Radiologists who use AI will replace those who don’t
There is a lot of hype and plenty of fear around artificial intelligence and its impact on the future of healthcare. There are many signs pointing towards the fact that AI will completely move the world of medicine. As deep learning algorithms and narrow AI started to buzz especially around the field of medical imaging, many radiologists went into panic mode. In his presentation at the GPU Tech Conference in San Jose in May 2017, Curtis Langlotz, Professor of Radiology and Biomedical Informatics at Stanford University, mentioned how he received an e-mail from one of his students saying he was thinking about going into radiology but does not know whether it is a viable profession anymore. But the assumption that the radiologist profession is dying, is just plain wrong.
Bradley Erickson, Director of the Radiology Informatics Lab at Mayo Clinic told me that some of the hype we hear from some of the machine learning and deep learning experts saying that AI would replace radiologists is for them looking at radiologists as just looking at pictures. That would be me saying while I look at programmers, all they do is typing, so we can replace a programmer with a speech recognition system, he added. Langlotz compared the situation to that of the autopilot in aviation. The innovation did not replace real pilots, it augmented their tasks. On very long flights, it is handy to turn on the autopilot, but they are useless when you need rapid judgment. So, the combination of humans and machines is the winner solution. And it will be the same in healthcare.
Thus, I agree with Langlotz completely when he says that artificial intelligence will not replace radiologists. Yet, those radiologists who use AI will replace the ones who don’t. Let me show you why.
What do cat intestines, X-ray lamps and the history of medical imaging have in common?
The field of clinical radiology started obviously with the quite coincidental discovery of the X-ray by Wilhelm Conrad Röntgen on 8 November 1895 in Würzburg, Germany. Within two months, the X-ray mania ran over the world. Sensational headlines in newspapers propagated the “new light seeing through flesh to bones”, while one inventor even speculated that “soon every house will have a cathode-ray machine”. Any similarities about hyped technologies coming to mind?
Thomas Edison became so excited about the new discovery that he even wanted to create a commercial “X-ray lamp” (unfortunately, his efforts failed) and tried to get an X-ray of the human brain in action (sadly, that was not a success either). His latter endeavor let story-driven reporters go nuts: they were allegedly waiting for the innovation outside his laboratory for weeks in vain. Some went as far as to fabricate images about the human brain. One of them turned out to be a pan of cat intestines radiographed in 1896 by H. A. Falk!!
While some early efforts turned out to be huge blows and impossible projects, X-ray got acclimatized in medicine. Something similar will happen with AI and healthcare soon. I hope with fewer cat intestines, though.
Radiology has been the playfield of technological development since the beginnings
In the TV series, The Knick depicting the first decades of modern surgery and healthcare, an inventor gets in touch with the hospital manager in his office to present him a new idea, the X-ray machine. It turns out, it takes an hour or so for the brand-new machine to take the picture! Currently, if you go to the hospital to do the annual check-up on your lungs, the X-ray procedure will take a couple of minutes in a fortunate situation, and some more until you get the results.
Plenty has changed since those experiments with the ‘X-ray lamp’, but one thing was constant: rapid technological development in radiology.
Bigger range of tools and higher precision
Approximately half a century after the discovery of the X-ray, the ultrasound joined the methods of medical imaging. From the mid-sixties onwards the advent of commercially available systems allowed a wider dissemination. Rapid technological advances in electronics and piezoelectric materials provided further improvements from bistable to greyscale images and from still images to real-time moving images. And it is also amazing to see how we went from room-sized, clumsy ultrasound machines to portable ones circa within another half of a century! In 2016, Clarius Mobile Health introduced the world’s first handheld ultrasound scanner with a mobile application. The doctor can carry around the personal ultrasound device for quick exams and to guide procedures such as nerve blocks and targeted injections.
Now, let’s look at body scanners. The first CT scanners were introduced in 1971 with a single detector for brain study under the leadership of Godfrey Hounsfield, an electrical engineer at EMI (Electric and Musical Industries, Ltd). The very first MRI scanner was built by Raymond Damadian in the 1970s by hand, assisted by his students at New York’s Downstate Medical Center. He achieved the first MRI scan of a healthy human body in 1977 and a human organism with cancer in 1978. The first functional MR imaging of the human brain is produced in the early 1990s. By the early 2000s, cardiac MRI, body MRI, fetal imaging, functional MR imaging became routine exams in many imaging centers.
Along with precision comes automation
Thus, the history of radiology shows the expansion of means as well as the increase in precision so far. While the latter is still in focus, there is also a visible shift towards making radiologists’ lives easier by automation. As radiologists need to go through more and more images every day, it becomes inevitable that part of their job can be automated. When we can train algorithms to spot and detect many types of abnormalities based on radiology images, why wouldn’t we let it do the time-consuming job so we can let radiologists dedicate their precious focus to the hardest issues?
When deep learning becomes possible and the algorithm could teach itself while radiologists rate its effectiveness, it’s going to get better just by working more. This is an opportunity we have to grab. This way radiology would be one of the most creative specialties in which problem-solving and the holistic approach would be the key.
So, it certainly would not mean that AI would take over all the tasks of radiologists. As Erickson put it, if you look at the frequency of findings and diagnosis on medical images, there are the common ones where AI could help, but there is a really long tail, uncommon but really important things that we cannot miss. He believes that it is going to be difficult for deep learning algorithms to identify those. But where do we stand with technology at the moment?
Could AI predict whether you would die soon?
Scientists at the University of Adelaide have been experimenting with an AI system that is said to be able to tell if you are going to die. By analyzing CT scans from 48 patients, the deep learning algorithms could predict whether they’d die within five years with 69 percent accuracy. It is “broadly similar” to scores from human diagnosticians, the paper says. It is an impressive achievement. The deep learning system was trained to analyze over 16,000 image features that could indicate signs of disease in those organs. Researchers say that their goal is for the algorithm to measure overall health rather than spot a single disease.
But this is just the tip of the iceberg. There is also plenty of on-going research to teach algorithms to detect various diseases. IBM launched an algorithm called Medical Sieve qualified to assist in clinical decision making in radiology and cardiology. The “cognitive health assistant” is able to analyze radiology images to spot and detect problems faster and more reliably. IBM’s flagship AI analytics platform, Watson is also utilized in the field of radiology. After the company purchased Merge Health in 2015, Watson got access to millions of radiology studies and a vast amount of existing medical record data to help train the AI in evaluating patient data and get better at reading imaging exams.
Not only IBM but also giants like Philips, Agfa and Siemens have already started integrating AI into their medical imaging software systems. GE is developing a predictive analytics software using elements of AI for the impact on imaging departments when someone calls in sick, or if patient volumes increase. Vital also has a similar work-in-progress predictive analytics software for imaging equipment utilization. Not to speak about the dozens of bigger and smaller start-ups trying to utilize the power of AI also for radiology.
But the on-going research does not mean that we are already at the stage where average patients will have to face their exact life expectancy based on their medical images when they go to the hospital.
What are the challenges in introducing AI to the radiology department?
In order to have some estimation when machine learning might be introduced on a wider scale, we have to look at how machine learning takes place in radiology. The process usually goes like this: the algorithm should be fed by thousands, if not millions of images and learn to spot differences regarding tissues. Just as in the case of computers recognizing images of dogs and cats. If the algorithm makes a mistake, the researcher notices it and adjusts the code. Thus, it is a rather lengthy process what needs tons of available data. Erickson believes that the result will look like the following: we’ll do the high volume exam, and the algorithm will probably create a structured, minable, preliminary report. So it will do the quantification, that most humans hate to do and it will do that very well, he noted.
Anna Fernandez, Health Informatics/Precision Medicine Lead at Booz Allen Hamilton told me though that there are several challenges in building these discovery and analytic platforms – from acquiring access and ingesting the data, sufficiently annotating the data, storage strategy, governance/policy use throughout, and types of analysis enabled via the platform. The biggest challenge is sufficiently annotating the data to allow different views of it (full right to owners, restricted subset to others) and enable discovery across the connected data sets in the platform.
Moreover, hospitals also need to be convinced that the AI algorithms work. Fernandez believes that it will be a step-wise process by for example taking advantage of hybrid internal and external “crowdsourcing” with sufficiently anonymized data.
For example, a vendor can have established data science algorithms based on anonymized data from their hospital network, then the new hospital can employ the algorithm and further refine it to the anonymized “local” data sets (that may include additional patient variables) to customize it to their population. As the hospitals see a “win,” they may be encouraged to release a more restricted anonymized data set to contribute back to the vendor solution. So it’s a little bit similar to how you try to go into the cold water on a hot summer day. First, you look at other people doing it, then you realize it’s safe, so you put your toes in the water before entirely going under.
When will we get to have AI analyzing our CT scans?
We are getting closer to clinical use day by day. The 2017 Data Science Bowl organized by Booz Allen Hamilton and Kaggle aimed at detecting lung cancer with the help of smart algorithms based on more than 1000 anonymized lung scans provided by the US National Cancer Institute. Participants developed 18,000 unique algorithms during the challenge. Fernandez told me that their ultimate goal was to identify the path to delivering the algorithms to the systems that can affect clinical care, thus they included members of the FDA along with the American College of Radiology connected to imaging system vendors and the radiologists who would adopt the use of these algorithms.
The FDA approved the first cloud-based deep learning algorithm for cardiac imaging developed by Arterys in 2017; so step by step we are getting there. She added that according to her estimations, within 3 years we’ll have many machine learning algorithms in active clinical pilot testing and in approved use. I would expect that within this time frame there will be a low dose CT lung cancer machine learning/deep learning algorithm output added to the radiology’s toolkit for assessing an individual’s risk of lung cancer (complement or added to LUNG-RADs or similar guidelines), she added.
Erickson did not mention a concrete estimation, he believes it will be a multi-step process where certain sub-fields will develop a lot faster than others, for example in mammography it is more likely to have AI sooner than in CT scanning. It has a potential for acceleration, and while it is unthinkable at the moment that it could create preliminary reports for everything in 10 years, there is a pretty good chance for it in certain fields, he explained.
Radiology’s Future is A.I.
All in all, research trends and experts underline how AI will revolutionize radiology in the long term. Thus, rather than neglecting it or feeling threatened by it, the medical community should embrace its achievements.
As Erickson put it, rather than pushing off machine intelligence as being a threat to their job, instead, radiologists should engage it, because it’s something that can really help patients. I’m sure it will dramatically change what radiologist do over the next ten years, but you should also keep in mind that eventually, radiology ten years ago was nothing like what it is today. So it is just one of those things where we need to make sure that we keep at the forefront; that we keep in mind that what matters most is taking care of patients. I could not agree more and could not express it better. Nurture that thought and let’s make a better future for radiology with AI!