MOSCOW, December 13. Russian scientists have found a new way to use a neural network in medicine. According to researchers from Samara University. Korolev, the use of this technology will increase the accuracy of MRI scanning. The research results were published in the journal Computer Optics.
According to scientists from Samara University. Korolev, with the help of the proposed technology it will be possible to detect interference caused by the patient’s movements directly during an MRI scanning session. Now the measurement is complicated by the fact of the patient’s voluntary movements in the device; they have to be taken into account manually during data post-processing.
Artificial errors, also called MRI artifacts, often make tomography difficult to interpret. They cause the patient's head to move during the procedure. Interference may cause the scan to be terminated early due to inaccurate results.
«
“In existing technology, motion artifacts are filtered out at the data processing stage of an MRI experiment when setting brain volumes to the first or average volume of the series using a rigid body transformation. They include three displacement parameters and three rotation parameters for each volume of a temporary MRI series,” said senior lecturer at the Department of Technical Cybernetics, employee of the Institute of Artificial Intelligence at Samara University. Queen Nikita Davydov.
According to the expert, work towards creating a technology for tracking interference during an MRI procedure has been carried out by the university since 2019, and has now been completed and implemented. At Samara University, neural networks were trained to detect step artifacts of head movement in fMRI data and adapted to a set of real data.
“First, the neural network model is trained on a large amount of synthetic data generated with parameters close to real ones, then on a small number of sets of real head movement data corresponding to different people, and after that the model already works with a small part of real data corresponding to a specific experiment,” – Davydov explained. This approach is called “meta-learning from a small amount of data” and has previously been used in the task of image restoration as part of research at the Institute of Artificial Intelligence.
The further task of the research team is to increase the accuracy of the neural network model by creating more procedure for generating synthetic data close to real data. The next step for further analysis will be the ability to calculate such characteristics as coordinates, heights, and duration of the anomaly, which will allow us to filter out the number of false positives of the classification.
Samara University. Koroleva is a participant in the Russian state program for supporting universities «Priority 2030» of the national project «Science and Universities».