Abstract:
In the area of 3D neural reconstruction, or connectomics, there is a significant gap between the time required for data acquisition and dense reconstruction of the neural processes. Manual labeling techniques have been used to generate highly accurate segmentations but are time intensive. Automatic methods are able to significantly reduce the time required, but the state-of-the-art accuracy is so far insufficient for use without user corrections.
My research has focused on reducing the time required to generate manually labeled datasets as well as manual corrections of automatic results. In the current state, manually labeled datasets are generally used as training for automatic methods to process larger volumes. I will present a novel method that relies on sparse sampling of the dataset with guided proofreading to generate a complete 3D dataset that can be used for training these automatic methods. Additionally, I will present a new visualization technique that utilizes the results from training to allow for more efficient correction of the final result.