딥러닝/segmentation

[Segmentation] H-DenseUnet 2D

ufris 2018. 2. 28. 17:39

논문 PDF - H-DenseUNet: Hybrid Densely Connected UNet for Liver and Liver Tumor Segmentation from CT Volumes

논문의 주제는 간의 종양을 segmentation 하는 것입니다. 제목에서 볼 수 있듯이 Unet 에 Densenet 을 넣어서 만든 모델로 DenseUnet 이라고 부릅니다. H-Densenet 에서 H 는 hydrid 로 2D 와 3D 를 합친 모델이기 때문에 H-Densenet 이라고 명명했습니다.

[2D 아키텍처]

loss 값도 상당히 낮고 정확도도 높기 때문에 2D 부분만 가져와서 brats 2015 데이터에 적용해 봤습니다.

class Denseunet_Model:

def __init__(self):

self.x_data = tf.placeholder(dtype=tf.float32,shape=[None,224,224,1])

self.y_data = tf.placeholder(dtype=tf.float32,shape=[None,224,224,2])

self.train_bool = tf.placeholder(dtype=bool)

self.class_num = 2

self.growth = 24

self.first_filter = 32

# result

self.logit = self.denseunet()

def denseunet(self):

def block(in_layer, out_filter, cnt, is_train):

for i in range(cnt):

temp_layer = batch_norm(in_layer, is_train)

temp_layer = tf.nn.leaky_relu(temp_layer)

temp_layer = ori_conv2(temp_layer, 4 * out_filter, 1, 1, 1, pad='same')

temp_layer = batch_norm(temp_layer, is_train)

temp_layer = tf.nn.leaky_relu(temp_layer)

temp_layer = ori_conv2( temp_layer, out_filter, 3, 3, 1, pad='same')

in_layer = tf.concat([in_layer, temp_layer], axis=3)

return in_layer

def transition(in_layer, out_filter, is_train):

in_layer = batch_norm(in_layer, is_train)

in_layer = tf.nn.leaky_relu(in_layer)

in_layer = ori_conv2(in_layer, out_filter, 1, 1, 1, pad='same')

in_layer = batch_norm(in_layer, is_train)

in_layer = tf.nn.leaky_relu(in_layer)

in_layer = tf.layers.average_pooling2d(in_layer, 2, 2, padding='same')

return in_layer

with tf.name_scope(name='down1') as scope:

down_layer = conv2(self.x_data, self.first_filter, 7, 7, 2, self.train_bool)

print(down_layer.shape)

down_layer_pool = tf.layers.max_pooling2d(down_layer, 3, 2, padding='same', name='down1_layer2')

print(down_layer_pool.shape)

with tf.name_scope(name='block1') as scope:

down_layer1 = block(down_layer_pool, self.growth, 4, self.train_bool)

print(down_layer1.shape)

with tf.name_scope(name='transition1') as scope:

down_transition1 = transition(down_layer1, down_layer1.shape[3]//2, self.train_bool)

print(down_transition1.shape)

with tf.name_scope(name='block2') as scope:

down_layer2 = block(down_transition1, self.growth, 8, self.train_bool)

print(down_layer2.shape)

with tf.name_scope(name='transition2') as scope:

down_transition2 = transition(down_layer2, down_layer2.shape[3]//2, self.train_bool)

print(down_transition2.shape)

with tf.name_scope(name='block3') as scope:

down_layer3 = block(down_transition2, self.growth, 24, self.train_bool)

print(down_layer3.shape)

with tf.name_scope(name='transition3') as scope:

down_transition3 = transition(down_layer3, down_layer3.shape[3]//2, self.train_bool)

print(down_transition3.shape)

with tf.name_scope(name='block4') as scope:

down_layer4 = block(down_transition3, self.growth, 16, self.train_bool)

print(down_layer4.shape)

with tf.name_scope(name='up4') as scope:

up_layer3 = up_conv2(down_layer4, down_layer2.shape[3], 2, 2, 2, self.train_bool)

print(up_layer3.shape)

up_layer3 = tf.concat([down_layer3, up_layer3], axis=3)

print(up_layer3.shape)

up_layer3 = conv2(up_layer3, up_layer3.shape[3], 3, 3, 1, self.train_bool)

print(up_layer3.shape)

with tf.name_scope(name='up3') as scope:

up_layer2 = up_conv2(up_layer3, down_layer1.shape[3], 2, 2, 2, self.train_bool)

print(up_layer2.shape)

up_layer2 = tf.concat([down_layer2, up_layer2], axis=3)

print(up_layer2.shape)

up_layer2 = conv2(up_layer2, up_layer2.shape[3], 3, 3, 1, self.train_bool)

print(up_layer2.shape)

with tf.name_scope(name='up2') as scope:

up_layer1 = up_conv2(up_layer2, down_layer.shape[3], 2, 2, 2, self.train_bool)

print(up_layer1.shape)

up_layer1 = tf.concat([down_layer1, up_layer1], axis=3)

print(up_layer1.shape)

up_layer1 = conv2(up_layer1, up_layer1.shape[3], 3, 3, 1, self.train_bool)

print(up_layer1.shape)

with tf.name_scope(name='up1') as scope:

up_layer = up_conv2(up_layer1, down_layer.shape[3], 2, 2, 2, self.train_bool)

print(up_layer.shape)

up_layer = tf.concat([down_layer, up_layer], axis=3)

print(up_layer.shape)

up_layer = conv2(up_layer, up_layer.shape[3], 3, 3, 1, self.train_bool)

print(up_layer.shape)

with tf.name_scope(name='up') as scope:

up_end = up_conv2(up_layer,int(self.first_filter * (2/3)),2,2,2,self.train_bool)