Exploring Compressive Sensing in 4D-STEM Cryo-ET and FIB for Structural Biology

Cryo–electron microscopy (cryo-EM) enables visualisation of biological systems in vitrified ice, preserving structures under conditions that more closely resemble their native environment. In recent years, single-particle analysis (SPA) in cryo-EM has achieved remarkable improvements in resolution through advances in hardware and software [1]. While cryo-electron tomography (cryo-ET) has followed a similar trajectory [2], resolution remains limited, in part because thicker in situ specimens degrade imaging quality due to dynamical scattering. One method to address this has been focused ion beam (FIB) milling to thin vitrified samples and, more recently, scanning transmission electron microscopy (STEM) rather than conventional Transmission Electron Microscopy (TEM). STEM is a well-established technique in materials science that is finding increasing applications in structural biology. It can image thicker samples compared to traditional TEM [3]. Four-dimensional STEM (4D-STEM), combined with emerging approaches such as ptychography, aims to further improve resolution for radiation-sensitive materials, such as biological samples [4]. The major limitations of using 4D-STEM for cryo-ET are the lack of real-time data visualisation, the acquisition times required to scan the sample area, and the volume of data collected. A single 4D-STEM tilt series collected at a 256x256 fully rastered field of view with a 256x256-pixel Merlin detector can generate approximately 328 GB of data. In this work, we employed a subsampling acquisition strategy to shorten acquisition times and substantially reduce storage requirements for large 4D-STEM datasets, thereby enabling a cryo-ET workflow. Currently, the data size scales with the subsampling rate; for example, 50% subsampling yields approximately 164 GB of data under the previous parameters. FIB-Scanning Electron Microscope (SEM) tomography (also known as slice-and-view tomography) is another area that can greatly benefit from compressive sensing [5]. Like in 4D-STEM, collecting large, high-resolution FIB-SEM tomography data requires hundreds of slices to capture as much ultrastructure information as possible. This results in prolonged imaging times and stability issues associated with such long sessions. Because collection speed is directly proportional to the sampling quantity, a 25% subsampled scan can reduce data collection time by up to 75%. These datasets were collected and reconstructed using SenseAI’s imaging and inpainting software (Sense AI Innovations Ltd.). The image reconstructions were done using a Beta-Process Factor Analysis algorithm [6]. Together, these approaches highlight practical advances toward more efficient imaging of structural biological specimens. Full rastered 256x256 image, 66% sampled image using linehop pattern, reconstructed subsampled image of microbacteria