Generalize TestTimeAugmentation to non-spatial predictions (#8715)

Fixes #8276

### Description

- Added a new argument `apply_inverse_to_pred`. Defaults to `True` to
preserve backward compatibility. When set to `False`, it skips the
inverse transformation step and aggregates the model predictions
directly.
- Added a new unit test to simulate a classification task with spatial
augmentation, verifying that the aggregation works correctly without
spatial inversion.

### Types of changes
<!--- Put an `x` in all the boxes that apply, and remove the not
applicable items -->
- [x] Non-breaking change (fix or new feature that would not break
existing functionality).
- [ ] Breaking change (fix or new feature that would cause existing
functionality to change).
- [x] New tests added to cover the changes.
- [ ] Integration tests passed locally by running `./runtests.sh -f -u
--net --coverage`.
- [ ] Quick tests passed locally by running `./runtests.sh --quick
--unittests --disttests`.
- [ ] In-line docstrings updated.
- [ ] Documentation updated, tested `make html` command in the `docs/`
folder.

---------

Signed-off-by: ytl0623 <david89062388@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Eric Kerfoot <17726042+ericspod@users.noreply.github.com>

Author: 3 people

monai/data/test_time_augmentation.py

@@ -16,7 +16,6 @@
 from copy import deepcopy
 from typing import TYPE_CHECKING, Any
 
-import numpy as np
 import torch
 
 from monai.config.type_definitions import NdarrayOrTensor
@@ -68,7 +67,7 @@ class TestTimeAugmentation:
     Args:
         transform: transform (or composed) to be applied to each realization. At least one transform must be of type
         `RandomizableTrait` (i.e. `Randomizable`, `RandomizableTransform`, or `RandomizableTrait`).
-            . All random transforms must be of type `InvertibleTransform`.
+        When `apply_inverse_to_pred` is True, all random transforms must be of type `InvertibleTransform`.
         batch_size: number of realizations to infer at once.
         num_workers: how many subprocesses to use for data.
         inferrer_fn: function to use to perform inference.
@@ -92,6 +91,11 @@ class TestTimeAugmentation:
             will return the full data. Dimensions will be same size as when passing a single image through
             `inferrer_fn`, with a dimension appended equal in size to `num_examples` (N), i.e., `[N,C,H,W,[D]]`.
         progress: whether to display a progress bar.
+        apply_inverse_to_pred: whether to apply inverse transformations to the predictions.
+            If the model's prediction is spatial (e.g. segmentation), this should be `True` to map the predictions
+            back to the original spatial reference.
+            If the prediction is non-spatial (e.g. classification label or score), this should be `False` to
+            aggregate the raw predictions directly. Defaults to `True`.
 
     Example:
         .. code-block:: python
@@ -125,6 +129,7 @@ def __init__(
         post_func: Callable = _identity,
         return_full_data: bool = False,
         progress: bool = True,
+        apply_inverse_to_pred: bool = True,
     ) -> None:
         self.transform = transform
         self.batch_size = batch_size
@@ -134,6 +139,7 @@ def __init__(
         self.image_key = image_key
         self.return_full_data = return_full_data
         self.progress = progress
+        self.apply_inverse_to_pred = apply_inverse_to_pred
         self._pred_key = CommonKeys.PRED
         self.inverter = Invertd(
             keys=self._pred_key,
@@ -152,20 +158,23 @@ def __init__(
 
     def _check_transforms(self):
         """Should be at least 1 random transform, and all random transforms should be invertible."""
-        ts = [self.transform] if not isinstance(self.transform, Compose) else self.transform.transforms
-        randoms = np.array([isinstance(t, Randomizable) for t in ts])
-        invertibles = np.array([isinstance(t, InvertibleTransform) for t in ts])
-        # check at least 1 random
-        if sum(randoms) == 0:
+        transforms = [self.transform] if not isinstance(self.transform, Compose) else self.transform.transforms
+        warns = []
+        randoms = []
+
+        for idx, t in enumerate(transforms):
+            if isinstance(t, Randomizable):
+                randoms.append(t)
+                if self.apply_inverse_to_pred and not isinstance(t, InvertibleTransform):
+                    warns.append(f"Transform #{idx} (type {type(t).__name__}) is random but not invertible.")
+
+        if len(randoms) == 0:
+            warns.append("TTA usually requires at least one `Randomizable` transform in the given transform sequence.")
+
+        if len(warns) > 0:
             warnings.warn(
-                "TTA usually has at least a `Randomizable` transform or `Compose` contains `Randomizable` transforms."
+                "TTA has encountered issues with the given transforms:\n  " + "\n  ".join(warns), stacklevel=2
             )
-        # check that whenever randoms is True, invertibles is also true
-        for r, i in zip(randoms, invertibles):
-            if r and not i:
-                warnings.warn(
-                    f"Not all applied random transform(s) are invertible. Problematic transform: {type(r).__name__}"
-                )
 
     def __call__(
         self, data: dict[str, Any], num_examples: int = 10
@@ -199,7 +208,10 @@ def __call__(
         for b in tqdm(dl) if has_tqdm and self.progress else dl:
             # do model forward pass
             b[self._pred_key] = self.inferrer_fn(b[self.image_key].to(self.device))
-            outs.extend([self.inverter(PadListDataCollate.inverse(i))[self._pred_key] for i in decollate_batch(b)])
+            if self.apply_inverse_to_pred:
+                outs.extend([self.inverter(PadListDataCollate.inverse(i))[self._pred_key] for i in decollate_batch(b)])
+            else:
+                outs.extend([i[self._pred_key] for i in decollate_batch(b)])
 
         output: NdarrayOrTensor = stack(outs, 0)
 

tests/integration/test_testtimeaugmentation.py

@@ -104,7 +104,7 @@ def test_test_time_augmentation(self):
         # output might be different size, so pad so that they match
         train_loader = DataLoader(train_ds, batch_size=2, collate_fn=pad_list_data_collate)
 
-        model = UNet(2, 1, 1, channels=(6, 6), strides=(2, 2)).to(device)
+        model = UNet(2, 1, 1, channels=(6, 6), strides=(2,)).to(device)
         loss_function = DiceLoss(sigmoid=True)
         optimizer = torch.optim.Adam(model.parameters(), 1e-3)
 
@@ -181,6 +181,43 @@ def test_image_no_label(self):
         tta = TestTimeAugmentation(transforms, batch_size=5, num_workers=0, inferrer_fn=lambda x: x, orig_key="image")
         tta(self.get_data(1, (20, 20), include_label=False))
 
+    def test_non_spatial_output(self):
+        """
+        Test TTA for non-spatial output (e.g., classification scores).
+        Verifies that setting `apply_inverse_to_pred=False` correctly aggregates
+        predictions without attempting spatial inversion.
+        """
+        input_size = (20, 20)
+        data = {"image": np.random.rand(1, *input_size).astype(np.float32)}
+
+        transforms = Compose(
+            [EnsureChannelFirstd("image", channel_dim="no_channel"), RandFlipd("image", prob=1.0, spatial_axis=0)]
+        )
+
+        def mock_classifier(x):
+            batch_size = x.shape[0]
+            return torch.tensor([[0.2, 0.8]] * batch_size, dtype=torch.float32, device=x.device)
+
+        tt_aug = TestTimeAugmentation(
+            transform=transforms,
+            batch_size=2,
+            num_workers=0,
+            inferrer_fn=mock_classifier,
+            device="cpu",
+            orig_key="image",
+            apply_inverse_to_pred=False,
+            return_full_data=False,
+        )
+        mode, mean, std, vvc = tt_aug(data, num_examples=4)
+
+        self.assertEqual(mean.shape, (2,))
+        np.testing.assert_allclose(mean, [0.2, 0.8], atol=1e-6)
+        np.testing.assert_allclose(std, [0.0, 0.0], atol=1e-6)
+
+        tt_aug.return_full_data = True
+        full_output = tt_aug(data, num_examples=4)
+        self.assertEqual(full_output.shape, (4, 2))
+
 
 if __name__ == "__main__":
     unittest.main()