Skip to content

HINEC Seeding Strategy Architecture

Design Principle

Centralized seeding - All seeding decisions happen in runTractography.m, not in the tractography algorithm itself.

Architecture

runTractography.m                    nim_tractography_standard.m
     │                                        │
     ├─ Loads nim data                        │
     ├─ Configures seeding strategy           │
     ├─ Creates seed_mask                     │
     ├─ Sets options.seed_mask                │
     │                                        │
     └─► Calls nim_tractography_standard()    │
                                              │
                                              ├─ Validates seed_mask exists
                                              ├─ Calls generate_seed_points_fact()
                                              └─ Executes tracking algorithm

Rationale

Why Centralize Seeding?

  1. Separation of Concerns

    • runTractography.m = Policy (what to seed)
    • nim_tractography_standard.m = Mechanism (how to track)

  2. Flexibility

    • Different pipelines can use different seeding strategies
    • IronTract can modify seeds without changing algorithm
    • Easy to test different seeding approaches

  3. Maintainability

    • Single source of truth for seeding decisions
    • No duplicate fallback logic
    • Clear error messages when seed mask missing

Seeding Strategy Hierarchy

runTractography.m implements a priority hierarchy:

Strategy 1: Preprocessed Brain Mask (BEST)

if isfield(nim, 'mask') && ~isempty(nim.mask)
    brain_mask = nim.mask > 0.5;
    brain_mask = brain_mask & (nim.FA > 0.05);  % Exclude CSF
  • Source: FSL brain extraction or manual mask
  • Coverage: Complete brain with clean boundaries
  • Quality: Best anatomical accuracy

Strategy 2: Expanded Parcellation Mask

elseif isfield(nim, 'parcellation_mask')
    parcel_mask = nim.parcellation_mask > 0;
    brain_mask = imdilate(parcel_mask, strel('sphere', 3));
    brain_mask = brain_mask & (nim.FA > 0.05);
  • Source: Atlas-based parcellation
  • Coverage: Labeled regions + surrounding tissue
  • Quality: Good for known anatomical structures

Strategy 3: FA Threshold Fallback (CONSERVATIVE)

else
    brain_mask = nim.FA > 0.10;
  • Source: Fractional anisotropy map
  • Coverage: White matter core only
  • Quality: Misses low-anisotropy regions (fornix, cingulum)
  • WARNING: This was causing the ISMRM validation failure

Implementation Details

runTractography.m

% Lines 76-138: Seeding strategy configuration
%% CENTRALIZED SEEDING STRATEGY
fprintf('\n=== Configuring Seeding Strategy ===\n');

% Try strategies in priority order
if isfield(nim, 'mask') ...
    % Strategy 1
elseif isfield(nim, 'parcellation_mask') ...
    % Strategy 2
else
    % Strategy 3
end

% Quality check
if sum(seed_mask(:)) == 0
    error('Seed mask is empty!');
end

options.seed_mask = seed_mask;

nim_tractography_standard.m

% Lines 152-159: Validation only (no fallback)
if isempty(options.seed_mask)
    error('No seed mask provided. Configure in runTractography.m');
end
options.seed_mask = logical(options.seed_mask > 0);

Special Cases

IronTract Challenge

IronTract requires seeding from injection sites, which may be outside brain mask.

Handled by: nim_irontract_submit.m (lines 102-128) 

% Add injection site to seed mask
seed_mask = seed_mask | (injection_mask > 0);

This doesn't modify runTractography.m seeding - it happens downstream in challenge-specific code.

Testing Seeding Strategies

Check Current Strategy

runTractography('data.mat');
% Look for output: "Strategy: brain_mask" or "expanded_parcellation" or "fa_threshold"

Force Specific Strategy

% Force FA-threshold (for comparison)
nim = load('data.mat');
nim.mask = []; % Remove brain mask
save('data_no_mask.mat', 'nim');
runTractography('data_no_mask.mat');

Performance Impact

Brain Mask vs FA Threshold

Metric Brain Mask FA Threshold Improvement
Seed voxels ~50,000 ~30,000 +67%
Coverage Whole brain WM core only Complete
Bundle detection All bundles WM only Fornix, cingulum detected
False positives Low Very low Slight increase

Seed Density

Current setting: 4 seeds/voxel - 50,000 voxels × 4 = 200,000 seeds - Bidirectional tracking = 400,000 potential tracks

Common Issues

"No seed mask provided" Error

Cause: Called nim_tractography_standard() directly without setting options.seed_mask

Solution: Always use runTractography() wrapper, not direct call

ISMRM Validation Shows 0% Coverage

Cause: FA-threshold seeding misses low-anisotropy bundles

Solution: Ensure brain mask exists (check nim.mask field)

IronTract Seeds Not Covering Injection Site

Cause: Injection site outside brain mask

Solution: Already handled by nim_irontract_submit.m - it adds injection voxels

Future Improvements

  1. Adaptive seeding density based on local FA
  2. Multi-resolution seeding (coarse in uniform regions, dense in complex areas)
  3. Anatomically-informed seeding using known bundle endpoints
  4. Probabilistic seeding with importance sampling

References

  • ISMRM Tractography Challenge: Whole-brain seeding required for full bundle coverage
  • IronTract Challenge: Injection-site seeding for anatomical accuracy
  • HINEC validation: Low FA seeding caused failure to detect 25/26 bundles