def train_Segment_GBM(data_directory):
# Define input modalities to load.
training_modality_dict = {'input_modalities':
['*FLAIR_pp.*', '*T2_pp.*', '*T1_pp.*', '*T1post_pp.*', '*full_edemamask_pp.*'],
'ground_truth': ['*full_edemamask_pp.*']}
# Create a Data Collection
training_data_collection = DataCollection(data_directory, training_modality_dict, verbose=True)
training_data_collection.fill_data_groups()
# Add left-right flips
flip_augmentation = Flip_Rotate_2D(flip=True, rotate=False, data_groups=['input_modalities', 'ground_truth'])
training_data_collection.append_augmentation(flip_augmentation, multiplier=2)
# Define patch sampling regions
def brain_region(data):
return (data['ground_truth'] != 1) & (data['input_modalities'] != 0)
def roi_region(data):
return data['ground_truth'] == 1
# Add patch augmentation
patch_augmentation = ExtractPatches(patch_shape=(32,32,32), patch_region_conditions=[[brain_region, .3], [roi_region, .7]], data_groups=['input_modalities', 'ground_truth'])
training_data_collection.append_augmentation(patch_augmentation, multiplier=70)
# Write the data to hdf5
training_data_collection.write_data_to_file('./test.h5')
# Define model parameters
model_parameters = {input_shape: (32, 32, 32, 4),
downsize_filters_factor: 1,
pool_size: (2, 2, 2),
filter_shape: (3, 3, 3),
dropout: .1,
batch_norm: False,
initial_learning_rate: 0.00001,
output_type: 'binary_label',
num_outputs: 1,
activation: 'relu',
padding: 'same',
implementation: 'keras',
depth: 4,
max_filter=512}
# Create U-Net
if True:
unet_model = UNet(**model_parameters)
# Or load an old one
else:
unet_model = load_old_model('model.h5')
# Define training parameters
training_parameters = {}
# Define training generators
training_generator = None
def train_Segment_GBM(data_directory, val_data_directory):
# Define input modalities to load.
training_modality_dict = {
'input_modalities':
['FLAIR_pp.*', 'T2_pp.*', 'T1_pp.*', 'T1post_pp.*'],
'ground_truth': ['enhancingmask_pp.nii.gz']
}
load_data = False
train_model = False
load_test_data = False
predict = True
training_data = '/mnt/jk489/QTIM_Databank/DeepNeuro_Datasets/BRATS_enhancing_prediction_only_data.h5'
model_file = '/mnt/jk489/QTIM_Databank/DeepNeuro_Datasets/BRATS_enhancing_prediction_only_model.h5'
testing_data = '/mnt/jk489/QTIM_Databank/DeepNeuro_Datasets/BRATS_enhancing_prediction_only_data.h5'
# Write the data to hdf5
if (not os.path.exists(training_data) and train_model) or load_data:
# Create a Data Collection
training_data_collection = DataCollection(
data_directory, modality_dict=training_modality_dict, verbose=True)
training_data_collection.fill_data_groups()
# Define patch sampling regions
def brain_region(data):
return (data['ground_truth'] != 1) & (data['input_modalities'] !=
0)
def roi_region(data):
return data['ground_truth'] == 1
def empty_region(data):
return data['input_modalities'] == 0
# Add patch augmentation
patch_augmentation = ExtractPatches(
patch_shape=(32, 32, 32),
patch_region_conditions=[[empty_region, .05], [brain_region, .25],
[roi_region, .7]],
data_groups=['input_modalities', 'ground_truth'],
patch_dimensions={
'ground_truth': [1, 2, 3],
'input_modalities': [1, 2, 3]
})
training_data_collection.append_augmentation(patch_augmentation,
multiplier=2000)
# Write data to hdf5
training_data_collection.write_data_to_file(training_data)
if train_model:
# Or load pre-loaded data.
training_data_collection = DataCollection(data_storage=training_data,
verbose=True)
training_data_collection.fill_data_groups()
# Add left-right flips
flip_augmentation = Flip_Rotate_2D(
flip=True,
rotate=False,
data_groups=['input_modalities', 'ground_truth'])
# flip_augmentation = Flip_Rotate_3D(data_groups=['input_modalities', 'ground_truth'])
training_data_collection.append_augmentation(flip_augmentation,
multiplier=2)
# Define model parameters
model_parameters = {
'input_shape': (32, 32, 32, 4),
'downsize_filters_factor': 1,
'pool_size': (2, 2, 2),
'filter_shape': (5, 5, 5),
'dropout': 0,
'batch_norm': True,
'initial_learning_rate': 0.000001,
'output_type': 'regression',
'num_outputs': 1,
'activation': 'relu',
'padding': 'same',
'implementation': 'keras',
'depth': 4,
'max_filter': 512
}
# Create U-Net
unet_model = UNet(**model_parameters)
plot_model(unet_model.model,
to_file='model_image_dn.png',
show_shapes=True)
training_parameters = {
'input_groups': ['input_modalities', 'ground_truth'],
'output_model_filepath': model_file,
'training_batch_size': 64,
'num_epochs': 1000,
'training_steps_per_epoch': 20
}
unet_model.train(training_data_collection, **training_parameters)
else:
unet_model = load_old_model(model_file)
# Define input modalities to load.
testing_modality_dict = {
'input_modalities':
['FLAIR_pp.*', 'T2_pp.*', 'T1_pp.*', 'T1post_pp.*']
}
if predict:
testing_data_collection = DataCollection(
val_data_directory,
modality_dict=testing_modality_dict,
verbose=True)
testing_data_collection.fill_data_groups()
if load_test_data:
# Write data to hdf5
testing_data_collection.write_data_to_file(testing_data)
testing_parameters = {
'inputs': ['input_modalities'],
'output_filename': 'brats_enhancing_only_prediction.nii.gz',
'batch_size': 250,
'patch_overlaps': 1,
'output_patch_shape': (26, 26, 26, 4),
'save_all_steps': True
}
prediction = ModelPatchesInference(**testing_parameters)
label_binarization = BinarizeLabel(postprocessor_string='_label')
prediction.append_postprocessor([label_binarization, largest_island])
unet_model.append_output([prediction])
unet_model.generate_outputs(testing_data_collection)
def train_Segment_GBM(data_directory, val_data_directory):
# Define input modalities to load.
training_modality_dict = {
'input_modalities': [
'*FLAIR*', ['*T2SPACE*', '*T2_pp*'], ['*T1_pp.*', '*MPRAGE_Pre*'],
['*T1post_pp.*', '*MPRAGE_POST*'], ['enhancing*'],
['wholetumor*', 'full_edemamask*']
],
'ground_truth': [['enhancing*'], ['wholetumor*', 'full_edemamask*']]
}
load_data = False
train_model = False
load_test_data = True
predict = True
training_data = '/mnt/jk489/QTIM_Databank/DeepNeuro_Datasets/enhancing_label_upsampling_323232.h5'
model_file = 'label_upsampling_323232_model_correct.h5'
testing_data = './FLAIR_upsampling_323232_test.h5'
# Write the data to hdf5
if (not os.path.exists(training_data) and train_model) or load_data:
# Create a Data Collection
training_data_collection = DataCollection(
data_directory, modality_dict=training_modality_dict, verbose=True)
training_data_collection.fill_data_groups()
# Define patch sampling regions
def brain_region(data):
return (data['ground_truth'] != 1) & (data['input_modalities'] !=
0)
def roi_region(data):
return data['ground_truth'] == 1
# Add patch augmentation
patch_augmentation = ExtractPatches(
patch_shape=(32, 32, 32),
patch_region_conditions=[[roi_region, 1]],
data_groups=['input_modalities', 'ground_truth'],
patch_dimensions={
'ground_truth': [0, 1, 2],
'input_modalities': [0, 1, 2]
})
training_data_collection.append_augmentation(patch_augmentation,
multiplier=70)
# Write data to hdf5
training_data_collection.write_data_to_file(training_data)
if train_model:
# Or load pre-loaded data.
training_data_collection = DataCollection(data_storage=training_data,
verbose=True)
training_data_collection.fill_data_groups()
# Choose a modality
choice_augmentation = ChooseData(
axis={
'input_modalities': -1,
'ground_truth': -1
},
choices=[-1, -2],
data_groups=['input_modalities', 'ground_truth'],
random_sample=False)
training_data_collection.append_augmentation(choice_augmentation,
multiplier=2)
# Add down-sampling
mask_augmentation = Downsample(channel=4,
axes={'input_modalities': [-4, -3, -2]},
factor=3,
data_groups=['input_modalities'])
training_data_collection.append_augmentation(mask_augmentation,
multiplier=4)
# Add left-right flips
flip_augmentation = Flip_Rotate_2D(
flip=True,
rotate=False,
data_groups=['input_modalities', 'ground_truth'])
training_data_collection.append_augmentation(flip_augmentation,
multiplier=2)
# Define model parameters
model_parameters = {
'input_shape': (32, 32, 32, 5),
'downsize_filters_factor': 1,
'pool_size': (2, 2, 2),
'filter_shape': (5, 5, 5),
'dropout': 0,
'batch_norm': True,
'initial_learning_rate': 0.000001,
'output_type': 'binary_label',
'num_outputs': 1,
'activation': 'relu',
'padding': 'same',
'implementation': 'keras',
'depth': 4,
'max_filter': 512
}
# Create U-Net
unet_model = UNet(**model_parameters)
plot_model(unet_model.model,
to_file='model_image_dn.png',
show_shapes=True)
training_parameters = {
'input_groups': ['input_modalities', 'ground_truth'],
'output_model_filepath': model_file,
'training_batch_size': 64,
'num_epochs': 1000,
'training_steps_per_epoch': 20
}
unet_model.train(training_data_collection, **training_parameters)
else:
unet_model = load_old_model(model_file)
# Load testing data..
if not os.path.exists(testing_data) or load_test_data:
# Create a Data Collection
testing_data_collection = DataCollection(
val_data_directory,
modality_dict=training_modality_dict,
verbose=True)
testing_data_collection.fill_data_groups()
# Write data to hdf5
testing_data_collection.write_data_to_file(testing_data)
if predict:
testing_data_collection = DataCollection(data_storage=testing_data,
verbose=True)
testing_data_collection.fill_data_groups()
# Choose a modality
choice_augmentation = ChooseData(
axis={
'input_modalities': -1,
'ground_truth': -1
},
choices=[-1, -2],
data_groups=['input_modalities', 'ground_truth'],
random_sample=False)
testing_data_collection.append_augmentation(choice_augmentation,
multiplier=2)
# Add down-sampling
mask_augmentation = Downsample(channel=4,
axes={'input_modalities': [-4, -3, -2]},
factor=3,
data_groups=['input_modalities'],
random_sample=False)
testing_data_collection.append_augmentation(mask_augmentation,
multiplier=3)
testing_parameters = {
'inputs': ['input_modalities'],
'output_filename': 'deepneuro-label.nii.gz',
'batch_size': 250,
'patch_overlaps': 6,
'output_patch_shape': (26, 26, 26, 4)
}
prediction = ModelPatchesInference(testing_data_collection,
**testing_parameters)
unet_model.append_output([prediction])
unet_model.generate_outputs()
def train_Segment_GBM(data_directory, val_data_directory):
# Define input modalities to load.
if True:
training_modality_dict = {
'input_modalities':
['*FLAIR_pp.*', '*T2_pp.*', '*T1_pp.*', '*T1post_pp.*'],
'ground_truth': ['*full_edemamask_pp.*']
}
else:
training_modality_dict = {
'input_modalities': [['*FLAIR_pp.*', 'FLAIR_norm2*'],
['*T1post_pp.*', 'T1post_norm2*']],
'ground_truth': ['*full_edemamask_pp.*', 'FLAIRmask-label.nii.gz']
}
load_data = True
train_model = True
load_test_data = True
predict = True
training_data = './wholetumor_predict_patches_test3.h5'
model_file = 'wholetumor_segnet-58-0.38.h5'
testing_data = './brats_test_case.h5'
# Write the data to hdf5
if (not os.path.exists(training_data) and train_model) or load_data:
# Create a Data Collection
training_data_collection = DataCollection(
data_directory, modality_dict=training_modality_dict, verbose=True)
training_data_collection.fill_data_groups()
# Define patch sampling regions
def brain_region(data):
return (data['ground_truth'] != 1) & (data['input_modalities'] !=
0)
def roi_region(data):
return data['ground_truth'] == 1
# Add patch augmentation
patch_augmentation = ExtractPatches(
patch_shape=(32, 32, 32),
patch_region_conditions=[[brain_region, 1]],
data_groups=['input_modalities', 'ground_truth'])
training_data_collection.append_augmentation(patch_augmentation,
multiplier=200)
# Add left-right flips
flip_augmentation = Flip_Rotate_2D(
flip=True,
rotate=False,
data_groups=['input_modalities', 'ground_truth'])
training_data_collection.append_augmentation(flip_augmentation,
multiplier=2)
# Write data to hdf5
training_data_collection.write_data_to_file(training_data)
# Or load pre-loaded data.
training_data_collection = DataCollection(data_storage=training_data,
verbose=True)
training_data_collection.fill_data_groups()
# Define model parameters
model_parameters = {
'input_shape': (32, 32, 32, 4),
'downsize_filters_factor': 1,
'pool_size': (2, 2, 2),
'filter_shape': (3, 3, 3),
'dropout': 0,
'batch_norm': True,
'initial_learning_rate': 0.000001,
'output_type': 'binary_label',
'num_outputs': 1,
'activation': 'relu',
'padding': 'same',
'implementation': 'keras',
'depth': 4,
'max_filter': 512
}
# Create U-Net
if train_model:
unet_model = UNet(**model_parameters)
plot_model(unet_model.model,
to_file='model_image_dn.png',
show_shapes=True)
training_parameters = {
'input_groups': ['input_modalities', 'ground_truth'],
'output_model_filepath':
'wholetumor_segnet-{epoch:02d}-{loss:.2f}.h5',
'training_batch_size': 2,
'num_epochs': 100,
'training_steps_per_epoch': 200,
'save_best_only': False
}
unet_model.train(training_data_collection, **training_parameters)
else:
unet_model = load_old_model(model_file)
# Load testing data..
if not os.path.exists(testing_data) or load_test_data:
# Create a Data Collection
testing_data_collection = DataCollection(
val_data_directory,
modality_dict=training_modality_dict,
verbose=True)
testing_data_collection.fill_data_groups()
# Write data to hdf5
testing_data_collection.write_data_to_file(testing_data)
if predict:
testing_data_collection = DataCollection(data_storage=testing_data,
verbose=True)
testing_data_collection.fill_data_groups()
testing_parameters = {
'inputs': ['input_modalities'],
'output_filename': 'deepneuro.nii.gz',
'batch_size': 200,
'patch_overlaps': 1
}
prediction = ModelPatchesInference(testing_data_collection,
**testing_parameters)
unet_model.append_output([prediction])
unet_model.generate_outputs()
def train_Segment_GBM(data_directory, val_data_directory):
# Define input modalities to load.
training_modality_dict = {'input_modalities':
['*FLAIR*nii.gz', ['*T2SPACE*nii.gz'], ['*MPRAGE_POST*nii.gz'], ['*MPRAGE_Pre*nii.gz']],
'ground_truth': ['*SUV_r_T2_raw.nii.gz*']}
load_data = False
train_model = False
load_test_data = True
predict = True
training_data = '/mnt/jk489/QTIM_Databank/DeepNeuro_Datasets/TMZ_4_323232.h5'
model_file = 'TMZ_4_323232_model.h5'
testing_data = './TMZ_4_323232_test.h5'
# Write the data to hdf5
if (not os.path.exists(training_data) and train_model) or load_data:
# Create a Data Collection
training_data_collection = DataCollection(data_directory, modality_dict=training_modality_dict, verbose=True)
training_data_collection.fill_data_groups()
# Define patch sampling regions
def brain_region(data):
return (data['ground_truth'] != 1) & (data['input_modalities'] != 0)
def roi_region(data):
return data['ground_truth'] >= 1.5
# Add patch augmentation
patch_augmentation = ExtractPatches(patch_shape=(32, 32, 32), patch_region_conditions=[[brain_region, .5], [roi_region, .5]], data_groups=['input_modalities', 'ground_truth'], patch_dimensions={'ground_truth': [0,1,2], 'input_modalities': [0,1,2]})
training_data_collection.append_augmentation(patch_augmentation, multiplier=2000)
# Write data to hdf5
training_data_collection.write_data_to_file(training_data)
if train_model:
# Or load pre-loaded data.
training_data_collection = DataCollection(data_storage=training_data, verbose=True)
training_data_collection.fill_data_groups()
# Add left-right flips
flip_augmentation = Flip_Rotate_2D(flip=True, rotate=False, data_groups=['input_modalities', 'ground_truth'])
training_data_collection.append_augmentation(flip_augmentation, multiplier=2)
# Define model parameters
model_parameters = {'input_shape': (32, 32, 32, 4),
'downsize_filters_factor': 1,
'pool_size': (2, 2, 2),
'filter_shape': (5, 5, 5),
'dropout': 0,
'batch_norm': True,
'initial_learning_rate': 0.000001,
'output_type': 'regression',
'num_outputs': 1,
'activation': 'relu',
'padding': 'same',
'implementation': 'keras',
'depth': 4,
'max_filter': 512}
# Create U-Net
unet_model = UNet(**model_parameters)
plot_model(unet_model.model, to_file='model_image_dn.png', show_shapes=True)
training_parameters = {'input_groups': ['input_modalities', 'ground_truth'],
'output_model_filepath': model_file,
'training_batch_size': 64,
'num_epochs': 1000,
'training_steps_per_epoch': 20}
unet_model.train(training_data_collection, **training_parameters)
else:
unet_model = load_old_model(model_file)
# Load testing data..
if not os.path.exists(testing_data) or load_test_data:
# Create a Data Collection
testing_data_collection = DataCollection(val_data_directory, modality_dict=training_modality_dict, verbose=True)
testing_data_collection.fill_data_groups()
# Write data to hdf5
testing_data_collection.write_data_to_file(testing_data)
if predict:
testing_data_collection = DataCollection(data_storage=testing_data, verbose=True)
testing_data_collection.fill_data_groups()
flip_augmentation = Copy(data_groups=['input_modalities', 'ground_truth'])
testing_data_collection.append_augmentation(flip_augmentation, multiplier=1)
testing_parameters = {'inputs': ['input_modalities'],
'output_filename': 'deepneuro_suv_4.nii.gz',
'batch_size': 50,
'patch_overlaps': 6,
'output_patch_shape': (26,26,26,4)}
prediction = ModelPatchesInference(testing_data_collection, **testing_parameters)
unet_model.append_output([prediction])
unet_model.generate_outputs()
def build_tensorflow_model(self, batch_size):
""" Break it out into functions?
"""
# Set input/output shapes for reference during inference.
self.model_input_shape = tuple([batch_size] + list(self.input_shape))
self.model_output_shape = tuple([batch_size] + list(self.input_shape))
self.latent = tf.placeholder(tf.float32, [None, self.latent_size])
self.reference_images = tf.placeholder(tf.float32, [None] + list(self.model_input_shape)[1:])
self.synthetic_images = generator(self, self.latent, depth=self.depth, name='generator')
_, _, _, self.discriminator_real_logits = discriminator(self, self.reference_images, depth=self.depth + 1, name='discriminator')
_, _, _, self.discriminator_fake_logits = discriminator(self, self.synthetic_images, depth=self.depth + 1, name='discriminator', reuse=True)
self.basic_loss = tf.reduce_mean(tf.square(self.reference_images - self.synthetic_images))
# Loss functions
self.D_loss = tf.reduce_mean(self.discriminator_fake_logits) - tf.reduce_mean(self.discriminator_real_logits)
self.G_loss = -tf.reduce_mean(self.discriminator_fake_logits)
# Gradient Penalty from Wasserstein GAN GP, I believe? Check on it --andrew
# Also investigate more what's happening here --andrew
self.differences = self.synthetic_images - self.reference_images
self.alpha = tf.random_uniform(shape=[tf.shape(self.differences)[0], 1, 1, 1], minval=0., maxval=1.)
interpolates = self.reference_images + (self.alpha * self.differences)
_, _, _, discri_logits = discriminator(self, interpolates, reuse=True, depth=self.depth + 1, name='discriminator')
gradients = tf.gradients(discri_logits, [interpolates])[0]
# Some sort of norm from papers, check up on it. --andrew
slopes = tf.sqrt(tf.reduce_sum(tf.square(gradients), reduction_indices=[1, 2, 3]))
self.gradient_penalty = tf.reduce_mean((slopes - 1.) ** 2)
tf.summary.scalar("gp_loss", self.gradient_penalty)
# Update Loss functions..
self.D_origin_loss = self.D_loss
self.D_loss += 10 * self.gradient_penalty
self.D_loss += 0.001 * tf.reduce_mean(tf.square(self.discriminator_real_logits - 0.0))
# vgg_model = tf.keras.applications.VGG19(include_top=False,
# weights='imagenet',
# input_tensor=self.synthetic_images,
# input_shape=(64, 64, 3),
# pooling=None,
# classes=1000)
# print(vgg_model)
# self.load_reference_model()
input_tensor = keras.layers.Input(tensor=self.synthetic_images, shape=self.input_shape)
model_parameters = {'input_shape': self.input_shape,
'downsize_filters_factor': 1,
'pool_size': (2, 2),
'kernel_size': (3, 3),
'dropout': 0,
'batch_norm': True,
'initial_learning_rate': 0.00001,
'output_type': 'binary_label',
'num_outputs': 1,
'activation': 'relu',
'padding': 'same',
'implementation': 'keras',
'depth': 3,
'max_filter': 128,
'stride_size': (1, 1),
'input_tensor': input_tensor}
unet_output = UNet(**model_parameters)
unet_model = keras.models.Model(input_tensor, unet_output.output_layer)
unet_model.load_weights('retinal_seg_weights.h5')
if self.hyperverbose:
self.model_summary()
# self.find_layers(['sampling'])
self.activated_tensor = self.grab_tensor(self.activated_tensor_name)
print self.activated_tensor
self.activated_tensor = tf.stack([self.activated_tensor[..., self.filter_num]], axis=-1)
print self.activated_tensor
# self.input_tensor = self.grab_tensor(self.input_tensor_name)
self.activation_loss = -1 * tf.reduce_mean(self.activated_tensor)
self.activaton_graidents = tf.gradients(self.activation_loss, self.synthetic_images)
print self.activaton_graidents
# Hmmm.. better way to do this? Or at least move to function.
t_vars = tf.trainable_variables()
self.d_vars = [var for var in t_vars if 'discriminator' in var.name]
self.g_vars = [var for var in t_vars if 'generator' in var.name]
# Create save/load operation
self.saver = tf.train.Saver(self.g_vars + self.d_vars)
self.G_activation_loss = self.G_loss + .000 * self.activation_loss
# Create Optimizers
self.opti_D = tf.train.AdamOptimizer(learning_rate=self.initial_learning_rate, beta1=0.0, beta2=0.99).minimize(
self.D_loss, var_list=self.d_vars)
self.opti_G = self.tensorflow_optimizer_dict[self.optimizer](learning_rate=self.initial_learning_rate, beta1=0.0, beta2=0.99).minimize(self.G_activation_loss, var_list=self.g_vars)
self.combined_loss = 1 * self.activation_loss + 1 * self.basic_loss
self.combined_optimizer = self.tensorflow_optimizer_dict[self.optimizer](learning_rate=self.initial_learning_rate, beta1=0.0, beta2=0.99).minimize(self.combined_loss, var_list=self.g_vars)
self.basic_optimizer = self.tensorflow_optimizer_dict[self.optimizer](learning_rate=self.initial_learning_rate, beta1=0.0, beta2=0.99).minimize(self.basic_loss, var_list=self.g_vars)
self.activation_optimizer = self.tensorflow_optimizer_dict[self.optimizer](learning_rate=self.initial_learning_rate, beta1=0.0, beta2=0.99).minimize(self.activation_loss, var_list=self.g_vars)