Cryo-ET membrane segmentation goes two-stage: Dice rises from 0.6381 to 0.8116, at 79.97 seconds per tomogram

Membrane segmentation can now run fully automatically, without manual prompts or per-tomogram parameter tuning. Traditional methods required human intervention and parameter tuning, while deep learning methods struggled to reach high precision and recall together, with some missing critical membrane regions and others producing noisy predictions. On 17 independent experimental tomograms with ground-truth annotations, ETSAM reaches Dice 0.8116, IoU 0.7059, precision 0.8123 and recall 0.8430, above Membrain-Seg at 0.6381, 0.4999, 0.5846 and 0.7371 and TARDIS at 0.5980, 0.4519, 0.6916 and 0.5788.

The approach treats a three-dimensional tomogram as a video so that neighboring slices constrain one another. SAM2 is a foundational image and video segmentation model for natural images and videos and cannot be applied to tomograms directly, so its video segmentation pipeline has to be repurposed. Training used 83 experimental tomograms from the CryoET Data Portal plus 28 PolNet simulated tomograms, 111 in total; the test set is 17 experimental tomograms from 9 CDP, 4 EMDB and 1 EMPIAR datasets, spanning 12 cellular organisms and 1 virus.

The second stage mainly suppresses false positives rather than raising recall. Single-stage ETSAM reaches precision 0.7763, Dice 0.7892 and recall 0.8445; the second stage lifts precision to 0.8123 and Dice to 0.8116 while recall slips to 0.8430. Across the 17 test tomograms, the two-stage pipeline improves Dice on 12 and precision on 11, while recall improves on only nine.

The reported scope is cell membrane segmentation in cryo-electron tomograms, evaluated on 17 independent experimental tomograms with manually curated ground truth, spanning 12 cellular organisms and 1 virus. Training used 83 experimental tomograms and 28 PolNet simulated tomograms, and the trained model outputs a binary membrane mask, while users can also store logits and adjust the threshold to trade precision against recall. On five tomograms of cellular organisms unseen in training, the model produced qualitatively less noisy predictions, but those tomograms have no ground-truth annotations, so no quantitative metrics are reported.

The model's high sensitivity can misclassify membrane-like artifacts as membranes, including ice contamination, border noise and carbon film holes, and these false positives require manual inspection and cleaning. For precision, the difference from TARDIS does not reach the Bonferroni-corrected significance threshold; for recall, the difference from Membrain-Seg likewise does not. No single logit threshold performs best across all tomograms, so users must choose a threshold per tomogram.