Latent Back-projection Network: a missing-wedge generative model for cryo-electron tomography

Gabriel Meyer-Lee, Robert Kiewisz, Tristan Bepler

Keywords: deep learning, cryo-em, cryo-et, autoencoder, representation learning

Abstract: Cryo-electron tomography (Cryo-ET) is hindered by the missing wedge, a gap in Fourier space information caused by limited tilt series angular coverage, leading to anisotropic resolution loss and artifacts. We introduce a Latent Back-projection Network (LBN) which learns a latent representation of the imaged volume, enabling re-projecion of micrographs at tilt angles which cannot be experimentally collected. Our model is trained as a masked autoencoder on a large dataset of real Cryo-ET tilt series. This allows the model to learn to realistically impute missing information, with the appropriate physical constraints imposed by our back-projection and re-projection architecture. Through evaluations on real tilt series with excluded tilts, the missing wedge of simulated tilt series, and synthetic augmentation to tomogram reconstruction, LBN demonstrates generative capabilities which overcome limitations of prior approaches. Additionally, LBN is designed as a general model, providing a path toward an easily distributed pre-trained method for missing-wedge correction.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 23645