Automatic determination of b-values from diffusion-weighted magnetic resonance images
A mechanism is provided in a data processing system for automatic determination of b-value difference from diffusion-weighted (DW) images. The mechanism receives a series of images wherein a first image has a first b-value and a second image has an unknown b-value. The mechanism applies a generative adversarial network (GAN) model to estimate a difference between b-values in the series of images. The mechanism determines a b-value for the second image based on the first b-value and the estimated difference between b-values.
1. A method, in a data processing system, for automatic determination of b-value difference from diffusion-weighted (DW) images, the method comprising:
receiving a series of DW images wherein a first image has a first b-value and a second image has an unknown b-value;
applying a generative adversarial network (GAN) model to estimate a difference between b-values in the series of DW images; and
determining a b-value for the second image based on the first b-value and the estimated difference between b-values.
2. The method of claim 1 , wherein applying the GAN model comprises constructing a latent space with a latent variable for learning to generate a synthetic apparent diffusion coefficient (ADC) map for identifying unknown b-values in the series of DW images and a loss function based on a relationship between the ADC map and the spatial relationship between images in the series of DW images.
3. The method of claim 2 , wherein constructing the latent space comprises applying encoders to the images in the series of DW images.
4. The method of claim 2 , wherein the latent space comprises a latent variable and a difference value representing a difference between b-values.
5. The method of claim 1 , wherein the GAN model comprises a generator, a discriminator, and an estimator.
6. The method of claim 5 , wherein GAN model has an identify-preserving loss (L identity ) function as follows:
L
identity
=
(
2
α
Z
α
S
A
D
C
Δ
b
)
-
1
❘
"\[LeftBracketingBar]"
❘
"\[LeftBracketingBar]"
Z
-
E
(
S
A
D
C
Δ
b
)
❘
"\[RightBracketingBar]"
❘
"\[RightBracketingBar]"
2
2
where the α values are parameters for normalization and are as follows:
α x =σ x −1 ∥x−μ x ∥.
7. The method of claim 5 , wherein the GAN model has a pixel loss (L pixel ) function as follows:
L
pixel
=
1
Δ
b
×
L
×
W
×
H
G
(
z
,
Δ
b
)
-
❘
"\[LeftBracketingBar]"
ln
(
S
b
2
/
S
b
1
)
❘
"\[RightBracketingBar]"
2
2
.
8. The method of claim 5 , wherein the GAN model has a regression loss (L regression ) function as follows:
L regression = z˜p z ∥R ( G ( z,Δb ))−Δ b∥ 2 2 .
9. The method of claim 5 , wherein the GAN model has discriminator and generator losses (L GAN-D , L GAN-G ) as follows:
L GAN-D = z,Δb˜p (z,Δb) [log D ( z,Δb )]+ z,Δb˜p (z,Δb) [−log(1− D ( G ( z,b )))]
L GAN-G = z,Δb˜p (z,Δb) [log(1− D ( G ( z,Δb )))].
10. The method of claim 5 , wherein the GAN model has objective functions for the generator and the discriminator as follows:
L G =γ i L identity +γ p L pixel +γ r L regression +γ g L GAN-G
L D =L GAN-D ,
wherein the objective is to minimize both objective functions L G and L D .
11. A computer program product comprising a computer readable storage medium having a computer readable program stored therein, wherein the computer readable program, when executed on a computing device, causes the computing device to:
receive a series of DW images wherein a first image has a first b-value and a second image has an unknown b-value;
apply a generative adversarial network (GAN) model to estimate a difference between b-values in the series of DW images; and
determine a b-value for the second image based on the first b-value and the estimated difference between b-values.
12. The computer program product of claim 11 , wherein applying the GAN model comprises constructing a latent space with a latent variable for learning to generate a synthetic apparent diffusion coefficient (ADC) map for identifying unknown b-values in the series of DW images and a loss function based on a relationship between the ADC map and the spatial relationship between images in the series of DW images.
13. The computer program product of claim 12 , wherein constructing the latent space comprises applying encoders to the images in the series of DW images and wherein the latent space comprises a latent variable and a difference value representing a difference between b-values.
14. The computer program product of claim 11 , wherein the GAN model comprises a generator, a discriminator, and an estimator.
15. The computer program product of claim 14 , wherein GAN model has an identify-preserving loss (L identity ) function as follows:
L
identity
=
(
2
α
Z
α
S
A
D
C
Δ
b
)
-
1
❘
"\[LeftBracketingBar]"
❘
"\[LeftBracketingBar]"
Z
-
E
(
S
A
D
C
Δ
b
)
❘
"\[RightBracketingBar]"
❘
"\[RightBracketingBar]"
2
2
where the α values are parameters for normalization and are as follows:
α x =σ x −1 ∥x−μ x ∥.
16. The computer program product of claim 14 , wherein the GAN model has a pixel loss (L pixel ) function as follows:
L
pixel
=
1
Δ
b
×
L
×
W
×
H
G
(
z
,
Δ
b
)
-
❘
"\[LeftBracketingBar]"
ln
(
S
b
2
/
S
b
1
)
❘
"\[RightBracketingBar]"
2
2
.
17. The computer program product of claim 14 , wherein the GAN model has a regression loss (L regression ) function as follows:
L regression = z˜p z ∥R ( G ( z,Δb ))−Δ b∥ 2 2 .
18. The computer program product of claim 14 , wherein the GAN model has discriminator and generator losses (L GAN-D , L GAN-G ) as follows:
L GAN-D = z,Δb˜p (z,Δb) [−log D ( z,Δb )]+ z,Δb˜p (z,Δb) [−log(1− D ( G ( z,b )))]
L GAN-G = z,Δb˜p (z,Δb) [log(1− D ( G ( z,Δb )))].
19. The computer program product of claim 14 , wherein the GAN model has objective functions for the generator and the discriminator as follows:
L G =γ i L identity +γ p L pixel +γ r L regression +γ g L GAN-G
L D =L GAN-D ,
wherein the objective is to minimize both objective functions L G and L D .
20. An apparatus comprising:
a processor; and
a memory coupled to the processor, wherein the memory comprises instructions which, when executed by the processor, cause the processor to:
receive a series of DW images wherein a first image has a first b-value and a second image has an unknown b-value;
apply a generative adversarial network (GAN) model to estimate a difference between b-values in the series of DW images; and
determine a b-value for the second image based on the first b-value and the estimated difference between b-values.