Teaching

Teaching and education are an integral part of our institute's mission. All of our courses—which are heavily influenced by our research—are taught in English. We offer lectures for students from various disciplines, but our core lectures are aimed at computer science students:

Winter semester
Summer semester 2025 (SS25)

In addition to the main lecture series, the individual sub-areas and subject areas are explored in depth through practical sessions and seminars. Through direct interaction with our lecturers, students can deepen and apply their acquired knowledge.

As an elective for medical students, we offer:

Computer Science for Medical Students

This course offers students exciting insights into the world of AI methods (especially neural networks) and their applications in medicine. In addition to acquiring basic theoretical knowledge, they gain initial practical experience with Python programming and have the opportunity to train their own neural networks.

In addition to lectures, seminars, and internships, we offer a range of bachelor's and master's projects. You can find our current calls for open projects here:

Master's Thesis

Enhancing Neuroradiology Reports with Agentic Deep Learning and Vision-Language Models

 

This project is conducted in collaboration with Cornell University and the Radiology Department at Presbyterian Hospital in New York City. It can be pursued as a Master’s Thesis, Guided Research/IDP, or a Working Student position—or any combination of these formats (a total duration of more than 6 months is preferred). Read more

 

Unsupervised 3D Clustering and Quantification of Pathological Hotspots in Orthopaedic SPECT/CT Imaging


This project aims to develop an unsupervised learning pipeline for the automatic detection and quantification of abnormal uptake regions ("hotspots") in 3D SPECT/CT scans used in orthopaedic diagnostics. Unlike traditional methods that rely on manual interpretation or fixed thresholds, the model will extract clinically relevant features—such as volume, intensity, and signal variation—to enable more objective and reproducible assessments. Read more

 

Multimodal Learning from Pre- and Postoperative Imaging for Orthopedic Surgery

 

This project aims to enhance orthopedic surgical planning and assessment—particularly for knee arthroplasty—by leveraging machine learning models. Radiographs, though widely used, often require expert interpretation, leading to variability across clinicians and institutions. The project addresses challenges in implant sizing and device identification by using real-world multimodal data and tackling issues like uncertainty and bias. By jointly analyzing pre- and postoperative radiographs with modern ML techniques, the goal is to support more accurate, consistent, and automated clinical decision-making. Read more

 

Early Alzheimer’s Disease Prediction Using Multimodal Longitudinal Data

 

Alzheimer’s Disease remains a devastating and incurable condition, where early prediction can make a meaningful difference in patient care. This thesis explores a novel direction by moving beyond static baseline data and tapping into the power of longitudinal multimodal information—spanning clinical and MRI measurements—from the ADNI dataset. By capturing temporal patterns over time, it seeks to not only forecast the likelihood of developing Alzheimer’s within a five-year window but also estimate when the disease might begin. Read more

 

LLM/VLM-based AI Agents Workflow for Simplifying Medical Image Analysis

 

This project, in collaboration with the University of Oxford, aims to simplify medical image analysis by using LLM/VLM-based agent systems. These models reduce the need for deep learning expertise, automating pipeline development for tasks like cardiac MRI analysis. The project will evaluate the performance of different LLM/VLM backends based on success rates, cost-effectiveness, and prompt clarity. Read more

Medical Image Segmentation Topics

 

This thesis delves into advanced medical image segmentation, focusing on soft segmentation, label masking, resolution optimization, and point-of-interest prediction. It aims to enhance segmentation accuracy, detect mislabels through gradient analysis, and explore model extrapolation for unseen structures. Read more

 

Multi-modal Longitudinal Heterogeneous Aging in UK Biobank

 

This thesis aims to enhance age prediction by combining imaging and non-imaging biomarkers. By training models on a healthy cohort, it seeks to identify accelerated and decelerated aging. The plan includes a literature review, developing multi-modal models with MRI and radiomics, and analyzing organ-specific aging patterns and lifestyle factors over time. Read more

 

Multimodal Deep Learning for Predicting Osteoarthritis Pain Trajectories

 

Knee osteoarthritis (OA) affects over 650 million people globally. Although radiographic severity is commonly used for assessment, pain is the most impactful symptom and directly affects patients’ function and quality of life. Pain progression varies widely which indicates the presence of distinct pain trajectories.
This thesis aims to use deep learning with multimodal data to identify and predict individual OA pain progression patterns. Read more

 

Scaling foundation models for time series analysis

 

In our previous work (Turgut et al., 2024), we developed a foundation model for time series analysis capable of handling diverse tasks, ranging from weather forecasting to disease classification. 
Our approach has shown strong generalisation capabilities across multiple domains, and we are now looking to scale and optimise our model to unlock the next generation of time series models. The goal of this master thesis is to successfully scale our time series foundation model, building upon state-of-the-art practices from large language models (LLMs) and recent research in model optimisation. Read more

 

Automated medical report generation from doctor-patient conversations in an orthopaedic ambulatory clinic

 

The increasing documentation burden in everyday clinical practice is considered a factor contributing to the growing strain and exhaustion of medical staff [1]. Studies show that physicians spend an average of around 37% of their working time on documentation in both inpatient and outpatient care [2]. At the same time, recent advances in natural language processing (NLP) and automatic speech recognition (ASR) are enabling new approaches to support medical documentation. The aim of this master's thesis is the automatic generation of medical reports from recorded doctor-patient conversations, thereby helping to reduce the documentation workload in clinical practice. Read more

 

Data-Efficient Alignment Techniques for Text Generation Tasks

 

Aligning large language models (LLMs) with human preferences is essential for safe and useful text generation. Traditional approaches like Reinforcement Learning with Human Feedback (RLHF) are data- and resource-intensive. Recent research explores hybrid alignment methods that combine small amounts of annotated data with model-generated supervision. Techniques such as SPPO, I-SHEEP, and RS-DPO show that models can iteratively refine themselves by generating, evaluating, and learning from their own outputs, often through mechanisms that allow gradients to pass through the model multiple times. This project investigates such techniques to achieve efficient alignment with minimal human input. Read more.

 

Activation Steering for Alignment of Medical Large Language Models

 

Large language models (LLMs) are increasingly being used in medical contexts, from diagnostic assistance to patient communication. However, LLMs can exhibit unpredictable behavioral shifts that pose serious risks in healthcare settings - from overconfident medical advice to biased treatment recommendations. Read more

 

Super-Resolution MRI Guided by Cardiac Scout Imaging

 

In cardiac MRI protocols, low-resolution non-diagnostic scout scans are typically acquired prior to the more detailed, diagnostic cardiac cine MRI sequences. Although these scout images are static (i.e., possessing no temporal information) and exhibit lower overall image quality, they provide a better 3D overview of the heart when compared to conventional cardiac cine MRI. This project aims to explore super-resolution techniques to reconstruct high-quality 2D and potentially 3D cardiac views from these scout scans. The goal is to bridge the quality gap between scout and cine acquisitions, potentially enabling faster or even scout-only cardiac imaging pipelines in the future. Read more

 

Multi-Plane Cardiac View Synthesis from Cardiac Scout MRI

 

Scout MRI images, while lower in quality, offer a high density of slices through the cardiac volume, making them a valuable yet underutilized source for advanced image generation. This project focuses on developing methods to synthesize arbitrary or novel cardiac planes—including oblique or anatomically standard views (e.g., 4-chamber, short-axis)—from scout acquisitions. The goal is to enable view synthesis and plane interpolation, possibly powered by generative models, to support planning, diagnosis, or even act as a substitute for software-based acquisitions in time-constrained scenarios. Read more

IDP/Thesis

Physics-based deep learning for hyperspectral neuronavigation

 

Hyperspectral imaging (HSI) analyzes the electromagnetic spectrum to detect physical and biochemical properties. In the HyperProbe project, the goal is to develop an AI-powered imaging system for brain tumor surgery, using HSI to identify biomarkers of healthy and tumor tissue. Read more