MIDAS.LAB

Available Bachelor/research/Master thesis

Some of the topics are in collaboration with the Institute of Signal Processing and System Theory, University of Stuttgart and the Max Planck Institute for Intelligent Systems, Empirical Inference, Tübingen.
The following section gives an overview of the potential topics.

Overview: AI-assisted data processing

The inclusion of artifical intelligence (AI) into medical data processing can help to improve performance by (but not limited to) increasing precision, boosting quality of service, easing processing and reducing computational times. It enables to develop patient-centered workflows with personalized treatments that include: data processing from multiple imaging modalities acquired with multi-parametric imaging sequences, advanced reconstruction, post-processing and anylsis techniques.
In acquisition step:

• Sequence development for multi-parametric and motion-resolved MRI
• Patient-adaptive sampling optimization including online feedback according to movement cycle, SNR, parametric information
• Monitor imaging state by external sensors (Microsoft Kinect camera, respiratory belt, ...) to match e.g. to a motion model

In reconstruction step:

• Conventional and deep-learning based image reconstruction: Inclusion of multi-parametric and motion information, high-dimensional data processing
• Non-rigid motion estimation and correction: Deep-learning based image registration and motion correction
• Sensor fusion to map imaging states to external surrogate signals
• Framework development for Gadgetron

In post-processing step:

• Convolutional neural networks for MR image artifact localization and quantification
• Generative adversarial networks (GANs) for MR motion correction
• Semantic segmentation of organs and tissues
• Automatic quality control measures
• Biomarker feature extraction
• Framework development for NORA browser

In analysis step:

• Treatment response predicition
• Combination of imaging and non-imaging data
• GANs for image synthesis
• Biological age estimation

Proposed topics

The proposed topics here are a selection of the current research projects. If you are interested in any other topic, please refer to the Github projects or software and code documentation and contact one of the employees directly.

Motion correction

Contact: ; 

Patient, respiratory and cardiac movements produce motion artefacts in the final image which reduce the quality and impair reliable diagnosis. The task is to accurately resolve, detect and correct the movements in acquisition, reconstruction and post-processing. High spatial-and temporal-resolved MR images need to beacquired over several respiratory and cardiac cycles under free movement conditions. Motion-resolved images are then reconstructed under the consideration of a derived motion model and surrogate signals guiding the motion model. Motion models are estimated by image registrations (conventional and deep learning) and self-supervised random walks. Motion correction can also be paired with image reconstruction.

• deep-learning based reconstruction
• random walks
• optical-flow and deep-learning image registration
• generative adversarial networks
• sensor fusion

Image reconstruction

Contact: ; 

Magnetic resonance (MR) imaging plays an important role in medical diagnostic imaging due to its high soft-tissue contrast and its non-invasive acquisition. One major drawback, however, is often the relatively long examination time. Therefore, acceleration of MR data acquisition is of great interest. One way to achieve this is to reduce the number of phase-encoding steps for a fixed resolution in Cartesian raw data acquisition. The resulting sub-Nyquist acceleration introduces aliasing artifacts in the image. Aliasing-free images can be recovered if (1) the images are compressible, i.e. can be sparsely represented in some transform domain, if (2) are randomly and incoherently sampled and if (3) a non-linear reconstruction is employed. Further consideration need to be taken for sharing of spatio-temporal-quantiative information and to handle the high-dimensional data processing efficiently.

• Compressed Sensing
• deep-learning based image reconstruction
• model-based reconstruction (motion, parametric)
• low-rank + sparse methods
• super-resolution

Image quality assessment and control

Contact:

Magnetic resonance imaging is a flexible imaging modality, but can be prone to a variety of imaging artefacts which can be grouped as: patient-related, hardware/scanner-related and signal-processing-related. The artefacts can degrade the derived image quality and impair a reliable diagnosis. In order to provide a quantitative score, the human visual system needs to be mimicked which requires manual labeling by expert readers. Since expert labeling can be a time- and cost-intensive task, strategies for minimizing labeling effort need to be explored. The work deals with accurately detecting and correcting artifacts to provide an objective and reliable quality control measure during acquisition and/or post-processing.

• feature extraction
• classification
• reinforcement learning
• active learning
• semi-supervised learning

Semantic segmentation

Contact:

Automatic segmentation of organs or tissue compartments in whole-body imaging is an important pre-requisite for any further analysis. We develop Deep learning based algorithms for the automatic detection of landmarks and the segmentation of organs, tissue compartiments or tumors. Our work focuses on the development and transfer of computer vision and machine learning techniques to achieve accurate results with a small number of labels and to enable the generalization to new environments.

• Self supervision and contrastive learning
• Weakly- and semi supervised learning for computer vision tasks
• Attenion mechanisms and Graph Neural Networks
• Domain adaptation and domain generalization
• Geometrical and statistical shape analysis

Probabilistic machine learning

Contact: ; ; ;

A variety of problems in medical imaging can be traced back to elementary problems from statistics and machine learning. The goal of our research is to relate fundamental theoretical insights to real-world problems. Our key topics are the quantification of uncertainty and the question of how causal relationships in observational data can be identified and leveraged in machine learning applications.

• Uncertainty and calibration for DL
• Generative models, identifiability and approximative inference
• Disentanglement and (causal) representation learning
• Causal discovery with hidden confounders
• Invariant mechanisms for domain generalization

For student projects/thesis, expertise in probability theory, mathematical statistics, stochastic processes and/or probabilistic/statistical ML are of advantage. Very good grades in undergraduate mathematics and related courses are expected.

Generative models

Contact:

The task of translating medical images between different domains has numerous useful applications. One example is the correction and restoration of artifact-corrupted images. Another potential application is the generation of novel image data between different modalities. The work focuses on the development and refinement of medical image translation frameworks and their applications - specifically, GAN-based MR motion correction, GAN-based PET attenuation correction and GAN-based image inpainting. Furthermore, the modelling and Inclusion of the uncertainties can steer the generative task and provide additional information.

• variational autoencoders
• generative adversarial networks
• uncertainty estimation

Prerequisites

• Highly motivated, independent and structured way of working
• Interest in machine learning, deep learning and signal processing
• Studies in the field of electrical engineering, informatics or Medizintechnik
• Good German and/or English skills (spoken and written)
• Programming expertise in Python is beneficial

---

Current ongoing and previous thesis

2021

2020