Vision-Threatening Diabetic Retinopathy Detection Using EfficientNetB4: A Two-Stage Transfer Learning Approach

Abstract

Vision-Threatening Diabetic Retinopathy (VTDR) marks a stage of diabetic eye disease that is characterized by a need for urgent medical treatment to avoid permanent sight impairment. In this paper, we have proposed an automated deep learning framework for binary classification of retinal fundus images in to Non-VTDR (Healthy/Mild DR) and VTDR(Moderate/Severe/Proliferate DR) categories. The transfer learning employs a two-stage strategy with EfficientNetB4 backbone architecture with weights pretrained on ImageNet. The dataset has 2,750 retinal images, and it is split into training (70%), validation (20%), and test (10%) subsets with classes balanced as best as possible. By using data augmentation strategies such as rotation, zoom, brightness, and horizontal flip, our approach increases generalization and helps the model generalize better. The model first does feature extraction with frozen base layers (10 epochs), and then fine-tunes the top 120 layers (20 epochs) with class weighted binary cross-entropy loss. Experimental results suggest that summary accuracy of 86.18% supports our method potential for automating VTDR screening. This system has potential to be used as a part of the real- time clinical workflow to help ophthalmologists detect sight-threatening conditions at an early stage mainly in the resource-limited settings with no access to specialists.