SubCellSeg: A generalist model for subcellular structure segmentation in 3D fluorescence microscopy

SubCellSeg, a single 3D fluorescence segmentation model for hiPS cells, achieves a mean Dice score of 0.909 across 13 subcellular structures, outperforming 13 specialist models.

• Why it matters: Structure-by-structure training limits scalability and robustness when morphologies, microscopes, or perturbations change. A generalist model could reduce the need for per-structure retraining.
• What they did: The authors trained SubCellSeg on expert-curated fluorescence images and computational reference masks for 13 morphologically distinct structures in hiPS cell lines from the Allen Cell Collection. They evaluated mean Dice versus specialist models and tested zero-shot transfer across microscopes, imaging configurations, drug perturbations, and hiPS-derived cardiomyocytes.
• The result: SubCellSeg’s specialist mean Dice was 0.773, and it also performs zero-shot on morphologies absent from training. When zero-shot is insufficient, fine-tuning with 10% labeled images (16 images) outperforms training from scratch, and the model is available via a napari plugin with pretrained checkpoints.

AI summary of the abstract · may contain errors · not medical advice · read the paper.