tensorlayer / dagan Goto Github PK

The implementation code for "DAGAN: Deep De-Aliasing Generative Adversarial Networks for Fast Compressed Sensing MRI Reconstruction"

Python 100.00%

generative-adversarial-network mri-reconstruction computer-vision deep-learning

dagan's Introduction

DAGAN

This is the official implementation code for DAGAN: Deep De-Aliasing Generative Adversarial Networks for Fast Compressed Sensing MRI Reconstruction published in IEEE Transactions on Medical Imaging (2018).
Guang Yang*, Simiao Yu*, et al.
(* equal contributions)

If you use this code for your research, please cite our paper.

@article{yang2018_dagan,
	author = {Yang, Guang and Yu, Simiao and Dong, Hao and Slabaugh, Gregory G. and Dragotti, Pier Luigi and Ye, Xujiong and Liu, Fangde and Arridge, Simon R. and Keegan, Jennifer and Guo, Yike and Firmin, David N.},
	journal = {IEEE Trans. Med. Imaging},
	number = 6,
	pages = {1310--1321},
	title = {{DAGAN: deep de-aliasing generative adversarial networks for fast compressed sensing MRI reconstruction}},
	volume = 37,
	year = 2018
}

If you have any questions about this code, please feel free to contact Simiao Yu ([email protected]).

Prerequisites

The original code is in python 3.5 under the following dependencies:

tensorflow (v1.1.0)
tensorlayer (v1.7.2)
easydict (v1.6)
nibabel (v2.1.0)
scikit-image (v0.12.3)

Code tested in Ubuntu 16.04 with Nvidia GPU + CUDA CuDNN (whose version is compatible to tensorflow v1.1.0).

How to use

Prepare data
1. Data used in this work are publicly available from the MICCAI 2013 grand challenge (link). We refer users to register with the grand challenge organisers to be able to download the data.
2. Download training and test data respectively into data/MICCAI13_SegChallenge/Training_100 and data/MICCAI13_SegChallenge/Testing_100 (We randomly included 100 T1-weighted MRI datasets for training and 50 datasets for testing)
3. run 'python data_loader.py'
4. after running the code, training/validation/testing data should be saved to 'data/MICCAI13_SegChallenge/' in pickle format.
Download pretrained VGG16 model
1. Download 'vgg16_weights.npz' from this link
2. Save 'vgg16_weights.npz' into 'trained_model/VGG16'
Train model
1. run 'CUDA_VISIBLE_DEVICES=0 python train.py --model MODEL --mask MASK --maskperc MASKPERC' where you should specify MODEL, MASK, MASKPERC respectively:
- MODEL: choose from 'unet' or 'unet_refine'
- MASK: choose from 'gaussian1d', 'gaussian2d', 'poisson2d'
- MASKPERC: choose from '10', '20', '30', '40', '50' (percentage of mask)
Test trained model
1. run 'CUDA_VISIBLE_DEVICES=0 python test.py --model MODEL --mask MASK --maskperc MASKPERC' where you should specify MODEL, MASK, MASKPERC respectively (as above).

Results

Please refer to the paper for the detailed results.

dagan's People

Contributors

Stargazers

Watchers

dagan's Issues

Error while loading data_loader.py

while loading data_loader.py it showing up dimension error in " img_2d = np.transpose(img_2d, (1, 0))" . Also while trying to do after reducing number of data set for fast training it shows up error
"Traceback (most recent call last):
File "sef_data_loader.py", line 94, in
X_train = X_train[:, :, :, np.newaxis]
IndexError: too many indices for array
"

how to calculate psnr in your paper?

After reading your paper and codes, i'm strange about the calculation of PSNR. I use some mri data to test the ZF reconstruction and try to calculate the NMSE in your codes, it seems right. But when I calculate the PSNR based on your NMSE or RMSE defined by myself, it seems wrong. What's more, don't you think the PSNR is too high via ZF reconstruction with 20% data kept, because when PSNR reaches 34dB, pictures may seem very similar.

Is this applicible on different type of dataset?

hello,
my question is that can we apply this model on generating the new dataset which related to other applications? such as Phase contrast microscopy images such as live cells dataset?

Some questions about the vgg_prepro function in utils.py files

Issue Description

Hello,
when I read your code, I have some questions about the vgg_prepro functions in utils.py
This function is defined as follows：

def vgg_prepro(x):
    x = imresize(x, [244, 244], interp='bilinear', mode=None)
    x = np.tile(x, 3)
    x = x / 127.5 - 1
    return x

I know this function is to preprocess an image, such as to Changing an image from [256,256,1] to [244,244,3], but I don't understand why do we need to use x/127.5 - 1 , here we have already to scale the image to [-1,1].

In addition, I would also like to ask why we cut the size of 244 instead of 224 which is the size of the initial VGG paper.

I was hoping you could help me with this problem, I can't get it right.

Updated Version of Code

Hi Do you have a updated version of the code?

Since Tensorflow 1.0 is too old and I am suffering from a layer building problem.

Thanks,
Junyu

How to calculate the reconstruction time

Hi, @nebulaV, I run your code and reconstruct one image, I use the following code during evaluate time,

    start_time = time.time()
    evaluate_restore_img = sess.run(net.outputs, {evaluate_image: evaluate_samples_bad})
    print("took: %4.4fs" % (time.time() - start_time))

I run on the GPU platform, but the time is about 16s, but your code is 5.4ms , I don not understand how to calculate that results.

Data size and range

What is the size and value range of HR images in your training?

FHD(1920x1080) size image

Hello. Your paper is very helpful and thankful.
However, I want to learn a bigger size image. I need advice on how to do it.

Can you tell me how to configure the network?

For reference, I am studying to eliminate aliasing of Aliasing images with Aliasing and Ground truth images.

name 'tl' is not defined

Hello,

I had come accross your work and was very excited to run it. Unfortunately I keep getting an error message "name 'tl' is not defined", this occurs when I run train.py here https://github.com/tensorlayer/DAGAN/blob/master/train.py#L475, I ran this according to the README. It seems that this t1 object has many inbuilt functions so doesnt look like I can get away with commenting all the t1 lines. Can you please help ?

questions about mask

could you share the code how to create mask such as Gauss2D?

up7 layer's output is not equal to input.

Hello, May I ask a problem?
Thank you for your code, I have been learn AI from about 3 weeks ago.
During training, It occur the error above like title.
Help me, please.

up7 = {DeConv2d} Last layer is: DeConv2d (u_net/deconv7) [25, 2, 2, 512]
all_drop = {dict} {}
all_layers = {list} <class 'list'>: [<tf.Tensor 'u_net/conv1/Identity:0' shape=(25, 1, 1, 64) dtype=float32>, <tf.Tensor 'u_net/conv2/Identity:0' shape=(25, 1, 1, 128) dtype=float32>, <tf.Tensor 'u_net/bn2/lrelu:0' shape=(25, 1, 1, 128) dtype=float32>, <tf.Tensor 'u_net/conv3/Identity:0' shape=(25, 1, 1, 256) dtype=float32>, <tf.Tensor 'u_net/bn3/lrelu:0' shape=(25, 1, 1, 256) dtype=float32>, <tf.Tensor 'u_net/conv4/Identity:0' shape=(25, 1, 1, 512) dtype=float32>, <tf.Tensor 'u_net/bn4/lrelu:0' shape=(25, 1, 1, 512) dtype=float32>, <tf.Tensor 'u_net/conv5/Identity:0' shape=(25, 1, 1, 512) dtype=float32>, <tf.Tensor 'u_net/bn5/lrelu:0' shape=(25, 1, 1, 512) dtype=float32>, <tf.Tensor 'u_net/conv6/Identity:0' shape=(25, 1, 1, 512) dtype=float32>, <tf.Tensor 'u_net/bn6/lrelu:0' shape=(25, 1, 1, 512) dtype=float32>, <tf.Tensor 'u_net/conv7/Identity:0' shape=(25, 1, 1, 512) dtype=float32>, <tf.Tensor 'u_net/bn7/lrelu:0' shape=(25, 1, 1, 512) dtype=float32>, <tf.Tensor 'u_net/conv8/lrelu:0' shape=(25, 1, 1, 51...
all_params = {list} <class 'list'>: [<tf.Variable 'u_net/conv1/kernel:0' shape=(4, 4, 1, 64) dtype=float32_ref>, <tf.Variable 'u_net/conv1/bias:0' shape=(64,) dtype=float32_ref>, <tf.Variable 'u_net/conv2/kernel:0' shape=(4, 4, 64, 128) dtype=float32_ref>, <tf.Variable 'u_net/conv2/bias:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'u_net/bn2/beta:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'u_net/bn2/gamma:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'u_net/bn2/moving_mean:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'u_net/bn2/moving_variance:0' shape=(128,) dtype=float32_ref>, <tf.Variable 'u_net/conv3/kernel:0' shape=(4, 4, 128, 256) dtype=float32_ref>, <tf.Variable 'u_net/conv3/bias:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'u_net/bn3/beta:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'u_net/bn3/gamma:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'u_net/bn3/moving_mean:0' shape=(256,) dtype=float32_ref>, <tf.Variable 'u_net/bn3/moving_variance:0' shape=(256,) dtype=float3...

inputs = {Tensor} Tensor("u_net/conv8/lrelu:0", shape=(25, 1, 1, 512), dtype=float32)
name = {str} 'u_net/deconv7'
outputs = {Tensor} Tensor("u_net/deconv7/Identity:0", shape=(25, 2, 2, 512), dtype=float32)
w_init = {TruncatedNormal} <tensorflow.python.ops.init_ops.TruncatedNormal object at 0x7fd754f26358>
x = {Tensor} Tensor("bad_image:0", shape=(25, 1, 1, 1), dtype=float32)

Questions about the weight for pixel loss

Hello,

I noticed that you use 15 as the weight for pixel loss, which is much larger than other weights such as for perceptual loss, frequency loss and also generator loss. If such a large coefficient is used, the influence of the discriminator will be reduced and the generator might become a direct estimator approximately. I would like to know if mode collapse can happen when the large weight for pixel loss is used.

BTW, I implemented conditional WGAN for MRI without pixel loss, but the image quality is not good enough.

Thanks!