Abstract

Diabetic retinopathy (DR) is a serious condition that damages retinal blood vessels, potentially leading to blindness. Traditional diagnosis involves manual analysis of colored fundus images by clinicians, which is error-prone and time-consuming. To mitigate these challenges, computer vision techniques have been utilized for automating DR detection. However, existing methods often struggle with computational complexity and inadequate feature extraction for precise DR stage classification. This paper introduces a novel approach for classifying DR stages with minimal learnable parameters to enhance training efficiency and model convergence. The VGG16 architecture, augmented with a spatial pyramid pooling layer (SPP) and network-innetwork (NiN) structures, constitutes the VGG-NiN model, capable of effectively processing DR images across different scales due to the adaptability of the SPP layer. Moreover, the incorporation of NiN enhances the model's ability to capture nonlinear features, thereby improving classification accuracy. Experimental findings validate the efficacy of the proposed model, demonstrating superior performance in terms of accuracy and computational efficiency compared to existing techniques.