Fix RandomAnisotropy with copy=False producing wrong shape

Capture spatial_shape and affine as values before downsampling instead
of reading them from the (now-mutated) image reference afterward.

When copy=False, the downsample Resample modifies the image in-place,
so the previously captured reference already reflects the downsampled
state by the time the upsample target is constructed. This caused the
upsample to be a no-op, leaving images at the smaller shape.

Closes #1436

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: fepegar

src/torchio/transforms/augmentation/spatial/random_anisotropy.py

@@ -101,6 +101,8 @@ def apply_transform(self, subject: Subject) -> Subject:
         # NOTE: If copy=False, the underlying image data will be modified in place.
         # We have to obtain the target spatial shape and affine before the transform
         image = subject.get_first_image()
+        original_shape = image.spatial_shape
+        original_affine = image.affine.copy()
         downsample = Resample(
             target=(
                 float(target_spacing[0]),
@@ -118,7 +120,7 @@ def apply_transform(self, subject: Subject) -> Subject:
         )
         downsampled = downsample(subject)
         upsample = Resample(
-            target=(image.spatial_shape, image.affine),
+            target=(original_shape, original_affine),
             image_interpolation=self.image_interpolation,
             scalars_only=self.scalars_only,
             copy=self.copy,

tests/transforms/augmentation/test_random_anisotropy.py

@@ -46,6 +46,22 @@ def test_2d_rgb(self):
         image = ScalarImage(tensor=torch.rand(3, 4, 5, 6))
         RandomAnisotropy()(image)
 
+    def test_copy_false_preserves_shape(self):
+        """Output shape and spacing must match input when copy=False (#1436)."""
+        subject = tio.Subject(
+            t1=tio.ScalarImage(tensor=torch.randn(1, 20, 22, 18)),
+        )
+        transform = RandomAnisotropy(
+            axes=1,
+            downsampling=(2, 2),
+            copy=False,
+        )
+        original_shape = subject.shape
+        original_spacing = subject.spacing
+        result = transform(subject)
+        assert result.shape == original_shape
+        assert result.spacing == original_spacing
+
     def test_2d_with_axis_2_warns(self):
         """Applying to 2D image with axis 2 in axes warns and excludes it."""
         image = ScalarImage(tensor=torch.rand(1, 10, 10, 1))