'''
'''
from Vnet.layer import (conv3d, deconv3d, normalizationlayer, crop_and_concat, resnet_Add,
weight_xavier_init, bias_variable, save_images)
from Vnet.loss_metric import (categorical_crossentropy, mean_iou, mean_dice, categorical_dice, categorical_focal_loss,
generalized_dice_loss_w, categorical_tversky, weighted_categorical_crossentropy,
categorical_dicePcrossentroy, categorical_dicePfocalloss, multiscalessim2d_loss,
ssim2d_loss)
import tensorflow as tf
import numpy as np
import os
def conv_bn_relu_drop(x, kernal, phase, drop, image_z=None, height=None, width=None, scope=None):
with tf.name_scope(scope):
W = weight_xavier_init(shape=kernal, n_inputs=kernal[0] * kernal[1] * kernal[2] * kernal[3],
n_outputs=kernal[-1], activefunction='relu', variable_name=scope + 'conv_W')
B = bias_variable([kernal[-1]], variable_name=scope + 'conv_B')
conv = conv3d(x, W) + B
conv = normalizationlayer(conv, is_train=phase, height=height, width=width, image_z=image_z, norm_type='group',
G=20, scope=scope)
conv = tf.nn.dropout(tf.nn.relu(conv), drop)
return conv
def down_sampling(x, kernal, phase, drop, image_z=None, height=None, width=None, scope=None):
with tf.name_scope(scope):
W = weight_xavier_init(shape=kernal, n_inputs=kernal[0] * kernal[1] * kernal[2] * kernal[3],
n_outputs=kernal[-1],
activefunction='relu', variable_name=scope + 'W')
B = bias_variable([kernal[-1]], variable_name=scope + 'B')
conv = conv3d(x, W, 2) + B
conv = normalizationlayer(conv, is_train=phase, height=height, width=width, image_z=image_z, norm_type='group',
G=20, scope=scope)
conv = tf.nn.dropout(tf.nn.relu(conv), drop)
return conv
def deconv_relu(x, kernal, samefeture=False, scope=None):
with tf.name_scope(scope):
W = weight_xavier_init(shape=kernal, n_inputs=kernal[0] * kernal[1] * kernal[2] * kernal[-1],
n_outputs=kernal[-2], activefunction='relu', variable_name=scope + 'W')
B = bias_variable([kernal[-2]], variable_name=scope + 'B')
conv = deconv3d(x, W, samefeture, True) + B
conv = tf.nn.relu(conv)
return conv
def conv_softmax(x, kernal, scope=None):
with tf.name_scope(scope):
W = weight_xavier_init(shape=kernal, n_inputs=kernal[0] * kernal[1] * kernal[2] * kernal[3],
n_outputs=kernal[-1], activefunction='sigomd', variable_name=scope + 'W')
B = bias_variable([kernal[-1]], variable_name=scope + 'B')
conv = conv3d(x, W) + B
conv = tf.nn.softmax(conv)
return conv
def _create_conv_net(X, image_z, image_width, image_height, image_channel, phase, drop, n_class=2):
inputX = tf.reshape(X, [-1, image_z, image_width, image_height, image_channel]) # shape=(?, 32, 32, 1)
# Vnet model
# layer1->convolution
layer0 = conv_bn_relu_drop(x=inputX, kernal=(3, 3, 3, image_channel, 20), phase=phase, drop=drop,
scope='layer0')
layer1 = conv_bn_relu_drop(x=layer0, kernal=(3, 3, 3, 20, 20), phase=phase, drop=drop,
scope='layer1')
layer1 = resnet_Add(x1=layer0, x2=layer1)
# down sampling1
down1 = down_sampling(x=layer1, kernal=(3, 3, 3, 20, 40), phase=phase, drop=drop, scope='down1')
# layer2->convolution
layer2 = conv_bn_relu_drop(x=down1, kernal=(3, 3, 3, 40, 40), phase=phase, drop=drop,
scope='layer2_1')
layer2 = conv_bn_relu_drop(x=layer2, kernal=(3, 3, 3, 40, 40), phase=phase, drop=drop,
scope='layer2_2')
layer2 = resnet_Add(x1=down1, x2=layer2)
# down sampling2
down2 = down_sampling(x=layer2, kernal=(3, 3, 3, 40, 80), phase=phase, drop=drop, scope='down2')
# layer3->convolution
layer3 = conv_bn_relu_drop(x=down2, kernal=(3, 3, 3, 80, 80), phase=phase, drop=drop,
scope='layer3_1')
layer3 = conv_bn_relu_drop(x=layer3, kernal=(3, 3, 3, 80, 80), phase=phase, drop=drop,
scope='layer3_2')
layer3 = conv_bn_relu_drop(x=layer3, kernal=(3, 3, 3, 80, 80), phase=phase, drop=drop,
scope='layer3_3')
layer3 = resnet_Add(x1=down2, x2=layer3)
# down sampling3
down3 = down_sampling(x=layer3, kernal=(3, 3, 3, 80, 160), phase=phase, drop=drop, scope='down3')
# layer4->convolution
layer4 = conv_bn_relu_drop(x=down3, kernal=(3, 3, 3, 160, 160), phase=phase, drop=drop,
scope='layer4_1')
layer4 = conv_bn_relu_drop(x=layer4, kernal=(3, 3, 3, 160, 160), phase=phase, drop=drop,
scope='layer4_2')
layer4 = conv_bn_relu_drop(x=layer4, kernal=(3, 3, 3, 160, 160), phase=phase, drop=drop,
scope='layer4_3')
layer4 = resnet_Add(x1=down3, x2=layer4)
# down sampling4
down4 = down_sampling(x=layer4, kernal=(3, 3, 3, 160, 320), phase=phase, drop=drop, scope='down4')
# layer5->convolution
layer5 = conv_bn_relu_drop(x=down4, kernal=(3, 3, 3, 320, 320), phase=phase, drop=drop,
scope='layer5_1')
layer5 = conv_bn_relu_drop(x=layer5, kernal=(3, 3, 3, 320, 320), phase=phase, drop=drop,
scope='layer5_2')
layer5 = conv_bn_relu_drop(x=layer5, kernal=(3, 3, 3, 320, 320), phase=phase, drop=drop,
scope='layer5_3')
layer5 = resnet_Add(x1=down4, x2=layer5)
# layer9->deconvolution
deconv1 = deconv_relu(x=layer5, kernal=(3, 3, 3, 160, 320), scope='deconv1')
# layer8->convolution
layer6 = crop_and_concat(layer4, deconv1)
_, Z, H, W, _ = layer4.get_shape().as_list()
layer6 = conv_bn_relu_drop(x=layer6, kernal=(3, 3, 3, 320, 160), image_z=Z, height=H, width=W, phase=phase,
drop=drop, scope='layer6_1')
layer6 = conv_bn_relu_drop(x=layer6, kernal=(3, 3, 3, 160, 160), image_z=Z, height=H, width=W, phase=phase,
drop=drop, scope='layer6_2')
layer6 = conv_bn_relu_drop(x=layer6, kernal=(3, 3, 3, 160, 160), image_z=Z, height=H, width=W, phase=phase,
drop=drop, scope='layer6_3')
layer6 = resnet_Add(x1=deconv1, x2=layer6)
# layer9->deconvolution
deconv2 = deconv_relu(x=layer6, kernal=(3, 3, 3, 80, 160), scope='deconv2')
# layer8->convolution
layer7 = crop_and_concat(layer3, deconv2)
_, Z, H, W, _ = layer3.get_shape().as_list()
layer7 = conv_bn_relu_drop(x=layer7, kernal=(3, 3, 3, 160, 80), image_z=Z, height=H, width=W, phase=phase,
drop=drop, scope='layer7_1')
layer7 = conv_bn_relu_drop(x=layer7, kernal=(3, 3, 3, 80, 80), image_z=Z, height=H, width=W, phase=phase,
drop=drop, scope='layer7_2')
layer7 = conv_bn_relu_drop(x=layer7, kernal=(3, 3, 3, 80, 80), image_z=Z, height=H, width=W, phase=phase,
drop=drop, scope='layer7_3')
layer7 = resnet_Add(x1=deconv2, x2=layer7)
# layer9->deconvolution
deconv3 = deconv_relu(x=layer7, kernal=(3, 3, 3, 40, 80), scope='deconv3')
# layer8->convolution
layer8 = crop_and_concat(layer2, deconv3)
_, Z, H, W, _ = layer2.get_shape().as_list()
layer8 = conv_bn_relu_drop(x=layer8, kernal=(3, 3, 3, 80, 40), image_z=Z, height=H, width=W, phase=phase,
drop=drop, scope='layer8_1')
layer8 = conv_bn_relu_drop(x=layer8, kernal=(3, 3, 3, 40, 40), image_z=Z, height=H, width=W, phase=phase,
drop=drop, scope='layer8_2')
layer8 = conv_bn_relu_drop(x=layer8, kernal=(3, 3, 3, 40, 40), image_z=Z, height=H, width=W, phase=phase,
drop=drop, scope='layer8_3')
layer8 = resnet_Add(x1=deconv3, x2=layer8)
# layer9->deconvolution
deconv4 = deconv_relu(x=layer8, kernal=(3, 3, 3, 20, 40), scope='deconv4')
# layer8->convolution
layer9 = crop_and_concat(layer1, deconv4)
_, Z, H, W, _ = layer1.get_shape().as_list()
layer9 = conv_bn_relu_drop(x=layer9, kernal=(3, 3, 3, 40, 20), image_z=Z, height=H, width=W, phase=phase,
drop=drop, scope='layer9_1')
layer9 = conv_bn_relu_drop(x=layer9, kernal=(3, 3, 3, 20, 20), image_z=Z, height=H, width=W, phase=phase,
drop=drop, scope='layer9_2')
layer9 = conv_bn_relu_drop(x=layer9, kernal=(3, 3, 3, 20, 20), image_z=Z, height=H, width=W, phase=phase,
drop=drop, scope='layer9_3')
layer9 = resnet_Add(x1=deconv4, x2=layer9)
# layer14->output
output_map = conv_softmax(x=layer9, kernal=(1, 1, 1, 20, n_class), scope='output')
return output_map
# Serve data by batches
def _next_batch(train_images, train_labels, batch_size, index_in_epoch):
start = index_in_epoch
index_in_epoch += batch_size
num_examples = train_images.shape[0]
# when all trainig data have been already used, it is reorder randomly
if index_in_epoch > num_examples:
# shuffle the data
perm = np.arange(num_examples)
np.random.shuffle(perm)
train_images = train_images[perm]
train_labels = train_labels[perm]
# start next epoch
start = 0
index_in_epoch = batch_size
assert batch_size <= num_examples
end = index_in_epoch
return train_images[start:end], train_labels[start:end], index_in_epoch
# convet label to one hot type
def convert_to_one_hot(y, numclass):
"""
convert y array to one-hot array
:param y:[batch size,z,x,y,channel]
:param numclass:number class
:return:[batch size,z,x,y,numclass]
"""
one_hoty = np.reshape(y, (-1,))
one_hoty = np.eye(numclass)[one_hoty.reshape(-1).astype(np.int)]
return one_hoty
class Vnet3dModuleMultiLabel(object):
"""
A Vnet3dMultiLabel implementation
:param image_height: number of height in the input image
:param image_width: number of width in the input image
:param image_depth: number of depth in the input image
:param channels: number of channels in the input image
:param costname: name of the cost function.Default is "dice coefficient"
"""
def __init__(self, image_height, image_width, image_depth, channels=4, numclass=4,
costname=("categorical_crossentropy",), inference=False, model_path=None):
self.image_width = image_width
self.image_height = image_height
self.image_depth = image_depth
self.channels = channels
self.numclass = numclass
self.labelchannels = 1
self.X = tf.placeholder("float", shape=[None, self.image_depth, self.image_height, self.image_width,
self.channels])
self.Y_gt = tf.placeholder("float", shape=[None, self.image_depth, self.image_height, self.image_width,
self.numclass])
self.lr = tf.placeholder('float')
self.phase = tf.placeholder(tf.bool)
self.drop = tf.placeholder('float')
self.weight_loss = [0.1, 1., 1., 1.]
self.Y_pred = _create_conv_net(self.X, self.image_depth, self.image_width, self.image_height, self.channels,
self.phase, self.drop, self.numclass)
self.cost1 = self.__get_cost(self.Y_pred, self.Y_gt, costname[0], gamma=2)
self.cost_re = multiscalessim2d_loss(self.Y_pred, self.Y_gt, self.numclass - 1)
self.cost = self.cost1 + self.cost_re
self.Y_pred_arg = tf.reshape(tf.argmax(self.Y_pred, axis=-1),
(-1, self.image_depth, self.image_height, self.image_width, self.labelchannels))
self.accuracy = self.__get_metrics(self.Y_pred, self.Y_gt, "mdice")
if inference:
init = tf.global_variables_initializer()
saver = tf.train.Saver()
self.sess = tf.InteractiveSession()
self.sess.run(init)
saver.restore(self.sess, model_path)
else:
self.sess = tf.InteractiveSession(
config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False))
def __get_cost(self, Y_pred, Y_gt, cost_name, gamma=2):
if cost_name == "categorical_crossentropy":
loss = categorical_crossentropy(Y_pred, Y_gt)
if cost_name == "weighted_categorical_crossentropy":
loss = weighted_categorical_crossentropy(Y_pred, Y_gt, self.weight_loss)
if cost_name == "categorical_dice":
loss = categorical_dice(Y_pred, Y_gt, self.weight_loss)
if cost_name == "generalized_dice_loss_w":
loss = generalized_dice_loss_w(Y_pred, Y_gt)
if cost_name == "categorical_focal_loss":
loss = categorical_focal_loss(Y_pred, Y_gt, gamma, self.weight_loss)
if cost_name == "categorical_tversky":
loss = categorical_tversky(Y_pred, Y_gt, beta=0.25, weight_loss=self.weight_loss)
if cost_name == "categorical_dicePcrossentroy":
loss = categorical_dicePcrossentroy(Y_pred, Y_gt, self.weight_loss)
if cost_name == "categorical_dicePfocalloss":
loss = categorical_dicePfocalloss(Y_pred, Y_gt, self.weight_loss, 0.6, 3.)
return loss
def __get_metrics(self, Y_pred, Y_gt, metric_name="miou"):
if metric_name == "miou":
metric = mean_iou(Y_pred, Y_gt)
if metric_name == "mdice":
metric = mean_dice(Y_pred, Y_gt)
return metric
def __loadnumtraindata(self, train_images, train_lanbels, num_sample, num_sample_index_in_epoch):
subbatch_xs = np.empty((num_sample, self.image_depth, self.image_height, self.image_width, self.channels))
subbatch_ys = np.empty((num_sample, self.image_depth, self.image_height, self.image_width, self.labelchannels))
batch_xs_path, batch_ys_path, num_sample_index_in_epoch = _next_batch(train_images, train_lanbels,
num_sample, num_sample_index_in_epoch)
for num in range(len(batch_xs_path)):
image = np.load(batch_xs_path[num])
label = np.load(batch_ys_path[num])
# prepare 3 model output
batch_ys_tmp = np.zeros(label.shape, np.int)
batch_ys_tmp[label == 1.] = 1
batch_ys_tmp[label == 2.] = 2
batch_ys_tmp[label == 4.] = 3
subbatch_xs[num, :, :, :, :] = np.reshape(image,
(self.image_depth, self.image_height, self.image_width,
self.channels))
subbatch_ys[num, :, :, :, :] = np.reshape(batch_ys_tmp,
(self.image_depth, self.image_height, self.image_width,
self.labelchannels))
subbatch_xs = subbatch_xs.astype(np.float)
subbatch_ys = subbatch_ys.astype(np.float)
return subbatch_xs, subbatch_ys, num_sample_index_in_epoch
def train(self, train_images, train_labels, model_path, logs_path, learning_rate,
dropout_conv=0.8, train_epochs=5, batch_size=1, showwind=[6, 8]):
num_sample = 1
if not os.path.exists(logs_path):
os.makedirs(logs_path)
if not os.path.exists(logs_path + "model\\"):
os.makedirs(logs_path + "model\\")
model_path = logs_path + "model\\" + model_path
train_op = tf.train.AdamOptimizer(self.lr).minimize(self.cost)
init = tf.global_variables_initializer()
saver = tf.train.Saver(tf.all_variables(), max_to_keep=10)
tf.summary.scalar("loss", self.cost)
tf.summary.scalar("accuracy", self.accuracy)
merged_summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(logs_path, graph=tf.get_default_graph())
self.sess.run(init)
ckpt = tf.train.get_checkpoint_state(logs_path + "model\\")
if ckpt and ckpt.model_checkpoint_path:
print('Checkpoint file: {}'.format(ckpt.model_checkpoint_path))
saver.restore(self.sess, ckpt.model_checkpoint_path)
DISPLAY_STEP = 1
num_sample_index_in_epoch = 0
index_in_epoch = 0
train_epochs = train_images.shape[0] * train_epochs
for i in range(train_epochs):
# Extracting num_sample images and labels from given data
if i % num_sample == 0 or i == 0:
subbatch_xs, subbatch_ys, num_sample_index_in_epoch = self.__loadnumtraindata(train_images,
train_labels, num_sample,
num_sample_index_in_epoch)
# get new batch
batch_xs, batch_ys, index_in_epoch = _next_batch(subbatch_xs, subbatch_ys, batch_size, index_in_epoch)
# convert label to one hot type
batch_ys_onehot = convert_to_one_hot(batch_ys, self.numclass)
batch_ys_onehot = np.reshape(batch_ys_onehot,
(-1, self.image_depth, self.image_height, self.image_width, self.numclass))
# check progress on every 1st,2nd,...,10th,20th,...,100th... step
if i % DISPLAY_STEP == 0 or (i + 1) == train_epochs:
train_loss, train_accuracy = self.sess.run([self.cost, self.accuracy],
feed_dict={self.X: batch_xs,
self.Y_gt: batch_ys_onehot,
self.lr: learning_rate,
self.phase: 1,
self.drop: dropout_conv})
print('epochs %d training_loss ,Training_accuracy => %.5f,%.5f ' % (i, train_loss, train_accuracy))
pred_arg = self.sess.run(self.Y_pred_arg, feed_dict={self.X: batch_xs,
self.Y_gt: batch_ys_onehot,
self.phase: 1,
self.drop: 1})
batch_ys_tmp = np.argmax(batch_ys_onehot, axis=-1)
gt = np.reshape(batch_ys_tmp[0], (self.image_depth, self.image_height, self.image_width))
gt = gt.astype(np.float)
save_images(gt, showwind, path=logs_path + 'gt_%d_epoch.png' % (i), pixelvalue=85)
result = np.reshape(pred_arg[0], (self.image_depth, self.image_height, self.image_width))
result = result.astype(np.float)
save_images(result, showwind, path=logs_path + 'predict_%d_epoch.png' % (i), pixelvalue=85)
save_path = saver.save(self.sess, model_path, global_step=i)
print("Model saved in file:", save_path)
if i % (DISPLAY_STEP * 10) == 0 and i:
DISPLAY_STEP *= 10
# train on batch
_, summary = self.sess.run([train_op, merged_summary_op], feed_dict={self.X: batch_xs,
self.Y_gt: batch_ys_onehot,
self.lr: learning_rate,
self.phase: 1,
self.drop: dropout_conv})
summary_writer.add_summary(summary, i)
summary_writer.close()
save_path = saver.save(self.sess, model_path)
print("Model saved in file:", save_path)
def prediction(self, test_images):
test_images = np.reshape(test_images,
(test_images.shape[0], test_images.shape[1], test_images.shape[2], self.channels))
test_images = test_images.astype(np.float)
y_dummy = np.zeros((test_images.shape[0], test_images.shape[1], test_images.shape[2], self.numclass))
pred_arg = self.sess.run(self.Y_pred_arg, feed_dict={self.X: [test_images],
self.Y_gt: [y_dummy],
self.phase: 1,
self.drop: 1})
result = pred_arg.astype(np.float)
result = np.clip(result, 0, 255).astype('uint8')
result = np.reshape(result, (test_images.shape[0], test_images.shape[1], test_images.shape[2]))
return result
