Scalable screening for diabetic retinopathy remains difficult to deliver where it is most needed. We present a hardwareefficient convolutional neural network (CNN) designed for reliable, on-device DR detection. Our approach integrates a compact CNN architecture with structured pruning and a Nondominated Sorting Genetic Algorithm II (NSGA-II) to work under strict memory and compute budgets. We validate the approach on standard DR datasets and demonstrate deployment on an ARM microcontroller, highlighting its practical feasibility for portable screening tools in rural and underserved clinics. This work contributes (1) an end-to-end, resource-aware pipeline that couples architecture search with pruning, (2) a principled optimization strategy that balances diagnostic accuracy and efficiency, and (3) an embedded deployment that illustrates scalable, accessible AI-driven DR screening.

Keywords