Halo: a pretrained model for whole-cell segmentation from nuclei images in spatial transcriptomics.

Area of Science:

• Genomics
• Bioinformatics
• Cell Biology

Background:

• Spatial transcriptomics measures gene expression within tissue context.
• Accurate whole-cell segmentation is crucial for single-cell transcriptomics but challenging with nuclear-only images.

Purpose of the Study:

• Introduce Halo, a novel pretrained segmentation model for whole-cell reconstruction in spatial transcriptomics.
• Improve cell boundary inference and RNA-to-cell assignment using integrated nuclear and transcriptomic data.

Main Methods:

• Halo integrates nuclear morphology (DAPI images) with RNA transcript spatial distribution.
• It converts transcript coordinates into density maps, processed with a Cellpose-SAM architecture.
• The model is pretrained on diverse multimodal Xenium data, enabling direct application without retraining.

Main Results:

• Halo outperforms traditional nuclear expansion methods across various tissues.
• Achieves higher accuracy in matching ground-truth cell boundaries and assigning RNA to cells.
• Demonstrates improved cell type identification and morphological feature estimation.

Conclusions:

• Halo provides a generalizable, pretrained solution for scalable and reproducible whole-cell segmentation in spatial transcriptomics.
• Enables more accurate downstream analyses like cell type identification and morphological profiling.