TY  - CONF
AU  - Oberstraß, Alexander
AU  - Vaca Cerda, Esteban Alejandro
AU  - Upschulte, Eric
AU  - Niu, Meiqi
AU  - Palomero-Gallagher, Nicola
AU  - Grässel, David
AU  - Schiffer, Christian
AU  - Axer, Markus
AU  - Amunts, Katrin
AU  - Dickscheid, Timo
TI  - Image-to-Image Translation for Virtual Cresyl Violet Staining From 3D Polarized Light Imaging
M1  - FZJ-2025-02901
PY  - 2025
AB  - Characterizing the structure of cortical networks in the brain requires complementary imaging techniques and the integration of different aspects such as fiber and cell body distributions. Ideally, different methods are applied to the same tissue for direct comparison. 3D polarized light imaging (3D-PLI) visualizes nerve fibers in brain tissue at high resolution based on optical properties alone. This enables subsequent staining of the same tissue for cell bodies after 3D-PLI measurement. However, this process is time-consuming, technically challenging, and requires nonlinear cross-modal registration to obtain pixel correspondence.Here we investigate image-to-image translation methods to predict the results of cell body staining directly from 3D-PLI, using generative adversarial networks (GANs) and neural style transfer (NST). We use 11 coronal sections of a vervet monkey brain for training, each imaged with 3D-PLI and subsequently stained with Cresyl violet for cell bodies. Since pixel-accurate registration of entire sections may be difficult and error-prone, we introduce an online registration head to linearly align model predictions for local image patches to the post-staining during network training. This exploits the fact that local deformations can be approximated by a linear model when a coarse pre-registration is available. The online alignment improves with predictions during training and ultimately converges to an accurate registration. We use a Fourier-based registration approach that is computationally efficient and GPU-parallelizable.We quantify model performance by comparing the predicted virtual staining to post-staining after 3D-PLI measurement. Our best model localizes the majority of larger cell instances (>100 µm² in-plane) segmented by a contour proposal network (CPN) with an F1 score of 63.1. The proposed online registration head significantly improves the performance of all investigated models, increasing F1 scores from 40.6 to 63.1 for NST and from 22.2 to 50.3 for a GAN.The applied virtual staining enables automatic localization of larger cell instances in unstained 3D-PLI images. Since the model predictions are pixel-aligned with 3D-PLI, they enable joint analysis of fiber tracts and cell bodies and may also serve as targets for registration of real post-staining. Future work will extend the training data to include more sections, brains and species, with potential applications to other imaging modalities.
T2  - Helmholtz AI Conference 2025
CY  - 3 Jun 2025 - 5 Jun 2025, Karlsruhe (Germany)
Y2  - 3 Jun 2025 - 5 Jun 2025
M2  - Karlsruhe, Germany
LB  - PUB:(DE-HGF)24
UR  - https://juser.fz-juelich.de/record/1043525
ER  -