Segmentation of Solid Oxide Cell Electrodes by Patch Convolutional Neural Network

A framework for the automatic segmentation of large 3-D datasets of microscopic images is designed and tested for microstructures of solid oxide cells (SOC) electrodes reconstructed using focused ion beam-scanning electron microscopy (FIB-SEM). The developed algorithm utilizes a simple, yet very effective deep neural network based on the patch-convolutional layers (patch-CNN) in the encoder-decoder configuration. The guidelines for a selection of network architecture and for preparing and minimizing the training dataset are given. The proposed methodology is tested for various SOC electrode microstructures. The analyzed FIB-SEM tomography data have different resolutions, material properties and measurement artefacts. Pixel-based accuracies of the validation datasets are over 97.5%, and the 3-D microstructural parameters calculated from the ground-truth data and CNN-data show good agreement. The proposed automatic patch-CNN microstructure reconstruction shortens the image processing time by two orders of magnitude. The proposed method can achieve high accuracy which has not been reported in previous studies. In addition, the proposed CNN can serve as an artificial pore-infiltration technique that will help to reconstruct samples without epoxy resin infiltration. The developed framework can be easily extended for other multiphase porous materials with different 3-D imagining techniques that require image processing.

Automated summary

The study presents a framework for automatically segmenting large 3-D datasets of microscopic images using a deep neural network, achieving strong results in microstructure analysis. The proposed automatic patch-CNN microstructure reconstruction significantly reduces image processing time. The suggested CNN can also serve as an artificial pore-infiltration technique, aiding in sample reconstruction with improved accuracy.