Guest Post: AI vs MD – What it Means for Pathology
AI vs MD: What it means for Pathology
By David West, CEO, Proscia
Will computers replace doctors? It’s the question on everyone’s mind when reading Siddhartha Mukherjee’s article in the New Yorker “AI vs MD”, which pits human doctors against computers. It’s a beautifully composed summary on the role of software algorithms in healthcare, examining the issue through the most controversial lens. Perhaps some radiologists read it and had an existential crisis. Surely some machine learning Ph.D.s read it and rolled their eyes with the gross simplicity with which deep learning was described. It’s a shocking and provocative concept for any reader of The New Yorker – doctors are the best and brightest among us, even they could be outsmarted by a machine?
Mukherjee mentions pathology only once and it’s in a favorable light (“Notably, there was a set of two thousand lesions that had also been biopsied and examined by pathologists, and thereby diagnosed with near-certainty”). But with the first FDA approval of WSI announced last month and a trend towards digitization making pathology more and more like radiology, we ask, “What are the implications of AI for pathologists?”
In Dr. Kaplan’s Part 1 response to “AI vs MD” on the Digital Pathology Blog, he touches on the role of the pathologist – does the pathologist “diagnose” or do more? I’m a technologist, and will keep myself out of that debate, but I think it’s important to understand what role software algorithms play in pathology.
Why All The Hype?
It’s no secret there’s a lot of buzz around AI these days, especially in medical imaging. In Part 1 of his response, Dr. Kaplan pointed out:
This article in particular has many facets addressing the use of computer aided technologies, be it “artificial intelligence” or “machine learning” or “deep learning” or “deep learning automated analysis” and so forth. Very popular buzz words right now to optimize your SEO rankings. All the rage, IBM, Google and dozens of other companies, big and small are creating warehouses of data – clinical, molecular, -omics, images, outcomes, treatments, diet behaviors, social behaviors, etc… and trying to find pieces that fit together to be personal, predictive and pre-emptive. This was the official war cry of the NIH I think more than 25 years ago – “the 3 P’s” as I recall as a resident at Bethesda Naval Hospital across Wisconsin Avenue from NIH.
Why is this? After all, Artificial Intelligence has been a concept for decades loosely defined as computers mimicking cognitive function. We’ve gone through periods of massive investment and experienced multiple “AI-Winters”. Methods for learning from past data to make decisions or predictions about the future have been discovered, studied, and utilized before computers even existed. So why is everyone talking about this now?
The hype comes largely from a confluence of two advancements that happened very recently, as in the past six years:
- The first is processing power, democratized by cloud computing, which made cheap and available the computing resources required to train large network-based algorithms.
- The second is a new machine learning method that is light-years beyond previous methods. In 2011, three scholars used a new kind of machine learning to win an international image classification competition. Alex Krizhevsky and Ilya Sutskever (under the supervision of Geoffrey Hinton) didn’t just win the competition, they beat five other prestigious teams in dramatic fashion. These new Convolutional Neural Networks, or CNNs, opened up a new field of deep learning. (For more on this, Nvidia has a great history of deep learning here)
We have yet to see all the effects of these advancements in our lives, but what’s particularly relevant to pathology is that deep learning lends itself very well to images and image classification. Combine this with a trend towards laboratory digitization, and suddenly people care about AI in pathology.
You can expect to hear more about this, and maybe even expect the buzzwords to get old at some point, but only because technology adoption cycles are longer than media cycles, and because it may not take the form that many of us expect. Labs will likely never purchase “artificially intelligent pathologist robots” to take the place of human experts. So what will machine learning look like in pathology?
What AI Might Look Like In Pathology
“Intelligent” algorithms for pathology are very much in the works, and things are moving more quickly than most imagine. There are a number of software algorithms in development, serving many different functions, but generally falling into the categories of a) automation or b) augmentation. There are some instances where the line between these categories is very fine, but existing regulatory commercialization frameworks are our best clue as to how things will look in pathology.
On the augmentation front and perhaps in automation, it’s likely we’ll see many intelligent algorithms enter the market as image-based diagnostics–a far cry from the pathologist robot that is often portrayed. Proscia’s CSO Hunter Jackson wrote a piece on image-based diagnostics assays, comparing algorithms that operate on images (with standard histology prep) to molecular assays already in practice, but filling the gap where molecular biomarkers fall short.
The “AI vs MD” article points out a precedent in radiology that could suggest what automation algorithms might look like: “Pattern-recognition software highlights suspicious areas, and radiologists review the results.” Software algorithms for pathology are already quite good at this, with companies already having developed technologies that can identify metastatic tissue with very high sensitivity and specificity. And the algorithms can even go beyond this by predicting outcomes. Notably, when combined with software, human pathologists outperform both the software-only and human-only systems. At the end of the day, even automation takes more of a form of augmentation.
If you’re a pathologist, take assurance that your profession is here to stay despite the technological evolution that is taking place. If anything, with anticipated demand-supply gaps of 5,700 pathologists over the next decade, there will be a greater need for more pathologists as software begins to play a larger role. The promise of exciting careers in a cutting edge field with new technologies at their disposal will attract the best and brightest residents to pathology. Embracing artificial intelligence will put laboratory medicine even closer to the center of the patient experience. There are many forward-thinking pathologists, young and old, who are working towards the future, recognizing that the field has changed in the past and will continue to change.
As internet pioneer and venture capitalist Marc Andreessen said, software is eating the world. Technology-driven shifts often start quietly and are over before we can even blink our eyes. Initialing change in healthcare is remarkably challenging and doesn’t happen overnight, but there are some exciting developments to look forward to in pathology as it dives into the digital realm.
——
David West is the Founder and CEO of Proscia Inc, a digital pathology software company bringing computer intelligence to pathology.
