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

Halo achieves over 20% higher accuracy than existing methods in whole-cell segmentation across diverse tissues using only nuclear images and RNA data.

• Why it matters: Accurate cell boundary reconstruction is essential for reliable spatial transcriptomics analysis, but current methods often require extensive dataset-specific training, limiting scalability and reproducibility.
• What they did: The model integrates nuclear morphology and RNA transcript distributions by converting transcript coordinates into molecular density maps, processed with DAPI images using a Cellpose-SAM architecture, and is pretrained on 12 tissue types.
• The result: Halo provides a ready-to-use, generalizable tool that improves cell boundary and RNA-to-cell assignment accuracy, enabling more precise cell-type identification and morphological analysis in spatial transcriptomics studies.