ResNet18 with Triplet Attention Mechanism for Diabetic Retinopathy Detection

Patients with diabetes may develop diabetic retinopathy (DR), which can cause visual loss and eventually blindness. This damage can be prevented if DR is identified initially. Current DR detection techniques based on fundus images are labour-intensive, complex, and prone to errors. In this paper, the ResNet18 architecture combined with the triplet attention mechanism is used to develop a deep learning based framework for DR detection and its classification. Its lightweight design suppresses the background noise and concentrates on disease-relevant retinal regions with improved classification performance. The proposed method employed pre-processing steps to enhance vessel visibility and reduce illumination variability. Furthermore, class weighting and augmentation strategies are also used to minimise the class imbalance problem. Based on experimental results, this framework, ResNet18 enhanced with Triplet attention, outperforms the other existing techniques and achieves 98.91% accuracy, 98.49% specificity, 99.40% sensitivity, 98.22% precision, 98.81% F1-score, and 99.74% AUC. The method provides an interpretable and computationally effective solution appropriate for mobile healthcare applications and large-scale screening.