Artificial Win(telligence)

With several groups and researchers developing artificial intelligence (AI) and deep learning systems for ophthalmic applications, diagnosis by machine is on the cards. Researchers at the Singapore National Eye Center and National University Singapore School of Computing have brought us one step closer with a deep-learning system that detects diabetic retinopathy and related eye diseases (glaucoma and AMD). But how does it compare with professional human graders? Using 494,661 retinal images from multiethnic (Chinese, Indian, Malay, Hispanic, African-American and White) patients with diabetes, the system demonstrated high sensitivity (≥90.5 percent) and specificity (≥87.2 percent) for identifying retinal diseases, comparable with the professional graders (≥88.5 percent and ≥99.3 percent, respectively). Daniel Ting, lead author on the corresponding paper (1), tells us more.

What impact will your AI system have on clinical practice?

It could potentially reduce total workload by 50–70 percent simply by removing non-referable images and allowing human graders to focus on the retinal images that need more attention. An established AI system could also be useful in conducting lifelong monitoring.

Any notable challenges in the course of your work?

Our team has come a long way. Together with another four co-inventors (Professor Tien Wong, Professor Wynne Hsu, Professor Mong Li Lee and Dr Gilbert Lim), we started developing and testing this AI system five years ago using retinal images that have been collected for over 10 years. Enormous financial and manpower resources have been poured into this AI project, and I am glad that our team has managed to overcome the initial obstacles and share our results.

What lies ahead for AI in medicine?

AI is the fourth industrial revolution in human history, and it will definitely revolutionize medicine in the next few decades. By having a robust AI algorithm, we also hope to deliver personalized medicine to the global population with diabetes, and we are certainly seeing similar trends in other medical specialties, such as dermatology, pathology and radiology. In ophthalmology, we certainly hope that that AI can help with repetitive workloads; for example, screening for diabetic retinopathy, glaucoma and AMD.

Pros and cons of using AI as a diagnostic tool?

Pros include cost- and time-savings, and zero intra-rater variability. Cons include the need for a large training dataset, technical expertise and supporting infrastructure.

Next steps?

We are currently in the midst of developing more algorithms for other retinal conditions, including retinal vein occlusions and retinal detachment.