Development and validation of a deep learning model for early detection and screening of diabetic retinopathy

Early diagnosis and screening of diabetic retinopathy (DR) are crucial for reducing medical burdens and conserving healthcare resources. This study introduces an advanced AI-assisted recognition system designed to enhance the detection of DR lesions through innovative automatic learning methods. Central to our approach are agnostic text instruction templates, which facilitate zero-shot DR detection by integrating text embeddings with visual information. Our system performs comprehensive lesion detection by leveraging similarity mapping at both the image and patch levels, enabling it to identify a wide range of diabetic retinopathy (DR) lesions without the need for extensive annotated data. This AI-assisted system distinguishes itself from traditional fully supervised models and few-shot learning approaches by addressing the complexities of DR image annotation and safeguarding patient privacy. To validate the system's effectiveness, we conducted extensive experiments across five internal and publicly available test sets, as well as an external test set captured using smartphone devices. Our evaluation involved performance analysis of various pre-training methods, including detailed patch-level visualizations and t-SNE clustering techniques to assess the quality of feature embeddings. The results of our zero-shot experiments reveal that our system outperforms conventional transfer learning-based DR detection methods. This superiority is evident in both the pre-training and testing phases, showcasing the system's ability to deliver accurate and reliable DR lesion detection while circumventing the limitations of traditional approaches.