refactor: extract the nnU-Net transform tail shared by five trainers

[Hendrik-code]

Stacked on #63.

Five get_training_transforms methods across nnUNetTrainerDAExt.py and nnUNetTrainerTest.py ended with a character-identical copy of the same nnU-Net sequence — intensity masking, the -1 label removal, the two cascade transforms, region conversion, deep-supervision downsampling. That's trainers/utils.py::nnunet_tail_transforms now. −255 / +134.

The one thing the copies differed in

Whether the DownsampleSegForDSTransform block was live or commented out:

# NOTE: DownsampleSegForDSTransform is now handled in train_step for GPU augmentations
# if deep_supervision_scales is not None:
#     transforms.append(DownsampleSegForDSTransform(ds_scales=deep_supervision_scales))

The GPU trainers carry it commented because with GPU augmentations the mask is still being deformed after this point, so the multi-scale targets have to be built from the augmented mask in train_step. That's now expressed by passing deep_supervision_scales=None rather than by a comment.

What is deliberately left alone

get_validation_transforms is not folded in. Its cascade branch adds only MoveSegAsOneHotToDataTransform, without the two RandomTransform wrappers — so it's a different sequence, not a sixth copy. Forcing it through the helper would need a flag that changes that branch, costing more than it saves.

How I know it changes nothing

I enumerated the transform list every trainer builds across 160 argument combinations — dummy-2D on/off × use_mask_for_norm × cascade × regions × deep supervision, for all five trainers — and compared the type and repr of every entry before and after:

IDENTICAL: all 160 trainer transform lists unchanged (addresses normalised)

(The only raw differences were <function ... at 0x...> addresses.)

That check is local-only and cannot run in CI. nnunetv2 is an optional extra and pip install -e ".[dev]" does not pull it, so nothing under trainers/ is imported by the test suite at all — CI will go green here without exercising a single line of this diff. That is exactly why this is a separate PR from #63, which CI does cover.

Full suite: 105 passed, 200 subtests (unchanged — none of them touch the trainers). ruff and mypy clean.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.