Frederic Mrozinski

M.Sc.

Research Interests

Hi, I'm Frederic and my current research centers around medical technology. I am working on bringing SOTA machine learning models to LMU's dermatology clinic for better cancer detection. To be honest, I mostly work with vision at the moment, but I do also integrate some NLP multimodal approaches into my research. I do love to work with motivated students and supervise their theses. So if you're interested, you may familiarize yourself with my research interests and see open topics below. Research interests:

• Classification and Segmentation of Whole Slide Images, i.e. microscopic tissue scans
• Image generation and style transfer
• Few-shot learning
• Text summarization
• Vision-Language-Models
• ...

Supervision of Theses

Are you interested in writing a Master's Thesis in this field? Below you will find a list of currently open topics. If it's empty, you can still text me, sometimes I forget to update or maybe you even have your own idea? At the moment, I don't supervise Bachelor students. Guided Research and IDP are possible if the topic allows. When you text me, I would like you to send me your transcript of records and a little bit about yourself (nothing ChatGPT-generated and nothing too formal, I just want to get to know you and your background).

Currently open topics

• (Master's Thesis) Building a stlye-transfer image-generating pipeline based on Diffusion compatible with images of gigapixel resolutions. Publication intended. For that you'll ideally bring a solid understanding of image processing, diffusion networks and hands-on PyTorch experience. Or if neither of these, then motivation and an extra month to learn it all :-)
• (Master's Thesis or Guided Research) Benchmarking advanced Whole Slide Image patch-subsampling strategies: In Multiple-Instance-Learning, there is a trend of only considering a subset of instances to increase computational efficiency. While the performance of uniform subsampling is highly dataset-dependent, the goal of your work would be researching and benchmarking more elaborate subsampling strategies. This thesis would be in computer vision, mostly with some cool applications of statistics and attention models. Publication intended.