How artificial intelligence is transforming the radiologist profession
The phrase āartificial intelligence in medicineā has moved from futurological fantasies and discussions to real life in the last ten years. This is particularly true in radiology, where computer vision technologies are already in clinical use. Artificial intelligence processes studies such as MRI, CT, ultrasound, and others for signs of pathologies. Is this to say that a radiologist is no longer required? What impact does the introduction of these technologies have on the profession?
Celsus AI has been developing and implementing artificial intelligence systems for radiology for over five years.The longer we work in this field, the more we interact with radiologists in the process of developing our solutions, the more convinced we are that artificial intelligence does not threaten, but rather expands, the radiologistās profession, allowing the doctor to engage in new types of professional activities.
In this article, we will discuss the main ābranchesā of the profession that have emerged at the intersection of radiology and artificial intelligence, as well as the role of the doctors with whom we work every day.
Radiologist-labeler
In another sense, this is a āradiologist-teacherā path, because a doctor is directly involved in the creation of data for artificial intelligence training. But first, we recommend that developers understand why they require this information in the first place.
Our āCelsusā system is a multi-layered neural network with a complex architecture. Developers feed a large amount of data into its input layer, where artificial intelligence analyzes it, looks for common patterns, and learns to interpret data in a specific way based on the task at hand. Itās similar to teaching a child: at first, he has no idea what a cat looks like (or what a cat is in general). However, if his parents show him a cat thousands of times and say āitās a cat,ā eventually some cat-specific signs form in his head ā and he can distinguish them, for example, from dogs.
It is not enough to show artificial intelligence many different studies (mammograms, fluorograms, radiographs) to teach it to detect X-ray signs of malignant neoplasms and other pathologies on medical images. It is necessary to mark up this data, in order to inform the neural network about which images have pathology and which do not. Even better, āshowā her the specific areas of interest in the image.
This is precisely why we require radiologists-labelers: the developers lack medical education and clinical experience. As a result, we realized from the start that we couldnāt do without doctors and that we needed to establish a continuous process of working on data markup for development.
We currently have over 40 marking doctors with whom we work on a commercial basis. Because our solutions work in four diagnostic directions, each doctor specializes in one of them: someone marks mammographic studies, someone fluorographic studies, and so on in all modalities. The doctor selects the number of studies he wishes to mark up and is paid based on the amount of work completed.
What are we looking for when recruiting radiologists to mark up data? To begin, work experience must be greater than five years. Second, the doctor must first pass the verification (confirmed) data entry control. Third, a radiologist must have a medical practice in addition to working with us. Furthermore, we value the radiologistās willingness to clearly follow the marking instructions, because any inaccuracies will have a negative impact on the neural networkās training.
Medical consultant
A medical consultant serves as a liaison between machine learning developers and a team of markup doctors. Although his responsibilities do not end there. They are traditionally classified into three groups.
A medical consultantās primary responsibility is to respond to medical-related questions from the ML team. Questions can be of various types, such as what a lung pattern is, how to recognize a malignant pathology in the study, whether the lungs of a young and an elderly person differ, and so on.
Another crucial aspect of his work is visual product quality testing. The medical consultant examines the errors in the serviceās operation, attempting to identify a common pattern and developing a hypothesis for how to solve the problem.
For example, the service detects pathology of a specific class where none previously existed. The medical consultant discovered that such errors occur in studies of older people after analyzing such errors. As a result, the ML team decides to include the patientās age in training to improve prediction accuracy.
The third direction is an examination of doctor marking. Because we have markup doctors from various levels of healthcare and with varying levels of experience, their approach to marking research differs. As a result, the consultant examines their markup, identifies disagreements, and clarifies the markup requirements, thereby assisting in the creation of the most understandable and accurate instructions for doctors.
Furthermore, the consultant assists us with cross-marking, which occurs when we assign a single study for analysis to multiple doctors at the same time. Often, the results of doctors differ, and we cannot figure out what is wrong without medical expertise: perhaps there are signs of some complex pathology on the study, and the doctor cannot make an accurate conclusion without additional information, or perhaps some doctors did not fully understand how these objects need to be marked up.
All of this contributes to our understanding of the medical consultant. In general, it is worth noting that, over time, he begins to understand the field of medical artificial intelligence in addition to the medical field. He already has a general understanding of how we train neural networks, set up experiments, and what is important to us in general. This is his specialized knowledge.
Our development managers (in each diagnostic area) have stated that prior to the appearance of a medical consultant, they felt like āblind kittensā because they had nothing to rely on and it was frequently difficult to achieve a full understanding with the marking doctors. They now interact with the consultant on a daily basis and can quickly get an answer to any question.
Medical Director (MD)
The doctors mentioned above are directly involved in the development of our solutions. The medical director is also well-versed in our internal cuisine: he understands how products develop and function in practice, and he occasionally performs some functions of a medical consultant. In addition, he contributes to the companyās scientific work, pondering the strategy for future product development and selecting the critical functionality of the service.
Its distinguishing feature, however, is that it also represents us āin the outside world.ā After all, developing a good artificial intelligence for radiology isnāt enough; you also need to convince the professional community of its importance. As a result, the medical director speaks at specialized conferences on a regular basis, talks about us to the professional community, and tries to convey the idea that artificial intelligence is not a replacement for a doctor, but rather a tool. In fact, he is the face of our teamās medical team.
Conclusion
Thus, the advancement of artificial intelligence in radiology has provided doctors with several new avenues for advancement. We are confident that this is the future, because products like our Celsus are never static, but are constantly refined. Other specialists with medical education and experience are very likely to be in demand in this area in the future, for example, specialists in evaluating artificial intelligence systems, introducing them into clinical practice, and selecting optimal use cases.
We would like to take this opportunity to thank all of the doctors with whom we work and about whom we have written in this article. You are true heroes; without you, nothing would have happened!