Connected Pathology in the UK: Part 5

“I'm all about images, quality, and helping to save the NHS through technology,” says David Brettle, quality lead for the UK-wide National Pathology Imaging Cooperative (NPIC). So far the program has rolled out connected digital pathology systems to more than thirty hospitals in the National Health Service (NHS), with another forty set to join over the next year or so.

Brettle has the mammoth task of leading on quality standards and consistency for NPIC that will impact across the whole network. What’s more, because the NPIC network is on an unprecedented scale, these standards are serving as a blueprint for digital pathology around the world.

In the fifth episode in our interview series with the NPIC team, we found out why Brettle and colleagues are the guardians of end-to-end quality in digital pathology.

Tell me about your background and your role on the NPIC team
 

I'm the Chief Scientific Officer for Leeds Teaching Hospitals (TH), which means I am the professional lead for the NHS healthcare science workforce of around 1,280 across the Trust. I am also leading on technology innovation/innovative technology for the two new hospitals we are planning to build in Leeds. The technology for the new hospitals has to be digital and sustainable, promote economic growth, and push the boundaries. As the new hospitals are going to be smaller than the estate they replace, with more patients going through, technology will be essential to meet the demand.

My role in NPIC is to lead on the quality side of work, which to some may sound like the least exciting part of NPIC, but I think it’s the best bit!  My background is in medical physics. I used to be the head of Medical Physics in Leeds TH, but have spent most of my career in imaging of one kind or another, and that's how I met NPIC leader Darren Treanor. We had a similar interest in image quality, and were involved in putting grant proposals together. And, suddenly… I'm in NPIC.

What are the main issues or concerns around quality that you've encountered from the labs?
 

The process of making histological slides hasn’t changed much since its first development around 180 years ago. We do have external technical and clinical validation to make sure our techniques are acceptable, and internal controls for consistency, but it's all very subjective. To date there has been a paucity of quantitative methods for slide and stain quality. That results in huge slide variability between laboratories, which is a known problem but, to a large extent, has been tolerated as humans are able to cope relatively well with variation. To add to that though, we also now have digital processes that have their own variations.

AI works better with consistent data, both for training and deployment. There are methods to cope with variation, but with varying levels of success. So, we have been looking at the end-to-end quality in digital pathology. That includes looking at traditional histology, where quality control has been largely subjective in the past, and trying to move that into a more quantitative framework. We are also looking at new methods of quality control for the digital elements such as the scanners and display. And I'm pleased to say that we have developed several tools to improve quality in these areas – for example the online display QC check called POUQA which is freely available to use. And, for the first time, we have developed a practical quantitative tool for measuring stain variability in the laboratory with a stain assessment tool called “Tango.”

This allows labs to potentially save on reagents, optimize staining, and reduce recutting and restaining of slides, reducing delays in diagnosis. We can now also look to set standards and guidelines around staining, to give us a more consistent dataset, which in turn should allow AI to work better, as long as you use a quality system.

We believe good quality management in histology will be transformative for accelerating AI development and adoption.

Tell me more about the standards that the team is developing
 

We’re looking beyond the UK, and developing world standards. Our strategy has been to identify any quality control gaps in the digital pathology pathway: where there are already quantitative tools to address those gaps  – to use them; where there are not – to develop our own. We then will build evidence to support the development of guidelines and standards moving forwards. To date we have developed tools in staining, display, and digitization.

We have also been collaborating internationally to evaluate these tools and, ultimately, to help grow the culture of quality, and the culture of innovation, in pathology worldwide.

There are many who say, “Slide staining hasn’t been a problem in 180 years, why are we looking at it now?” Well, actually I believe we have a duty to apply quality as part of our covenant with patients. In addition, we believe it may be possible to reduce costs in the laboratory by reagent optimization. My rough estimates are that we can save over a million pounds for the NHS for every five tonnes of chemical wastage prevented and, at the same time, make AI more realistic via more consistent datasets. That sounds like a win–win to me.

Basically, we are trying to change the world and bring everybody up to the highest standard in quality. And we have won multiple awards for our quality initiatives including a chemical industry award for co-operation with the NHS and industry.

Can pathologists trust AI to take over some of their tasks and do them as well as humans can?
 

That's the same question as “Can technology save the NHS?” The answer to that is “yes” but “no.” We not only need to ensure we have the right technology that is reliable and trustworthy, but we also need that culture of innovation to optimize the benefits – and for that I believe we need clinicians to be part of its development. However, I would say that if we continue to develop as we are then, yes, AI will have a chance of working well in pathology. And, if we can develop that trust and understanding, we can start to realize real staff and patient benefits.

Personally, I think there's nothing scary about using AI – it's basically just math, at the end of the day. What's scary is not understanding how it works and not knowing how it affects your workflow. I think that evolution is still ongoing. Explainable, transparent AI is part of that evolution and in turn that will help pathologists to gain trust and experience in AI implementation to ensure its continuing development. I believe quality underpins all of that.

What are the challenges with validation of AI?
 

That can be complicated because AI is developed from a dataset. Say we develop an AI application on a dataset in Leeds, and then install it in Manchester, their data might be different, and it may not work. One way to get around this is by using lots of training data from multiple institutions, or by local retraining. This should make the algorithm more transportable between institutions. In other words – AI is only as good as its training dataset.

Going back to the trust issue, because there's a lack of transparency in AI and, on occasion, it can behave unpredictably, this makes people nervous. I think we need to either develop ways of having that transparency and communicating how safe AI is, or share real-world experience of using it safely and effectively. Those are two things that need to happen before pathologists will lean on AI and trust it completely.

What are the next steps for the team?
 

We are currently in a five-year plan around commercializing our stain assessment tool to allow the tool to become widely available. Once the tool becomes widely used, we can start to realize the vision of improving quality at scale and pace.

When we start to deploy our stain assessment tool, we can start to look at standards. We are gathering real-world evidence to show that quality achieves better consistency of slides. Then we can present the evidence to the professional bodies, standards organizations, and regulators, and work towards agreed standards and guidelines.

Once the standards are in place, then we will see this wonderful situation of more consistent data, and faster AI. And that will accelerate AI adoption and development. That's the plan.

In parallel, we're developing a new display unit for digital pathology. The prototype arrived just recently and we’re very excited about it.

That demonstrates an important part of NPIC’s remit. We look at what labs need for digital pathology and, if it doesn't exist, we'll develop it and make it available. Our skill is in translating that problem solving into real technology. To improve staining consistency we had to develop a product that's really hard to develop, and then we had to validate it and commercialize it to make it available worldwide. I believe we really have achieved a lot already.

What should the take-home message be?
 

We're very proud of what we do at NPIC. I am very proud of the quality team and what they have achieved. It is sometimes easy to lose sight of the impact of your work on patient outcomes when you are not patient facing. But I truly believe quality is not only expected by our patients as part of our covenant with them, but that it will be a key part in realizing the benefits of digital and AI, which will have direct benefits for staff and patients. So I think that quality is actually pretty cool!