Q&A: The Radiology-Pathology Merger

For the past few years, there have been murmurs throughout the industry that the future of radiology includes a merger with pathology. The specialties share several characteristics as predominantly information-gathering professions. But, recently, another reason has emerged that supports this mesh – the advent of greater artificial intelligence and deep learning.

Diagnostic Imaging spoke with Eric J. Topol, MD, director of the Scripps Translational Science Institute, and Saurabh Jha, MBBS, MRCS, MS, associate professor of radiology at the Hospital of the University of Pennsylvania, about their recent editorial in the Journal of the American Medical Association about the looming possibility of a merger between radiology and pathology, as well as the role artificial learning can play in this future scenario. Here are their thoughts.

DI: Why are we seeing this big push – especially this year – for greater adoption of artificial intelligence?[[{"type":"media","view_mode":"media_crop","fid":"55092","attributes":{"alt":"","class":"media-image media-image-right","id":"media_crop_1597867393688","media_crop_h":"0","media_crop_image_style":"-1","media_crop_instance":"6897","media_crop_rotate":"0","media_crop_scale_h":"0","media_crop_scale_w":"0","media_crop_w":"0","media_crop_x":"0","media_crop_y":"0","style":"height: 267px; width: 200px; border-width: 0px; border-style: solid; margin: 1px; float: right;","title":"Eric J. Topol, MD","typeof":"foaf:Image"}}]]

Topol: This has obviously been talked about for many years. It goes back decades that the machine or computer would take on greater functions, but only in recent years have we seen a big jump in capability that’s demonstrated, including the ability to read X-rays or slides as good as or better than doctors. There’s been a cluster of new reports coming from first-rate centers and investigators that have proven what’s previously been theoretical. The other thing – just to mention the other reason that this has become a reality – is because start-up companies, as well as tech titans are really building up for this capability. You have companies like Enlitic and Merge, that was acquired by IBM Watson, that have shown us a market increase in efficiency and an interest in the business of taking this on.

Jha: I think machine learning has made tremendous progress and is coming into commercial use. There have been three trends. First, the progress in machine learning has, undoubtedly, been exponential over the past three-to-four years. Second, there’s a realization that a lot of what we do in radiology doesn’t require a high level of cognitive work. Radiology is becoming much more of a data-rich science than it has been in the past. When we used to look at a chest X-ray, there was an art to what we did. We had to extract as much clinical information as possible and decide what the problem in the chest or lungs was. That was subjectivity, and you needed to have someone who was willing to be subjective. Now, what’s happening progressively – once we went to CT – is that subjectivity has slowly reduced. We started to see things more clearly and started to see that things were really significant pathologies. That came with a price. The third trend is that price was that data simply increased markedly. Instead of having two slices with a chest X-ray, you have 4,000, and it keeps increasing. Now, we have quantifications, and none of it really requires a terrible amount of cognitive work. It’s just muscle learning. When you have something like that, it can be automated.

DI: What is it about radiology and pathology that seem to make them a good merge?

Topol: These two are truly the information specialties. They are purely diagnostic information specialties that are unlike other areas of medicine. They will also have the marked increase in the use of artificial intelligence and machine learning. But, most other specialties aren’t purely diagnostic, and there are parts of radiology and pathology that aren’t diagnostic, such as interventional radiology and lab medicine in pathology. But, the parts of radiology and pathology that are purely diagnostic lend themselves well to the coalescence.

Jha: The two are inter-dependent. We’ve always aspired to be like pathologists. There is a correlation between radiology and pathology – ultimately, there is truth in that what we do and say comes from that pathology correlation. The two fields have been working independently despite radiology having a certain reverence for pathology. The two fields are shrinking dramatically. For example, what radiologists often do before anyone was a biopsy of the heart, we often discuss whether it’s worthwhile because errors are possible. But, we can help point out where the biopsy can be taken, so that increases the yield for sure and provides feedback on the disease. In oncology imaging, with the pathology correlation, the needle sample has really refined radiology’s ability to signal progression of disease and get high yield specimens for pathology analysis. The two really must sit next to each other and work together. It’s culturally difficult because the radiologist is in the dark room, and the pathologist has been with dead bodies. That kind of barrier of where we work has to be broken down somehow.[[{"type":"media","view_mode":"media_crop","fid":"55093","attributes":{"alt":"Saurabh Jha, MBBS, MRCS, MS ","class":"media-image media-image-right","id":"media_crop_4346084765942","media_crop_h":"0","media_crop_image_style":"-1","media_crop_instance":"6898","media_crop_rotate":"0","media_crop_scale_h":"0","media_crop_scale_w":"0","media_crop_w":"0","media_crop_x":"0","media_crop_y":"0","style":"height: 269px; width: 200px; border-width: 0px; border-style: solid; margin: 1px; float: right;","title":"Saurabh Jha, MBBS, MRCS, MS ","typeof":"foaf:Image"}}]]

DI: How would the merge work?

Topol: It’s attractive to think we’ll need people who are specially-trained in computing in artificial intelligence and deep and machine learning. Not only are they trained in the specialty of radiology or pathology, but they also have the background so they can really use it to become the information specialists of the future.

Jha: It has to be an integrated department. The two fields are similar because their primary task is information extraction. They’re mainly non-facing fields. One of the major things that we’re going to have to face is the emergence of computer algorithms and machine learning. Will there be a role for non-patient facing clinicians or will machines and patient-facing physicians be sufficient? There will be a role if it’s carefully crafted. We believe there will be a role because it’s a lot of information for any one clinician to master. The two fields have a natural need symbiosis, and they will exist so long as they decide what their added value is going to be.

DI: What would change in practice for radiologist?

Topol: In JAMA, there was another editorial that talked about they saw the logic of radiology and pathology coming together. So, you can start to see how big of a deal this is. The editorial from a group at Harvard, and the line that caught my eye was that 260 million images a day could be read for the cost of $1,000. That’s what taking artificial intelligence to radiology could achieve. This wasn’t our idea. This is how they calculated it – 3,000 images per second. You could read all the images in the United States for $1,000 a day rather than billions of dollars. But, you still need oversight. You need radiologists and pathologists who are trained to understand the liability and caveats.

Jha: First, you have to be aware of disparate sources of information and correlate them with other sources. Understand the added value of any added information. Ultimately, any information means subtracting to enhance the signal. That’s what we’re fundamentally taught in imaging. The job is to sweep away the noise. It takes a lot of practice. For example, you start off with a patient who is a young athlete who faints. You do an EKG and find he has an EKG pattern seen in atypical cardiomyopathy. We can correlate that data from wearables or with genomics. Your job is to know when the situation is crying wolf and when information is pointing to a clinical finding. What tends to happen now is that we look at the image and say a patient has this condition or that without contextualizing it with the first or second source of data.

DI: What difference would there be for patients?

Topol: They would see an almost non-existent turnaround time. Everything has to be validated so that it’s highly accurate, but things could be more accurate with speed and cost for the patient. I hope that would be the case because patients, first and foremost, are entitled to getting their data. They should have their results and their scans. But, secondly, you hope that this machine integration of interpretation of medical scans will lead to lower costs. One of the biggest costs is professional costs, and now you have a machine doing that.

Jha: Patients who have rare diseases will see an improvement in their prognoses. It’s the nature of data to be able to capture rare disease. The flipside is what will happen for someone who is healthy. What chance do they have to be misdiagnosed as unhealthy? We need to have interactions between the patient and the clinicians. Our role is to give them the information so they can decide what to do with it. While we use our judgement to analyze images, the patient makes the ultimate decisions.

DI: What other benefits are there to merging these two specialties?

Topol: The benefits of this new era will be a whole new type of medicine. One where we combine the unique capability of machines with the unique expertise of humans. We haven’t even started really doing that. But, it’s time to wake up and realize we can be much more efficient. Rather than writing articles about doom and gloom and physician shortages and the cost of care, we have the solution in our sights that we’re not taking advantage of.

Jha: Whatever changes we make have to be made at the organizational level and the pedagogical level. People are starting to think about merging and starting off with something simple like radiology having a pathology integration in training. Our next generation of medical students could be exposed to both as an integrated specialty. The road to the merging of two fields is a very long road, and it’s one that has multiple stops. You have to stop at all the destinations on the road and regroup and reboot. The timing is right for this now, and we need to start thinking about it.

DI: Are there any downsides or challenges to this type of merger?

Topol: There are plenty of challenges. First, all of this has to be validated. Second, we’re asking the medical community to do something they’re not very good at – change. We’re very resistant to it. This is a very big threat to reimbursement because it challenges how the specialists get paid and reimbursed. We need to validate that this works really well at scale and that patients are really helped by this rather than find any evidence of harm. But, beyond that, we have to educate and provide new training for the specialty. We don’t even have this in existence today – an information specialty hasn’t been used before. It’s a new concept that we’re laying out, and there’s no training program for it yet. Continuing on to combining radiology and pathology is very attractive because they’re so well suited. We’re close to really zooming in on the sweet spot of where artificial intelligence was headed first. It’s going to hit every aspect of medicine, but this is one of the most appealing, ideal ways to harness the power.

There are always downsides when things are going toward big changes. You can have errors, but they can’t be tolerated. I fight concern with fact. We have millions of serious medical errors yearly in the United States. Not many are through scans, but we have to make those better. Obviously, one of the problems we have is with the interpretation of slides and, with pathology, there’s lots of heterogeneity. Some providers could read cancer, and others not. There are as many different opinions in pathology. Human oversight will still be important for years.

We still have to figure out a way to have human oversight, and that’s not been done yet. When IBM Watson first purchased Merge with the capability for artificial intelligence for scans, they didn’t want to replace the radiologist. But, in effect, we’re headed to that unless we figure out a better plan – unless we figure out how to integrate humans with machine capabilities.

Jha: There are going to be regulatory roadblocks. We have to have robust regulations in terms of how to know what artificial intelligence technology is doing what it claims to be. This takes some time to develop. The second is going to be a source of having artificial intelligence assistance – we already do that. I take advantage of these techniques, but there was a time when we didn’t take advantage of it at all. Any measures that improve the situation will take time to fully develop. The biggest challenge is going to be actually getting people to accept this without the fear that the sky is falling on their head.

The challenge here is to convince radiologists that their future is bright, but it’s not going to be the same. We can’t deny the role of artificial intelligence anymore. We can quibble about whether it will be 40 years or 20 years, but we have to prepare for a future in which artificial intelligence plays a significant part in the workforce. We have to decide what our role is going to be and accept that it’s not going to be what it is now.