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

Abstract

ABSTRACT Threatening Diabetic Retinopathy (VTDR) remains an advanced stage of ophthalmic complications secondary to diabetes and is characterized as a major, preventable cause of irreversible vision loss. In this paper, we propose an automated deep learning framework to classify retinal fundus images into Non-VTDR (Healthy/Mild DR) and VTDR (Moderate/Severe/Proliferate DR). We use a two-stage transfer learning with EfficientNetB4 backbone architecture set up during training with ImageNet weights. A dataset of 2,750 retinal images is split into balanced training (70%), validation(20%), and testing (10%) sets. We applied several data augmentation techniques using rotation, zoom, brightness and horizontal flipping to improve model generalization. We freeze all base layers for 10 epochs, use class-weighted binary cross-entropy loss, then fine-tune the top 120 layers of VGG16 (20 epochs). Experimental results show a total test accuracy of 86.18% proving the efficiency of our method for automating VTDR screening. Conclusion: A network that maintains relatively high accuracy, is easy-to-use, and takes only a few milliseconds to classify an image can be easily integrated into clinical workflows to provide ophthalmologists with decision support for finding sight-threatening disease, especially in underprivileged areas where access to specialist care is limited. Keywords: Deep Learning, EffientNetB4, Transfer Learning, Machine learning (ML), Medical image classification, Retinal Fundus Imaging,, Computer-Aided Diagnosis,(CAC), Binary Classification Multiclass Classification Diabetic Eye Disease Screening