A Data-Efficient Deep Learning Workflow for Atomic-Level Segmentation of In Situ High-Resolution Transmission Electron Microscopy Images With Heterogeneous Backgrounds.

While in situ high-resolution transmission electron microscopy (HRTEM) allows atomic-resolution observation of dynamic processes such as chemical reactions and phase transformations, quantitative analysis of such data is often hindered by its complexity and volume that make manual labeling or traditional segmentation methods ineffective. Here, we demonstrate a data-efficient deep learning workflow using a standard U-Net architecture to segment atomic positions from in situ HRTEM images of copper oxides on copper substrates and gas interfaces-a system with challenging heterogeneous backgrounds and complex interfaces. The workflow is optimized for small datasets, achieving high accuracy after being trained on fewer than 50 manually labeled images augmented with rotation, scaling, and Gaussian filtering. The resulting U-Net model achieved test accuracies of approximately 92% for bulk atoms and 63% for interfacial atoms on images similar to the training dataset, with most misclassifications occurring at the complex interfacial regions. Moreover, our approach offers an approximately 300,000-fold improvement in labeling speed compared to manual labeling. This work presents a practical and adaptable framework for accelerating quantitative analysis of experimental systems for which acquiring large, annotated datasets is prohibitive, demonstrating the effective application of U-Net for a common challenge in electron microscopy.