Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Parameter and Uncertainty Estimation for MR Relaxometry
The primary aim of the project is to adapt and characterize an existing framework for Bayesian parameter and uncertainty estimation that utilizes machine-learning for posterior estimation.
Relaxometry describes a plethora of quantitative magnetic resonance imaging techniques that aim at determining relaxation times (T1, T2, T2*) and proton density, which can be indicative of pathological changes of tissue. Classical Relaxometry techniques rely on finding relaxation times using mostly analytical, simplified MR signal models, which are fitted to measured multi-contrast data on a per-pixel basis (see e.g. MOLLY (Messroghli et al., 2004), TESS (Heule, Ganter, & Bieri, 2014)). Typical simplifications entail modelling of noise as normal distributed, neglection of the imaging and signal reconstruction process, as well as, the influence of confounding factors such as motion induced artifacts, leading to biases in their estimates.
To correctly account for the measurement noise of the used MR magnitude images, Bayesian methods can be used that directly take the noise distribution into account. However, sampling of the posterior distribution can be computationally expensive. In this project, we would like to explore machine learning for the posterior estimation as was previously performed for intra-voxel incoherent-motion (IVIM) imaging (Zhang, Vishnevskiy, Jakab, & Goksel, 2018).
Contrary to conventional imaging, MRI measures the Fourier domain representation of an object (its “k-space”). As such, inconsistencies and truncation of k-space leads to artifacts, which are spread over the whole image domain (e.g. ringing artifacts) leading to biased Relaxometry results. One way to correct for such imaging based artifacts would be to include neighboring pixel information into the Relaxometry reconstruction.
Relaxometry describes a plethora of quantitative magnetic resonance imaging techniques that aim at determining relaxation times (T1, T2, T2*) and proton density, which can be indicative of pathological changes of tissue. Classical Relaxometry techniques rely on finding relaxation times using mostly analytical, simplified MR signal models, which are fitted to measured multi-contrast data on a per-pixel basis (see e.g. MOLLY (Messroghli et al., 2004), TESS (Heule, Ganter, & Bieri, 2014)). Typical simplifications entail modelling of noise as normal distributed, neglection of the imaging and signal reconstruction process, as well as, the influence of confounding factors such as motion induced artifacts, leading to biases in their estimates.
To correctly account for the measurement noise of the used MR magnitude images, Bayesian methods can be used that directly take the noise distribution into account. However, sampling of the posterior distribution can be computationally expensive. In this project, we would like to explore machine learning for the posterior estimation as was previously performed for intra-voxel incoherent-motion (IVIM) imaging (Zhang, Vishnevskiy, Jakab, & Goksel, 2018).
Contrary to conventional imaging, MRI measures the Fourier domain representation of an object (its “k-space”). As such, inconsistencies and truncation of k-space leads to artifacts, which are spread over the whole image domain (e.g. ringing artifacts) leading to biased Relaxometry results. One way to correct for such imaging based artifacts would be to include neighboring pixel information into the Relaxometry reconstruction.
The goals of the project can be adapted to the background and interest of the applying student. Tentative goals:
1. Review of literature regarding existing reconstruction algorithms and Bayesian parameter estimation approaches (mandatory)
2. Implementation and validation of machine learning based reconstruction for current state-of-the-art Relaxometry techniques that can perform (either) of two tasks:
a. Per-pixel relaxometry quantitation including an error estimate
b. Inclusion of neighboring pixel information to increase robustness of parameter quantitation
3. Characterization and comparison of the developed algorithms to classical, yet state-of-the-art reconstruction algorithms
The goals of the project can be adapted to the background and interest of the applying student. Tentative goals:
1. Review of literature regarding existing reconstruction algorithms and Bayesian parameter estimation approaches (mandatory)
2. Implementation and validation of machine learning based reconstruction for current state-of-the-art Relaxometry techniques that can perform (either) of two tasks:
a. Per-pixel relaxometry quantitation including an error estimate
b. Inclusion of neighboring pixel information to increase robustness of parameter quantitation
3. Characterization and comparison of the developed algorithms to classical, yet state-of-the-art reconstruction algorithms
Programming experience in MATLAB or Python is required. Experience in machine learning and the use of Tensorflow would be helpful. The candidate presents with a keen interest in medical imaging and the application of machine learning to medical image reconstruction. Knowledge of medicine / anatomy is not required.
Please apply with transcript of records, CV and a short motivational statement.
Programming experience in MATLAB or Python is required. Experience in machine learning and the use of Tensorflow would be helpful. The candidate presents with a keen interest in medical imaging and the application of machine learning to medical image reconstruction. Knowledge of medicine / anatomy is not required.
Please apply with transcript of records, CV and a short motivational statement.