def re_label(label_input, num_data, labels):
relabel = []
for i in range(num_data):
data_label = np.load(label_input[i])['vol_data']
data_label = generators._relabel(data_label, labels)
relabel.append(data_label)
return relabel
def vol_and_label_generator_patch(vol_name,
label_name,
trf_vol_model,
trf_label_model,
resize_model,
num_data,
patch_size,
labels,
num_labels=30,
stride_patch=32):
vol_patch = np.empty([120, 64, 64, 64, 1])
labels_patch = np.empty([120, 64, 64, 64, 30])
count1 = 0
count2 = 0
vol_size = (160, 192, 224)
i = np.random.randint(num_data)
# for i in range(num_data):
da_factor = np.random.random(1) * 5
# print(da_factor)
rand_phi = np.random.randn(10, 12, 14, 3) * da_factor
rand_phi = rand_phi[np.newaxis, ...]
resize_vol = resize_model.predict([rand_phi])
data_vol = np.load(
vol_name[i])['vol_data'] # load the volume data from the list
# print("volume data",i,":",vol_name[i]) # print the volume data used for training
##### data agmentation part
data_vol = data_vol[np.newaxis, ..., np.newaxis]
data_vol = trf_vol_model.predict([data_vol, resize_vol])[0, ..., 0]
######
data_label = np.load(label_name[i])['vol_data'] # load the label data
data_label = generators._relabel(data_label, labels) # relabel
# print("label data",i,":",label_name[i])
##### data agmentation part
data_label = to_categorical(data_label, num_labels) # to one_hot
data_label = data_label[np.newaxis, ...]
data_label = trf_label_model.predict([data_label, resize_vol
])[0, ..., :] # spatial tranform
data_label = np.argmax(data_label, axis=-1)
######
for item in patchlib.patch_gen(data_vol, patch_size, stride=stride_patch):
vol_patch[count1, :, :, :, 0] = item
count1 += 1
for item in patchlib.patch_gen(data_label, patch_size,
stride=stride_patch):
item = to_categorical(item,
num_labels) # change to one-hot representation
labels_patch[count2, :, :, :, :] = item
count2 += 1
return vol_patch, labels_patch
def vol_and_label_generator_patch(vol_name,
label_name,
trf_vol_model,
trf_label_model,
resize_model,
num_data,
patch_size,
labels,
num_labels=30,
stride_patch=32):
# random field deformation augmentation is performed which makes model robust.
vol_patch = np.empty(
[120, 64, 64, 64,
1]) # patch size is 64,64,64 and 120 patches per volume
labels_patch = np.empty([120, 64, 64, 64, 30]) # 30 label is used
vol_size = (160, 192, 224) # size of the image
i = np.random.randint(num_data) # to randomly pick the data
da_factor = np.random.random(
1) * 5 # control the deformation by restrict the max
rand_phi = np.random.randn(
10, 12, 14, 3
) * da_factor # play around the initial random array to find best fit for data
rand_phi = rand_phi[np.newaxis, ...]
resize_vol = resize_model.predict(
[rand_phi]) # resize the random field to become smooth
data_vol = np.load(
vol_name[i])['vol_data'] # load the volume data from the list
data_vol = data_vol[np.newaxis, ..., np.newaxis]
data_vol = trf_vol_model.predict(
[data_vol, resize_vol])[0, ...,
0] # warp the the image to random field
data_label = np.load(label_name[i])['vol_data'] # load the label data
data_label = generators._relabel(data_label, labels) # relabel
data_label = to_categorical(data_label, num_labels) # to one_hot matrix
data_label = data_label[np.newaxis, ...]
data_label = trf_label_model.predict(
[data_label, resize_vol])[0, ..., :] # warp label to random field
data_label = np.argmax(data_label, axis=-1)
for idx, item in enumerate(
patchlib.patch_gen(data_vol, patch_size, stride=stride_patch)):
vol_patch[idx, :, :, :, 0] = item
for idx, item in enumerate(
patchlib.patch_gen(data_label, patch_size, stride=stride_patch)):
item = to_categorical(item,
num_labels) # change to one-hot representation
labels_patch[idx, :, :, :, :] = item
return vol_patch, labels_patch
def re_label(label_input, num_data, labels):
'''
This function relabels the given label
input - label_input: input label data(3D)
num_data: number of data in label_input
labels: labels of interest
This function usese generators function from Neuron toolbox
'''
relabel = []
for i in range(num_data):
data_label = np.load(label_input[i])['vol_data']
data_label = generators._relabel(data_label, labels)
relabel.append(data_label)
return relabel
def label_generator_patch(label_name,
num_data,
patch_size,
labels,
num_labels,
stride_patch=32,
out=1):
# 120 = number of patches for one volume, 30 = number of labels of interest
labels_patch = np.empty([num_data * 120, 64, 64, 64, 30])
labels_patch2 = []
patch_loc = []
count = 0 # count the batch size for the network
for i in range(num_data):
data_label = np.load(label_name[i])['vol_data']
print("label data", i, ":", label_name[i])
data_label = generators._relabel(data_label, labels)
temp = []
loc_temp = []
if out == 2:
# generate the patch and store them in a list
for item, loc in patchlib.patch_gen(data_label,
patch_size,
stride=stride_patch,
nargout=out):
item = to_categorical(
item, num_labels) # change to one-hot representation
item = np.reshape(item, (1, ) + item.shape)
temp.append(item)
loc_temp.append(loc)
labels_patch2.append(temp)
patch_loc.append(loc_temp)
elif out == 1:
for item in patchlib.patch_gen(data_label,
patch_size,
stride=stride_patch):
item = to_categorical(
item, num_labels) # change to one-hot representation
# vol_patch = [batch size, (dimension), channel(30 as default)]
labels_patch[count, :, :, :, :] = item
count += 1
if out == 1:
return labels_patch
elif out == 2:
return labels_patch2, patch_loc