For the first time, Russian scientists from the National Research Nuclear University MEPhI (NRNU MEPhI) taught a neural network to analyze and search for nanoparticles through a microscope.
In order to teach a neural network to perform this task, it needs to be shown several tens of thousands of marked up photos. These are special images that show what task the neural network will perform. This, according to scientists, is extremely difficult for a number of highly specialized scientific problems.
To get around this problem, scientists did not mark up real photographs to train the neural network, but generated images that simulate them on a computer.
SEM (scanning electron microscope), which uses an electron beam instead of visible light, is used to study nanoparticles synthesized for medicine and other purposes. Analysis of SEM images consists in the detection of particles and their size distribution. Neural network approaches in this area are not developed, and standard methods of image processing do not provide the required quality.
Alexander Kharin, specialist of the Engineering Physics Institute of Biomedicine, National Research Nuclear University MEPhI
The results of the study will make it possible to automate the processing of SEM images, revolutionizing the standard methods for studying new materials, scientists are sure.
This work, the authors believe, will help not only to reduce the research time, but also to increase the number of analyzed particles – from hundreds of units to tens of thousands. In the future, the research team is going to automate the classification of nanoparticles.