Master/Bachelor Thesis Projects at the Image Analysis and Learning / Hamprecht lab.

If you are interested, send an email to Fred Hamprecht. Our topics are particularly suitable for students of physics, mathematics, and computer science, but motivated people from other fields are also welcome!

Neuro Segmentation

A major goal of neuro-science is the understanding of basic neural circuits. To this end, novel electron microscopic methods acquire 3-dimensional images of the brain at a resolution of 10-20 nm and a field of view of up to 0.1 mm. These huge datasets can only be analyzed automatically, and our group is a major algorithm provider for this application. We continuously offer interesting master projects dealing with specific subproblems, for example:
  • Design and implementation of new features that capture important image properties such as the shape and layout of candidate regions.
  • Detection methods for organelles such as mitochondria and vesicles.
  • Novel optimization algorithms that efficiently find the best among many candidate segmentations.
  • Development of semi-automatic tools that optimally combine automatic methods with interactive user guidance to arrive at high quality results as quickly as possible.

The Digital Embryo

Biologist are making rapid progress towards the understanding of embryonic development. Thanks to novel light-sheet microscopes that can capture fully 3-dimensional volume data at a temporal resolution of under one minute over several days, they can now study how initially independent cells manage to group themselves into all the different organs of an organism. Our group addresses the algorithmic challenge to track all visible cells in such a dataset over time. We continuously offer interesting master projects dealing with specific subproblems, for example:
  • Design of a reliable detector for cell division events, where the track of a single cell splits up into two child tracks.
  • Hierarchical segmentation algorithms that reliably find all visible cells even under high noise levels.
  • Efficient methods to estimate the uncertainty of the results of our analysis.
  • Algorithms for the intuitive visualization of 3D+time data on 2-dimensional displays.

Development of Novel Machine Learning Methods

Machine learning has become the method of choice to solve complex image analysis problems. It allows non-expert users to train a generic analysis algorithm by showing examples of desired outcomes. To extend the scope of this approach to more complex problems, we investigate new learning methods and novel ways to train them. Interesting master projects in this context include:
  • Development of active learning methods that minimize training time by guiding the user towards particularly valuable examples.
  • New image features that capture important aspects of the data for subsequent use by learning methods.
  • Intuitive training strategies for non-standard types of training data such as object shape and layout.
  • Design of novel ensemble methods that combine analysis results from several sources into an integrated result of much higher quality than any of the inputs.

Scientific Software Design

Our image analysis solutions apply complex algorithms to huge amounts of data and yet must be intuitively operable by non-experts. This poses great challenges on careful software design, and our group maintains a number of major open-source projects to address these challenges. Fast and flexible modules for fundamental image analysis and optimization algorithms are mainly implemented in the VIGRA and OpenGM C++ libraries, whereas the Python-based ilastik project takes care of the high-level organization, parallelization, and graphical user interface. Students who want to master advanced software development techniques can contribute to these projects in many ways, for example:
  • Design of easy-to-use workflows for recurring tasks in the application fields we address.
  • Reusable integration of novel algorithms developed by our group or elsewhere into these projects.
  • Implementation of advanced parallelization strategies to fully utilize the power of modern multi-core CPUs and clusters.
  • Powerful 3-dimensional visualization of analysis results for rapid user feedback and presentation.