Improving Inference Time in Diabetic Retinopathy—Recent Trends and Future Directions

Diabetic Retinopathy (DR), a leading cause of blindness worldwide, requires an early and accurate diagnosis. Deep learning models have advanced automated DR detection, but their reliance on high-dimensional features creates computational bottlenecks, limiting real-world deployment. This literature review (2019–2024) evaluates optimized Feature Selection (FS) techniques to address these challenges. We analyzed 70 studies focusing on FS methods integrated with deep learning models for DR classification. Our findings reveal that hybrid FS methods achieve higher accuracies, up to 98.9%, by combining filter-based dimensionality reduction with optimization-driven feature refinement. These methods reduce computational overhead by 40–60% compared to traditional techniques, enabling faster inference on resource-constrained devices. Classical techniques like LDA also show promising results but risk omitting non-linear biomarkers. Despite progress, challenges persist—class imbalance in public datasets skews models toward majority classes, while aggressive FS risks discarding critical features like microaneurysms. This survey provides a detailed analysis of FS techniques that reduce inference time and discusses accuracy-efficiency trade-offs, inference latency reporting, and future directions for real-time clinical deployment in the domain.