BUG: Guard empty re-estimation pass in ConfidenceConnected filters#6578
Merged
hjmjohnson merged 1 commit intoJul 10, 2026
Conversation
github-actions
Bot
added
type:Bug
This comment was marked as resolved.
This comment was marked as resolved.
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
The ConfidenceConnected and VectorConfidenceConnected filters fix the threshold before the iteration loop (to the seed-inclusion-boosted multiplier) but re-estimate the mean/covariance from the previous pass's segmentation every iteration. Because the statistics move while the threshold does not, a re-estimated pass can produce statistics that exclude even the seed pixel; that pass's flood then segments nothing, and the following pass counts zero samples. The accumulation loops then divided the summed mean/covariance by that zero count with no guard. In the vector filter this makes the mean NaN, which makes every Mahalanobis squared distance NaN; MahalanobisDistanceThresholdImageFunction treats a non-positive (and NaN is not greater than zero) squared distance as distance 0.0, which passes "<= threshold" for every pixel, so the next flood includes the ENTIRE image reachable from the seeds. In the scalar filter the same missing guard makes the mean/variance NaN, the threshold bounds NaN, and every comparison false, silently emptying the output; the division is meaningless either way. Guard the re-estimation loop in both filters: when the previous pass segmented no pixels there is nothing to re-estimate from, so break out of the iteration loop. This leaves the segmentation exactly as the last completed flood produced it (empty), which is the converged answer under the fixed threshold. Add a deterministic regression test (itkVectorConfidenceConnectedImageFilterEmptyReestimationTest) that engineers a single-outlier vector image where pass 0's re-estimated mean moves off the seed, pass 0's re-flood segments nothing, and the next pass would divide by zero. It asserts the output is not the whole image (exactly the empty converged segmentation) and that the re-estimated mean/covariance contain no NaN. The scalar guard prevents the same division-by-zero; its self-limiting empty-output failure mode does not warrant a second engineered scenario. Addresses item B33 of InsightSoftwareConsortium#6575
hjmjohnson
force-pushed
the
bug-confidence-connected-empty-reestimation
branch
from
d9fc995 to
58a0789
Compare
hjmjohnson
approved these changes
Jul 10, 2026
physwkim
added a commit
to physwkim/sitk-rs
that referenced
this pull request
Jul 12, 2026
…pass The re-estimation loop divides mean/covariance by num with no num == 0 guard (unlike the initial statistics' seed_cnt == 0 guard). When a pass's mask is empty -- reachable because the multiplier is fixed pre-loop while mean/covariance drift each round -- num == 0 makes the statistics NaN, which clamps every Mahalanobis distance to 0 and floods the whole image. Stop the loop at num == 0, keeping the previous pass's statistics and mask (upstream fix PR InsightSoftwareConsortium/ITK#6578). The scalar confidence_connected's num == 0 is already unreachable. Filed as #6575 item B33.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Addresses item B33 of #6575.
VectorConfidenceConnectedImageFilterre-estimates the mean and covariance from the previous pass's segmentation each iteration, but the Mahalanobis threshold is fixed before the loop (the seed-inclusion boost runs pre-loop only). Because the statistics move while the threshold does not, a pass can re-estimate statistics that exclude even the seeds; that pass's flood segments nothing, and the next pass countsnum == 0samples and divides by it. The resulting NaN mean makes every Mahalanobis squared distance NaN, whichMahalanobisDistanceThresholdImageFunctionclamps to a distance of0.0(NaN fails its> 0test), and0.0 <= thresholdholds for any positive threshold — the next flood includes every pixel in the image.The scalar sibling
ConfidenceConnectedImageFilterhas the same unguarded division in its re-estimation loop (both filters do guard the equivalent empty case in their initial statistics). Its failure mode is self-limiting — NaN bounds make every comparison false and the output goes silently empty — but the division is equally meaningless.The fix adds the missing guard to both filters: when a re-estimation pass finds zero samples, stop iterating instead of dividing by zero. The segmentation is left exactly as the last completed flood produced it — the converged answer under the fixed threshold — and the previously estimated statistics remain finite.
The new regression test engineers the empty pass deterministically (single outlier seed against a uniform background; hand-traced pass by pass in the test's header comment): the pre-loop flood covers the whole image, the first re-estimation moves the mean far enough that the seed fails the fixed threshold (empty flood), and the next pass hits
num == 0. Unfixed, the final output floods all 256 pixels; fixed, the filter stops at the converged empty segmentation with NaN-free statistics.Verified locally (Linux, GCC 13, Release): the new test fails on unmodified
main(Segmented pixels: 256 of 256) and passes with the fix; all 12 ITKRegionGrowing and 24 ITKDisplacementField tests pass with the fix applied.PR Checklist
AI assistance