Algorithm Profiles

This document describes the algorithm profiling system used to optimize segmentation parameters.

Overview

Algorithm profiles provide data-driven recommendations for algorithm and parameter selection based on:

Automated optimization - Optuna-based parameter search against gold standards
Modality characteristics - Known imaging physics and typical intensity patterns
Structure properties - Shape, size, and boundary characteristics

Profile Data Structure

Algorithm profiles are stored in algorithm_profiles.json at the project root.

Parameter Importance

From optimization studies, we found:

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.

Algorithm Characteristics

Each algorithm has profiled characteristics:

{
  "watershed": {
    "speed": "medium",
    "precision": "high",
    "boundary_adherence": "excellent",
    "recommended_for": ["tumor", "lesion", "brain_tissue"],
    "parameters": {
      "gradient_scale": {"range": [0.5, 3.0], "optimal": 1.1},
      "smoothing": {"range": [0.1, 1.0], "optimal": 0.78}
    }
  }
}

Modality Recommendations

Recommendations are organized by imaging modality and structure:

{
  "ct": {
    "bone": {
      "primary": "threshold_brush",
      "alternatives": ["connected_threshold"],
      "preset": "ct_bone"
    }
  }
}

Using Profiles Programmatically

Loading Profiles

import json
from pathlib import Path

profiles_path = Path(__file__).parent / "algorithm_profiles.json"
with open(profiles_path) as f:
    profiles = json.load(f)

Getting Recommendations

def get_recommendation(modality: str, structure: str) -> dict:
    """Get algorithm recommendation for modality/structure pair."""
    mod_recs = profiles.get("modality_recommendations", {})
    if modality in mod_recs and structure in mod_recs[modality]:
        return mod_recs[modality][structure]
    return profiles.get("structure_recommendations", {}).get(structure, {})

# Example usage
rec = get_recommendation("ct", "bone")
print(f"Primary algorithm: {rec['primary']}")  # threshold_brush
print(f"Preset to apply: {rec['preset']}")      # ct_bone

Applying Recommendations

def apply_recommendation(effect, modality: str, structure: str):
    """Apply algorithm recommendation to effect."""
    rec = get_recommendation(modality, structure)

    # Apply the preset (sets common parameters)
    if "preset" in rec:
        effect.applyPreset(rec["preset"])

    # Algorithm is selected via UI - log the recommendation
    print(f"Recommended algorithm: {rec.get('primary', 'watershed')}")

Generating New Profiles

Running Optimization

Use the optimization script to generate benchmark data:

Slicer --python-script scripts/run_optimization.py

This runs Optuna optimization against gold standards and outputs:

results.json - Full trial data
parameter_importance.json - Relative parameter importance
Screenshots and segmentations for each trial

Adding Benchmarks

After running optimization, add benchmark results to the profile:

{
  "algorithms": {
    "watershed": {
      "benchmarks": {
        "MRBrainTumor1_tumor": {
          "dice": 0.9991,
          "edge_sensitivity": 70,
          "threshold_zone": 60,
          "brush_radius_mm": 25.0
        }
      }
    }
  }
}

Profile Evolution

Adding New Modalities

Create recipes for the new modality
Run optimization with gold standards
Analyze results and add to modality_recommendations
Update documentation

Refining Recommendations

Collect user feedback on segmentation quality
Run A/B tests with alternative algorithms
Update primary/alternative recommendations based on results

Integration with Presets

Presets define common parameters (edge sensitivity, sampling method, etc.) that work well for a modality/structure combination.

Algorithm selection is independent of presets - users choose the algorithm via the UI dropdown based on:

Speed requirements - Fast vs. precise
Boundary characteristics - Sharp vs. diffuse
Algorithm profiles - Data-driven recommendations

Preset to Algorithm Mapping

While presets don’t force an algorithm, there are natural pairings:

Preset	Recommended Algorithm
`ct_bone`	Threshold Brush
`ct_soft_tissue`	Watershed
`ct_lung`	Connected Threshold
`ct_vessel_contrast`	Geodesic Distance
`mri_t1gd_tumor`	Watershed
`mri_t2_lesion`	Watershed

Future Work

Auto-selection - Automatically suggest algorithm based on modality detection
Online learning - Update profiles based on user corrections
Per-structure profiles - Build profiles for specific anatomical structures
Multi-algorithm fusion - Combine results from multiple algorithms