Diabetic Retinopathy Aided Diagnosis System Based on Improved P-ResNet50

Diabetic Retinopathy (DR), a significant complication of diabetes mellitus, can induce irreversible visual impairment. Given the escalating number of diabetic patients globally, the demand for early DR screening and diagnosis is surging. Traditional DR diagnosis, hinging on fundus examination by professional ophthalmologists, is not only time-consuming and labor-intensive but also necessitates extensive physician experience. Against this backdrop, this paper puts forward an intelligent assisted diagnosis system for DR, which is based on an improved P-ResNet50 incorporating a pulse-coupled neural network (PCNN). This system follows a two-phase framework: in the image enhancement phase, PCNN is utilized to boost the contrast of low-contrast lesions in fundus images and quell noise; in the classification phase, the P-ResNet50 model, enhanced by an attention mechanism, zeroes in on key features. Experiments conducted on the Fundus Image dataset, comprising 650 clinical fundus images, yielded promising results. The system achieved a precision of 0.923 and a specificity of 98% when classifying positive images among lesion images, significantly surpassing other classic models. Moreover, the integration of Pulse-Coupled Neural Network (PCNN) preprocessing further elevated the model's performance, offering highly accurate technical support for automated Diabetic Retinopathy (DR) screening. Looking ahead, the system's clinical application value can be further enhanced through model optimization for lightweighting and expanded multi-center data validation.