Add pre-cropping

docs/customization.rst

@@ -154,6 +154,17 @@ Feature Extractor Level
 .. note::
     Resampling is disabled when either `resampledPixelSpacing` or `interpolator` is set to `None`
 
+*Pre-Cropping*
+
+- preCrop [False]: Boolean, if true and resampling is disabled, crops the image onto the bounding box with additional
+  padding as specified in ``padDistance``. Similar to padding after resampling, padding does not exceed original image
+  bounds after pre-cropping. Setting ``preCrop`` to true speeds up extraction and makes it less memory intensive,
+  especially in the case of large images with only small ROIs.
+
+.. note::
+  Because image and mask are also cropped onto the bounding box before they are passed to the feature classes,
+  pre-crop is only beneficial when filters are enabled.
+
 *Resegmentation*
 
 - resegmentRange [None]: List of 2 floats, specifies the lower and upper threshold, respectively. Segmented voxels

examples/exampleSettings/exampleMR_NoResampling.yaml

@@ -75,7 +75,9 @@ setting:
 
   # Resampling:
   # Not enabled in this example. However, because texture calculation assumes isotropic spacing, a forced 2D extraction
-  # is used, therefore only requiring the voxels to be isotropic in-plane.
+  # is used, therefore only requiring the voxels to be isotropic in-plane. Enable pre-cropping to reduce memory
+  # footprint and speed up applying the filters.
+  preCrop: true
 
   # Forced 2D extracion:
   # This allows to calculate texture features using anisotropic voxels (although it assumes that voxels are isotropic

radiomics/featureextractor.py

@@ -85,6 +85,7 @@ class specific are defined in the respective feature classes and and not include
             'removeOutliers': None,
             'resampledPixelSpacing': None,  # No resampling by default
             'interpolator': 'sitkBSpline',  # Alternative: sitk.sitkBSpline
+            'preCrop': False,
             'padDistance': 5,
             'distances': [1],
             'force2D': False,
@@ -446,6 +447,15 @@ def loadImage(self, ImageFilePath, MaskFilePath):
                                                   self.settings['interpolator'],
                                                   self.settings['label'],
                                                   self.settings['padDistance'])
+    elif self.settings['preCrop']:
+      bb, correctedMask = imageoperations.checkMask(image, mask, **self.settings)
+      if correctedMask is not None:
+        # Update the mask if it had to be resampled
+        mask = correctedMask
+      if bb is None:
+        # Mask checks failed
+        return None, None
+      image, mask = imageoperations.cropToTumorMask(image, mask, bb, self.settings['padDistance'])
 
     return image, mask
 

radiomics/imageoperations.py

@@ -363,7 +363,7 @@ def _checkROI(imageNode, maskNode, label):
   return bb
 
 
-def cropToTumorMask(imageNode, maskNode, boundingBox):
+def cropToTumorMask(imageNode, maskNode, boundingBox, padDistance=0):
   """
   Create a sitkImage of the segmented region of the image based on the input label.
 
@@ -382,8 +382,12 @@ def cropToTumorMask(imageNode, maskNode, boundingBox):
   maskNode = sitk.Cast(maskNode, sitk.sitkInt32)
   size = numpy.array(maskNode.GetSize())
 
-  ijkMinBounds = boundingBox[0::2]
-  ijkMaxBounds = size - boundingBox[1::2] - 1
+  ijkMinBounds = boundingBox[0::2] - padDistance
+  ijkMaxBounds = size - boundingBox[1::2] - padDistance - 1
+
+  # Ensure cropped area is not outside original image bounds
+  ijkMinBounds = numpy.maximum(ijkMinBounds, 0)
+  ijkMaxBounds = numpy.maximum(ijkMaxBounds, 0)
 
   # Crop Image
   logger.debug('Cropping to size %s', (boundingBox[1::2] - boundingBox[0::2]) + 1)

radiomics/schemas/paramSchema.yaml

@@ -68,6 +68,8 @@ mapping:
       resegmentRange:
         seq:
           - type: float
+      preCrop:
+        type: bool
       sigma:
         seq:
           - type: float