def run_postprocessing(self):
print('Post-processing of segmented images...')
print('Clipping images... ')
filepaths = util.get_nifti_filepaths('predictions/')
filenames = util.get_sub_dirs(self.input_dir)
for i in tqdm(range(len(filepaths))):
self.remove_small_objects(filepaths[i], filenames[i],
self.input_dir)
# Delete folder created by miscnn
shutil.rmtree('predictions/')
def run(self):
# Get filenames
filenames = util.get_paths_from_tree(self.input_dir, 'segmentation')
print('\n\nRunning clustering on samples...\n\n')
# Loop over images
for i, filename in enumerate(tqdm(filenames)):
# Print output
img = nib.load(filename)
img_data = img.get_data()
hd = img.header
M = img.affine[:3, :3]
iM = np.linalg.inv(M)
abc = img.affine[:3, 3]
# Get coordinate data of image
print('\n\nGet coordinate matrix from segmentation.. ')
data = get_coordinates_from_segmentation(img_data, M, iM, abc)
# Shuffle the data
#np.random.shuffle(data)
# Set random seed for reproducibility
random.seed(0)
# Initialise centroids
centroids = data[random.sample(range(data.shape[0]),
self.num_clusters)]
# Create a list to store which centroid is assigned to each dataset
assigned_centroids = np.zeros(len(data), dtype=np.int32)
# Number of dimensions in centroid
num_centroid_dims = data.shape[1]
# List to store SSE for each iteration
sse_list = []
# Start clustering time
tic = time.time()
# Main Loop
print('\nStart clustering image...')
for n in tqdm(range(self.num_iters), leave=False):
# Get closest centroids to each data point
assigned_centroids = get_closest_centroid(
data[:, None, :], centroids[None, :, :])
# Compute new centroids
for c in range(centroids.shape[1]):
# Get data points belonging to each cluster
cluster_members = data[assigned_centroids == c]
# Compute the mean of the clusters
cluster_members = cluster_members.mean(axis=0)
# Update the centroids
centroids[c] = cluster_members
# Compute SSE
sse = compute_sse(data.squeeze(), centroids.squeeze(),
assigned_centroids)
sse_list.append(sse)
print('\tFinished clustering image.')
# End clustering time
toc = time.time()
# Print output
print("Image clustering-time: " + str(round(toc - tic, 2)))
print("Image clustering-time per iteration: " +
str(round(toc - tic, 2) / self.num_iters))
print('\nSaving clustered image to nifti label-map... ')
# Start saving time
tic = time.time()
# Save clustering to nifti label-map
label_map = np.zeros_like(img_data)
# Join centeroids and coordinates to dict
merged = defaultdict(list)
for a, b in zip(assigned_centroids, data):
merged[a].append(b)
for key, cords in merged.items():
cluster = key + 1
for cord in cords:
ijk = invf(cord[0], cord[1], cord[2], iM, abc)
label_map[int(ijk[0]), int(ijk[1]), int(ijk[2])] = cluster
# End clustering time
toc = time.time()
# Print output
print("Time spent saving image to labelmap: " +
str(round(toc - tic, 2)))
"""
# Start saving time
tic = time.time()
# Make dict for cent
label = 1
print('\nSaving clustered image to nifti label-map... \n')
for c in tqdm(range(len(centroids)), leave=False):
cluster_members = [data[i] for i in range(len(data)) if assigned_centroids[i] == c]
cluster_members = np.array(cluster_members)
for cord in cluster_members:
ijk = invf(cord[0], cord[1], cord[2], iM, abc)
label_map[int(ijk[0]),int(ijk[1]),int(ijk[2])] = label
label+=1
# End clustering time
toc = time.time()
# Print output
print("Time spent saving image to labelmap: " + str(round(toc - tic, 2)))
"""
folder_name = util.get_sub_dirs(self.input_dir)[i]
self.save_clustered_image(img, label_map, hd, folder_name)
print('File saved.')
def evaluate(model,
config,
experiment,
validation_directory,
file_identifier=''):
missclassified = {}
# get number of classes in model
number_of_classes = config['dataset']['number_of_classes']
# image dimensions
image_width = config['image_processing']['image_width']
image_height = config['image_processing']['image_height']
image_channels = config['image_processing']['image_channels']
# get class directory names from validation directory
class_names = get_sub_dirs(validation_directory)
class_names.sort()
# get keras labels in label-index format
label_index = {
class_name: index
for index, class_name in enumerate(class_names)
}
index_label = {
index: class_name
for index, class_name in enumerate(class_names)
}
# prepare confusion table
confusion = np.zeros((number_of_classes, number_of_classes))
# iterate over each class name
for class_name in class_names:
print(f'Starting {class_name}')
# set path to class directory
class_dir = os.path.join(validation_directory, class_name)
# iterate over each image in class directory
for file_name in os.listdir(class_dir):
# models class prediction for image
prediction = None
# process image before passing it through the network
image = imread(os.path.join(class_dir, file_name), mode='RGB')
image = imresize(image,
(image_width, image_height, image_channels))
image = image.reshape(1, image_width, image_height, image_channels)
image = np.true_divide(image, 255.)
with tf.get_default_graph().as_default():
predictions = model.predict(image)[0]
prediction = np.argmax(predictions)
# check prediction against ground truth, i.e, if it equals the class directory name
if (prediction != label_index[class_name]):
# initialize empty list of fist missclassified of class
if class_name not in missclassified:
missclassified[class_name] = {}
missclassified[class_name][file_name] = {
'prediction': index_label[prediction],
'predictions': {
index_label[class_index]: pred
for class_index, pred in enumerate(predictions)
}
}
# update confusion table
confusion[prediction][label_index[class_name]] += 1
# calculate FP, FN, TP and TN based on confusion table
FP = confusion.sum(axis=0) - np.diag(confusion)
FN = confusion.sum(axis=1) - np.diag(confusion)
TP = np.diag(confusion)
TN = confusion.sum() - (FP + FN + TP)
print(f"True Positives: { TP }")
print(f"True Negatives: { TN }")
print(f"False Positives: { FP }")
print(f"False Positives: { FN }")
# calculate metrics based on FP, FN, TP and TN
f1 = np.nan_to_num(f1score(TP, TN, FP, FN))
rec = np.nan_to_num(recall(TP, TN, FP, FN))
acc = np.nan_to_num(accuracy(TP, TN, FP, FN))
prec = np.nan_to_num(precision(TP, TN, FP, FN))
spec = np.nan_to_num(specificity(TP, TN, FP, FN))
mcc = np.nan_to_num(matthews_correlation_coefficient(TP, TN, FP, FN))
# bundle metrics into dictionary
metrics = {
'FP': FP,
'FN': FN,
'TP': TP,
'TN': TN,
'f1': f1,
'rec': rec,
'acc': acc,
'prec': prec,
'spec': spec,
'mcc': mcc
}
# save missclassified images to file together with class
for class_name in missclassified:
log_misclassifications(
f'{file_identifier}_class_misclassifications.txt',
missclassified[class_name], class_name, index_label)
# write kvasir legend to results file
log_class_legend(f'{file_identifier}_split_evaluation_summary.txt',
class_names)
# write confusion table to results file
log_confusion_table(f'{file_identifier}_split_evaluation_summary.txt',
confusion)
# write model summary to results file
log_model_results(f'{file_identifier}_split_evaluation_summary.txt',
metrics, file_identifier)
# write summaries for each class
for class_name in class_names:
# class index
class_index = label_index[class_name]
class_metrics = {
key: value[class_index]
for key, value in metrics.items()
}
# write class summary to results file
log_class_results(f'{file_identifier}_class_results.txt',
class_metrics, class_name, class_index)
evaluation_path = config['evaluation']['path']
print("starting test validation...")
for file_name in os.listdir(evaluation_path):
prediction = None
prediction_time = None
image = imread(os.path.join(evaluation_path, file_name), mode='RGB')
image = imresize(image, (image_width, image_height, image_channels))
image = image.reshape(1, image_width, image_height, image_channels)
image = np.true_divide(image, 255.)
with tf.get_default_graph().as_default():
start_time = time.time()
prediction = model.predict(image)[0]
prediction_time = time.time() - start_time
prediction_index = np.argmax(prediction)
prediction_label = index_label[prediction_index]
log_file_evaluation(f'{file_identifier}_test_evaluation_results.txt',
file_name, prediction_label,
prediction[prediction_index], prediction_time)
# add evaluation files to experiment
experiment.add_artifact(
f'../tmp/{file_identifier}_split_evaluation_summary.txt')
experiment.add_artifact(
f'../tmp/{file_identifier}_class_misclassifications.txt')
experiment.add_artifact(f'../tmp/{file_identifier}_class_results.txt')
experiment.add_artifact(
f'../tmp/{file_identifier}_test_evaluation_results.txt')
# return evaluation metrics
return {
'f1': np.mean(f1),
'rec': np.mean(rec),
'acc': np.mean(acc),
'prec': np.mean(prec),
'spec': np.mean(spec),
'mcc': np.mean(mcc)
}
def run(self, batch_size):
# Get filenames
img_filenames = util.get_paths_from_tree(self.input_dir, 'imaging')
clus_filenames = util.get_paths_from_tree(self.input_dir, 'cluster')
folder_names = util.get_sub_dirs(self.input_dir)
print('\n\nRunning feature extraction on samples.. \n\n')
for i in tqdm(range(len(img_filenames))):
# Create out dir for patches
save_dir = util.create_fe_patch_dir(self.patch_size)
# Append filename to new list to match input of create 2_patches
img = [img_filenames[i]]
clus = [clus_filenames[i]]
# Extract 3d patches from img
img_ids, label_ids = extract_patches.patch_sampler(
img_filenames=img,
labelmap_filenames=clus,
patch_size=self.patch_size,
out_dir=save_dir,
sampler_type='grid',
voxel_spacing=self.resample,
patch_overlap=self.patch_overlap,
save_patches=True,
inference=True)
# Save info
info = {
'patch_size': self.patch_size,
'cluster_selection': self.cluster_selection,
'num_clusters': self.num_clusters,
'patch_overlap': self.patch_overlap
}
util.save_dict(info, self.out_dir, 'info.csv')
# Add partition to dict and save
partition = dict()
partition['image'] = img_ids
partition['cluster'] = label_ids
# Create generators
image_generator = DataGenerator(partition['image'],
data_dir=save_dir,
shuffle=False,
batch_size=batch_size)
cluster_generator = DataGenerator(partition['cluster'],
data_dir=save_dir,
shuffle=False,
batch_size=batch_size)
# Load all data from generator
image_data = [x[0] for x in image_generator]
cluster_data = [x[0] for x in cluster_generator]
image_data = np.array(image_data)
cluster_data = np.array(cluster_data)
if self.patch_size[0] == 1:
image_data = image_data.reshape(
image_data.shape[0] * image_data.shape[1],
image_data.shape[2], image_data.shape[3], 1)
cluster_data = cluster_data.reshape(
cluster_data.shape[0] * cluster_data.shape[1],
cluster_data.shape[2], cluster_data.shape[3], 1)
else:
image_data = image_data.reshape(
image_data.shape[0] * image_data.shape[1],
image_data.shape[2], image_data.shape[3],
image_data.shape[4], 1)
cluster_data = cluster_data.reshape(
cluster_data.shape[0] * cluster_data.shape[1],
cluster_data.shape[2], cluster_data.shape[3],
cluster_data.shape[4], 1)
print(image_data.shape)
# Make predictions
pred = self.encoder.predict(image_data,
verbose=1,
batch_size=batch_size)
# Calculate center index
x_center = 0
y_center = 0
z_center = 0
if self.patch_size[0] != 1:
x_center = int(self.patch_size[0] / 2 - 1)
y_center = int(self.patch_size[1] / 2 - 1)
z_center = int(self.patch_size[2] / 2 - 1)
else:
x_center = int(self.patch_size[1] / 2 - 1)
y_center = int(self.patch_size[2] / 2 - 1)
# Start feature extraction time
tic = time.time()
feature = dict().fromkeys(range(1, self.num_clusters + 1), 0)
print('\nStart extracting features from image...')
for j in range(len(image_data)):
# Find matching cluster based on center voxel
if self.cluster_selection == 'center':
cluster = int(cluster_data[j, x_center, y_center,
z_center])
# Find the matching cluster based on highest share
elif self.cluster_selection == 'highest_share':
cluster_patch = cluster_data[j].astype(int)
cluster = np.argmax(np.bincount(cluster_patch.flat))
# Calculate and add max stds for each cluster
if cluster != 0:
std = np.std(pred[j])
max_std = feature.get(cluster)
if std > max_std:
feature[cluster] = std
# End clustering time
toc = time.time()
print("Image feature-extraction-time: " + str(round(toc - tic, 2)))
# Save features to disk
util.save_features(list(feature.values()), folder_names[i],
self.out_dir)
# Delete last img patches
shutil.rmtree(save_dir)