A Neural Network for Medical Image Registration Without Pretraining on Large Datasets Has Been Developed in Russia
Researchers at the AI Center of Lomonosov Moscow State University (MSU) and the university's Faculty of Computational Mathematics and Cybernetics have developed a neural network method for registering medical images that does not require pretraining on a large image dataset. The approach was tested on lung CT scans and MRI scans of the brain and heart. A research paper reporting the test results was published in the journal Machine Learning and Knowledge Extraction.
Artificial intelligence in medicine is most often associated with detecting abnormalities: an algorithm analyzes a scan and identifies a tumor. But doctors sometimes face an equally challenging task: tracking how an organ changes over time. The lungs deform as a person breathes, the heart contracts, and brain structures change as diseases progress. Simply overlaying the scans won't work. Researchers at MSU's AI Center and Faculty of Computational Mathematics and Cybernetics have proposed a way to address this problem while bypassing one of medical AI's biggest bottlenecks: the need for enormous datasets.
Anatomy in Motion
Conventional neural networks used for medical image registration require pretraining on tens of thousands of scans. This creates several challenges. Collecting medical images is difficult because of patient privacy requirements, and scans produced by different imaging systems are often incompatible.
The MSU team took a different approach, building on Kolmogorov–Arnold Networks (KANs). Its KAN-IDIR algorithm and lightweight version, RandKAN-IDIR, do not need to learn from millions of examples. Instead, they are optimized directly on the specific pair of images that needs to be aligned.
"In our experiments, the proposed approach combined accurate image registration with low GPU memory consumption, moderate runtime and less sensitivity to random initialization than competing methods," explains Nikita Drozdov, a Ph.D. student at the Faculty of Artificial Intelligence of Lomonosov Moscow State University and a junior researcher at the university's AI Center.
Tests using lung CT scans and brain and cardiac MRI scans produced impressive results. The lightweight version processes a pair of scans in less than a minute, with an image registration error of about 1 mm at anatomical landmarks. The algorithm is also highly robust and requires less computing power.
Reducing Dependence on Big Data
This development has strategic significance for Russia's medical technology sector. It reduces reliance on large training datasets, which domestic hospitals and research centers often find extremely difficult to assemble.
The technology's lower GPU memory requirements could also make it accessible to hospitals and research labs without supercomputers. The algorithm could become a core component of Russian clinical decision support systems, helping physicians accurately assess disease progression and plan treatment.
From Mathematics to Clinical Practice
For now, this is not a finished medical product but a fundamental research development. There is still a long way to go from a laboratory algorithm to a tool displayed on a physician's monitor at a regional hospital.
According to Roszdravnadzor (the Federal Service for Surveillance in Healthcare), as of August 2026, Russia had registered 59 domestically developed medical solutions that use AI. The market is taking shape, but deploying the MSU method will require independent clinical trials using equipment from different manufacturers, followed by government registration as a medical device. The researchers themselves emphasize that the method needs further testing on datasets with different levels of noise.
Global Context and Russia's Computer Vision Research
Globally, the Russian development is entering the field of deformable image registration, where Swiss developers of TotalSegmentator are also active. The Russian method's strengths lie in its instance-specific optimization and robustness to random initialization. In the longer term, this could create opportunities to export not hardware but Russian algorithms and research expertise. Potential partners could include countries in the Eurasian Economic Union (EAEU) and BRICS, which are actively developing their own medical IT systems. Commercial exports, however, will require resolving licensing issues and adapting the algorithm to international standards.
MSU's work is part of an emerging Russian field of medical computer vision. Developments over the past several years show a progression from projects focused on microscopic image segmentation (2022–2024) and the FNOReg algorithm, which reduced memory requirements eightfold (2024), to three-level models for brain MRI registration developed at Peter the Great St. Petersburg Polytechnic University (SPbPU) in 2026.
Russian researchers are moving beyond simply building "disease detectors." They are developing fundamental analytical tools that make it possible to observe the human body as it moves and changes over time. If clinical trials confirm that Kolmogorov–Arnold Networks remain robust in real-world hospital settings, this technology could become an important part of digital healthcare. It would demonstrate that medical breakthroughs do not always require endless volumes of data – sometimes, mathematical precision is what matters most.
