Lightweight AI Fundus Image Model for Early Screening of Diabetic Retinopathy

Diabetic retinopathy (DR) is now the main cause of preventable blindness in working-age adults worldwide. With the rising prevalence of diabetes, the need for DR screening is becoming increasingly urgent. Traditional DR screening requires professional ophthalmologists; thus, there are problems of a shortage of expert resources, uneven distribution, and lengthy manual interpretation. Most of the current deep-learning models have too many parameters and are computationally expensive; therefore, they are not suitable for deployment on edge devices. To solve the problem of deployment bottlenecks, the development of lightweight artificial intelligence models for fundus images has become necessary. This paper systematically reviews the recent developments in lightweight artificial intelligence models for DR screening based on fundus photographs. Custom image pre-processing, a lightweight network design, advanced model compression techniques and interpretability mechanisms are employed to reduce both the number of parameters and computational cost substantially without losing the sensitivity to early retinopathy. The model's diagnostic reliability and computational resource consumption have been examined. Ablation studies and visual analyses show the respective contributions of each module to the model's performance and efficiency. Additionally, to improve the interpretability of the model and boost clinical trust, Gradient-weighted Class Activation Mapping (Grad-CAM) is introduced. Future research will enhance the adaptability of cross-device and cross-domain applications, expand the scope of clinical verification, and integrate multimodal clinical data for reliable and fair ophthalmic AI deployment.