Related Experiment Video
Updated: May 5, 2026
06:17
Analysis of Multidimensional Microscopy Data Using Cell-ACDC
Published on: November 7, 2025
776
Foundation cell segmentation models performance on live microscopy and spatial-omics data
Yang Miao1, Nick Surguladze2, Josh Lerner1
1Department of Biomedical Engineering, Duke University, Durham, NC 27708, USA.
Biorxiv : the Preprint Server for Biology
|May 4, 2026
Summary
Choosing the right deep learning cell segmentation model is crucial for biological image analysis. Different models excel on specific imaging types, impacting downstream analyses like cell identification.
Area of Science:
- Computational Biology
- Bioimaging Analysis
- Deep Learning Applications
Background:
- Accurate cell segmentation is vital for quantitative biological imaging.
- Deep learning models show promise across various imaging modalities.
- Limited systematic comparisons exist for downstream biological analysis.
Purpose of the Study:
- Evaluate recent deep learning cell segmentation models.
- Compare model performance on phase contrast, fluorescence, and multiplexed tissue imaging.
- Assess impact on downstream analyses like clustering and cell identification.
Main Methods:
- Tested Cellpose cyto3, Cellpose-SAM, μSAM, CellSAM on phase contrast and fluorescence images.
- Benchmarked Mesmer and InstanSeg on CO-Detection by IndEXing (CODEX) tissue images.
- Analyzed downstream effects on clustering and cell type identification.
Main Results:
- Cellpose-SAM performed well on phase contrast images.
- SAM-based models showed strong results on fluorescence cell culture data.
- No single model dominated CODEX datasets; performance varied by model strengths and limitations.
Conclusions:
- Model selection for cell segmentation should align with dataset characteristics and analytical objectives.
- A universal approach to cell segmentation is not optimal.
- Understanding model-specific strengths is key for reliable biological insights.
