Code to enable NN to read other file types with support for SimpleITK

loading if installed

data/PROMISE12/setup.py

@@ -6,19 +6,3 @@
   zip_ref = zipfile.ZipFile(os.path.join(zip_dir,zip_filename), 'r')
   zip_ref.extractall(target_dir)
   zip_ref.close()
-
-import SimpleITK
-import glob
-for fn in glob.glob('*.mhd'):
-  if 'segmentation' in fn:
-    fn_out=fn[:-4]+'.nii.gz'
-  else:
-    fn_out=fn[:-4]+'_T2.nii.gz'
-  SimpleITK.WriteImage(SimpleITK.ReadImage(fn),fn_out)
-  try:
-    os.remove(fn)
-    os.remove(fn[:-4]+'.raw')
-  except:
-    pass
-
-

demos/PROMISE12/promise12_demo_train_config.txt

@@ -1,17 +1,18 @@
 [image modality 1]
 path_to_search = ./data/PROMISE12
-filename_contains = T2
-filename_not_contains = Case2
+filename_contains = Case,mhd
+filename_not_contains = Case2,segmentation
 
 [label modality 1]
 path_to_search = ./data/PROMISE12
-filename_contains = segmentation
+filename_contains = Case,segmentation,mhd
 filename_not_contains = Case2
 
 [settings]
 cuda_devices = ""
 model_dir = ./models/model_vnet
 net_name = vnet
+activation_function = prelu
 
 # preprocessing threads parameters
 queue_length = 8
@@ -56,6 +57,9 @@ max_angle = 10.0
 spatial_scaling = True
 min_percentage = -10.0
 max_percentage = 10.0
+min_numb_labels = 2
+min_sampling_ratio = 0.000001
+window_sampling = uniform
 
 # ** training only ** gradient descent and loss parameters
 lr = 0.0001

utilities/misc_io.py

@@ -8,6 +8,14 @@
 import numpy as np
 import scipy.ndimage
 
+image_loaders = [nib.load]
+try:
+    import utilities.simple_itk_as_nibabel
+    image_loaders.append(utilities.simple_itk_as_nibabel.SimpleITKAsNibabel)
+except ImportError:
+    raise ImportWarning('SimpleITK adapter failed to load, reducing the supported file formats.')
+
+
 warnings.simplefilter("ignore", UserWarning)
 
 FILE_EXTENSIONS = [".nii.gz", ".tar.gz"]
@@ -23,10 +31,29 @@ def create_affine_pixdim(affine, pixdim):
         np.expand_dims(np.append(np.asarray(pixdim), 1), axis=1), [1, 4])
     return np.multiply(np.divide(affine, to_divide.T), to_multiply.T)
 
-
+def load_image(filename):
+    # load an image from a supported filetype and return an object
+    # that matches nibabel's spatialimages interface
+    for image_loader in image_loaders:
+      try:
+        img=image_loader(filename)
+        img = correct_image_if_necessary(img)
+        return img
+      except nib.filebasedimages.ImageFileError: # if the image_loader cannot handle the type continue to next loader
+        pass
+    raise nib.filebasedimages.ImageFileError('No loader could load the file') # Throw last error
+
+def correct_image_if_necessary(img):
+  # Check that affine matches zooms
+  pixdim = img.header.get_zooms()
+  if not np.array_equal(np.sqrt(np.sum(np.square(img.affine[0:3, 0:3]), 0)), np.asarray(pixdim)):
+    if img.hasattr('get_sform'): 
+      # assume it is a malformed NIfTI and try to fix it
+      img=rectify_header_sform_qform(img)
+  return img
+
 def rectify_header_sform_qform(img_nii):
     # TODO: check img_nii is a nibabel object
-    pixdim = img_nii.header.get_zooms()
     sform = img_nii.get_sform()
     qform = img_nii.get_qform()
     norm_sform = np.sqrt(np.sum(np.square(sform[0:3, 0:3]), 0))
@@ -155,7 +182,7 @@ def csv_cell_to_volume_5d(csv_cell):
                 data_array[t][m] = expand_to_5d(np.zeros(dimensions))
                 continue
             # load a 3d volume
-            img_nii = nib.load(csv_cell()[t][m])
+            img_nii = load_image(csv_cell()[t][m])
             img_data_shape = img_nii.header.get_data_shape()
             assert np.prod(img_data_shape) > 1
 

utilities/simple_itk_as_nibabel.py

@@ -0,0 +1,27 @@
+import SimpleITK as sitk
+import nibabel
+import numpy as np
+
+
+class SimpleITKAsNibabel(nibabel.spatialimages.SpatialImage):
+  ''' Minimal interface to use a SimpleITK image as if it were
+      a nibabel object. Currently only supports the subset of the 
+      interface used by NiftyNet and is read only
+  '''
+  def __init__(self, filename):
+    try:
+      self._SimpleITKImage = sitk.ReadImage(filename)
+    except RuntimeError as err:
+      if 'Unable to determine ImageIO reader' in str(err):
+        nibabel.filebasedimages.ImageFileError(str(err))
+      else:
+        raise
+    self._header = SimpleITKAsNibabelHeader(self._SimpleITKImage)
+    # get affine transform
+    c=np.array([self._SimpleITKImage.TransformContinuousIndexToPhysicalPoint(p) for p in ((1,0,0),(0,1,0),(0,0,1),(0,0,0))])
+    affine = np.transpose(np.concatenate([np.concatenate([c[0:3]-c[3:],c[3:]],0),[[0.],[0.],[0.],[1.]]],1))
+    super(SimpleITKAsNibabel,self).__init__(sitk.GetArrayFromImage(self._SimpleITKImage), affine)
+class SimpleITKAsNibabelHeader(nibabel.spatialimages.SpatialHeader):
+  def __init__(self, image_reference):
+    super(SimpleITKAsNibabelHeader,self).__init__(data_dtype=sitk.GetArrayViewFromImage(image_reference).dtype,shape=sitk.GetArrayViewFromImage(image_reference).shape,
+                 zooms=image_reference.GetSpacing())

utilities/subject.py

@@ -132,7 +132,7 @@ def _read_original_affine(self):
         and update the corresponding field if not done yet
         """
         img_object = self.__find_first_nibabel_object()
-        util.rectify_header_sform_qform(img_object)
+        util.correct_image_if_necessary(img_object)
         return img_object.affine
 
     @CacheFunctionOutput
@@ -159,9 +159,10 @@ def __find_first_nibabel_object(self):
         list_files = [item for sublist in input_image_files for item in sublist]
         for filename in list_files:
             if not filename == '' and os.path.exists(filename):
-                path, name, ext = util.split_filename(filename)
-                if 'nii' in ext:
-                    return nib.load(filename)
+                try:
+                  return util.load_image(filename)
+                except:
+                  pass # ignore failures here
         return None
 
     def __reorient_to_stand(self, data_5d):