AFM image analysis of porous structures by means of neural networks

In this work, we presented a method of finding and characterising transmembrane porous structures, called fenestrations, with the help of convolutional neural networks. Case studies are performed on high resolution AFM images of murine liver sinusoidal endothelial cells (LSECs). At first, we evaluated different kinds of noise occurring in the LSEC AFM measurements. Next, we proposed a schematic structure of the neural network suitable for our purpose. We examined different loss functions, optimising the accuracy of fenestration detection. Finally, we presented the method of rough calculation of fenestration size distributions. We demonstrated that the accuracy of this method surpasses 90%. Furthermore, it is fast, not sensitive to the chosen image contrast and fully deterministic. The simple scheme can be easily modified to different objects of interest, which promotes the use of neural networks as a universal tool for the analysis of various kinds of mi-croscopy images.

Automated summary

In this study, a method utilizing convolutional neural networks to find and characterize transmembrane porous structures in high resolution AFM images of murine liver sinusoidal endothelial cells (LSECs) was presented. Through evaluation, proposing network structure, optimizing loss functions, and rough calculation steps, the method achieved an accuracy exceeding 90%, is fast, insensitive to image contrast choices, and easily modifiable for different objects of interest, promoting neural networks as a universal tool for analyzing microscopy images.