Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Physiologically informed neural networks for cardiac Diffusion Tensor Imaging
The aim of this project is to investigate the possible improvements of parameter estimation in cDTI by training neural networks on synthetic data.
Keywords: Master Thesis
Cardiac Diffusion Tensor Imaging (cDTI) provides valuable information about the microstructure of the heart. This information is obtained by fitting a parametric signal model to a series of diffusion weighted MR images per pixel. As respiratory motion and image distortions can corrupt the parameter estimation, it is beneficial to incorporate prior knowledge into the fitting process. Simulating MR images from randomly generated ground truth, allows training Neural Networks to do the parameter inference. By controlling the statistics of the ground truth and including image distortions and motion into the simulation, the prior assumptions of the training data are transferred to the network inference on in-vivo data. Such a robust algorithm could reduce the required amount of data and thereby increase the clinical value of cDTI.
Cardiac Diffusion Tensor Imaging (cDTI) provides valuable information about the microstructure of the heart. This information is obtained by fitting a parametric signal model to a series of diffusion weighted MR images per pixel. As respiratory motion and image distortions can corrupt the parameter estimation, it is beneficial to incorporate prior knowledge into the fitting process. Simulating MR images from randomly generated ground truth, allows training Neural Networks to do the parameter inference. By controlling the statistics of the ground truth and including image distortions and motion into the simulation, the prior assumptions of the training data are transferred to the network inference on in-vivo data. Such a robust algorithm could reduce the required amount of data and thereby increase the clinical value of cDTI.
The goal of this project is to improve the network-based parameter estimation for cDTI data to enable a robust and fast inference. The project will make use of our simulation framework implemented in Python and TensorFlow.
Potential tasks include:
- Segmentation of the heart for varying diffusion contrasts
- Extend simulation pipeline (e.g. off-resonance effects)
- Train a variational network for parameter inference
- Implementing a Bayesian CNN for uncertainty estimation
The project’s focus can be adapted according to your preferences. If you are interested and like to know more please contact us.
**Requisites**
Programming experience in Python, familiarity with TensorFlow or Pytorch is beneficial. Interest in clean implementations and use of version control.
The goal of this project is to improve the network-based parameter estimation for cDTI data to enable a robust and fast inference. The project will make use of our simulation framework implemented in Python and TensorFlow. Potential tasks include: - Segmentation of the heart for varying diffusion contrasts - Extend simulation pipeline (e.g. off-resonance effects) - Train a variational network for parameter inference - Implementing a Bayesian CNN for uncertainty estimation
The project’s focus can be adapted according to your preferences. If you are interested and like to know more please contact us.
**Requisites** Programming experience in Python, familiarity with TensorFlow or Pytorch is beneficial. Interest in clean implementations and use of version control.
Please contact: Jonathan Weine (weine@biomed.ee.ethz.ch). To apply for this project please send a copy of your CV and transcripts of your Bachelor and Master studies.
Supervising Professor: Prof. Dr. Sebastian Kozerke (kozerke@biomed.ee.ethz.ch).
Please contact: Jonathan Weine (weine@biomed.ee.ethz.ch). To apply for this project please send a copy of your CV and transcripts of your Bachelor and Master studies. Supervising Professor: Prof. Dr. Sebastian Kozerke (kozerke@biomed.ee.ethz.ch).