Preserve inverse transform image scope

[gaoflow]

Summary

• record the image names affected by each applied transform
• restrict rebuilt inverse transforms to those names when they exist on the current data
• add Gamma include/exclude and Flip include regression coverage

Fixes #1480.

Tests

• uv run --group test pytest tests/test_gamma.py tests/test_flip.py -q -k 'inverse_respects_include_scope or inverse_respects_exclude_scope'
• uv run --group test pytest tests/test_gamma.py tests/test_flip.py tests/test_inverse.py tests/test_transforms_base.py -q
• uv run --group test pytest tests/test_bias_field.py tests/test_normalize.py tests/test_standardize.py tests/test_remap_labels.py tests/test_one_hot.py tests/test_sequential_labels.py tests/test_transpose.py tests/test_crop.py tests/test_pad.py -q
• uv run --group test pytest -m 'not slow' -q
• uv run --group quality ruff check
• uv run --group quality ruff format --check
• uv run python -m compileall -q src tests
• uvx --with tox-uv tox -e lint,types,format

AI assistance was used under my direction.

[Copilot]

Pull request overview

This PR fixes a correctness bug in TorchIO’s invertible transform history: inverse transforms are now reconstructed with the same image scope (include/exclude) as the original forward transform, preventing unintended mutation of untouched images during apply_inverse_transform().

Changes:

• Extend AppliedTransform history records to include the names of images affected by each transform.
• Rebuild inverse transforms with an include list derived from the recorded affected image names, intersected with the current data’s images.
• Add regression tests covering Gamma include/exclude scoping and Flip include scoping during inversion.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/torchio/transforms/transform.py	Records affected image names into AppliedTransform during forward passes.
src/torchio/transforms/inverse.py	Reapplies recorded image scope when reconstructing inverse transforms; threads current data into inverse construction.
src/torchio/data/invertible.py	Passes self into inverse reconstruction so current image names are available.
tests/test_gamma.py	Adds regression tests ensuring inverse respects include/exclude scope for Gamma.
tests/test_flip.py	Adds regression test ensuring inverse respects include scope for Flip.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[gaoflow]

Handled in the latest push. _restrict_inverse_to_trace_images() now returns whether the inverse should be appended; when recorded image names have no match in the current data, that inverse step is skipped instead of falling back to all images. I also added a regression test for the removed-image case and adjusted the copied-history test to require matching image names.

[fepegar]

[Generated by a coding agent]

---

For the record: the skip-when-no-overlap behavior added in response to this comment is exactly what introduced the TTA regression flagged in the later review (history copied onto a differently-named pred image gets its spatial inverse silently dropped).

Recording the forward include/exclude instead avoids needing a skip at all. The removed-image case you raised here stays safe because _get_images only returns images that exist, so an include naming an absent image is already a no-op.

[Copilot]

Pull request overview

Copilot reviewed 8 out of 8 changed files in this pull request and generated 1 comment.

[fepegar]

[Generated by a coding agent]

---

Good catch, and I verified it empirically:

• On main, the documented TTA snippet (history copied onto pred=..., then apply_inverse_transform(ignore_intensity=True)) correctly inverts the spatial transform.
• On this branch the inverse is silently skipped: restored.pred equals the still-augmented data, because the forward recorded image_names=['t1'] and there is no overlap with {pred}.

On the suggested remedy, though, I'd avoid the intensity/non-intensity split with a "fall back to all current images" branch: that reintroduces the over-application this PR is trying to prevent (for spatial transforms) and is hard to reason about.

A simpler fix that satisfies both your review comments at once: record the forward transform's own include/exclude (not the resolved image names) and set them on the inverse. Then:

• The original Gamma(include=['a']) corruption is fixed: the inverse gets include=['a'].
• TTA is preserved: a forward with no scope records include=None, so the inverse applies to whatever images are present, including a renamed pred.
• The removed-image case from the earlier thread stays safe without any explicit skip: _get_images only returns images that exist, so include=['a'] on a subject without a is a natural no-op.
• No data= threading, no current_image_names, no intensity/spatial branching.

Intensity inverses re-apply their own scalar-only filter, so label maps still won't be touched.

[fepegar]

[Generated by a coding agent]

---

Thanks for tackling this, the underlying bug (#1480) is real and the new include/exclude regression tests are valuable.

My one blocking concern is a regression in the documented test-time-augmentation workflow, which I verified empirically:

scenario	main	this branch
copy history onto a renamed pred image, then apply_inverse_transform	inverse applied ✅	inverse silently skipped ❌

The TTA guide (docs/how-to/tta.md, lines 37-46) wraps the model output as pred=tio.ScalarImage(...), copies applied_transforms, and inverts. Because the forward Flip recorded image_names=['t1'] and there's no overlap with {pred}, _restrict_inverse_to_trace_images returns False and the spatial inverse is dropped, so predictions are never mapped back to the original space.

Two things mask this in the PR:

• tests/test_flip.py::test_inverse_on_image_via_subject was changed to rename pred -> t1 so the names match, which hides the regression rather than addressing it.
• docs/how-to/tta.md still documents the now-broken pred=... pattern and isn't updated.

Suggested approach

Record the forward transform's own include/exclude in the trace and set them on the inverse, instead of recording resolved image names and intersecting them with the current data. This:

• fixes the original corruption (Gamma(include=['a']) -> inverse include=['a']);
• preserves TTA, since a scopeless forward records include=None and the inverse applies to whatever images are present (including a renamed pred);
• keeps the removed-image case safe with no explicit skip, because _get_images only returns images that exist, so an include naming an absent image is a no-op;
• drops the data= threading, current_image_names, _get_image_names, and the intensity/spatial special-casing entirely.

Intensity inverses re-apply their own scalar-only filter, so label maps stay untouched either way.

If you'd like, I'm happy to push a commit implementing this and add a TTA regression test (renamed prediction image) plus keep the original pred-named test_inverse_on_image_via_subject.

[fepegar]

Thanks for the PR, @gaoflow. Would you like to keep addressing the clankers' reviews? Otherwise I'm happy to take over.

[gaoflow]

I kept going and pushed 16a9175b with the suggested approach:

• AppliedTransform now records the forward transform’s include / exclude scope instead of resolved image names.
• inverse reconstruction applies that recorded scope to the inverse transform and drops the data= / current-image-name intersection path.
• restored the copied-history prediction case to use a renamed pred image, so the documented TTA pattern is covered again.

Verification:

uv run tox -e test -- tests/test_flip.py tests/test_gamma.py tests/test_inverse.py tests/test_pad.py -q
uv run tox -e lint
uv run tox -e format
uv run tox -e prek

[fepegar]

Looking good. Thanks for your contribution, @gaoflow. Thanks also for disclosing your usage of AI.

[fepegar]

@all-contributors please add @gaoflow for bug, code.

[allcontributors]

@fepegar

I've put up a pull request to add @gaoflow! 🎉