FST JOURNAL

There is a real challenge in sourcing data in sufficient volumes to train AI models in a safe, secure and appropriate way. That is, one which maintains the integrity of the data, while understanding that if we do this badly, we create biased solutions that will not work effectively and will actually increase health inequalities.

Canon Medical is a division of Canon, the multinational organisation that makes printers, scanners, photocopiers and document-imaging solutions. The company was originally formed over 30 years ago by two graduates from the University of Edinburgh. It now employs about 140 people, is still based in Edinburgh and is now part of Canon. We work with universities every day and we work very closely with NHS and Scottish Government colleagues.

So our world is imaging. We can produce the most modern 3D visualisations. Yet beautiful pictures are not the point. We want to help humans to work as efficiently as possible by automating some of the more mundane tasks and allow patients to progress through the healthcare system as rapidly as possible. The demographic challenge facing the health service is unsolvable without more technology. There are just too many of us living too long in need of healthcare – and not enough doctors and care staff to look after us. It is simple to state the problem but not so easy to solve. Part of the solution, though, is AI.

Canon, as a multinational healthcare business, has decided to put its AI Centre of Excellence in Edinburgh. That is because of both history and also the opportunities in Scotland to work with the universities, the NHS and the Scottish government.

The timing of that decision was fortuitous. About five years ago, money became available from Innovate UK as part of the Industrial Strategy, which included a lifesciences sector deal. This stressed the place of AI and agreed to fund a number of centres to look at AI within healthcare. One of these was in Scotland. The Industrial Centre for Artificial Intelligence Research in Digital Diagnostics (iCAIRD) brought together different organisations supported by Scottish Government, including multinationals like Canon and Philips, as well as Nvidia who make the computer platforms used for AI nowadays.

It took a couple of years to advance from the original idea to a large-scale project. Funding was originally about £15 million, with £10 million from Innovate UK who recognised our ability to create this consortium in Scotland.

ICAIRD

 

Shutterstock

 Two parts

There were two parts to the project. The first was radiology imaging, the second was pathology, examining histology and microscope slides.

The radiology started off as a relatively modest project, with three different exemplars in the Glasgow and Grampian regions, supported by the relevant NHS bodies and universities. The project ran for an extra year because of Covid and it delivered far more than originally envisaged. It resulted in a Covid work package in addition to other work packages as well, a much bigger and more comprehensive programme of work that is still continuing.

There were even smaller ambitions for pathology, working initially in Glasgow with Philips. Again, though, a large network of projects developed.

Success was due to the fact that the model was different from that previously used for AI training in healthcare. Instead of a company approaching the NHS and offering to use their data in the company’s data centre, on company in-house computers, we did the opposite. The NHS kept their data in their own safe and secure environment. We went to them, logging into their systems securely and we carried out our AI training within that safe haven.

Transformational

That approach has been transformational. It respects the ownership and privacy that are so important to citizens: this is their data and nobody else's.

Also, the AI is not being trained with cleansed exported data, it is taking the real data as it is. That is important because an algorithm trained on real data will work on real data. Create a completely sanitised view of the world and train the AI in that world, then it will likely fail in the real world because the real world is not sanitised.

Other people recognised that what we were doing was different and they came here to Scotland to work in this way too.

Canon has helped develop the safe haven AI training platform called SHAIP. Pathology Lab was able to digitise a large number of slides, creating a digital archive that is now available for research and ongoing use. A number of different partnerships and collaboration models have also been established: we started off with 15 partners and £15 million funding but we have ended up with 40 partners, £25 million and about 250 staff. We also have access to 75 million medical images and when it comes to training data, the volume is the most important factor.

A further offshoot from the original project has been delivered in Aberdeen under the Opportunity Northeast programme. This was funded under the Small Business Research Initiative (SBRI) with support from Scottish Government. Around 100 companies came to Scotland for the teaser session to understand the opportunity. We have created a reason for people to come here: there is the availability of data and a triple helix model of collaboration – academia, business and the NHS.

It is said that imitation is the most sincere form of flattery. So we were delighted when NHS England decided to launch a scheme including a network of ‘secure data environments’. They backed that with a budget of up to £200 million in funding. This is exactly the model that has been piloted in Scotland with iCAIRD. There will be regional centres where innovators can access data, be they businesses, academics or just an independent person.

Federated learning

A technology called federated learning allows individual pools of data to be connected together. So this is not just about creating a piece of AI that is trained in Aberdeen but which will only ever work in Aberdeen. If AI is to apply globally, it has to be trained on a representative global population or it will not work.

Now, it is just not realistic to have large international pools of data that people can access. What is realistic, though, is for a company or an innovator to do some training here in the UK with the NHS, then go to the US, to China, to Russia in order to build a representative population for the world. England is doing it. Scotland is committed to doing this as well.  Making this data available will enable the level of applicability of AI algorithms to be much higher. That will benefit users of the health service across the country.