Moorfields Eye Hospital in London is nearly nine months into a collaboration protect with the deep learning specialists DeepMind, and is aiming to have a working algorithm in production that can make diagnoses from eye scans by the end of the year.

Speaking at the Re:Work Deep Learning in Healthcare Summit yesterday, just around the corner from the hospital in Old Street, London, Pearse Keane — consultant ophthalmologist at Moorfields Eye Hospital and National Institute for Health Research (NIHR) clinician scientist at University College London, Institute of Ophthalmology — explained how the collaboration with Alphabet's DeepMind took 12 months to get off the ground and why AI is so well suited to the field of ophthalmology.

iStock Photo: Rocky89

Specifically, Moorfields is working with DeepMind to create a "general purpose algorithm which can look at Optical Coherence Tomography (OCT) scans and diagnose anything that you would expect a retinal specialist". This could be used to spot conditions such as age-related macular degeneration and diabetic eye disease, and according to Keane it should be ready by the end of 2017. This will help cut down on the amount of unnecessary referrals being made to the NHS off the back of OCT scans, freeing up clinicians to focus on urgent cases.

Applying AI to healthcare

Ophthalmology is a medical field focused on diagnosing and treating any disorders to the eye, and Keane believes it to be the best "test bed" for AI techniques like machine and deep learning within the healthcare space.

Why? Firstly, there is the sheer amount of data generated by ophthalmology. Moorfields alone sees 600,000 patients per year and conducts more than 3,000 OCT scans a week, a figure Keane says is likely "an underestimate". OCT scans use light waves to take a cross-section of images of the retina and are useful in spotting glaucoma and retinal diseases.

Keane added that it is "likely within the next few years that every patient will have a scan as matter of recourse. So this will only increase, so we were buried beneath the data we collected."

The rise of the OCT scan, especially at high street opticians, has created a deluge of outpatients for the NHS eye service, as any scan that deviates from the norm is deemed an urgent referral. Keane says this issue motivated him to seek out a solution using artificial intelligence.

Data preparation and governance

The announcement of the collaboration between DeepMind the AI startup acquired by Google for £400 million in 2014 and Moorfields Eye Hospital made national headlines last year. What wasn’t reported at the time was the amount of time-consuming data preparation work that was required in the background to get a complex project like this off the ground.

Keane said that the hospital "approached DeepMind and said we should apply deep learning to the interpretation and triage of OCT scans" in July 2015. During his first meeting with DeepMind cofounder Mustafa Suleyman he was asked how many scans they had in total, a figure Keane didn’t have to hand, and then, crucially, how many were labelled, again a figure he didn’t have to hand.

Read next: Google’s DeepMind promises openness as it begins public consultation over healthcare plans

So hospital staff had to manually convert the files to an open source format and anonymise the data robustly before any processing could be done. As Keane said: "In that 12 month period it required a huge amount of work and man hours and effort to curate a data set that would be tractable for machine learning."

Keane said that initially he had lots of volunteers for the project, "who kind of get excited initially when they hear Google or a big AI project and are enthusiastic for a week."

"But then they realise that it is incredibly tedious work and their contribution tails off," Keane said. So even for an organisation the size of Moorfields it was a challenge to find the man power to do this task, and particularly with healthcare data where you "can’t just allow anyone to come in.

Read next: Google DeepMind: What is it, how does it work and should you be scared?

"Then we have all of the issues related to ethics, consent, information governance and contractual governance," he added. "I had no clue how time consuming and how much hard work would be required to do this."

Only once all of this data cleansing work was completed could a data sharing agreement be put in place, allowing Moorfields to share 1.2 million 3D OCT scans with paired retinal photographs, all anonymised, with DeepMind.

In practice this data is walled off from DeepMind, bundled together and wiped of any patient-identifiable information and it even gets a final check from an information governance department at the hospital to ensure the data set is completely anonymised before it is sent to the external partner.

What next?

Longer term Keane wants to apply machine learning algorithms "to derive new clinical and scientific insight from the data".

As OCT scans are taken often, Keane believes that they can start to use "novel deep learning approaches to try and figure out information about the path of physiology, the natural history and prognosis of these conditions."

It doesn’t end with OCT scan images either, as Keane sees the long-term vision of the relationship with DeepMind as picking up other data sets, like those taken from glaucoma patients and adding them to the AI pipeline. "But what we are still thinking about is," he said, "what's the best way to do that?"

Read next: Five real-life use cases for Google DeepMind’s machine learning systems

Find your next job with techworld jobs