| Description |
In connectomics the goal is to generate a full 3D reconstruction of neurons within the brain. This reconstruction will help neuroscientists better understand how the brain functions and also how it fails in the case of dementia and other neurodegenerative diseases. Attempts for this reconstruction are ongoing at varying levels of resolution from the millimeter scale of magnetic resonance imaging (MRI) to the nanometer scale of electron microscopy. In this dissertation, we develop tools that improve the ability of researchers to trace neurons through a volume in vitro at a resolution sufficient for the identification of synapses. The first toolset will speed up the process for generating training datasets in newly acquired volumes to be used in supervised learning methods. Current methods of training dataset generation require dense labeling by a trained research scientist over hundreds of hours. This toolset will reduce the time required for training data generation by using sparse sampling and reduce the cost of that time by using a priori knowledge of the structure to allow for less specialized researchers to assist in the process. The second toolset is targeted at speeding up the correction of errors introduced in automatic neuron segmentation methods. Often referred to as proofreading, current methods require significant user input and fail to fully incorporate the information generated by the automatic segmentation method. I will use novel 2D and 3D visualization techniques to take advantage of the information generated during automatic segmentation into the proofreading process. This toolset will consist of two different applications, with one process targeting 2D proofreading with 3D linking and the other process targeting direct 3D proofreading. |
