Algorithms

SlicerAdaptiveBrush provides eight segmentation algorithms. Each has different characteristics suited to different segmentation tasks.

Algorithm Comparison

Algorithm	Speed	Precision	Boundary	Best For
Watershed	Medium	High	Excellent	General use (default)
Connected Threshold	Very Fast	Low	Poor	Quick rough segmentation
Region Growing	Fast	Medium	Fair	Homogeneous regions
Threshold Brush	Very Fast	Variable	N/A	Simple threshold painting
Geodesic Distance	Medium	High	Good	Structures with clear edges
Level Set	Slow	Very High	Excellent	High precision needs
Random Walker	Slow	High	Good	Complex boundaries

Watershed

Recommended for most use cases.

The watershed algorithm treats the image as a topographic surface where intensity values represent elevation. It finds boundaries by simulating water flooding from marker points.

Parameters

• Gradient Scale (0.5-3.0): Controls smoothing of gradient magnitude

  • Lower values: More detailed boundaries

  • Higher values: Smoother boundaries, less noise sensitivity

• Smoothing (0.1-1.0): Gaussian smoothing before gradient computation

  • Lower values: Preserve fine details

  • Higher values: Reduce noise

When to Use

• General segmentation tasks

• Structures with clear boundaries

• When boundary adherence is important

Example

Watershed performs well on brain tumor segmentation where the tumor has a clear boundary with surrounding tissue:

Best parameters from optimization:
- Algorithm: watershed
- Edge sensitivity: 70
- Gradient scale: 1.06
- Smoothing: 0.78
- Dice score: 99.91%

Connected Threshold

Fastest algorithm, lowest precision.

Segments all connected voxels within an intensity range from the seed point. Simple but effective for homogeneous regions.

Parameters

Uses the global Edge Sensitivity to determine threshold range around seed intensity.

When to Use

• Quick rough segmentation

• Highly homogeneous regions (e.g., air, fluid)

• When speed is more important than precision

Limitations

• No boundary refinement

• Sensitive to intensity variations

• May leak through weak boundaries

Region Growing

Fast with confidence-based growing.

Uses SimpleITK’s ConfidenceConnected filter which iteratively grows the region based on mean and standard deviation of included voxels.

Parameters

• Multiplier: Implicit from edge sensitivity

• Iterations: Number of growing iterations (default: 5)

When to Use

• Regions with consistent intensity statistics

• When connected threshold leaks too much

• Moderate precision requirements

Threshold Brush

Simple intensity thresholding with auto-detection.

Paints all voxels within a threshold range under the brush. Can automatically determine whether to segment above or below threshold based on seed intensity.

Parameters

• Auto-threshold method: Otsu, Huang, Triangle, Maximum Entropy, IsoData, Li

• Manual mode: Set lower/upper thresholds directly

• Set from seed: Calculate thresholds from seed intensity with tolerance

When to Use

• Simple intensity-based segmentation

• When you know the exact intensity range

• Quick selection of high/low intensity structures

Geodesic Distance

Good for structures with clear edges.

Combines intensity similarity and gradient information using fast marching. The “speed” function rewards paths through similar intensities and penalizes crossing edges.

Parameters

• Edge Weight (0-1): Balance between intensity and gradient

  • 0: Pure intensity-based (like region growing)

  • 1: Pure edge-based (stops at all edges)

  • 0.5: Balanced (default)

When to Use

• Structures with clear edges

• When watershed over-segments

• Tubular structures (vessels)

Level Set

Highest precision, slowest speed.

Uses geodesic active contours that evolve a curve/surface to minimize an energy function. Provides smooth, accurate boundaries.

Parameters

• Iterations (30-150): Number of evolution iterations

  • More iterations: More accurate but slower

• Propagation (0.5-2.0): Expansion/contraction force

  • > 1: Tends to expand

  • < 1: Tends to contract

When to Use

• When precision is paramount

• Smooth, well-defined boundaries needed

• Willing to wait for results

Limitations

• Significantly slower than other methods

• May require parameter tuning

• Can get stuck in local minima

Random Walker

Probabilistic segmentation for complex cases.

Treats segmentation as a probability diffusion problem. Labels propagate from seed points based on image gradients.

Parameters

• Beta (10-200): Controls sensitivity to gradients

  • Higher values: More sensitive to edges

  • Lower values: Smoother propagation

When to Use

• Complex, intertwined structures

• Multiple competing regions

• When deterministic methods fail

Requirements

Requires scikit-image. Falls back to region growing if not available.

Algorithm Selection Guide

By Structure Type

Structure	Recommended Algorithm
Tumor/Lesion	Watershed
Bone (CT)	Threshold Brush or Connected Threshold
Brain tissue	Watershed or Level Set
Blood vessels	Geodesic Distance
Fluid/CSF	Connected Threshold
Complex anatomy	Random Walker

By Priority

Priority	Recommended Algorithm
Speed	Connected Threshold or Threshold Brush
Precision	Level Set
Balanced	Watershed (default)
Boundary accuracy	Watershed or Geodesic Distance

From Optimization Results

Based on automated optimization against gold standards:

Parameter Importance:
- Algorithm choice: 73.1%
- Brush radius: 18.6%
- Threshold zone: 5.1%
- Edge sensitivity: 3.2%

Key insight: Algorithm selection is by far the most important parameter. Watershed consistently outperforms other algorithms for boundary-sensitive tasks.