optimize-segmentation

Optimize AdaptiveBrush segmentation parameters for a specific target structure.

Usage

code

/optimize-segmentation [sample-data]

Where sample-data is optional:

•MRBrainTumor1 (default) - Brain tumor MRI
•MRHead - Normal brain MRI
•Or specify a loaded volume name

What This Skill Does

•Loads sample data (or uses specified volume)
•Prompts user to place 5 reference points on target structure
•Runs all working algorithms with those points
•Compares results (voxel count, timing, consistency)
•Recommends best algorithm and parameters
•Generates optimization report

Workflow

Step 1: Load Data and Get Reference Points

Launch Slicer for interactive point placement:

bash

# Read SLICER_PATH from .env first
SLICER_PATH=$(cat .env | grep SLICER_PATH | cut -d= -f2)
$SLICER_PATH

Instruct user:

•Load sample data: File > Download Sample Data > [dataset]
•Go to Markups module
•
Place 5 fiducial points on the target structure:
- •Point 1: Center of structure
- •Point 2: Superior edge
- •Point 3: Anterior edge
- •Point 4: Posterior edge
- •Point 5: Lateral edge
•Export markups: Right-click > Export as... → save as CSV or copy coordinates

Step 2: Convert Coordinates (if needed)

CRITICAL: Slicer markups export in LPS format, but the API uses RAS.

Conversion formula:

•R = -L (negate first coordinate)
•A = -P (negate second coordinate)
•S = S (keep third coordinate)

Example:

code

LPS: (5.31, -34.77, 20.83)
RAS: (-5.31, 34.77, 20.83)

Step 3: Update Test Coordinates

Edit SegmentEditorAdaptiveBrushTester/TestCases/test_optimization_tumor.py:

python

TUMOR_CLICK_POINTS = [
    # Convert user's LPS points to RAS
    (-L1, -P1, S1),  # Point 1: Center
    (-L2, -P2, S2),  # Point 2: Superior
    (-L3, -P3, S3),  # Point 3: Anterior
    (-L4, -P4, S4),  # Point 4: Posterior
    (-L5, -P5, S5),  # Point 5: Lateral
]

Also update parameters to test:

python

OPTIMIZATION_PARAMS = {
    "brush_radius_mm": 20.0,  # Adjust based on structure size
    "edge_sensitivity": 35,   # 0-100, lower = more propagation
    "inner_radius_ratio": 0.3,
}

Step 4: Run Optimization Test

bash

SLICER_PATH=$(cat .env | grep SLICER_PATH | cut -d= -f2)
$SLICER_PATH --python-script scripts/run_tests.py --exit optimization

Step 5: Analyze Results

Read the results file:

bash

# Find latest optimization run
ls -t test_runs/ | grep optimization | head -1

Then read test_runs/<run>/results.json to get:

•Voxel counts per algorithm
•Timing data
•Screenshots

Step 6: Generate Report

Create optimization report with:

•
Data Summary
- •Dataset used
- •Reference points (RAS coordinates)
- •Structure being segmented
•
Algorithm Comparison Table

Algorithm Voxels Time (ms) Voxels/ms Rank
•
Recommendations
- •Best algorithm for this structure
- •Optimal brush radius
- •Optimal edge sensitivity
•
Screenshots
- •Link to screenshots for visual verification

Algorithms

Algorithm	Description	Best For	Status
`connected_threshold`	Flood fill within intensity range	Maximum coverage	✓ Working
`threshold_brush`	Threshold + brush mask	Consistent results	✓ Working
`region_growing`	Iterative region expansion	Homogeneous regions	✓ Working
`watershed`	Gradient-based boundaries	Edge-sensitive segmentation	✓ Working
`geodesic_distance`	Distance-weighted propagation	Complex boundaries	✓ Working
`random_walker`	Probabilistic label propagation	Uncertain boundaries	✓ Working
`level_set`	Level set evolution	High precision	✓ Working

Parameter Guidelines

Brush Radius

•Small structures (< 10mm): 5-10mm brush
•Medium structures (10-30mm): 15-20mm brush
•Large structures (> 30mm): 25-35mm brush

Edge Sensitivity (0-100)

•Low (10-30): More propagation, may leak
•Medium (30-50): Balanced
•High (50-80): Stops at weak edges, may under-segment

Number of Clicks

•Minimum: 3 (center + 2 edges)
•Recommended: 5 (center + 4 cardinal edges)
•Maximum useful: 8-10 for complex shapes

Example Output

code

OPTIMIZATION RESULTS: MRBrainTumor1
===================================
Reference Points: 5 (user-placed on tumor)
Brush Radius: 20mm
Edge Sensitivity: 35

ALGORITHM COMPARISON:
---------------------
1. connected_threshold: 22,696 voxels (1,252ms) - BEST COVERAGE
2. threshold_brush:     15,298 voxels (1,095ms) - MOST CONSISTENT
3. watershed:           13,159 voxels (1,808ms) - GOOD EDGES
4. region_growing:      11,694 voxels (1,172ms) - BALANCED

RECOMMENDATIONS:
----------------
- For maximum recall: connected_threshold
- For consistency: threshold_brush
- For edge precision: watershed

Optimal parameters for this structure:
- Brush: 20mm
- Sensitivity: 35
- Algorithm: threshold_brush (recommended)

Troubleshooting

"0 voxels" for all algorithms

•Check coordinate conversion (LPS vs RAS)
•Verify points are inside the volume
•Check brush radius is large enough

Segmentation leaks outside structure

•Increase edge sensitivity
•Decrease brush radius
•Use threshold_brush instead of connected_threshold

Inconsistent results between clicks

•Increase brush radius for more overlap
•Use region_growing for homogeneous structures
•Place points more centrally in the structure