Comparison of segmentation algorithms for FIB-SEM tomography of porous polymers: Importance of image contrast for machine learning segmentation

Abstract

Focused-ion-beam/scanning-electron-microscopy (FIB-SEM) tomography has become increasingly important to investigate the three-dimensional (3D) distribution of different phases in inhomogeneous materials. However, quantitative analysis of 3D structures by FIB-SEM tomography requires adequate image segmentation that is typically based on thresholds in the intensity histograms of SEM images. Recently machine learning (ML) segmentation algorithms have emerged with the opportunity to tune the algorithm based on prior knowledge of SEM image contrast. In this work we have performed FIB-SEM tomography of 3D printed nanoporous polymer structures with a special focus on SEM image segmentation and pore size determination. By applying Monte Carlo (MC) simulations, SEM image contrast of the pore/polymer interface was analysed. The performance of several traditional segmentation algorithms on simulated SEM images was found to lead to erroneous results on pore sizes. Training the ML segmentation algorithm by the knowledge obtained from MC simulations yields more reliable segmentation results. Evaluated pore sizes and analysis of further 3D material properties show a strong dependence on the segmentation method, which emphasize the importance of the segmentation process in the 3D reconstruction of FIB-SEM data.