Cellpose-SAM hits 0.163 error rate, beating expert disagreement in cellular segmentation
RichardsonDx · x · 2026-11-10
HHMI Janelia researchers (Marius Pachitariu, Michael Rariden, Carsen Stringer) released Cellpose-SAM (bioRxiv, May 2025).
Key findings:
- Two experts tracing the same cells disagree on 25% of them; Cellpose-SAM cuts the error rate to 0.163 — below human disagreement and near the human-consensus bound
- Method: adapts the pretrained transformer backbone of Meta's SAM foundation model to the Cellpose framework
- Robustness training against channel shuffling, cell size variation, shot noise, downsampling, and isotropic/anisotropic blur boosts generalization to unseen stains and microscopes
- Drops into the Cellpose ecosystem: finetuning, human-in-the-loop training, image restoration, and 3D segmentation
Related event: Cellpose-SAM Beats Inter-Expert Disagreement in Cell Segmentation(2 posts)→
More from Research
- ProtoMotions Deployment Stack: One Exported Graph, 1.3 ms C++ Loop for Real Robots — carlosdponx · 2026-11-11
- MKA deep dive: top models cluster at 95-97% on basics but split 28-57% on hard tasks — ArtificialAnlys · 2026-11-11
- MKA Benchmark by Output Type: No Model Tops 52% on Word or Text Files — ArtificialAnlys · 2026-11-11
- DeepMind-backed MKA benchmark crowns Claude Fable 5.2 at 57%, with 30x cost spread — ArtificialAnlys · 2026-11-11
- Walmart new grad turns solo RTX 3090 Ti side project into SIGGRAPH 2026 paper — teortaxesTex · 2026-11-11
- Emergent misalignment is predictable OOD; later-layer fine-tuning reduces transfer — AdtRaghunathan · 2026-11-11