CECAD: Katarzyna Bozek

Research Focus Our Goals Key Publications

Data Science of Bioimages

With the use of deep learning, we develop new data-driven approaches to study image and video data in biology. Our applications span cancer, nephrology, eye medicine, neurology and behaviour research.

Scatter plot of neural-network derived representations of immunohistochemistry images of HER2-stained esophagus samples. The image representations reflect histological scoring of these images.

Research Focus

We are interested in resolving large data challenges in biology and medicine with the major focus on image and video data. We strive to bring scientific advances in computer vision to the field of biomedical research. Unlike benchmark datasets in computer vision, biomedical images are massive in size and contain complex visual information that might either be dense or sparse and difficult to detect. We adapt ideas and solutions from basic computer vision to the challenges and specificities of bioimage data. Our machine learning solutions are aimed at finding representations of visual data that are well suited for resolving specific scientific questions, such as detecting and classifying disease entity, quantifying complex molecular morphologies, detecting subtle changes in motion that precede neurodegenerative diseases, etc. Our research is strongly collaborative, we work with experimental and medical experts in CECAD and several departments of the University Hospital Cologne.

Our vision is to elevate analysis of biomedical image data to the scale and resolution of other high-throughput data types in biology, such as sequencing or mass spectrometry data.

Our Goals

Segmentation of foot processes of podocytes in kidney. Our segmentation method allows to automate the analysis of foot process morphology.

Apart from automating image analysis tasks such as segmentation and classification, we are interested in using deep learning to explore complex visual features encoded in images which might defy human vision. As an example, we build models predicting cancer patient therapy outcomes based on microscopy images of their primary biopsies. Our results in cutaneous squamous cell carcinoma and esophageal adenocarcinoma demonstrate that this is indeed possible.
We also develop self-supervised models for representing complex morphologies of structures such as biological membranes, nerves, or blood vessels in a comprehensive, parameter-free manner. Such structures are unconstrained, forming sometimes dense and branching architectures which are difficult to describe using a set of pre-defined parameters. Our self-supervised methods allow to capture this complexity in a comprehensive manner and to quantitatively link morphology to the disease type and advancement.
Several imaging techniques allow to image in biological specimens multiple fluorescent channels, where each channel contains signal from a specific protein. Resulting images are not only spatially large but also multi - up to 50- or 100-channel. We are devising techniques to segment and quantify protein signal and analyze such multi-dimensional spatial data.

Key Publications

Deserno M, Bozek K. WormSwin: Instance segmentation of C. elegans using vision transformer. Sci Rep (2023) 13, 11021. doi.org/10.1038/s41598-023-38213-7
Butt L, Unnersjö-Jess D, Höhne M, Sergei G, Witasp A, Wernerson A, Patrakka J, Hoyer PF, Blom H, Schermer B, Bozek K*, Benzing T*. Deep learning-based segmentation and quantification of podocyte foot process morphology Kidney International (2023) doi.org/10.1016/j.kint.2023.03.013
vom Stein A, Rebollido-Rios R, Lukas A, Koch M, von Lom A, Reinartz S, Bachurski D, Rose F, Bozek K, Abdallah A, Kohlhas V, Saggau J, Zölzer R, Zhao Y, Bruns C, Bröckelmann P, Lohneis P, Buettner R, Häupl B, Oellerich T, Nguyen P-H, Hallek M. LYN kinase programs stromal fibroblasts to facilitate leukemic survival via regulation of c-JUN and THBS1 Nat Commun (2023) 14, 1330 doi.org/10.1038/s41467-023-36824-2
Pisula JI, Datta RR, Börner-Valdez L, Avemarg J-R, Jung J-O, Plum P, Löser H, Lohneis P, Meuschke M, Pinto dos Santos D, Gebauer F, Quaas A, Bruns CJ, Walch A, Lawonn K, Popp FC, Bozek K. Predicting the HER2 status in esophageal cancer from tissue microarrays using convolutional neural networks, Br J Cancer (2023) doi.org/10.1038/s41416-023-02143-y
Hanuscheck N, Thalman C, Domingues M, Schmaul S, Muthuraman M, Hetsch F, Ecker M, Endle H, Oshaghi M, Martino G, Kuhlmann T, Bozek K, van Beers T, Bittner S, von Engelhardt J, Vogt J, Vogelaar CF, Zipp F. Interleukin-4 receptor signaling modulates neuronal network activity, J Exp Med (2022) 219 (6): e20211887. doi.org/10.1084/jem.20211887

Katarzyna Bozek

CMMC, Institute for Biomedical Informatics

Dr. Katarzyna Bozek

Research Areas