Demystifying AI for Pathologists – and for Patients
Pathologists face a wide variety of challenges, and thoughtful application of nascent artificial intelligence (AI) has the potential to address specific use cases. Balancing the future promise of AI against skepticism amongst clinicians, AI practitioners must embrace an approach of humility and partnership when supplementing pathologists’ tried-and-true workflows
Chad Salinas | | Interview
sponsored by Leica Biosystems
What are the potential benefits of AI?
In the future, AI may help pathologists in several ways. Foundational AI could answer basic questions like: is there a tumor in this whole slide image? And, if so, is it benign or malignant? If it is malignant, can we assist the pathologist by highlighting the pixel region(s) most indicative of malignancy? This may save a lot of work because much of what pathologists see is benign – and a quick slide view can confirm the computer’s verdict. In addition, huge global discrepancies exist regarding patients’ access to pathology services. When the nearest fully equipped and staffed pathology department is thousands of miles away, AI may reduce workloads and offer patients quicker turnaround times by helping local medical professionals determine which cases require further expert examination.
I think there’s also a real opportunity for behavior-changing AI. For instance, few pathologists want to be hunched over a microscope counting mitotic figures any longer than necessary. In contrast, AI never gets tired, so it can rapidly count all the mitotic figures in a given area – changing the diagnostic approach from “sufficient counting” to “comprehensive information.”
AI technology should enhance and extend pathologists’ capabilities and potential, not seek to replace pathologists. I like to say to pathologists, “AI isn’t here to out-doctor the doctors.” If we collaborate closely with pathologists to understand their existing workflows, we can add value – not with a killer application, but by giving them a portfolio of tools to aid their diagnostics.
Why should a pathologist trust AI?
I see two sources of skepticism regarding AI. The first is experience; some pathologists who don’t use or may not fully understand the technology are hesitant to adopt it. The second is haste; many vendors approach the market so quickly and with such extensive promises that pathologists may become apprehensive. When radiology first digitized, some computer scientists proclaimed, “You don’t need to train radiologists anymore; you can just let the machine do the job.” In the years since this proclamation, more radiologists have joined the field than there were prior to digitization (1).
It is important to demystify the “black box” of AI. We need to find the right balance between transparency and protection of privacy and intellectual property. Through a candid dialogue between industry and the pathology community, we can create trust around AI and its applications in pathology.
Much of the AI community comes from a mathematical background, so we provide mathematical proof that our algorithms work – but that might not be the right approach for pathologists. Pathologists will likely want to see AI tried, tested, and true. For example, once an algorithm has reviewed 10,000 slides and evidenced no false positives or negatives, pathologists can have confidence in that algorithm.
A “thought experiment” I like to offer is as follows: let’s say you’ve been offered the opportunity to fly to the moon on an experimental rocket. If you only have mathematical proofs assuring you that the rocket won’t fall out of the sky, will you still board it? Or will you wait for a rocket that has made the trip 10,000 times? Most of us would choose the latter. AI in the healthcare space has similar stakes; an effective AI must prove itself over and over before we feel comfortable applying it to diagnostics.
What’s next in the AI space?
There are significant differences between humans and computers. Humans learn broadly; we can learn from small datasets and transfer our learning to other domains very quickly. Machines learn differently; they need a lot of examples of each case before they can recognize patterns – and, even then, the tasks they perform are quite narrow. AI is not a silver bullet; it’s a tool to augment the pathology workflow.
In the future, we hope to see more behavior-changing AI – not just for diagnostics, but also to help identify the best possible treatment for each patient and to provide some information about potential outcomes based on previous patient data.
Developing AI is not a simple task; there are countless variables for the industry to consider as we work toward integrating AI into the pathology workflow, four of which I outline in a blog post on the Leica Biosystems website (2). However, if we maintain a position of humility and partner with pathologists, I’m optimistic that AI can truly make a difference in advancing healthcare.