It seems that the BN layer has some problems for 3D cases. It does not work as I expected, which may affect the accuracy. I'll try to fix it.

Thanks for update your work. I try to run your work in my machine and I got some errors and this is how I fix it.

1. python main.py --action='predict'/'test' changed to python main.py --option='predict'/'test' in ReadMe

2. from utils.* in network.py , you may get error ImportError: No module named utils.*. To fix it, you just add a file __init__.py in the folder utils. You don't need to put anything into the file.

3. When you perform python main.py --option='predict'. You may get the error about inconsistent shape. (5x32x32x32x2) to the target shape (1x32x32x32x2) . I did not know how to fix it. :(

4. For generating the hdf5, I think we should you 9 subjects, instead of 10 subject as you did. Because you will use one subject for validation. Hence, the validation subjects should not include in the hdf5 data that used for training.

Thanks!

@zhengyang-wang
I am interested in this repo unet_3d, I'd like to run it, but no input file, I was wondering how can I make a simple dataset, such as generating a random input tensor rather than reading from the file, so that I could run it well?

As title, it seems that in test/predict mode the output of batch_norm is very weird, I believe it is due to the moving mean and variances are not saved.

so in network.py use the following code:

def configure_networks(self):
self.build_network()
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
optimizer = tf.train.AdamOptimizer(self.conf.learning_rate)
self.train_op = optimizer.minimize(self.loss_op, name='train_op')
tf.set_random_seed(self.conf.random_seed)
self.sess.run(tf.global_variables_initializer())
self.saver = tf.train.Saver()
self.writer = tf.summary.FileWriter(self.conf.logdir, self.sess.graph)

insted of using :
trainable_vars = tf.trainable_variables()
self.saver = tf.train.Saver(var_list=trainable_vars, max_to_keep=0)
use :
self.saver = tf.train.Saver()
so that the moving mean and variace are saved.

also, for batch_norm function, use a train_flag and put it into a bool placeholder(self.phase), remember to modify all calls to conv, deconv, pixel_dcl, ipixel_cl, ipixel_dcl functions:
in network.py:
def build_network(self):
......
self.phase = tf.placeholder(tf.bool) #define a bool placeholder
...
conv1 = ops.conv(inputs, out_num, self.conv_size,
name+'/conv1', self.conf.data_type, train_flag=self.phase)
...

in ops.py:
def conv(inputs, out_num, kernel_size, scope, data_type='2D', norm=True, train_flag=False):
......
if norm:
return tf.contrib.layers.batch_norm(
outs, decay=0.9, epsilon=1e-5, activation_fn=tf.nn.relu,
updates_collections=tf.GraphKeys.UPDATE_OPS, scope=scope+'/batch_norm', is_training=train_flag)
......

when train:
feed_dict = {self.inputs: inputs,
self.annotations: annotations,
self.phase: True}

when test/predict:
feed_dict = {self.inputs: inputs,
self.annotations: annotations,
self.phase: False}

Thanks for sharing your work!

I would like to ask you two questions:

1. In your preprocessing>generate_h5.py, you have

      if SUBSTRACT_MEAN:
           inputs -= MEAN
       if NORMALIZE:
           inputs /= 1000.

What is value 1000? Is it maximum intensity? As my knowledge, normalize with zero mean and unit variance must be (inputs-MEANS)/std, hence the value 1000 must be standard variance value. Is it right?

2. I am regarding your performance with/without data augment. In my testing, with/without augment data have the same performance. How about you?