Bottom-Up to 2020
How can digital champions rise up and lead the way with AI for primary diagnosis?
“When will we see mainstream adoption of artificial intelligence (AI) in pathology?” It’s a question I hear at every digital pathology meeting across the world – but one that has no definite answer. I know of a handful of labs that use AI every day; some use it prospectively before the pathologist sees the slides, whereas others use algorithms retrospectively to confirm diagnoses. Although small in number, these labs demonstrate that there is a real appetite for AI in pathology and that it can be used successfully. But exactly how long do I think widespread adoption will take? In my view, at least five years. Many people speak of a “third revolution” in pathology; however, before labs can capitalize on AI, it’s crucial to first have a proper digital pathology platform – a key obstacle for many slow adopters.
Our advantage at the University of Pittsburgh Medical Center (UPMC) is that we already have a digital pathology platform in place. Although the journey to a fully digital workflow is difficult for any lab (not least because the chance of reimbursement is initially minimal), one of AI’s benefits is that it gives clear justification for the implementation of a fully digital platform. The combination of AI and digital pathology makes it a much more exciting journey and increases the chance of buy-in from both pathologists and non-pathologists.
The UPMC healthcare system has 41 hospitals and travel between them can take up to five hours. We run a very distributed operation, with generalists on the periphery and academic medical centers in the heart of Pittsburgh, so we often move difficult cases to the central hospitals. AI will support our distributed model by providing expertise to community pathologists, preventing the need to routinely send cases elsewhere. I am particularly excited by the increased accuracy that AI promises for every single case – and the potential cost savings that result.
As an academic medical center, AI’s potential in research is an appealing opportunity for our staff. We’ve already started to probe images with AI tools to make discoveries not possible through human investigation alone. We have also experienced the reputational benefits of AI. We now commonly have prospective residents ask us whether we are implementing AI. Doing so makes ours a highly attractive program and increases enrolment. Residents want to be trained for the future.
Pathologists are often scared to be the first to do something because of the possible risks (bad press, medico-legal risk, and potential patient harm). For these reasons, we tend to follow the flock. And that’s where I think pathology colleges – both in the US and around the world – could do more in terms of setting guidelines for the use of AI. Leading the charge in this way would endorse the adoption of AI and prevent it from being perceived as a rogue activity. And, crucially, it would provide some much-needed clearance on regulatory methods.
Because there isn’t a wealth of experience to draw upon when using AI, it’s important to have guidance on validation – something I struggled with when validating an algorithm for clinical use at UPMC. I was unable to seek advice from other labs, so I contacted various organizations to ask what guidelines I should follow. There were none. By applying good scientific and laboratory practice, we successfully ensured that the algorithm is safe to use at UPMC – but it certainly delayed the validation process. College guidelines would undoubtedly simplify and expedite the adoption of AI and ensure that everything is standardized for safe practice. We faced a similar hurdle when whole-slide imaging first became commercially available; many colleges stepped up and provided guidelines for validation, which improved adoption and made pathologists feel more comfortable using the technology. That should act as a precedent for AI.
Our ultimate aim is to implement AI for primary diagnosis in 2020. We know that will be a challenge – not least because some of our pathologists are determined to stick to traditional microscopes – but we have noticed an increase in requests for digital pathology tools. Instead of a top-down approach, where digital pathology would be forced upon our staff, we’ve instead opted for a bottom-up strategy that allows champions to rise up, request digital pathology, and start using AI for routine diagnosis. Such an approach will allow a much smoother transition with stepwise changes across different subspecialties – and we hope to reap its rewards in the near future!
Professor of Pathology, Department of Pathology and Clinical Labs, University of Michigan, Ann Arbor, Michigan, USA.