Machine vision system for recognition of three-dimensional structure of perineuronal networks

The article proposes a system representing a set of algorithms for detecting the boundaries of perineuronal network cells on different layers of microscopic images, representing a cell in a three-dimensional structure, and three-dimensional visualization of a perineuronal network cell. The initial data are layers of confocal stacks of the mouse brain. In the course of the work, the applicability of neural networks for solving the problem of extracting masks of the internal structure of perineural network cells was studied. This article presents an algorithm for extracting masks based on solving the problem of semantic segmentation using a neural network on the U-Net architecture, which is popular in biomedicine. Architectural solutions are proposed that allow mitigating the problem of overfi tting in conditions of a small sample size. Two algorithms are proposed for studying perineural network cells based on discrete measurements of the color signal distribution across the thickness of the confocal stack, as well as an algorithm for detecting the cell itself in the thickness of the obtained image layers. An algorithm for processing the obtained masks to create a three-dimensional point cloud and a method for subsequent reconstruction of the cell using alpha forms for three-dimensional visualization are proposed.