readme

niftynet/contrib/segmentation_selective_sampler/README.md

@@ -0,0 +1,8 @@
+### Training segmentation network with selective sampling of image windows
+
+To run this application (from NiftyNet source code root folder)
+with a config file (e.g., `niftynet/contrib/segmentation_selective_sampler/selective_seg.ini`)
+```bash
+python net_run.py train -a niftynet.contrib.segmentation_selective_sampler.ss_app.SelectiveSampling \
+                        -c niftynet/contrib/segmentation_selective_sampler/selective_seg.ini
+```

niftynet/contrib/segmentation_selective_sampler/selective_seg.ini

@@ -0,0 +1,80 @@
+############################ input configuration sections
+[modality1]
+csv_file=
+path_to_search = ./example_volumes/monomodal_parcellation
+filename_contains = T1
+filename_not_contains =
+spatial_window_size = (20, 42, 42)
+interp_order = 3
+pixdim=(1.0, 1.0, 1.0)
+axcodes=(A, R, S)
+
+[label]
+path_to_search = ./example_volumes/monomodal_parcellation
+filename_contains = Label
+filename_not_contains =
+spatial_window_size = (20, 42, 42)
+interp_order = 0
+pixdim=(1.0, 1.0, 1.0)
+axcodes=(A, R, S)
+
+############################## system configuration sections
+[SYSTEM]
+cuda_devices = ""
+num_threads = 2
+num_gpus = 1
+model_dir = ./models/model_monomodal_toy
+
+[NETWORK]
+name = toynet
+activation_function = prelu
+batch_size = 1
+decay = 0.1
+reg_type = L2
+
+# volume level preprocessing
+volume_padding_size = 21
+# histogram normalisation
+histogram_ref_file = ./example_volumes/monomodal_parcellation/standardisation_models.txt
+norm_type = percentile
+cutoff = (0.01, 0.99)
+normalisation = False
+whitening = False
+normalise_foreground_only=True
+foreground_type = otsu_plus
+multimod_foreground_type = and
+
+queue_length = 20
+
+
+[TRAINING]
+sample_per_volume = 32
+rotation_angle = (-10.0, 10.0)
+scaling_percentage = (-10.0, 10.0)
+random_flipping_axes= 1
+lr = 0.01
+loss_type = Dice
+starting_iter = 0
+save_every_n = 100
+max_iter = 10
+max_checkpoints = 20
+
+[INFERENCE]
+border = (0, 0, 1)
+#inference_iter = 10
+save_seg_dir = ./output/toy
+output_interp_order = 0
+spatial_window_size = (0, 0, 3)
+
+############################ custom configuration sections
+[SEGMENTATION]
+image = modality1
+label = label
+output_prob = False
+num_classes = 160
+label_normalisation = True
+
+rand_samples=0
+num_min_labels=1
+proba_connect=True
+compulsory_labels=0, 1

niftynet/contrib/segmentation_selective_sampler/selective_sampling_app.py → niftynet/contrib/segmentation_selective_sampler/ss_app.py

@@ -27,11 +27,11 @@
 SUPPORTED_INPUT = set(['image', 'label', 'weight', 'sampler'])
 
 
-class SegmentationApplication(BaseApplication):
+class SelectiveSampling(BaseApplication):
     REQUIRED_CONFIG_SECTION = "SEGMENTATION"
 
     def __init__(self, net_param, action_param, is_training):
-        super(SegmentationApplication, self).__init__()
+        super(SelectiveSampling, self).__init__()
         tf.logging.info('starting segmentation application')
         self.is_training = is_training
 
@@ -155,7 +155,7 @@ def initialise_selective_sampler(self):
             self.segmentation_param.min_ratio_sampling,
             self.segmentation_param.num_min_labels,
             self.segmentation_param.proba_connect)
-        self.sampler = [
+        self.sampler = [[
             SelectiveSampler(
                 reader=reader,
                 data_param=self.data_param,
@@ -164,7 +164,7 @@ def initialise_selective_sampler(self):
                 constraint=selective_constraints,
                 random_windows_per_image=self.segmentation_param.rand_samples,
                 queue_length=self.net_param.queue_length)
-            for reader in self.readers]
+            for reader in self.readers]]
 
     def initialise_grid_sampler(self):
         self.sampler = [[GridSampler(
@@ -185,9 +185,9 @@ def initialise_grid_aggregator(self):
 
     def initialise_sampler(self):
         if self.is_training:
-            self.SUPPORTED_SAMPLING[self.net_param.window_sampling][0]()
+            self.SUPPORTED_SAMPLING['selective'][0]()
         else:
-            self.SUPPORTED_SAMPLING[self.net_param.window_sampling][1]()
+            self.SUPPORTED_SAMPLING['selective'][1]()
 
     def initialise_network(self):
         w_regularizer = None