A Study on Deep Learning-Based Diabetic Retinopathy Classification Using Fundus Images

Diabetic Retinopathy is a disease that damages microvasculature, leading to vision loss and poten- tially blindness if not treated early. Traditional Fundus examinations are time-consuming, prone to hu- man error, and limited by the shortage of medical professionals. In contrast, Deep Learning technology can automatically detect pathological features in fundus images, significantly aiding early diagnosis and treatment decisions, especially in resource-constrained environments. However, conventional CNN- based models have limitations in effectively capturing critical details and spatial information. To address this, this study proposes a model that integrates ECA(Efficient Channel Attention) and Spatial Attention into EfficientNet. Experimental results on the APTOS 2019 and DDR datasets demonstrate that the pro- posed model outperforms Swin Transformer, ConvNeXt, and EfficientNet in key performance metrics. Future research should include additional experiments using diverse datasets and address the issue of class imbalance effectively.