Method of quantifying magnetic resonance diffusion parameters by using diffusion weighted magnetic resonance images
Disclosed is an unsupervised deep learning method, which simultaneously performs registration between diffusion weighted magnetic resonance images and quantification of diffusion parameters. The unsupervised deep learning method decreases a registration error caused by a contrast difference by comparing a similarity between images of the same contrast, increases the performance of registration, and increases the accuracy of quantifying magnetic resonance diffusion parameters.
1. A method of quantifying magnetic resonance diffusion parameters, as a method of quantifying the magnetic resonance diffusion parameters by using diffusion weighted magnetic resonance images, the method comprising:
(i) preparing a reference image and diffusion weighted magnetic resonance images;
(ii) controlling the reference image to be input to first neural networks and the diffusion weighted magnetic resonance images to be respectively input to the first neural networks, wherein the first neural networks are configured to estimate transformation functions representing a spatial transformation to the reference image from the diffusion weighted magnetic resonance images respectively corresponding to the first neural networks, and the first neural networks are configured to respectively provide first output images in response to the input of the reference image and the diffusion weighted magnetic resonance images respectively corresponding to the first neural networks;
(iii) controlling the first output images to be input to second neural networks, wherein the second neural networks are configured to provide at least one second output image in response to the input of the first output images;
(iv) providing third output images of which the number is the same as the number of first output images by applying intravoxel incoherent motion-diffusion kurtosis imaging (IVIM-DKI) to the reference image and the at least one second output image;
(v) providing fourth output images by applying inverse functions of the transformation functions to the third output images;
(vi) updating the first and second neural networks by using the diffusion weighted magnetic resonance images, the first output images, the third output images, and the fourth output images; and
(vii) repeating operations (ii) to (vi) a plurality of times,
wherein, the first neural networks include one or more neural networks, each of which is distinguished based on the direction of diffusion gradient and the second neural networks include one or more neural networks, each of which is for quantifying diffusion parameters.
2. The method of claim 1 , wherein
the reference image is a magnetic resonance imaging (MRI) image obtained in the setting of a b value to 0[s/mm 2 ], and
the diffusion weighted magnetic resonance images are MRI images obtained while changing a b value to different values except 0[s/mm 2 ].
3. The method of claim 1 , wherein,
as operation (vii) is performed, the first output images get closer to the diffusion weighted magnetic resonance images and closely registered with the reference image.
4. The method of claim 1 , wherein
the diffusion weighted magnetic resonance images are respectively generated by using diffusion weighted magnetic resonance images each obtained in a readout direction, in a phase encoding direction, and in a slice selection direction.
5. The method of claim 1 , wherein
the at least one second output image is a diffusion parameter image representing values of diffusion parameters on biometrics represented by the diffusion weighted magnetic resonance images.
6. The method of claim 5 , wherein
the at least one second output image comprises at least one of:
a diffusion coefficient image representing values of diffusion coefficients D on the biometrics represented by the diffusion weighted magnetic resonance images;
a perfusion coefficient image representing values of perfusion coefficients D p on the biometrics represented by the diffusion weighted magnetic resonance images;
a kurtosis image representing values of kurtoses K on the biometrics represented by the diffusion weighted magnetic resonance images; and
a perfusion fraction image representing values of perfusion fractions f on the biometrics represented by the diffusion weighted magnetic resonance images.
7. The method of claim 1 , wherein, as operation (vii) is performed, the third output images get closer to the first output images.
8. The method of claim 1 , wherein, as operation (vii) is performed, the fourth output images get closer to the diffusion weighted magnetic resonance images.
9. The method of claim 1 , wherein:
updating the first and second neural networks includes: configuring a loss function by using the diffusion weighted magnetic resonance images, the first output images, the third output images, and the fourth output images; and calculating the first weights and the second weights to minimize the loss function,
the first weights include one or more neural network weights related to the first neural works, and
the second weights include one or more neural network weights related to the second neural networks.
10. The method of claim 9 , wherein
the loss function is defined by the following equations:
loss
=
loss
1
+
λ
loss
2
=
∑
b
l
NCC
(
S
in
(
b
)
;
S
^
out
(
b
)
)
+
λ
S
a
(
b
)
-
S
^
a
(
b
)
l
NCC
(
X
;
Y
)
=
-
log
(
NCC
(
X
;
Y
)
+
1
2
)
,
wherein loss denotes the loss function, loss 1 denotes a first loss function, loss 2 denotes a second loss function, S in (b) denotes the diffusion weighted magnetic resonance images, Ŝ out (b) denotes the fourth output images, S a (b) denotes the first output images, Ŝ a (b) denotes the third output images, λ denotes a weight, and NCC(X; Y) denotes a normalized cross correlation between X and Y.