Accurate Nuclei Detection in 3D Cell Systems Through Centroid Prediction

Session

Poster Session

Authors

Tim Van De Looverbosch (2), Sarah De Beuckeleer (2), Jan Sijbers (1), Winnok De Vos (2, 3, 4, 5)

Affiliations

1. Imec-Vision Lab, University of Antwerp
2. Laboratory of Cell Biology and Histology, University of Antwerp
3. IMARK, University of Antwerp
4. Antwerp Centre of Andanced Microscopy, University of Antwerp
5. µNeuro Research Centre of Excellence, University of Antwerp

Keywords

3D Cell detection; Light sheet microscopy; Organoid phenotyping

Abstract text

3D cell systems, such as organoids, are next generation model systems for fundamental and preclinical research. To optimize their production and to take full advantage of these information rich samples, understanding their composition, organization, and changes their off is of tremendous value. Light sheet microscopy can be used to acquire in-toto images at cellular resolution, however subsequent analysis is not straightforward. Cell detection, a common first step in analysis pipelines, is often done using instance segmentation methods. While powerful methods are available for 2D images, instance segmentation in large 3D cell systems remains challenging due to the high cell density, the heterogeneous resolution and contrast across the image volume, and the difficulty in generating reliable and sufficient ground truth data for model training. Reasoning that many applications rely on nuclear segmentation but actually do not necessarily require an accurate delineation of their shapes, we implemented a 3D U-Net based method that rapidly provides the position of their centroids by predicting centroid probability maps. The ease of data annotation allows for fast application to new datasets. We show that our method outperforms available methods and that performance can be improved using weak supervised pretraining. We apply it on diverse images and show that it can be used to map subpopulations in spheroids and organoids. Finally, we demonstrate that our model can serve as seed generator for seeded watershed segmentation, or as a lightweight prompt generator for automated nuclei segmentation using foundation models.