Invited Speaker

Carsten Marr

Bio:

Carsten Marr is the founding director of the Institute of AI for Health at Helmholtz Munich, a European center for applied artificial intelligence. His goal is to develop AI-based methods to improve the diagnosis, treatment, and understanding of diseases.

After studying theoretical physics at the Technical University of Munich and completing his diploma thesis at the Max Planck Institute of Quantum Optics, Carsten switched from physics to theoretical biology. His PhD thesis at the Technical University of Darmstadt focused on the architecture of biological networks and was awarded the best of its year in the Department of Biology. After postdoctoral research stays in Munich, Bremen, and Edinburgh, he started his research group at the Helmholtz Munich in 2013 and became deputy head of the Institute of Computational Biology. In interdisciplinary projects with experimentalists, biomedical experts, and clinicians, he pioneered the training of deep neural networks on life science data for the prediction of stem cell decisions from microscopic images and the identification of leukemia from blood and bone marrow smears. He has received several awards for his research and analysis of single-cell data, as well as ERC Consolidator and Proof of Concept grants on training AI models for the automated analysis of blood diseases.

Presentation abstract:

Diagnosing hematologic malignancies still relies heavily on the subjective visual assessment of cytological and histological images. Experts are increasingly challenged by large volumes of data, the rarity of diagnostic cell types, and the heterogeneous presentation of disease. Despite the availability of comprehensive patient data, advanced deep learning algorithms, and a solid understanding of hematopoiesis, there is currently no robust model capable of automatically analyzing and predicting disease dynamics from blood smears or bone marrow aspirates. In this talk, I will present recent advances in AI-based hematopathology that aim to address key challenges such as model robustness, generalization to real-world data, bias mitigation, and the integration of multimodal sources. I will highlight three promising directions: (1) efficient single-cell detection using neural cellular automata, (2) interpretable feature learning via sparse autoencoders, and (3) the integration of biomedical prior knowledge into model training through customized loss functions. These developments illustrate how tailored AI solutions can bridge the gap between machine learning algorithms and clinical decision-making, paving the way toward more accurate, scalable, and explainable diagnostics in hematology.

Contacts

Albert Comelli <acomelli@fondazionerimed.com>

Cecilia Di Ruberto <dirubert@unica.it>

Andrea Loddo <andrea.loddo@unica.it>

Lorenzo Putzu <lorenzo.putzu@unica.it>

Alessandro Stefano <alessandro.stefano@cnr.it>

Luca Zedda <luca.zedda@unica.it>

Page updated

Google Sites

Report abuse