imlab-uiip/lung-segmentation-3d

Lung fields segmentation on tomography images using convolutional neural networks.

Lung Segmentation (3D)

Repository features UNet inspired architecture used for segmenting lungs on chest 3D tomography images.

Demo

Run inference.py to see the application of the model on Demo files.

Implementation

Implemented in Keras(2.0.4) with TensorFlow(1.1.0) as backend.

To use this implementation one needs to load and preprocess data (see load_data.py), train new model if needed (train_model.py) and use the model for generating lung masks (inference.py).

trained_model.hdf5 and trained_model_wc.hdf5 contain models trained on private data set without and with coordinates channels.

Segmentation

lung-segmentation-3d's Issues

Will 50 epoch be enough？

hi. eclique. What is the numbers did you use？
Is it enough epoch = 50 ？

load model problem when run inference.py

Hello,
I also cannot run inference.py directly without changing a thing.
And then I trained a new model , but it also cannot load successfully.

I run this on keras 2.0.8, tensorflow 1.0.0.
Can you help me with this problem? Thanks a lot.

do you have some suggestions in this situation of difference size data?

hello. eclique. thank you for your sharing of the codes. I have a question. If the size of my input data is not your data size of 12812864, I have to crop or reshape my data into this size for training. But I feel this method might miss some important detailed information. So, do you have some suggestions in this situation of difference size data?

append coordinates?

hi. eclique. i do not understand your code about appending coordinates (i cut the code from this Repository). because i think xx and yy should exchange their coordinates. could you please tell me why do you do that? why not xx[i, :, :, 0] = i and yy[:, i, :, 0] = i ?
if append_coord:
n = X.shape[0]
inpShape = X.shape[1:]
xx = np.empty(inpShape)
for i in xrange(inpShape[1]):
xx[:, i, :, 0] = i
yy = np.empty(inpShape)
for i in xrange(inpShape[0]):
yy[i, :, :, 0] = i
zz = np.empty(inpShape)
for i in xrange(inpShape[2]):
zz[:, :, i, 0] = i
X = np.concatenate([X, np.array([xx] * n), np.array([yy] * n), np.array([zz] * n)], -1)

data acquisition question

Hi,

I checked your 3D lung segmentation repository and it is really amazing. We are now doing research on organ segmentation and believe your repository is a very good reference.

I see you mentioned in ReadMe that you are using private dataset and I guess we may not able to obtain the same dataset as you used? But also I find that there are some requirement on data format and ground truth file format, that I would like to ask whether you know any public dataset might have similar format as the one you used for this repository?

Thanks,
Jin

how can i download the datasets? ask for help!

the inference just can not run in my computer, and the demo contains a litter datasets....

slice indices must be integers or None or have an __index__ method

Hi
trying to reproduce the result.
As "from scipy.misc import imresize" can't be used any more .... I am trying to using
from skimage.transform import resize.

changes

lungs[:img_size, :img_size] = imresize(saggital(t_img), [img_size, img_size]) * 1. / 256

mask[:img_size, :img_size][imresize(saggital(pred[i]), [img_size, img_size]) > 128] = 1

gt_mask[:img_size, :img_size][imresize(saggital(y[i][..., 1]), [img_size, img_size]) > 128] = 1

lungs[:img_size, :img_size] = resize(saggital(t_img), [img_size, img_size]) * 1. / 256
mask[:img_size, :img_size][resize(saggital(pred[i]), [img_size, img_size]) > 128] = 1
gt_mask[:img_size, :img_size][resize(saggital(y[i][..., 1]), [img_size, img_size]) > 128] = 1

got below error:
t_img = exposure.rescale_intensity(nib.load(path + df.ix[i]['path']).get_data(), out_range=(0, 1))
Traceback (most recent call last):
File "/user/iibi/skuanar/Downloads/lung-segmentation-3d/lung-segmentation-3d-master/inference.py", line 101, in
lungs[yy * img_size: (yy + 1) * img_size, xx * img_size: (xx + 1) * img_size] = t_img[:, :, -zz]
TypeError: slice indices must be integers or None or have an index method

l want to know how many accuracy about the model.

np.concatenate((1 - y, y), -1)?

hi. eclique. After reading your code, I don't understand much about it. I hope to get your advice. I don't quite understand 'np.concatenate (1-y, y),-1)',What does this code mean?and Why do you want to do the reverse operation?I implore your answer.
y = np.expand_dims(y, -1)
y = np.array(y)
y = np.concatenate((1 - y, y), -1)
y = np.array(y)

Questions about loading model

Hello,

I cannot run inference.py directly without changing a thing. I run in python3, and I do not know if it is the version that matters? It seems that there exists problem in loading the model, I guess.

Can you help me with my problem? I am a machine learning beginner and it may be a stupid problem and I will be glad to get your feedback!!

incomplete segmentation

Hi,
Have you seen any hole within the segmentation due to nodules or artifact? If yes, how do you get the perfect filled-out mask?

what's the effect of 'append_coords' in dataload.py

Hey, many thanks to you. I have read your code ,I can not understand the effect of the parameter ‘append_coords’ in dataload.py. Can you tell me how to use this paramater.