Fix Invertd confusion with postprocessing transforms

- Modified Invertd to only invert preprocessing transforms
- When orig_key == key, limits transform_info to preprocessing transforms only
- Counts invertible transforms in preprocessing and uses only that many from applied_operations
- Fixes issue #8396 where Lambdad before Invertd caused errors
- Added test case to verify the fix

The issue occurred because Invertd would try to invert all transforms
in applied_operations, including those from postprocessing. When Lambdad
was applied in postprocessing before Invertd, it would be at the top of
the stack, causing ID mismatch errors when preprocessing transforms tried
to pop themselves.

DCO Remediation Commit for sewon.jeon <sewon.jeon@connecteve.com>

I, sewon.jeon <sewon.jeon@connecteve.com>, hereby add my Signed-off-by to this commit: 342ff38

Signed-off-by: sewon.jeon <sewon.jeon@connecteve.com>

Author: pre-commit-ci[bot]

monai/transforms/post/dictionary.py

@@ -689,7 +689,7 @@ def __call__(self, data: Mapping[Hashable, Any]) -> dict[Hashable, Any]:
             if orig_key in d and isinstance(d[orig_key], MetaTensor):
                 all_transforms = d[orig_key].applied_operations
                 meta_info = d[orig_key].meta
-                
+
                 # If orig_key == key, the data at d[orig_key] may have been modified by
                 # postprocessing transforms. We need to exclude any transforms that were
                 # added after the preprocessing pipeline completed.
@@ -704,7 +704,7 @@ def __call__(self, data: Mapping[Hashable, Any]) -> dict[Hashable, Any]:
                                 num_preproc_transforms += 1
                     elif isinstance(self.transform, InvertibleTransform):
                         num_preproc_transforms = 1
-                    
+
                     # Use only the first N transforms from applied_operations,
                     # where N is the number of preprocessing transforms
                     # This excludes any postprocessing transforms like Lambdad

tests/transforms/inverse/test_invertd.py

@@ -140,10 +140,10 @@ def test_invert(self):
     def test_invert_with_postproc_lambdad(self):
         """Test that Invertd works correctly when postprocessing contains invertible transforms like Lambdad."""
         set_determinism(seed=0)
-        
+
         # Create test images
         im_fname, seg_fname = (make_nifti_image(i) for i in create_test_image_3d(101, 100, 107, noise_max=100))
-        
+
         # Define preprocessing transforms
         preproc = Compose([
             LoadImaged(KEYS, image_only=True),
@@ -152,7 +152,7 @@ def test_invert_with_postproc_lambdad(self):
             ScaleIntensityd("image", minv=1, maxv=10),
             ResizeWithPadOrCropd(KEYS, 100),
         ])
-        
+
         # Define postprocessing with Lambdad before Invertd (the problematic case)
         from monai.transforms import Lambdad
         postproc = Compose([
@@ -161,14 +161,14 @@ def test_invert_with_postproc_lambdad(self):
             # Invertd should only invert the preprocessing transforms
             Invertd(["pred"], preproc, orig_keys=["image"], nearest_interp=True),
         ])
-        
+
         # Apply preprocessing
         data = {"image": im_fname, "label": seg_fname}
         preprocessed = preproc(data)
-        
+
         # Create prediction (copy from preprocessed image)
         preprocessed["pred"] = preprocessed["image"].clone()
-        
+
         # Apply postprocessing with Lambdad before Invertd
         # This should work without errors - the main issue was that it would fail
         result = postproc(preprocessed)
@@ -177,7 +177,7 @@ def test_invert_with_postproc_lambdad(self):
         # Check that the shape was correctly inverted
         self.assertTupleEqual(result["pred"].shape[1:], (101, 100, 107))
         # The fact that we got here without an exception means the fix is working
-        
+
         set_determinism(seed=None)