imlab-uiip / lung-segmentation-2d

Lung fields segmentation on CXR images using convolutional neural networks.

Lung Segmentation (2D)

Repository features UNet inspired architecture used for segmenting lungs on chest X-Ray images.

Demo

See the application of the model in Demo folder.

Implementation

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

Use of data augmentation for training required slight changes to keras ImageDataGenerator. Generator in image_gen.py applies same transformation to both the image and the label mask.

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 contains model trained on both data sets mentioned below.

Segmentation

Scores achieved on Montgomery and JSRT(With these masks. See preprocess_JSRT.py.) (Measured using 5-fold cross-validation):

	JSRT	Montgomery
IoU	0.971	0.956
Dice	0.985	0.972

lung-segmentation-2d's Issues

`preprocess_JSRT.py` cannot `reshape`

I'm guessing output of preprocess_JSRT.py must be used as input images for train_model.py. Is this correct? If yes, then on line 8, reshaping the ndarray to (2048, 2048) gives ValueError: cannot reshape array of size ${size_of_ndarray} into shape (2048,2048) on all original images from JSRT.
Furthermore, I don't really follow why img = 1.0 - np.fromfile(path + filename, dtype='>u2') * 1. / 4096 (changing the values like that, regardless of the reshaping) is happening as preprocessing. Could you please give a quick explanation?

Dataset

please tell me where I can get the lung-dataset,thank you !

Value error while running model.fit

Did you run into any such errors?

ValueError: A target array with shape (8, 512, 512, 3) was passed for an output of shape (None, 512, 512, 1) while using as loss binary_crossentropy. This loss expects targets to have the same shape as the output.

License

Hey there, thanks for the great repo.

What is the license for code reuse?

Cheers

The Concat layer have dimension problem.

The Concat layer has a dimension problem. The layer to be merged is of different dimension (I have just read the build model function)

Hi, what data did you use for training?

I see that you used Montgomery for test but you have other datasets as well in your load_data.py.
Could you mention which ones?

Training on different structures

Thank you for sharing this. Is there anyway to go about if you want to train the model to recognize several structures (e.g. lungs and heart)?

How to combine the resulting hdf5 models into one?

I get models 10- model 100, but I can’t figure out how to combine all of them into trained_model.

what is in th 'idx.cvs' and 'left lung'?

I don't Know what kind of data for this code.
csv_path = '/path/to/JSRT/idx.csv'
what kind of data in idx.cvs?
Does 'filenames' include suffix?
and what's in 'left lung' ?
left = io.imread('/path/to/JSRT/Masks/left lung/' + filename[:-4] + '.gif')
right = io.imread('/path/to/JSRT/Masks/right lung/' + filename[:-4] + '.gif')

idx.csv

Where can i get idx.csv for the JSRT's database?
I only downloaded "All247images.zip" and the SCR's data。

Implementing the model on Montgomery Dataset

Thank you for sharing the code for lung segmentation here. I am trying to implement the code for Montgomery Dataset. Do I simply change JSRT with Montgomery in order to do that? Also, there is no csv file for this dataset, do I have to create it manually?