An Efficient Deep Learning Framework for Automated Diabetic Retinopathy Detection Using OCTA Images

Resumen

Diabetic retinopathy (DR) is a leading cause of vision impairment among individuals with diabetes, making early detection and accurate severity grading essential for effective clinical intervention. This work proposes an automated deep learning–based framework for the detection and classification of diabetic retinopathy using Optical Coherence Tomography Angiography (OCTA) images from the ROAD dataset. The proposed methodology consists of three main stages: image preprocessing, deep feature extraction, and multi-class classification. In the preprocessing stage, image resizing using bilinear interpolation, Z-score normalization, and Contrast Limited Adaptive Histogram Equalization (CLAHE) are employed to enhance retinal microvascular structures and suppress illumination variations. Deep features are extracted using convolutional neural network architectures, including ResNet and EfficientNet variants, and optimized using the Adam optimizer. A Softmax classifier is then utilized to categorize OCTA images into No Diabetic Retinopathy, Mild Non-Proliferative Diabetic Retinopathy, and Proliferative Diabetic Retinopathy classes. The performance of the proposed framework is evaluated using accuracy, sensitivity, specificity, precision, and F1-score metrics. Comparative analysis demonstrates that the EfficientNetB4 + Softmax model achieves superior performance, with an accuracy of 98.97%, sensitivity of 97.26%, specificity of 98.54%, precision of 98.66%, and F1-score of 98.25%. The obtained results confirm the effectiveness and robustness of the proposed framework.