Synthesising training data with generative adversarial networks (GANs) in computed tomography perfusion.
Master thesis
Permanent lenke
https://hdl.handle.net/11250/2787141Utgivelsesdato
2021Metadata
Vis full innførselSamlinger
- Studentoppgaver (TN-IDE) [823]
Sammendrag
Brain stroke is seen as a very vital problem due to its possible health consequencesand incidence. It is the second cause of premature death worldwide. Systems thatcan assist the decision process of doctors are of vital importance, as they need interventionas soon as possible after a possible brain stroke.
Studies on detecting brain stroke in the fields of image processing and machinelearning often encounter the problem of not having enough labeled data. This has asignificant impact on the success percentages of the studies carried out.
Based on this problem, this thesis aims to synthetically generate Perfusion CT imagesof patients affected by brain stroke with the Generative adversarial networks(GANs). In order to do that, different architectures such as Deep Convolutional GenerativeAdversarial Network (DCGAN), Wasserstein Generative Adversarial Network(WGAN), and Motion and Content Decomposed Generative Adversarial Network(mocoGAN) are investigated and modified for this problem and as a result,Perfusion CT images are generated.