%0 Generic
%A Yu, Qin
%C Heidelberg
%D 2025
%F heidok:37375
%K Bioimage, Bioimage Informatics, Bioimage Analysis, Image Analysis, Bioimage-Informatik, Bioimage-Analyse, Bildanalyse
%R 10.11588/heidok.00037375
%T Robust and Accessible Segmentation of Cells and Nuclei in 3D Microscopy
%U https://archiv.ub.uni-heidelberg.de/volltextserver/37375/
%X Volumetric microscopy reveals unprecedented details of cells and subcellular structures across diverse biological tissues. However, harnessing these complex datasets requires robust segmentation methods that maximise the use of multi-channel information while minimising the need for expert annotation. In this thesis, I introduce scalable workflows that reduce reliance on extensive expert labelling while enhancing segmentation accuracy under realistic imaging conditions.    First, I present GoNuclear, a versatile toolkit for three-dimensional (3D) nuclear segmentation of plant tissues stained with the affordable, broadly applicable DNA-binding dye TO-PRO-3. Unlike genetically encoded markers, which require laborious transformations, TO-PRO-3 can be directly applied to fixed and cleared tissues, greatly simplifying nuclear segmentation in non-model species. By leveraging human-in-the-loop annotations and carefully curated datasets, GoNuclear provides accurate segmentation from weak, noisy signals and generalises effectively across diverse tissues and staining modalities. This enables downstream analyses such as nuclear size control, nuclear-to-cell volume ratios, and spatial gene expression.    Next, I describe substantial enhancements to PlantSeg, a deep-learning-based toolkit for 3D tissue segmentation. Version 2.0 features an interactive Napari-based interface, integration with the BioImage Model Zoo, sparse instance segmentation support, automatic optimisation of patch and halo sizes, and powerful proofreading tools. These improvements enable the accurate segmentation of both cells and nuclei in complex microscopy volumes, making advanced computational methods more accessible to the scientific community.    I then introduce SPOCO, an embedding-based instance segmentation method requiring minimal annotations. Through targeted transfer learning, SPOCO adapts models trained on limited annotated datasets to new imaging domains, significantly reducing annotation requirements while preserving segmentation quality. I also demonstrate computational efficiency gains through dimensionality reduction and novel clustering strategies, establishing SPOCO as a practical solution for large-scale analyses.    Finally, I show how multi-channel bioimage analysis can substantially improve segmentation and tracking accuracy. By integrating complementary chemical stains, multiple imaging modalities, temporal sequences, and various biological structures into a single dataset, these strategies enhance both the accuracy and interpretability of segmentation outcomes, particularly under challenging imaging conditions.    Together, these contributions advance 3D bioimage segmentation by minimising annotation burdens and improving segmentation performance under realistic conditions. This work provides practical, user-friendly solutions for precise, large-scale biological analyses, opening new avenues for the investigation of developmental biology, cellular architecture, and morphogenesis in a wide range of plant and animal systems.