Optimizing Deep Learning Models for Ophthalmic Disease Detection on Resource-Constrained Devices

This paper explores the optimization of convolutional neural networks (CNNs) for ophthalmic disease detection via fundus image classification on resource-constrained edge devices. The study investigates the impact of pruning, quantization, and knowledge distillation on reducing model size and inference time while maintaining high classification accuracy. Using a dataset of retinal fundus images, the optimized models were deployed on a Jetson Nano microcontroller. Experimental results demonstrate that post-training quantization achieves the best trade-off between efficiency and accuracy, reducing model size to 11.22MB while maintaining a validation accuracy of 97.39%. Hybrid approaches combining knowledge distillation and quantization further reduced parameters while preserving performance, highlighting promising strategies for deploying deep learning in edge AI healthcare applications.