AI in Diabetic Retinopathy

Diabetic retinopathy (DR) is a diabetes-related eye condition caused by damage to the blood vessels in the retina (the light-sensitive tissue at the back of the eye). Over time, high blood sugar levels can weaken and block these vessels, leading to vision impairment and in severe cases, blindness. DR often develops without early symptoms, making regular screening essential for timely detection and treatment.

As the global prevalence of diabetes continues to rise, so does the number of individuals at risk for diabetic retinopathy. Screening guidelines typically recommend eye exams every 1–2 years, but health systems are struggling to keep pace with the demand. In this context, artificial intelligence (AI) is gaining attention as a scalable and efficient solution to support DR screening efforts, offering the potential to improve access, reduce costs, and detect disease earlier and more accurately.

When implementing AI in this sector, it is important to consider the specificity and sensitivity of AI. Specificity refers to how well the system can correctly identify images that do not have the thing it's looking for. A high specificity means there are fewer false alarms. Sensitivity is how good the AI is at correctly finding something when it is actually there. A high sensitivity means there are fewer missed cases. 

An Australian study conducted by Scheetz et al., demonstrated that a particular AI system used for diabetic retinopathy screening had 96.9% and 87.7% sensitivity and specificity, respectively. These results overwhelmingly surpassed the United States Food and Drug Administration (FDA) recommendations with strong sensitivity and specificity scores.

According to another Thai study, the AI algorithm showed 91.4% sensitivity and 95.4% specificity. In comparison, retinal specialists had slightly lower sensitivity at 84.8%, with a similar specificity of 95.5%. These differences are significant, they highlight how capable AI models can be in detecting diabetic retinopathy and suggest real potential for wider use in the future.

However, it is important to consider the disadvantages that come with using AI for DR screening. A major drawback is that doctors often induce mydriasis to improve screening accuracy by reducing ungradable images. Mydriasis is when the pupil of the eye becomes abnormally dilated or widened and is unresponsive to light. The issue with this however, is that this can cause temporary vision issues that interfere with daily activities. Unlike manual graders, AI systems struggle with blurred images since they are trained on ideal conditions and lack the flexibility to adjust viewing techniques. As a result, relying on AI for diabetic retinopathy screening may increase the need for mydriasis, which can disrupt the patient’s quality of life.

In conclusion, AI shows great promise in improving diabetic retinopathy screening by offering high accuracy, scalability, and efficiency. Studies demonstrate that AI can match or surpass human specialists in detecting the disease. However, its limitations with poor-quality images and the possible increased need for mydriasis highlight the importance of combining AI tools with clinical oversight for optimal patient care.