Advancing X-ray Tomography using Deep Generative ...

Advancing X-ray Tomography using Deep Generative Adversarial Networks (TomoGAN)

Zhengchun Liu

Research Scientist at the University of Chicago (now)

Assistant Computer Scientist at Data Science and Learning Division (soon)

July 25, 2019

PSE AI Townhall meeting at Argonne National Laboratory

Collaborators

Tekin Bicer DSL, XSD

Raj Kettimuthu DSL

Doga Gursoy XSD

Francesco De Carlo XSD

Ian Foster DSL

Full text: Liu et al. arXiv: 1902.07582

Motivation

(1) lower X-ray dosage for sensitive sample like bio-sample;

(2) faster experiment to capture dynamic features, like in fast chemical reaction processes;

(3) smaller dataset and less computation for [near] realtime tomography imaging.

On the left, the results of conventional reconstruction, which are highly noisy. On the right, those same results after denoising with TomoGAN.

Model is trained with one shale sample imaged at APS and tested with another shale sample imaged at Swiss Light Source (SLS).

Method

A generative adversarial network (GAN) is a class of machine learning systems in which two neural networks, generator (G) and discriminator (D), contest with each other in a game (in the sense of game theory, often but not always in the form of a zero-sum game).

Low dose projections

Normal dose projections

>tixetal/")llun("=46esab_1ahstixetaltixetal/")llun("=46esab_1ahstixetaltixetal/")llun("=46esab_1ahstixetaltixetal/")llun("=46esab_1ahstixetal< >tixetal/")llun("=46esab_1ahstixetaltixetal/")llun("=46esab_1ahstixetaltixetal/")llun("=46esab_1ahstixetaltixetal/")llun("=46esab_1ahstixetal< >tixetal/")llun("=46esab_1ahstixetaltixetal/")llun("=46esab_1ahstixetaltixetal/")llun("=46esab_1ahstixetaltixetal/")llun("=46esab_1ahstixetal< >tixetal/")llun("=46esab_1ahstixetaltixetal/")llun("=46esab_1ahstixetaltixetal/")llun("=46esab_1ahstixetaltixetal/")llun("=46esab_1ahstixetal<

IL[iD

b

d 2

c,i+b

d 2

c]

Tomographic Reconstruction

Generator

Tomographic INi D Reconstruction

Iteration

back propagation and updating weights

Denoised image

Pixel L2 Loss LMSE Content

Pre-trained

VGG

LV GG

Distance

Adversarial

Loss

G Loss

Discriminator

Wasserstein Distance

D Loss

back propagation and updating weights

In our model, the discriminator's job remains unchanged, but the generator is tasked not only with fooling (indistinguishable) the discriminator but also with being near the ground truth output in an L2 sense.

The discriminator works as a helper to train the generator that we need to denoise images.

Adjacent d d noisy images

8 32 32 Down sampling

Up sampling

64 32 32 16 1 Enhanced image

1x1

C

relu

Legend 1024 x 1024

2x2 Max pooling

2x Bilinear upsampling

3x3

C 3x3 Conv. + ReLU

1x1

C 1x1 Convolution + ReLU

relu 1x1

C 1x1 Convolution + Linear

tanh

10242 10242 10242 10242

3x3 3x3

CC

64 64

3x3 3x3

CC

Copy Copy

3x3 3x3

CC

1x1 1x1

CC

relu tanh

1024 x 1024

10242

m 10242 10242 10242 10242

128 32 32

3x3 3x3

CC

5122 5122 5122

5122 5122 5122

3x3 1283x3 128

CC

Copy

256 3x3 643x364

CC

2562 2562 2562

2562 2562 2562

128 3x3

128

C

1282

1282

Our Generator Architecture

Training

Discriminator

Wasserstein GAN [1] + gradient penalty [2]

L (D)

=

1 m

m

i=1

[D

(G

(ILi D))

-

D

(INi D)]

+

D

1 m

m

i=1

[(

I D (Ii)

2

2 - 1) ],

Generator

Weighted average of Adversarial loss, Perceptual loss, and Pixel-wise MSE

G = gadv + pmse + vvgg

adv (G)

=

-

1 m

m

D (G (ILi D))

i=1

Wf Hf

2

vgg

=

i=1

i=1

(Vvgg

(IND)i,j

-

Vvgg

(GG

(ILD))i,j)

WH

2

mse = (IcN,rD - GG (ILD)c,r)

c=1 r=1

[1] Wasserstein GAN. M. Arjovsky, S. Chintala, L. Bottou. arXiv:1701.07875

[2] Improved Training of Wasserstein GANs. I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, A. Courville. arXiv:1704.00028

 ................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download