BUG: Skip voting-undecided pixels in MultiLabelSTAPLE seeding

[physwkim]

Addresses item B10 of #6575.

MultiLabelSTAPLEImageFilter seeds each rater's confusion matrix from a LabelVotingImageFilter pass over the same inputs. On voting ties that filter writes its undecided label, which equals the total label count — one past the confusion matrix's last valid column. Array2D::operator[] is an unchecked row pointer into a flat row-major buffer, so ++CM[k][in][out] with out == m_TotalLabelCount lands on CM[k][in + 1][0]: inside the allocation (the matrix has a spare row), but silently corrupting the NEXT observed label's row with phantom "rater observed in+1, estimated truth 0" counts. Every tie pixel in the seeding vote injects one such phantom count, skewing the row-normalized initial confusion matrices the EM iteration starts from — and the filter's public GetConfusionMatrix() output.

The EM loop and both normalizations only ever index rows and columns by real labels, so an undecided voting pixel has no legitimate slot in the matrix: it carries no (observed label, estimated true label) evidence. The fix skips such pixels in the seeding loop instead of letting the increment run past the row.

The new regression test engineers deterministic ties (two raters, four pixels, two of them split 1–1) and sets MaximumNumberOfIterations = 0, which leaves the matrices exactly as seeded. Unfixed, both raters' matrices come out with row 1 = [0.5, 0.5] and row 2 = [1, 0] (the leaked counts) and the output labels are wrong at three of four pixels ([0, 2, 0, 0]); fixed, the matrices are the exact row-normalized vote statistics (row 0 = [1, 0], row 1 = [0, 1], row 2 = [0, 0]) and the output is [0, 1, 2, 2] — agreements labeled, ties undecided. The full pass-by-pass trace is in the test's header comment.

Verified locally (Linux, GCC 13, Release): the new test fails on unmodified main (confusion matrix (1,0) = 0.5 expected 0, (2,0) = 1 expected 0; output [0, 2, 0, 0] expected [0, 1, 2, 2]) and passes with the fix; the full ITKLabelVoting suite plus the ITKImageStatistics, ITKReview, ITKDisplacementField, and ITKRegionGrowing suites pass with the fix applied.

PR Checklist

• No API changes were made (or the changes have been approved)
• No major design changes were made (or the changes have been approved)
• Added test (or behavior not changed)
• Updated API documentation (or API not changed)
• Added license to new files (if any)
• Added Python wrapping to new files (if any) as described in ITK Software Guide Section 9.5
• Added ITK examples for all new major features (if any)

AI assistance

• Tool: Claude Code (claude-fable-5, with claude-opus-4-8 subagents)
• Role: implemented the fix and regression test from the analysis previously filed as item B10 of BUG: Findings collected while porting ~60 filters — reported together in one issue #6575; hand-traced the seeded matrices and final labeling (fixed and unfixed) before implementation
• Testing: built locally and verified the regression test fails before / passes after the fix; five module test suites pass with the fix
• Note: clang-format 19.1.7 was not available locally; formatting was hand-matched to ITK style and may need the CI formatter's touch-up

[greptile-apps]

Greptile Summary

This PR fixes how MultiLabelSTAPLEImageFilter seeds confusion matrices when the initial voting pass produces tie labels. The main changes are:

• Skip voting-undecided pixels during confusion-matrix seeding.
• Keep seeded matrix counts limited to real output labels.
• Add a GoogleTest regression for tied votes, zero EM iterations, seeded matrices, and undecided output labels.
• Register the new LabelVoting GoogleTest driver in CMake.

Confidence Score: 5/5

Safe to merge with minimal risk.

The change is narrow, matches the existing matrix dimensions, and the test directly covers the corrected seeding behavior.

No files require special attention.

T-Rex Logs

What T-Rex did

• Checked the multilabelstaple tool availability and found that cmake, ctest, ninja, and pixi were not found, indicating the build environment lacks core configuration tools.
• Attempted the narrow CMake configure command specified for the PR, and the configure step exited with code 127.
• Inspected the LabelVoting CMake test registration and the intended GTest assertions that would run in a properly provisioned build.

_{Ran code and verified through T-Rex}

Important Files Changed

Filename	Overview
Modules/Segmentation/LabelVoting/include/itkMultiLabelSTAPLEImageFilter.hxx	Skips voting-undecided labels during confusion-matrix seeding, preventing one-past-column writes into the next row.
Modules/Segmentation/LabelVoting/test/itkMultiLabelSTAPLEImageFilterGTest.cxx	Adds a focused regression test that verifies tie pixels do not corrupt seeded confusion matrices and remain undecided in the output.
Modules/Segmentation/LabelVoting/test/CMakeLists.txt	Adds a GoogleTest driver for the new LabelVoting regression test following existing ITK module patterns.

Sequence Diagram

%%{init: {'theme': 'neutral'}}%%
sequenceDiagram
participant Inputs as Rater input images
participant STAPLE as MultiLabelSTAPLEImageFilter
participant Voting as LabelVotingImageFilter
participant CM as Confusion matrices
participant Output as Output image

STAPLE->>Voting: Run initial label voting over inputs
Voting-->>STAPLE: Voted labels, including undecided ties
loop Each rater and pixel
    STAPLE->>STAPLE: "Check voted label < total label count"
    alt Real voted label
        STAPLE->>CM: Increment observed label / voted truth count
    else Undecided tie label
        STAPLE->>STAPLE: Skip seeding update
    end
end
STAPLE->>CM: Normalize seeded rows
STAPLE->>Output: Generate final labels from seeded/EM matrices

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
sequenceDiagram
participant Inputs as Rater input images
participant STAPLE as MultiLabelSTAPLEImageFilter
participant Voting as LabelVotingImageFilter
participant CM as Confusion matrices
participant Output as Output image

STAPLE->>Voting: Run initial label voting over inputs
Voting-->>STAPLE: Voted labels, including undecided ties
loop Each rater and pixel
    STAPLE->>STAPLE: "Check voted label < total label count"
    alt Real voted label
        STAPLE->>CM: Increment observed label / voted truth count
    else Undecided tie label
        STAPLE->>STAPLE: Skip seeding update
    end
end
STAPLE->>CM: Normalize seeded rows
STAPLE->>Output: Generate final labels from seeded/EM matrices

Loading

_{Reviews (2): Last reviewed commit: "BUG: Skip voting-undecided pixels in Mul..." | Re-trigger Greptile}