Merge branch 'MIC-DKFZ:master' into master

nnunetv2/inference/predict_from_raw_data.py

@@ -422,6 +422,16 @@ def predict_single_npy_array(self, input_image: np.ndarray, image_properties: di
                                  output_file_truncated: str = None,
                                  save_or_return_probabilities: bool = False):
         """
+        WARNING: SLOW. ONLY USE THIS IF YOU CANNOT GIVE NNUNET MULTIPLE IMAGES AT ONCE FOR SOME REASON.
+
+
+        input_image: Make sure to load the image in the way nnU-Net expects! nnU-Net is trained on a certain axis
+                     ordering which cannot be disturbed in inference,
+                     otherwise you will get bad results. The easiest way to achieve that is to use the same I/O class
+                     for loading images as was used during nnU-Net preprocessing! You can find that class in your
+                     plans.json file under the key "image_reader_writer". If you decide to freestyle, know that the
+                     default axis ordering for medical images is the one from SimpleITK. If you load with nibabel,
+                     you need to transpose your axes AND your spacing from [x,y,z] to [z,y,x]!
         image_properties must only have a 'spacing' key!
         """
         ppa = PreprocessAdapterFromNpy([input_image], [segmentation_previous_stage], [image_properties],

nnunetv2/inference/readme.md

@@ -147,6 +147,11 @@ cons:
 
 tldr:
 - you give one image as npy array
+- axes ordering must match the corresponding training data. The easiest way to achieve that is to use the same I/O class
+                     for loading images as was used during nnU-Net preprocessing! You can find that class in your
+                     plans.json file under the key "image_reader_writer". If you decide to freestyle, know that the
+                     default axis ordering for medical images is the one from SimpleITK. If you load with nibabel,
+                     you need to transpose your axes AND your spacing from [x,y,z] to [z,y,x]!
 - everything is done in the main process: preprocessing, prediction, resampling, (export)
 - no interlacing, slowest variant!
 - ONLY USE THIS IF YOU CANNOT GIVE NNUNET MULTIPLE IMAGES AT ONCE FOR SOME REASON
@@ -160,9 +165,20 @@ cons:
 - never the right choice unless you can only give a single image at a time to nnU-Net
 
 ```python
-    # predict a single numpy array
+    # predict a single numpy array (SimpleITKIO)
     img, props = SimpleITKIO().read_images([join(nnUNet_raw, 'Dataset003_Liver/imagesTr/liver_63_0000.nii.gz')])
     ret = predictor.predict_single_npy_array(img, props, None, None, False)
+
+    # predict a single numpy array (NibabelIO)
+    img, props = NibabelIO().read_images([join(nnUNet_raw, 'Dataset003_Liver/imagesTr/liver_63_0000.nii.gz')])
+    ret = predictor.predict_single_npy_array(img, props, None, None, False)
+
+    # The following IS NOT RECOMMENDED. Use nnunetv2.imageio!
+    # nibabel, we need to transpose axes and spacing to match the training axes ordering for the nnU-Net default:
+    nib.load('Dataset003_Liver/imagesTr/liver_63_0000.nii.gz')
+    img = np.asanyarray(img_nii.dataobj).transpose([2, 1, 0])  # reverse axis order to match SITK
+    props = {'spacing': img_nii.header.get_zooms()[::-1]}      # reverse axis order to match SITK
+    ret = predictor.predict_single_npy_array(img, props, None, None, False)
 ```
 
 ## Predicting with a custom data iterator

nnunetv2/training/dataloading/base_data_loader.py

@@ -19,7 +19,6 @@ def __init__(self,
                  pad_sides: Union[List[int], Tuple[int, ...], np.ndarray] = None,
                  probabilistic_oversampling: bool = False):
         super().__init__(data, batch_size, 1, None, True, False, True, sampling_probabilities)
-        assert isinstance(data, nnUNetDataset), 'nnUNetDataLoaderBase only supports dictionaries as data'
         self.indices = list(data.keys())
 
         self.oversample_foreground_percent = oversample_foreground_percent