Lightweight Deep Learning for Early Diabetic Retinopathy Detection: Benchmarking, Efficiency, and Applicability in Resource-Constrained Healthcare

Diabetic retinopathy (DR) remains a leading cause of preventable blindness worldwide, with its burden most acute in resource-limited settings where access to specialist care and advanced diagnostic tools is restricted. Early detection is vital to mitigate vision loss, yet most state-of-the-art deep learning models demand high computational resources, hindering deployment in such environments. This paper proposes and validates a lightweight convolutional neural network (CNN) for DR detection that balances diagnostic accuracy with computational efficiency. Using a balanced dataset of 4217 retinal images, the model achieved an accuracy of 81.1%, a macro F1-score of 0.8125, an inference time of just 12 ms per image, and a compact 11 MB model size. To ensure robustness, we conducted comparative benchmarking against widely used architectures. ResNet, GoogLeNet, and VGGNet, demonstrating that while these deeper models achieved higher accuracy (up to 88.7%), they required significantly larger memory footprints and slower inference speeds. By contrast, the lightweight model maintained competitive performance while being substantially more efficient. These results establish the proposed model as particularly well-suited for low-resource healthcare environments, including mobile health platforms, telemedicine applications, and rural clinics lacking high-end infrastructure. Beyond technical contributions, this work addresses a critical gap in the literature by explicitly validating lightweight CNNs as feasible, scalable, and equitable solutions for global healthcare challenges.