Deep Neural Networks for Image Processing and Biomolecular Single-Particle 3D Reconstruction from Electron Micrographs

Electron microscopy (EM) has become a powerful tool for materials characterization. However, the existence of limiting factors in imperfect imaging systems, such as aberrations, astigmatism, detection noise, etc., makes extracting sample information from noisy electron micrographs an inverse problem. Differences in image formation models and features compared to optical cameras hinder the direct application of advanced computer vision techniques developed for visual objects to electron microscopy. This Thesis focuses on developing novel deep learning models for image processing of electron micrographs to enable faster and more accurate characterization of diverse samples. Specifically, (1) U-Net and its variants were employed for high-performance semantic/instance segmentation and shape classification, facilitating automatic quantification and classification of microplastics. (2) Noise2Noise, zero-shot Noise2Noise and Noise2SR were adapted to eliminate common corruptions in experimental SEM images of minerals, carbon nanotubes and mesoporous carbon. (3) Three deep learning models (VampPrior-SPR, ExemplarPrior-SPR and LSGM-SPR) were proposed for biomolecular heterogeneous 3D reconstruction from low-SNR 2D cryo-EM images. These models successfully resolved discrete compositional and continuous conformational heterogeneity across various simulated and experimental datasets. These advancements contribute to enhancing the capabilities of electron microscopy for broader applications with higher inference accuracy and speed.