Enable global coordinates in spatial crop transforms

Add convenience transforms for converting points between world and image
coordinate spaces, and extend SpatialCropd to accept string dictionary
keys for ROI parameters, enabling deferred coordinate resolution at
call time.

New transforms:
- TransformPointsWorldToImaged: world-to-image coordinate conversion
- TransformPointsImageToWorldd: image-to-world coordinate conversion

SpatialCropd changes:
- roi_center, roi_size, roi_start, roi_end now accept string keys
- When strings are provided, coordinates are resolved from the data
  dictionary at __call__ time (zero overhead for existing usage)
- Tensors from ApplyTransformToPoints are automatically flattened and
  rounded to integers

Author: emanuilo

monai/transforms/__init__.py

@@ -562,6 +562,12 @@
     ApplyTransformToPointsd,
     ApplyTransformToPointsD,
     ApplyTransformToPointsDict,
+    TransformPointsWorldToImaged,
+    TransformPointsWorldToImageD,
+    TransformPointsWorldToImageDict,
+    TransformPointsImageToWorldd,
+    TransformPointsImageToWorldD,
+    TransformPointsImageToWorldDict,
     AsChannelLastd,
     AsChannelLastD,
     AsChannelLastDict,

monai/transforms/croppad/dictionary.py

@@ -431,17 +431,33 @@ class SpatialCropd(Cropd):
         - a spatial center and size
         - the start and end coordinates of the ROI
 
+    ROI parameters (``roi_center``, ``roi_size``, ``roi_start``, ``roi_end``) can also be specified as
+    string dictionary keys. When a string is provided, the actual coordinate values are read from the
+    data dictionary at call time. This enables pipelines where coordinates are computed by earlier
+    transforms (e.g., :py:class:`monai.transforms.TransformPointsWorldToImaged`) and stored in the
+    data dictionary under the given key.
+
+    Example::
+
+        from monai.transforms import Compose, TransformPointsWorldToImaged, SpatialCropd
+
+        pipeline = Compose([
+            TransformPointsWorldToImaged(keys="roi_start", refer_keys="image"),
+            TransformPointsWorldToImaged(keys="roi_end", refer_keys="image"),
+            SpatialCropd(keys="image", roi_start="roi_start", roi_end="roi_end"),
+        ])
+
     This transform is capable of lazy execution. See the :ref:`Lazy Resampling topic<lazy_resampling>`
     for more information.
     """
 
     def __init__(
         self,
         keys: KeysCollection,
-        roi_center: Sequence[int] | int | None = None,
-        roi_size: Sequence[int] | int | None = None,
-        roi_start: Sequence[int] | int | None = None,
-        roi_end: Sequence[int] | int | None = None,
+        roi_center: Sequence[int] | int | str | None = None,
+        roi_size: Sequence[int] | int | str | None = None,
+        roi_start: Sequence[int] | int | str | None = None,
+        roi_end: Sequence[int] | int | str | None = None,
         roi_slices: Sequence[slice] | None = None,
         allow_missing_keys: bool = False,
         lazy: bool = False,
@@ -450,19 +466,75 @@ def __init__(
         Args:
             keys: keys of the corresponding items to be transformed.
                 See also: :py:class:`monai.transforms.compose.MapTransform`
-            roi_center: voxel coordinates for center of the crop ROI.
+            roi_center: voxel coordinates for center of the crop ROI, or a string key to look up
+                the coordinates from the data dictionary.
             roi_size: size of the crop ROI, if a dimension of ROI size is larger than image size,
-                will not crop that dimension of the image.
-            roi_start: voxel coordinates for start of the crop ROI.
+                will not crop that dimension of the image. Can also be a string key.
+            roi_start: voxel coordinates for start of the crop ROI, or a string key to look up
+                the coordinates from the data dictionary.
             roi_end: voxel coordinates for end of the crop ROI, if a coordinate is out of image,
-                use the end coordinate of image.
+                use the end coordinate of image. Can also be a string key.
             roi_slices: list of slices for each of the spatial dimensions.
             allow_missing_keys: don't raise exception if key is missing.
             lazy: a flag to indicate whether this transform should execute lazily or not. Defaults to False.
         """
-        cropper = SpatialCrop(roi_center, roi_size, roi_start, roi_end, roi_slices, lazy=lazy)
+        self._roi_center = roi_center
+        self._roi_size = roi_size
+        self._roi_start = roi_start
+        self._roi_end = roi_end
+        self._roi_slices = roi_slices
+        self._has_str_roi = any(isinstance(v, str) for v in [roi_center, roi_size, roi_start, roi_end])
+
+        if not self._has_str_roi:
+            cropper = SpatialCrop(roi_center, roi_size, roi_start, roi_end, roi_slices, lazy=lazy)
+        else:
+            # Placeholder cropper for the string-key path. Actual cropping is done with a
+            # local SpatialCrop created at __call__ time once string keys are resolved.
+            # Crop.inverse() reads crop info from the MetaTensor's transform stack (not from
+            # cropper state), so the placeholder still works correctly for inverse operations.
+            cropper = SpatialCrop(roi_start=[0], roi_end=[1], lazy=lazy)
         super().__init__(keys, cropper=cropper, allow_missing_keys=allow_missing_keys, lazy=lazy)
 
+    @staticmethod
+    def _resolve_roi_param(val, d):
+        """Resolve an ROI parameter: if it's a string, look it up in the data dict and flatten."""
+        if not isinstance(val, str):
+            return val
+        if val not in d:
+            raise KeyError(f"ROI key '{val}' not found in the data dictionary.")
+        resolved = d[val]
+        # ApplyTransformToPoints outputs tensors of shape (C, N, dims).
+        # A single coordinate like [142.5, -67.3, 301.8] becomes shape (1, 1, 3).
+        # Flatten to 1-D and round to integers for compute_slices.
+        # Uses banker's rounding (torch.round) to avoid systematic bias in spatial coordinates.
+        if isinstance(resolved, torch.Tensor):
+            resolved = torch.round(resolved.flatten()).to(torch.int64)
+        return resolved
+
+    @property
+    def requires_current_data(self):
+        return self._has_str_roi
+
+    def __call__(self, data: Mapping[Hashable, torch.Tensor], lazy: bool | None = None) -> dict[Hashable, torch.Tensor]:
+        if not self._has_str_roi:
+            return super().__call__(data, lazy=lazy)
+
+        d = dict(data)
+        roi_center = self._resolve_roi_param(self._roi_center, d)
+        roi_size = self._resolve_roi_param(self._roi_size, d)
+        roi_start = self._resolve_roi_param(self._roi_start, d)
+        roi_end = self._resolve_roi_param(self._roi_end, d)
+
+        lazy_ = self.lazy if lazy is None else lazy
+        cropper = SpatialCrop(
+            roi_center=roi_center, roi_size=roi_size,
+            roi_start=roi_start, roi_end=roi_end,
+            roi_slices=self._roi_slices, lazy=lazy_,
+        )
+        for key in self.key_iterator(d):
+            d[key] = cropper(d[key], lazy=lazy_)
+        return d
+
 
 class CenterSpatialCropd(Cropd):
     """

monai/transforms/utility/dictionary.py

@@ -192,6 +192,12 @@
     "ApplyTransformToPointsd",
     "ApplyTransformToPointsD",
     "ApplyTransformToPointsDict",
+    "TransformPointsWorldToImaged",
+    "TransformPointsWorldToImageD",
+    "TransformPointsWorldToImageDict",
+    "TransformPointsImageToWorldd",
+    "TransformPointsImageToWorldD",
+    "TransformPointsImageToWorldDict",
     "FlattenSequenced",
     "FlattenSequenceD",
     "FlattenSequenceDict",
@@ -1910,6 +1916,86 @@ def inverse(self, data: Mapping[Hashable, torch.Tensor]) -> dict[Hashable, torch
         return d
 
 
+class TransformPointsWorldToImaged(ApplyTransformToPointsd):
+    """
+    Dictionary-based transform to convert points from world coordinates to image coordinates.
+
+    This is a convenience subclass of :py:class:`monai.transforms.ApplyTransformToPointsd` with
+    ``invert_affine=True``, which transforms world-space coordinates into the coordinate space of a
+    reference image by inverting the image's affine matrix.
+
+    Args:
+        keys: keys of the corresponding items to be transformed.
+            See also: monai.transforms.MapTransform
+        refer_keys: The key of the reference image used to derive the affine transformation.
+            This is required because the affine must come from a reference image.
+            It can also be a sequence of keys, in which case each refers to the affine applied
+            to the matching points in ``keys``.
+        dtype: The desired data type for the output.
+        affine_lps_to_ras: Defaults to ``False``. Set to ``True`` if your point data is in the RAS
+            coordinate system or you're using ``ITKReader`` with ``affine_lps_to_ras=True``.
+        allow_missing_keys: Don't raise exception if key is missing.
+    """
+
+    def __init__(
+        self,
+        keys: KeysCollection,
+        refer_keys: KeysCollection,
+        dtype: DtypeLike | torch.dtype = torch.float64,
+        affine_lps_to_ras: bool = False,
+        allow_missing_keys: bool = False,
+    ):
+        super().__init__(
+            keys=keys,
+            refer_keys=refer_keys,
+            dtype=dtype,
+            affine=None,
+            invert_affine=True,
+            affine_lps_to_ras=affine_lps_to_ras,
+            allow_missing_keys=allow_missing_keys,
+        )
+
+
+class TransformPointsImageToWorldd(ApplyTransformToPointsd):
+    """
+    Dictionary-based transform to convert points from image coordinates to world coordinates.
+
+    This is a convenience subclass of :py:class:`monai.transforms.ApplyTransformToPointsd` with
+    ``invert_affine=False``, which transforms image-space coordinates into world-space coordinates
+    by applying the reference image's affine matrix directly.
+
+    Args:
+        keys: keys of the corresponding items to be transformed.
+            See also: monai.transforms.MapTransform
+        refer_keys: The key of the reference image used to derive the affine transformation.
+            This is required because the affine must come from a reference image.
+            It can also be a sequence of keys, in which case each refers to the affine applied
+            to the matching points in ``keys``.
+        dtype: The desired data type for the output.
+        affine_lps_to_ras: Defaults to ``False``. Set to ``True`` if your point data is in the RAS
+            coordinate system or you're using ``ITKReader`` with ``affine_lps_to_ras=True``.
+        allow_missing_keys: Don't raise exception if key is missing.
+    """
+
+    def __init__(
+        self,
+        keys: KeysCollection,
+        refer_keys: KeysCollection,
+        dtype: DtypeLike | torch.dtype = torch.float64,
+        affine_lps_to_ras: bool = False,
+        allow_missing_keys: bool = False,
+    ):
+        super().__init__(
+            keys=keys,
+            refer_keys=refer_keys,
+            dtype=dtype,
+            affine=None,
+            invert_affine=False,
+            affine_lps_to_ras=affine_lps_to_ras,
+            allow_missing_keys=allow_missing_keys,
+        )
+
+
 class FlattenSequenced(MapTransform, ReduceTrait):
     """
     Dictionary-based wrapper of :py:class:`monai.transforms.FlattenSequence`.
@@ -1975,4 +2061,6 @@ def __call__(self, data: Mapping[Hashable, NdarrayOrTensor]) -> dict[Hashable, N
 AddCoordinateChannelsD = AddCoordinateChannelsDict = AddCoordinateChannelsd
 FlattenSubKeysD = FlattenSubKeysDict = FlattenSubKeysd
 ApplyTransformToPointsD = ApplyTransformToPointsDict = ApplyTransformToPointsd
+TransformPointsWorldToImageD = TransformPointsWorldToImageDict = TransformPointsWorldToImaged
+TransformPointsImageToWorldD = TransformPointsImageToWorldDict = TransformPointsImageToWorldd
 FlattenSequenceD = FlattenSequenceDict = FlattenSequenced