Transfer Learning for Image Segmentation using Convolutional Neural Networks

SCS - Studienarbeiten

Today’s Machine Learning (ML) success is often limited by a lack of labelled ground truth (GT) data to train the models. This is especially true for applications in medical imaging.Transfer learning (TL) is a state-of-art ML-technique that can be useful to overcome this problem for similar yet distinct tasks. TL tries to apply knowledge or models gained from a task to a second one. In the field of image recognition, TL-methods for deep neural nets re-use parameters trained on source domain data except for the output layer (see image on the right).In medical optical coherence tomography (OCT) of the eye only a limited set of GT-labeled images exists. In addition, there are several device manufacturers and imaging approaches. Thus, transfer learning could 1) enable an easier addition of new device types or imaging techniques, 2) improve the quality of available segmentations even with limited data.


  • Explore current methods and approaches to transfer learning with convolutional neural networks (CNNs)
  • Apply promising approaches to existing ground truth data of OCT images from different machines
  • Evaluate results of transfer learning


  • Fluency in a high-level programming language (e.g. Python)
  • Strong interest in Machine/Deep Learning and willingness to learn about frameworks (MXNet, tensorflow, etc.)


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

Fluent in a high level programming language, ML/DL basic knowledge helpful

Time & Effort
Master’s Thesis, 1 Person