BUG: Findings collected while porting ~60 filters — reported together in one issue

[physwkim]

Following up on #6569: while porting SimpleITK's filter surface to another language, I kept a ledger of everything in the ITK sources that looked wrong or surprising. It has grown to 92 suspected bugs and 118 quirks (originally 30 and 24; follow-up batches B31–B92 and Q25–Q118) — too many to file individually without flooding the tracker, so I am reporting them together here for triage. The quirks now live in #6684; this issue tracks the 92 suspected bugs only. Happy to split any subset into standalone issues if that is more useful; I am not planning PRs for all of these, so please treat this as raw material.

Status at a glance (maintainer triage — updated 2026-07-23 by @hjmjohnson)

Bucket	Count
✅ Bugs fixed and merged	82 of 92
🚀 Bugs fixed, PR open	3 (#6681, #6682, #6683)
🔄 Bugs fixed on a branch, no ITK PR yet	0
⬜ Bugs unclaimed (all vendored ThirdParty)	1 (B64 MetaIO LIST is in upstream PR)
❌ Bugs withdrawn — investigated, not defects	4
Quirks (moved to #6684)	16 of 118 dispositioned

Every ledger PR that has completed review has merged (latest: #6672 (B34) and #6673 (B35), split from superseded draft #6669). Three are open now: #6681 (B63/Q82), #6682 (B91), #6683 (B52). The completed work is collapsed below; what follows is only what still needs someone.

⬅ Outstanding work

Bugs fixed, PR open (3)

#	Site	One-line symptom
🚀 B52	niftilib nifti_read_buffer (vendored)	maps every non-finite float/double/complex component to 0 on read wherever isfinite is a macro — always, since it is a C99 <math.h> macro everywhere — so NaN/±Inf pixels are silently destroyed; confirmed empirically (1.5 nan inf -inf 2.5 reads back 1.5 0 0 0 2.5) — fix: #6683 (open), via ISC/nifti_clib f24a6078 + UpdateFromUpstream.sh
🚀 B91	vnl vnl_matlab_readhdr::read_hdr (vendored)	The no-swap switch lists 7 of the 8 legal type_t encodings — 1010 (big-endian, column-wise, single-precision), which vnl's own writer emits on a big-endian host, is byte-swapped as though foreign — fix: #6682 (open), vendored-only by maintainer decision
🚀 B63	NrrdImageIO::Write metadata dispatch	NRRD_-prefixed keys are matched by strncmp of the candidate's length — NRRD_space directions is swallowed by the space handler, and unmatched NRRD_ keys (including the reader's own NRRD_pixel_original_axis) are silently dropped — fix: #6681 (open)

Ledger item	PR	Author	State
B52 (niftilib non-finite pixel zeroing)	#6683	@hjmjohnson	PR open — fixed in ISC/nifti_clib (f24a6078, both nifti1 and nifti2) and imported via UpdateFromUpstream.sh; opt-in fix_floats option keeps the old default for other consumers; 49/49 NIfTI tests
B91 (vnl matlab reader native 1010 code)	#6682	@hjmjohnson	PR open — endian-agnostic fail-before/pass-after GTests; vendored-only fix, no upstream vxl submission planned
B63, Q82 (NrrdImageIO, non-vendored)	#6681	@hjmjohnson	PR open — rebased on the merged #6674 snapshot; also adds ITK-side regression GTests for B67–B69

Vendored MetaIO — upstream PRs open at Kitware/MetaIO (3)

All three are fixed in draft PRs against Kitware/MetaIO, the repo Modules/ThirdParty/MetaIO/UpdateFromUpstream.sh imports from. ITK picks them up by re-running that script once they merge; no ITK-side change is needed. Each carries a new MetaIO regression test verified fail-before/pass-after.

#	Site	One-line symptom
🚀 B64	MetaIO ElementDataFile = LIST (vendored)	out-of-range file dimension (LIST -1 indexes m_DimSize[-1]/m_SubQuantity[-1]; LIST NDims reads nothing) and a short slice list both report success — fix: Kitware/MetaIO#141
🚀 B65	MetaIO M_ReadElements (vendored)	a compressed image whose header omits CompressedDataSize measures from the start of the file, so a LOCAL image inflates its own ASCII header — fix: Kitware/MetaIO#142
🚀 B66	MetaIO MET_PerformUncompression (vendored)	unconditionally return true; also its dest_pos accounting skipped the final inflate call, so no byte-count check was possible — both fixed, and both call sites now honour the result — fix: Kitware/MetaIO#142

Quirks — moved to #6684

All 118 quirks were split into #6684 so this issue can be driven to completion. They are lower priority and deliberately deferred until the bug ledger closes. Of them, 16 are already dispositioned (Q25–Q38 in open PR #6671, Q82 in open PR #6681, Q84 merged via #6674); the remaining 102 are untriaged and tracked there.

✅ Completed — 82 merged bug fixes (table)

#	Site	One-line symptom
✅ B1	ProjectionImageFilter	Collapsed-axis origin shift uses the axis index, unsigned; axis 0 wraps to a ~2³¹·spacing shift — fix: #6594 (merged)
✅ B2	N4BiasFieldCorrectionImageFilter::SharpenImage	Degenerate bin range → 0.0/0.0 = NaN → NaN→int cast (UB) → unchecked histogram index — fix: #6589 (merged)
✅ B3	SLICImageFilter	Empty cluster → 0/0 NaN centroid → NaN→integer index cast (UB) — fix: #6589 (merged)
✅ B4	WienerDeconvolutionImageFilter	Pf == Pn → complex division by zero → NaN output (guard passes on inf) — fix: #6589 (merged)
✅ B6	GridImageSource	Kernel loop walks only the first line per axis; origin cancels out of the pattern — fix: #6600 (merged)
✅ B7	MinMaxCurvatureFlowFunction (3-D)	acos taken on a gradient rescaled to length r, not 1 — wrong polar angle for radius ≥ 2 — fixed: #6603 (merged)
✅ B8	MinMaxCurvatureFlowFunction	Neighborhood scales divide by the stencil radius but only ±1 neighbors are sampled — update r² too small — fix: #6604 (merged)
✅ B9	AdaptiveHistogramEqualization	Constant input → iscale == 0 → NaN everywhere — fix: #6589 (merged)
✅ B10	MultiLabelSTAPLEImageFilter	Confusion-matrix increment walks into the next row on any voting tie — fix: #6579 (merged)
✅ B11	STAPLEImageFilter	All-background inputs → p_denom == 0 → NaN, unguarded — fix: #6589 (merged)
✅ B12	LabelOverlapMeasuresImageFilter	false_positive_error guard tests the wrong quantity — NaN reachable — fix: #6589 (merged)
✅ B13	ThresholdMaximumConnectedComponentsImageFilter	Initial bisection midpoint is (upper−lower)/2; unsigned wrap when it lands below lower — fixed: #6595 (merged)
✅ B14	ScalarImageKmeansImageFilter	UseNonContiguousLabels interval underflows to 0xFFFFFFFF for > 255 classes — fix: #6596 (merged)
✅ B15	LabelVotingImageFilter	Undecided label max_label+1 wraps to 0 for a UInt8 image using all 256 labels — fix: #6597 (merged)
✅ B16	MaskedFFTNormalizedCorrelationImageFilter	Pixel squaring runs in the input integer type — int32 overflow before the cast to real — fix: #6598 (merged)
✅ B17	IterativeDeconvolutionImageFilter	Output regions clobbered before PadInput/CropOutput read them — OutputRegionMode silently ignored — fixed: #6599 (merged)
✅ B18	FastApproximateRankImageFilter::SetRank	Rank forwarded to only the first Dim−1 per-axis filters — last axis always median — fix: #6580 (merged)
✅ B19	AttributeMorphologyBaseImageFilter	stable_sort end iterator off by one — last raster pixel never sorted — fix: #6581 (merged)
✅ B20	BoxSigmaImageFilter	Radius 0 → 0.0/0.0 → NaN for every pixel — fix: #6589 (merged)
✅ B21	LevelSetNeighborhoodExtractor	Strict sign tests: a pixel exactly on the contour neutralizes its neighborhood; grid-aligned contour starves the outward march — fixed: #6601 (merged)
✅ B22	StructureTensorImageFilter (AnisotropicDiffusionLBR)	Tensor smoothing K_ρ applied along axis 0 only (single-direction filter, SetDirection never called) — fix: #6602 (merged)
✅ B23	CoherenceEnhancingDiffusionImageFilter	Constant input → ∞ rescale → NaN tensors → negative step count via UB cast → zero iterations by accident — fix: #6589 (merged)
✅ B24	FastMarchingUpwindGradientImageFilter	AllTargets stop condition counts arrival events, not distinct targets — a duplicated target point ends the march early — fix: #6658 (merged)
✅ B25	Transform::ApplyToImageMetadata	Null GetInverseTransform() (singular or non-linear transform) dereferenced without a check — fix: #6583 (merged)
✅ B26	PatchBasedDenoisingImageFilter (POISSON)	static_cast<PixelValueType>(0.99999) is 0 for integer types — step size collapses to 1e-5 always — fix: #6633 (merged)
✅ B27	PatchBasedDenoisingImageFilter	Update write-back casts an unclamped negative float to an unsigned pixel type — UB — fix: #6589 (merged)
✅ B28	UniformRandomSpatialNeighborSubsampler	Search loops forever when the search box is exactly the query point (kernel-bandwidth path) — fix: #6605 (merged)
✅ B29	GaussianRandomSpatialNeighborSubsampler	Negative Gaussian variate cast to unsigned int — UB; rejection loop relies on the wrap — fix: #6589 (merged)
✅ B30	RegionBasedLevelSetFunction	Reinitialization smoothing term silently degenerates to a bare Laplacian when CurvatureWeight == 0 — fix: #6606 (merged)
✅ B31	IterativeInverseDisplacementFieldImageFilter	smallestError carries over from the previous pixel when the mapped point starts outside the buffer — output depends on raster order — fix: #6576 (merged)
✅ B32	InverseDisplacementFieldImageFilter	the field-subsampling resampler never gets the input's direction — with non-identity direction cosines the spline landmarks are not on the input lattice at all — fix: #6577 (merged)
✅ B33	VectorConfidenceConnectedImageFilter	a re-estimation pass that visits zero pixels produces NaN statistics, which invert the Mahalanobis guard — the next flood includes every pixel in the image — fix: #6578 (merged)
✅ B34	WarpImageFilter	the same-information fast-path test omits size, so a valid larger-output configuration dies in an iterator containment check instead of taking the general path — fix: #6672 (merged)
✅ B35	ExponentialDisplacementFieldImageFilter	the field is divided by 1 << numiter — UB when numiter >= 31; unreachable by default (cap 20, clamped), but an explicit SetMaximumNumberOfIterations(>= 31) makes it live — fix: #6673 (merged)
✅ B36	FastMarchingImageFilterBase::UpdateNeighbors	the neighbor loop gates on the center being strictly interior, so a border seed never propagates along the boundary-normal axis — a face seed confines the march to that face, a corner seed marches nowhere — fixed: #6613 (merged)
✅ B37	FastMarchingImageFilterBase (NoHandles)	marched-alive voxels never receive a connected-component label, so the facing-neighbor label test always passes and NoHandles degenerates to Strict — fixed: #6613 (merged)
✅ B38	FastMarchingImageFilterBase::Solve	the zero-speed guard substitutes a finite coefficient, giving zero-speed pixels an arrival time ≈ 2⁵² instead of the old filter's ∞ — speed barriers leak — fixed: #6613 (merged)
✅ B39	ScalarToRGBColormapImageFilter / ColormapFunction	NumericTraits<float/double>::min() (smallest positive) used as a "most negative" seed in two places — wrong rescale for non-positive float images, and a ≈FLT_MAX default input range — fixed: #6615 (merged)
✅ B40	MergeLabelMapFilter (KEEP)	the deferred-collision deque is never cleared between inputs — with ≥ 3 inputs one LabelObject ends up aliased under two label keys — fix: #6590 (merged)
✅ B41	LabelMapContourOverlayImageFilter (SLICE_CONTOUR)	the slice-radius loop tests the write cursor against m_SliceDimension where it means the read index — SliceDimension == 0 draws no contour at all — fix: #6590 (merged)
✅ B42	LabelMapMaskImageFilter (Crop)	an empty/zero-line object set leaves the bounding-box seeds untouched — signed overflow (UB), then a 2×2 region at index 2⁶³−1 — fix: #6590 (merged)
✅ B43	AttributeUniqueLabelMapFilter	the "remove objects without lines" pass is while (it.IsAtEnd()) — inverted, never executes — fix: #6590 (merged)
✅ B44	LabelMap::PushLabelObject	on a completely full label range the gap scan finds nothing, skips SetLabel, and silently overwrites the entry at the object's own label instead of throwing "map is full" — fix: #6590 (merged)
✅ B45	StatisticsLabelMapFilter	weighted elongation/flatness: no eigenvalue clamp, one shared guard, no ratio > 0 check — an indefinite weighted covariance yields NaN where ShapeLabelMapFilter yields 0 — fix: #6590 (merged)
✅ B46	NormalizedCorrelationImageFilter	both divisions unguarded — a constant template or a locally-constant neighborhood produces NaN — fixed: #6612 (merged)
✅ B47	TransformIOBaseTemplate::CorrectTransformPrecisionType	find("float") result used unchecked — a transform-type string naming neither precision throws std::out_of_range past every ITK handler — fixed: #6608 (merged)
✅ B48	NiftiImageIO::Read	rescale cast buffer sized per voxel, consumed per component — heap over-read for any multi-component image with scl_slope/scl_inter — fix: #6625 (merged)
✅ B49	NiftiImageIO::Read	RescaleFunction is given the voxel count — only the first numElts of a multi-component/complex buffer are rescaled — fix: #6625 (merged)
✅ B50	NiftiImageIO::Read	RAS↔LPS conversion never checks the 3-component requirement its own exception text names — 2/4-component vectors get a stride-3 sign flip across voxel boundaries — fix: #6625 (merged)
✅ B53	GiplImageIO::Write	INT/UINT images throw "Pixel Type Unknown" only after the target file has been truncated and the header written — destroys existing data, leaves a 256-byte stub — fixed: #6611 (merged)
✅ B54	GiplImageIO::Read	success test is !bad(); a short pixel block sets failbit/eofbit, never badbit — truncated file reads as success over an uninitialized tail — fixed: #6611 (merged)
✅ B55	GiplImageIO::Read (compressed)	success flag is p != nullptr on the caller's always-non-null output buffer — gzread's byte count is never inspected — fixed: #6611 (merged)
✅ B56	GiplImageIO::Write	header image_type is derived from the component type alone while the byte count multiplies by components — a vector image writes a scalar header plus N× the bytes, reading back as garbage — fixed: #6611 (merged)
✅ B57	VTKImageIO::Read	ReadBufferAsBinary's bool (false on short read) is discarded — truncated BINARY file reads as success — fixed: #6610 (merged)
✅ B58	VTKImageIO::InternalReadImageInformation	every sscanf return is discarded over uninitialized locals — an under-filled DIMENSIONS/SPACING/ORIGIN line sizes the image from indeterminate values (UB) — fixed: #6610 (merged)
✅ B59	VTKImageIO::InternalReadImageInformation	attribute dispatch is an ordered substring test — SCALARS vector_field float matches the vector arm and is misparsed as a 3-component VECTORS line — fixed: #6610 (merged)
✅ B60	VTKImageIO ASCII write	precision(16) is set on a throwaway header stream; pixel data is emitted through a second stream at the default 6 significant digits — silent precision loss — fixed: #6610 (merged)
✅ B61	ImageIOBase::ReadBuffer (ASCII)	extraction variable declared outside the loop, stream state never checked — one bad/overflowing field latches into every later component (5 99999 7 8 as short reads [5, 32767, 32767, 32767]) — fixed: #6609 (merged)
✅ B62	VTKImageIO::CanReadFile	four GetNextLine calls outside the try block — a .vtk shorter than four lines throws "Premature EOF" through the factory instead of returning false — fixed: #6610 (merged)
✅ B67	NrrdIO per-axis field parsers (vendored)	sizes/spacings/thicknesses/axis mins/axis maxs parse dim + 1 values into a NRRD_DIM_MAX(16)-element stack array — a 16-D header with a 17th token is a stack buffer overflow — fixed: #6674 (merged), upstream teem branch bug-nrrd-parse-axis-overflow on sourceforge
✅ B68	NrrdIO nrrdOriginCalculate (vendored)	the gotMin loop tests axis[0]->min for every axis (literal 0, not the loop variable) — origins computed from NaN mins, or usable mins ignored — fixed: #6674 (merged), upstream teem branch bug-nrrd-origin-calculate on sourceforge
✅ B69	NrrdIO rnParse_kinds (vendored)	the none token arm assigns the axis's centering, not its kind — kinds: none … clobbers a previously parsed centers: value — fixed: #6674 (merged), upstream teem branch bug-nrrd-kinds-none-centering on sourceforge
✅ B70	PNGImageIO::WriteSlice	the file is opened "wb" — truncating it — before the component-type switch whose default: throws; writing an int16 image destroys any existing .png and leaves it zero-byte — fixed: #6614 (merged)
✅ B72	PNGImageIO::WriteSlice	any component count > 4 declares 4 channels (PNG_COLOR_TYPE_RGB_ALPHA) while row pointers advance by the actual count — each row's trailing components are silently dropped — fixed: #6614 (merged)
✅ B73	HDF5ImageIO::WriteImageInformation	the C-string meta-data branch calls constCstringObj->GetMetaDataObjectValue() under a guard that admits constCstringObj == nullptr — null dereference for a MetaDataObject<char *> — fix: #6626 (merged)
✅ B74	HDF5ImageIO::ReadImageInformation	every geometry dataset's length is taken on trust — spacing[i] / Dims[i] are indexed to numDims regardless of the file's actual dataset length, reading past the end of a heap vector before any pixel is touched — fix: #6626 (merged)
✅ B76	JPEGImageIO::WriteSlice	the "prefer centimetres" density branch tags the JFIF header density_unit = 0 ("aspect ratio only") while storing real dots-per-cm values — the reader's own density_unit > 0 guard then discards them, so the spacing is lost on a round-trip through ITK — fix: #6627 (merged)
✅ B77	JPEGImageIO::Read	the CMYK→RGB conversion assumes all four channels are inverted; that holds for plain CMYK, but for a YCCK source libjpeg un-inverts C/M/Y and passes K through — a white YCCK pixel decodes to black — fix: #6627 (merged)
✅ B78	TIFFImageIO::ReadVolume	each kept page is written at its raw directory index while the volume depth is sized from m_SubFiles — one ignored or untagged page before a SUBFILETYPE == 0 page and the last slice is written past the end of the buffer — fix: #6636 (merged)
✅ B79	TIFFImageIO::ReadGenericImage	the scanline buffer is sized from the current directory, but width/SamplesPerPixel/BitsPerSample stay pinned to directory 0 — a multi-page file whose later pages are narrower over-reads the buffer — fix: #6636 (merged)
✅ B80	TIFFImageIO::ReadImageInformation	the 32-bit arm of the component-type ladder has no default:, and CanRead never constrains SampleFormat — a 32-bit SAMPLEFORMAT_VOID page is read as whatever m_ComponentType already held — fix: #6636 (merged)
✅ B81	TIFFImageIO::ReadGenericImage	PHOTOMETRIC_MINISWHITE is read out un-inverted — the // check inverted comment above the plain-copy grayscale arm never grew a body — fix: #6636 (merged)
✅ B82	TIFFImageIO::GetFormat	MINISBLACK/MINISWHITE map to GRAYSCALE regardless of SamplesPerPixel, so a 2-sample grey+alpha page is declared 1-component and each row keeps only its first half — fix: #6636 (merged)
✅ B83	TIFFImageIO::InternalWrite	PHOTOMETRIC_RGB is written for every component count other than 1 — a 2-component image becomes an RGB-tagged file with SamplesPerPixel = 2 and no EXTRASAMPLES — fix: #6636 (merged)
✅ B84	MultiLabelSTAPLEImageFilter	label loop counters are typed InputPixelType but bounded by m_TotalLabelCount — a uint8 input using all 256 labels wraps the counter 255 → 0 and InitializePriorProbabilities() never terminates (reproduced as a live hang, not source analysis) — fix: #6607 (merged)
✅ B85	MultiLabelSTAPLEImageFilter::InitializeConfusionMatrixArrayFromVoting	the internal seeding LabelVotingImageFilter never receives the outer filter's SetLabelForUndecidedPixels(), so the seed pass always uses automatic label selection regardless of the caller's setting — fix: #6607 (merged)
✅ B86	GiplImageIO::ReadImageInformation / CanReadFile	the 256-byte header is consumed as 15 separate gzread/istream::read pairs, none of whose byte counts is checked — a truncated or non-GIPL file populates dimensions, spacing and origin from uninitialized locals (same shape as B54/B55, but no success flag is computed at all) — fix: #6628 (merged)
✅ B87	TileImageFilter	A 0 in any non-last Layout axis is never validated: with a non-zero last axis it indexes an empty std::vector and runs a ~2³² iteration loop; with a zero last axis (the constructor default) it divides by zero — fix: #6630 (merged)
✅ B88	GDCMImageIO::Read (SINGLEBIT)	The 1-bit expansion walks the buffer as an unpadded bitstream, but GDCM pads each row to a byte boundary — every row after the first is decoded from the wrong bits when the width is not a multiple of 8 — fix: #6639 (merged), replacing the reverted #6629
✅ B89	GDCMImageIO::InternalReadImageInformation (ultrasound/hardcopy arm)	A single-valued (0028,0030) leaves the second Element slot never written — axis-0 spacing is read from indeterminate heap memory; the assert(GetLength() == 2) guard is vacuous and cannot fire even in a Debug build — fix: #6639 (merged), replacing the reverted #6629
✅ B90	LabelSetErodeImageFilter / LabelSetDilateImageFilter	GenerateData only calls AllocateOutputs() and every write is gated on m_Scale[d] > 0; under UseImageSpacing a zero radius on the last axis skips erosion's only output-writing pass — the filter returns uninitialized memory — fix: #6631 (merged)
✅ B92	ImageSeriesReader::GenerateOutputInformation	An ITK_ImageOrigin override array shorter than the output dimension is silently shrunk by Array::operator=, and the next loop still indexes to ImageDimension - 1 — unchecked out-of-bounds read. No in-tree ImageIO writes that key, so it is live only for an out-of-tree/custom ImageIO — fix: #6632 (merged)

✅ Completed — PR-by-PR triage record (51 settled rows)

#6629 (B88/B89, GDCM) was reverted by #6638 — its single-bit decode assumed row-padded 1-bit Pixel Data, but DICOM PS3.5 packs contiguously — and is now replaced by #6639, which decodes contiguously (verified pixel-for-pixel against pydicom on a real 13×12×8 highdicom SEG file) and validates PixelSpacing multiplicity via gdcm::VM::GetNumberOfElementsFromArray.

Every @physwkim PR carries a single commit; each regression test was verified locally to fail on the pre-fix code and pass on the fix.

Ledger item	PR	Author	State
B10 (MultiLabelSTAPLE tie increment)	#6579	@physwkim	Merged
B24 (FastMarchingUpwindGradient AllTargets duplicates)	#6658	@hjmjohnson	Merged
B67–B69, Q84 (vendored NrrdIO via ISC teem fork snapshot)	#6674	@hjmjohnson	Merged
B18 (FastApproximateRank last axis)	#6580	@physwkim	Merged
B19 (AttributeMorphologyBase sort range)	#6581	@physwkim	Merged
B25 (ApplyToImageMetadata null inverse)	#6583	@dzenanz	Merged
B31 (IterativeInverseDisplacementField stale error bar)	#6576	@physwkim	Merged
B32 (InverseDisplacementField subsampler direction)	#6577	@physwkim	Merged; adjacent geometry gaps split to #6582, closed by #6588
B32 follow-up (lattice centering, output-direction API, subsampling guards)	#6588	@hjmjohnson	Merged — closed #6582
B2, B3, B4, B9, B11, B12, B20, B23, B27, B29 (NaN / UB-cast guards)	#6589	@hjmjohnson	Merged
B40–B45 (LabelMap / label-statistics correctness)	#6590	@hjmjohnson	Merged
B1 (ProjectionImageFilter collapsed-axis origin)	#6594	@physwkim	Merged — baselines uploaded
B13 (ThresholdMaximumConnectedComponents bisection seed)	#6595	@physwkim	Merged
B14 (ScalarImageKmeans non-contiguous label interval)	#6596	@physwkim	Merged
B15 (LabelVoting / MultiLabelSTAPLE undecided-label wrap)	#6597	@physwkim	Merged
B16 (MaskedFFTNormalizedCorrelation integer squaring)	#6598	@physwkim	Merged
B17 (IterativeDeconvolution ignores OutputRegionMode)	#6599	@physwkim	Merged
B6 (GridImageSource ignores the image origin)	#6600	@physwkim	Merged
B7 (MinMaxCurvatureFlow 3-D polar angle)	#6603	@physwkim	Merged
B8 (CurvatureFlowFunction derivative scale vs radius)	#6604	@physwkim	Merged
B21 (LevelSetNeighborhoodExtractor zero contour)	#6601	@physwkim	Merged
B22 (StructureTensor single-axis smoothing)	#6602	@physwkim	Merged — the 16 AnisotropicDiffusionLBR baselines were regenerated and uploaded
B28 (UniformRandomSpatialNeighborSubsampler infinite Search)	#6605	@physwkim	Merged
B30 (RegionBasedLevelSetFunction reinitialization term)	#6606	@physwkim	Merged
B36, B37, B38 (FastMarchingImageFilterBase boundary / topology / zero speed)	#6613	@physwkim	Merged
B39 (colormap NumericTraits::min as a most-negative seed)	#6615	@physwkim	Merged
B46 (NormalizedCorrelationImageFilter unguarded divisions)	#6612	@physwkim	Merged
B47 (TransformIOBase CorrectTransformPrecisionType)	#6608	@physwkim	Merged
B53–B56 (GiplImageIO write truncation, short-read detection, vector header)	#6611	@physwkim	Merged
B57–B60, B62 (VTKImageIO short read, sscanf, attribute dispatch, ASCII precision, CanReadFile)	#6610	@physwkim	Merged
B61 (ImageIOBase::ReadBuffer ASCII stream state)	#6609	@physwkim	Merged
B70, B72 (PNGImageIO truncates the file before validating)	#6614	@physwkim	Merged
B71 (PNGImageIO writes only the first slice)	—	@physwkim	Withdrawn — not a defect, the contract is documented and tested
B84, B85 (MultiLabelSTAPLE counter wrap, undecided label not propagated)	#6607	@physwkim	Merged
B48–B50 (NIfTI multi-component rescale over-read)	#6625	@physwkim	Merged
B73, B74 (HDF5 null metadata cast, unchecked geometry lengths)	#6626	@physwkim	Merged
B76, B77 (JPEG centimetre density tag, YCCK conversion)	#6627	@physwkim	Merged
B86 (GIPL header read in one checked call)	#6628	@physwkim	Merged
B88, B89 (GDCM single-bit packing, single-valued PixelSpacing)	#6629 (reverted by #6638), replaced by #6639	@physwkim, @hjmjohnson	Merged — #6639 decodes 1-bit Pixel Data contiguously per PS3.5 (including the trailing partial byte, which the pre-#6629 code dropped) and counts PixelSpacing values with GDCM's own multiplicity counter, closing the trailing-backslash gap. Verified against pydicom on a real multi-frame SEG file.
B87 (TileImageFilter zero Layout axis)	#6630	@physwkim	Merged
B90 (LabelSet erode/dilate uninitialized output)	#6631	@physwkim	Merged
B92 (ImageSeriesReader origin-override out-of-bounds)	#6632	@physwkim	Merged
B26 (PatchBasedDenoising POISSON step collapse)	#6633	@physwkim	Merged
B78–B83 (TIFF volume packing, per-page geometry, SampleFormat, MINISWHITE, grey+alpha, 2-component write)	#6636	@physwkim	Merged
B39 family (NumericTraits::min as a most-negative seed, 5 further sites)	#6634	@physwkim	Merged
B39 family (entropy threshold calculators)	#6635	@physwkim	Merged — expected constants and baselines updated
B51 (NIfTI 2-D vector rank), B71 (PNG first slice), B75 (HDF5 memory type)	—	@physwkim	Withdrawn — not defects, see the bug table
B5 (CannySegmentationLevelSetFunction speed image)	—	@physwkim	Withdrawn — not a defect, see comment below
B33 (ConfidenceConnected empty re-estimation)	#6578	@physwkim	Merged
Precursors	#6569 (JHMI marginals + derivative)	@physwkim	Merged
	#6572 (ScaleLogarithmicTransform Jacobian)	@physwkim	Merged

❌ Withdrawn — 4 items investigated and found not to be defects (table)

#	Site	One-line symptom
❌ B5	CannySegmentationLevelSetFunction	Propagation weight 0 → advection generated from a never-allocated speed image — not a defect; withdrawn (UpdateOutputInformation resets the empty requested region)
❌ B51	NiftiImageIO	spatial-dimension inference for intent-vector images never consults dim[1] — a 2-D vector image one row tall reads back as 1-D — not a defect; withdrawn: the write side collapses trailing degenerate dimensions too, so a 2-D 5x1 vector image and a 1-D 5 vector image produce byte-identical dim[]; the rank is lost at write time and no read-side change can recover it
❌ B71	PNGImageIO::WriteSlice	height comes from GetDimensions(1) alone, no axis beyond it is consulted — a 3-D image writes only its first slice, with no error — not a defect; withdrawn: writing the first slice of an image of 3 or more dimensions is the documented, tested contract (Modules/IO/PNG/test/itkPNGImageIOTest.cxx:152-160)
❌ B75	HDF5ImageIO::Read	the file datatype is passed as H5Dread's memory type, making the conversion the identity — a VoxelData stored STD_I16BE is copied without a byte swap and read back byte-swapped, silently — not a defect; withdrawn: ReadImageInformation() requires an exact H5Tequal match against a NATIVE_* predtype and throws otherwise, and it is what populates m_VoxelDataSet — so Read() is unreachable with a foreign-endian on-disk type (verified: writing STD_I16BE throws during ReadImageInformation())

How these were found, and what was checked before filing

Each filter was reimplemented from the .hxx sources and the two implementations compared on synthetic inputs; every divergence was traced back to the ITK code path that caused it. Before filing, every item was re-checked against main at e46eb723a5: the cited lines re-read in context, declared types and constructor defaults verified, and each claimed consequence re-derived from the quoted code. Items whose original analysis did not survive that re-check were corrected or dropped — some of my own earlier conclusions turned out to be wrong, so I would not be surprised if a few of the remaining ones are too. Corrections welcome.

One item is flagged inline as source-analysis only (B28: non-termination identified from the loop structure, not run to a live hang).

AI assistance

• Tool: Claude Code (claude-opus-4-8)
• Role: found the defects while cross-checking an independent reimplementation of these filters against this source; performed the pre-filing re-verification pass (re-reading every cited line on main at e46eb723a5, checking declared types and constructor defaults, re-deriving each consequence)
• Contribution: the inventory above
• Testing: source-level verification against e46eb723a5; behavioral claims cross-checked against the independent reimplementation's pinned tests. B28 is source analysis only (not run to a live hang) and is labeled as such inline.
• Items whose original analysis did not survive re-verification were corrected or dropped before filing.

[github-actions]

Thank you for contributing an issue! 🙏

Welcome to the ITK community! 🤗👋☀️

We are glad you are here and appreciate your contribution. Please keep in mind our community participation guidelines. 📜
Also, please check existing open issues and consider discussion on the ITK Discourse. 📖

This is an automatic message. Allow for time for the ITK community to be able to read the issue and comment on it.

[physwkim]

Following up on this issue: continuing the porting work, I have a second batch of findings, all in the displacement-field / warp / vector-filter area. Same methodology and caveats as the original report — source-level reading against master e46eb723a5, independently re-verified before posting, no runtime reproduction. Numbering continues from the original lists (B31–B34, Q25–Q29).

Bugs

#	Filter / site	Finding
B31	IterativeInverseDisplacementFieldImageFilter	smallestError carries over from the previous pixel when the mapped point starts outside the buffer — output depends on raster order
B32	InverseDisplacementFieldImageFilter	the field-subsampling resampler never gets the input's direction — with non-identity direction cosines the spline landmarks are not on the input lattice at all
B33	VectorConfidenceConnectedImageFilter	a re-estimation pass that visits zero pixels produces NaN statistics, which invert the Mahalanobis guard — the next flood includes every pixel in the image
B34	WarpImageFilter	the same-information fast-path test omits size, so a valid larger-output configuration dies in an iterator containment check instead of taking the general path

Quirks / doc

#	Site	Finding
Q25	IterativeInverseDisplacementFieldImageFilter	the coordinate-descent probe step is axis 0's spacing on every axis
Q26	DisplacementFieldJacobianDeterminantFilter	SetDerivativeWeights silently flips UseImageSpacing off — precedence between the two is last-writer-wins, while the class doc describes only one ordering
Q27	WarpImageFilter	m_OutputSize[0] == 0 alone keys "size unset"; output-grid defaults are unit/zero/identity from neither input; the field is sampled with edge replication, never out-of-domain
Q28	VectorConnectedComponentImageFilter	zero vectors join anything under the default threshold (no special case); concept macro name typo InputValyeTypeIsFloatingCheck
Q29	MahalanobisDistanceMembershipFunction	the det < 0 exception branch is unreachable; GetCovariance() never reflects the singular-case substitute

Details below.

B31 — IterativeInverseDisplacementField: stale smallestError makes output raster-order dependent

itkIterativeInverseDisplacementFieldImageFilter.hxx:96 declares the residual bar outside the per-pixel loop:

double smallestError = 0;
InputIt.GoToBegin();
OutputIt.GoToBegin();
while (!OutputIt.IsAtEnd())

and it is (re)assigned only inside the inside-buffer branch (:122-132):

if (inputFieldInterpolator->IsInsideBuffer(mappedPoint))
{
  ...
  smallestError = 0;
  ...
  smallestError = std::sqrt(smallestError);
}

Nothing at the top of the pixel iteration resets it. For a pixel whose initial mapped point falls outside the buffer, the branch is skipped and smallestError still holds the previous pixel's residual. The coordinate-descent probes are individually buffer-guarded (:147, :167), so a probe that lands inside the buffer is compared against that stale bar (:156 if (tmp < smallestError)) and accepted or rejected based on how well the previous pixel converged — the inverse field's values at such pixels depend on raster order.

B32 — InverseDisplacementField: subsampling resampler never gets the input's direction

itkInverseDisplacementFieldImageFilter.hxx, PrepareKernelBaseSpline configures the field-subsampling resampler at :119-141:

resampler->SetInput(inputImage);
resampler->SetOutputOrigin(inputImage->GetOrigin());
...
resampler->SetSize(size);
resampler->SetOutputStartIndex(subsampledRegion.GetIndex());
resampler->SetOutputSpacing(spacing);

SetOutputDirection is never called, and ResampleImageFilter's constructor leaves it identity (itkResampleImageFilter.hxx:51 m_OutputDirection.SetIdentity();; no reference image is set, so nothing overrides it). For an input field with non-identity direction cosines two things go wrong at once:

1. the resampler samples along an axis-aligned physical grid; interior sample points can fall outside the input buffer and silently take the default pixel value (zero, :65, applied at :411);
2. the landmark read-back at :168 (sampledInput->TransformIndexToPhysicalPoint(ot.GetIndex(), targetPoint)) uses the subsampled image's (identity) direction, so the target landmarks handed to the kernel transform are those axis-aligned points — not a subset of the input field's lattice.

The kernel-spline inverse is then anchored on landmarks that don't correspond to the field being inverted.

B33 — VectorConfidenceConnected: empty re-estimation pass floods the whole image via NaN statistics

itkVectorConfidenceConnectedImageFilter.hxx: the re-estimation loop recomputes mean/covariance from the previous segmentation with no guard on the visited-pixel count (SizeValueType num{}; at :264):

mean[ii] /= static_cast<double>(num);              // :289
covariance[ii][jj] /= static_cast<double>(num);    // :292

(The initial statistics do guard the equivalent case at :179-184; the loop has no such guard.) The empty case is reachable: the threshold is fixed before the loop (SetThreshold(m_Multiplier), :198/:226; the seed-inclusion boost :209-222 runs pre-loop only), while mean and covariance are recomputed each round (:304-305) — the acceptance interval moves, so a round can exclude even the seeds; the flood loop then simply never executes (:319-323 over an all-background image filled at :314) and the next pass has num == 0.

With num == 0, mean and covariance are 0.0/0.0 = NaN. In MahalanobisDistanceThresholdImageFunction the NaN squared distance then fails the guard at itkMahalanobisDistanceThresholdImageFunction.hxx:113:

const double mahalanobisDistance = (mahalanobisDistanceSquared > 0.0) ? std::sqrt(...) : 0.0;

so the distance clamps to 0.0, and 0.0 <= m_Threshold (:87) is true for any positive threshold — the next flood accepts every pixel in the image. (Mechanism note: the NaN reaches the distance chiefly through the NaN mean — a NaN determinant fails det > singularThreshold in itkMahalanobisDistanceMembershipFunction.hxx:107 and routes to the singular branch, which substitutes a finite inverse; the measurement − NaN-mean differences carry the NaN regardless.)

B34 — WarpImageFilter: same-information fast-path test omits size, so a valid configuration throws

itkWarpImageFilter.hxx:387-391:

this->m_DefFieldSameInformation =
  (outputPtr->GetOrigin()   ... is_equal(fieldPtr->GetOrigin(),   ...)) &&
  (outputPtr->GetSpacing()  ... is_equal(fieldPtr->GetSpacing(),  ...)) &&
  (outputPtr->GetDirection()... is_equal(fieldPtr->GetDirection(), ...));

Size is not part of the test, but the fast path iterates the field over the output's thread region (:294 ImageRegionConstIterator fieldIt(fieldPtr, outputRegionForThread);). The output can legitimately be larger than the field — :428-438 takes the output region from the user's SetOutputSize/SetOutputStartIndex when set — and GenerateInputRequestedRegion explicitly anticipates the mismatch by falling back to the field's largest possible region (:407-410), which silences the pipeline's requested-region check.

What happens then is not a buffer over-read — the iterator's own containment assert catches it (itkImageConstIterator.h:207-212, itkAssertOrThrowMacro(bufferedRegion.IsInside(m_Region)), which throws in release builds and aborts in debug). But the user-visible result is that a configuration the general path handles fine (it clamps field sampling to the field domain, :200-216) instead dies with an internal iterator error ("Region ... is outside of buffered region") whenever the output grid shares origin/spacing/direction with a smaller field. Adding a size/region comparison to the fast-path test would route this case to the general path.

Q25 — IterativeInverseDisplacementField: probe step is axis 0's spacing on every axis

itkIterativeInverseDisplacementFieldImageFilter.hxx:89 is the only per-axis spacing read in GenerateData:

const double spacing = inputPtr->GetSpacing()[0];

and the coordinate-descent loop applies it along every physical axis (:144-146 mappedPoint[k] += step;, :166, :186). For anisotropic fields every axis but the first is probed with the wrong step size.

Q26 — DisplacementFieldJacobianDeterminant: SetDerivativeWeights silently flips UseImageSpacing off

itkDisplacementFieldJacobianDeterminantFilter.hxx:55: SetDerivativeWeights begins with m_UseImageSpacing = false; (unconditionally, not via SetUseImageSpacing). The effective precedence between the two parameters is therefore last-writer-wins: SetUseImageSpacing(true) after SetDerivativeWeights re-derives weights from spacing at update time (:155-168), while the reverse order keeps the explicit weights and drops the spacing request. The class doc (.h:78-82) describes only the first ordering ("if UseImageSpacing is set to TRUE (ON), then these weights will be overridden by weights derived from the image spacing when the filter is updated").

Downstream note: SimpleITK's generated wrapper sets members in yaml order — UseImageSpacing before DerivativeWeights — so a SimpleITK user who sets both gets the UseImageSpacing(true) request silently dropped.

Q27 — WarpImageFilter parameter quirks

Three small ones, all in itkWarpImageFilter.hxx:

1. :428 — if (this->m_OutputSize[0] == 0 && fieldPtr != nullptr) keys "size unset" on element 0 alone; an explicit SetOutputSize({0, 5, 5}) silently discards the 5s and inherits the field's region.
2. :43-48 — constructor defaults are unit spacing, zero origin, identity direction, zero size: borrowed from neither the input image nor the field (asymmetric with ResampleImageFilter's reference-image convention).
3. :200-216 — displacement-field sampling clamps baseIndex/neighIndex into [m_StartIndex, m_EndIndex], i.e. edge-replicates, and never goes out-of-domain; only the input-image probe (m_Interpolator->IsInsideBuffer, :312/:342) can send a pixel to EdgePaddingValue. Worth a doc note since the two lookups follow different out-of-domain policies.

Q28 — VectorConnectedComponent: zero vectors, and a concept-macro typo

itkVectorConnectedComponentImageFilter.h:77-87: the functor is an unnormalized dot product (documented — "Assumes that vectors are normalized", and Math::Absolute deliberately makes antiparallel vectors similar per .h:100-101). A zero vector, however, has no documented story: it yields dotProduct = 0 against anything, so 1 - |0| = 1, and with the wrapped default DistanceThreshold = 1 it joins any neighbor — zero-padding or masked-out regions merge with whatever they touch.

Trivial: .h:145 — itkConceptMacro(InputValyeTypeIsFloatingCheck, ...) misspells "Value" as "Valye".

Q29 — MahalanobisDistanceMembershipFunction: dead det check; GetCovariance hides the substitute

itkMahalanobisDistanceMembershipFunction.hxx:

1. :98-103 — const double det = inv_cov.determinant_magnitude(); followed by if (det < 0.) { itkExceptionStringMacro(...) }. determinant_magnitude() is a product of SVD singular values, hence non-negative; the branch is unreachable (a NaN det also fails < 0.).
2. :92 caches the raw covariance; the singular-case substitution (:114-124) replaces only m_InverseCovariance with aLargeDouble·I. GetCovariance() (a GetConstReferenceMacro) therefore always reports the literal input matrix, with no indication the function is actually operating with the substituted inverse.

(Same AI-assistance disclosure as the issue body applies to this comment.)

[physwkim]

A third batch from the continuing port, this time in the demons-registration / displacement-field-exponential area. Same methodology and caveats as before — source-level reading against master e46eb723a5, independently re-verified before posting, no runtime reproduction. Numbering continues (B35, Q30–Q37).

Bug (latent, explicit-configuration-only)

#	Filter / site	Finding
B35	ExponentialDisplacementFieldImageFilter	the field is divided by 1 << numiter — UB when numiter >= 31; unreachable by default (cap 20, clamped), but an explicit SetMaximumNumberOfIterations(>= 31) makes it live

Quirks / doc

#	Site	Finding
Q30	FastSymmetricForcesDemonsRegistrationFilter	its difference function is ESMDemonsRegistrationFunction; the identically-named FastSymmetricForcesDemonsRegistrationFunction class is orphaned from its own filter
Q31	ExponentialDisplacementFieldImageFilter	"take the ceil" comment vs + 1.0-then-truncate code (integral values overshoot by 1); a zero field still runs one squaring iteration
Q32	WarpVectorImageFilter (as wired by DiffeomorphicDemonsRegistrationFilter)	under the nearest-neighbor-extrapolating interpolator, IsInsideBuffer is always true and m_EdgePaddingValue is dead code
Q33	ESMDemonsRegistrationFunction	the WarpedMoving and MappedMoving gradient options coincide exactly at a zero displacement field (interior pixels, default linear interpolation, coincident grids)
Q34	LevelSetMotionRegistrationFilter/Function	dt = 1/maxL1 makes the applied step at the governing pixel exactly Alpha-independent; thresholded pixels count toward the metric while out-of-buffer pixels don't; Halt adds an RMS==0 clause the base can never fire
Q35	LevelSetMotionRegistrationFunction	one-sided difference steps add index-axis spacing to a physical coordinate — wrong direction and step under non-identity direction cosines
Q36	demons family	zero-iteration runs report GetRMSChange() == 0 but GetMetric() == NumericTraits<double>::max() in the two filters that don't override GetRMSChange
Q37	LevelSetMotionRegistrationFunction smoothing	the smoothed moving image is re-quantized to the moving image's own pixel type (u8 in → u8 smoothed) before the gradient; and any axis shorter than 4 pixels is a hard error for this family member only

Details below.

B35 — ExponentialDisplacementField: signed shift UB behind SetMaximumNumberOfIterations ≥ 31

itkExponentialDisplacementFieldImageFilter.hxx:160 (and the inverse at :164):

m_Divider->SetInput2(static_cast<InputPixelRealValueType>(1 << numiter));

numiter itself is unsigned int (:70); the UB kernel is the signed literal 1 — the shift's result type is int, so 1 << 31 is undefined. To be clear about reachability: this is not hittable by default. The constructor caps m_MaximumNumberOfIterations at 20 (:32), and the automatic count is hard-clamped:

numiter = std::min(static_cast<unsigned int>(numiterfloat + 1.0), m_MaximumNumberOfIterations);  // :112

so no input field, however extreme, raises it past the cap. It becomes live only when a user explicitly calls SetMaximumNumberOfIterations with ≥ 31 (clamp target in automatic mode; taken verbatim in non-automatic mode, :121). 1u << numiter or computing the power in floating point would close it.

Q30 — FastSymmetricForcesDemonsRegistrationFunction is orphaned from its own filter

itkFastSymmetricForcesDemonsRegistrationFilter.h:120-121:

using DemonsRegistrationFunctionType =
  ESMDemonsRegistrationFunction<FixedImageType, MovingImageType, DisplacementFieldType>;

with the in-source comment at :117: FIXME: Why is this the only permissible function ?. Meanwhile the identically-named FastSymmetricForcesDemonsRegistrationFunction (itkFastSymmetricForcesDemonsRegistrationFunction.h:47) is a complete, instantiable class that its own filter never uses. It is not the default force function of any filter either (CurvatureRegistrationFilter defaults to MeanSquareRegistrationFunction, itkCurvatureRegistrationFilter.h:96-99); its only uses anywhere in the tree are as an explicit CurvatureRegistrationFilter template argument in two tests (itkCurvatureRegistrationFilterTest.cxx:114, RegistrationITKv3/DeformableRegistration10.cxx:60). Worth either wiring it back to its filter, documenting the ESM substitution, or deprecating the orphan.

Q31 — ExponentialDisplacementField iteration count: not-quite-ceil, and the zero-field case

itkExponentialDisplacementFieldImageFilter.hxx:111-112:

// take the ceil and threshold
numiter = std::min(static_cast<unsigned int>(numiterfloat + 1.0), m_MaximumNumberOfIterations);

For an exactly-integral numiterfloat == 2.0 this yields 3 where ceil would give 2 — a harmless extra squaring, but not what the comment says.

Zero displacement field: maxnorm2 == 0, so the ternary at :106-107 assigns numiterfloat = NumericTraits<...>::min() (the smallest positive normal), which passes numiterfloat >= 0.0 (:109) and truncates to numiter = 1 — the identity field is still divided by 2, squared once, and composed. Correct result, one wasted composition pass.

Q32 — WarpVectorImageFilter: EdgePaddingValue is dead under the extrapolating interpolator

All three IsInsideBuffer overloads of VectorLinearInterpolateNearestNeighborExtrapolateImageFunction return true unconditionally (.h:99-118), and EvaluateAtContinuousIndex clamps the base index and zeroes the out-of-range fractional distance (.hxx:51-67) — out-of-buffer points extrapolate the nearest pixel. WarpVectorImageFilter's per-pixel test (itkWarpVectorImageFilter.hxx:171-186):

if (m_Interpolator->IsInsideBuffer(point)) { ... } else { outputIt.Set(m_EdgePaddingValue); }

is then always-true, and the m_EdgePaddingValue branch is unreachable. The plain filter's default interpolator (VectorLinearInterpolateImageFunction, .h:141) does real bounds checks, so this is configuration-dependent — but DiffeomorphicDemonsRegistrationFilter installs exactly the extrapolating one (itkDiffeomorphicDemonsRegistrationFilter.h:196-199, ctor wiring .hxx:38-39), so for the demons field warper the edge-padding parameter silently does nothing.

Q33 — ESMDemonsRegistrationFunction: WarpedMoving and MappedMoving gradients coincide at a zero field

WarpedMoving hand-computes an index-space central difference over the warped moving image (itkESMDemonsRegistrationFunction.hxx:294-296 — deliberately not via CentralDifferenceImageFunction, per the comment at :206-207), while MappedMoving calls CentralDifferenceImageFunction::Evaluate, which samples the interpolant at p ± 0.5·spacing in physical space (itkCentralDifferenceImageFunction.hxx:264-300). Under linear interpolation, M(p + 0.5s) − M(p − 0.5s) = ½(m[i+1] − m[i−1]) at a pixel centre, so divided by s the two are exactly equal at a zero displacement field — for interior pixels, default linear interpolation on both paths, and coincident fixed/moving grids (the boundary rules and spacing divisors differ). Users choosing between the two enum options may expect a difference the defaults don't deliver; a doc note would cover it.

Q34 — LevelSetMotion: time-step consequences the docs don't state

The only family member whose global time step is data-dependent: dt = 1/m_MaxL1Norm (itkLevelSetMotionRegistrationFunction.hxx:352-367; the other four functions all return a constant m_TimeStep = 1.0). Three consequences:

1. Per-pixel update is speed * gradient[j] / (gradientMagnitude + m_Alpha) (:330) and the L1 norm feeds m_MaxL1Norm (:339, :345-348). At the pixel that attains the max, with one dominant gradient direction, dt · update[k] = ±mSpacing[k] — the (|g| + Alpha) factor cancels exactly, so at the governing pixel the applied displacement is independent of Alpha (other pixels carry the ratio (|g_max|+α)/(|g_p|+α), so it is approximate field-wide). The Alpha doc ("small positive constant to avoid division by zero") does not mention that the global step rescales its attenuation back out.
2. A thresholded pixel still increments SumOfSquaredDifference/NumberOfPixelsProcessed (:313-325 — counted before the guards zero the update), while an out-of-buffer pixel returns before counting (:220-228) — the reported metric mixes the two exclusion policies.
3. LevelSetMotionRegistrationFilter::Halt (.hxx:76-89) adds m_RMSChange == 0.0 && GetElapsedIterations() != 0; the base's convergence test is strictly GetMaximumRMSError() > m_RMSChange (itkFiniteDifferenceImageFilter.hxx:225), which at the default MaximumRMSError == 0.0 can never fire on 0.0 > 0.0 — so this subclass converges on an exactly-zero iteration where every sibling keeps iterating to the count limit.

Q35 — LevelSetMotion: one-sided differences step a physical coordinate by index-axis spacing

itkLevelSetMotionRegistrationFunction.hxx:253 (loop :251-277):

PointType mPoint(mappedPoint);          // physical point
...
mPoint[j] += mSpacing[j];               // index-axis j's spacing added to physical coordinate j

mappedPoint comes from TransformIndexToPhysicalPoint plus the displacement (:213), and mSpacing is the moving image's spacing (:240-245). With a non-identity direction matrix, physical axis j is not index axis j — the probe walks the wrong direction by the wrong step. (Same borrowed-axis pattern as CentralDifferenceImageFunction::Evaluate's point[dim] ± 0.5 * spacing[dim], so it may be a known family-wide convention — flagging it because here it also interacts with UseImageSpacing, which selects both this step and the L1-norm divisor of the time step, :339. The demons/ESM functions have no such toggle; they fold spacing in unconditionally via m_Normalizer, itkDemonsRegistrationFunction.hxx:122-127.)

Q36 — Demons family: zero-iteration metric reporting is inconsistent

SymmetricForcesDemonsRegistrationFilter (.h:118), FastSymmetricForcesDemonsRegistrationFilter (.h:113), and DiffeomorphicDemonsRegistrationFilter (.h:125) override GetRMSChange() to forward to the function; DemonsRegistrationFilter and LevelSetMotionRegistrationFilter do not. For the latter two, a run with zero elapsed iterations reports the base's value-initialized m_RMSChange == 0.0 (itkFiniteDifferenceImageFilter.h:361) — "converged" — while GetMetric() forwards (itkLevelSetMotionRegistrationFilter.hxx:91-103) to the function's constructor sentinel m_Metric = NumericTraits<double>::max() (itkLevelSetMotionRegistrationFunction.hxx:50). Two query methods on the same freshly-run filter thus disagree about whether anything happened.

Q37 — LevelSetMotion smoothing: re-quantization, and a 4-pixel minimum unique in the family

itkLevelSetMotionRegistrationFunction.h:116 instantiates the smoother with one template argument:

using MovingImageSmoothingFilterType = SmoothingRecursiveGaussianImageFilter<MovingImageType>;

so TOutputImage = TInputImage (itkSmoothingRecursiveGaussianImageFilter.h:50): the filter computes internally in a real-typed image but casts back (CastingFilterType, .h:91) — a u8 moving image is smoothed and then re-quantized to u8 before the gradient interpolator reads it (.hxx:181, :185). And through RecursiveSeparableImageFilter it inherits the hard error on any filtered axis shorter than 4 pixels (itkRecursiveSeparableImageFilter.hxx:193-201, in BeforeThreadedGenerateData) — a 3-pixel-thick volume is a throw for LevelSetMotionRegistrationFilter and for no other member of the family (the others take gradients via CentralDifferenceImageFunction, no minimum size).

(Same AI-assistance disclosure as the issue body applies to this comment.)