Medical Artificial Intelligence
Needle Detection in Interventional MRI (2026)
This project addressed reliable needle localization in low-quality, strongly undersampled real-time MRI. I worked across the full pipeline—from acquisition design (synthetic data, ex-vivo and in-vivo phantoms) with MR physicists, through phantom and in-vivo studies, to selecting architectures and training early AI models that distinguish needle artifacts from undersampling artifacts. The result was a working prototype evaluated on real data. One of those rare projects where everything from physics to learning actually had to come together.
EuCanImage (2024)
In a large EU consortium, the challenge was to process and utilize heterogeneous CT and MR datasets for liver lesion analysis. I focused on data ingestion, conversion, and curation, and developed a detection and classification pipeline. The outcome was a functioning system operating on curated multi-center data. Working with that many partners mostly means alignment and communication with partners as diverse as small companies, universities, hospitals and data centers, as well as legal teams to get data transfer agreements in order.
CLINIC 5.1 (2023)
This project aimed to analyze longitudinal electronic health records in FHIR format to predict prostate cancer recurrence. I contributed to developing and deploying a Kubernetes-based on-site web service and coordinated a small team building the system. The result was a deployed prototype operating within a clinical environment. Getting models to run is easy—getting them to run where they are actually needed is quite a process.
CancerScout (2020)
This work focused on large-scale whole slide imaging (WSI) for cancer detection and subtype prediction. I handled hosting and maintenance of the EXACT annotation platform on Azure and implemented processing pipelines for large pyramidal tiled TIFF data. CNN-based models were applied for segmentation and classification.
My Health My Data (2019)
This project explored privacy-preserving AI infrastructure under GDPR constraints. I worked on a Node.js service for data access and model training, while engaging with concepts like federated learning and homomorphic encryption. A good reminder that legal and infrastructure constraints often shape what AI can realistically do.
More info →
Information Storage at Molecular Level (2012)
This side-project was a collaboration with a chemistry department and analyzed very special microscopy data showing individual molecules that could encode binary states through orientation. I developed image analysis software to study flipping behavior at room temperature. The result provided quantitative insight into molecular stability. Seeing individual atoms on a screen still feels slightly unreal.
Calibration in Computed Tomography
Building a CT Scanner (2019)
This project involved designing and assembling a portable CT scanner for field research on ants in Panama (no kidding). I contributed calibration methods and hardware integration using custom phantoms and software. The system was successfully deployed in a non-laboratory environment. Not a typical CT application—but definitely one of the more memorable ones.
View larger image →
Flexible CT Calibration (2018)
I invented this calibration method using projective invariants and custom phantoms with metal bead configurations. I developed the approach to establish 2D–3D correspondences for non-standard CT trajectories. The method was implemented and patented by Siemens Healthineers.
Estimating Heart Phase in Rotational Angiography (2017)
This collaboration applied epipolar consistency to angiography data under truncation constraints. A colleague came up with the idea of using digital subtraction angiography to address the truncation problem. The result demonstrated feasibility in a setting where standard assumptions break down.
Epipolar Consistency in Transmission Imaging (2015)
My PhD work focused on detecting inconsistencies in CT projections using epipolar geometry. I implemented GPU-based computation (CUDA) and supporting tooling, and released the framework publicly. The method was later used in industry for motion and artifact correction and as of writing this, seven years later, there is still enough interest - I was recently asked by a company to freelance and give a seminar on the matter.
Thesis →
Measuring Reproducibility of Medical C-arm Motion (2014)
This project evaluated whether C-arm systems can reproduce trajectories accurately enough for one-time calibration. I analyzed geometric consistency and its impact on reconstruction. The results clarified limits of current systems and provided me with an in-depth understanding of CT geometry.
Multi-modal 3D-3D Registration (2009)
During an internship, I developed an optimization framework for CT–ultrasound registration and worked with GPU programming beyond graphics. This contributed to aligning different imaging modalities. Also my first time in the US, which didn’t hurt.
Augmented Reality and Visualization
AR for Orthodontic Applications (2012)
My MSc thesis focused on registering dental CT data with RGB video for AR-based brace placement. I implemented the registration pipeline and visualization. The system demonstrated feasibility and was technically interesting, though practical relevance is limited.
X-ray Visualization (2011)
This project explored enhancing depth perception in X-ray imaging using prior CT information. I contributed to methods for blending 2D ultrasound and colorizing based on depth. The result improved interpretability of fluoroscopic data. Although this actually works, I doubt this will ever be adapted in clinical practice.
Example →
Direct Volume Rendering (2010)
I built a volume rendering library from scratch using Eigen and GLUT. It supported multiple rendering modes, transfer functions, and depth handling with geometry. The result was a flexible visualization framework. Writing everything yourself is painful—but you a personal relatoinship to every pixel on the screen.
View larger image →
Projective Invariants for Tracking Targets (2010)
I developed a method using geometric invariants to detect markers and establish correspondences for AR tracking. It was applied in a custom AR setup. The result was a working tracking approach under constrained conditions. Geometry strikes again.
Self-build Augmented Reality Hardware (2010)
Before consumer AR existed, I built a custom system using a head-mounted display and PlayStation cameras. I implemented tracking and calibration for a functional prototype. The result was a fully working, if improvised, AR system. Definitely more enthusiasm than engineering discipline—but effective.
Depth Perception in Augmented Reality (2008)
I suggested improving depth perception in AR by segmenting a colored glove to ensure correct occlusion. This was my first contribution to a research paper. The method worked reliably in the setup. Simple idea, surprisingly effective.
Focus & Context for AR (2007)
I integrated OpenGL/GLSL-based visualization techniques into an AR framework about two years into my bachelor studies. This involved combining rendering and interaction concepts. Making this work in 2007 involves detailed understanding of tansformations from multipe tracking systems, managing latency and tuning compute to mach hardware constraints.
More info →Other Projects
LibGetSet (2012)
I developed a header-only C++ library for managing hierarchical parameter trees with typed properties. It has been used in many of my projects for building simple GUIs. The result is a reusable and lightweight tool. Still using it today, which probably says enough.
GitHub →
BOMBAflop (2003)
My first complete software project at age 16: a multiplayer game written in C with OpenGL and particle effects. It supports four players on a single keyboard. The result was a fully playable game. Not sharing the code—but still proud of it.
View larger image →