fix: state whether the second channel is labels instead of guessing

[Hendrik-code]

The defect

if img.shape[0] == 2:
    subject = transform(tio.Subject(
        image=tio.ScalarImage(tensor=img[0:1]),
        discs=tio.LabelMap(tensor=img[1:2]),   # <- decided by channel count alone
        seg=tio.LabelMap(tensor=seg)))

The totalspineseg "step 2" layout really does carry an odd-disc segmentation in channel 1. But a genuine two-modality input — T1+T2, in-phase/out-of-phase — has exactly the same shape, and got its second modality registered as a tio.LabelMap. torchio then resamples it nearest-neighbour and skips every intensity artifact on it, so the two modalities diverge in both interpolation and augmentation while the caller believes both were augmented together.

Channel count cannot distinguish the two cases; nothing in the tensor can.

The fix

second_channel_is_labels, threaded from ArtifactTransform / SpatialCustomTransform through apply_enabled into apply_tio, where a config can set it.

It defaults to True — the long-standing behaviour — so the step 2 pipeline is unchanged and test_the_second_channel_is_treated_as_labels passes untouched. A multi-modality caller now has a way to say so, which was previously impossible to express.

all_augmentations.json is regenerated for the new parameter.

Tests

Added to unit_tests/test_torchio_ops.py: the default still treats channel 1 as labels; second_channel_is_labels=False lets an intensity artifact reach it (the case that used to be inexpressible); the default leaves it alone under the same call, as the control; and a single-channel image is unaffected either way.

Verified failing before the fix (3 failed) and passing after. Full suite unchanged against baseline; ruff check . clean.

🤖 Generated with Claude Code