Graphcut Segmentation¶

Semi-automated image segmentation using the max-flow/min-cut graphcut method.

Overview¶

The Graphcut segmentation tool provides semi-automated segmentation using the max-flow/min-cut algorithm. It is based on the Max-flow/min-cut algorithm by Yuri Boykov and Vladimir Kolmogorov, with a MATLAB implementation by Michael Rubinstein.

Instead of segmenting individual pixels, it operates on superpixels (2D) or supervoxels (3D), generated via the SLIC algorithm by Radhakrishna Achanta et al. or the Watershed algorithm. SLIC superpixels excel for objects with intensity contrast, while Watershed superpixels are better for objects with distinct boundaries.
Precomputing superpixels requires initial processing time but accelerates subsequent segmentation.

General example¶

Watch on YouTube

Mode panel¶

The Mode panel lets you select the segmentation scope.

• 2D, current slice only — segments only the current slice in the Image View panel
• 2D, slice-by-slice — applies 2D segmentation to each slice individually
• 3D, volume — performs 3D segmentation on the entire dataset or a subarea (see Subarea panel below)
• 3D, volume, grid — segments a large dataset by dividing it into subvolumes (defined by Chop fields); the subvolume centred in the Image View panel is processed (enable the centre marker via on the Quick Access Bar).
  Use Segment All to process all subvolumes.

Subarea panel¶

The Subarea panel defines a dataset subset for processing, useful for large datasets or binning.

• X sets the width range (e.g., 1:512)
• Y sets the height range
• Z sets the z-slice range
• from Selection fills X, Y, and Z with coordinates from the Selection layer's bounding box
• Current View limits X and Y to the visible area in the Image View panel
• Reset restores full dataset dimensions
• Bin, xtimes: binning factor XY; Z to reduce resolution for faster processing (e.g., 2; 1 halves XY)

Calculation of superpixels/supervoxels¶

Superpixels (2D) or supervoxels (3D) are precomputed using SLIC or Watershed algorithms. SLIC suits intensity-contrast objects (e.g., lipid droplets); Watershed excels for boundary-defined objects.

• Superpixels: superpixel type — SLIC or Watershed
• Size of superpixels: approximate size of each superpixel (2D) or supervoxel (3D) in pixels. Smaller values give finer segmentation but increase computation time.
• Compactness: controls SLIC superpixel squareness (higher values yield squarer shapes). Not applicable to Watershed.
• Color channel: color channel for superpixel calculation.
• Type of signal: signal polarity — black-on-white (electron microscopy) or light-on-black (light microscopy, fluorescence).
• Calculate Superpixels: computes superpixels/supervoxels and builds the boundary graph. The button turns green on completion.

• Chop X/Y/Z: number of grid tiles in each dimension for 3D, volume, grid mode. Also controls SLIC supervoxel grid tiling.
• autosave: automatically prompts to save the graphcut structure to disk (*.graph) after superpixels are computed.
• parallel: enables parallel processing (parfor) for Watershed clustering in 3D, volume, grid mode. Starts a parallel pool if not already running.
• use PixelIdxList: pre-computes pixel index lists per superpixel. Speeds up mask updates at the cost of extra memory.
• Preview superpixels: overlays the computed superpixel boundaries on the current image.
• Import: loads a previously saved graphcut structure from disk (*.graph) or the MATLAB workspace.
• Export: saves superpixels and the graph. Options:
  • Export to Matlab — exports the Graphcut struct to the MATLAB workspace.
  • Save to a file — saves to a *.graph file (excluding the sparse Graph field).
  • Export to a model — writes the superpixel labels into the model layer as materials.
  • Export to 3DLines — converts the adjacency graph to a Lines3D object (not recommended for large numbers of supervoxels; see this video).

Recalculate Graph panel

• Coef: edge weight scaling factor (default 25). Graph edge weights are computed as exp(−normalised_intensity_difference × Coef). Increase to sharpen boundary cuts; decrease to allow softer cuts.
• Recalculate: rebuilds the sparse boundary graph from the existing edge list without recomputing superpixels. Use this after changing Coef.

Image segmentation settings¶

The Graphcut workflow uses labeled areas (Background and Object) for segmentation. It handles objects with both boundaries and intensity contrast, making it versatile.

• Background: model material that labels background areas.
• Object: model material that labels objects of interest.
• Update lists: refreshes the material dropdowns from the current model.
• Auto update: re-runs segmentation automatically after each paint stroke on the currently shown slice. Best for small datasets (~400×400×400 px).

Image segmentation example¶

Steps for segmenting mitochondria using Graphcut:

• Load a sample dataset: Ribbon → Home → Import image from → URL / Zarr, enter http://mib.helsinki.fi/tutorials/WatershedDemo/watershed_demo1.tif
• Add a material named Background in the Segmentation panel with + (right-click to rename)
• Use the Brush tool to label cytoplasm, then press A to add it to Background

• Add another material named Seeds with +
• Label mitochondria interiors, then press A to add to Seeds

• Start Graphcut via Ribbon → Tools → Semi-automatic segmentation → Graphcut
• Select Watershed in Superpixels
• Ensure Background and Seeds are set in the Image segmentation settings panel
• Click Segment to segment mitochondria
• Add more seeds to refine, then re-segment with Segment or enable Auto update
• Results appear in the Mask layer
• Optionally smooth the mask: Ribbon → Mask → Smooth Mask

References¶

• Max-flow/min-cut algorithm by Yuri Boykov and Vladimir Kolmogorov (research license only)
• MATLAB wrapper for maxflow by Michael Rubinstein
• SLIC superpixels and supervoxels by Radhakrishna Achanta et al.
• Region Adjacency Graph (RAG) by David Legland, INRA, France (modified for Watershed)

Back to MIB | User interface | Ribbon | Tools