Student Projects

Automatically and robustly extracting anatomical information from cardiac MR images is an area of ongoing research, which has potential application in a medical setting and in personalised modelling by using real anatomical morphology to construct a realistic synthetic anatomy, which can then be used in bio-physical simulations. We are developing a framework for fitting anatomical meshes to CMR data, and one key point for the correct quantification of clinical biomarkers is the identification of anatomical landmarks. This project aims to make use of state of the art machine vision approaches to locate and track heart valves in cardiac MR images from 2D sets of images. The student will first evaluate existing solutions available in the literature and if necessary, will train a new neural network using our clinical dataset which will be manually processed to generate the requited training, validation and testing dataset. In this project you will work extensively with python, numpy and pytorch, and prior experience with these is required. Additionally, prior experience with (cardiac) MRI or convolutional neural networks would be helpful.

The aim of this project is to develop a method that is able to automatically and robustly locate the heart valves in 2D cardiac MR images. Further, for time resolved data the method should be able to track the valves over the full cardiac cycle.

Supervisors: Dr. Stefano Buoso (buoso@biomed.ee.ethz.ch). To apply for this project please email a copy of your CV and transcripts of your Bachelor and/or Master studies.

Cardiovascular Magnetic Resonance MR Imaging (MRI) allows for the acquisition of subject-specific anatomical, functional as well as microstructural data of the heart. In the last years, several frameworks based on machine learning have been proposed to classify, label and segment MRI cardiac data. In a second step, personalised in-silico models are generated to analyse and enrich he available information from MRI. Conventionally, in these models, discretise the partial differential equations describing cardiac mechanics, electrophysiology and haemodynamics are solved relying on Finite Elements or Finite Volumes. In this project, we want the explore the challenges and requirements to extend these neural networks frameworks to the generation of personalised in-silico models. Particularly, we are interested in physics informed neural networks. The project will make use of our existing numerical tools based on open-source libraries for Finite Elements (FEniCS), Finite Volumes (openFOAM) and machine learning (Tensorflow, Pytorch, Keras).

Our investigation has multiple objectives and can accommodate for the student’s interests. Please contact us if you are interested and would like to know more. Potential projects include: - Modelling of cardiac mechanics, - Atlas based and reduced order modelling of cardiac mechanics, - Simulation of blood flow in arteries and left ventricles, - Reduced order modelling of blood flow, - Inverse problems in cardiac mechanics and fluid dynamics, - Generation of synthetic MRI images based on biophysical models and training of neural networks for functional evaluation of clinical images. **Requisites** Programming experience in Python is required. Familiarity with Tensorflow or Pythorch or Keras is a plus.

Supervisors: Dr. Stefano Buoso (buoso@biomed.ee.ethz.ch). To apply for this project please send a copy of your CV and transcripts of your Bachelor and Master studies.

Standardised performance assessment of image acquisition, reconstruction and processing methods is limited by the absence of images paired with ground truth reference values. Synthetic datasets could therefore be bridging this cap since they are associated with known parameters that could be used for computation of error metrics. However, these datasets are characterised by poor anatomical and functional variability which is usually limited to healthy conditions. With this project, we want to develop a generative model, based on deep-learning methods, for the generation of synthetic phantoms with sufficient anatomical variability to realistically represent population statistics. The project will be based on our recent work on generative models based and implemented in Pytorch.

The project aims at: - Building a generative model for left-ventricular 2D masks using convolutional variational autoencoders - Building a conditional generative model for the human torso tissues using convolutional variational autoencoders - Apply realistic tissue properties to the images using our texturizer network - Generating a dataset of representative cases including healthy and pathological left ventricles

Supervisors: Dr. Stefano Buoso (buoso@biomed.ee.ethz.ch). To apply for this project please send a copy of your CV and transcripts of your Bachelor and Master studies.

Image registration plays a crucial role in the realm of medical image acquisition and analysis. It encompasses various complex processes including image alignment, optical flow estimation, tissue tracking, motion compensation, and deformation estimation. Particularly, within dynamic MRI reconstruction, the focus lies in motion tracking of the subject during acquisition to enhance the quality of reconstruction. We have been successfully utilizing a method known as pTV registration in our workflow. This method has also been incorporated into other libraries, including Matlab, for broader usage. Further information and examples on pTV registration can be found on the following GitHub page: https://github.com/visva89/pTVreg (For examples: https://github.com/visva89/pTVreg/blob/master/Examples.md ) The project at hand consists of two major tasks: * Since Matlab is gradually becoming obsolete, a significant portion of the codebase needs to be ported to Python. This will ensure the continued accessibility and usability of the pTV registration. * Additionally, the pTV method needs to be rewritten using PyTorch to make it compatible with the context of differentiable programming. This project is ideal for those who possess strong programming skills and a comprehensive understanding of numerical programming. Given the wide usage and citation of the code, the ported version will be made public, offering visibility to the individual who successfully completes the task. This project can be undertaken as a semester-long assignment or as a master's project. Reference: Vishnevskiy, V. (2016). Effective Joint Regularization in Medical Imaging for Nonsmooth Nonlinear Priors (pp.1-36). ETH Zurich. https://www.research-collection.ethz.ch/handle/20.500.11850/155735

Valery Vishnevskiy vishnevskiy@biomed.ee.ethz.ch

Phase Contrast Cardiovascular Magnetic Resonance MR Imaging (PC-MRI) allows for the subject specific quantitative assessment of blood flow, with cerebrovascular applications including dementia, aneurysms, arteriovenous malformations, and atherosclerosis among others. However, the required high spatial resolution (~0.8 mm3) to assess smaller vessels results in a high sensitivity to subject motion (random but also breathing). To overcome this issue, sub-voxel motion during acquisition can be measured using a camera, which can be used to retrospectively correct the data before image reconstruction. Application: https://doi.org/10.1016/j.neuroimage.2022.119711 Methods overview: http://dx.doi.org/10.1088/0031-9155/61/5/R32

The aim of this project is to develop a camera-based solution for motion correction of cerebrovascular 4D flow MRI. The student will investigate the possible setups proposed in literature, assemble the hardware of the camera-based solution, develop the data processing software (with the possibility to explore denoising algorithms), and test in-vivo using a 3T MRI.

Supervisors: Manuel Christanell (mchristanel@ethz.ch) and Luuk Jacobs (ljacobs@ethz.ch). To apply for this project please email a copy of your CV and transcripts of your Bachelor and/or Master studies.

Phase Contrast Cardiovascular Magnetic Resonance MR Imaging (PC-MRI) allows for the subject specific quantitative assessment of blood flow. As abnormal flow patterns are related to the evolution of cardiovascular disease, automatic tools for hemodynamic and flow analysis are a fundamental step in diagnosing and monitoring disease progression. However, limited data availability and lack of ground truth information prohibit the development of generalized and robust models.

The aim of this project is to develop a method using physics-informed neural networks to learn aortic flow behaviour and identify aortic hemodynamic properties, which hold important information used for the assessment of aortic stenosis. The student will develop and expand existing physics-informed neural network approaches, as proposed in the literature, to analyse gemetrical and velocity field properties from in-silico generated aortae. Synthetic and subsequently real data will be used to test the approach. The project will be carried out in Python and make use of open-source libraries such as Pytorch and Pyvista.

Gloria Wolkerstorfer (wolkerstorfer@biomed.ee.ethz.ch), Dr. Stefano Buoso (buoso@biomed.ee.ethz.ch). To apply for this project, please email a copy of your CV and transcripts of your Bachelor and/or Master studies.

This project is set to limited visibility by its publisher. To see the project description you need to log in at SiROP. Please follow these instructions: - Click link "Open this project..." below. - Log in to SiROP using your university login or create an account to see the details. If your affiliation is not created automatically, please follow these instructions: http://bit.ly/sirop-affiliate

In 4D Flow MRI respiratory gating or triggering is required to suppress respiratory motion. Recent work has aimed at resolving rather than suppressing respiratory motion and hence has been referred to as 5D Flow MRI. To achieve image reconstruction from undersampled data, data correlations along space, cardiac and respiratory phases have been exploited by favouring low-rank representations. However, it has been shown that temporal smoothing of high fidelity image features can occur leading to temporal low-pass filtering of velocity data. To address the issue, low-rank properties have been refined by determining and incorporating transformations between motion states.

• Investigate and report on the accuracy of our current self-gating method o Coil clustering o Automatic coil detection o Limited field-of-view • Investigate and report on the motion-informed LLR algorithm compare to LLR baseline o Motion artifacts reduction o Underestimation of peak velocities o Potential turbulence estimation benefits

Supervisors: Sébastien Emery (emery@biomed.ee.ethz.ch); Prof. Dr. Sebastian (kozerke@biomed.ee.ethz.ch).