Deep learning methods aid diagnosis of eye disease

28 Aug 2018

AI tools from DeepMind were able to interpret OCT retinal scans rapidly and accurately.

A research partnership between Moorfields Eye Hospital and UK-based artificial intelligence company DeepMind has released the first results of its investigation into the use of AI for diagnosis and referral in retinal disease.

Having applied its novel deep learning architecture to OCT scans from patients treated at the hospital, the team found that the AI platform's performance in making a referral recommendation reached or exceeded that of experts on a range of retinal diseases. The findings were published in Nature Medicine.

"The results show that our AI system can quickly interpret eye scans from routine clinical practice with unprecedented accuracy," commented the DeepMind team announcing the findings. "It can correctly recommend how patients should be referred for treatment for over 50 sight-threatening eye diseases as accurately as world-leading expert doctors."

If the system proves able to handle the wide variety of patients found in routine clinical practice, then it could in the long term help doctors quickly prioritise patients who need urgent treatment, and hence ultimately save patients' sight.

AI has been a promising solution to the interpretation of medical images for some time, but applying it in real-world clinical scenarios has remained challenging.

In its paper, the DeepMind team note that AI systems must be trained on hundreds of thousands of examples from one canonical dataset, and must then generalize to new populations and devices without a substantial loss of performance or prohibitive data requirements.

In addition, the AI tools need to be applicable to real-world scans and designed for clinical deployment, while matching or exceeding the performance of human experts in such real-world situations. Recent work applying AI to OCT had shown promise in resolving some of these criteria in isolation, but has not until now shown clinical applicability by resolving all three.

Clinical confidence

DeepMind's solution combines two different neural networks. The first, known as the segmentation network, analyses the OCT scan to provide a map of the different types of eye tissue and the features of disease it sees, such as hemorrhages, lesions, irregular fluid or other symptoms of eye disease - a process the company describes as "allowing eyecare professionals to gain insight into the system’s thinking."

The second network, known as the classification network, analyses this map to present clinicians with diagnoses and a referral recommendation. Crucially, the network expresses this recommendation as a percentage, allowing clinicians to assess the system’s confidence in its own analysis.

"This functionality is critically important, since eyecare professionals are always going to play a key role in deciding the type of care and treatment a patient receives," commented the DeepMind team. "Enabling them to scrutinize the technology's recommendations is key to making the system usable in practice."

Initial tests using an OCT platform at Moorfields showed that after suitable training an error rate on referral decision of 5.5 percent was achieved. Subsequent modifications to the procedure, and a switch to a different commercial OCT system, reduced this error rate to 3.4 percent.

"Our technology can be easily applied to different types of eye scanners, and not just the specific type of device it was trained on at Moorfields," noted the team. "This might seem inconsequential, but it means that the technology could be applied across the world with relative ease, massively increasing the number of patients who could potentially benefit."