Lea Bogensperger

I am a postdoctoral researcher in the Krauthammer Lab at the Department of Quantitative Biomedicine at the University of Zurich. Prior to that, I worked as a doctoral researcher at the Institute of Computer Graphics and Vision, Graz University of Technology. I defended my PhD thesis "Variational Methods in Imaging Meet Machine Learning" in June 2024, which was supervised by Prof. Thomas Pock and with Prof. Carola-Bibiane Schönlieb as my external referee. During my PhD I worked on inverse problems in medical imaging, bilevel optimization as well as generative modeling using diffusion models and flow matching for image segmentation.

Email  /  Scholar  /  GitHub

Research

I'm interested in protein design, particularly in optimizing protein fitness using generative modeling and protein language models. Additionally, I have a strong interest in inverse problems in medical imaging. * indicates equal contribution.

We introduce RestoraFlow, a training-free method for masked-based inverse problems that guides flow matching sampling by a degradation mask and incorporates a trajectory correction mechanism to enforce consistency with degraded inputs.

We propose an improved MeanFlow training strategy that rapidly stabilizes instantaneous velocity before progressively emphasizing long-interval averages, enabling faster convergence and higher-quality one-step generation.

An energy matching framework is introduced, combining optimal transport paths far from the data manifold with an entropic energy term to explicitly capture data likelihood, enabling flexible priors and high-fidelity generation without auxiliary networks.

We introduce a multi-channel generative approach based on Flow Matching to synthesize medical images paired with heatmaps, enabling robust data augmentation and enhanced generalization for anatomical landmark localization - especially in settings with limited training data or occlusions.

CRISPR-PAMdb, a large-scale database of Cas9 proteins and PAM profiles, is introduced alongside CICERO, a protein language model–based predictor that enables accurate PAM preference inference and broad exploration of PAM diversity for genome editing.

A variational framework for generative protein fitness optimization using a flow matching prior and a classifier-guidance model in a continuous latent space is introduced.

A flow matching framework for generative medical image segmentation using signed distance functions that enables canonical smoothing of SDF mask distributions through noise injection.

We introduce a bilevel learning framework for variational image reconstruction that uses a primal-dual approach with a-posteriori error bounds and adaptive step sizes.

A DDPM-based framework for generating medical images with landmark heatmaps, using a Markov Random Field for matching and a Statistical Shape Model for plausibility checks.

A score-based generative model for binary medical image segmentation using signed distance functions, where diffusion corrupts SDF masks instead of binary ones for more natural distortions.

We introduce a framework for incorporating general discretizations of second-order TGV with variational consistency, and learn interpolation filters via a piggyback algorithm.

We develop a joint alignment and reconstruction algorithm for electron tomography without fiducial markers, applied to studying immune–beta cell interactions in NOD mice for type 1 diabetes research.

We analyze a piggyback-style method for differentiating saddle-point problems, and apply it to learning optimized shearlet transforms for imaging.

We use deep learning to study insulin granules in NOD mouse beta cells, where a multi-task regression aids in distinguishing healthy from diabetic samples.

We show that deep neural networks can estimate grain density in austenitic steel, with classification and regression models learning distinct feature representations.

Teaching