Artificial intelligence can now match or outperform human experts in diagnosis and referral on eye diseases, suggests a new paper from DeepMind. The UK-based, Google-owned research institute today released joint research results with the UK’s Moorfields Eye Hospital and UCL Institute of Ophthalmology, which present a new AI technique in the context of OCT imaging. The paper was published on Nature Medicine’s website.
OCT (optical coherence tomography), also known as ‘optical ultrasound’, is a critical imaging tool for evaluating eye health that captures high-resolution 2D/3D images from biological tissues. The dearth of specialists with OCT diagnosis expertise has driven research into AI applications, but how to apply AI in 3D diagnostic scans remained an unsolved challenge.
“The number of eye scans we’re performing is growing at a pace much faster than human experts are able to interpret. There is a risk that this may cause delays in the diagnosis and treatment of sight-threatening diseases, which can be devastating for patients,” says Moorfields Eye Hospital Consultant Ophthalmic Surgeon Dr. Pearse Keane.
The paper’s authors apply a novel deep learning architecture to a heterogenous set of 3D OCT scans from Moorfields Eye Hospital patients. After training on less than 15,000 scans, the AI model delivered a referral suggestion for over 50 critical eye diseases with 94 percent accuracy, achieving and in some cases exceeding human experts’ performance.
The DeepMind AI architecture generates a segmentation network that can transform OCT scans into a map of different types of eye tissues and disease features; and a classification network that can provide diagnoses and referral suggestions with expert-level accuracy. The network can be integrated into a physician’s workflow to present recommendations in the form of percentages.
“One of the most important aspects of the work is the interpretability of the system, so clinicians can see and understand how the system is making its recommendations,” says Demis Hassabis, CEO and co-founder of DeepMind.
The AI system will also automatically suggest treatments for patients who are in critical condition and need urgent care.
While the research is still in its first phase, DeepMind plans to promote the technology across all Moorfield’s UK hospitals and community clinics, which serve 300,000 patients a year and receive over 1,000 OCT scan referrals every day. Moorfields Eye Hospital and DeepMind Health announced their five-year partnership to explore the potential of AI-applied healthcare technologies in 2016.
This year, DeepMind parent company Google stepped up its efforts to apply advanced AI technology to eye care. In April, the Mountain View tech giant began using AI to treat detect diabetic retinopathy — high blood sugar levels which can result in blindness — at three eye hospitals in India. The company also demonstrated how AI models can use retinal images to predict a patient’s risk of heart attack or stoke.
In the US, FDA approvals for AI in medicine are increasing, according to physician-scientist Eric Topol. The FDA has approved 11 uses of AI in medicine this year comparing to only two last year.
But recent layoffs at IBM Watson Health are a setback for the AI healthcare market, which is estimated to reach almost US$8 billion by 2022. Founder of Landing.ai and Deeplearning.ai Andrew Ng tweeted: “We need to set realistic expectations. AI will transform industries, but the path is not always obvious.”
Although the new DeepMind paper represents a leap forward for AI in eye health and medicine in general, there are still many obstacles to ultimately transforming patient care with AI.
Journalist: Tony Peng | Editor: Michael Sarazen