2525from monai .transforms .transform import Randomizable , Transform
2626from monai .transforms .utils import (
2727 compute_divisible_spatial_size ,
28+ generate_label_classes_crop_centers ,
2829 generate_pos_neg_label_crop_centers ,
2930 generate_spatial_bounding_box ,
3031 is_positive ,
3132 map_binary_to_indices ,
33+ map_classes_to_indices ,
3234 weighted_patch_samples ,
3335)
3436from monai .utils import Method , NumpyPadMode , ensure_tuple , ensure_tuple_rep , fall_back_tuple
4648 "CropForeground" ,
4749 "RandWeightedCrop" ,
4850 "RandCropByPosNegLabel" ,
51+ "RandCropByLabelClasses" ,
4952 "ResizeWithPadOrCrop" ,
5053 "BoundingRect" ,
5154]
@@ -766,7 +769,11 @@ def randomize(
766769 ) -> None :
767770 self .spatial_size = fall_back_tuple (self .spatial_size , default = label .shape [1 :])
768771 if fg_indices is None or bg_indices is None :
769- fg_indices_ , bg_indices_ = map_binary_to_indices (label , image , self .image_threshold )
772+ if self .fg_indices is not None and self .bg_indices is not None :
773+ fg_indices_ = self .fg_indices
774+ bg_indices_ = self .bg_indices
775+ else :
776+ fg_indices_ , bg_indices_ = map_binary_to_indices (label , image , self .image_threshold )
770777 else :
771778 fg_indices_ = fg_indices
772779 bg_indices_ = bg_indices
@@ -802,12 +809,7 @@ def __call__(
802809 raise ValueError ("label should be provided." )
803810 if image is None :
804811 image = self .image
805- if fg_indices is None or bg_indices is None :
806- if self .fg_indices is not None and self .bg_indices is not None :
807- fg_indices = self .fg_indices
808- bg_indices = self .bg_indices
809- else :
810- fg_indices , bg_indices = map_binary_to_indices (label , image , self .image_threshold )
812+
811813 self .randomize (label , fg_indices , bg_indices , image )
812814 results : List [np .ndarray ] = []
813815 if self .centers is not None :
@@ -818,6 +820,139 @@ def __call__(
818820 return results
819821
820822
823+ class RandCropByLabelClasses (Randomizable , Transform ):
824+ """
825+ Crop random fixed sized regions with the center being a class based on the specified ratios of every class.
826+ The label data can be One-Hot format array or Argmax data. And will return a list of arrays for all the
827+ cropped images. For example, crop two (3 x 3) arrays from (5 x 5) array with `ratios=[1, 2, 3, 1]`::
828+
829+ image = np.array([
830+ [[0.0, 0.3, 0.4, 0.2, 0.0],
831+ [0.0, 0.1, 0.2, 0.1, 0.4],
832+ [0.0, 0.3, 0.5, 0.2, 0.0],
833+ [0.1, 0.2, 0.1, 0.1, 0.0],
834+ [0.0, 0.1, 0.2, 0.1, 0.0]]
835+ ])
836+ label = np.array([
837+ [[0, 0, 0, 0, 0],
838+ [0, 1, 2, 1, 0],
839+ [0, 1, 3, 0, 0],
840+ [0, 0, 0, 0, 0],
841+ [0, 0, 0, 0, 0]]
842+ ])
843+ cropper = RandCropByLabelClasses(
844+ spatial_size=[3, 3],
845+ ratios=[1, 2, 3, 1],
846+ num_classes=4,
847+ num_samples=2,
848+ )
849+ label_samples = cropper(img=label, label=label, image=image)
850+
851+ The 2 randomly cropped samples of `label` can be:
852+ [[0, 1, 2], [[0, 0, 0],
853+ [0, 1, 3], [1, 2, 1],
854+ [0, 0, 0]] [1, 3, 0]]
855+
856+ If a dimension of the expected spatial size is bigger than the input image size,
857+ will not crop that dimension. So the cropped result may be smaller than expected size, and the cropped
858+ results of several images may not have exactly same shape.
859+
860+ Args:
861+ spatial_size: the spatial size of the crop region e.g. [224, 224, 128].
862+ if a dimension of ROI size is bigger than image size, will not crop that dimension of the image.
863+ if its components have non-positive values, the corresponding size of `label` will be used.
864+ for example: if the spatial size of input data is [40, 40, 40] and `spatial_size=[32, 64, -1]`,
865+ the spatial size of output data will be [32, 40, 40].
866+ ratios: specified ratios of every class in the label to generate crop centers, including background class.
867+ if None, every class will have the same ratio to generate crop centers.
868+ label: the label image that is used for finding every classes, if None, must set at `self.__call__`.
869+ num_classes: number of classes for argmax label, not necessary for One-Hot label.
870+ num_samples: number of samples (crop regions) to take in each list.
871+ image: if image is not None, only return the indices of every class that are within the valid
872+ region of the image (``image > image_threshold``).
873+ image_threshold: if enabled `image`, use ``image > image_threshold`` to
874+ determine the valid image content area and select class indices only in this area.
875+ indices: if provided pre-computed indices of every class, will ignore above `image` and
876+ `image_threshold`, and randomly select crop centers based on them, expect to be 1 dim array
877+ of spatial indices after flattening. a typical usage is to call `ClassesToIndices` transform first
878+ and cache the results for better performance.
879+
880+ """
881+
882+ def __init__ (
883+ self ,
884+ spatial_size : Union [Sequence [int ], int ],
885+ ratios : Optional [List [Union [float , int ]]] = None ,
886+ label : Optional [np .ndarray ] = None ,
887+ num_classes : Optional [int ] = None ,
888+ num_samples : int = 1 ,
889+ image : Optional [np .ndarray ] = None ,
890+ image_threshold : float = 0.0 ,
891+ indices : Optional [List [np .ndarray ]] = None ,
892+ ) -> None :
893+ self .spatial_size = ensure_tuple (spatial_size )
894+ self .ratios = ratios
895+ self .label = label
896+ self .num_classes = num_classes
897+ self .num_samples = num_samples
898+ self .image = image
899+ self .image_threshold = image_threshold
900+ self .centers : Optional [List [List [np .ndarray ]]] = None
901+ self .indices = indices
902+
903+ def randomize (
904+ self ,
905+ label : np .ndarray ,
906+ indices : Optional [List [np .ndarray ]] = None ,
907+ image : Optional [np .ndarray ] = None ,
908+ ) -> None :
909+ self .spatial_size = fall_back_tuple (self .spatial_size , default = label .shape [1 :])
910+ indices_ : List [np .ndarray ]
911+ if indices is None :
912+ if self .indices is not None :
913+ indices_ = self .indices
914+ else :
915+ indices_ = map_classes_to_indices (label , self .num_classes , image , self .image_threshold )
916+ else :
917+ indices_ = indices
918+ self .centers = generate_label_classes_crop_centers (
919+ self .spatial_size , self .num_samples , label .shape [1 :], indices_ , self .ratios , self .R
920+ )
921+
922+ def __call__ (
923+ self ,
924+ img : np .ndarray ,
925+ label : Optional [np .ndarray ] = None ,
926+ image : Optional [np .ndarray ] = None ,
927+ indices : Optional [List [np .ndarray ]] = None ,
928+ ) -> List [np .ndarray ]:
929+ """
930+ Args:
931+ img: input data to crop samples from based on the ratios of every class, assumes `img` is a
932+ channel-first array.
933+ label: the label image that is used for finding indices of every class, if None, use `self.label`.
934+ image: optional image data to help select valid area, can be same as `img` or another image array.
935+ use ``image > image_threshold`` to select the centers only in valid region. if None, use `self.image`.
936+ indices: list of indices for every class in the image, used to randomly select crop centers.
937+
938+ """
939+ if label is None :
940+ label = self .label
941+ if label is None :
942+ raise ValueError ("label should be provided." )
943+ if image is None :
944+ image = self .image
945+
946+ self .randomize (label , indices , image )
947+ results : List [np .ndarray ] = []
948+ if self .centers is not None :
949+ for center in self .centers :
950+ cropper = SpatialCrop (roi_center = tuple (center ), roi_size = self .spatial_size ) # type: ignore
951+ results .append (cropper (img ))
952+
953+ return results
954+
955+
821956class ResizeWithPadOrCrop (Transform ):
822957 """
823958 Resize an image to a target spatial size by either centrally cropping the image or
0 commit comments