Bachelor and master thesis will be offered from various research areas. Specific topics will be defined by the chair or in collaboration with the student.

To write a thesis at the chair of Explainable Machine Learning, the following qualifications need to be fulfilled:

  • successful test in a modul with lecture and exercise in Deep Learning, Machine Learning or Introduction into AI
  • successful participation at one of the chair offered seminars or projects

Open theses

Please refer to VC [Link] for further details.

Current theses

  • "Unveiling CNN Layer Contributions: Application of Feature Visualization in Medical Image Classification Tasks" - Jonida Mukaj supervised by Ines Rieger
  • "Generative Data Augmentation in the Embedding Space of Vision Foundation Models to Address Long-Tailed Learning and Privacy Constraints" - David Elias Tafler supervised by Francesco Di Salvo
  • "Evaluation and feasibility of selected data-driven Machine Learning approaches for Production Planning to enhance Order Sequencing and to improve OEE in Manufacturing" - Nicolai Christian Frosch supervised by Christian Ledig
  • Markus Brücklmayr - supervised by Christian Ledig
  • Erik-Jonathan Schmidt, Zusammenarbeit mit Stabilo- supervised by Christian Ledig
  • Aaron-Lukas Pieger, Zusammenarbeit mit Stabilo- supervised by Christian Ledig
  • Junquan Pan - supervised by Christian Ledig
  • Pascal Cezanne - supervised by Sebastian Dörrich
  • Michael Sebastian Schick - supervised by Sebastian Dörrich
  • Julius Stutz - supervised by Sebastian Dörrich
  • Marius Ludwig Bachmeier - supervised by Sebastian Dörrich
  • Peiyao Mao - supervised by Francesco Di Salvo

Finished theses

„Human Activity Recognition via Deep Learning based on active exoskeleton data” - Christoph Zink

Autor: Christoph Zink, betreut von Prof. Dr. Christian Ledig

To optimize the behavior of exoskeletons with built-in motors, so-called active exoskeletons, real-time classification of the user's activity has become increasingly popular in recent years. However, no studies have evaluated the performance of different neural network architectures for this task in a real-world scenario where both the observed locations and the observed subjects are absent from the training data. To fill this gap, this study trained four architectures of neural networks and compared their performance on a self-recorded test set containing 10 new subjects at 5 previously unseen locations. In addition, a comparison was made with a model representing the standard approach prior to the advent of deep learning models in the field, to answer whether deep learning models consistently classify better in a robust manner.

The results indicate that deep learning models are overall well suited to this task, i.e. all neural networks outperformed the baseline method. Regarding the robustness of the models, it appears that the neural networks can generalize well beyond the single location and few subjects observed within the training data. This ability to robustly generalize appears to be strongly dependent on the overall amount of training data available, i.e. the models generalized rather poorly when applied to activities rarely observed within the training data.

The study resulted from a cooperation with the company German Bionic, an Augsburg-based producer of active exoskeletons intended for the relief of the lower back. For the study data coming from the IoT-connected device Cray-X was used, which is depicted in the attached picture.

Link to theses (2.1 MB)

"CNN-based Classification of I-123 ioflupane dopamine transporter SPECT brain images to support the diagnosis of Parkinson’s disease with Decision Confidence Estimation"- Aleksej Kucerenko

Author: Aleksej Kucerenko, supervised by Prof. Dr. Christian Ledig and Dr. Ralph Buchert

Parkinson's disease (PD) is a prevalent neurodegenerative condition posing significant challenges to individuals and societies alike.

It is anticipated to become a growing burden on healthcare systems as populations age.
The differentiation between PD and secondary parkinsonian syndromes (PS) is crucial for effective treatment, yet it remains challenging,
particularly in cases of clinically uncertain parkinsonian syndromes (CUPS).
Dopamine transporter single-photon emission computed tomography (DAT-SPECT) is a widely used diagnostic tool for PD,
offering high accuracy but also presenting interpretational challenges, especially in borderline cases.

This study aims to develop reliable automated classification methods for DAT-SPECT images, particularly targeting inconclusive cases,
which may be misclassified by conventional approaches.
Convolutional neural networks (CNNs) are investigated as promising tools for this task.
The study proposes a novel performance metric, the area under balanced accuracy (AUC-bACC) over the percentage of inconclusive cases,
to compare the performance of CNN-based methods with benchmark approaches (SBR and Random Forest).
A key focus is the training label selection strategy, comparing majority vote training (MVT) with random label training (RLT),
which aims to expose the model to the uncertainty inherent in borderline cases.
The study evaluates the methods on internal and external testing datasets to assess generalizability and robustness.

The research was conducted in collaboration with the University Medical Center Hamburg-Eppendorf (UKE).
The dataset utilized for model training originated from clinical routine at the Department of Nuclear Medicine, UKE.
The attached figure showcases augmented versions for two sample cases from the dataset:
a healthy control case ('normal') and a Parkinson's disease case ('reduced') with reduced availability of DAT in the striatum.

The study addresses the need for reliable and automated classification of DAT-SPECT images,
providing insights into improving diagnostic accuracy,
reducing the burden of misclassifications and minimizing the manual inspection effort.

Link to thesis(12.5 MB)

Benchmarking selected State-of-the-Art Baseline Neural Networks for 2D Biomedical Image Classification, Inspired by the MedMNIST v2 Framework"- Julius Brockmann

Author: Julius Brockmann, supervised by Sebastian Dörrich

This thesis examines the benchmarking of state-of-the-art baseline neural networks in the field of 2D biomedical image classification. Focusing on the effectiveness of deep learning models on high-quality medical databases, the study employs pre-trained baseline networks to establish benchmarks. The research investigates four convolutional neural
networks and one transformer-based architecture, exploring how changes in image resolution affect performance. The findings highlight the advanced capabilities of newer convolutional networks and demonstrate the effectiveness of transformer architectures for handling large datasets. Common misclassifications and their causes are also briefly analyzed, offering insights into potential areas for improvement in future studies.

Link to thesis(8.4 MB)

"Development of a dataset and AI-based proof-of-concept algorithm for the classification of digitized whole slide images of gastric tissue"- Tom Hempel

Author: Tom Hempel, supervised by Prof. Dr. Christian Ledig

The thesis focuses on the development of a dataset and AI algorithms for classifying digitized whole slide images (WSIs) of gastric tissue. It details the creation and meticulous annotation of the dataset, which is crucial for effectively training the AI. The process involved gathering, anonymizing, and annotating a vast array of WSIs, aimed at building robust AI models that can accurately classify different regions of the stomach and identify inflammatory conditions.

Two AI models were developed, one for assessing gastric regions and another for inflammation detection, achieving high accuracy in areas like the antrum and corpus but facing challenges with intermediate regions due to dataset limitations and the specificity of training samples.

The challenges encountered during the dataset creation, such as data collection and the necessity for detailed annotation to ensure data integrity and privacy, highlight the complexity of this research.

The dataset and initial models serve as a foundation for further research by Philipp Andreas Höfling in his master thesis, aiming to refine these AI algorithms and enhance their utility in medical diagnostics.

Link to thesis(5.1 MB)

"Development of an AI-based algorithm for the classification of gastric tissue for computational pathology"- Philipp Andreas Höfling

Author: Philipp Andreas Höfling, supervised by Prof. Dr. Christian Ledig

Computational pathology has significantly advanced in recent years, yet a notable gap exists in the specific area of gastric tissue research. To address this issue, this study focuses on developing AI algorithms specifically for classifying gastric tissue types and inflammation caused by gastritis.

Using two ResNet18 models, trained on annotated tiles from over 200 slides, high accuracy in both inflammation and tissue type classification has been achived. However, challenges remain, especially in generalizing inflammation detection across all types of gastritis.

While promising, further research is needed, including expanding datasets and refining annotations, to fully harness AI's potential in gastric tissue analysis.

Link to thesis(7.7 MB)

"Component ldentification for Geometrie Measurements in the Vehicle Development Process Using Machine Learning" - Tobias Koch

Autor: Tobias Koch, supervised by Prof. Dr. Christian Ledig

Geometric measurements are frequently performed along the virtual vehicle development chain to monitor and confirm the fulfillment of dimensional requirements for purposes like safety and comfort. The current manual measuring process lacks in comparability and quality aspects and involves high time and cost expenditure due to the repetition across different departments, engineers, and vehicle projects.

Thereby motivated, this thesis presents an automated approach to component identification, leveraging the power of Machine Learning (ML) in combination with rule-based filters. It streamlines the geometric measurement process by classifying vehicle components as relevant or not and assigning uniformly coded designations. To determine the most effective approach, the study compares various ML models regarding performance and training complexity, including Light Gradient-Boosting Machines (LightGBMs), eXtreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), and Feedforward Neural Networks (FNNs).

The results indicate that the integration of ML models can substainally improve the geometric measurement process in the virtual vehicle development process. Especially LightGBM and CatBoost proved to be the most capable models for this tasks and offer promising progress in the virtual development of vehicles.

Link to the code

Link to the thesis(3.6 MB)

Further finished theses