On all fronts: Will AI change cancer treatment
Artificial intelligence in modern medicine is increasingly ceasing to be an auxiliary tool and becoming a full-fledged participant in the scientific process. One example of such a transition was the development of the SPARK system, an AI approach capable of working with scientific ideas, formulating and testing hypotheses related to the study and treatment of cancer. Whether AI can become a key tool in the fight against cancer, how the new system works and when to expect its application in real clinical practice — in the Izvestia article.
What is the advantage of the new system?
Until recently, artificial intelligence (AI) in medicine was perceived mainly as an auxiliary tool. He helped doctors analyze images faster and more accurately: to find tumors, identify suspicious areas, and reduce the risk of diagnostic errors. However, a new study published in the journal Nature Medicine shows that the role of AI can be much broader.
We are talking about an attempt to create a system that is able to participate in a full-fledged scientific search and help make new discoveries.
The SPARK system developed by the researchers is an example of the so—called agent-based approach. It uses not one model, but several specialized modules that interact with each other. Some of them formulate hypotheses, others refine them, and others test them on real medical data.
At the same time, the system uses plain text as a universal language. She can describe the intended idea, then translate it into measurable parameters and check whether it is confirmed on medical images of tumors. In other words, the AI does not just analyze the images, but tries to explain what biological processes may be behind the changes it sees.
This approach differs significantly from traditional algorithms that are trained to find previously known features, such as cell shape or tissue density. The new system is capable of offering potentially significant characteristics that could escape human attention.
How AI can change the fate of a cancer patient
Experts say that one of the main advantages of systems like SPARK is an increase in the accuracy of cancer diagnosis and faster detection of dangerous changes.
As Alexey Trukhin, associate professor of the Department of Medical Physics at the Engineering Physics Institute of Biomedicine at the National Research Nuclear University "MEPhI", explains, any neural network is based on a complex mathematical model that allows analyzing large amounts of information and identifying patterns that are invisible to humans. According to him, today such technologies are already helping doctors to focus on suspicious areas of medical images, standardize documentation and interpretation of research.
— Early and accurate diagnosis largely determines the patient's chances of successful treatment. The earlier a tumor is detected and the better its nature is understood, the more likely it is to choose effective therapy and avoid aggressive treatments where they are not needed. The development of AI can increase the targeting and safety of therapeutic methods, as doctors will receive more accurate information about the disease, the expert notes.
The use of neural networks looks particularly promising in pathomorphology, a field of medicine where specialists study tumor tissues and cells under a microscope. Modern digital images of such drugs have ultra-high resolution and contain thousands of objects interconnected by a complex structure. Manually analyzing such data requires a lot of time and high concentration.
— A pathologist works with images with a resolution of up to 200 nanometers. Neural networks make it possible to automate localization, segmentation and accounting of the number of objects in such images, and quantify their organization," explains Trukhin.
At the same time, experts emphasize that this is not about replacing a doctor, but about changing the nature of his work. AI takes over part of the routine analysis, allowing the specialist to focus on interpreting the results and making decisions.
How fast technology is reaching people
The introduction of medical technologies based on artificial intelligence is a sequential chain of stages. According to Alexey Trukhin, the key first step is to register the technology as a medical device. Depending on the complexity of the system and the quality of the evidence base, this process can take from 6-12 months in optimal cases to 2-3 years in practice. At this level, the Roszdravnadzor regulator evaluates algorithms according to formalized criteria: sensitivity, specificity, and accuracy.
However, registration is only the beginning of the path. This is followed by implementation into clinical practice, which may take several more years. It requires the restructuring of medical processes, the training of specialists and the adaptation of the healthcare system to new tools.
As Yuri Komarov, Deputy Director for Organizational and Methodological Work at the N.N. Petrov National Medical Research Center for Oncology, told Izvestia, artificial intelligence has already entered everyday medical practice today.
— AI can help with the formation of recommendations, and at the moment, systems are being implemented that, according to the instructions for use, can provide information on drug compatibility, and taking into account concomitant diseases or research results, suggest possible side effects, but the final decision will be made by a specialist in any case, — the expert notes.
He emphasizes that AI systems are already used in almost every region, but the very concept of "implementation" remains vague: it can be either installing a software solution on a server or using it fully in clinical practice, which requires much more time.
Komarov also draws attention to the fact that the integration of AI into real work includes not only technical configuration, but also training of doctors, checking the quality of systems after implementation and the formation of sustainable skills in their use.
Who is responsible for AI decisions
As neural networks are increasingly used in diagnosis and treatment selection, it becomes especially important to define the boundaries of responsibility between the doctor and the technology developer.
According to Yuri Komarov, a doctor should not be responsible for the quality of the algorithm, since artificial intelligence remains a tool, and not an independent subject of decision-making. The specialist's role is to be able to use such systems, but not to formally confirm their conclusions.
— If a specialist only "signs" the AI's conclusion without actually analyzing the result, it devalues the very idea of using technology. The doctor will still double—check the data," Komarov notes.
He emphasizes that a more stable model is a system of constant feedback, in which all controversial or erroneous results of AI work are considered jointly by doctors and developers. This approach allows not only to identify problems, but also to consistently improve the algorithms themselves.
— It is necessary to build a system in which any deviations in the results of AI research will be dealt with by specialists together with developers. This should be a continuous process of technology improvement," the expert emphasizes.
At the same time, according to Alexey Trukhin, it is impossible to completely eliminate errors in the operation of neural networks. Any such system is trained on limited datasets, which means there is always a risk of both false positive and false negative results. This is especially true for rare diseases, which may be poorly represented in training samples and may be lost among typical cases.
"The same data may correspond to different human conditions, which is why an error appears, but this error is controllable in the long run, unlike a subjective opinion," the Izvestia interlocutor summarizes.
That is why medical systems with AI belong to the category of high-risk solutions and require particularly rigorous testing before being implemented in clinical practice.
Переведено сервисом «Яндекс Переводчик»
