Harmonize qform/sform of inputs to avoid mislocated segmentation (#44)

[CedricConday]

Problem (#44)

A NIfTI stores two affines. The pipeline reads images with nibabel's img.affine, which prefers the sform when sform_code > 0, but greedy (registration) reads the qform. When an earlier co-registration (e.g. SPM) updates only the sform, the two disagree, so the warped/annotated mask is mislocated in FLAIR space. The reporter confirmed that setting qform = sform resolves it.

Quick demonstration of the divergence:

import numpy as np, nibabel as nib
img = nib.Nifti1Image(np.zeros((4, 4, 4), np.float32), np.eye(4))
img.set_qform(np.eye(4), code=1)
sf = np.eye(4); sf[:3, 3] = [10, -5, 3]; img.set_sform(sf, code=2)
img.affine[:3, 3]       # [10, -5, 3]  (nibabel / this pipeline: sform)
img.get_qform()[:3, 3]  # [0,  0,  0]  (greedy: qform)

Fix

Add harmonize_affines() (in utils.py) and call it once on the --t1/--flair inputs, right after the work dir is created, before any tool reads them. It sets both the qform and sform to nibabel's resolved affine, so greedy and nibabel agree. It is a no-op when the two forms already match, so existing inputs are unaffected.

Tests

Added tests/test_utils.py: one test reproducing the SPM-style mismatch (asserts qform == sform == resolved affine after harmonization), one asserting it is a no-op (data + affine unchanged) when the forms already agree. Both pass.

Note: I validated the affine harmonization in isolation; I do not have a greedy/GPU setup to run the full pipeline end-to-end, so a sanity check on a real mismatched scan would be worth confirming. The fix matches the reporter's verified qform = sform workaround.

---

Disclosure: authored with AI assistance (Claude Code); reviewed by me.

[jqmcginnis]

Hey @CedricConday, thank you very much for the PR and contribution. We will have a look asap and integrate with the next version bump!

[jqmcginnis]

Adressed in #48, closing this.

[jqmcginnis]

Closing as superseded — but merged in substance, with your authorship: the v2.0.0 line (PR #48) includes your fix ported to the PyTorch codebase, committed with you as the git author (2ecc545), extended to --existing_seg inputs, and with your tests included. It ships in the just-published v2.0.0rc1 (pip install --pre lst-ai).

One finding from validating it end to end that you'll appreciate, given your PR noted you couldn't run the full pipeline: on the v2 stack the original mislocation no longer reproduces even on a deliberately corrupted input (FLAIR qform shifted 10 mm against a correct sform) — picsl-greedy 1.4.0.1 resolves the sform the same way nibabel does, unlike the v1.x compiled greedy. So on v2 your harmonization lands as defense in depth; for the v1.x line, where #44 was reported, it remains the actual fix. Thank you for the contribution and the careful writeup!