SALEM²: Segment Anything in Light and Electron Microscopy via Membrane Guidance

This repository has been archived. Please instead refer SAEM2 in SAvEM3.

Comparison with General-purpose Methods

Quantitative comparison: SAM, HQ-SAM, Mobile-SAM (from HQ-SAM) and Micro-SAM, as well as trained SAM and HQ-SAM in both LM and EM datasets.

Comparison with Specialized Supervised Methods

Qualitative comparison: CellPose in LM, ilastik (pretrained models from https://bioimage.io/#/?partner=ilastik) and Superhuman (onnx models from https://github.com/seung-lab/DeepEM/releases) in EM.

Elongated cells in LM:

Raw	GT	CellPose (Nat. Methods 2021)	OmniPose (Nat. Methods 2022)	SALM2 (Ours)

Weak boundaries in LM:

Raw	GT	CellPose (Nat. Methods 2021)	SALM2 (Ours)

CREMI-B (blur and misalignment) in EM:

Raw (sections 12-18)	GT (sections 12-18)	Superhuman (section 15) (IEEE Trans. Med. Imaging 2021)	SAEM2 (section 15) (Ours)

CREMI-C (blur and missing sections) in EM:

Raw (sections 100-106)	GT (sections 100-106)	ilastik (section 103) (Nat. Methods 2019)	SAEM2 (section 103) (Ours)

More results can be found in Google Drive: BBC039, NeurIPS22-CellSeg and CREMI. We used CellPose and OmniPose with the configurations of nuclei (flow_threshold=0.3) for Bare Nuclei, tissuenet (flow_threshold=0.6) for Weak Boundary, nuclei (flow_threshold=0.6) and bact_phase_omni for Elongated. We used the open-sourced trained models of ilastik and Superhuman with alternative preprocessing steps.

Contributors

• Hao Zhai
• Jinyue Guo
• Yanchao Zhang

Thanks to the other authors and MiRA Team for their support and resources.

Acknowledgments

• SAM
• HQ-SAM
• Micro-SAM
• Mobile-SAM

Thanks for their public code and released models.