Latest News

How AI Will Support – Not Substitute – Pathologists

Contributed Commentary By Ralf Huss 

May 10, 2017 | A study that appeared in New England Journal of Medicine a few months ago has caused ripples throughout the industry about whether artificial intelligence has the potential to replace the work being done by pathologists within the next few years. The study, conducted by Drs. Obermeyer and Emanuel, raises some great points, namely that AI and machine learning bring powerful improvements to the profession, with algorithms that can “reliably predict outcomes, improve prognosis…and improve diagnostic accuracy.” But when I see articles with headlines like, “Why AI is about to make doctors obsolete”, I have to disagree. Many are viewing the AI revolution as the end of the pathologist, but in actuality, AI has the ability to help pathologists and make them even more effective at their jobs.

How AI Can Help

Pathologists face a number of challenges both in the clinical trial and clinical practice settings. First, big data provides big opportunities for breakthroughs, but it’s a challenge to wrangle. In order to make the best decisions, pathologists need to collect and analyze all the available information on a patient, making use of electronic health records, previous treatments, genetic backgrounds, etc., and correlate that with any relevant data from other labs or databases.

Machines are capable of mining and crunching huge amounts of data and identifying patterns that humans cannot, giving us a way to deal with the the vast amount of data that’s being accumulated. Artificial intelligence and machine learning thus provide significant opportunities for advancement; if you take enormous computing power and feed it tremendous amounts of data, you end up with an AI network that pathologists can interact with in a very helpful way.

Not only do pathologists have to contend with analyzing an increasingly overwhelming amount of data and samples, but many studies have emphasized the high inter- and intra-observer variability in anatomical pathology, which can lead to an alarming frequency of diagnostic errors, putting patients’ lives at risk. Pathologists are expected to effectively support clinical decision making, and machine learning will help improve diagnostic accuracy and quality control so they can have more confidence in what they’re communicating and the guidance they’re providing on appropriate treatment options. Algorithms will soon generate differential diagnoses to guide decision-making, and suggest relevant tests that confirm the working hypothesis while ignoring those which will have no value. If the system excludes or confirms a hypothesis, it will make the pathologist much more comfortable and confident in making a decision.

Pathologists are also increasingly being requested to predict therapy response, which is not only about identifying biomarker over-expression, but also the absence of certain markers in the context of genetics. For example, a patient with highly inflamed colon cancer with mucus has a high likelihood of having a genetic predisposition; when no mucus is present, the patient is unlikely to have a genetic predisposition, which would change the guidance on treatment—but this is all we really know so far. With more data and machine learning, we can learn so much more.

In the NEJM study, Drs. Obermeyer and Emanuel predicted that machine learning will dramatically improve prognosis because it can identify morphological features, or phenes in the tissue, that have been largely ignored so far, or could not have been detected by simple pathological insights. The availability of big data sets and machine learning programs that can mine and learn from them will make it possible to validate such new “signatures,” which will become clinically meaningful prognostic algorithms in the not-so-distant future.

No Need for Fear

It’s not just the NEJM study or magazine articles that are suggesting AI might make pathologists obsolete. I was recently at a workshop discussing digital health and EHRs when machine learning came up—and it’s clear there’s a lot of angst among pathologists on the topic. However, we have to remember that we’ve already seen big evolutions in pathology before, all of which have improved—not replaced—the pathologist’s role.

In the early days of pathology, we relied only on what we could see with microscopes. Then molecular pathology arrived on the scene. And for years digital pathology companies have provided technology that automates tissue image analysis and "sees" things in the tissue that the human eye cannot, essentially, automating big pieces of the pathologist's job. There was fear by some early on that this kind of technology might replace pathologists. But on the contrary, it has helped pathologists find answers to their questions faster, and do more with less time. Pathologists are still needed to guide the technology, and ultimately it has enabled the industry to identify biomarkers and develop precision drugs and diagnostics. It just required a shift in how the job is done.

AI is simply the natural next step in the evolution of pathology. It’ll just become part of the way we work, an interface running in the background of our computer systems like an Apple OS, or a tool like Microsoft Excel. We should embrace it. It will make our jobs easier and will improve our services to patients in the end, because they rely on the fact that we’re considering all the available data, and using AI is the only way we can effectively mine that data to ensure we are making the right decisions about diagnosis and predictions on therapy response. That’s not to say we shouldn’t take a critical view of anything suggested by the system—just as you’d double check any “fact” you find on the Internet, we need to do a plausibility check of anything the machine learning system spits out—but having access to significantly more information by which to make a decision will certainly be helpful.

In the end, AI will will not put pathologists out of work, it will simply change the processes. I, for one, don’t intend to become a machine learning expert, but I do intend to use it.

About the Author

Ralf Huss is Chief Medical Officer of Definiens and has more than 20 years of training and experience in histopathology and cancer research. Prior to joining Definiens, he also served as Global Head of Histopathology and Tissue Biomarkers at Roche Diagnostics. He can be reached at