def generate_volumes_lesion(test_data_dir, n_neighboringslices=1, new_postfix='', prediction_postfix='prediction', image_size=320):
dict_name = os.path.basename(os.path.dirname(test_data_dir))
print('DICT_NAME: ', dict_name)
tst = pipe.Reader(seg.file_paths_ordered(test_data_dir, iterations=1, image_identifier='volume', label_identifier='segmentation', prediction_postfix=prediction_postfix, new_postfix=new_postfix), name="Read File Names")
tst = tst.multiply(custom_numpy_load_all_slices(n_neighboringslices=n_neighboringslices), name="Load Slices")
tst = _test_val_tail(tst, None, label_of_interest=2, label_required=1, fuse_labels=False, apply_crop=True, discard_labels=False, image_size=image_size)
return tst
def test(test_data_dir, label_of_interest=2, label_required=1,fuse_labels=False):
dict_name = os.path.basename(os.path.dirname(test_data_dir))
# scaling_dict = get_scaling_dict(dict_name + "_scaling", test_data_dir)
# crop_dict = get_crop_dict(dict_name + "_cropping", test_data_dir)
tst = pipe.Reader(seg.file_paths_ordered(test_data_dir, iterations=1, image_identifier='volume', label_identifier='segmentation'), name="Read File Names")
tst = tst.run_on(1)
tst = tst.multiply(seg.numpy_load_all_slices(), name="Load Slices")
# tst = _test_val_tail(tst, scaling_dict, crop_dict)
tst = _test_val_tail(tst, None, label_of_interest=label_of_interest, label_required=label_required,fuse_labels=fuse_labels)
tst = tst.run_on(4)
return tst
def validation(validation_data_dir, slice_type='axial', n_neighboringslices=5, image_size=320, label_of_interest=2,
label_required=1, max_tries=100, fuse_labels=False, apply_crop=False):
dict_name = os.path.basename(os.path.dirname(validation_data_dir))
# scaling_dict = get_scaling_dict(dict_name + "_scaling", validation_data_dir)
# crop_dict = get_crop_dict(dict_name + "_cropping", validation_data_dir)
vld = pipe.Reader(seg.file_paths_random(validation_data_dir, iterations=0, image_identifier='volume',
label_identifier='segmentation'), name="Read File Names")
vld = vld.transform(
custom_load_slices(label_of_interest=label_of_interest, label_required=label_required, min_frequency=1.0,
max_tries=max_tries, slice_type=slice_type, n_neighboringslices=n_neighboringslices),
name="Load Slices")
vld = _test_val_tail(vld, None, image_size=image_size, label_of_interest=label_of_interest,
label_required=label_required, fuse_labels=fuse_labels, apply_crop=apply_crop)
return vld
def get_scaling_dict(name, data_dir):
name = name + '.p'
if os.path.isfile(name):
print('Found dictionary file "{0}". Loading dictionary...'.format(name))
infile = open(name, 'rb')
dictionary = pickle.load(infile)
infile.close()
else:
print('No dictionary file "{0}" found. Creating dictionary...'.format(name))
dictionary = {}
scale = prep.RobustScaler(copy=False)
#LITS
dct = pipe.Reader(seg.file_paths_ordered(data_dir, iterations=1, image_identifier='volume', label_identifier='segmentation'), name="Read File Names")
#3dircad
#dct = pipe.Reader(seg.file_paths_ordered(data_dir, iterations=1, image_identifier='image', label_identifier='label'), name="Read File Names")
dct = dct.transform(seg.numpy_load_volume(), name="Load Volumes")
dct = dct.transform(seg.numpy_clip(-100, 400), name="Clip Pixel Values")
dct = dct.transform(seg.numpy_mask_background(0, -100.))
for inputs, parameters in dct:
file_name = parameters["file_names"][0]
volume = inputs[0]
volume = np.ma.masked_values(volume, -100.)
volume = np.ma.masked_values(volume, 400., copy=False)
volume = np.ma.compressed(volume)
volume = volume.reshape(-1, 1)
print('Filename %s has %s labels' % (file_name,np.unique(inputs[1], return_counts=True)))
scale.fit(volume)
dictionary[file_name] = (scale.center_, scale.scale_)
outfile = open(name, 'wb')
pickle.dump(dictionary, outfile)
outfile.close()
print('Dictionary loaded.')
return dictionary
def get_crop_dict(name, data_dir):
name = name + '.p'
if os.path.isfile(name):
print('Found dictionary file "{0}". Loading dictionary...'.format(name))
infile = open(name, 'rb')
dictionary = pickle.load(infile)
infile.close()
else:
print('No dictionary file "{0}" found. Creating dictionary...'.format(name))
dictionary = {}
dct = pipe.Reader(seg.file_paths_ordered(data_dir, iterations=1, image_identifier='volume', label_identifier='segmentation'), name="Read File Names")
dct = dct.transform(seg.numpy_load_volume(), name="Load Volumes")
for inputs, parameters in dct:
file_name = parameters["file_names"][0]
volume = inputs[1]
x = np.any(volume, axis=(1, 2))
y = np.any(volume, axis=(0, 2))
z = np.any(volume, axis=(0, 1))
xmin, xmax = np.where(x)[0][[0, -1]]
ymin, ymax = np.where(y)[0][[0, -1]]
zmin, zmax = np.where(z)[0][[0, -1]]
dictionary[file_name] = (xmin, xmax, ymin, ymax, zmin, zmax)
outfile = open(name, 'wb')
pickle.dump(dictionary, outfile)
outfile.close()
print('Dictionary loaded.')
return dictionary
def training(train_data_dir,slice_type='axial',n_neighboringslices=5, image_size=320, oversample=True,
label_of_interest=2, label_required=1, max_tries=500, min_frequency=1.0, fuse_labels=False,
apply_crop=True, data_augmentation=False):
print(train_data_dir, slice_type, n_neighboringslices, image_size, oversample, label_of_interest, label_required,
max_tries, min_frequency, fuse_labels, apply_crop, data_augmentation)
dict_name = os.path.basename(os.path.dirname(train_data_dir))
# scaling_dict = get_scaling_dict(dict_name + "_scaling", train_data_dir)
sampler = None
# Instantiate the oversampling class which is used in the custom slice loader to oversample outliers
if oversample:
sampler = oversampler.Oversampler(label_of_interest=2, label_required=label_required, train_data_dir=train_data_dir, debug=False)
# crop_dict = get_crop_dict(dict_name + "_cropping", train_data_dir)
# Load slice
# 构建一个reader的对象,他的next属性可以获取下一个图像的路径
tr = pipe.Reader(seg.file_paths_random(train_data_dir, iterations=0, image_identifier='volume', label_identifier='segmentation'), name="Read File Names")
tr = tr.transform(custom_load_slices(label_of_interest=label_of_interest, label_required=label_required,
min_frequency=min_frequency, max_tries=max_tries, slice_type=slice_type,
n_neighboringslices=n_neighboringslices, oversampler=sampler),
name="Load Slices")
# Random transformations
tr = tr.transform(seg.numpy_rotation2D(1.0, upper_bound=90, min_val=-350.), name="Random Rotation")
if data_augmentation:
tr = tr.transform(seg.numpy_clip_scale(-300, 500, 255), name='Clip Scale')
tr = tr.transform(seg.numpy_image_augmentation(0.5), name='Image Augmentation')
tr = tr.transform(seg.numpy_random_zoom2D(1.0, [image_size, image_size], lower_bound=0.8, upper_bound=1.2),
name="Random Slice Scaling")
tr = tr.transform(seg.numpy_translation2D(0.5, factor=0.25, default_border=0.25, label_of_interest=1),
name="Random Translation")
if data_augmentation:
tr = _test_val_tail(tr, None, image_size=image_size, label_of_interest=label_of_interest,
label_required=label_required, fuse_labels=fuse_labels, apply_crop=apply_crop,
s_min=0, s_max=255)
else:
tr = _test_val_tail(tr, None, image_size=image_size, label_of_interest=label_of_interest,
label_required=label_required, fuse_labels=fuse_labels, apply_crop=apply_crop)
return tr
def build_oversampling_dict(self):
dict_name = os.path.basename(os.path.dirname(self.train_data_dir))
name = dict_name + '_oversampling' + '.p'
if os.path.isfile(name):
print(
'Found oversampling file "{0}". Loading dictionary...'.format(
name))
infile = open(name, 'rb')
self.cache, self.buckets, self.global_intensity = pickle.load(
infile)
infile.close()
else:
print('No oversampling file "{0}" found. Creating dictionary...'.
format(name))
#LITS
dct = pipe.Reader(seg.file_paths_ordered(
self.train_data_dir,
iterations=1,
image_identifier='volume',
label_identifier='segmentation'),
name="Read File Names")
dct = dct.multiply(seg.numpy_gen_slices(), name="Load Slices")
for outputs, parameters, inputs, slice_index in dct:
if self.label_required in outputs[1]:
self.register_slice(inputs[1], slice_index, outputs)
self.calculate_global_intensity()
outfile = open(name, 'wb')
pickle.dump([self.cache, self.buckets, self.global_intensity],
outfile)
outfile.close()
print('Oversampling loaded.')