Accelerating Large Volume Electron Microscopy Through Sparse Sampling

Many important processes occur over vastly differing length scales, with critically important information stretching from the nanoscale to the macroscale. Volume electron microscopy (vEM) is a powerful tool for studying biological systems in three dimensions across precisely these length scales, forming a critical part of the information pipeline that aids understanding both subcellular and ultrastructural relationships within cells and tissues [1]. To optimize vEM for any application requires a balance between four competing factors; field-of-view, spatial resolution, electron dose, and time-to-solution. Existing low-dose imaging techniques typically limit one or more of these aspects to preserve the rest, maintaining whichever imaging parameters are considered most critical to ensure a successful experiment. An ideal imaging method would therefore impose no additional restriction on these four aspects, and would increase the imaging efficiency by some other means. Recently, applications of subsampling and inpainting to electron microscopy have revealed that many of the above issues that limit the resolution and sensitivity of electron microscopy images can be alleviated [2,3]. Subsampling, along with a data recovery method such as inpainting, can reduce the number of data points required to perform a given analysis of a material [4]. In the context of scanning electron microscopy, this is achieved by directing the scan coils to acquire data at only a fraction of the candidate pixels, and using a computational method such as dictionary learning and sparse coding methods to recover the missing data [5]. As volume electron microscopy suffers from many issues, especially in the context of electron beam sensitive samples, subsampling and inpainting may be a viable method for increasing experimental efficiency and lowering the barrier of entry. In this presentation we will show how subsampling and inpainting workflows can be implemented in vEM to assist in reducing data acquisition times and required electron doses, without sacrificing field-of-view or spatial resolution. Examples of 25% subsampled cryo FIB SEM tomography data (left) and data recovery by inpainting (right). Inset squares show zoomed crop of both the input and reconstruction.