How and why doctors’ attitudes towards artificial intelligence have changed: an interview with a radiologist.
OLEG BRONOV — MD at Celsus AI
Artificial intelligence systems for analysing digital medical images have come a long way in the last five years — from distrust and apprehension on the medical community part to implementation into clinical practice. The latter applies, above all, to the capital: since 2020, all radiology departments in Moscow have been using AI services as an experiment part by the Health Department. Moreover, in 2021, the Roszdravnadzor registered Russia’s first AI system for mammography and fluorography (CXR) analysis as a medical device.
However, what do doctors themselves think about it? About the first “encounter” with artificial intelligence and how its perception has changed since then, about the benefits for the doctor and possible risks from the AI use — told Oleg Yurievich Bronov, the highest category radiologist, medical sciences Candidate, the company scientific director that develops the AI system for radiology.
- Oleg Yurievich, when you first were exposed to artificial intelligence for radiology and what were your impressions?
- 6–7 years ago, I was at the annual conference in the USA — RSNA (The North America Radiological Society), this is the largest and most famous radiological conferences one in the world. The RSNA a special feature is the exhibition material large amount, including from radiology equipment manufacturers. In addition, since about 2016, software developers, including those related to the AI technology, have been active there. The famous IBM Watson was also present there — on the radiology subject.
Impressions were mixed. Like probably many doctors, there were thoughts that in the near future, radiologists will be replaced by artificial intelligence. Because radiologists are a 100% digital profession: every image we work with, whether it is a CT scan, MRI, or ultrasound, is all digital data. Hence the question: what’s the then having a human being point, if all this data can now be processed by a system?
Then I went back to Russia, where, at that time, a movement to develop artificial intelligence in medicine was already underway. I have to know several development companies’ representatives (they some are still working now, and some are no longer in operation), and I began to learn more about it.
It is important to say that over the past six years, the scenarios understanding for the artificial intelligence application in medicine has changed, and doctors’ fears have mostly gone away. The rhetoric in the professional community on this has changed from ‘an AI will replace a doctor’ to ‘a doctor with an AI will replace a doctor without an AI’.
- And how have the AI application scenarios understanding changed?
- It used to be imagined that an artificial intelligence did practically all the work for you. In this “ideal” world, you get everything at a button push: the examination description, the report, all the metrics you can think. In addition, in such a scenario, it’s really not clear why a human being, a radiologist, is needed at all.
However, the industry is moving away from such global ideas to narrower, more specialized tasks. Here, artificial intelligence will not do your job for you — it will not write the conclusion for the radiologist, it will not diagnose the doctor. Such services are called medical decision support systems for a reason. Yes, they are! Only the doctor makes the final decision and only the doctor is responsible for it.
Artificial Intelligence assists him in certain, above all, routine tasks for which the radiologist spends a lot of time. In addition to routine tasks, it also helps in some complicated ones: when some calculations and measurements are needed — which also takes a lot of time.
I admit (and I am almost certain) that in the future, medical artificial intelligence tasks will scale back. When we learn how to use it to solve narrow tasks well (and, just as importantly, to integrate this into the overall process), it will be possible to combine such developments and move on to more global tasks.
- Everyone talks about the using artificial intelligence advantages in radiology. Are there any downsides or potential dangers?
Firstly, there may be an over-reliance on artificial intelligence. The radiologist cannot rely 100% on it. Even if the system has high metrics (sensitivity, specificity, etc.), the doctor must double-check — because he will be responsible for the result, not the artificial intelligence.
This fear comes from the AI other applications — such as unmanned driving. Uber robot cars stopped driving in major the US cities after an accident caused a cyclist death. She was trying to cross the road in a wrong place, and the person in the driver’s seat, who was supposed to take over control, failed to react in time. Such vehicles other manufacturers experience — Google and Waymo — also shows that people trust too much in technology, do not keep their eyes on the road and sometimes even fall asleep while driving.
In the radiographer case, this does not mean he has to measure and recalculate everything. You should at least look at what areas the system has highlighted on the image, what objects it has found — and if you agree with the preliminary conclusion, then all is well. Re-checking will still take less time than if you were to do it all manually — so you’ve already gained in time.
Secondly, information security is important. Here, in turn, two aspects can be distinguished. The first is the personal medical data security. You want to be sure that the studies that are uploaded to cloud storage for subsequent processing are truly anonymised, and that after processing, they do not go anywhere. In addition, medical artificial intelligence is not necessarily radiological: some solutions use textual data, including medical history, test results and so on. In addition, it is important to secure this patient data.
The second aspect concerns the software itself security — in the sense that it must be impossible for third parties to make any changes to it.
There was a telling story in Israel 2–3 years ago when cyber security researchers developed malware to expose vulnerabilities in medical imaging equipment. The virus “grafted” kidney cancer onto “normal” CT scan data and pathology removed signs from studies that originally had them. When radiologists were asked to analyse these studies, they reached the wrong conclusion in cases with “drawn” abnormalities 99% and in cases with deleted ones 95%. This once again shows the information security importance in relation to medical software.
- You said that radiologists’ attitudes towards AI are changing, that there is less apprehension. The Moscow experiment results on the services use in clinical practice also show that doctors’ attitudes have improved. What is the reason for this change?
- First, with improving the services themselves. When many developers started participating in the Moscow experiment in 2020, there were nuances with usability. The radiologist, after all, wants everything to be fast, for the study to be sent to the artificial intelligence for one-click verification.
If you go to radiologists a group with your software, the first question they’ll ask you is “how long will it take?” Because when you have to offload studies from the PACS system, archive them, anonymised them open the AI service in a browser, drag and drop studies into the cloud… It takes both time and the desire to use the service.
Another thing is when the service is built into the same interface in which you work, and there is a “send AI” button. That is how it is now implemented in Moscow, and I am sure that use ease is the key point’s one that builds doctors’ loyalty to AI.
In addition, AI services in general have become better: they analyse better, calculate better. When a doctor assumes something and the AI confirms it to him, it certainly creates more trust in the service, gives the doctor the feeling that he is not alone, that there is additional support.
This is especially important for radiologists with experience (3–5 years or more), because they are more sceptical and their trust really needs to be earned. In addition, a person without experience is less sceptical, he needs advice — so he will rely on artificial intelligence more than an experienced specialist.
In addition, according to my observations, doctors’ attitudes towards artificial intelligence are influenced by their awareness. It is normal for people to be something wary they know little about. Moreover, vice versa: the more such systems are used in clinical practice, the more familiar they become to doctors and the less apprehensive they become.
- When do you think the AI use in clinical practice will become widespread?
- If we are talking about Russia, where state medical institutions, then the artificial intelligence implementation scale in doctors’ work represent a significant part of medicine depends largely on the authorities’ initiative. Right now, as we can see, the technology is being used on a large scale only in Moscow.
In addition, in order to extend this practice to the whole country, we need appropriate initiatives. One option (which, as I know, is already being discussed) is the services inclusion for processing medical images with artificial intelligence in the obligatory medical insurance rate. In addition, for the AI widespread use in public medicine, a common centralised infrastructure should be created (as, again, was done in Moscow).
Eventually, the artificial intelligence use by a doctor will become commonplace. This usual symbiosis will lead to a lesser workload on the doctor, studies will be viewed in more detail and the error rate will decrease. There will be time for self-education, for banal relaxation. From the state viewpoint, I see the main advantages: diagnosis and treatment itself lower costs — diseases will be diagnosed at an early stage, which will ultimately lead to lower mortality rates and increased life expectancy. And patients will get better medical service in every sense. The waiting time for test results will be reduced, and patient satisfaction will increase — and course, medicine for patients will become truly high-tech.