Enhanced phase segmentation of scanning electron microscopy images using U-Net with energy dispersive X-ray spectroscopy-guided labelling

This study introduces a deep learning-based methodology that uses U-Nets for the semantic segmentation of Scanning Electron Microscopy (SEM) images of Solid Oxide Cells (SOCs). The objective is the identification of the four key phases in hydrogen SOC electrode (porosity, yttria-stabilised zirconia (YSZ), nickel (Ni), and nickel oxide (NiO)). While the database used for training is built using Energy Dispersive X-ray Spectroscopy (EDS) maps to help label the SEM images using Random Forest classification, the model is trained to segment SEM images directly, without the need for EDS. The results show that the model achieves high segmentation accuracy on SEM images without any manual pre-labelling stepand demonstrates robustness to brightness variations, while challenges such as blur and zoom are addressed through sensitivity analysis. Additionally, a “sliding window” post-processing approach for segmenting large-scale images is used to reduce edge artefacts. The benefit of this approach is to provide an automatic workflow that does not require long EDS acquisitions or a time-consuming manual labelling procedure, offering a scalable solution for microstructural analysis of SOC electrodes. • U-Net trained with EDS-guided labels segments SEM phases with high accuracy. • Model achieves 90% accuracy, removing the need for EDS during mappings. • Deep learning analysis is 60 times faster than semi-manual segmentation. • Model outperforms EDS-based labels, correcting segmentation misclassifications. • U-Net shows sensitivity to blur, zoom, and acquisition conditions.