A Newer Frontier

Digital pathology is no longer a “new frontier” in the field – in fact, it’s growing increasingly mainstream. But even for those who have readily adopted the digital side of the discipline, artificial intelligence (AI) often remains a mystery. Why should laboratories consider “going AI?” What benefits does it offer? And what about laboratories with fewer resources, higher case volumes, or less digital infrastructure? At first glance, AI may not seem accessible or equitable – but the reality might surprise you.

Pathology’s biggest challenges
 

Many places currently face pathologist shortages – with too few staff members to tackle the laboratory’s volume of work. That’s not necessarily the case everywhere but, even in places with sufficient staff, labs are under increasing pressure due to not just the growing volume of cases, but also their complexity. That’s why a lot of labs are now looking to get into digital and computational pathology – but that, too, raises a lot of questions. How do I evaluate AI algorithms? How do I choose the right digital pathology vendor for my work? What IT infrastructure do I need? It’s easy to talk about getting started, but harder to actually do it.

Conversations about digitization often compare pathology to radiology – but I think AI will play somewhat of a different role for pathologists than it does for radiologists. That’s partly because, when radiology went digital, they simply didn’t have the AI capabilities we have today. AI and pathology are growing and learning together, which means we have an opportunity to really add value by incorporating AI into our workflows. That has led to a reluctance among pathologists to just “go digital and then figure out AI later.” They want to incorporate both at the same time because they feel that, without AI, they won’t see the full benefits of digitization.

AI brings with it three main value propositions:

1. Improved diagnostic accuracy.

Pathologists are often busy and overworked. Every day, we see AI catching potential diagnostic errors and alerting pathologists to their presence, giving them the opportunity to revisit the diagnosis and reduce the likelihood of error.

2. Increased efficiency.

There are two ways in which AI can make diagnosis more efficient. The first is at the individual level; AI is a decision support tool that can help pathologists make diagnoses faster; you can think of it as a “thought assistant.” The second is at the lab level – using AI to optimize the lab workflow. It can help you ensure that you address the most urgent cases first, assign specific cases to specific pathologists, or even order stains in advance, all of which can have a significant impact on turnaround time.

3. A new frontier in cancer diagnosis.

Given a new ability to treat pathology slides quantitatively, what new insights can we glean from the data? That opens up huge opportunities for AI to support us in broadening our horizons. Computers can see things our eyes cannot – and that means they can allow us to make use of diagnostic information we could never access before.

Taking the plunge
 

“Artificial intelligence” is a big – and potentially scary – concept. It puts in mind the idea that technology might one day replace pathologists – a misconception not shared by most machine learning experts. But AI is really just another set of tools – one that happens to be digital. So what are the challenges you might face if your laboratory wants to implement AI?

First of all, you need digital pathology infrastructure, which goes beyond purchasing the right hardware and software products. You need a way to embed AI within your existing workflow. And you need to overcome potential resistance from your clinicians.

Some take that first step by implementing AI for case review. Basically, the algorithm performs an additional review – a second read - of all the cases going through the lab and raises an alert if there is a possibility of diagnostic error. It’s easy to bring on board because it doesn’t require any changes to the primary diagnostic process; it’s simply an add-on step at the end that acts as a “safety net.” Clinicians also really like this application because they tell us it “helps them sleep a bit better at night.” Within this framework, the subsequent expansion of the AI platform toward assisting in primary diagnosis – which may lead to additional gains – would become more natural.

Every day, I see misdiagnoses – cases in which samples have been reported as benign until our system alerts on a potential false-negative error, at which point they are revisited and diagnosed as cancer. Recently, I encountered a situation in which the cancerous cells were not in the area of the biopsy the pathologists were examining; as a result, they were missed until the AI algorithm alerted to their presence. We took the case to an advisor – Stuart Schnitt, a breast pathologist at Brigham and Women’s Hospital – who confirmed a diagnosis of breast cancer, but emphasized that it was a very subtle case. For us, that was incredible. Thanks to AI, the patient was able to avoid misdiagnosis and potentially a severely worsened outcome.

Thinking ahead
 

The transition to remote working has definitely accelerated the adoption of digital pathology – and of AI as well. The downside, though, is that many labs have been overwhelmed by the volume of COVID-19 testing they’ve been asked to do, making it a less-than-ideal time to deploy a new technology. My hope is that the acceleration of pathology’s digitization will help the discipline move more quickly and efficiently toward its future.

And what is that future?

It comes in two parts: the adoption of digital pathology and AI into diagnostic practice and our increasing ability to extract data from slides. The first is already happening around the world – we’re past the point of early adoption. There will be an inflection point where AI suddenly becomes a standard part of routine diagnostic practice. The second will be another modality that joins our existing troves of data – clinical data, genomic data, and now quantitative features extracted from slides by AI. Ultimately, having access to this kind of data, which enables new insights, will allow us to develop newer and better tests for not just diagnosis, but also prognosis and treatment selection.

There’s one other important aspect to this – and that’s the equity it brings to diagnostics around the world. Right now, if you’re a patient in a developed country with access to top hospitals and world experts, you stand a high chance of getting the right diagnosis and treatment quickly. But not all patients are in these settings – and not all pathologists have the luxury of working in these environments. Many work under severe time and resource constraints and more pressure, which unfortunately creates more opportunities for things to go wrong. The beauty of AI is that it can be deployed everywhere. In fact, it brings the most value in labs with higher volumes and fewer resources, because that’s where the likelihood of error is highest – so that’s where an extra layer of oversight can yield the greatest benefit. All patients should have equitable access to diagnostics – and, even though it seems counterintuitive, new technologies can pave the way.