Themen in Magdeburg

Master Thesis: Deep Learning Based Segmentation Task of medical CT-Images based on advanced Preprocessing

Current state:

The chances of success of tumor treatment are highly dependent on the patient’s physical condition. In everyday clinical practice, the patient’s BMI is calculated for this purpose. However, this is a rather inaccurate measure, since the distribution of muscle to fat tissue is a decisive indicator. For a more accurate evaluation, the patient’s CT images must be evaluated. However, this is a time-consuming task.

Scope of the thesis:

This work is intended to address the problem. Currently, data are being acquired in clinical practice and segmented by experts. These are CT data sets in which muscle and fur tissues were segmented in one layer. Your task is to create an automatic segmentation using Deep Learning methods. Subsequently, the segmented regions are to be evaluated with the help of a measure. The explicitly mentioned preprocessing step is to split the given segmentation (symmetry of the body) to provide more data to the network during the learning process. An optional extension would be the automatic selection of the layer in which the evaluation should take place.

We offer:

  • interesting clinically relevant research
  • support in technical questions and writing of the thesis

We expect:

  • good programming skills (Python)
  • knowledge of image processing
  • experience with Deep Learning and frameworks (Pytorch, Tensorflow, Keras)
  • good study achievements

3D Deep learning for wall shear stress prediction of intracranial aneurysms

Wall shear stress is a parameter derived from hemodynamic simulation and can be used in the diagnosis of intracranial aneurysms. We want to train a neural net to predict areas of high wall shear stress in intracranial aneurysms. This is a research oriented topic. Beside familiarization with recent research in deep learning on 3d structures it requires initiative and own ideas to advance the ongoing research.
Material: surface meshes of (artificial) aneurysms and results of hemodynamic simulations
Requirements: Programming experience (python), Experience with deep learning We expect high-qualified students interested in this project (team projects, bachelor or master thesis. Please send your application!

Master Thesis: Bridging the Domain Gap: Visualization Support for Transfer Learning on Time Series Data

End-of-line testing is an essential step in the production process to validate the functionality of units near the end of the production line. Defective products or those not matching manufacturing tolerances must be rejected before the products are shipped. To shorten production cycles, automatic testing increasingly replaces manual inspection performed by human operators. A unit under test is exposed to a stimulus and its response is recorded by different sensors. The resulting multivariate time signals are analyzed for defect identification and classification.
The goal is to bridge the domain gap between simulated and measured unit responses, such that a classification algorithm learned on simulation data can be applied to new (unlabeled) real-world data from sensor measurements. The units under test are electric motors, whose current dt. Strom signals are to be used for a defect classification. The transfer learning problem is characterized by a shift in the input domain (i.e. simulated vs. measured signals of the same motor) while the analysis task remains the same (i.e. classify the defect). The data are multivariate time series that are simulated/measured across various operating conditions. For defect-free motors, both simulated and corresponding measured signals are available. For defective motors, however, only simulated time series with corresponding defect labels are available. The lack of real-world training data for defective motors raises the need to learn from the simulated data a well-performing classifier that can be applied to real-word measurements.

Interactive VR Visualization to Assess the Collision Probability with Space Debris (joint topic with the German Aerospace Center)

Description: The increasing amount of space debris in earth orbit poses a growing threat to space travel. Therefore, it is very important to know where space debris is located in orbit and whether there is a possibility of collisions with satellites or space crafts in operation.
The goal is to develop a VR software that visualizes orbital objects. The prototype should facilitate the evaluation of possible collisions of objects and support the decision whether, for example, a course correction of a satellite is necessary. However, the determination of the position of objects in orbit is associated with uncertainties. Various influences on objects in Earth orbit, such as the interaction with the atmosphere or variations in the gravitational field, lead to a deviation between the actual position and the position calculated from the observation data.


  • Research on existing debris visualizations and satellite propagation
  • Implementation of a real-time visualization of debris positions and collision probabilities of objects in orbit
  • Development of methods for the targeted, user-guided interactive analysis of space situations


  • Study of computer science or comparable fields of study
  • Knowledge of computer graphics
  • Experience with software development in Unity/C#

Master Thesis: Mesh generation with machine learning

Extension of Shrinkingtubemesh-generation (as illustrated below) with machine learning. The current version was written in matlab and is only suitable for cylinder-like structures, for example vessels. The program should be adjusted to fit a wider variation of shapes.
Subtasks: The task can be solved in two ways: using classical machine learning or with deep learning.

Option 1: Maschine Learning

  • including development of a range of suitable startshapes, definition of point cloud features for machine learning, generation of a Testdatabase, usage of Machine Learning to predict a suitable startshape and shrinkingtubemesh-algorithm parameter for a given pointcloud .

Option 2: Deep Learning:

  • Generate startshapes (simple, roughly the pointcloud describing meshes) using deep learning (for example using a pointcloud to mesh approach like AtlasNet);
  • use these startshapes for the shrinkingtubemesh generation and compare to other mesh generation approaches .

Requirements: Knowledge of Python (Pytorch) and Matlab; Experience in Machine Learning/Deep Learning

We expect high-qualified students interested in this project (hiwi job / student assistant or team projects, bachelor or master thesis). Please send your application!

DL Segmentation of Meningiomas

We need you for our brain tumor segmentation project!
We want to support our clinical cooperation partners from the University Hospital in Magdeburg. You will work with real medical data sets and you should develop a Deep Learning-based solution. Advantages: We have a Deep Learning server for remote work and the clinicians already provide sufficient ground truth data, so the data augmentation will be possible in feasible time.
We expect high-qualified students interested in this project (hiwi job / student assistant or team projects, bachelor or master thesis). Please send your application!

Klinische Entscheidungsunterstützung für die Therapie zerebraler Aneurysmen

Bei der klinischen Entscheidungsfindung werden klinische Richtlinien herangezogen, die auf Evidenzen basieren und Empfehlungen eines Gremiums von Experten beinhalten. Einige der Richtlinien sind auf logischen „Wenn-dann“-Regeln und komplexeren, mehrstufige Regeln aufgebaut. Obwohl diese Regeln als Algorithmus zur Entscheidungsunterstützung formalisiert werden können, liegen die Richtlinien zumeist nur in Textform vor und müssen für die klinische Routine in übersichtliche Handlungsempfehlungen „übersetzt“ werden. Innerhalb der Arbeit soll ein durch Ärzte bedienbarer Prototyp entwickelt werden, der die „Übersetzung“ einer klinischen Richtlinie in einen Algorithmus zur Entscheidungsunterstützung ermöglicht.

Anforderungen: Gute bis sehr gute Programmierkenntnisse

Interactive Blood Flow Exploration – In collaboration with Dept. of Neurology, OVGU and Inria, France

In cerebral aneurysm research, CFD simulations allow us to gain a better understanding of the dynamics of the blood flow. The simulated flow is often visualized using integral curves resulting in cluttered “spaghetti plots”. Advanced approaches group similar curves and show only selected representatives (image). These approaches however, fail in showing the clusters’ spatial extent. In this thesis, an interactive approach facilitating a continuous transition between the full set of integral curves and an uncluttered abstracted visualization shall be developed. Browsing back and forth through various levels of abstraction shall allow the user to grasp both, the general structure of the blood flow pattern as well as the spatial extent of individual substructures.

Requirements: Good to very good programming skills (C++) are mandatory

Detektion von Aneurysmen mit Deep Learning

Im Rahmen aktueller Forschungsprojekte werden Aneurysmen und ihre Durchblutung untersucht. Dabei besitzt ein Patient häufig multiple Aneurysmen, welche erst in 3D Bilddaten detektiert werden können. Ziel des Projekts ist der Einsatz von Deep Learning Techniken zur automatischen Aneurysmadetektion basierend auf einer annotierten Trainingsdatenbank. Die Aufgabe eignet sich als Teamprojekt, kann aber auch für eine Hiwistelle oder Abschlussarbeit angepasst werden.

Anforderungen: Gute bis sehr gute Programmierkenntnisse (Python / Matlab) sind erforderlich

3D-Stereoverfahren für die Herzchirurgie

Im Rahmen der Arbeit sollen Stereoverfahren für die 3D-Rekonstruktion von Strukturen aus intraoperativen Endoskopiebildern entwickelt werden. Die Arbeit wird in enger Kooperation zwischen der Fakultät Informatik der OvGU Magdeburg (Dr. Sandy Engelhardt) und der Herzchirurgie des Universitätsklinikums Heidelberg (Prof. De Simone) durchgeführt. Weitere Themen für Abschlussarbeiten sind vorhanden. Melden Sie sich gern bei Interesse.


  • Meshing einer rekonstruierten 3D-Punktewolke
  • Texturierung der Oberfläche
  • Fusion von verschiedenen Ansichten zu einem Mesh

Anforderungen: Programmiererfahrung in C++ (OpenCV- Kenntnisse hilfreich)