Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Machine Learning for Advanced Magnetic Resonance Perfusion Quantification
Magnetic Resonance (MR) perfusion imaging is a clinically established technique to detect coronary artery disease by visual assessment. However, quantitative evaluation of myocardial blood flow is desirable as it promotes objective diagnosis. Here, ML may help to avoid expensive fitting procedures.
Keywords: Magnetic Resonance Imaging, Machine Learning, Myocardial perfusion, Numerical models, Computer simulations, MATLAB, Python, C
A decrease in myocardial tissue perfusion is one consequence of coronary artery disease, which, if untreated, can lead to myocardial infarction. In clinical practice, perfusion is routinely examined using time resolved cardiovascular perfusion magnetic resonance imaging. However, the diagnosis is mainly based on visual assessment.
Fully quantitative evaluation of myocardial blood flow (MBF) in ml/g/min is desirable as it promotes objective diagnosis and simplifies analysis. Moreover, diagnostic accuracy benefits as MBF assessment helps to distinguish between actual myocardial deficits and image artefacts.
Generally, myocardial perfusion quantification relies on a kinetic tracer model in the tissue of interest, e.g. the Fermi or the more complex blood tissue exchange model (BTEX). Different models offer different trade-offs between model complexity and robustness against noise and artefacts. Pixel-wise iterative fitting is especially prone to noise and computationally expensive.
Most recent work indicates, however, that machine learning based approaches allow to directly map pharmacokinetic parameter estimates from noisy tissue signal time curves at low computational cost.
The objective of the present project is to use the MRXCAT model as a platform to test a machine learning approach for absolute MBF quantification. Different quantification models are already implemented in the MRXCAT framework. This at hand, perfusion training (N=1000) and validation data with varying lesion deficits in the myocardium are generated. Corresponding tissue residual concentration time curves will be simulated at various noise levels. MBF quantification will be performed via the tissue models to provide synthetic training datasets. A suitable convolutional neural network (CNN) is already implemented. It will be trained, tested and validated with the simulated data. Attention will be directed to the statistical evaluation of the pharmacokinetic parameter estimates.
In this student project, the candidate will work in an interdisciplinary environment. We offer a spectrum of expertise from theoretical description and implementation of new MR sequences, to the development of acquisition and image reconstruction pipelines as well as latest Machine Learning applications.
A decrease in myocardial tissue perfusion is one consequence of coronary artery disease, which, if untreated, can lead to myocardial infarction. In clinical practice, perfusion is routinely examined using time resolved cardiovascular perfusion magnetic resonance imaging. However, the diagnosis is mainly based on visual assessment.
Fully quantitative evaluation of myocardial blood flow (MBF) in ml/g/min is desirable as it promotes objective diagnosis and simplifies analysis. Moreover, diagnostic accuracy benefits as MBF assessment helps to distinguish between actual myocardial deficits and image artefacts.
Generally, myocardial perfusion quantification relies on a kinetic tracer model in the tissue of interest, e.g. the Fermi or the more complex blood tissue exchange model (BTEX). Different models offer different trade-offs between model complexity and robustness against noise and artefacts. Pixel-wise iterative fitting is especially prone to noise and computationally expensive. Most recent work indicates, however, that machine learning based approaches allow to directly map pharmacokinetic parameter estimates from noisy tissue signal time curves at low computational cost. The objective of the present project is to use the MRXCAT model as a platform to test a machine learning approach for absolute MBF quantification. Different quantification models are already implemented in the MRXCAT framework. This at hand, perfusion training (N=1000) and validation data with varying lesion deficits in the myocardium are generated. Corresponding tissue residual concentration time curves will be simulated at various noise levels. MBF quantification will be performed via the tissue models to provide synthetic training datasets. A suitable convolutional neural network (CNN) is already implemented. It will be trained, tested and validated with the simulated data. Attention will be directed to the statistical evaluation of the pharmacokinetic parameter estimates.
In this student project, the candidate will work in an interdisciplinary environment. We offer a spectrum of expertise from theoretical description and implementation of new MR sequences, to the development of acquisition and image reconstruction pipelines as well as latest Machine Learning applications.
• Review literature on absolute quantification strategies in MRI
and CMR perfusion.
• Familiarize with the MRXCAT phantom model with multi-model perfusion quantification.
• Employ our previously implemented CNN.
• Train, test and validate the network with the simulated data.
• Compare influence of estimation model for training dataset on final estimate.
• Statistically evaluate parameter prediction depending on various noise levels.
The here listed student project is subject to further modifications based on the student’s background, interest and motivation. Please do not hesitate to contact hoh@biomed.ee.ethz.ch if you are interested.
• Review literature on absolute quantification strategies in MRI and CMR perfusion.
• Familiarize with the MRXCAT phantom model with multi-model perfusion quantification.
• Employ our previously implemented CNN.
• Train, test and validate the network with the simulated data.
• Compare influence of estimation model for training dataset on final estimate.
• Statistically evaluate parameter prediction depending on various noise levels.
The here listed student project is subject to further modifications based on the student’s background, interest and motivation. Please do not hesitate to contact hoh@biomed.ee.ethz.ch if you are interested.
Supervisor: Tobias Hoh (hoh@biomed.ee.ethz.ch); Professor: Prof. Dr. Sebastian Kozerke;
Supervisor: Tobias Hoh (hoh@biomed.ee.ethz.ch); Professor: Prof. Dr. Sebastian Kozerke;