Reconstruction of subcellular structures from resolution-limited volume electron microscopy

Volume electron microscopy (vEM) of biological samples achieves high resolution with generic contrast, providing access to the intricate 3D organization of tissues and cells. It has been used extensively for connectomics—the reconstruction of biological nervous systems—but also serves as a powerful tool for the life sciences more broadly. However, the size and complexity of these data make information extraction a significant image analysis challenge. In this PhD work, I developed deep learning (DL) methods for three such problems. Serial-section vEM produces data with highly anisotropic resolution. We show that high isotropic resolution can be reliably reconstructed if adequate training data is available. Since such data can only be simulated, we developed self-supervised training strategies that eliminate the need for isotropic ground truth, with promising early results. Connectome reconstruction depends on accurately segmenting neurons and detecting all synapses. We developed a method to systematically adapt a successful DL architecture to anisotropic data and reformulated segmentation as a regression on a distance map. This significantly improved state-of-the-art performance, and we used it to generate a synaptic cleft segmentation of the entire Drosophila brain—a valuable resource for neuroscience. We further extended the method to encode synaptic partners at each cleft, enabling assignment of synaptic connections to neurons. Most automated vEM tools focus on neurons and synapses, leaving much of the rich subcellular information untapped. Standard EM staining reveals organelles and other structures distinguishable by morphology and texture. The third project in this thesis makes this information accessible using DL to segment such structures. We manually annotated up to 36 organelle classes across diverse high-resolution datasets and trained networks that generalize across 4 nm and 8 nm data, demonstrating their utility in deriving new biological insights.