A few weeks ago, we wrote about the possible implications of robots assisting with surgeries. But the operating room isn’t the only place where rapidly advancing technology has the potential to produce serious disruptions. Pathology could be one of the next proving grounds for artificial intelligence (AI).
AIs have been making news all over the place lately, with varied results; from the simple and goofy, like the neural network that created a series of new paint colours and proceeded to give them outrageously goofy names like Snowbonk, to the surreal and abstract, such as Google DeepMind’s foray into artificial dreaming, to the disturbing example of Microsoft’s AI Chatbot that quickly turned racist after spending less than a day on Twitter, AI certainly seems to be poised as the next big breakthrough in technology. And every day there seems to be some new news article about it. Whether it’s scientists like Stephen Hawking or celebrity entrepreneurs like Elon Musk warning of the threat of an AI super-intelligence, or controversies about massive and widespread unemployment in the wake of faster, cheaper and more efficient robots, the internet is brimming with articles about the advances of artificial intelligence.
Machine Learning and Healthcare
When most people think of artificial intelligence disrupting industries these days, they think of self-driving cars or factory robots taking on more complex and complicated assembly procedures. What most don’t realize is that artificial intelligence is already having a dramatic impact in places they never expected it too, even healthcare. And while some may currently view the role of the artificial intelligence and machine learning in healthcare to remain relegated to primarily simple, repetitive support tasks, such as mining medical records or managing medications and prescriptions, the pace of advancement of artificial intelligence is staggering. Some experts predict that machine learning will be widespread by 2025 (if not earlier). And with systems like IBM’s Watson Health and Google’s Deepmind Health already making a splash in healthcare circles, it’s not hard to understand why.
And now, it’s pathology’s turn to face the specter of artificial intelligence and machine learning. The Journal of the American Medical Association recently published a study detailing the results of the CAMELYON16 challenge, a two-year research project that pit multiple deep-learning algorithms against a team of 11 skilled human pathologists in a two-hour contest to review and accurately assess over a 100 different whole-slide images with the goal of identifying both macro- and micro-metastases in the images. The results? The top-performing algorithms were able diagnose micro-metastases more accurately than their human competitors.
Of course, the CAMYLEON16 environment was by no means a precise replication of how pathologists work, particularly considering the unrealistic timetable. In fact, when the 2hour time constraint was lifted, the top-performing algorithm performed no better than its human competitor. And in addition to the tight schedule, there were plenty of other discrepancies. For example, pathologists would almost certainly ask for special stains, more sections and additional information when dealing with unclear findings, none of which was available to the team.
Currently, no one seriously expects the machines to be coming for pathologists’ jobs anytime soon, and with good reason. The most obvious reason is currently cost; the ability of algorithms to accurately diagnose, while impressive, isn’t necessarily an indication of added value, and that is absolutely requisite for anyone, pathologist or otherwise, to make a serious investment in the technology. Until pathology algorithms prove their value, they’re unlikely to be a serious threat to trained experts.
What’s more, many see the rise of artificial intelligences to be disruptive to pathology workflows, but not necessarily destructive, treating them as tools that will improve pathologists’ efficiency and accuracy by letting them spend less time interpreting slides and more time synthesizing data from disparate sources into personalized diagnoses for patients.
Given these issues, it’s fairly obvious that real, live pathologists aren’t going anywhere for now, but the CAMYLEON16 study clearly demonstrated that advancements in AI and machine learning may soon be a significant part of the pathologist’s laboratory. What’s more, CAMYLEON17 is already underway. This time, the organizers aim to create a challenge more clinically relevant, pushing the boundaries of pathology algorithms even further. But regardless of the outcome, it’s clear that machine learning and artificial intelligences are virtually guaranteed to have a role of some kind in the pathology laboratory in the future. And that future may be nearer than we think.