Advances in Computational Cryo-Electron Tomography ---Model-based and Neural Reconstructions

  • Yuanhao Wang

Student thesis: Doctoral Thesis


Tilt-series cryo-electron tomography (cryo-ET) is an established imaging technique used in several fields like biology and material science. Despite its success, cryo-ET remains an arduous task. The missing-wedge acquisition, the motion, and the high-level noise are the main challenges existing in this field. In this dissertation, we tackle these challenges through the exploration of three distinct approaches: plug-and-play approach, adaptive differentiable density grids, and adaptive tensorial density field representation. Firstly, using embedded fiducial markers, our framework first estimates the motion field in the sample in order to correct the captured projections, and then reconstructs the sample using an iterative plug-and-play approach. Secondly, we propose an adaptive representation based on an octree structure. Each block of the structure represents a 3D density grid that is optimized from the captured projections. To ensure a denoised output, we update the octree in a multi-scale fashion. We also combined the differentiable image formation model with cross-nodes non-local constraint, total variation, and boundary consistency priors in the loss function. Thirdly, we propose an adaptive tensorial density field representation. Considering that cryo-ET is thin along the z-axis, we use a quadtree structure instead of initializing a nulled octree, and represent each node with a vector-matrix factorization. Combined with total variation, boundary consistency, and an isotropic Fourier prior, our framework allows us to reconstruct clean cryo-ET results. Experiments show considerable improvement in cryo-ET reconstructions.
Date of AwardSep 2023
Original languageEnglish (US)
Awarding Institution
  • Computer, Electrical and Mathematical Sciences and Engineering
SupervisorWolfgang Heidrich (Supervisor)

Cite this