Skip to content

Membrane Refinement

Membrane refinement is a post-processing step that cleans up organelle and membrane segmentation pairs, enforcing topological consistency between them. It is optional — only relevant if you have segmented both an organelle class and its bounding membrane.

Membrane Refinement Example

Left: raw binary segmentations with spurious membrane fragments. Right: refined masks with clean boundaries and unique instance labels matched to each organelle.

Why Refine Membranes?

Raw segmentation outputs frequently contain artifacts that affect downstream morphological analysis:

Common artifacts without refinement

  • Spurious fragments: Small membrane pieces not attached to any organelle
  • Boundary overlap: Organelle masks extending slightly beyond their membrane
  • Internal membranes: Fragments inside an organelle that should be removed
  • Edge artifacts: Masks that clip at the volume boundary

Refinement resolves all of these by using the organelle mask to constrain the membrane and vice versa, producing paired masks that are geometrically consistent.

Generating Initial Membrane Segmentations

SABER does not train membrane detectors — use MemBrain-seg to generate high-quality initial membrane segmentations from cryo-ET data. MemBrain-seg is purpose-built for this task and integrates with Copick natively.

MemBrain-seg documentation

Running Refinement

CLI

saber analysis refine-membranes \
    --config config.json \
    --org-info "organelles,saber,1" \
    --mem-info "membranes,membrane-seg,1" \
    --voxel-size 10 \
    --save-session-id "1"

Specifying input segmentations — the --org-info and --mem-info flags accept either a simple name or a full name,userID,sessionID triplet:

--org-info "organelles"
--mem-info "membranes"

SABER picks the first matching user ID and session ID in the project. Use this for simple single-user projects.

--org-info "mitochondria,saber,3"
--mem-info "membranes,membrain-seg,2"

Explicit control over which user's and which session's segmentation to use. Recommended for multi-user projects or when you have multiple refinement runs.

What gets saved:

Input Output
organelles (user saber, session 1) organelles (user saber-refined, session 1)
membranes (user membrane-seg, session 1) membranes (user membrane-seg-refined, session 1)
saber analysis refine-membranes Parameters
Parameter Description Example
--config Copick config file config.json
--org-info Organelle segmentation query "organelles,saber,1"
--mem-info Membrane segmentation query "membranes,membrane-seg,1"
--voxel-size Voxel size in Å 10
--save-session-id Session ID for refined outputs "1"

Python API

For custom refinement workflows or parameter sweeps:

from saber.analysis.refine_membranes import OrganelleMembraneFilter, FilteringConfig

config = FilteringConfig(
    ball_size=5,                  # (1)
    min_membrane_area=10000,      # (2)
    edge_trim_z=5,                # (3)
    edge_trim_xy=3,
    batch_size=8,                 # (4)
    keep_surface_membranes=False, # (5)
)

filter_obj = OrganelleMembraneFilter(config, gpu_id=0)
results = filter_obj.run(organelle_seg, membrane_seg)

refined_organelles = results['organelles']
refined_membranes  = results['membranes']
  1. Morphological kernel radius. Larger values smooth boundaries more aggressively.
  2. Minimum voxel count for a membrane component to be kept.
  3. Voxels to trim from Z-edges to remove boundary artifacts.
  4. GPU batch size. Reduce if running out of memory.
  5. Set to True to keep only surface-facing membranes and remove internal fragments.
FilteringConfig Parameters
Parameter Type Default Description
ball_size int 3 Morphological kernel radius — larger values smooth boundaries more aggressively
min_membrane_area int 10000 Minimum voxel count for a membrane component to be kept
edge_trim_z int 5 Voxels to trim from Z-edges to remove boundary artifacts
edge_trim_xy int 3 Voxels to trim from XY-edges
batch_size int 8 GPU batch size — reduce if running out of memory
keep_surface_membranes bool False Set True to keep only surface-facing membranes and remove internal fragments
Parameter tuning guide
Symptom Adjustment
Too many small membrane fragments remaining Increase min_membrane_area
Valid structures being removed Decrease ball_size or min_membrane_area
Artifacts at volume edges Increase edge_trim_z or edge_trim_xy
Internal membrane fragments inside organelles Set keep_surface_membranes=False
GPU out of memory Decrease batch_size

Next Steps

  • Export Statistics

    Extract per-organelle volume, diameter, and coordinates from refined segmentations.

  • MemBrain-seg

    Generate the initial membrane segmentations that this workflow refines.