In the TorchIO v2 pre-release, the option that sets the order of two image transforms was inverted.

Does TorchIO’s combined spatial transform apply its two steps in the order its option says?

TorchIO prepares and augments 3D medical images for training deep-learning models. Its Spatial transform combines an affine step (rotation, scaling and shift) with an elastic deformation in one resampling, and the order matters: the other order gives a different image. A training pipeline would run normally with the wrong order, and nothing in the output would flag it.

Reproduced: A reported behaviour was reproduced from a minimal, public input. Fix submitted, not yet merged: A fix was submitted to the maintainers with its evidence. It is not merged or released.

1. Compared the combined transform with the same two transforms applied one after the other, in each order, on a synthetic 3D volume.
2. Repeated the comparison when resampling to a different target grid, against a reference written independently in numpy and scipy.
3. Checked the released v1 code, which the issue had been filed against.
4. Added four test cases, then broke the fix deliberately in three ways to check that the tests notice.

• On the v2 code, affine_first=False reproduced affine-then-elastic (mean difference 0.00026). The default, affine_first=True, was 0.0049 from one order and 0.0142 from the other.
• With the fix, each setting matches its requested order: mean differences of 0.00026 and 0.00041, where the two orders differ from each other by 0.0154.
• The released v1 already used the correct order, and the swap proposed in the issue would have broken it, so the change targets v2 only.
• The four new test cases fail on the v2 code and pass with the fix. The full suite passes: 1,476 tests, 8 skipped.

The objections that would most easily undo the conclusion, and what answered each.

Is the swap proposed in the issue the right fix?

Not for the released version: v1 was already correct, and the swap would have broken it.

Could the new tests pass on the broken code?

No. All four fail on the v2 code, and each of the three deliberate breakages is caught.

Synthetic volumes on macOS. No clinical images were used, and no effect on any trained model, diagnosis or patient was measured or is claimed. Applies to the v2 pre-release, not the stable release.

Run from a TorchIO checkout of the pull request branch with its test dependencies installed. Re-run by OMEGA at this revision on 18 October 2026 (macOS): 115 passed.

1. 6 July 2026Issue #1491 opened by another user
2. 18 October 2026Fix submitted for review (pull request #1536)