def test_net(net, p, outputname):
c = color_codes()
options = parse_inputs()
patch_width = options['patch_width']
patch_size = (patch_width, patch_width, patch_width)
batch_size = options['test_size']
p_name = p[0].rsplit('/')[-2]
patient_path = '/'.join(p[0].rsplit('/')[:-1])
outputname_path = os.path.join(patient_path, outputname + '.nii.gz')
pr_outputname_path = os.path.join(patient_path, outputname + '.pr.nii.gz')
try:
image = load_nii(outputname_path).get_data()
except IOError:
print('%s[%s] %sTesting the network%s' % (c['c'], strftime("%H:%M:%S"), c['g'], c['nc']))
nii = load_nii(p[0])
roi = nii.get_data().astype(dtype=np.bool)
centers = get_mask_voxels(roi)
test_samples = np.count_nonzero(roi)
image = np.zeros_like(roi).astype(dtype=np.uint8)
pr = np.zeros_like(roi).astype(dtype=np.float32)
print('%s[%s] %s<Creating the probability map %s%s%s%s - %s%s%s%s (%d samples)>%s' % (
c['c'], strftime("%H:%M:%S"),
c['g'], c['b'], p_name, c['nc'],
c['g'], c['b'], outputname, c['nc'],
c['g'], test_samples, c['nc']
))
n_centers = len(centers)
image_list = [load_norm_list(p)]
for i in range(0, n_centers, batch_size):
print(
'%f%% tested (step %d/%d)' % (100.0 * i / n_centers, (i / batch_size) + 1, -(-n_centers/batch_size)),
end='\r'
)
sys.stdout.flush()
centers_i = [centers[i:i + batch_size]]
x = get_patches_list(image_list, centers_i, patch_size, True)
x = np.concatenate(x).astype(dtype=np.float32)
y_pr_pred = net.predict(x, batch_size=options['batch_size'])
[x, y, z] = np.stack(centers_i[0], axis=1)
# We store the results
image[x, y, z] = np.argmax(y_pr_pred, axis=1).astype(dtype=np.int8)
pr[x, y, z] = y_pr_pred[:, 1].astype(dtype=np.float32)
print(' '.join([''] * 50), end='\r')
sys.stdout.flush()
# Post-processing (Basically keep the biggest connected region)
# image = get_biggest_region(image)
print('%s -- Saving image %s%s%s' % (c['g'], c['b'], outputname_path, c['nc']))
nii.get_data()[:] = image
nii.to_filename(outputname_path)
nii.get_data()[:] = pr
nii.to_filename(pr_outputname_path)
return image
def vox_generator_test(all_files):
path = options['root_path']
while 1:
for file in all_files:
p = file
coll = glob.glob(os.path.join(path, file) + '/*')
for c in coll:
if 'flair.' in c or 'flair_corrected.' in c:
flair_path = c
if 't1.' in c or 't1_corrected.' in c:
t1_path = c
if 't2.' in c or 't2_corrected.' in c:
t2_path = c
if 't1ce.' in c or 't1ce_corrected.' in c:
t1ce_path = c
flair = load_nii(flair_path).get_data()
t2 = load_nii(t2_path).get_data()
t1 = load_nii(t1_path).get_data()
t1ce = load_nii(t1ce_path).get_data()
data = np.array([flair, t2, t1, t1ce])
data = np.transpose(data, axes=[1, 2, 3, 0])
data_norm = np.array([norm(flair), norm(t2), norm(t1), norm(t1ce)])
data_norm = np.transpose(data_norm, axes=[1, 2, 3, 0])
labels = load_nii(os.path.join(path, p,
p + '_seg.nii.gz')).get_data()
yield data, data_norm, labels
def vox_generator_test(all_files):
path = options['root_path']
while 1:
for file in all_files:
p = file
if options['correction']:
flair = load_nii(os.path.join(path, file, file + '_flair_corrected.nii.gz')).get_data()
t2 = load_nii(os.path.join(path, file, file + '_t2_corrected.nii.gz')).get_data()
t1 = load_nii(os.path.join(path, file, file + '_t1_corrected.nii.gz')).get_data()
t1ce = load_nii(os.path.join(path, file, file + '_t1ce_corrected.nii.gz')).get_data()
else:
flair = load_nii(os.path.join(path, p, p + '_flair.nii.gz')).get_data()
t2 = load_nii(os.path.join(path, p, p + '_t2.nii.gz')).get_data()
t1 = load_nii(os.path.join(path, p, p + '_t1.nii.gz')).get_data()
t1ce = load_nii(os.path.join(path, p, p + '_t1ce.nii.gz')).get_data()
data = np.array([flair, t2, t1, t1ce])
data = np.transpose(data, axes=[1, 2, 3, 0])
data_norm = np.array([norm(flair), norm(t2), norm(t1), norm(t1ce)])
data_norm = np.transpose(data_norm, axes=[1, 2, 3, 0])
labels = load_nii(os.path.join(path, p, p + '_seg.nii.gz')).get_data()
yield data, data_norm, labels
def get_patient_survival_features(path, p, p_features, test=False):
# Init
options = parse_inputs()
roi_sufix = '_seg.nii.gz' if not test else '.nii.gz'
roi = load_nii(os.path.join(path, p, p + roi_sufix)).get_data()
brain = load_nii(os.path.join(path, p, p + options['t1'])).get_data()
brain_vol = np.count_nonzero(brain)
vol_features = map(lambda l: np.count_nonzero(roi == l) / brain_vol, [1, 2, 4])
age_features = [float(p_features['Age']) / 100]
features = [age_features + vol_features]
return features
def get_atlas_vectors(options, dir_name, centers, t1_names):
"""
Generate training data vectors from probabilistic atlases. These vectors are concatenated with fully-connected layers.
"""
subjects = [
f for f in sorted(os.listdir(dir_name))
if os.path.isdir(os.path.join(dir_name, f))
]
atlas_names = [
os.path.join(dir_name, subject, 'tmp', 'MNI_sub_probabilities.nii.gz')
for subject in subjects
]
atlas_images = []
# load atlas, register if does not exist
for t1, atlas, subject in zip(t1_names, atlas_names, subjects):
if os.path.exists(atlas) is False:
print " --> registering priors for scan {}".format(subject)
t = register_masks(t1)
print " (elapsed time {} min.".format(t / 60.0)
atlas_images.append(fix_shape(load_nii(atlas).get_data()))
# ATLAS probabilities (centered voxel)
# convert lesion centers
lc = map(lambda l: np.asarray(l), centers)
atlas_vectors = [
a[c[:, 0], c[:, 1], c[:, 2]] for a, c in zip(atlas_images, lc)
]
# correct for background. if no probability exists for any class, set as background
for index in range(len(atlas_vectors)):
if np.sum(atlas_vectors[index]) == 0:
atlas_vectors[index][14] = 1
return atlas_vectors
def eval_one(npy_path, name, save_nii, save_root, post_process):
pred_path = os.path.join(npy_path, name)
predict = np.load(pred_path).astype(np.float) / 255
label_path = '/data/lbw/structseg2019/data_ori/Task3_Thoracic_OAR/{}/label.nii.gz'.format(name[5:-4])
label = load_nii(label_path).get_data().astype(np.uint8)
label = np.eye(7)[label].transpose([3, 0, 1, 2])[np.newaxis, :]
if post_process:
ori_shape = predict.shape
predict = F.interpolate(torch.from_numpy(predict), [512, 512, 512], mode='trilinear')
predict = F.interpolate(predict, ori_shape[2:], mode='trilinear').numpy()
predict = postprocessing(predict)
if save_nii:
predict_nii = predict.copy()
predict_nii = np.argmax(predict_nii.squeeze(), axis=0).astype(np.uint8)
# predict_npy = predict_npy.transpose([1, 0, 2])
predict_nii = nib.Nifti1Image(predict_nii, None)
nib.save(predict_nii, os.path.join(save_root, name[:-4]))
dices = compute_multi_dice(predict, label)
print(name[:-4])
for i in range(7):
print(all_organs[i], float(dices[i]))
return [dices, '{0:02d}'.format(int(name[5:-4]))]
def get_isbi_data(
d_path='data/ISBI',
images=None,
verbose=1,
preload=True,
):
"""
Function that loads the images and masks of a list of patients.
:param d_path: Path to the LSI data.
:param images: Image names (prefixes).
:param verbose: Level of verbosity
:return: list of numpy arrays for the concatenated images, lesion
mask and brain mask.
"""
lesion_mask_name = 'mask1.nii'
tmp = get_dirs(d_path) # if p_tag in p
p_train = sorted([p for p in tmp],
key=lambda p: int(''.join(filter(str.isdigit, p))))
if images is None:
images = ['flair', 'mprage']
p_trains = []
# Finally we either load all images and normalise them (a lot of RAM) or we
# leave that job to the dataset object.
if verbose > 1:
print('Loading the images')
lesion_names = []
data = []
for p_path in p_train:
tmp_stages = set()
for file in os.listdir(os.path.join(d_path, p_path, 'preprocessed')):
if file.endswith(".nii"):
tmp_stages.add(file.split('_')[1])
for stage in tmp_stages:
lesion_names.append(
os.path.join(d_path, p_path, 'masks',
'_'.join([p_path, stage, lesion_mask_name])))
data.append(
np.stack([
get_normalised_image(
os.path.join(d_path, p_path, 'preprocessed',
'%s_%s_%s_pp.nii' % (p_path, stage, im)),
None,
) for im in images
],
axis=0))
p_trains.append(p_path + '_' + stage)
# Lesion masks (we are using this function for training, so there should
# always be a lesion mask).
if verbose > 1:
print('Loading the lesion masks')
lesions = list(map(get_mask, lesion_names))
brains = [np.full(lesions[0].shape, 1, dtype=int)] * len(lesions)
example_nii = load_nii(
os.path.join(d_path, p_path, 'preprocessed',
'%s_%s_%s_pp.nii' % (p_path, stage, images[0])))
return data, lesions, brains, p_trains, example_nii
def select_voxels_from_previous_model(model, train_x_data, options):
"""
Select training voxels from image segmentation masks
"""
# get_scan names and number of modalities used
scans = train_x_data.keys()
modalities = train_x_data[scans[0]].keys()
# select voxels for training. Discard CSF and darker WM in FLAIR.
# flair_scans = [train_x_data[s]['FLAIR'] for s in scans]
# selected_voxels = select_training_voxels(flair_scans, options['min_th'])
# evaluate training scans using the learned model and extract voxels with probability higher than 0.5
seg_mask = [
test_scan(model,
dict(train_x_data.items()[s:s + 1]),
options,
save_nifti=False) > 0.5 for s in range(len(scans))
]
# check candidate segmentations:
# if no voxels have been selected, return candidate voxels on FLAIR modality > 2
flair_scans = [train_x_data[s]['FLAIR'] for s in scans]
images = [load_nii(name).get_data() for name in flair_scans]
images_norm = [(im.astype(dtype=np.float32) - im[np.nonzero(im)].mean()) /
im[np.nonzero(im)].std() for im in images]
seg_mask = [
im > 2 if np.sum(seg) == 0 else seg
for im, seg in zip(images_norm, seg_mask)
]
return seg_mask
def test_scan(model, test_x_data, options, save_nifti= True, candidate_mask = None):
"""
Test data based on one model
Input:
- test_x_data: a nested dictionary containing training image paths:
train_x_data['scan_name']['modality'] = path_to_image_modality
- save_nifti: save image segmentation
- candidate_mask: a binary masks containing voxels to classify
Output:
- test_scan = Output image containing the probability output segmetnation
- If save_nifti --> Saves a nifti file at specified location options['test_folder']/['test_scan']
"""
# get_scan name and create an empty nifti image to store segmentation
scans = test_x_data.keys()
flair_scans = [test_x_data[s]['FLAIR'] for s in scans]
flair_image = load_nii(flair_scans[0]).get_data()
seg_image = np.zeros_like(flair_image)
# get test paths
test_folder, scan = os.path.split(flair_scans[0])
# compute lesion segmentation in batches of size options['batch_size']
for batch, centers in load_test_patches(test_x_data, options['patch_size'], options['batch_size'], candidate_mask):
y_pred = model.predict_proba(np.squeeze(batch))
[x, y, z] = np.stack(centers, axis=1)
seg_image[x, y, z] = y_pred[:, 1]
if save_nifti:
out_scan = nib.Nifti1Image(seg_image, np.eye(4))
#out_scan.to_filename(os.path.join(options['test_folder'], options['test_scan'], options['experiment'], options['test_name']))
out_scan.to_filename(os.path.join(test_folder, options['experiment'], options['test_name']))
return seg_image
def get_normalised_image(
image_name, mask=None, dtype=np.float32, masked=False
):
"""
Function to a load an image and normalised it (0 mean / 1 standard
deviation)
:param image_name: Path to the image to be noramlised
:param mask: Mask defining the region of interest
:param dtype: Data type for the final image
:param masked: Whether to mask the image or not
:return:
"""
image = load_nii(image_name).get_fdata().astype(dtype)
# If no mask is provided we use the image as a mask (all non-zero values)
if mask is None:
mask_bin = image.astype(np.bool)
else:
mask_bin = mask.astype(np.bool)
# Parameter estimation using the mask provided
image_mu = np.mean(image[mask_bin])
image_sigma = np.std(image[mask_bin])
norm_image = (image - image_mu) / image_sigma
if masked:
output = norm_image * mask_bin.astype(dtype)
else:
output = norm_image
return output
def test_scan(model,
test_x_data,
options,
save_nifti=True,
candidate_mask=None):
"""
Test data based on one model
Input:
- test_x_data: a nested dictionary containing training image paths:
train_x_data['scan_name']['modality'] = path_to_image_modality
- save_nifti: save image segmentation
- candidate_mask: a binary masks containing voxels to classify
Output:
- test_scan = Output image containing the probability output segmetnation
- If save_nifti --> Saves a nifti file at specified location
options['test_folder']/['test_scan']
"""
# get_scan name and create an empty nifti image to store segmentation
scans = list(test_x_data.keys())
flair_scans = [test_x_data[s]['FLAIR'] for s in scans]
flair_image = load_nii(flair_scans[0])
seg_image = np.zeros_like(flair_image.get_data().astype('float32'))
if candidate_mask is not None:
all_voxels = np.sum(candidate_mask)
else:
all_voxels = np.sum(flair_image.get_data() > 0)
if options['debug'] is True:
print("> DEBUG ", scans[0], "Voxels to classify:", all_voxels)
# compute lesion segmentation in batches of size options['batch_size']
batch, centers = load_test_patches(test_x_data, options['patch_size'],
options['batch_size'], candidate_mask)
if options['debug'] is True:
print("> DEBUG: testing current_batch:", batch.shape)
y_pred = model['net'].predict(np.squeeze(batch), options['batch_size'])
[x, y, z] = np.stack(centers, axis=1)
seg_image[x, y, z] = y_pred[:, 1]
if options['debug'] is True:
print("...done!")
# check if the computed volume is lower than the minimum accuracy given
# by the min_error parameter
if check_min_error(seg_image, options, flair_image.header.get_zooms()):
if options['debug']:
print("> DEBUG ", scans[0], "lesion volume below ",
options['min_error'], 'ml')
seg_image = np.zeros_like(flair_image.get_data().astype('float32'))
if save_nifti:
out_scan = nib.Nifti1Image(seg_image, affine=flair_image.affine)
out_scan.to_filename(
os.path.join(options['test_folder'], options['test_scan'],
options['experiment'], options['test_name']))
return seg_image
def get_patches_from_name(filename, centers, patch_size):
image = load_nii(filename).get_data()
patches = get_patches(image, centers,
patch_size) if len(patch_size) == 3 else [
get_patches2_5d(image, centers, patch_size)
]
return patches
def post_process_segmentation(image_folder, input_mask):
"""
doc
"""
filtered_mask = np.zeros_like(input_mask)
atlas = load_nii(
os.path.join(image_folder, 'tmp',
'MNI_subcortical_mask.nii.gz')).get_data()
for l in range(1, 15):
th_label = input_mask == l
labels, num_labels = ndimage.label(th_label)
label_list = np.unique(labels)
# filter candidates by size. Only storing the biggest one
num_elements_by_lesion = ndimage.labeled_comprehension(
np.logical_and(th_label, atlas), labels, label_list, np.sum, float,
0)
argmax = np.argmax(num_elements_by_lesion)
# assign voxels to output
current_voxels = np.stack(np.where(labels == argmax), axis=1)
filtered_mask[current_voxels[:, 0], current_voxels[:, 1],
current_voxels[:, 2]] = l
return filtered_mask
def generate_vfn_label():
root = '/data/lbw/structseg2019/data_ori/Task3_Thoracic_OAR'
# root = '/data/lbw/structseg2019/data_ori/re_labeling'
target = '/data/lbw/structseg2019/data_ori/trachea_vfn_data/label'
with open('./data_split/img_spacing.json', 'r') as f:
spacings = json.load(f)
with open('./data_split/label_range.json', 'r') as f:
label_range = json.load(f)
for i in range(1, 51):
# for i in [3, 5, 28, 32]:
print(i)
label_path = '{}/{}/label.nii.gz'.format(root, i)
label = np.uint8(load_nii(label_path).get_data())
label[label != 4] = 0
label[label == 4] = 1
this_range = label_range[str(i)]
# min_rl_h, min_rl_w, min_rl_t, max_rl_h, max_rl_w, max_rl_t = this_range['RightLung']
# min_ll_h, min_ll_w, min_ll_t, max_ll_h, max_ll_w, max_ll_t = this_range['LeftLung']
min_tr_h, min_tr_w, min_tr_t, max_tr_h, max_tr_w, max_tr_t = this_range[
'Trachea']
min_h, min_w, min_t, max_h, max_w, max_t = min_tr_h - 10, min_tr_w - 10, min_tr_t - 5, max_tr_h + 11, max_tr_w + 11, max_tr_t + 6
# min_h, min_w, max_h, max_w = min(min_rl_h, min_ll_h), min(min_rl_w, min_ll_w), max(max_rl_h, max_ll_h), max(
# max_rl_w, max_ll_w)
# min_t, max_t = min_he_t - 3, max_he_t + 4
label = label[min_h:max_h, min_w:max_w, min_t:max_t]
spacing = spacings[str(i)]['spacing']
label = equal_spacing(label, spacing, 1, mode='nearest')
target_path = os.path.join(target, 'label_{}'.format(i))
np.save(target_path, label)
def load_and_stack_iter2(names_lou, mask_names, roi_names, patch_size, old):
if old:
rois = load_thresholded_images_by_name(roi_names, threshold=0.5)
print(' Loading FLAIR images')
flair, y_train = load_patch_vectors_by_name(names_lou[0, :], mask_names, patch_size, rois)
print(' Loading PD images')
pd, _ = load_patch_vectors_by_name(names_lou[1, :], mask_names, patch_size, rois)
print(' Loading T2 images')
t2, _ = load_patch_vectors_by_name(names_lou[2, :], mask_names, patch_size, rois)
print(' Loading T1 images')
t1, _ = load_patch_vectors_by_name(names_lou[3, :], mask_names, patch_size, rois)
else:
pr_maps = [load_nii(roi_name).get_data() for roi_name in roi_names]
print(' Loading FLAIR images')
flair, y_train = load_patch_vectors_by_name_pr(names_lou[0, :], mask_names, patch_size, pr_maps)
print(' Loading PD images')
pd, _ = load_patch_vectors_by_name_pr(names_lou[1, :], mask_names, patch_size, pr_maps)
print(' Loading T2 images')
t2, _ = load_patch_vectors_by_name_pr(names_lou[2, :], mask_names, patch_size, pr_maps)
print(' Loading T1 images')
t1, _ = load_patch_vectors_by_name_pr(names_lou[3, :], mask_names, patch_size, pr_maps)
x_train = [np.stack(images, axis=1) for images in zip(*[flair, pd, t2, t1])]
return x_train, y_train
def load_test_patches(test_x_data,
patch_size,
batch_size,
voxel_candidates=None,
datatype=np.float32):
"""
Function generator to load test patches with size equal to patch_size,
given a list of selected voxels. Patches are returned in batches to reduce
the amount of RAM used
Inputs:
- x_data: list containing all subject image paths for a single modality
- selected_voxels: list where each element contains the subject binary
mask for selected voxels [len(x), len(y), len(z)]
- tuple containing patch size, either 2D (p1, p2, 1) or 3D (p1, p2, p3)
- Voxel candidates: a binary mask containing voxels for testing
Outputs (in batches):
- X: Train X data matrix for the each channel [num_samples, p1, p2, p3]
- voxel_coord: list of tuples with voxel coordinates (x,y,z) of
selected patches
"""
# get scan names and number of modalities used
scans = list(test_x_data.keys())
modalities = list(test_x_data[scans[0]].keys())
# load all image modalities and normalize intensities
images = []
for m in modalities:
raw_images = [load_nii(test_x_data[s][m]).get_data() for s in scans]
images.append([normalize_data(im) for im in raw_images])
# select voxels for testing. Discard CSF and darker WM in FLAIR.
# If voxel_candidates is not selected, using intensity > 0.5 in FLAIR,
# else use the binary mask to extract candidate voxels
if voxel_candidates is None:
flair_scans = [test_x_data[s]['FLAIR'] for s in scans]
selected_voxels = [
get_mask_voxels(mask)
for mask in select_training_voxels(flair_scans, 0.5)
][0]
else:
selected_voxels = get_mask_voxels(voxel_candidates)
# yield data for testing with size equal to batch_size
# for i in range(0, len(selected_voxels), batch_size):
# c_centers = selected_voxels[i:i+batch_size]
# X = []
# for m, image_modality in zip(modalities, images):
# X.append(get_patches(image_modality[0], c_centers, patch_size))
# yield np.stack(X, axis=1), c_centers
X = []
for image_modality in images:
X.append(get_patches(image_modality[0], selected_voxels, patch_size))
Xs = np.stack(X, axis=1)
return Xs, selected_voxels
def dump_spacing():
root = '/data/lbw/structseg2019/data_ori/Task3_Thoracic_OAR'
img_range = dict()
for i in range(1, 51):
print(i)
path = '{}/{}/data.nii.gz'.format(root, i)
img_range[i] = dict()
img = load_nii(path).get_data().astype(np.uint8)
spacing = load_nii(path).header['pixdim'][1:4]
img_range[i]['shape'] = (int(img.shape[0]), int(img.shape[1]),
int(img.shape[2]))
img_range[i]['spacing'] = (float(spacing[0]), float(spacing[1]),
float(spacing[2]))
with open('./data_split/img_spacing.json', 'w') as f:
json.dump(img_range, f)
def test_cascaded_model(model, test_x_data, options):
"""
Test the cascaded approach using a learned model
inputs:
- CNN model: a list containing the two cascaded CNN models
- test_x_data: a nested dictionary containing testing image paths:
test_x_data['scan_name']['modality'] = path_to_image_modality
- options: dictionary containing general hyper-parameters:
outputs:
- output_segmentation
"""
# print '> CNN: testing the model'
# organize experiments
exp_folder = os.path.join(options['test_folder'],
options['test_scan'],
options['experiment'])
if not os.path.exists(exp_folder):
os.mkdir(exp_folder)
# first network
options['test_name'] = options['experiment'] + '_debug_prob_0.nii.gz'
# only save the first iteration result if debug is True
save_nifti = True if options['debug'] is True else False
t1 = test_scan(model[0],
test_x_data,
options,
save_nifti=save_nifti)
# second network
options['test_name'] = options['experiment'] + '_prob_1.nii.gz'
t2 = test_scan(model[1],
test_x_data,
options,
save_nifti=True,
candidate_mask=(t1 > 0.5))
# postprocess the output segmentation
# obtain the orientation from the first scan used for testing
scans = test_x_data.keys()
flair_scans = [test_x_data[s]['FLAIR'] for s in scans]
flair_image = load_nii(flair_scans[0])
options['test_name'] = options['experiment'] + '_hard_seg.nii.gz'
out_segmentation = post_process_segmentation(t2,
options,
save_nifti=True,
orientation=flair_image.affine)
# return out_segmentation
return out_segmentation
def vox_generator(all_files, n_pos, n_neg, correction=False):
path = options["root_path"]
def get_filename(prefix):
return glob.glob(os.path.join(path, file, prefix))[0].split("/")[-1]
while 1:
for file in all_files:
if correction:
flair = load_nii(
os.path.join(path, file, file + "_flair_corrected.nii.gz")
).get_data()
t2 = load_nii(
os.path.join(path, file, file + "_t2_corrected.nii.gz")
).get_data()
t1 = load_nii(
os.path.join(path, file, file + "_t1_corrected.nii.gz")
).get_data()
t1ce = load_nii(
os.path.join(path, file, file + "_t1ce_corrected.nii.gz")
).get_data()
else:
flair = load_nii(
os.path.join(path, file, file + "_flair.nii.gz")
).get_data()
t2 = load_nii(os.path.join(path, file, file + "_t2.nii.gz")).get_data()
t1 = load_nii(os.path.join(path, file, file + "_t1.nii.gz")).get_data()
t1ce = load_nii(
os.path.join(path, file, file + "_t1ce.nii.gz")
).get_data()
data_norm = np.array([norm(flair), norm(t2), norm(t1), norm(t1ce)])
data_norm = np.transpose(data_norm, axes=[1, 2, 3, 0])
labels = load_nii(os.path.join(path, file, file + "_seg.nii.gz")).get_data()
foreground = np.array(np.where(labels > 0))
background = np.array(np.where((labels == 0) & (flair > 0)))
# n_pos = int(foreground.shape[1] * discount)
foreground = foreground[
:, np.random.permutation(foreground.shape[1])[:n_pos]
]
background = background[
:, np.random.permutation(background.shape[1])[:n_neg]
]
centers = np.concatenate((foreground, background), axis=1)
centers = centers[:, np.random.permutation(n_neg + n_pos)]
yield data_norm, labels, centers
def get_patient_roi_slice(path, p):
options = parse_inputs()
n_slices = options['n_slices']
# roi_sufix = '_seg.nii.gz' if not test else '.nii.gz'
roi_sufix = '.nii.gz'
roi = load_nii(os.path.join(path, p, p + roi_sufix)).get_data()
brain = load_nii(os.path.join(path, p, p + options['t1'])).get_data()
bounding_box_min = np.min(np.nonzero(brain), axis=1)
bounding_box_max = np.max(np.nonzero(brain), axis=1)
center_of_masses = np.mean(np.nonzero(roi), axis=1, dtype=np.int)
slices = [[
slice(bounding_box_min[0], bounding_box_max[0] + 1),
slice(bounding_box_min[1], bounding_box_max[1] + 1),
slice(center_of_masses[-1] - n_slices / 2, center_of_masses[-1] + n_slices / 2)
]]
return slices
def gen_test_data(p):
path = options['root_path']
flair = load_nii(os.path.join(path, p, p + '_flair.nii.gz')).get_data()
t2 = load_nii(os.path.join(path, p, p + '_t2.nii.gz')).get_data()
t1 = load_nii(os.path.join(path, p, p + '_t1.nii.gz')).get_data()
t1ce = load_nii(os.path.join(path, p, p + '_t1ce.nii.gz')).get_data()
data = np.array([flair, t2, t1, t1ce])
data = np.transpose(data, axes=[1, 2, 3, 0])
data_norm = np.array([norm(flair), norm(t2), norm(t1), norm(t1ce)])
data_norm = np.transpose(data_norm, axes=[1, 2, 3, 0])
labels = load_nii(os.path.join(path, p, p + '_seg.nii.gz')).get_data()
return data, data_norm, labels
def load_train_patches(x_data, y_data, selected_voxels, patch_size, random_state = 42, datatype=np.float32):
"""
Load train patches with size equal to patch_size, given a list of selected voxels
Inputs:
- x_data: list containing all subject image paths for a single modality
- y_data: list containing all subject image paths for the labels
- selected_voxels: list where each element contains the subject binary mask for selected voxels [len(x), len(y), len(z)]
- tuple containing patch size, either 2D (p1, p2, 1) or 3D (p1, p2, p3)
Outputs:
- X: Train X data matrix for the particular channel [num_samples, p1, p2, p3]
- Y: Train Y labels [num_samples, p1, p2, p3]
"""
# load images and normalize their intensties
images = [load_nii(name).get_data() for name in x_data]
images_norm = [(im.astype(dtype=datatype) - im[np.nonzero(im)].mean()) / im[np.nonzero(im)].std() for im in images]
# load labels
lesion_masks = [load_nii(name).get_data().astype(dtype=np.bool) for name in y_data]
nolesion_masks = [np.logical_and(np.logical_not(lesion), brain) for lesion, brain in zip(lesion_masks, selected_voxels)]
# Get all the x,y,z coordinates for each image
lesion_centers = [get_mask_voxels(mask) for mask in lesion_masks]
nolesion_centers = [get_mask_voxels(mask) for mask in nolesion_masks]
# load all positive samples (lesion voxels) and the same number of random negatives samples
np.random.seed(random_state)
x_pos_patches = [np.array(get_patches(image, centers, patch_size)) for image, centers in zip(images_norm, lesion_centers)]
y_pos_patches = [np.array(get_patches(image, centers, patch_size)) for image, centers in zip(lesion_masks, lesion_centers)]
indices = [np.random.permutation(range(0, len(centers1))).tolist()[:len(centers2)] for centers1, centers2 in zip(nolesion_centers, lesion_centers)]
nolesion_small = [itemgetter(*idx)(centers) for centers, idx in zip(nolesion_centers, indices)]
x_neg_patches = [np.array(get_patches(image, centers, patch_size)) for image, centers in zip(images_norm, nolesion_small)]
y_neg_patches = [np.array(get_patches(image, centers, patch_size)) for image, centers in zip(lesion_masks, nolesion_small)]
# concatenate positive and negative patches for each subject
X = np.concatenate([np.concatenate([x1, x2]) for x1, x2 in zip(x_pos_patches, x_neg_patches)], axis = 0)
Y = np.concatenate([np.concatenate([y1, y2]) for y1, y2 in zip(y_pos_patches, y_neg_patches)], axis= 0)
return X, Y
def mean_list(filename):
filelist = file_name(filename)
# img_arr = []
img_sum = []
img_list = []
for filename in filelist:
fsi = load_nii(filename).get_data()
img_arr = np.squeeze(fsi)
if len(img_list) == 0:
img_sum = img_arr
affine = load_nii(filename).affine
else:
img_sum += img_arr
img_list.append(img_arr)
img_mean = img_sum / len(img_list)
img_list = np.array(img_list)
print('img_list', img_list.shape)
print(img_mean.shape)
return img_mean, img_list, affine
def get_bounding_centers(image_names, patch_width, overlap=0, offset=0):
list_of_centers = map(
lambda names: get_bounding_blocks(
load_nii(names[0]).get_data(),
patch_width,
overlap=overlap,
offset=offset
),
image_names
)
return list_of_centers
def test():
path = '/media/lele/DATA/brain/Brats17TrainingData/HGG_test/Brats17_2013_10_1/'
gt = path + 'Brats17_2013_10_1_seg.nii.gz'
roi_nii = load_nii(gt)
roi = roi_nii.get_data().astype(dtype=np.bool)
centers = get_mask_voxels(roi)
test_samples = np.count_nonzero(roi)
image = np.zeros_like(roi).astype(dtype=np.uint8)
# print image
print test_samples
print image.shape
def check_image_list(patients_list, options):
sufix = get_sufix(options)
try:
for p in patients_list:
p_name, patient_path = get_patient_info(p)
outputname = os.path.join(patient_path,
'deep-' + p_name + sufix + 'brain.hdr')
np.squeeze(load_nii(outputname))
return True
except IOError:
return False
def generate_esophagus_data():
root = '/data/lbw/structseg2019/data_ori/Task3_Thoracic_OAR'
target_root = '/data/ybh/PublicDataSet/StructSeg2019/single_class_data/trachea'
target_img_root = os.path.join(target_root, 'img')
target_label_root = os.path.join(target_root, 'label')
if not os.path.exists(target_root):
os.mkdir(target_root)
if not os.path.exists(target_img_root):
os.mkdir(target_img_root)
if not os.path.exists(target_label_root):
os.mkdir(target_label_root)
for dir in os.listdir(root):
print(dir)
img_path = '{}/{}/data.nii.gz'.format(root, dir)
label_path = '{}/{}/label.nii.gz'.format(root, dir)
target_img_dir = os.path.join(target_img_root, dir)
target_label_dir = os.path.join(target_label_root, dir)
if not os.path.exists(target_img_dir):
os.mkdir(target_img_dir)
if not os.path.exists(target_label_dir):
os.mkdir(target_label_dir)
img = read_nii(img_path, 1180, -440)
label = load_nii(label_path).get_data().astype(np.uint8)
# rl_h, rl_w, rl_t = np.where(label == 1)
# ll_h, ll_w, ll_t = np.where(label == 2)
es_h, es_w, es_t = np.where(label == 5)
#
min_es_h, max_es_h, min_es_w, max_es_w, min_es_t, max_es_t = min(
es_h), max(es_h), min(es_w), max(es_w), min(es_t), max(es_t)
# # min_lung_h, max_lung_h = min(min(rl_h), min(ll_h)), max(max(rl_h), max(ll_h))
# # min_lung_w, max_lung_w = min(min(rl_w), min(ll_w)), max(max(rl_w), max(ll_w))
#
# img_crop = img[100: 420, 140: 410, min_es_t - 5: max_es_t + 6]
img_crop = img[min_es_h - 10:max_es_h + 11,
min_es_w - 10:max_es_w + 11, min_es_t - 5:max_es_t + 6]
label_crop = label[min_es_h - 10:max_es_h + 11,
min_es_w - 10:max_es_w + 11,
min_es_t - 5:max_es_t + 6]
# label_crop = label[100: 420, 140: 410, min_es_t - 5: max_es_t + 6]
# img_crop = img
# label[label != 5] = 0
# label[label == 5] = 1
# label_crop = label
print(img_crop.shape)
for i in range(img_crop.shape[2]):
cv2.imwrite(os.path.join(target_img_dir, '{}.png'.format(i)),
img_crop[:, :, i:i + 1])
cv2.imwrite(os.path.join(target_label_dir, '{}.png'.format(i)),
label_crop[:, :, i:i + 1])
def get_patches_list(list_of_image_names, centers_list, size):
patch_list = [
np.stack(
map(
lambda image: get_patches(norm(np.asarray(load_nii(image).dataobj)), centers, size),
image_names
),
axis=1
)
for image_names, centers in zip(list_of_image_names, centers_list) if centers
]
return patch_list
def load_patch_batch_percent(
image_names,
batch_size,
size,
defo_size=None,
d_names=None,
mask=None,
datatype=np.float32
):
images = [load_nii(name).get_data() for name in image_names]
defos = [load_nii(name).get_data() for name in d_names] if d_names is not None else []
images_norm = [(im - im[np.nonzero(im)].mean()) / im[np.nonzero(im)].std() for im in images]
defos_norm = [im / np.linalg.norm(im, axis=4).std() for im in defos]
mask = images[0].astype(np.bool) if mask is None else mask.astype(np.bool)
lesion_centers = get_mask_voxels(mask)
n_centers = len(lesion_centers)
for i in range(0, n_centers, batch_size):
centers = lesion_centers[i:i + batch_size]
x = get_image_patches(images_norm, centers, size).astype(dtype=datatype)
d = get_defo_patches(defos_norm, centers, size=defo_size) if defos else []
patches = (x, d) if defos else x
yield patches, centers, (100.0 * min((i + batch_size), n_centers)) / n_centers
def get_normalised_image(image_name, mask, dtype=np.float32, masked=False):
mask_bin = mask.astype(np.bool)
image = load_nii(image_name).get_data().astype(dtype)
image_mu = np.mean(image[mask_bin])
image_sigma = np.std(image[mask_bin])
norm_image = (image - image_mu) / image_sigma
if masked:
output = norm_image * mask_bin.astype(dtype)
else:
output = norm_image
return output
def vox_generator(all_files, n_pos, n_neg):
path = options['root_path']
while 1:
for file in all_files:
coll = glob.glob(os.path.join(path, file) + '/*')
for c in coll:
if 'flair.' in c or 'flair_corrected.' in c:
flair_path = c
if 't1.' in c or 't1_corrected.' in c:
t1_path = c
if 't2.' in c or 't2_corrected.' in c:
t2_path = c
if 't1ce.' in c or 't1ce_corrected.' in c:
t1ce_path = c
flair = load_nii(flair_path).get_data()
t2 = load_nii(t2_path).get_data()
t1 = load_nii(t1_path).get_data()
t1ce = load_nii(t1ce_path).get_data()
data_norm = np.array([norm(flair), norm(t2), norm(t1), norm(t1ce)])
data_norm = np.transpose(data_norm, axes=[1, 2, 3, 0])
labels = load_nii(os.path.join(path, file,
file + '_seg.nii.gz')).get_data()
foreground = np.array(np.where(labels > 0))
background = np.array(np.where((labels == 0) & (flair > 0)))
# n_pos = int(foreground.shape[1] * discount)
foreground = foreground[:,
np.random.permutation(foreground.shape[1]
)[:n_pos]]
background = background[:,
np.random.permutation(background.shape[1]
)[:n_neg]]
centers = np.concatenate((foreground, background), axis=1)
centers = centers[:, np.random.permutation(n_neg + n_pos)]
yield data_norm, labels, centers
def get_information_image(names, mask_name, patch_size, method='stdev'):
# Init
mask = load_nii(mask_name).get_data()
patch_size = patch_size if type(patch_size) is list else (patch_size,) * len(mask.shape)
patch_voxels = np.prod(patch_size)
n_images = len(names)
# ROI stuff
min_c = np.stack(np.nonzero(mask.astype(dtype=np.bool))).min(axis=1)
max_c = np.stack(np.nonzero(mask.astype(dtype=np.bool))).max(axis=1)
roi_shape = max_c - min_c
# Slicing
slicing = [slice(i, j) for i, j in zip(min_c, max_c)]
images = [load_nii(name).get_data() for name in names]
clipped_images = [np.pad(im[slicing], patch_size[0]/2, 'constant') for im in images]
n_slices = np.prod(roi_shape)
slices = np.zeros((n_slices, n_images, patch_voxels), dtype=np.int16)
for i, ci in enumerate(clipped_images):
slices[:, i, :] = get_rolling_patches(ci, patch_size).reshape((-1, patch_voxels))
slices.reshape((-1, n_images * patch_voxels))
min_s = np.amin(slices)
max_s = np.amax(slices)
hist = [np.histogram((x.astype(np.float32)-min_s)/(max_s-min_s), range=(0, 1))[0] for x in slices]
methods = {
'stdev': np.array([x.std() for x in slices]),
'entropy': np.array([entropy(x) * 512 for x in hist])
}
# Final image
im = np.zeros_like(mask, dtype=np.float32)
im[slicing] = methods[method].reshape(roi_shape)
im[np.logical_not(mask)] = 0
return im
def load_image_vectors(name, dir_name, min_shape, datatype=np.float32):
# Get the names of the images and load them
patients = [f for f in sorted(os.listdir(dir_name)) if os.path.isdir(os.path.join(dir_name, f))]
image_names = [os.path.join(dir_name, patient, name) for patient in patients]
images = [load_nii(image_name).get_data() for image_name in image_names]
# Reshape everything to have data of homogenous size (important for training)
# Also, normalize the data
if min_shape is None:
min_shape = min([im.shape for im in images])
data = np.asarray(
[nd.zoom((im - im.mean()) / im.std(),
[float(min_shape[0]) / im.shape[0], float(min_shape[1]) / im.shape[1],
float(min_shape[2]) / im.shape[2]]) for im in images]
)
return data.astype(datatype), image_names
def reshape_to_nifti(image, original_name):
# Open the original nifti
original = load_nii(original_name).get_data()
# Reshape the image and save it
reshaped = nd.zoom(
image,
[
float(original.shape[0]) / image.shape[0],
float(original.shape[1]) / image.shape[1],
float(original.shape[2]) / image.shape[2]
]
)
reshaped *= original.std()
reshaped += original.mean()
reshaped_nii = NiftiImage(reshaped, affine=np.eye(4))
return reshaped_nii
def get_cnn_rois(names, mask_names, roi_names=None, pr_names=None, th=1.0, balanced=True):
rois = load_thresholded_norm_images_by_name(
names[0, :],
threshold=th,
mask_names=roi_names
) if roi_names is not None else load_masks(names)
if pr_names is not None:
pr_maps = [load_nii(name).get_data() * roi for name, roi in izip(pr_names, rois)]
if balanced:
idx_sorted_maps = [np.argsort(pr_map * np.logical_not(lesion_mask), axis=None)
for pr_map, lesion_mask in izip(pr_maps, load_masks(mask_names))]
rois_n = [idx.reshape(lesion_mask.shape) > (idx.shape[0] - np.sum(lesion_mask) - 1)
for idx, lesion_mask in izip(idx_sorted_maps, load_masks(mask_names))]
else:
rois_n = [np.logical_and(np.logical_not(lesion_mask), pr_map > 0.5)
for pr_map, lesion_mask in izip(pr_maps, load_masks(mask_names))]
else:
rois_n = [np.logical_and(np.logical_not(lesion), brain)
for lesion, brain in izip(load_masks(mask_names), rois)]
rois_p = list(load_masks(mask_names))
return rois_p, rois_n
def main():
options = parse_inputs()
c = color_codes()
# Prepare the net architecture parameters
register = options['register']
multi = options['multi']
defo = options['deformation']
layers = ''.join(options['layers'])
greenspan = options['greenspan']
freeze = options['freeze']
balanced = options['balanced'] if not freeze else False
# Prepare the net hyperparameters
epochs = options['epochs']
padding = options['padding']
patch_width = options['patch_width']
patch_size = (32, 32) if greenspan else (patch_width, patch_width, patch_width)
pool_size = options['pool_size']
batch_size = options['batch_size']
dense_size = options['dense_size']
conv_blocks = options['conv_blocks']
n_filters = options['number_filters']
n_filters = n_filters if len(n_filters) > 1 else n_filters*conv_blocks
conv_width = options['conv_width']
conv_size = conv_width if isinstance(conv_width, list) else [conv_width]*conv_blocks
# Prepare the sufix that will be added to the results for the net and images
use_flair = options['use_flair']
use_pd = options['use_pd']
use_t2 = options['use_t2']
flair_name = 'flair' if use_flair else None
pd_name = 'pd' if use_pd else None
t2_name = 't2' if use_t2 else None
images = filter(None, [flair_name, pd_name, t2_name])
reg_s = '.reg' if register else ''
filters_s = 'n'.join(['%d' % nf for nf in n_filters])
conv_s = 'c'.join(['%d' % cs for cs in conv_size])
im_s = '.'.join(images)
mc_s = '.mc' if multi else ''
d_s = 'd%d.' % (conv_blocks*2+defo) if defo else ''
sufix = '.greenspan' if greenspan else '%s.%s%s%s.p%d.c%s.n%s.d%d.e%d.pad_%s' %\
(mc_s, d_s, im_s, reg_s, patch_width, conv_s, filters_s, dense_size, epochs, padding)
# Prepare the data names
mask_name = options['mask']
wm_name = options['wm_mask']
sub_folder = options['sub_folder']
sub_name = options['flair_sub']
dir_name = options['dir_name']
patients = [f for f in sorted(os.listdir(dir_name))
if os.path.isdir(os.path.join(dir_name, f))]
n_patients = len(patients)
names = get_names_from_path(dir_name, options, patients)
defo_names = get_defonames_from_path(dir_name, options, patients) if defo else None
defo_width = conv_blocks*2+defo if defo else None
defo_size = (defo_width, defo_width, defo_width)
# Random initialisation
seed = np.random.randint(np.iinfo(np.int32).max)
# Metrics output
metrics_file = os.path.join(dir_name, 'metrics' + sufix)
with open(metrics_file, 'w') as f:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + 'Starting leave-one-out' + c['nc'])
# Leave-one-out main loop (we'll do 2 training iterations with testing for each patient)
for i in range(0, n_patients):
# Prepare the data relevant to the leave-one-out (subtract the patient from the dataset and set the path)
# Also, prepare the network
case = patients[i]
path = os.path.join(dir_name, case)
names_lou = np.concatenate([names[:, :i], names[:, i + 1:]], axis=1)
defo_names_lou = np.concatenate([defo_names[:, :i], defo_names[:, i + 1:]], axis=1) if defo else None
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['nc'] + 'Patient ' + c['b'] + case + c['nc'] +
c['g'] + ' (%d/%d)' % (i+1, n_patients))
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<Running iteration ' + c['b'] + '1' + c['nc'] + c['g'] + '>' + c['nc'])
net_name = os.path.join(path, 'deep-longitudinal.init' + sufix + '.')
if greenspan:
net = create_cnn_greenspan(
input_channels=names.shape[0]/2,
patience=25,
name=net_name,
epochs=500
)
images = ['axial', 'coronal', 'sagital']
else:
if multi:
net = create_cnn3d_det_string(
cnn_path=layers,
input_shape=(None, names.shape[0], patch_width, patch_width, patch_width),
convo_size=conv_size,
padding=padding,
dense_size=dense_size,
pool_size=2,
number_filters=n_filters,
patience=10,
multichannel=True,
name=net_name,
epochs=100
)
else:
net = create_cnn3d_longitudinal(
convo_blocks=conv_blocks,
input_shape=(None, names.shape[0], patch_width, patch_width, patch_width),
images=images,
convo_size=conv_size,
pool_size=pool_size,
dense_size=dense_size,
number_filters=n_filters,
padding=padding,
drop=0.5,
register=register,
defo=defo,
patience=10,
name=net_name,
epochs=100
)
names_test = get_names_from_path(path, options)
defo_names_test = get_defonames_from_path(path, options) if defo else None
outputname1 = os.path.join(path, 't' + case + sufix + '.iter1.nii.gz') if not greenspan else os.path.join(
path, 't' + case + sufix + '.nii.gz')
# First we check that we did not train for that patient, in order to save time
try:
net.load_params_from(net_name + 'model_weights.pkl')
except IOError:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' +
c['g'] + 'Loading the data for ' + c['b'] + 'iteration 1' + c['nc'])
# Data loading. Most of it is based on functions from data_creation that load the data.
# But we also need to prepare the name list to load the leave-one-out data.
paths = [os.path.join(dir_name, p) for p in np.concatenate([patients[:i], patients[i+1:]])]
mask_names = [os.path.join(p_path, mask_name) for p_path in paths]
wm_names = [os.path.join(p_path, wm_name) for p_path in paths]
pr_names = [os.path.join(p_path, sub_folder, sub_name) for p_path in paths]
x_train, y_train = load_lesion_cnn_data(
names=names_lou,
mask_names=mask_names,
defo_names=defo_names_lou,
roi_names=wm_names,
pr_names=pr_names,
patch_size=patch_size,
defo_size=defo_size,
random_state=seed
)
# Afterwards we train. Check the relevant training function.
if greenspan:
x_train = np.swapaxes(x_train, 1, 2)
train_greenspan(net, x_train, y_train, images)
else:
train_net(net, x_train, y_train, images)
with open(net_name + 'layers.pkl', 'wb') as fnet:
pickle.dump(net.layers, fnet, -1)
# Then we test the net. Again we save time by checking if we already tested that patient.
try:
image_nii = load_nii(outputname1)
image1 = image_nii.get_data()
except IOError:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<Creating the probability map ' + c['b'] + '1' + c['nc'] + c['g'] + '>' + c['nc'])
image_nii = load_nii(os.path.join(path, options['image_folder'], options['flair_f']))
mask_nii = load_nii(os.path.join(path, wm_name))
if greenspan:
image1 = test_greenspan(
net,
names_test,
mask_nii.get_data(),
batch_size,
patch_size,
image_nii.get_data().shape,
images
)
else:
image1 = test_net(
net,
names_test,
mask_nii.get_data(),
batch_size,
patch_size,
defo_size,
image_nii.get_data().shape,
images,
defo_names_test
)
image_nii.get_data()[:] = image1
image_nii.to_filename(outputname1)
if greenspan:
# Since Greenspan did not use two iterations, we must get the final mask here.
outputname_final = os.path.join(path, 't' + case + sufix + '.final.nii.gz')
mask_nii.get_data()[:] = (image1 > 0.5).astype(dtype=np.int8)
mask_nii.to_filename(outputname_final)
else:
# If not, we test the net with the training set to look for misclassified negative with a high
# probability of being positives according to the net.
# These voxels will be the input of the second training iteration.
''' Here we get the seeds '''
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' +
c['g'] + '<Looking for seeds for the final iteration>' + c['nc'])
patients_names = zip(np.rollaxis(names_lou, 1), np.rollaxis(defo_names_lou, 1)) if defo\
else np.rollaxis(names_lou, 1)
for patient in patients_names:
if defo:
patient, d_patient = patient
else:
d_patient = None
patient_path = '/'.join(patient[0].rsplit('/')[:-1])
outputname = os.path.join(patient_path, 't' + case + sufix + '.nii.gz')
mask_nii = load_nii(os.path.join('/'.join(patient[0].rsplit('/')[:-3]), wm_name))
try:
load_nii(outputname)
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' +
c['g'] + ' Patient ' + patient[0].rsplit('/')[-4] + ' already done' + c['nc'])
except IOError:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' +
c['g'] + ' Testing with patient ' + c['b'] + patient[0].rsplit('/')[-4] + c['nc'])
image_nii = load_nii(patient[0])
image = test_net(
net,
patient,
mask_nii.get_data(),
batch_size,
patch_size,
defo_size,
image_nii.get_data().shape,
images,
d_patient
)
print(c['g'] + ' -- Saving image ' + c['b'] + outputname + c['nc'])
image_nii.get_data()[:] = image
image_nii.to_filename(outputname)
''' Here we perform the last iteration '''
# Finally we perform the final iteration. After refactoring the code, the code looks almost exactly
# the same as the training of the first iteration.
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<Running iteration ' + c['b'] + '2' + c['nc'] + c['g'] + '>' + c['nc'])
f_s = '.f' if freeze else ''
ub_s = '.ub' if not balanced else ''
final_s = f_s + ub_s
outputname2 = os.path.join(path, 't' + case + final_s + sufix + '.iter2.nii.gz')
net_name = os.path.join(path, 'deep-longitudinal.final' + final_s + sufix + '.')
if multi:
net = create_cnn3d_det_string(
cnn_path=layers,
input_shape=(None, names.shape[0], patch_width, patch_width, patch_width),
convo_size=conv_size,
padding=padding,
pool_size=2,
dense_size=dense_size,
number_filters=n_filters,
patience=50,
multichannel=True,
name=net_name,
epochs=epochs
)
else:
if not freeze:
net = create_cnn3d_longitudinal(
convo_blocks=conv_blocks,
input_shape=(None, names.shape[0], patch_width, patch_width, patch_width),
images=images,
convo_size=conv_size,
pool_size=pool_size,
dense_size=dense_size,
number_filters=n_filters,
padding=padding,
drop=0.5,
register=register,
defo=defo,
patience=50,
name=net_name,
epochs=epochs
)
else:
net.max_epochs = epochs
net.on_epoch_finished[0].name = net_name + 'model_weights.pkl'
for layer in net.get_all_layers():
if not isinstance(layer, DenseLayer):
for param in layer.params:
layer.params[param].discard('trainable')
try:
net.load_params_from(net_name + 'model_weights.pkl')
except IOError:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' +
c['g'] + 'Loading the data for ' + c['b'] + 'iteration 2' + c['nc'])
roi_paths = ['/'.join(name.rsplit('/')[:-1]) for name in names_lou[0, :]]
paths = [os.path.join(dir_name, p) for p in np.concatenate([patients[:i], patients[i + 1:]])]
ipr_names = [os.path.join(p_path, sub_folder, sub_name) for p_path in paths] if freeze else None
pr_names = [os.path.join(p_path, 't' + case + sufix + '.nii.gz') for p_path in roi_paths]
mask_names = [os.path.join(p_path, mask_name) for p_path in paths]
wm_names = [os.path.join(p_path, wm_name) for p_path in paths]
x_train, y_train = load_lesion_cnn_data(
names=names_lou,
mask_names=mask_names,
defo_names=defo_names_lou,
roi_names=wm_names,
init_pr_names=ipr_names,
pr_names=pr_names,
patch_size=patch_size,
defo_size=defo_size,
random_state=seed,
balanced=balanced
)
train_net(net, x_train, y_train, images)
with open(net_name + 'layers.pkl', 'wb') as fnet:
pickle.dump(net.layers, fnet, -1)
try:
image_nii = load_nii(outputname2)
image2 = image_nii.get_data()
except IOError:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<Creating the probability map ' + c['b'] + '2' + c['nc'] + c['g'] + '>' + c['nc'])
image_nii = load_nii(os.path.join(path, options['image_folder'], options['flair_f']))
mask_nii = load_nii(os.path.join(path, wm_name))
image2 = test_net(
net,
names_test,
mask_nii.get_data(),
batch_size,
patch_size,
defo_size,
image_nii.get_data().shape,
images,
defo_names_test
)
image_nii.get_data()[:] = image2
image_nii.to_filename(outputname2)
image = image1 * image2
image_nii.get_data()[:] = image
outputname_mult = os.path.join(path, 't' + case + final_s + sufix + '.iter1_x_2.nii.gz')
image_nii.to_filename(outputname_mult)
image = (image1 * image2) > 0.5
image_nii.get_data()[:] = image
outputname_final = os.path.join(path, 't' + case + final_s + sufix + '.final.nii.gz')
image_nii.to_filename(outputname_final)
# Finally we compute some metrics that are stored in the metrics file defined above.
# I plan on replicating Challenge's 2008 evaluation measures here.
gt = load_nii(os.path.join(path, mask_name)).get_data().astype(dtype=np.bool)
seg1 = image1 > 0.5
if not greenspan:
seg2 = image2 > 0.5
dsc1 = dsc_seg(gt, seg1)
if not greenspan:
dsc2 = dsc_seg(gt, seg2)
if not greenspan:
dsc_final = dsc_seg(gt, image)
else:
dsc_final = dsc1
tpf1 = tp_fraction_seg(gt, seg1)
if not greenspan:
tpf2 = tp_fraction_seg(gt, seg2)
if not greenspan:
tpf_final = tp_fraction_seg(gt, image)
fpf1 = fp_fraction_seg(gt, seg1)
if not greenspan:
fpf2 = fp_fraction_seg(gt, seg2)
if not greenspan:
fpf_final = fp_fraction_seg(gt, image)
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<DSC ' + c['c'] + case + c['g'] + ' = ' + c['b'] + str(dsc_final) + c['nc'] + c['g'] + '>' + c['nc'])
f.write('%s;Test 1; %f;%f;%f\n' % (case, dsc1, tpf1, fpf1))
if not greenspan:
f.write('%s;Test 2; %f;%f;%f\n' % (case, dsc2, tpf2, fpf2))
if not greenspan:
f.write('%s;Final; %f;%f;%f\n' % (case, dsc_final, tpf_final, fpf_final))
def main():
# Parse command line options
parser = argparse.ArgumentParser(description='Test different nets with 3D data.')
parser.add_argument('-f', '--folder', dest='folder', default='/home/mariano/DATA/Challenge2016/',
help="read data from FOLDER")
parser.add_argument('-v', '--verbose', action='store_true', dest='verbose', default=False)
parser.add_argument('-p', '--patch-size', action='store', type=int, nargs='+',
dest='patch_size', default=(15, 15, 15))
parser.add_argument('--use-gado', action='store_true', dest='use_gado')
parser.add_argument('--no-gado', action='store_false', dest='use_gado', default=False)
parser.add_argument('--gado', action='store', dest='gado', default='GADO_preprocessed.nii.gz')
parser.add_argument('--use-flair', action='store_true', dest='use_flair')
parser.add_argument('--no-flair', action='store_false', dest='use_flair', default=True)
parser.add_argument('--flair', action='store', dest='flair', default='FLAIR_preprocessed.nii.gz')
parser.add_argument('--use-pd', action='store_true', dest='use_pd')
parser.add_argument('--no-pd', action='store_false', dest='use_pd', default=True)
parser.add_argument('--pd', action='store', dest='pd', default='DP_preprocessed.nii.gz')
parser.add_argument('--use-t2', action='store_true', dest='use_t2')
parser.add_argument('--no-t2', action='store_false', dest='use_t2', default=True)
parser.add_argument('--t2', action='store', dest='t2', default='T2_preprocessed.nii.gz')
parser.add_argument('--use-t1', action='store_true', dest='use_t1')
parser.add_argument('--no-t1', action='store_false', dest='use_t1', default=True)
parser.add_argument('--t1', action='store', dest='t1', default='T1_preprocessed.nii.gz')
parser.add_argument('--mask', action='store', dest='mask', default='Consensus.nii.gz')
options = vars(parser.parse_args())
dir_name = options['folder']
files = sorted(os.listdir(dir_name))
patients = [f for f in files if os.path.isdir(os.path.join(dir_name, f))]
n_patients = len(patients)
for patient, i in zip(patients, range(n_patients)):
patient_folder = os.path.join(dir_name, patient)
print('\033[36m[' + strftime("%H:%M:%S") + '] \033[0mPatient \033[1m' + patient +
'\033[0m\033[32m (%d/%d)\033[0m' % (i + 1, n_patients))
mask_nii = load_nii(os.path.join(patient_folder, options['mask']))
mask_img = mask_nii.get_data()
lesion_centers = get_mask_voxels(mask_img)
flair = None
pd = None
t1 = None
t2 = None
gado = None
patch_size = tuple(options['patch_size'])
if options['use_flair']:
flair = get_patches_from_name(os.path.join(patient_folder, options['flair']),
lesion_centers,
patch_size
)
if options['use_pd']:
pd = get_patches_from_name(os.path.join(patient_folder, options['pd']),
lesion_centers,
patch_size
)
if options['use_t1']:
t1 = get_patches_from_name(os.path.join(patient_folder, options['t1']),
lesion_centers,
patch_size
)
if options['use_t2']:
t2 = get_patches_from_name(os.path.join(patient_folder, options['t2']),
lesion_centers,
patch_size
)
if options['use_gado']:
gado = get_patches_from_name(os.path.join(patient_folder, options['gado']),
lesion_centers,
patch_size
)
patches = np.stack([np.array(data) for data in [flair, pd, t2, gado, t1] if data is not None], axis=1)
print 'Our final vector\'s size = (' + ','.join([str(num) for num in patches.shape]) + ')'
def get_patches_from_name(filename, centers, patch_size):
image = load_nii(filename).get_data()
patches = get_patches(image, centers, patch_size) if len(patch_size) == 3 \
else [get_patches2_5d(image, centers, patch_size)]
return patches
def load_thresholded_images(name, dir_name, threshold=2.0, datatype=np.float32):
patients = [f for f in sorted(os.listdir(dir_name)) if os.path.isdir(os.path.join(dir_name, f))]
image_names = [os.path.join(dir_name, patient, name) for patient in patients]
images = [load_nii(image_name).get_data() for image_name in image_names]
rois = [image.astype(dtype=datatype) > threshold for image in images]
return rois
def main():
c = color_codes()
patch_size = (15, 15, 15)
dir_name = '/home/sergivalverde/w/CNN/images/CH16'
patients = [f for f in sorted(os.listdir(dir_name)) if os.path.isdir(os.path.join(dir_name, f))]
names = np.stack([name for name in [
[os.path.join(dir_name, patient, 'FLAIR_preprocessed.nii.gz') for patient in patients],
[os.path.join(dir_name, patient, 'DP_preprocessed.nii.gz') for patient in patients],
[os.path.join(dir_name, patient, 'T2_preprocessed.nii.gz') for patient in patients],
[os.path.join(dir_name, patient, 'T1_preprocessed.nii.gz') for patient in patients]
] if name is not None], axis=1)
seed = np.random.randint(np.iinfo(np.int32).max)
''' Here we create an initial net to find conflictive voxels '''
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] + '<Running iteration ' + c['b'] + '1>' + c['nc'])
net_name = '/home/sergivalverde/w/CNN/code/CNN1/miccai_challenge2016/deep-challenge2016.init.'
net = NeuralNet(
layers=[
(InputLayer, dict(name='in', shape=(None, 4, 15, 15, 15))),
(Conv3DDNNLayer, dict(name='conv1_1', num_filters=32, filter_size=(5, 5, 5), pad='same')),
(Pool3DDNNLayer, dict(name='avgpool_1', pool_size=2, stride=2, mode='average_inc_pad')),
(Conv3DDNNLayer, dict(name='conv2_1', num_filters=64, filter_size=(5, 5, 5), pad='same')),
(Pool3DDNNLayer, dict(name='avgpool_2', pool_size=2, stride=2, mode='average_inc_pad')),
(DropoutLayer, dict(name='l2drop', p=0.5)),
(DenseLayer, dict(name='l1', num_units=256)),
(DenseLayer, dict(name='out', num_units=2, nonlinearity=nonlinearities.softmax)),
],
objective_loss_function=objectives.categorical_crossentropy,
update=updates.adam,
update_learning_rate=0.0001,
on_epoch_finished=[
SaveWeights(net_name + 'model_weights.pkl', only_best=True, pickle=False),
SaveTrainingHistory(net_name + 'model_history.pkl'),
PlotTrainingHistory(net_name + 'training_history.png'),
EarlyStopping(patience=10)
],
verbose=10,
max_epochs=50,
train_split=TrainSplit(eval_size=0.25),
custom_scores=[('dsc', lambda p, t: 2 * np.sum(p * t[:, 1]) / np.sum((p + t[:, 1])))],
)
try:
net.load_params_from(net_name + 'model_weights.pkl')
except IOError:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' +
c['g'] + 'Loading the data for ' + c['b'] + 'iteration 1' + c['nc'])
# Create the data
(x, y, _) = load_patches(
dir_name=dir_name,
use_flair=True,
use_pd=True,
use_t2=True,
use_t1=True,
use_gado=False,
flair_name='FLAIR_preprocessed.nii.gz',
pd_name='DP_preprocessed.nii.gz',
t2_name='T2_preprocessed.nii.gz',
t1_name='T1_preprocessed.nii.gz',
gado_name=None,
mask_name='Consensus.nii.gz',
size=patch_size
)
print('-- Permuting the data')
np.random.seed(seed)
x_train = np.random.permutation(np.concatenate(x).astype(dtype=np.float32))
print('-- Permuting the labels')
np.random.seed(seed)
y_train = np.random.permutation(np.concatenate(y).astype(dtype=np.int32))
y_train = y_train[:, y_train.shape[1] / 2 + 1, y_train.shape[2] / 2 + 1, y_train.shape[3] / 2 + 1]
print('-- Training vector shape = (' + ','.join([str(length) for length in x_train.shape]) + ')')
print('-- Training labels shape = (' + ','.join([str(length) for length in y_train.shape]) + ')')
print c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +\
'Training (' + c['b'] + 'initial' + c['nc'] + c['g'] + ')' + c['nc']
# We try to get the last weights to keep improving the net over and over
net.fit(x_train, y_train)
''' Here we get the seeds '''
print c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] + '<Looking for seeds>' + c['nc']
for patient in names:
output_name = os.path.join('/'.join(patient[0].rsplit('/')[:-1]), 'test.iter1.nii.gz')
try:
load_nii(output_name)
print c['c'] + '[' + strftime("%H:%M:%S") + '] ' \
+ c['g'] + '-- Patient ' + patient[0].rsplit('/')[-2] + ' already done' + c['nc']
except IOError:
print c['c'] + '[' + strftime("%H:%M:%S") + '] '\
+ c['g'] + '-- Testing with patient ' + c['b'] + patient[0].rsplit('/')[-2] + c['nc']
image_nii = load_nii(patient[0])
image = np.zeros_like(image_nii.get_data())
for batch, centers in load_patch_batch(patient, 100000, patch_size):
y_pred = net.predict_proba(batch)
[x, y, z] = np.stack(centers, axis=1)
image[x, y, z] = y_pred[:, 1]
print c['g'] + '-- Saving image ' + c['b'] + output_name + c['nc']
image_nii.get_data()[:] = image
image_nii.to_filename(output_name)
''' Here we perform the last iteration '''
print c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] + '<Running iteration ' + c['b'] + '2>' + c['nc']
net_name = '/home/sergivalverde/w/CNN/code/CNN1/miccai_challenge2016/deep-challenge2016.final.'
net = NeuralNet(
layers=[
(InputLayer, dict(name='in', shape=(None, 4, 15, 15, 15))),
(Conv3DDNNLayer, dict(name='conv1_1', num_filters=32, filter_size=(5, 5, 5), pad='same')),
(Pool3DDNNLayer, dict(name='avgpool_1', pool_size=2, stride=2, mode='average_inc_pad')),
(Conv3DDNNLayer, dict(name='conv2_1', num_filters=64, filter_size=(5, 5, 5), pad='same')),
(Pool3DDNNLayer, dict(name='avgpool_2', pool_size=2, stride=2, mode='average_inc_pad')),
(DropoutLayer, dict(name='l2drop', p=0.5)),
(DenseLayer, dict(name='l1', num_units=256)),
(DenseLayer, dict(name='out', num_units=2, nonlinearity=nonlinearities.softmax)),
],
objective_loss_function=objectives.categorical_crossentropy,
update=updates.adam,
update_learning_rate=0.0001,
on_epoch_finished=[
SaveWeights(net_name + 'model_weights.pkl', only_best=True, pickle=False),
SaveTrainingHistory(net_name + 'model_history.pkl'),
PlotTrainingHistory(net_name + 'training_history.png'),
],
batch_iterator_train=BatchIterator(batch_size=4096),
verbose=10,
max_epochs=2000,
train_split=TrainSplit(eval_size=0.25),
custom_scores=[('dsc', lambda p, t: 2 * np.sum(p * t[:, 1]) / np.sum((p + t[:, 1])))],
)
try:
net.load_params_from(net_name + 'model_weights.pkl')
except IOError:
pass
print c['c'] + '[' + strftime("%H:%M:%S") + '] '\
+ c['g'] + 'Loading the data for ' + c['b'] + 'iteration 2' + c['nc']
(x, y, names) = load_patches(
dir_name='/home/sergivalverde/w/CNN/images/CH16',
use_flair=True,
use_pd=True,
use_t2=True,
use_t1=True,
use_gado=False,
flair_name='FLAIR_preprocessed.nii.gz',
pd_name='DP_preprocessed.nii.gz',
t2_name='T2_preprocessed.nii.gz',
gado_name=None,
t1_name='T1_preprocessed.nii.gz',
mask_name='Consensus.nii.gz',
size=patch_size,
roi_name='test.iter1.nii.gz'
)
print '-- Permuting the data'
np.random.seed(seed)
x_train = np.random.permutation(np.concatenate(x).astype(dtype=np.float32))
print '-- Permuting the labels'
np.random.seed(seed)
y_train = np.random.permutation(np.concatenate(y).astype(dtype=np.int32))
y_train = y_train[:, y_train.shape[1] / 2 + 1, y_train.shape[2] / 2 + 1, y_train.shape[3] / 2 + 1]
print '-- Training vector shape = (' + ','.join([str(length) for length in x_train.shape]) + ')'
print '-- Training labels shape = (' + ','.join([str(length) for length in y_train.shape]) + ')'
print c['c'] + '[' + strftime("%H:%M:%S") + '] '\
+ c['g'] + 'Training (' + c['b'] + 'final' + c['nc'] + c['g'] + ')' + c['nc']
net.fit(x_train, y_train)
def load_thresholded_norm_images_by_name(image_names, mask_names=None, threshold=2.0):
masks = [load_nii(mask).get_data() for mask in mask_names] if mask_names else None
return threshold_image_list(norm_image_generator(image_names), threshold, masks)
def norm_image_generator(image_names):
for name in image_names:
im = load_nii(name).get_data()
yield (im - im[np.nonzero(im)].mean()) / im[np.nonzero(im)].std()
def main():
@contextlib.contextmanager
def dummy_file():
yield None
# Parse command line options
parser = argparse.ArgumentParser(description='Test different nets with 3D data.')
group_in = parser.add_mutually_exclusive_group()
group_out = parser.add_mutually_exclusive_group()
folder_help = 'Folder with the files to evaluate. Remember to include a init_names.py for the evaluation pairs.'
files_help = 'Pair of files to be compared. The first is the GT and the second the file you wnat to evaluate.'
group_in.add_argument('-f', '--folder', help=folder_help)
group_in.add_argument('--files', nargs=2, help=files_help)
general_help = 'General evaluation. Based on the 2008 MS challenge''s measures. ' \
'Volume and detection absolute measures are also included for all the images.'
sizes_help = 'Evaluation based on region sizes. ' \
'Includes TPF, FPF and DSC for detection and DSC for segmentation. ' \
'The size of TP is determined by the GT size, while the FP size is determined by the FP lesion.'
group_out.add_argument('-g', '--general', action='store_true', help=general_help)
group_out.add_argument('-s', '--sizes', dest='sizes', nargs='+', type=int, default=[3, 11, 51], help=sizes_help)
args = parser.parse_args()
if args.folder:
folder_name = args.folder
sys.path = sys.path + [folder_name]
from init_names import get_names_from_folder
gt_names, all_names = get_names_from_folder(folder_name)
elif args.files:
folder_name = os.getcwd()
gt_names = [args.files[0]]
all_names = [[args.files[1]]]
if args.general:
print('\033[32;1mGeneral\033[0m\033[32m evaluation\033[0m')
results_name = 'results.g.csv'
else:
print('\033[32;1mLesion size\033[0m\033[32m evaluation\033[0m')
sizes = args.sizes
results_name = 'results.s%s.csv' % '.'.join(['%d' % s for s in sizes])
with open(os.path.join(folder_name, results_name), 'w') if args.folder else dummy_file() as f:
for gt_name, names in zip(gt_names, all_names):
print('\033[32mEvaluating with ground truth \033[32;1m' + gt_name + '\033[0m')
gt_nii = load_nii(gt_name)
gt = gt_nii.get_data()
spacing = dict(gt_nii.header.items())['pixdim'][1:4]
for name in names:
name = ''.join(name)
print('\033[32m-- vs \033[32;1m' + name + '\033[0m')
lesion = load_nii(name).get_data()
if args.general:
dist = average_surface_distance(gt, lesion, spacing)
tpfv = tp_fraction_seg(gt, lesion)
fpfv = fp_fraction_seg(gt, lesion)
dscv = dsc_seg(gt, lesion)
tpfl = tp_fraction_det(gt, lesion)
fpfl = fp_fraction_det(gt, lesion)
dscl = dsc_det(gt, lesion)
tp = true_positive_det(lesion, gt)
gt_d = num_regions(gt)
lesion_s = num_voxels(lesion)
gt_s = num_voxels(gt)
pdsc = probabilistic_dsc_seg(gt, lesion)
if f:
measures = (gt_name, name, dist, tpfv, fpfv, dscv, tpfl, fpfl, dscl, tp, gt_d, lesion_s, gt_s)
f.write('%s;%s;%f;%f;%f;%f;%f;%f;%f;%d;%d;%d;%d\n' % measures)
else:
measures = (dist, tpfv, fpfv, dscv, tpfl, fpfl, dscl, tp, gt_d, lesion_s, gt_s, pdsc)
print('SurfDist TPFV FPFV DSCV TPFL FPFL DSCL TPL GTL Voxels GTV PrDSC')
print('%f %f %f %f %f %f %f %d %d %d %d %f' % measures)
else:
tpf, fpf, dscd, dscs = analysis_by_sizes(gt, lesion, sizes)
names = '%s;%s;' % (gt_name, name)
measures = ';'.join(['%f;%f;%f;%f' % (tpf_i, fpf_i, dscd_i, dscs_i)
for tpf_i, fpf_i, dscd_i, dscs_i in zip(tpf, fpf, dscd, dscs)])
if f:
f.write(names + measures + '\n')
else:
intervals = ['\t\t[%d-%d)\t\t|' % (mins, maxs) for mins, maxs in zip(sizes[:-1], sizes[1:])]
intervals = ''.join(intervals) + '\t\t[%d-inf)\t|' % sizes[-1]
measures_s = 'TPF\tFPF\tDSCd\tDSCs\t|' * len(sizes)
measures = ''.join(['%.2f\t%.2f\t%.2f\t%.2f\t|' % (tpf_i, fpf_i, dscd_i, dscs_i)
for tpf_i, fpf_i, dscd_i, dscs_i in zip(tpf, fpf, dscd, dscs)])
print(intervals)
print(measures_s)
print(measures)
def test_all_nets(
path,
names_test,
defo_names_test,
roi_name,
nets,
case,
batch_size,
patch_sizes,
defo_sizes,
dense_sizes,
n_filters,
sufixes,
iter_name,
train_case=False
):
c = color_codes()
net_combos = itertools.product(zip(patch_sizes, defo_sizes), n_filters, dense_sizes)
image_nii = load_nii(names_test[0])
mask_nii = load_nii(roi_name)
images = list()
for net, sufix, ((patch_size, defo_size), _, _) in zip(nets, sufixes, net_combos):
outputname = os.path.join(path, 't' + case + sufix + iter_name + '.nii.gz')
# We save time by checking if we already tested that patient.
try:
image_nii = load_nii(outputname)
image = image_nii.get_data()
if train_case:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' +
c['g'] + ' Patient ' + names_test[0].rsplit('/')[-4] + ' already done' + c['nc'])
except IOError:
if train_case:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' +
c['g'] + ' Testing with patient ' + c['b'] + names_test[0].rsplit('/')[-4] + c['nc'])
else:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<Creating the probability map for net: ' +
c['b'] + sufix + c['nc'] + c['g'] + '>' + c['nc'])
image = test_net(
net=net,
names=names_test,
mask=mask_nii.get_data(),
batch_size=batch_size,
patch_size=patch_size,
defo_size=defo_size,
image_size=image_nii.get_data().shape,
images=['flair', 'pd', 't2'],
d_names=defo_names_test
)
if train_case:
print(c['g'] + ' -- Saving image ' + c['b'] + outputname + c['nc'])
image_nii.get_data()[:] = image
image_nii.to_filename(outputname)
images.append(image)
return images
def main():
parser = argparse.ArgumentParser(description='Test different nets with 3D data.')
parser.add_argument('-f', '--folder', dest='dir_name', default='/home/sergivalverde/w/CNN/images/CH16')
parser.add_argument('--flair', action='store', dest='flair', default='FLAIR_preprocessed.nii.gz')
parser.add_argument('--pd', action='store', dest='pd', default='DP_preprocessed.nii.gz')
parser.add_argument('--t2', action='store', dest='t2', default='T2_preprocessed.nii.gz')
parser.add_argument('--t1', action='store', dest='t1', default='T1_preprocessed.nii.gz')
parser.add_argument('--mask', action='store', dest='mask', default='Consensus.nii.gz')
parser.add_argument('--old', action='store_true', dest='old', default=False)
options = vars(parser.parse_args())
c = color_codes()
patch_size = (15, 15, 15)
batch_size = 100000
# Create the data
patients = [f for f in sorted(os.listdir(options['dir_name']))
if os.path.isdir(os.path.join(options['dir_name'], f))]
flair_names = [os.path.join(options['dir_name'], patient, options['flair']) for patient in patients]
pd_names = [os.path.join(options['dir_name'], patient, options['pd']) for patient in patients]
t2_names = [os.path.join(options['dir_name'], patient, options['t2']) for patient in patients]
t1_names = [os.path.join(options['dir_name'], patient, options['t1']) for patient in patients]
names = np.stack([name for name in [flair_names, pd_names, t2_names, t1_names]])
seed = np.random.randint(np.iinfo(np.int32).max)
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + 'Starting leave-one-out' + c['nc'])
for i in range(0, 15):
case = names[0, i].rsplit('/')[-2]
path = '/'.join(names[0, i].rsplit('/')[:-1])
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['nc'] + 'Patient ' + c['b'] + case + c['nc'])
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<Running iteration ' + c['b'] + '1' + c['nc'] + c['g'] + '>' + c['nc'])
net_name = os.path.join(path, 'deep-challenge2016.init.')
net = NeuralNet(
layers=[
(InputLayer, dict(name='in', shape=(None, 4, 15, 15, 15))),
(Conv3DDNNLayer, dict(name='conv1_1', num_filters=32, filter_size=(5, 5, 5), pad='same')),
(Pool3DDNNLayer, dict(name='avgpool_1', pool_size=2, stride=2, mode='average_inc_pad')),
(Conv3DDNNLayer, dict(name='conv2_1', num_filters=64, filter_size=(5, 5, 5), pad='same')),
(Pool3DDNNLayer, dict(name='avgpool_2', pool_size=2, stride=2, mode='average_inc_pad')),
(DropoutLayer, dict(name='l2drop', p=0.5)),
(DenseLayer, dict(name='l1', num_units=256)),
(DenseLayer, dict(name='out', num_units=2, nonlinearity=nonlinearities.softmax)),
],
objective_loss_function=objectives.categorical_crossentropy,
update=updates.adam,
update_learning_rate=0.0001,
on_epoch_finished=[
SaveWeights(net_name + 'model_weights.pkl', only_best=True, pickle=False),
EarlyStopping(patience=10)
],
verbose=10,
max_epochs=50,
train_split=TrainSplit(eval_size=0.25),
custom_scores=[('dsc', lambda p, t: 2 * np.sum(p * t[:, 1]) / np.sum((p + t[:, 1])))],
)
flair_name = os.path.join(path, options['flair'])
pd_name = os.path.join(path, options['pd'])
t2_name = os.path.join(path, options['t2'])
t1_name = os.path.join(path, options['t1'])
names_test = np.array([flair_name, pd_name, t2_name, t1_name])
outputname1 = os.path.join(path, 'test' + str(i) + '.iter1.nii.gz')
try:
net.load_params_from(net_name + 'model_weights.pkl')
except IOError:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' +
c['g'] + 'Loading the data for ' + c['b'] + 'iteration 1' + c['nc'])
names_lou = np.concatenate([names[:, :i], names[:, i + 1:]], axis=1)
paths = ['/'.join(name.rsplit('/')[:-1]) for name in names_lou[0, :]]
mask_names = [os.path.join(p_path, 'Consensus.nii.gz') for p_path in paths]
x_train, y_train = load_iter1_data(
names_lou=names_lou,
mask_names=mask_names,
patch_size=patch_size,
seed=seed
)
print(' Training vector shape ='
' (' + ','.join([str(length) for length in x_train.shape]) + ')')
print(' Training labels shape ='
' (' + ','.join([str(length) for length in y_train.shape]) + ')')
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'Training (' + c['b'] + 'initial' + c['nc'] + c['g'] + ')' + c['nc'])
# We try to get the last weights to keep improving the net over and over
net.fit(x_train, y_train)
try:
image_nii = load_nii(outputname1)
image1 = image_nii.get_data()
except IOError:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<Creating the probability map ' + c['b'] + '1' + c['nc'] + c['g'] + '>' + c['nc'])
flair_name = os.path.join(path, options['flair'])
image_nii = load_nii(flair_name)
image1 = np.zeros_like(image_nii.get_data())
print(' 0% of data tested', end='\r')
sys.stdout.flush()
for batch, centers, percent in load_patch_batch_percent(names_test, batch_size, patch_size):
y_pred = net.predict_proba(batch)
print(' %f%% of data tested' % percent, end='\r')
sys.stdout.flush()
[x, y, z] = np.stack(centers, axis=1)
image1[x, y, z] = y_pred[:, 1]
image_nii.get_data()[:] = image1
image_nii.to_filename(outputname1)
''' Here we get the seeds '''
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' +
c['g'] + '<Looking for seeds for the final iteration>' + c['nc'])
for patient in np.rollaxis(np.concatenate([names[:, :i], names[:, i+1:]], axis=1), 1):
outputname = os.path.join('/'.join(patient[0].rsplit('/')[:-1]), 'test' + str(i) + '.iter1.nii.gz')
try:
load_nii(outputname)
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' +
c['g'] + ' Patient ' + patient[0].rsplit('/')[-2] + ' already done' + c['nc'])
except IOError:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' +
c['g'] + ' Testing with patient ' + c['b'] + patient[0].rsplit('/')[-2] + c['nc'])
image_nii = load_nii(patient[0])
image = np.zeros_like(image_nii.get_data())
print(' 0% of data tested', end='\r')
for batch, centers, percent in load_patch_batch_percent(patient, 100000, patch_size):
y_pred = net.predict_proba(batch)
print(' %f%% of data tested' % percent, end='\r')
[x, y, z] = np.stack(centers, axis=1)
image[x, y, z] = y_pred[:, 1]
print(c['g'] + ' -- Saving image ' + c['b'] + outputname + c['nc'])
image_nii.get_data()[:] = image
image_nii.to_filename(outputname)
''' Here we perform the last iteration '''
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<Running iteration ' + c['b'] + '2' + c['nc'] + c['g'] + '>' + c['nc'])
outputname2 = os.path.join(path, 'test' + str(i) + '.old.iter2.nii.gz') if options['old'] \
else os.path.join(path, 'test' + str(i) + '.new.iter2.nii.gz')
net_name = os.path.join(path, 'deep-challenge2016.final.old.') if options['old'] \
else os.path.join(path, 'deep-challenge2016.final.new.')
net = NeuralNet(
layers=[
(InputLayer, dict(name='in', shape=(None, 4, 15, 15, 15))),
(Conv3DDNNLayer, dict(name='conv1_1', num_filters=32, filter_size=(5, 5, 5), pad='same')),
(Pool3DDNNLayer, dict(name='avgpool_1', pool_size=2, stride=2, mode='average_inc_pad')),
(Conv3DDNNLayer, dict(name='conv2_1', num_filters=64, filter_size=(5, 5, 5), pad='same')),
(Pool3DDNNLayer, dict(name='avgpool_2', pool_size=2, stride=2, mode='average_inc_pad')),
(DropoutLayer, dict(name='l2drop', p=0.5)),
(DenseLayer, dict(name='l1', num_units=256)),
(DenseLayer, dict(name='out', num_units=2, nonlinearity=nonlinearities.softmax)),
],
objective_loss_function=objectives.categorical_crossentropy,
update=updates.adam,
update_learning_rate=0.0001,
on_epoch_finished=[
SaveWeights(net_name + 'model_weights.pkl', only_best=True, pickle=False),
EarlyStopping(patience=50)
],
batch_iterator_train=BatchIterator(batch_size=4096),
verbose=10,
max_epochs=2000,
train_split=TrainSplit(eval_size=0.25),
custom_scores=[('dsc', lambda p, t: 2 * np.sum(p * t[:, 1]) / np.sum((p + t[:, 1])))],
)
try:
net.load_params_from(net_name + 'model_weights.pkl')
except IOError:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' +
c['g'] + 'Loading the data for ' + c['b'] + 'iteration 2' + c['nc'])
names_lou = np.concatenate([names[:, :i], names[:, i + 1:]], axis=1)
paths = ['/'.join(name.rsplit('/')[:-1]) for name in names_lou[0, :]]
roi_names = [os.path.join(p_path, 'test' + str(i) + '.iter1.nii.gz') for p_path in paths]
mask_names = [os.path.join(p_path, 'Consensus.nii.gz') for p_path in paths]
x_train, y_train = load_iter2_data(
names_lou=names_lou,
mask_names=mask_names,
roi_names=roi_names,
patch_size=patch_size,
seed=seed,
old=options['old']
)
print(' Training vector shape = (' + ','.join([str(length) for length in x_train.shape]) + ')')
print(' Training labels shape = (' + ','.join([str(length) for length in y_train.shape]) + ')')
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' +
c['g'] + 'Training (' + c['b'] + 'final' + c['nc'] + c['g'] + ')' + c['nc'])
net.fit(x_train, y_train)
try:
image_nii = load_nii(outputname2)
image2 = image_nii.get_data()
except IOError:
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<Creating the probability map ' + c['b'] + '2' + c['nc'] + c['g'] + '>' + c['nc'])
image_nii = load_nii(flair_name)
image2 = np.zeros_like(image_nii.get_data())
print(' 0% of data tested', end='\r')
sys.stdout.flush()
for batch, centers, percent in load_patch_batch_percent(names_test, batch_size, patch_size):
y_pred = net.predict_proba(batch)
print(' %f%% of data tested' % percent, end='\r')
sys.stdout.flush()
[x, y, z] = np.stack(centers, axis=1)
image2[x, y, z] = y_pred[:, 1]
image_nii.get_data()[:] = image2
image_nii.to_filename(outputname2)
image = (image1 * image2) > 0.5
seg = np.roll(np.roll(image, 1, axis=0), 1, axis=1)
image_nii.get_data()[:] = seg
outputname_final = os.path.join(path, 'test' + str(i) + '.old.final.nii.gz') if options['old'] \
else os.path.join(path, 'test' + str(i) + '.new.final.nii.gz')
image_nii.to_filename(outputname_final)
gt = load_nii(os.path.join(path, 'Consensus.nii.gz')).get_data().astype(dtype=np.bool)
dsc = np.sum(2.0 * np.logical_and(gt, seg)) / (np.sum(gt) + np.sum(seg))
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<DSC value for ' + c['c'] + case + c['g'] + ' = ' + c['b'] + str(dsc) + c['nc'] + c['g'] + '>' + c['nc'])
def load_thresholded_images_by_name(image_names, threshold=2.0):
images = [load_nii(image_name).get_data() for image_name in image_names]
return threshold_image_list(images, threshold)
def patches_network_segmentation(options, mode):
c = color_codes()
image_sufix = get_sufix(
options['use_flair'],
options['use_pd'],
options['use_t2'],
options['use_gado'],
options['use_t1']
)
size_sufix = '.'.join([str(length) for length in tuple(options['patch_size'])])
sufixes = image_sufix + '.' + size_sufix
mode_write = mode + '.mc' if options['multi_channel'] else mode + '.sc'
print(c['g'] + 'Loading the data for the patch-based ' + c['b'] + mode + c['nc'])
# Create the data
(x, y, names) = load_patches(
dir_name=options['folder'],
use_flair=options['use_flair'],
use_pd=options['use_pd'],
use_t2=options['use_t2'],
use_gado=options['use_gado'],
use_t1=options['use_t1'],
flair_name=options['flair'],
pd_name=options['pd'],
t2_name=options['t2'],
gado_name=options['gado'],
t1_name=options['t1'],
mask_name=options['mask'],
size=tuple(options['patch_size'])
)
print(c['g'] + 'Starting leave-one-out for the patch-based ' + c['b'] + mode + c['nc'])
n_channels = x[0].shape[1]
channels = range(0, n_channels)
patch_size = tuple(options['patch_size'])
for x_train, y_train, i in leave_one_out(x, y):
print('Running patient ' + c['c'] + names[0, i].rsplit('/')[-2] + c['nc'])
seed = np.random.randint(np.iinfo(np.int32).max)
print('-- Permuting the data')
np.random.seed(seed)
x_train = np.random.permutation(np.concatenate(x_train).astype(dtype=np.float32))
print('-- Permuting the labels')
np.random.seed(seed)
y_train = np.random.permutation(np.concatenate(y_train).astype(dtype=np.int32))
y_train = y_train.reshape([y_train.shape[0], -1])
print('-- Training vector shape = (' + ','.join([str(length) for length in x_train.shape]) + ')')
print('-- Training labels shape = (' + ','.join([str(length) for length in y_train.shape]) + ')')
print(c['g'] + '-- Creating the ' + c['b'] + 'patch-based ' + c['b'] + mode + c['nc'])
# Train the net and save it
net_name = os.path.join(
os.path.split(names[0, i])[0], 'patches_' + mode + '.c' + str(i) + '.' + sufixes
)
net_types = {
'unet': create_unet3d_seg_string,
'unet-short': create_unet3d_shortcuts_seg_string
}
net = net_types[mode](
''.join(options['layers']),
x_train.shape,
options['convo_size'],
options['pool_size'],
options['number_filters'],
options['patience'],
options['multi_channel'],
net_name
)
print(c['g'] + '-- Training the ' + c['b'] + 'patch-based ' + c['b'] + mode + c['nc'])
# We try to get the last weights to keep improving the net over and over
try:
net.load_params_from(net_name + 'model_weights.pkl')
except IOError:
pass
if options['multi_channel']:
net.fit(x_train, y_train)
else:
x_train = np.split(x_train, n_channels, axis=1)
inputs = dict(
[('\033[30minput_%d\033[0m' % ch, channel) for (ch, channel) in zip(channels, x_train)])
net.fit(inputs, y_train)
print(c['g'] + '-- Creating the test probability maps' + c['nc'])
image_nii = load_nii(names[0, i])
image = np.zeros_like(image_nii.get_data())
for batch, centers, _ in load_patch_batch_percent(names[:, i], options['batch_size'], patch_size):
if options['multi_channel']:
y_pred = net.predict_proba(batch)
else:
batch = np.split(batch, n_channels, axis=1)
inputs = dict(
[('\033[30minput_%d\033[0m' % ch, channel) for (ch, channel) in zip(channels, batch)])
y_pred = net.predict_proba(inputs)
image += sum_patches_to_image(y_pred, centers, image)
image_nii.get_data()[:] = image
name = mode_write + '.c' + str(i) + '.' + sufixes + '.nii.gz'
path = '/'.join(names[0, i].rsplit('/')[:-1])
image_nii.to_filename(os.path.join(path, name))
def load_masks(mask_names):
for image_name in mask_names:
yield load_nii(image_name).get_data().astype(dtype=np.bool)
def norm_defo_generator(image_names):
for name in image_names:
im = load_nii(name).get_data()
yield im / np.linalg.norm(im, axis=4).std()
def main():
# Parse command line options
parser = argparse.ArgumentParser(description='Test different nets with 3D data.')
parser.add_argument('--flair', action='store', dest='flair', default='FLAIR_preprocessed.nii.gz')
parser.add_argument('--pd', action='store', dest='pd', default='DP_preprocessed.nii.gz')
parser.add_argument('--t2', action='store', dest='t2', default='T2_preprocessed.nii.gz')
parser.add_argument('--t1', action='store', dest='t1', default='T1_preprocessed.nii.gz')
parser.add_argument('--output', action='store', dest='output', default='output.nii.gz')
parser.add_argument('--no-docker', action='store_false', dest='docker', default=True)
c = color_codes()
patch_size = (15, 15, 15)
options = vars(parser.parse_args())
batch_size = 10000
min_size = 30
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<Loading the net ' + c['b'] + '1' + c['nc'] + c['g'] + '>' + c['nc'])
net_name = '/usr/local/nets/deep-challenge2016.init.model_weights.pkl' if options['docker'] \
else './deep-challenge2016.init.model_weights.pkl'
net = NeuralNet(
layers=[
(InputLayer, dict(name='in', shape=(None, 4, 15, 15, 15))),
(Conv3DDNNLayer, dict(name='conv1_1', num_filters=32, filter_size=(5, 5, 5), pad='same')),
(Pool3DDNNLayer, dict(name='avgpool_1', pool_size=2, stride=2, mode='average_inc_pad')),
(Conv3DDNNLayer, dict(name='conv2_1', num_filters=64, filter_size=(5, 5, 5), pad='same')),
(Pool3DDNNLayer, dict(name='avgpool_2', pool_size=2, stride=2, mode='average_inc_pad')),
(DropoutLayer, dict(name='l2drop', p=0.5)),
(DenseLayer, dict(name='l1', num_units=256)),
(DenseLayer, dict(name='out', num_units=2, nonlinearity=nonlinearities.softmax)),
],
objective_loss_function=objectives.categorical_crossentropy,
update=updates.adam,
update_learning_rate=0.0001,
verbose=10,
max_epochs=50,
train_split=TrainSplit(eval_size=0.25),
custom_scores=[('dsc', lambda p, t: 2 * np.sum(p * t[:, 1]) / np.sum((p + t[:, 1])))],
)
net.load_params_from(net_name)
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<Creating the probability map ' + c['b'] + '1' + c['nc'] + c['g'] + '>' + c['nc'])
names = np.array([options['flair'], options['pd'], options['t2'], options['t1']])
image_nii = load_nii(options['flair'])
image1 = np.zeros_like(image_nii.get_data())
print('0% of data tested', end='\r')
sys.stdout.flush()
for batch, centers, percent in load_patch_batch_percent(names, batch_size, patch_size):
y_pred = net.predict_proba(batch)
print('%f%% of data tested' % percent, end='\r')
sys.stdout.flush()
[x, y, z] = np.stack(centers, axis=1)
image1[x, y, z] = y_pred[:, 1]
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<Loading the net ' + c['b'] + '2' + c['nc'] + c['g'] + '>' + c['nc'])
net_name = '/usr/local/nets/deep-challenge2016.final.model_weights.pkl' if options['docker'] \
else './deep-challenge2016.final.model_weights.pkl'
net = NeuralNet(
layers=[
(InputLayer, dict(name='in', shape=(None, 4, 15, 15, 15))),
(Conv3DDNNLayer, dict(name='conv1_1', num_filters=32, filter_size=(5, 5, 5), pad='same')),
(Pool3DDNNLayer, dict(name='avgpool_1', pool_size=2, stride=2, mode='average_inc_pad')),
(Conv3DDNNLayer, dict(name='conv2_1', num_filters=64, filter_size=(5, 5, 5), pad='same')),
(Pool3DDNNLayer, dict(name='avgpool_2', pool_size=2, stride=2, mode='average_inc_pad')),
(DropoutLayer, dict(name='l2drop', p=0.5)),
(DenseLayer, dict(name='l1', num_units=256)),
(DenseLayer, dict(name='out', num_units=2, nonlinearity=nonlinearities.softmax)),
],
objective_loss_function=objectives.categorical_crossentropy,
update=updates.adam,
update_learning_rate=0.0001,
batch_iterator_train=BatchIterator(batch_size=4096),
verbose=10,
max_epochs=2000,
train_split=TrainSplit(eval_size=0.25),
custom_scores=[('dsc', lambda t, p: 2 * np.sum(t * p[:, 1]) / np.sum((t + p[:, 1])))],
)
net.load_params_from(net_name)
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<Creating the probability map ' + c['b'] + '2' + c['nc'] + c['g'] + '>' + c['nc'])
image2 = np.zeros_like(image_nii.get_data())
print('0% of data tested', end='\r')
sys.stdout.flush()
for batch, centers, percent in load_patch_batch_percent(names, batch_size, patch_size):
y_pred = net.predict_proba(batch)
print('%f%% of data tested' % percent, end='\r')
sys.stdout.flush()
[x, y, z] = np.stack(centers, axis=1)
image2[x, y, z] = y_pred[:, 1]
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<Saving to file ' + c['b'] + options['output'] + c['nc'] + c['g'] + '>' + c['nc'])
image = (image1 * image2) > 0.5
# filter candidates < min_size
labels, num_labels = ndimage.label(image)
lesion_list = np.unique(labels)
num_elements_by_lesion = ndimage.labeled_comprehension(image, labels, lesion_list, np.sum, float, 0)
filt_min_size = num_elements_by_lesion >= min_size
lesion_list = lesion_list[filt_min_size]
image = reduce(np.logical_or, map(lambda lab: lab == labels, lesion_list))
image_nii.get_data()[:] = np.roll(np.roll(image, 1, axis=0), 1, axis=1)
path = '/'.join(options['t1'].rsplit('/')[:-1])
outputname = options['output'].rsplit('/')[-1]
image_nii.to_filename(os.path.join(path, outputname))
if not options['docker']:
path = '/'.join(options['output'].rsplit('/')[:-1])
case = options['output'].rsplit('/')[-1]
gt = load_nii(os.path.join(path, 'Consensus.nii.gz')).get_data().astype(dtype=np.bool)
dsc = np.sum(2.0 * np.logical_and(gt, image)) / (np.sum(gt) + np.sum(image))
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] +
'<DSC value for ' + c['c'] + case + c['g'] + ' = ' + c['b'] + str(dsc) + c['nc'] + c['g'] + '>' + c['nc'])