[Paper] Region of Interest Segmentation and Morphological Analysis for Membranes in Cryo-Electron Tomography

Source: arXiv - 2602.21195v1

Overview

The paper introduces TomoROIS‑SurfORA, a two‑stage pipeline that lets researchers isolate and quantitatively analyze specific membrane regions directly from cryo‑electron tomography (cryo‑ET) volumes. By bypassing the traditional “segment‑then‑post‑process” workflow, the authors make ROI extraction fast, shape‑agnostic, and usable even with very small training sets—opening the door for routine, high‑throughput membrane analysis in both academic labs and industry R&D.

Key Contributions

• TomoROIS: a deep‑learning module that segments user‑defined regions of interest (ROIs) on raw tomograms without needing a full‑volume membrane segmentation first.
• SurfORA: a geometry‑focused toolkit that converts the ROI masks into point clouds and surface meshes, then computes morphometrics such as inter‑membrane distance, curvature, and surface roughness.
• Shape‑agnostic training: the network can be trained from scratch on as few as a handful of manually annotated slices, making it practical for niche or novel cryo‑ET datasets.
• Support for open surfaces: special handling of the “missing wedge” artifact common in cryo‑ET, allowing accurate analysis of partially observed membranes.
• Open‑source reference implementation: code and example notebooks are released under a permissive license, encouraging integration into existing cryo‑ET processing pipelines.

Methodology

1. ROI Segmentation (TomoROIS)

   • A 3‑D U‑Net‑style convolutional neural network receives the raw tomogram and a binary mask indicating the desired ROI class (e.g., a vesicle, a contact site).
   • The model is trained end‑to‑end with a Dice loss, using data augmentation (rotations, intensity jitter) to compensate for the limited number of annotated volumes.
   • During inference, the network outputs a probability map that is thresholded to produce a voxel‑level ROI mask.

2. Morphological Surface Analysis (SurfORA)

   • The binary ROI mask is first converted to a point cloud via marching‑cubes surface extraction.
   • A lightweight mesh simplification step reduces computational load while preserving geometric fidelity.
   • Surface‑based metrics are then computed:
     • Inter‑membrane distance – pairwise nearest‑neighbor distances between two extracted surfaces.
     • Mean/Gaussian curvature – estimated from local mesh normals using discrete differential geometry formulas.
     • Roughness – quantified as the standard deviation of surface height relative to a fitted reference plane or sphere.
   • For open surfaces, SurfORA automatically caps the mesh edges or applies a “wedge‑aware” distance metric to avoid bias from missing data.

The whole pipeline can be scripted from the command line or called from Python, making it easy to embed in existing cryo‑ET workflows (e.g., after motion correction and tomogram reconstruction).

Results & Findings

• Accuracy: On a test set of 12 manually annotated vesicle ROIs, TomoROIS achieved a mean Dice coefficient of 0.89, comparable to state‑of‑the‑art full‑volume segmentation models that require far more training data.
• Robustness to limited data: Training with only 5 annotated volumes still yielded >0.80 Dice, demonstrating the model’s data efficiency.
• Morphometric fidelity: SurfORA’s curvature estimates matched ground‑truth analytical surfaces (synthetic spheres/torii) within 5 % error, and inter‑membrane distance measurements were within 2 nm of manual measurements (the typical resolution limit of cryo‑ET).
• Biological insight: Applying the pipeline to reconstituted vesicle systems revealed statistically significant differences in membrane roughness between wild‑type and mutant proteins that induce invagination, a result that previously required labor‑intensive manual tracing.

Practical Implications

• Accelerated discovery: Developers of membrane‑protein assays can now automate the quantification of contact sites, curvature changes, or vesicle remodeling, turning what used to be a weeks‑long manual effort into a few minutes of compute time.
• Integration with AI‑driven pipelines: TomoROIS can be chained after denoising or super‑resolution models (e.g., Cryo‑CARE) to provide end‑to‑end AI‑augmented analysis.
• Cross‑domain applicability: Because SurfORA works on generic point‑cloud meshes, the same surface‑analysis code can be reused for other 3‑D imaging modalities (e.g., micro‑CT of porous materials, light‑sheet microscopy of organoids).
• Low‑resource adoption: The small training footprint means labs without large GPU clusters can still train a useful model on a single workstation (≈8 GB VRAM).
• Open‑source ecosystem: The released Python package includes utilities for visualizing meshes in napari or PyVista, enabling rapid prototyping of custom metrics for product development or quality‑control pipelines.

Limitations & Future Work

• Missing wedge bias: Although SurfORA includes wedge‑aware corrections, severe missing‑wedge angles (> 60°) still degrade curvature estimates; the authors suggest integrating tilt‑series alignment improvements.
• ROI definition granularity: The current UI only supports binary ROI masks; multi‑class segmentation (e.g., distinguishing multiple vesicles simultaneously) will require extending the network architecture.
• Scalability to whole‑cell tomograms: The pipeline has been validated on relatively small, isolated membrane patches. Scaling to gigavoxel whole‑cell volumes will need hierarchical processing or distributed inference.
• Generalization to other organelles: While the authors demonstrate membrane applicability, further experiments are needed to confirm performance on highly irregular structures like the endoplasmic reticulum or cytoskeletal filaments.

Overall, TomoROIS‑SurfORA offers a practical, developer‑friendly toolkit that bridges deep learning segmentation and quantitative geometry, promising to streamline membrane analysis in cryo‑ET and beyond.

Authors

• Xingyi Cheng
• Julien Maufront
• Aurélie Di Cicco
• Daniël M. Pelt
• Manuela Dezi
• Daniel Lévy

Paper Information

• arXiv ID: 2602.21195v1
• Categories: cs.CV
• Published: February 24, 2026
• PDF: Download PDF