Attribution: ImageCreator – http://www.imagecreator.co.uk/
As I walk into my office each day I wave hello to a plethora of staff, all of whom contribute to the well-oiled machine that is a Nuclear Medicine department: Technologists, admin, support, physics and, of course, radiologists.
As I sit down to work on my long-term projects, it strikes me that their successful completion will inevitably result in job losses for some of these friendly faces.
This first dawned on me when I set up image reconstruction protocols in our image processing software. The aim was to simplify the procedure, so technologists could reconstruct and deliver the ideal image for reporting to the radiologists. It was a relatively simple process generating a series of steps technologists could easily follow which reduced the workload on the radiologists.
This is seen, rightly, as a positive. Reducing the burdensome but ultimately repetitive tasks from radiologists’ workload should allow them to use their expertise in more high-value tasks and patient facing roles. However, there comes a point at which the repetitive tasks are so automated that maintaining the radiologist workforce as is becomes economically unjustifiable to employers.
Technologists remain indispensable – no immediate threat of job losses
As with many other branches of medical imaging, the last 20 years have seen an explosion of automation. Pre-set scanning and reconstruction protocols, underpinned by black-box algorithms (usually a version of expectation maximisation) have removed clinicians and technologists further from influencing the final images than ever before.
Tube-current modulation in CT has minimised the need for radiographer decision-making. Acquiring an optimal image has become the work of computers, programmed and set-up by engineers and physicists, requiring only the push of a button as operator control.
Technologists, however, remain indispensable. Although much of what happens while patients are on the couch is now out of their hands, especially in nuclear medicine, much else is difficult to imagine being automated. The simple human interaction required to calm patients during injection, weighing or stress testing shouldn’t be replaced.
Radiologists too remain confident their interpretation skills will protect them from job losses. But is this confidence misplaced? There are good reasons to think it may be.
Innovation is propelled by demand in the UK’s National Health Service (NHS)
Innovation can be propelled by a number of different factors: brilliant individuals; high demand; or as I suggest, sheer necessity. The UK’s National Health Service (NHS) is an excellent precursor for healthcare around the world. As a highly streamlined semi-centralised national healthcare service, it often encounters issues related to high volume which also manifest in other countries. This is supported by EU data on bed occupancy, a highly contentious issue in UK healthcare. Of 27 EU countries, 22 saw their bed availability drop and just five saw (mostly modest) increases.
What can this trend of demand tell us about the likelihood of AI leading to radiologist job losses?
The Care Quality Commission (CQC), the independent regulator of all health and adult social care in the UK, recently released a review of radiology nationally [1]. This was largely in response to information exposing a large number of unreported scans in a southern NHS trust. Over 23,000 scans had gone unreported in a period of just 12 months. They were not the only ones; a London based trust also had over 21,000 unreported scans on its books across two sites.
This is endemic of the problem across the NHS. The volume of scans is simply too high for the current workforce to report on properly. The same CQC report stated: “97% of radiology departments in the UK […] were unable to meet reporting requirements”. This is a national problem which will only get worse.
Already we are seeing images being sent for reporting by untrained referrers. As more studies come out proving the proficiency of neural networks in reporting the majority of scans, struggling NHS Trusts may have no choice but to turn to these algorithms as a solution for the reporting problem.
Human interpretation is imperfect, no matter how skilled the human is. As such the temptation for NHS trusts, struggling to meet demand, to turn to AI is strong.
And why not? Currently these networks achieve somewhere in the region of 80-90% agreement with clinicians in a range of fields including brain tumour MRI’s, lung nodules from CT’s and PET scans of the brain [2].
There are still improvements to be made. Ideally these algorithms will highlight when they are unsure of their conclusions and become better at imitating the volumes that clinicians identify. However, these challenges are surmountable and the ability of neural networks to train is likely to overcome these issues given diverse enough data sets.
Compounding the appeal of automated reporting is the inconsistency in reporting by clinicians. It has been shown that, depending on the country, rates of false-positive reports in, for example, mammography are high (>12 % by some measurements) [3].
This echoes studies into the consistency of “contouring” in radiotherapy between clinicians, sometimes resulting in dose differences of >50% to organs at risk [4].
These studies simply prove what is already commonly known: human interpretation is imperfect, no matter how skilled the human is. As such the temptation for NHS trusts, struggling to meet demand, to turn to AI is strong.
Long-term pressures on the system and improvements to software will force movement in this field, and when it does the wave of change will be rapid
Once the technology is proven there will be no need to retain workforce purely for reporting. One radiologist could do the work of 10 with some reliable automated segmentation software.
So how will they transform their role to be useful in this new reality? Radiologists will still be useful pillars of responsibility, retaining the “practitioner” role outlined in legislation and authorising scans and operators. Some will be retained to focus on patient facing aspects of their roles, but it seems highly unlikely this will be enough to prevent job losses, given the required workforce to cover these roles coupled with fees demanded by such highly skilled professionals.
The outlook is not totally bleak. Currently workforce demands make immediate job losses from AI unlikely. Similarly, until the software is proven to be sufficiently capable of identifying its own short-falls, radiologists will be required to second check any reporting done in this way.
But long-term pressures on the system and improvements to software will force movement in this field, and when it does the wave of change will be rapid.
What are your thoughts on this article? Feel free to comment in the box below, or email responses to: [email protected]
If you want more information about how AI will impact radiology, then download AIMed Magazine issue 03 on Medical Imaging & Biomedical Diagnostics here.
Benjamin Fongenie is a Trainee Clinical Scientist in Medical Physics with the North London Medical Physics Training Consortium. He is specializing in Imaging with Ionising Radiation. He has a Masters in Physics from the University of Bristol, and is studying for a Masters in Clinical Science (Medical Physics) from King’s College London.
References:
[1] C. Q. Commission, “Radiology review A national review of radiology reporting within the NHS in England A national review of radiology reporting within the NHS in England,” 2018.
[2] H. Dong, G. Yang, F. Liu, Y. Mo, and Y. Guo, “Automatic Brain Tumor Detection and Segmentation Using U-Net Based Fully Convolutional Networks,” p. 1, 2011.
[3] H. D. Nelson, E. S. O’Meara, K. Kerlikowske, S. Balch, and D. Miglioretti, “Factors Associated With Rates of False-Positive and False-Negative Results From Digital Mammography Screening: An Analysis of Registry Data.,” Ann. Intern. Med., vol. 164, no. 4, pp. 226–35, Feb. 2016.
[4] B. E. Nelms, W. A. Tom E, G. Robinson, and J. Wheeler, “VARIATIONS IN THE CONTOURING OF ORGANS AT RISK: TEST CASE FROM A PATIENT WITH OROPHARYNGEAL CANCER,” Radiat. Oncol. Biol., vol. 82, pp. 368–378, 2012.
There is nothing new here as new technology has been impacting jobs and professions for thousands of years or more. The technologies evolve, and the cycle-time accelerates, reducing the time we have to understand and absorb the impact and some technologies (such as AI and robotics) may present more radical and fundamental challenges than others, ethical, economic or social.
It might be useful to stand back and look at other professions also facing radical change and job losses from automation and new technology, AI or otherwise. This is the case for air traffic controllers and pilots, truck drivers and policemen, and I’m sure other healthcare professionals such as those in primary care such as GP’s in the UK. Human error is the biggest single cause of airline accidents and passenger deaths and there is evidence that removing the pilot and flying the plane remotely (like a military drone) would now be a safer option. This is even clearer if we add-in deaths resulting from kamikaze crashes instigated by a pilot or onboard terrorist taking control of a plane, 9-11 being the worst example with almost 3,000 killed. If we assume this to be the case now or soon, why would we still have pilots when they could be replaced by safer (and most likely cheaper and more reliable technology)? There are several possible economic, social and psychological factors:
1. Sunk costs: all passenger planes are designed to be flown by pilots so removing the pilot would mean a lot of prohibitively expensive re-engineering of existing planes
2. Mixed mode risk: new pilot-less planes would have to co-exist in the air beside piloted ones for many years creating a risk of collision due to unpredictable pilot behaviour. This is the same problem for driverless cars and one reason that makes developing a driverless car so challenging and expensive- they must deal with all irrational driving and pedestrian behaviour from day one.
3. Attachment bias: research suggests passengers prefer to see or hear a pilot onboard even when they understand that a pilot-less plane is a safer alternative. This bias is deliberately enforced by airlines and pilots in numerous small ways, from the uniform and appearance to the tone of voice and language used. As humans, we have a deep need to know who is in charge, especially in unfamiliar or dangerous situations, situations where our dependency on others is high. Removing the human face, someone who can immediately explain what has happened and why in unusual circumstances, could have significant and unknown consequences on passenger behaviour on planes. No technology can replace this role today, or soon. Maybe the offer of free or very cheap travel would persuade some but AI does not offer this, the cost savings are too small to change the fundamental economics of air travel.
4. Job protection: pilots are often unionised and would fight to retain their jobs and status – many also like their jobs, they live to fly, not always the case for other professions!
Almost certainly there are more. It’s also worth noting that globally there is a shortage of trained pilots as passenger air traffic is rapidly growing and I understand the same is true globally for radiologists, certainly in the UK.
If we assume that AI can now outperform radiologists in some areas why would we as patients not prefer the AI to the radiologist in these areas? Patient power is not well developed in the UK, but it is real and growing, aided by increasingly open data, consumerisation of healthcare provision (patient first), digital devices and access to personal medical data, social media and a host of other techs, regulatory and social drivers. This is clearly evidenced by the rising number of legal cases brought against NHS Trusts and the increasing payments made by the NHS to successful claimants. The UK Government has set aside at least £65billion for such claims according to some analysis and the sum is set to rise. These legal claims may not be on the same scale as in the USA, but the trend is clear; patients are not willing to accept second-rate care and clear errors, whether of diagnosis or treatment. In this scenario, it is not tenable that hospitals will persist with human intervention in situations where automated procedures clearly outperform existing human ones, especially if also cheaper. This is particularly true in the UK where the NHS is a perennial political and social topic given a rapidly ageing population, a clear and growing expectation vs. delivery gap and a significant funding shortfall.
Unlike pilots, radiologists have almost no visibility to patients, at most a face behind a machine, an important cog in a complex machine fronted by GP’s and consultant clinicians: there is no social attachment or bond to be preserved or broken. Unlike with pilotless aeroplanes or driverless cars, there is also no technical contention or reason why new proven AI couldn’t be phased-in over time, gradually replacing radiologists in some areas – relieving recruitment shortages, saving money, improving performance, enabling more time for analysis where required – too often not the case today. This would also allow time to evolve roles, training and expectations of those impacted. Some radiologists may resist but I find it hard to imagine that when presented with the hard evidence, many would want to continue to provide an inferior patient service to that available with AI: they may not take the hypocritic oath (who does?) but I suspect not doing harm is instinctive in radiologists as it is in other health professionals. Imaging volumes and applications enabled by new tech will continue to rapidly grow and the professional opportunities associated with this growth significant. I think most radiologists will make a reasoned response and accept the inevitable and start to plan accordingly, especially those newest to the profession. By taking control of how and when to deal with the inevitable and preparing radiologists both old and new to 21st century digital realities it will continue to be possible to have a successful (if different) career in a world where less than 50% of the global population has access to any form of professional healthcare and the diagnostic services radiologists provide.