About Me

I work on computer vision problems. My interests include explainable AI, deep metric and representation learning and their interdisciplinary applications.

Resume [PDF]

Most Recent Work

I am currently working in Roche's phc department in the Data, Analytics & Imaging group. I utilize deep learning and self-supervised methods to create cutting-edge solutions for medical image analysis and various downstream tasks in digital pathology. This role combines healthcare, product development, deep learning and bio-medical image analysis, making it a challenging and rewarding place to work. Learn more about the work we do at Roche's phc here. Our first results on foundational models in digital pathology have been published on arXiv.

Shorted abstract of our paper Aggregation Model Hyperparameters Matter in Digital Pathology:

Traditional evaluation methods in digital pathology rely on fixed aggregation model hyperparameters, a framework we identify as potentially biasing the results. Our study uncovers a co-dependence between feature extractor models, often termed foundational models, and aggregation model hyperparameters, indicating that performance comparability can be skewed based on the chosen hyperparameters. By accounting for this co-dependency, we find that the performance of many current feature extractor models is notably similar.

Projects

Master thesis: “Tackling Challenges and Enhancing Disease Diagnosis in Medical Imaging with Deep Metric Learning” supervised by Prof. Ommer (LMU Munich) and examined by Prof. Hamprecht (University of Heidelberg)

• Proposed a novel deep metric learning approach for classifying lung diseases in X-rays using disease-conditioned similarity networks.
• Addressed challenges like image variability, class imbalance, and domain shifts.
• Outperformed or on par with state-of-the-art methods.
• Analyzed weakly-supervised and fully supervised methods, as well as different models and losses.

Bachelor thesis: “Conditional Similarity Learning for Multilabel Classification of Medical Images” supervised by Prof. Ommer (Machine Vision and Learning Group).

• Created new deep metric learning approach to classify thorax diseases using TensorFlow
• Made it possible to learn more about the similarity of diseases by creating a disentangled representation space
• Increased classification ROC AUC score by 6% over the baseline

Hackathon challenge: “Bias Free AI” by IBM at Q Summit in Mannheim

• Implemented efficient algorithm in Python to solve the problem of a fair assignment of teachers to a school
• Used machine learning techniques such as k means clustering and k nearest neighbor classification
• Applied solution to real world data and won 2nd place

University project: “Galaxy Detection and Classification in the Hubble Deep Field”

• Trained CNN to classify 10 different galaxy types
• Performed classification with accuracy of 86%
• Designed end-to-end approach to detect and classify galaxies in images of the Hubble Deep Field

University project: “Eigenfaces” R package

• Created R package in a team of 3 that enables simple principal component analysis for different datasets such as human faces
• Tested package and wrote extensive documentation
• Implemented interactive shiny application

Details