"If you can't stop AI – lead it": interview with a radiologist who became a consultant at a medtech startup

In 2020, radiologist Alexander became acquainted with AI for the first time in order to detect COVID-19. Many of his colleagues were concerned that AI would replace them. Alexander also contacted the developers and offered suggestions on how to improve the software.

You've probably heard that data is required for AI training. And high-quality medical data is required to train medical AI, such as detecting pathologies on X-ray images. Furthermore, they must be labeled in order to show neural networks: this is a lymph node, and this is most likely cancer. And this will not be possible without the assistance of a qualified specialist. As a result, responsible medical AI developers include doctors in the development process.

In the medtech startup "Celsus AI," Alexander Khomyakov is a consultant doctor and the coordinator of marking doctors in the direction of fluorography and chest X-ray. We interviewed him to learn more about how he works in the ML field and whether he believes that assisting AI is assisting a competitor.

— Tell us about yourself and what you did before starting the medtech startup.

— I am a five-year-experienced radiologist. I have graduated from I.M. Sechenov Moscow State Medical University in 2016 and completed my radiology residency at A.I. Evdokimov Moscow State Medical University in 2018. I continued to work in computed tomography and traditional X-ray modalities, first in one of Moscow's city polyclinics and then in the scientific and practical clinical center for diagnostics and telemedicine.

— When did you first come across AI? How did you find it?

— I first came into contact with AI while working in an outpatient CT center during the COVID-19 a pandemic Then, AI services aided in quickly assessing the extent of lung damage and devising treatment strategies (we're talking about the Moscow experiment with AI in clinical practice - approximately).

Most colleagues were skeptical of AI, which is understandable given that the services were in their early stages. They discovered pathology outside the target organ, discovered "frosted glass" in knee joint studies, calculated the volume of lung damage incorrectly (underestimated), and took a long time to process the studies.

But even then, it was clear to me that artificial intelligence should be used in radiology. Its potential (and now, apparently, real) speed of calculations and measurements exceeds human capabilities, allowing services to provide a preliminary conclusion even in the absence of a radiologist. They can also help radiologists by taking over routine tasks and paying attention to pathologies, which is important when working long hours without a break.

— Why did you decide to work in the field of artificial intelligence?

— I was initially interested in information technology and new diagnostic methods. I wanted to have an impact on the development of AI services, but when I first started working with them, I found it difficult to communicate with developers: all communication took place through chats with service reviews; it was not possible to communicate directly.

Soon, interested doctors were given contact information for representatives of the services of interest, and the interaction became more productive as it became possible to get quick answers to questions.

— What are your responsibilities at Celsus today?

— I answer medical questions, consult, and assess service quality. At Celsa, a team of doctors is working on data markup, and I interact with them, correct their work, and propose hypotheses for improvement.

— Have you personally contributed to the development of an AI product?

— The particulars of marking medical data are that each doctor has his or her own opinion about the marking of objects in the study. And this is detrimental to AI training.

One of my major responsibilities at Celsus was to "clean up" datasets (approx.) from inaccurate and incorrect markup, an explanation to the team of some of the features of a radiologist's work that he receives through experience and observation. As a result, we were able to compete successfully with other AI services, achieve high quality indicators, and lead the Moscow Experiment leaderboard in 2023.

— What was the most difficult adjustment?

— In the context of information technology in general, and machine learning in particular. This is not the case when you can read specialized literature and learn at least a little bit. Even before I met Celsus, I had a philistine interest in artificial intelligence. To comprehend the neural network algorithm, I needed to become at least superficially acquainted with probability theory, statistics, and mathematical analysis.

However, this is insufficient for specific suggestions for improving the work. My assumptions frequently appear philistine to me, as there is insufficient qualification. It's also difficult to get used to professional slang in a work chat - as in the meme "nothing is clear, but very interesting." But the guys and I are trying to find a common language:)

— Have you ever regretted your decision to work in artificial intelligence? Aren't you concerned that you're assisting a "competitor"?

— I don't regret it: at Celsus AI, I learned about IT product development and witnessed the project's transformation from a raw product to a leader in its field.

Concerning the "competition," I completely understand my colleagues who have a neutral or even negative attitude toward AI. To begin with, working with a raw product in the midst of an already heavy workload is a dubious pleasure. Second, if the product is already mature enough, it raises concerns that AI will eventually replace doctors.

But, having seen "from the inside" how neural networks are trained and achieve the necessary quality metrics, I believe AI will eventually take over in radiation diagnostics. If he does not replace the doctor, he will be able to filter out studies that do not conform to the "norm" - reducing the amount of research described by the doctor significantly. At the same time, the question of its medical ethics remains unanswered.

When I realized this, I thought to myself, "You can't stop the ugliness (that is, AI) - lead it!":)

— Doctors from various countries will read our interview. What would you recommend to radiologists who want to learn more about the principles of work and how to work with AI?

- I advise interested radiologists to communicate directly with developers, asking questions, offering ideas, and developing and mastering a new specialty.