%0 Conference Paper
%A Schiffer, Christian
%A Boztoprak, Zeynep
%A Kropp, Jan-Oliver
%A Thönnißen, Julia
%A Behr, Katja
%A Spitzer, Hannah
%A Amunts, Katrin
%A Dickscheid, Timo
%T CytoNet: A Foundation Model for Cytoarchitecture in the Human Cerebral Cortex
%M FZJ-2025-02894
%D 2025
%X Microscopic analysis of cytoarchitecture in the human cerebral cortex plays a central role for developing high-resolution microstructural human brain atlases. Cytoarchitecture is defined by the spatial organization of cells, including their shape, density, size, type, and their columnar and laminar arrangement, which varies between brain regions. Cytoarchitecture provides an important microstructural reference for brain connectivity and function and is therefore of great interest in brain research. Microscopic scans of histological human brain sections allow detailed analysis of cytoarchitecture. High-throughput microscopy scanners can digitize sections of an entire brain at 1 micrometer isotropic resolution in about a year, resulting in petabyte-scale image datasets that capture the brain’s complexity and variability across multiple subjects. These large imaging datasets offer great opportunities for brain research, but also pose novel methodological and technical challenges that require developing new analytical methods.Addressing these needs, we present CytoNet, a foundation model for cytoarchitecture in the human cerebral cortex. CytoNet is trained with a novel self-supervised learning task that promotes the extraction of cytoarchitectonic features from microscopic image patches extracted along the cortical ribbon. We demonstrate that CytoNet learns to extract powerful and anatomically plausible representations of cytoarchitecture that capture intra-subject variance and inter-subject variability. CytoNet is able to compute features that are 1) completely data-driven, 2) globally comparable across brain regions and subjects, 3) encode a wide range of cytoarchitectonic properties that facilitate relevant downstream analysis and whole-brain correlative analysis, and 4) can be computed at arbitrarily dense sampling intervals at any location within the cerebral cortex of any subject. CytoNet achieves state-of-the-art performance for brain area classification, cortical layer segmentation, morphology estimation, and data-driven discovery of cytoarchitectonic structures. Using embedding analysis, we show that CytoNet learns to map microscopic images into a semantically highly structured and anatomically plausible latent space that facilitates the aforementioned downstream tasks. Our work has broad implications for changing the way we analyze microscopic brain organization and will improve our ability to make discoveries from large datasets.
%B Helmholtz AI Conference 2025
%C 3 Jun 2025 - 5 Jun 2025, Karlsruhe (Germany)
Y2 3 Jun 2025 - 5 Jun 2025
M2 Karlsruhe, Germany
%F PUB:(DE-HGF)6
%9 Conference Presentation
%U https://juser.fz-juelich.de/record/1043518