Generative learning of morphological and contrast heterogeneities for self-supervised electron micrograph segmentation

Deep learning in electron microscopy (EM) data analysis is predominantly supervised, relying on manually labeled data. This dependence limits scalability and slows the development of high-throughput EM characterization of materials. While simulation-based approaches provide an alternative, they often struggle with morphological heterogeneity, contrast complexity, and experimental artifacts, reducing their real-world effectiveness. We introduce EMcopilot, a closed-loop generative learning framework that enables label-free EM segmentation. EMcopilot leverages the general vision model to extract morphological priors and employs a conditional generative adversarial network to generate contrast-aware images. An EM-specific domain adapter further enhances realism by modeling key microscope-specific perturbations. Benchmark results show that EMcopilot-trained models not only achieve segmentation accuracy comparable to human-annotated models but also outperform them in detecting nanoparticles in poor-contrast regions and spatially clustered configurations, overcoming inherent human biases in annotation. By illustrating how generative models distill and transform complex EM features into a robust training resource in a self-supervised manner, EMcopilot provides a scalable solution for automated microscopy analysis.