Regarding subimage patch and label size about brats17 HOT 7 CLOSED

To save memeory, the training and testing were based on image patches, not the entire image size. The convolution in the z-axis was based on 'valid' mode, that's why the output size is reduced by 8 in z-axis.

from brats17.

+1

from brats17.

+1
That is so wired!
I use the dataset of BRATS15, the original data shape is 155 * 240 * 240, but the sub image shape is 19 * 144 * 144.

According to his function ,
center_point = get_random_roi_sampling_center(volume_shape, sub_label_shape, batch_sample_model, boundingbox) #

and also
sub_data_moda = extract_roi_from_volume(transposed_volumes[moda],center_point,sub_data_shape)
it seems that your randomly get a sub image by cropping the original image, which I think may miss some information of the tumor.
How do you guarantee this cropping method will get all the tumor information we want?

from brats17.

I love this question, I run the MSNet and get 91.4% in whole tumor segmentation on BRATS2015 with this (19, 144, 144, 4), but I don't understand why 19 and 11.

from brats17.

I love this question, I run the MSNet and get 91.4% in whole tumor segmentation on BRATS2015 with this (19, 144, 144, 4), but I don't understand why 19 and 11.

Well, it is the problem of the model if self !
You could revise util/MSNet.py like

if __name__ == '__main__':
    x = tf.placeholder(tf.float32, shape = [1, 96, 96, 96, 1])
    y = tf.placeholder(tf.float32, shape = [1, 96, 96, 96, 2])
    net = MSNet(num_classes=2)
    predicty = net(x, is_training = True)
    print(x)
    print(predicty)
    print (Y)

ant run it like

python util/MSNet.py

You will find that the result is

shape .....
(1, 96, 96, 96, 1)
(1, 88, 96, 96, 2)
(1, 96, 96, 96, 2)

I hope this could help you solve the problem

from brats17.

I love this question, I run the MSNet and get 91.4% in whole tumor segmentation on BRATS2015 with this (19, 144, 144, 4), but I don't understand why 19 and 11.

Well, it is the problem of the model if self !
You could revise util/MSNet.py like

if __name__ == '__main__':
    x = tf.placeholder(tf.float32, shape = [1, 96, 96, 96, 1])
    y = tf.placeholder(tf.float32, shape = [1, 96, 96, 96, 2])
    net = MSNet(num_classes=2)
    predicty = net(x, is_training = True)
    print(x)
    print(predicty)
    print (Y)

ant run it like

python util/MSNet.py

You will find that the result is

shape .....
(1, 96, 96, 96, 1)
(1, 88, 96, 96, 2)
(1, 96, 96, 96, 2)

I hope this could help you solve the problem

我又思考了一遍，沿着axial轴截了19（155），但是后续沿着coronal, sagittal方向，也是截了19，三个长方体叠起来，是不是覆盖了大部分范围。

from brats17.

I love this question, I run the MSNet and get 91.4% in whole tumor segmentation on BRATS2015 with this (19, 144, 144, 4), but I don't understand why 19 and 11.

Well, it is the problem of the model if self !
You could revise util/MSNet.py like

if __name__ == '__main__':
    x = tf.placeholder(tf.float32, shape = [1, 96, 96, 96, 1])
    y = tf.placeholder(tf.float32, shape = [1, 96, 96, 96, 2])
    net = MSNet(num_classes=2)
    predicty = net(x, is_training = True)
    print(x)
    print(predicty)
    print (Y)

ant run it like
python util/MSNet.py
You will find that the result is

shape .....
(1, 96, 96, 96, 1)
(1, 88, 96, 96, 2)
(1, 96, 96, 96, 2)

I hope this could help you solve the problem

我又思考了一遍，沿着axial轴截了19（155），但是后续沿着coronal, sagittal方向，也是截了19，三个长方体叠起来，是不是覆盖了大部分范围。

yeah! As you know, the shape of raw data is 155 * 240 * 240, but we only randomly select some data with shape of 19 * 144 * 144 from the raw data. This is because with a large iteration (like here is 20000), we can cover the whole data （155 * 240 * 240）probabilistically speaking.

I guess he did so because it could help save the memory use when training or testing!

from brats17.