ENH: Implement dynamic binning

docs/customization.rst

@@ -290,6 +290,9 @@ Feature Class Level
 - ``binCount`` [None]: integer, > 0, specifies the number of bins to create. The width of the bin is
   then determined by the range in the ROI. No definitive evidence is available on which method of discretization is
   superior, we advise a fixed bin width. See more :ref:`here <radiomics_fixed_bin_width>`.
+- ``dynamicBinning`` [False]: Boolean, if set to true, scales the bin width for derived images by the ratio of the range
+  in the image (ROI) and the range of the original image (ROI). This setting has no effect when a fixed bin count is
+  used, or when a custom bin width has been specified for the filter. See also :py:func:`getBinEdges()`.
 
 *Forced 2D extraction*
 

radiomics/featureextractor.py

@@ -468,6 +468,13 @@ def execute(self, imageFilepath, maskFilepath, label=None, label_channel=None, v
     if not resegmentShape and resegmentedMask is not None:
       mask = resegmentedMask
 
+    dynamic_ref_range = None
+    if self.settings.get('dynamicBinning', False):
+      im_arr = sitk.GetArrayFromImage(image)
+      ma_arr = sitk.GetArrayFromImage(mask) == self.settings.get('label', 1)
+      target_voxel_arr = im_arr[ma_arr]
+      dynamic_ref_range = max(target_voxel_arr) - min(target_voxel_arr)
+
     # 6. Calculate other enabled feature classes using enabled image types
     # Make generators for all enabled image types
     self.logger.debug('Creating image type iterator')
@@ -476,6 +483,8 @@ def execute(self, imageFilepath, maskFilepath, label=None, label_channel=None, v
       args = self.settings.copy()
       args.update(customKwargs)
       self.logger.info('Adding image type "%s" with custom settings: %s' % (imageType, str(customKwargs)))
+      if 'binWidth' not in customKwargs and imageType != 'Original':
+        args['dynamic_ref_range'] = dynamic_ref_range
       imageGenerators = chain(imageGenerators, getattr(imageoperations, 'get%sImage' % imageType)(image, mask, **args))
 
     self.logger.debug('Extracting features')

radiomics/imageoperations.py

@@ -74,7 +74,7 @@ def getBinEdges(parameterValues, **kwargs):
   \mod W = 0`, the maximum intensity will be encoded as numBins + 1, therefore the maximum number of gray
   level intensities in the ROI after binning is number of bins + 1.
 
-  If dynamic binning is enabled (parameter `dynamicBinning`), and no custom binwidth has been defined for the filter,
+  If dynamic binning is enabled (parameter ``dynamicBinning``), and no custom binwidth has been defined for the filter,
   the actual bin width used (:math:`W_{dyn}`) is defined as:
 
   .. math::
@@ -118,13 +118,19 @@ def getBinEdges(parameterValues, **kwargs):
   global logger
   binWidth = kwargs.get('binWidth', 25)
   binCount = kwargs.get('binCount')
+  dynamic_ref_range = kwargs.get('dynamic_ref_range', None)
 
   if binCount is not None:
     binEdges = numpy.histogram(parameterValues, binCount)[1]
     binEdges[-1] += 1  # Ensures that the maximum value is included in the topmost bin when using numpy.digitize
   else:
     minimum = min(parameterValues)
     maximum = max(parameterValues)
+    if dynamic_ref_range is not None and dynamic_ref_range > 0 and minimum < maximum:
+      range_scale = (maximum - minimum) / dynamic_ref_range
+      binWidth = binWidth * range_scale
+      logger.debug('Applied dynamic binning (reference range %g, current range %g), scaled bin width to %g',
+                   dynamic_ref_range, maximum - minimum, binWidth)
 
     # Start binning form the first value lesser than or equal to the minimum value and evenly dividable by binwidth
     lowBound = minimum - (minimum % binWidth)

radiomics/schemas/paramSchema.yaml

@@ -39,6 +39,8 @@ mapping:
         type: int
         range:
           min-ex: 0
+      dynamicBinning:
+        type: bool
       normalize:
         type: bool
       normalizeScale: