Method for establishing 3D medical image segmentation model based on masked modeling and application thereof
View Patent ↗Disclosed is a method for establishing a 3D medical image segmentation model based on masked modeling and application thereof includes: establishing a semi-supervised learning network, wherein a student network includes an encoding module for extracting latent features and a segmentation decoder that predicts segmentation results, a teacher network includes an encoding module and a segmentation decoder that are structurally consistent with the student network; training the semi-supervised learning network, wherein during training, two random masking operations are performed on each image, and the image is input to the two networks respectively; optimizing and updating the weight of the student network, and transferring the updated weight to the teacher network, wherein the training loss function includes prototype representation loss, which is used to characterize the difference between the prototypes extracted and generated by the two networks; the student network may further include a reconstruction decoder and an auxiliary segmentation decoder.
1 . A method for establishing a 3D medical image segmentation model based on masked modeling, comprising:
(S1) establishing a semi-supervised learning network comprising a student network and a teacher network;
wherein the student network comprises: a first encoding module and a decoding module; the first encoding module comprises an encoder for extracting different size features of a 3D input image to obtain a latent feature; the decoding module comprises a first segmentation decoder, and the first segmentation decoder is disposed for performing feature extraction and up-sampling on the latent feature to obtain segmentation results;
the teacher network comprising: a second encoding module structurally consistent with the first encoding module, and a second segmentation decoder structurally consistent with the first segmentation decoder;
(S2) using a 3D medical image segmentation dataset comprising labeled images and unlabeled images to train the semi-supervised learning network, wherein the training method is as follows: fixing weights of the teacher network, and performing two random masking operations on each of the images, and the image is input to the student network and teacher network respectively, optimizing and updating a weight of the student network according to a preset training loss function, and transferring the updated weight to the teacher network; the training loss function comprises a prototype representation loss L p1 , which is disposed to characterize a difference between features in a corresponding area of a segmentation target in the latent features V s and V t extracted by the student network and the teacher network;
(S3) extracting the first encoding module and connecting the first encoding module to a first decoder to form the 3D medical image segmentation model.
2 . The method for establishing the 3D medical image segmentation model based on masked modeling according to claim 1 , wherein for the student network or teacher network, a feature p fg in the corresponding area of the segmentation target in the latent feature is calculated as follows:
p
fg
=
1
C
∑
j
=
1
C
Up
(
V
j
)
·
P
j
wherein V represents the latent feature, P represents the segmentation result; C represents a number of channels of the latent feature, V j represents a j-th channel of V, P j represents a j-th channel of P; and UP( ) represents an up-sampling operation; and
an expression of the prototype representation loss L p1 is as follows:
L
p
1
=
1
N
+
M
∑
i
=
1
N
+
M
L
mse
(
p
fg
s
,
p
fg
t
)
wherein
p
fg
s
and
p
fg
t
represent characteristics of the corresponding area of the segmentation target in the latent features V s and V t respectively, L mse represents a root mean square error, N and M respectively represent a number of the labeled images and the unlabeled images in the 3D medical image segmentation dataset.
3 . A 3D medical image segmentation method, comprising:
a 3D medical image to be segmented being input into the 3D medical image segmentation model established by the method for establishing the 3D medical image segmentation model based on masked modeling according to claim 2 , and a segmentation result is obtained from an output of the 3D medical image segmentation model.
4 . A non-transitory computer-readable storage medium, comprising a stored computer program, wherein when the computer program is executed by a processor, a device where the computer-readable storage medium is located is controlled to execute the method for establishing the 3D medical image segmentation model based on masked modeling according to claim 2 .
5 . The method for establishing the 3D medical image segmentation model based on masked modeling according to claim 1 , wherein the training loss function further comprises: a latent feature loss L fea ; the latent feature loss L fea is disposed to characterize a difference between the latent features extracted by the student network and the teacher network, and is expressed as follows:
L
fea
=
1
N
+
M
∑
i
=
1
N
+
M
L
mse
(
V
i
s
,
V
i
t
)
wherein L mse represents a root mean square error, N and M respectively represent a number of the labeled images and the unlabeled images in the 3D medical image segmentation dataset;
V
i
s
and
V
i
t
respectively represent the latent features extracted by the student network and the teacher network after an i-th image X i in the 3D medical image segmentation dataset is input.
6 . The method for establishing the 3D medical image segmentation model based on masked modeling according to claim 5 , wherein in the student network, the decoding module further comprises K auxiliary segmentation decoders; the auxiliary segmentation decoder is disposed to extract and up-sample the latent features to obtain the segmentation results; the up-sampling methods of the K auxiliary segmentation decoders are different from each other, and are different from that of the first segmentation decoder; and
the training loss function further comprises: a segmentation consistency loss L mc ; a segmentation consistency constraint is disposed to characterize a difference between segmentation results of the K auxiliary segmentation decoders and the first segmentation decoder, and is expressed as follows:
L
mc
=
1
N
+
M
∑
i
=
1
N
+
M
∑
m
,
n
=
1
&
m
≠
n
L
mse
(
P
i
.
m
s
,
P
i
,
n
s
_
sharp
)
wherein K is a positive integer,
A
=
C
K
+
1
2
;
P
i
.
m
s
and
P
i
.
n
s
represent segmentation results predicted by an m-th segmentation decoder and an n-th segmentation decoder represents a result after sharpening respectively after the image X i is input,
P
i
,
n
s
_
sharp
represents a result after sharpening
P
i
.
n
s
;
the segmentation decoder is an auxiliary segmentation decoder or the first segmentation decoder.
7 . The method for establishing the 3D medical image segmentation model based on masked modeling according to claim 6 , wherein in the student network, the decoding module further comprises: a reconstruction decoder; wherein the reconstruction decoder is disposed to extract and up-sample the latent features to restore an original image information and obtain a reconstructed image;
the training loss function further comprises: a reconstruction loss L sup1 ; the reconstruction loss is disposed to characterize a difference between the reconstructed image reconstructed by the student network and the original image, and is expressed as follows:
{
L
sup
1
=
α
N
+
M
∑
i
=
1
N
+
M
L
rec
(
Q
i
s
,
X
i
)
L
rec
=
1
N
+
M
∑
l
=
1
N
+
M
L
mse
(
Q
i
s
,
X
i
)
wherein
Q
i
s
represents the reconstructed image reconstructed by the student network after the image X i is input, and α represents a balance parameter.
8 . The method for establishing the 3D medical image segmentation model based on masked modeling according to claim 7 , wherein in the first encoding module, F Hybridformer modules connected successively are also comprised following the encoder; in the second encoding module, F Hybridformer modules connected successively are also comprised following the encoder;
the Hybridformer module is disposed to calculate self-attention in a pixel space and a sample dimension; and
the latent feature extracted by the student network is a feature image extracted by the encoder in the first encoding module and processed by the F HybridFormer modules, and the latent feature extracted by the teacher network is a feature image extracted by the encoder in the second encoding module and processed by the F HybridFormer modules;
wherein F is a positive integer.
9 . A 3D medical image segmentation method, comprising:
a 3D medical image to be segmented being input into the 3D medical image segmentation model established by the method for establishing the 3D medical image segmentation model based on masked modeling according to claim 5 , and a segmentation result is obtained from an output of the 3D medical image segmentation model.
10 . A non-transitory computer-readable storage medium, comprising a stored computer program, wherein when the computer program is executed by a processor, a device where the computer-readable storage medium is located is controlled to execute the method for establishing the 3D medical image segmentation model based on masked modeling according to claim 5 .
11 . A 3D medical image segmentation method, comprising:
a 3D medical image to be segmented being input into the 3D medical image segmentation model established by the method for establishing the 3D medical image segmentation model based on masked modeling according to claim 6 , and a segmentation result is obtained from an output of the 3D medical image segmentation model.
12 . A 3D medical image segmentation method, comprising:
a 3D medical image to be segmented being input into the 3D medical image segmentation model established by the method for establishing the 3D medical image segmentation model based on masked modeling according to claim 7 , and a segmentation result is obtained from an output of the 3D medical image segmentation model.
13 . A 3D medical image segmentation method, comprising:
a 3D medical image to be segmented being input into the 3D medical image segmentation model established by the method for establishing the 3D medical image segmentation model based on masked modeling according to claim 8 , and a segmentation result is obtained from an output of the 3D medical image segmentation model.
14 . The method for establishing the 3D medical image segmentation model based on masked modeling according to claim 1 , wherein the training loss function further comprises: a segmentation loss L sup2 , which is disposed to characterize a difference between a segmentation result predicted by the first segmentation decoder and a gold standard, and is expressed as follows:
L
sup
2
=
1
N
∑
i
=
1
N
L
seg
(
P
i
s
,
Y
i
)
wherein N represents a number of the labeled images in the 3D medical image segmentation dataset, Y i represents a gold standard for a segmented image corresponding to the i-th image X i in the 3D medical image segmentation dataset,
P
i
s
represents the segmentation result predicted by the first segmentation decoder after the image X i is input; and L seg represents a sum of a DICE loss and a cross-entropy loss.
15 . A 3D medical image segmentation method, comprising:
a 3D medical image to be segmented being input into the 3D medical image segmentation model established by the method for establishing the 3D medical image segmentation model based on masked modeling according to claim 14 , and a segmentation result is obtained from an output of the 3D medical image segmentation model.
16 . The method for establishing the 3D medical image segmentation model based on masked modeling according to claim 1 , wherein the random masking operation comprises:
dividing the 3D medical images into non-overlapping cubes of equal size, randomly selecting a proportion of cubs, and setting pixels in the corresponding area to zero.
17 . A 3D medical image segmentation method, comprising:
a 3D medical image to be segmented being input into the 3D medical image segmentation model established by the method for establishing the 3D medical image segmentation model based on masked modeling according to claim 16 , and a segmentation result is obtained from an output of the 3D medical image segmentation model.
18 . A 3D medical image segmentation method, comprising:
a 3D medical image to be segmented being input into the 3D medical image segmentation model established by the method for establishing the 3D medical image segmentation model based on masked modeling according to claim 1 , and a segmentation result is obtained from an output of the 3D medical image segmentation model.
19 . A non-transitory computer-readable storage medium, comprising a stored computer program, wherein when the computer program is executed by a processor, a device where the computer-readable storage medium is located is controlled to execute the 3D medical image segmentation method according to claim 18 .
20 . A non-transitory computer-readable storage medium, comprising a stored computer program, wherein when the computer program is executed by a processor, a device where the computer-readable storage medium is located is controlled to execute the method for establishing the 3D medical image segmentation model based on masked modeling according to claim 1 .