PARSEG: a novel approach to tackle overlapping segmentation in bioimages

Session

Poster Session

Authors

Vanessa Dao (1, 3), Georgios Sioutas (2), Kurt Anderson (1), David Barry (1), Todd Fallesen (1)

Affiliations

1. Crick Advanced Light Microscopy, The Francis Crick Institute
2. Sex Chromosome Biology Laboratory, The Francis Crick Institute
3. University of Bath

Keywords

segmentation, image analysis, colocalization, napari

Abstract text

Segmenting multi-scale biological structures is a common challenge in bioimage analysis. Tools such as StarDist and Cellpose identify objects within a predicted object size range, however, may generate mislabelled or split objects outside these boundaries. A potential solution is to generate and merge multiple segmentation maps tailored to different predicted object size. However, this approach results in overlapping segmentations.

We therefore introduce PARSEG (PArallelised Refinement of SEGmentations), a unique plugin for the Napari ecosystem. PARSEG filters overlapping segmentation masks based on colocalization statistics, such as percent overlap. By leveraging Dask, PARSEG handles data in a computationally efficient manner by processing individual 2D slices in parallel. This optimisation eliminates the requirement for manual curation of segmentations. Additionally, PARSEG allows for the preservation of data on overlapping segmentations for subsequent analysis.

We demonstrate the utility of PARSEG by applying it to mouse oocyte data exhibiting oocyte growth during follicle maturation. Using PARSEG to refine Cellpose segmentations of developing oocytes provides a significant improvement in segmentation accuracy. PARSEG can benefit research not only in developmental and reproductive biology, but in any image set with indefinable objects of different sizes, providing a tool for more accurate and efficient analysis.