UC San Diego

Electron microscopy (EM) facilitates analysis of the structure, distribution, and functional status of organelle networks within the nervous system. Recent breakthroughs in EM specimen preparation and instrumentation have furnished scientists with the ability to automatically collect volumetric datasets large enough to cover significant swaths of neuroanatomical subdivisions at nano-resolution. The quantification of biological morphologies from these data, however, typically requires image segmentation, which is a long- standing and well-recognized bottleneck. Though datasets may now be collected at rates exceeding teravoxels per day, the manual segmentation and analysis of all features from such a volume requires many years of human labor. As technological advances driven by the desire to reconstruct entire nervous systems continue to push instrument throughput skyward, it is clear that our ability to model brain ultrastructure will be limited by the rate of image analysis rather than that of image acquisition. The body of work described in this dissertation represents a contribution towards alleviating this impediment. A pipeline for the automatic segmentation, morphological quantification, and spatial characterization of organelles from high resolution datasets at the teravoxel-scale is presented. Segmentations were generated using a highly parallelized, supervised machine learning approach that reduces the required human effort from years to just a few hours. A host of generic and organelle-specific post- segmentation filters were developed, and it is shown that their application improves segmentation accuracy. Accelerated approaches for generating surface renderings from these large-scale segmentations are introduced, and a workflow for the automatic computation and reporting of morphological, topological, and spatial metrics is described. These methods were then applied to study the spatiotemporal changes of organelles in neurons of the mouse suprachiasmatic nucleus across the diurnal cycle. Novel findings pertaining to nuclear structure and organization are reported and discussed. Taken together, the methods described here provide a series of tools for expediting the quantitative analysis of organelle structure-function relationships in the current era of big data in biological microscopy