Sparse Scanning Electron Microscopy for Imaging and Segmentation in Connectomics

Abstract

Sparse Scanning Electron Microscopy can be used in combination with Inpainting algorithms to reduce acquisition time and electron dose. Especially for three-dimensional (3D) or very large field of view imaging, acquisition time reductions are of large importance to the community. Dose reduction is of importance in imaging material that is sensitive to electron radiation. In this study we demonstrate a workflow that acquires data by performing a sparse scan at random positions on a specimen. The sparse data is reconstructed to a full grid image by a GPU accelerated dictionary based inpainting algorithm. The reconstructed data is suitable to be used for automatic semantic segmentation of neuron structures. We demonstrate the procedure on two key segmentation applications in connectomics (cell membranes and mitochondria) and show that the overall segmentation quality improves notably compared to data from a conventional raster scan acquired with the same total dose per image. Alternatively, the total dwell time per pixel can be reduced by 33% while maintaining the same level of quality of the segmentation. These results demonstrate that sparse scanning and reconstruction can increase the effective data acquisition rates without sacrificing on quality for the end user segmentation application.