Education

Bachelor's degree in Artificial Intelligence from Johannes Kepler University Linz, and a master's degree with only the final exam left. Both theses were done with the Medical University of Vienna, on generative models for retinal imaging.

Degrees

Oct 2023 – present

MSc Artificial Intelligence

Johannes Kepler University Linz

Status: Final exam pending
Current average: 1.6 (Austrian scale), GPA 3.40 / 4.00
Thesis: Generating Counterfactual OCT Retinal Images: Training Generative AI Models (Diffusion) for Class-Specific Image Transformation
Partner: In cooperation with the Medical University of Vienna
Coursework: Computer Vision, AI in Life Sciences, Sequence Analysis

Oct 2020 – Sep 2023

BSc Artificial Intelligence

Johannes Kepler University Linz

Grade: 1.9 (Austrian scale), GPA 3.10 / 4.00
Thesis: Generative models, synthetic data and counterfactuals: applications in retinal imaging
Partner: In cooperation with the Medical University of Vienna
Coursework: Unsupervised Machine Learning, AI in Medicine, Natural Language Processing

Master's thesis

Illustration of the idea, not model output. Currently showing: healthy.

An OCT scan is a cross-section of the retina. I adapted NVIDIA's MAISI medical-image synthesis framework, built for 3D CT and MRI, to 2D OCT, and asked two questions. Can its generated scans help train disease classifiers? And can it edit a real scan into a counterfactual: the same eye, showing a different disease?

The pipeline was built from scratch on the Kermany dataset (CNV, DME, drusen, normal) and extended to RETOUCH scans with pathology masks, using patient-level splits so no patient appears in both training and test data.

+0.071: Mean Macro F1 gain for a disease classifier trained with half synthetic data when only 10% of the real data is available, over ten seeds.
0.892: SSIM of the 256 px adversarial VAE (reconstruction FID 5.00), whose 4 × 64 × 64 latent space made a 256 px diffusion model trainable.
1,000: Images in the balanced, patient-disjoint test set used for every evaluation.

What worked

A reproducible VAE and class-conditional latent diffusion model (rectified flow) for OCT.
Synthetic scans improved classifiers in low-data settings, both added to real data and replacing part of it.
Mask-based editing (Dif-fuse-style inversion with masked latent fusion) confined changes to fluid and detachment regions and left the rest of the scan intact.

What remains open

A generic healthy prior did not reliably reconstruct plausible retinal anatomy inside edited regions.
Edits could fool a classifier, especially for DME, without showing convincing disease morphology, so classifier scores alone cannot validate a counterfactual.
Next steps: anatomy-aware inpainting objectives, layer-aware evaluation and blinded review by retina specialists.

Research and challenges

Oct 2024

MARIO challenge finalist, MICCAI 2024

MICCAI 2024, Marrakech

Methods for diagnosing and monitoring age-related macular degeneration from OCT. Presented in Marrakech; the paper was published by Springer.

OCT
AMD
classification
medical imaging

Nov 2024 – Sep 2025

Student Researcher, Medical University of Vienna

Generative models for retinal OCT; the research behind my master's thesis. Details in the CV.

Apr 2024 – May 2024

Fellow, Linz Institute of Technology

A BERT-embedding recommender for non-numerical data, plus LLM-based accessible PDF conversion. Details in the CV.

Awards

Jun 2024

1st place, Danube Dynamics image segmentation challenge

Winning entry in an image segmentation competition.

Mar 2023

3rd place, AI Cloudflight Coding Contest

Timed AI coding contest at JKU.