Lisa Herzog

“In addition to conventional imaging analysis methods, we will apply newly developed deep learning algorithms for an unbiased identification of outcome predictors from patient perfusion imaging data.”

Homepage

Thesis: Deep learning approaches in medical image analysis

Supervisors:
Prof. Dr. Torsten Hothorn (UZH)
Prof. Dr. Beate Sick (ZHAW)

Deep-Learning methods are used to detect strokes in MRT images

The PhD Project is part of an approved SNF project. The title of the SNF project is “Predicting outcome after stroke: take a look at the other side”. The PI of the SNF project is Prof. Dr. Susanne Wegener, Department of Neurology, University Hospital Zurich and University of Zurich, CH-Switzerland. Collaboration partner in the SNF project is Beate Sick, EBPI, UZH.

Summary of SNF Project

Current imaging strategies predicting therapeutic success or failure in acute stroke patients remain insufficient. Here, we propose a novel prediction approach that is based on immediate and long-term vascular adaptations affecting the contralateral side of stroke. In this translational project, the molecular and cellular mechanisms of contralateral blood flow regulation that contribute to ischemia resistance will be studied. Imaging correlates of contralateral perfusion will be applied to predict success of thrombolytic treatments in stroke. In addition to conventional imaging analysis methods, we will apply newly developed deep learning algorithms for an unbiased identification of outcome predictors from patient perfusion imaging data. Using this approach, I envision i) to deepen the understanding of the basic mechanisms regulating brain perfusion after ischemic stroke and ii) to improve therapeutic decision-making in acute stroke patients.

Key research areas that will be addressed within the PhD project:

  • Developing deep learning architectures that take into account the special 3D structure of MRI data
  • Develop DL approaches for semi-supervised learning that allow to make use of vast amount of unlabeled MRI data in addition to the rather small labeled training data
  • Develop a probabilistic prediction model that yields confidence measures that allow deciding if the DL decision might not be reliable and the case should be handed over to a human expert.
  • Identify and visualize image features that are relevant for the decision drawn by the deep learning model
  • For an optimized outcome prediction for individual patients we will evaluate how to combining deep learning approaches with classical statistical models such as time to event models

Planned Duration: 1.1.2018 – 1.6.2021.