Application of Machine Learning Methods for Annotating Boundaries of Meshes of Perineuronal Nets

The article explores the use of neural networks to solve the problem of determining boundaries of meshes of perineuronal nets. The confocal stacks’ image layers of rat brains are used as initial data. This article presents a comparison of two alternative methods to solving the problem. The first method is based on the generation of boundaries through the use of a neural network based on the DCGAN architecture. The second method is based on solving the problem of semantic segmentation using a neural network based on the U-Net architecture, that is widely used in biomedicine. For both neural networks, architectural changes are presented to achieve greater generalizability of the models. Some learning strategies are considered to solve the problem of overfitting that is typical for small samples. Both solutions showed results comparable in quality to the semi-automatic algorithm. A solution based on the U-Net architecture provides good tools for further solving the problem of boundary ambiguity and allows customizing the algorithm for various criteria for boundaries of the perineuronal nets.