that DL achieved an accuracy of 79.34% with a 20.04% loss, whereas TL significantly improved performance, yielding 97.78% accuracy and only 6% loss. The study concludes that Transfer Learning with dataset balancing produces a more precise and efficient diagnostic aid compared to standard Deep Learning, offering a reliable tool to support medical professionals in accelerating screening processes and prioritizing patient care.