U-Net Network Enhancements to Facilitate Rapid Electron Microscopy Imaging for Connectomics

Abstract

Imaging the structural and functional connections between cells in the brain allows neuroscientists to understand the brain by studying neuronal wiring diagrams. To automatically segment and classify images that are used in the construction of these neuronal wiring diagrams, or connectomes today, machine learning segmentation techniques require an image scanned with an electron microscope at either a slow dwell time or with small pixel sizes. However, a scalable and more rapid implementation of connectome construction has not yet been realized because of the significant cost of multi-beam electron microscopes and the relatively slow time in which connectomes can be constructed using a single-beam electron microscope. Segmented connectomes include sections that can be segmented properly with a fast scanned image as well as sections that require slow scanning for proper segmentation. Due to this fact, a potential way to enhance the time in which connectomes can be produced and segmented is to first scan samples at fast resolution and perform segmentation using a convolutional neural network, identify those areas of interest that require more detailed imaging through a learning-based error detection network, and then rescan only those identified high interest areas to produce a fused image for segmentation. The proposed thesis will analyze various machine learning methods for segmentation using the U-Net network and review proposed enhancements to the U-Net network that can better utilize electron microscopy images for construction of segmented connectomes. The successful use of fused electron microscopy images will potentially enable higher speed and lower cost electron microscopy imaging for connectomics.