Jonathan Dietrich

Semester Work
Supervisors: Dr. Christian Baumgartner

Bridging the Gap Between 1.5 T and 3 T MRI Data Using Generative Adversarial Networks

Medical image analysis using machine learning techniques suffers from a lack of labeled data sets. Specif- ically, while some large annotated datasets exist, there are many datasets that have been acquired with dif- ferent protocols for which no annotations are available. Training machine learning algorithms on a source domain where labels are available and applying them to unlabeled target data is thus of great interest. The majority of existing approaches require at least some supervision in the target domain. Furthermore, only one related work takes advantage of the predictive power of deep neural networks. In this thesis we in- vestigate a method which uses generative adversarial networks to perform unsupervised domain adaptation, meaning no pairs of corresponding images are required. Images are translated from the source domain to the target domain. This allows translating a labeled data set to another domain, where no labeled images exist and training a machine learning algorithm on the translated data set. The effectiveness of this domain adaptation technique is tested by training on 1.5 T MRI data in the ADNI data set and testing on 3 T images or vice versa. We find that, after our preprocessing, there is only a small domain gap between 1.5 T images and 3 T. Nevertheless, the studied methods manage to largely bridge this domain gap. Furthermore we test the influence of conditioning image translation on noise. Whether noise helps to bridge the domain gap is inconclusive, but we observe that the noise changes the anatomical structure of the output images.