[demianw]

What was wrong

Several tract_math operations crashed with the supported dependencies (NumPy 2.5.4, SciPy 1.18.1, NiBabel 5.4.2 here). I ran each operation listed in #105 on a 100-tract subset of the test .trk file and the test atlas, calling the registered operations directly (tract_querier.tract_math.operations[...]) so the real exceptions were visible.

Listed in #105, all reproduced:

operation	error on master
tract_kappa	UnboundLocalError: cannot access local variable 'numpy'
tract_kappa_volume	UFuncTypeError (image compared with the threshold string)
tract_tract_confidence	TypeError: histogram() got an unexpected keyword argument 'normed'
tract_prototype_median, tract_prototype_mean	ImportError: cannot import name 'in1d' from 'numpy'
tract_prototype_mean	'0' > 0 compares str with int
tract_bhattacharyya_coefficient	TypeError: bins must be an integer; mn_ used where mx_ is meant
tract_flip_endpoints_in_label	TypeError: takes from 3 to 4 positional arguments but 5 were given; then iterates over dict keys
tract_generate_population_probability_map	returns a tuple, so no output file is written
tract_deform	and instead of or: a 4-D image whose last axis is not 3 passes the check and fails later with IndexError

Found while reproducing (each reproduced the same way):

• tract_prototype_*: after the in1d fix, AttributeError: 'numpy.ndarray' object has no attribute 'ptp' (removed in NumPy 2) in tract_obb.py.
• tract_tract_confidence: after the normed fix, saving failed in the trk writer because the new per-point data was 1-D; it is now (n, 1) like other scalars.
• tract_generate_population_probability_map with a single input tractography: IndexError from tractographies[1].
• scalar_mean_std: TypeError: only 0-dimensional arrays can be converted to Python scalars (float() of a shape-(1,) array).
• tract_bhattacharyya_coefficient: the tract file name was appended 3 times per tract, so with two or more other tracts every row after the first was labelled with the wrong file.
• tract_map_image with a 4-D image: iterated over the first spatial axis instead of the volumes, and stored the flat (n_points, 1) array as tract data (trk writer: Missing scalars for some points!).

The change

Smallest fix per item: in1d → isin, .ptp() → ptp(), normed → density, float(threshold), int(smooth_order), mx_, integer bins, optional_flags + .items(), no trailing comma, tractographies[-1], or, .item(). The identical kappa computation of tract_kappa and tract_kappa_volume is moved to one _kappa helper (not decorator-registered, so the operations list is unchanged). tract_map_image now reuses the 3-D branch's per-tract split for each volume image_data[..., i], and the 4-D outputs keep the input's affine/image_dims like the 3-D output does.

Behaviour changes: tract_bhattacharyya_coefficient rows after the first now show the right tract file (before, they repeated earlier names) and the bounding box now uses the true maximum (mx_) of all tracts. tract_map_image on 4-D images now writes one file per volume with the values of that volume; before it could not produce correct output.

Not fixed (maintainer decision needed)

• tract_kappa_volume still fails: tract_mask(mask, tractography) is given a NumPy array instead of an image (AttributeError: 'numpy.ndarray' object has no attribute 'affine'), and tract_mask returns an array, not the voxel set _kappa expects; its help text also omits the resolution parameter. The intended semantics (which voxel grid to compare in) are not recoverable from the code, so I left it. Hence "Part of BUG: several tract_math operations crash #105".
• The kappa normalisation in _kappa (unchanged code from both operations): N = (all_voxels.max(0) - all_voxels.min(0)).prod() undercounts the bounding-box voxels (inclusive bounds would need + 1), so nn can go negative and kappa leaves [-1, 1] on small inputs (on the synthetic test tracts I got -1.72, 2.92 and -inf). On the 100-tract test subset the value is 0.277. Fixing it changes reported kappa values, so I left it; test_tract_kappa therefore checks only the output rows, not the value.

Testing

• tests/test_tract_math.py: new smoke tests on a small synthetic tractography for tract_kappa, tract_tract_confidence, tract_prototype_median, tract_prototype_mean, tract_bhattacharyya_coefficient (rows labelled correctly), tract_flip_endpoints_in_label (points and data flipped), tract_generate_population_probability_map (single input, file written), tract_deform (non-deformation 4-D image rejected), scalar_mean_std, tract_map_image (synthetic 4-D image, per-tract data per volume). SciPy-dependent tests use pytest.importorskip('scipy'). All 10 fail on master and pass with this branch.
• python -m pytest -q -p no:cacheprovider tract_querier: 52 passed, 1 skipped (pre-existing unconditional skip in test_query_rewrite.py).
• Reproduction script re-run: all listed operations except tract_kappa_volume now run on the test data subset; tract_math a.trk tract_prototype_median out.trk and tract_math a.trk tract_kappa 2 b.trk work from the CLI.

Merge order

Needed for #123 (Tractography validating per-tract data makes the old tract_map_image 4-D branch raise). Conflicts with my upcoming PRs for #102 (length std labels in operations.py) and #107 (tract_obb.py); merge this first. #138 also appends tests to tests/test_tract_math.py (textual conflict only).

Part of #105

🤖 Generated with Claude Code