Membrane Detector¶

Automatic image segmentation using a train-and-predict scheme based on random forest classification.

Overview¶

The Membrane Detector is a pixel classifier for automatic image segmentation, based on Random Forest for Membrane Detection by Verena Kaynig, using the randomforest-matlab library by Abhishek Jaiantilal.

It excels at segmenting complex datasets where global thresholding fails due to varying background intensities or subtle intensity gradients.

Launch via Ribbon → Tools → Classifiers → Membrane detector.

Interface¶

The tool is divided into:

• Top fields — paths for the temp directory and classifier file
• Workflow panel — step-by-step guide to training and predicting
• Train classifier panel — training and prediction settings
• Log list — progress messages from the last operation

File paths¶

• Temp dir — directory for temporary feature files (.fm) generated during training and prediction. Defaults to RF_Temp/ next to the open dataset. Use the browse button to change it.
• Classifier filename — full path to the .forest classifier file. Use the browse button to pick a location.

Workflow panel¶

Guides the four-step process.

Step 1 — Train classifier... (toggle): activates Train mode. Training controls in the Train classifier panel are enabled. The main action button reads Train classifier.

Step 2 — Save classifier: saves the trained classifier to the file specified in Classifier filename.

Step 3 — Predict dataset... (toggle): activates Predict mode. Training-specific controls are disabled; Predict shown slice appears. The main action button reads Predict dataset.

Step 4 (optional) — Wipe Temp dir: deletes all files in Temp dir to free disk space.

Train classifier panel¶

Settings used for both training and prediction.

• Object — model material that marks the object (e.g., membrane).
• Background — model material that marks the background.
• Context size — context window radius in pixels for feature extraction. Smaller values suit short or highly curved structures; larger values capture more context.
• Membrane thickness — expected thickness of the target structure in pixels.
• Votes thresholding — threshold applied to the classifier vote probability (0–1) to produce a binary result. Values closer to 1 require stronger classifier confidence.
• Export votes to MATLAB — exports the raw per-pixel vote probability map to the MATLAB workspace as mibVotes after each training or prediction step.
• Get closed contours — applies skeletonise + dilate morphological closing to the prediction, useful for enforcing closed membrane outlines.
• Predict shown slice (predict mode only) — runs prediction on the currently displayed slice as a quick preview. Results appear in the Selection layer.
• Train classifier / Predict dataset — the main action button. Its label changes with the active mode:
  • Train mode: trains the classifier on the current slice using the labeled materials.
  • Predict mode: runs prediction across the entire dataset (or the subrange shown in the log).

Step-by-step example¶

This example segments endoplasmic reticulum from a time-lapse widefield dataset where global thresholding is ineffective due to background intensity gradients.

1 — Label training areas¶

• Start a new model in the Segmentation panel with New
• Add two materials with + and rename them Object and Background

• Use the Brush tool to mark regions:
  • Select endoplasmic reticulum profiles → press A to add to Object
  • Mark background areas → press A to add to Background

2 — Train the classifier¶

• Launch via Ribbon → Tools → Classifiers → Membrane detector
• Set Object to Object and Background to Background
• Adjust Context size and Membrane thickness as needed
• In Step 1, click Train classifier... to activate train mode
• Click Train classifier to train on the current slice

• If the result is unsatisfactory, add more labels (across multiple slices if needed) and retrain
• Once satisfied, click Save classifier (Step 2)

3 — Predict the whole dataset¶

• In Step 3, click Predict dataset... to activate predict mode
• Use Predict shown slice to preview the result on the current slice

• If acceptable, click Predict dataset to process the full dataset

Prediction results are written to the Selection layer. Transfer them to the Mask or Model layer for further refinement or saving.

4 — Clean up¶

• Click Wipe Temp dir (Step 4, optional) to remove the large feature files generated during prediction.

Batch scripting¶

This tool supports batch scripting for automation.

% Train on the current slice
BatchOpt.Mode                 = {'trainClassifier', {'trainClassifier','predictDataset'}};
BatchOpt.ObjectMaterial       = {'Object'};
BatchOpt.BackgroundMaterial   = {'Background'};
BatchOpt.ContextSize          = {29, [1 Inf], true};
BatchOpt.MembraneThickness    = {3, [1 Inf], true};
BatchOpt.VotesThreshold       = {0.5, [0 1], false};
BatchOpt.ExportVotes          = false;
BatchOpt.SkelClosed           = false;

obj.mibController.startController('controllers.MembranePixClassifier', [], BatchOpt);

% Predict the full dataset
BatchOpt.Mode{1} = 'predictDataset';
obj.mibController.startController('controllers.MembranePixClassifier', [], BatchOpt);

References¶

• Random Forest for Membrane Detection by Verena Kaynig
• randomforest-matlab by Abhishek Jaiantilal

Back to MIB | User interface | Ribbon | Tools