Volumetric Segmentation in Electron Microscopy Brain Imaging. In Data Science & Engineering Master of Advanced Study (DSE MAS) Capstone Projects

Advances in volume electron microscopy imaging have enabled the accumulation of large-scale, high-resolution biological data. In the neuroscience domain, the potential for these volumes to characterize the dense network of intertwining structures of the brain along with its subcellular components is within reach, but still limited in many ways by the size and complexity of the analysis. We have utilized the high-speed storage and GPU resources provided by the Pacific Research Platform (PRP) and CHASE-CI, managed by the Kubernetes engine, Nautilus, to implement and validate a 3D U-Net for multi-class volumetric segmentation trained on a scarcely labeled mouse brain dataset. A deep learning internal zero-shot superresolution was evaluated on downsampled volumes to simulate its effect on pixel classification after x-y resolution boosting. In addition to traditional model performance metrics, a volume rendering tool was created to visualize voxel-level predictions to better understand problematic structures and subcellular features. These models and tools extend the neuroimaging reconstruction framework, NeuroKube.