Deep Learning for Limited-Angle Electron Tomography (Master Project / Thesis)

Background

Electron microscopy (EM) tomography has become a key technique for visualizing biological structures at (sub)cellular scales. In practice, however, mechanical limits of the microscope, specimen geometry, and electron dose constraints restrict the tilt angles that can be acquired. This leads to the limited-angle tomography problem and the well-known missing-wedge in Fourier space, which in turn causes strong reconstruction artifacts and loss of structural information.

Figure 1. Examples of LACT

In recent years, many deep learning methods have been proposed to reduce these artifacts in limited-angle computed tomography (LACT), often by combining physical imaging models with data priors or learned regularizations. Yet, when the angular range is very small, LACT remains extremely challenging and current methods still struggle to produce reliable reconstructions.

Objective

The goal of this project is to investigate and develop an advanced reconstruction pipeline for limited-angle electron tomography of biological samples, with a particular focus on challenging angular ranges and realistic imaging conditions. To this end, cutting-edge techniques such as diffusion models, flow matching, NeRF, and 3D Gaussian splatting will be explored and integrated into the pipeline.
 
You will:
  • Investigate and implement deep learning–based approaches for LACT (e.g.,physics-informed networks, learned priors, or generative models).
  • Explore alternative formulations of the problem (e.g., reconstruction in image space vs. sinogram space, regularization in 3D vs. projection domains).
  • Assess reconstruction quality on real EM datasets of biological samples (e.g., mitochondria), not just on synthetic phantoms.

Student contribution

By the end of the project, you will have:

  • Contributed to the design and implementation of a LACT reconstruction pipeline in PyTorch.
  • Trained and tested your methods on real electron tomography data of biological samples.
  • Performed a quantitative and qualitative evaluation of the reconstructions, including the analysis of typical missing-wedge artifacts.

Prerequisites

  • Solid programming skills in Python.
  • Proficiency with PyTorch (or strong motivation to learn it quickly).
  • Basic knowledge of computer vision, image processing, or computer graphics is recommended.
  • Familiarity with deep learning and optimization is a plus but not strictly required if you are willing to learn during the project.

(This project can be adapted in scope and depth for a semester project or a master’s thesis,
depending on your background and available time.)

Contact

[email protected]