ETSAM

Effectively segmenting cell membranes in cryo-electron tomograms with a two-stage SAM2-based deep learning model.

Installation Tutorial Read Paper View on GitHub

quick start

# 1. Install ETSAM (Linux, NVIDIA GPU; for Windows, see Installation)
curl -fsSL https://jianlin-cheng.github.io/ETSAM/install.sh | sh

# 2. Run ETSAM on a tomogram, in a new terminal
etsam input_tomogram.mrc --output-dir results/

Two-stage refinement

Stage 1 generates an initial membrane mask from grid-point prompts. Stage 2 fuses that output with the normalized tomogram to refine membrane predictions.

SAM2-based architecture

Fine-tuned on top of SAM2’s video-prediction model, adapted for cryo-ET volume slices instead of standard RGB imagery.

Strong precision-recall trade-off

Benchmarked against TARDIS and MemBrain-Seg on independent simulated and experimental tomograms with state-of-the-art performance.

Practical inference speed

Processes a 500-slice tomogram in approximately 90 seconds on an NVIDIA RTX 4070 GPU.

Diverse training data

Trained on 83 experimental CryoET Data Portal tomograms and 28 PolNet simulated tomograms across diverse cell types.

Configurable outputs

Tune logit thresholds, save raw likelihood scores, and apply optional 3D connected-component post-processing.

pipeline

The ETSAM workflow

Stage 1 prompts SAM2 with a grid on the first tomogram slice and propagates the prediction through the volume. Stage 2 combines the normalized tomogram with Stage 1 output and runs a second pass to improve the final binary membrane mask.

ETSAM two-stage pipeline architecture — Two-stage ETSAM pipeline and block architecture.

About the Research

Cryogenic Electron Tomography (cryo-ET) visualizes cell structures and macromolecules in their native environment. Accurate membrane segmentation is critical for understanding cellular organization, but low signal-to-noise ratio, missing wedge artifacts, and other noise sources make reliable segmentation challenging.

ETSAM addresses these challenges by fine-tuning SAM2’s video-prediction architecture and using a cascaded two-stage design where each stage’s output informs the next. The model was developed and tested by Joel Selvaraj and Jianlin Cheng at the University of Missouri.

Cite

@article{Selvaraj2026,
    author  = {Selvaraj, Joel and Cheng, Jianlin},
    title   = "{ETSAM}: effectively segmenting cell membranes in cryo-electron tomograms",
    journal = {Nature Communications},
    year    = {2026},
    doi     = {10.1038/s41467-026-77962-7},
    url     = {https://doi.org/10.1038/s41467-026-77962-7},
    issn    = {2041-1723}
}