Generation of synthetic ground truth for Machine Learning

SCS - Studienarbeiten

Today’s Machine Learning algorithms rely on large data sets of ground truth. Especially for medical use cases the lack of appropriate data for training can be a limiting factor for successful Deep Learning. For some imaging problems, the creation of synthetic ground truth can effectively overcome the lack of original ground truth and enable successful learning.In this work we aim to explore the potential of synthetic ground truth for tumor segmentation in medical imagery of eyes. The student will work with 3D stacks of optical coherence tomography (OCT) images. The currently implemented Convolutional Neural Networks (CNNs, e.g. U-net) shall be extended to segment tumors using both synthetic ground truth data and state-of-art CNNs.


  • Explore current state-of-the-art research on synthetic ground truth both for medical and non-medical image data
  • Implement an (semi-)automated method to introduce artificial tumor artefacts into images of healthy eyes
  • Explore different neural network architectures for OCT image segmentation
  • Evaluate advantages and disadvantages of synthetic ground truth


  • Fluency in a high level programming language (e.g. Python)
  • Computer Vision basics very desirable
  • Machine/Deep Learning knowledge helpful, willingness to learn about frameworks


Kind of Work
30% Theory, 40% Implementation, 30% Evaluation

Fluent in a high level programming language, Computer Vision basics, ML/DL knowledge helpful

Time & Effort
Master’s Thesis, 1 Person