% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@INPROCEEDINGS{Kropp:1043527,
author = {Kropp, Jan-Oliver and Schiffer, Christian and Amunts,
Katrin and Dickscheid, Timo},
title = {{S}tep by {S}tep: {T}owards a gapless 1 micron {B}ig{B}rain
with {D}iffusion {M}odels},
reportid = {FZJ-2025-02903},
year = {2025},
abstract = {Advances in microscopic imaging and high-performance
computing allow analyzing the complex cellular structure of
the human brain in great detail. This progress has greatly
aided in brain mapping and cell segmentation, and the
development of automated analysis methods. However,
histological image data can contain data gaps due to
inevitable processing artifacts, which, despite careful
precautions, may arise during histological lab work, such as
missing sections, tissue tears, or inconsistent staining.To
address this issue, we presented a convolutional neural
network model that reconstructs corrupted data from
surrounding tissue, while preserving precise cellular
distributions. Our approach uses a denoising diffusion
probabilistic model trained on light-microscopy scans of
cell-body stained histological sections. We extended this
model with the RePaint method to impute corrupted image
data. We evaluate its performance with established deep
learning models trained on the same type of histological
data.A key challenge of our initial model was its difficulty
in accurately reconstructing tissue boundaries and larger
anatomical structures such as blood vessels. We address
these challenges by an enhanced diffusion-based model that
incorporates contextual information from adjacent sections
of the brain. This model integrates three tissue patches
from neighboring sections using a siamese network
architecture with cross-attention mechanisms. Leveraging
spatially aligned information across consecutive sections,
our approach achieves a more anatomically coherent
reconstruction.We demonstrate that our model significantly
improves realism and anatomical plausibility of
reconstructed cellular distributions, as measured by both
cell density prediction and brain area classification tasks.
The error in predicted cell density was reduced to below
$5\%$ across large inpainting regions, marking a notable
improvement over previous approaches. The model reliably
preserves tissue borders and reconstructs larger structures
like blood vessels, which are crucial for accurate
cytoarchitectonic mapping.These findings underscore the
potential of generative deep learning models for
cytoarchitectonic research, opening new avenues for the
automated reconstruction of histological data. Beyond
inpainting small regions, our approach paves the way for the
reconstruction of entirely missing brain slices, offering a
powerful tool for bridging data gaps in high-resolution
brain mapping efforts.},
month = {Jun},
date = {2025-06-03},
organization = {Helmholtz AI Conference 2025,
Karlsruhe (Germany), 3 Jun 2025 - 5 Jun
2025},
subtyp = {After Call},
cin = {INM-1},
cid = {I:(DE-Juel1)INM-1-20090406},
pnm = {5254 - Neuroscientific Data Analytics and AI (POF4-525) /
Helmholtz AI - Helmholtz Artificial Intelligence
Coordination Unit – Local Unit FZJ (E.40401.62) / HIBALL -
Helmholtz International BigBrain Analytics and Learning
Laboratory (HIBALL) (InterLabs-0015) / EBRAINS 2.0 - EBRAINS
2.0: A Research Infrastructure to Advance Neuroscience and
Brain Health (101147319) / JL SMHB - Joint Lab
Supercomputing and Modeling for the Human Brain (JL
SMHB-2021-2027)},
pid = {G:(DE-HGF)POF4-5254 / G:(DE-Juel-1)E.40401.62 /
G:(DE-HGF)InterLabs-0015 / G:(EU-Grant)101147319 /
G:(DE-Juel1)JL SMHB-2021-2027},
typ = {PUB:(DE-HGF)24},
url = {https://juser.fz-juelich.de/record/1043527},
}