Magnetic resonance parameter quantification method and device based on recurrent network, equipment and medium

Abstract

The application provides a magnetic resonance parameter quantification method and device based on a recurrent network, equipment and a medium, wherein the method comprises: acquiring a plurality of magnetic resonance weighted images of different contrasts; inputting the plurality of magnetic resonance weighted images into a constructed recurrent neural network, and obtaining a magnetic resonance parameter image after multiple cycles; wherein the recurrent neural network is constructed based on a physical model and a convolutional neural network; for each cycle of the recurrent neural network, the corresponding input is the plurality of magnetic resonance weighted images of different contrasts and the state parameters output by the last cycle of the recurrent neural network, and the corresponding output is the magnetic resonance parameter image and the state parameters calculated by the current cycle of the recurrent neural network; and the physical model is determined by the acquisition process of the plurality of magnetic resonance weighted images of different contrasts. The application solves the problem that the existing deep learning technology cannot perform high-quality parameter quantification when the acquisition parameters of the input magnetic resonance weighted image change greatly.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of magnetic resonance, and particularly relates to a magnetic resonance parameter quantification method and device based on a recurrent network, equipment and a medium. BACKGROUND

[0002] Magnetic resonance parameter quantification is a technique that calculates the physical parameters of tissue from multiple magnetic resonance weighted images of different contrasts. Magnetic resonance parameter images can reflect the spatial distribution of biophysical parameters in tissue, such as the transverse relaxation time (T2) and longitudinal relaxation time (T1) of tissue. Compared with traditional qualitative magnetic resonance weighted images, quantitative magnetic resonance parameters are more sensitive to physiological changes in tissue, and the obtained parameter values are comparable between multiple centers. Magnetic resonance parameter quantification has been widely used in the study of diseases such as multiple sclerosis, brain tumors, stroke, epilepsy, tissue iron overload, cartilage disease and cardiomyopathy.

[0003] Magnetic resonance parameter quantification requires the acquisition of multiple magnetic resonance weighted images of different contrasts. However, the acquisition parameters of the acquisition sequence of these images can be flexibly set. In different centers, different magnetic resonance scanners, and different research purposes, there are large differences in acquisition parameters. In recent years, many studies have explored the application of deep learning methods in the field of magnetic resonance parameter quantification. For example, fully connected neural networks, convolutional neural networks and other deep learning networks have achieved good results. However, these deep learning networks can usually only be applied to magnetic resonance weighted images acquired with specific acquisition parameters. When the acquisition parameters of the input images change significantly and are different from the acquisition parameters used in the training data, such as inputting magnetic resonance weighted images with different numbers, if the deep learning network is not retrained, their parameter quantification accuracy is low, or even they cannot calculate reasonable parameters. In addition, retraining of deep learning networks is usually very time-consuming, and it is also unrealistic to train all different acquisition parameters. SUMMARY

[0004] The purpose of the present application is to provide a magnetic resonance parameter quantification method and device based on a recurrent network, which solves the problem that existing deep learning technology cannot perform high-quality parameter quantification when the acquisition parameters of the input magnetic resonance weighted images change significantly, avoids the need to retrain the network, and at the same time can achieve good parameter quantification quality.

[0005] The present application is achieved by the following technical solutions:

[0006] A magnetic resonance parameter quantification method based on a recurrent network, comprising the following steps:

[0007] Acquire multiple magnetic resonance weighted images of different contrasts;

[0008] Multiple weighted magnetic resonance images are input into a pre-constructed recurrent neural network (RNN). After multiple iterations, magnetic resonance parameter images are obtained. The RNN is constructed based on a physical model and a convolutional neural network. For each iteration of the RNN, the corresponding inputs are multiple weighted magnetic resonance images with different contrasts and the state parameters output from the previous iteration of the RNN. The corresponding outputs are the magnetic resonance parameter images and state parameters calculated by the RNN in the current iteration. The physical model is determined by the acquisition process of multiple weighted magnetic resonance images with different contrasts.

[0009] Furthermore, the state parameters include a, b, and c;

[0010] A recurrent neural network is composed of gradient units, Hessian units, denoising units, and two scaling coefficients λ and ρ connected together.

[0011] For the i-th iteration of the recurrent neural network, the input to the gradient unit is the state parameter a output from the (i-1)-th iteration of the recurrent neural network. i-1 b i-1 c i-1 And multiple magnetic resonance weighted images with different contrasts acquired { q}, its output is the objective function f(a) with respect to the state parameter a in a i-1 gradient value at Where q = 1, ..., N q N q Given the number of input magnetic resonance weighted images with different contrasts, the objective function f(a) is:

[0012]

[0013] In the formula, S q The physical model is determined during the acquisition of the q-th input magnetic resonance weighted image. The input of the physical model is the state parameter a, and its output is the magnetic resonance weighted image calculated by applying the physical model.

[0014] The input to the Hessian unit is the state parameter a output of the (i-1)th iteration of the recurrent neural network. i-1 Its output is the objective function f(a) with respect to the state parameter a in a i-1 The Hessian matrix H at the location f (a i-1 );

[0015] The input to the denoising unit is the state parameter b output by the (i-1)th iteration of the recurrent neural network. i-1 Its output is the state parameter b i-1 State parameters D(b) after noise removali-1 )。

[0016] Further, the denoising unit is a convolutional neural network with a residual structure, which is composed of a plurality of convolutional layers connected in series, and a linear rectified activation function after each convolutional layer except the last one, and the output of the last convolutional layer is added to the input of the convolutional neural network to obtain the output of the convolutional neural network.

[0017] Further, the recurrent neural network calculates the state parameters a i , b i and c i of the i-th cycle output by using the following formula:

[0018]

[0019] b i = [p (a i + c i-1 ) + lD (b i-1 )] / (p + l);

[0020] c i = c i-1 + a i - b i ;

[0021] In the formula, H f -1 i-1 is the inverse matrix of the output Hessian matrix H i-1 of the Hessian unit, D (b i ) is the output result of the denoising unit, and i is an integer greater than or equal to 1.

[0022] Further, the recurrent neural network calculates the magnetic resonance parameter image x i of the i-th cycle output by using the following formula:

[0023] x i = a s .

[0024] Further, the proportionality coefficients l and p are learned by a weighted mean absolute error loss function, wherein the weighted mean absolute error loss function is:

[0025]

[0026] In the formula, N i is the number of cycles of the recurrent neural network, x i-1 is the magnetic resonance parameter image calculated and output by the i-th cycle of the recurrent neural network, D is the denoising unit, b gt is the state parameter output by the i-1-th cycle of the recurrent neural network, and xgt is a real parameter image.

[0027] Further, the multiple magnetic resonance weighted images of different contrasts are magnetic resonance weighted images acquired using different sequence parameters.

[0028] The application also provides a magnetic resonance parameter quantification device based on a recurrent network, comprising:

[0029] The acquisition module is configured to acquire the multiple magnetic resonance weighted images of different contrasts.

[0030] The input module is configured to input the multiple magnetic resonance weighted images into the constructed recurrent neural network, and obtain the magnetic resonance parameter image after multiple cycles.

[0031] The application also discloses an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any of the above methods when executing the computer program.

[0032] The application also discloses a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of any of the above methods.

[0033] Compared with the prior art, the application has the beneficial effects that: the recurrent neural network is obtained by combining the physical model and the convolutional neural network, compared with the existing deep learning network, the recurrent neural network can make full use of the known physical model relied on by the magnetic resonance signal acquisition to cope with the influence brought by the change of the acquisition parameters of the input magnetic resonance weighted image, so that the constructed recurrent neural network is more robust to the change of the acquisition parameters of the magnetic resonance weighted image, and good parameter quantification quality can be obtained under different acquisition parameters, and the convolutional neural network in the recurrent neural network can more effectively constrain the generated magnetic resonance parameter image, and further ensure high-quality magnetic resonance parameter quantification. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is a step flowchart of the magnetic resonance parameter quantification method based on the recurrent network of the application;

[0035] Figure 2 is a whole framework diagram of the magnetic resonance parameter quantification method based on the recurrent network of the application;

[0036] Figure 3 FIG. 1 is a structural schematic diagram of a recurrent neural network in a magnetic resonance parameter quantification method based on a recurrent network according to an embodiment of the present application;

[0037] Figure 4 FIG. 2 is a structural schematic diagram of a denoising unit in a recurrent neural network in a magnetic resonance parameter quantification method based on a recurrent network according to an embodiment of the present application;

[0038] Figure 5 FIG. 3 is a module schematic diagram of a magnetic resonance parameter quantification device based on a recurrent network according to an embodiment of the present application;

[0039] Figure 6 FIG. 4 is a structural schematic diagram of an electronic device according to an embodiment of the present application;

[0040] Figure 7 FIG. 5 is a structural schematic block diagram of a computer readable storage medium according to an embodiment of the present application. DETAILED DESCRIPTION

[0041] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations.

[0042] Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed application, but only represents selected embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative labor are within the scope of protection of the present application.

[0043] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. Meanwhile, in the description of the present application, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0044] It is to be noted that, in the present text, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by "comprises... a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the recited element.

[0045] In the description of the application, it should be noted that the terms "upper", "lower", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, or the orientation or positional relationship in which the product of the application is usually placed, only for the convenience of describing the application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the application.

[0046] Please refer to Figure 1 and Figure 2 , Figure 1 is a step flow chart of the magnetic resonance parameter quantification method based on the recurrent network of the application, Figure 2 is a whole framework diagram of the magnetic resonance parameter quantification method based on the recurrent network of the application. A magnetic resonance parameter quantification method based on a recurrent network comprises the following steps:

[0047] S1, acquiring a plurality of magnetic resonance weighted images of different contrasts;

[0048] S2, inputting the plurality of magnetic resonance weighted images into the constructed recurrent neural network, and after multiple cycles, obtaining a magnetic resonance parameter image; wherein the recurrent neural network is constructed based on a physical model and a convolutional neural network, for each cycle of the recurrent neural network, the corresponding input is a plurality of magnetic resonance weighted images of different contrasts and the state parameter output by the recurrent neural network in the last cycle, and the corresponding output is the magnetic resonance parameter image and the state parameter calculated by the recurrent neural network in this cycle; the physical model is determined by the acquisition process of the plurality of magnetic resonance weighted images of different contrasts.

[0049] In the step S1, the multiple magnetic resonance weighted images with different contrasts are magnetic resonance weighted images acquired by using different sequence parameters. The multiple magnetic resonance weighted images with different contrasts can be obtained by using a multi-echo gradient echo sequence to obtain multiple magnetic resonance weighted images with different echo times, thereby obtaining multiple magnetic resonance weighted images with different contrasts. The multi-echo gradient echo sequence has the characteristics of strong applicability, easy operation, low cost, fast acquisition speed, etc. It should be noted that the method of obtaining the magnetic resonance weighted image is not limited to the multi-echo gradient echo sequence in the embodiment, but can also be obtained by single-echo acquisition multiple times, etc. The obtained magnetic resonance weighted image can have any size, for example, 64x128 pixels, 128x128 pixels, 256x256 pixels, 512x512 pixels, etc. The obtained magnetic resonance weighted image can be a magnetic resonance weighted image of a human tissue organ such as the brain, liver, etc. In the embodiment, the obtained magnetic resonance weighted image is a magnetic resonance weighted image of the brain.

[0050] In the step S2, the recurrent neural network is constructed in advance. The recurrent neural network is constructed based on a physical model and a convolutional neural network. The physical model is a specific signal change physical model followed by the signal change acquired by the acquisition sequence in the magnetic resonance signal acquisition process. Therefore, the form of the physical model is determined by the acquisition sequence applied in the acquisition process of the input multiple magnetic resonance weighted images with different contrasts. It is derived from the classical Bloch equation of magnetic resonance and can be expressed by a physical formula. The acquisition sequence refers to an existing acquisition sequence dedicated to parameter quantization. For example, when a multi-echo spin echo sequence is used to acquire a magnetic resonance weighted image to quantify the R2 parameter, the physical model corresponding to the signal change of a specific pixel position can be expressed as S=0xp(-TR2), where S is the acquired signal value, S0 is the signal value at the quantized echo time of 0, and TE is the echo time. The type of parameter quantization is not limited to the R2 parameter, but can also be a relaxation parameter such as T1. The corresponding physical model in other parameter quantization processes can also be easily derived according to the Bloch equation and the form of the acquisition sequence.

[0051] The input of the first cycle of the recurrent neural network is the multiple magnetic resonance weighted images of different contrasts, the state parameters a0 of the input are all set to 1, the state parameters b0 of the input are all set to 1, and the state parameters c0 of the input are all set to 0, and the output of the first cycle of the recurrent neural network is the magnetic resonance parameter image calculated in this cycle and the state parameters. Starting from the second cycle of the recurrent neural network, the input of each cycle of the recurrent neural network is the multiple magnetic resonance weighted images of different contrasts and the state parameters output by the last cycle of the recurrent neural network, and the output of each cycle of the recurrent neural network is the magnetic resonance parameter image calculated in this cycle and the state parameters. Therefore, the multiple magnetic resonance weighted images obtained are input into the recurrent neural network, and the magnetic resonance parameter image is obtained after a preset number of cycles. Specifically, the preset number of cycles can be set according to actual conditions. In this embodiment, the number of cycles is set to 15, and the magnetic resonance parameter image output in the last cycle is taken as the final result after 15 cycles of the recurrent neural network, so that the final output magnetic resonance parameter image is obtained. The magnetic resonance parameter image is an image composed of quantified magnetic resonance parameters, and the value of each pixel of the parameter image is the magnetic resonance parameter at the corresponding pixel position.

[0052] Further, the state parameters output by the recurrent neural network include three, a, b and c;

[0053] Please refer to Figure 3 , Figure 3 The structure diagram of the recurrent neural network in the magnetic resonance parameter quantification method based on the recurrent network. The specific process of constructing the recurrent neural network includes:

[0054] The recurrent neural network is composed of a gradient unit G, a Hessian unit H, a denoising unit D and two scale factors λ and ρ.

[0055] For the i-th cycle of the recurrent neural network, the input of the gradient unit G is the state parameters a i-1 , b i-1 , c i-1 output by the (i-1)-th cycle of the recurrent neural network and the multiple magnetic resonance weighted images of different contrasts {y q}, and the output of the gradient unit G is the gradient value of the objective function f(a) with respect to the state parameter a at a i-1 . Wherein, q=1,…,N q , N q is the number of input magnetic resonance weighted images of different contrasts, and the objective function f(a) is:

[0056]

[0057] In the formula, S q is a physical model followed by the corresponding signal change in the qth input magnetic resonance weighted image acquisition process, and the input of the physical model is the state parameter a, and the output of the physical model is the magnetic resonance weighted image calculated by applying the physical model; specifically, in this embodiment, taking the transverse relaxation rate R2 parameter of the brain as an example, Where, TE q is the qth echo time, S0 and R2 are the expected quantified magnetic resonance parameters, a=(S0, R2), S0 represents the signal intensity of the magnetic resonance weighted image when the echo time is 0 ms (TE0=0), and R2 represents the tissue transverse relaxation rate parameter, y q is the magnetic resonance weighted image acquired at the qth echo time, N q is the number of input magnetic resonance weighted images acquired at different echo times. It should be noted that the quantified parameter is not limited to the transverse relaxation rate R2 in this embodiment, but can also be the longitudinal relaxation rate R1 and other relaxation parameters. For the quantification of other parameters, the corresponding existing special parameter quantification acquisition sequence is required.

[0058] The gradient value is calculated by the following formula

[0059]

[0060] In the formula, a p is the pth magnetic resonance parameter, N p is the number of expected quantified magnetic resonance parameters. In this embodiment, taking the transverse relaxation rate R2 parameter quantification as an example, N p = 2, a 1 =S0, a 2 =R2, S q , respectively, the derivative of a p can be obtained:

[0061]

[0062]

[0063] The input of the Hessian unit H is the state parameter a i-1 output of the recurrent neural network at the i-1th cycle, and the output of the Hessian unit H is the Hessian matrix H i-1 of the objective function f(a) with respect to the state parameter a at a f (a i-1 ).

[0064] Specifically, in this embodiment, taking the transverse relaxation rate R2 parameter of the brain as an example, the Hessian matrix can be estimated according to the Gauss-Newton method, and the calculation formula is as follows:

[0065]

[0066] The input of the denoising unit D is the state parameter b output by the i-1th cycle of the recurrent neural network i-1 , and the output is the state parameter b i-1 . The state parameter b i-1 after removing noise D(b i-1 ) is the output of the convolutional neural network.

[0067] Please refer to Figure 4 , Figure 4 for the structure diagram of the denoising unit in the recurrent neural network of the magnetic resonance parameter quantification method based on the recurrent network of the application. Further, the denoising unit D is a convolutional neural network with a residual structure, which is composed of a plurality of convolutional layers in series, and a linear rectifier activation function is added after each convolutional layer except the last convolutional layer. The output of the last convolutional layer is added to the input of the convolutional neural network to obtain the output of the convolutional neural network.

[0068] Specifically, in this embodiment, the denoising unit D is composed of seven convolutional layers in series, and a linear rectifier activation function ReLU is added after the first six convolutional layers, and the size of the convolution kernel is 3. The feature layer number of the output of the first six convolutional layers is 64, and the feature layer number of the output of the last convolutional layer is N p , which corresponds to different magnetic resonance parameters respectively. At the same time, the convolutional neural network has a residual structure, that is, the output of the last convolutional layer of the convolutional neural network is added to the input of the convolutional neural network to obtain the output of the convolutional neural network.

[0069] Further, the recurrent neural network calculates the state parameters a i , b i and c i of the i-th cycle output by the following formula:

[0070]

[0071] b i = [p(a i +c i-1 )+lD(b i-1 )] / (p+l);

[0072] c i =c i-1 +a i -b i ;

[0073] wherein, is the output Hessian matrix H of the Hessian unit f is the inverse matrix of D(b i-1 ), D(b i-1 ) is the output result of the denoising unit, and i is an integer greater than or equal to 1.

[0074] Further, the recurrent neural network calculates the magnetic resonance parameter image x i of the i-th cycle according to the following formula:

[0075] x i = a i .

[0076] After the recurrent neural network is constructed, a certain number of samples are required for training to learn the learnable parameters λ and ρ and the weight coefficients in the convolutional neural network in the denoising unit, so that the recurrent neural network can be used for magnetic resonance parameter quantization. Further, the learning is performed by using a weighted mean absolute error loss function, wherein the weighted mean absolute error loss function is:

[0077]

[0078] wherein, N s is the number of cycles of the recurrent neural network, x i is the magnetic resonance parameter image output by the i-th cycle of the recurrent neural network, D is the denoising unit, b i-1 is the state parameter output by the i-1-th cycle of the recurrent neural network, and x gt is the real parameter image.

[0079] The above-mentioned recurrent neural network combines the physical model followed in the magnetic resonance signal acquisition process and the convolutional neural network, so that the constructed recurrent neural network can make full use of the acquisition parameter information used in the magnetic resonance weighted image acquisition process, thereby enabling it to be applied to magnetic resonance weighted images acquired by different acquisition parameters, and having strong robustness. In addition, the addition of the learnable convolutional neural network further improves the accuracy of parameter quantization.

[0080] Please refer to Figure 5 , Figure 5 for the module schematic diagram of the magnetic resonance parameter quantization device based on the recurrent network of the present application. The present application also provides a magnetic resonance parameter quantization device based on a recurrent network, comprising:

[0081] The acquisition module 1 is used for acquiring a plurality of magnetic resonance weighted images of different contrasts;

[0082] The input module 2 is configured to input a plurality of magnetic resonance weighted images into the constructed recurrent neural network, and after multiple cycles, a magnetic resonance parameter image is obtained; wherein the recurrent neural network is constructed based on a physical model and a convolutional neural network; for each cycle of the recurrent neural network, the corresponding input is a plurality of magnetic resonance weighted images of different contrasts and a state parameter output by the recurrent neural network in the last cycle, and the corresponding output is a magnetic resonance parameter image calculated by the recurrent neural network in the current cycle and a state parameter; and the physical model is determined by the acquisition process of the plurality of magnetic resonance weighted images of different contrasts.

[0083] The plurality of magnetic resonance weighted images of different contrasts are magnetic resonance weighted images acquired using different sequence parameters. The acquisition module 1 can obtain a plurality of magnetic resonance weighted images with different echo times using a multi-echo gradient echo sequence, thereby obtaining a plurality of magnetic resonance weighted images of different contrasts. The multi-echo gradient echo sequence used has the characteristics of strong applicability, easy operation, low cost, fast acquisition speed, etc. It should be noted that the method of obtaining the magnetic resonance weighted image is not limited to the multi-echo gradient echo sequence in the embodiment, but can also be obtained by single-echo acquisition multiple times, etc. The obtained magnetic resonance weighted image can have any size, for example, 64x128 pixels, 128x128 pixels, 256x256 pixels, 512x512 pixels, etc. The obtained magnetic resonance weighted image can be a magnetic resonance weighted image of a human tissue organ such as the brain, liver, etc. In the embodiment, the obtained magnetic resonance weighted image is a magnetic resonance weighted image of the brain.

[0084] In the above input module 2, the recurrent neural network is constructed in advance, and the recurrent neural network is constructed by combining a physical model and a convolutional neural network. The physical model is a specific signal change physical model followed by the signal change obtained by the acquisition sequence in the magnetic resonance signal acquisition process, and therefore the form of the physical model is determined by the acquisition sequence applied in the acquisition process of the input plurality of magnetic resonance weighted images of different contrasts. It is derived from the classical Bloch equation of magnetic resonance and can be expressed by a physical formula. The acquisition sequence refers to an existing acquisition sequence dedicated to parameter quantization. For example, when a multi-echo spin echo sequence is used to acquire a magnetic resonance weighted image to quantify the R2 parameter, the physical model corresponding to the signal change of a specific pixel position can be expressed as S = S0xp(-TE R2), wherein S is the signal value obtained by acquisition, S0 is the signal value obtained by quantization when the echo time is 0, and TE is the echo time. The type of parameter quantization is not limited to the R2 parameter, but can also be a relaxation parameter such as T1. The corresponding physical model in other parameter quantization processes can also be easily derived according to the Bloch equation and the form of the acquisition sequence.

[0085] The input of the first cycle of the recurrent neural network is the multiple magnetic resonance weighted images of different contrasts, the state parameters a0 of the input are all set to 1, the state parameters b0 of the input are all set to 1, and the state parameters c0 of the input are all set to 0, and the output of the first cycle of the recurrent neural network is the magnetic resonance parameter image calculated in this cycle and the state parameters. Starting from the second cycle of the recurrent neural network, the input of each cycle of the recurrent neural network is the multiple magnetic resonance weighted images of different contrasts and the state parameters output by the last cycle of the recurrent neural network, and the output of each cycle of the recurrent neural network is the magnetic resonance parameter image calculated in this cycle and the state parameters. Therefore, the multiple magnetic resonance weighted images obtained are input into the recurrent neural network, and the magnetic resonance parameter image is obtained after a preset number of cycles. Specifically, the preset number of cycles can be set according to actual conditions. In this embodiment, the number of cycles is set to 15, and the magnetic resonance parameter image output in the last cycle is taken as the final result after 15 cycles of the recurrent neural network, so that the final output magnetic resonance parameter image is obtained. The magnetic resonance parameter image is an image composed of quantified magnetic resonance parameters, and the value of each pixel of the parameter image is the magnetic resonance parameter at the corresponding pixel position.

[0086] Further, in the input module 2, the state parameters output by the recurrent neural network include three, which are a, b and c;

[0087] Please refer to Figure 3 , Figure 3 The structure diagram of the recurrent neural network in the magnetic resonance parameter quantification method based on the recurrent network.

[0088] The recurrent neural network is composed of a gradient unit G, a Hessian unit H, a denoising unit D and two scale factors λ and ρ.

[0089] For the i-th cycle of the recurrent neural network, the input of the gradient unit G is the state parameters a i-1 , b i-1 , c i-1 output by the (i-1)-th cycle of the recurrent neural network and the multiple magnetic resonance weighted images of different contrasts {y q}, and the output of the gradient unit G is the gradient value of the objective function f(a) with respect to the state parameter a at a i-1 . Wherein, q=1,…,N q , N q is the number of input magnetic resonance weighted images of different contrasts, and the objective function f(a) is:

[0090]

[0091] In the formula, S q is a physical model followed by the corresponding signal change in the qth input magnetic resonance weighted image acquisition process, and the input of the physical model is the state parameter a, and the output is the magnetic resonance parameter image calculated by applying the physical model; specifically, in this embodiment, taking the transverse relaxation rate R2 parameter of the brain as an example, wherein TE q is the qth echo time, S0 and R2 are the magnetic resonance parameters to be quantified, a=(S0, R2), S0 represents the signal intensity of the magnetic resonance weighted image when the echo time is 0 ms (TE0=0), and R2 represents the tissue transverse relaxation rate parameter, y q is the magnetic resonance weighted image acquired at the qth echo time, N q is the number of input magnetic resonance weighted images acquired at different echo times. It should be noted that the quantified parameter is not limited to the transverse relaxation rate R2 in this embodiment, but can also be the longitudinal relaxation rate R1 and other relaxation parameters. For the quantification of other parameters, the corresponding existing special acquisition sequence is required.

[0092] The gradient value is calculated by the following formula

[0093]

[0094] In the formula, a p is the pth magnetic resonance parameter, N p is the number of magnetic resonance parameters to be quantified. In this embodiment, taking the transverse relaxation rate R2 parameter quantification as an example, N p = 2, a 1 =S0, a 2 =R2, S q , respectively, the derivative of a p can be obtained:

[0095]

[0096]

[0097] The input of the Hessian unit H is the state parameter a i-1 output of the recurrent neural network at the i-1th time, and the output of the Hessian unit H is the Hessian matrix H i-1 of the objective function f(a) with respect to the state parameter a at a f (a i-1 ).

[0098] Specifically, in this embodiment, taking the transverse relaxation rate R2 parameter of the brain as an example, the Hessian matrix can be estimated according to the Gauss-Newton method, and the calculation formula is as follows:

[0099]

[0100] The input of the denoising unit D is the state parameter b of the i-1th cycle output of the recurrent neural network i-1 , and the output of the denoising unit is the state parameter b i-1 after removing noise D(b i-1 ), D(b i-1 ) is the output of the convolutional neural network.

[0101] Please refer to Figure 4 , Figure 4 for the structure diagram of the denoising unit in the recurrent neural network of the magnetic resonance parameter quantification method based on the recurrent network of the application. Further, the denoising unit D is a convolutional neural network with a residual structure, which is composed of a plurality of convolutional layers connected in series, and a linear rectifier activation function is added after each convolutional layer except the last convolutional layer. The output of the last convolutional layer is added to the input of the convolutional neural network to obtain the output of the convolutional neural network.

[0102] Specifically, in this embodiment, the denoising unit D is composed of seven convolutional layers connected in series, and a linear rectifier activation function ReLU is added after the first six convolutional layers, and the size of the convolution kernel is 3. The feature layer number of the output of the first six convolutional layers is 64, and the feature layer number of the output of the last convolutional layer is N p , which corresponds to different magnetic resonance parameters respectively. At the same time, the convolutional neural network has a residual structure, that is, the output of the last convolutional layer of the convolutional neural network is added to the input of the convolutional neural network to obtain the output of the convolutional neural network.

[0103] Further, the recurrent neural network calculates the state parameters a i , b i and c i of the i-th cycle output according to the following formula:

[0104]

[0105] b i = [p(a i +c i-1 )+lD(b i-1 )] / (p+l);

[0106] c i =c i-1 +a i -b i ;

[0107] wherein, is the output Hessian matrix H of the Hessian unit f is the inverse matrix of D(b i-1 ), D(b i-1 ) is the output result of the denoising unit, and i is an integer greater than or equal to 1.

[0108] Further, the recurrent neural network calculates the magnetic resonance parameter image x i in the i-th cycle according to the following formula:

[0109] x i = i .

[0110] After the recurrent neural network is constructed, a certain number of samples are required for training to learn the learnable parameters λ and ρ and the weight coefficients in the convolutional neural network in the denoising unit, so that the recurrent neural network can be used for magnetic resonance parameter quantification. Further, the learning is performed by a weighted mean absolute error loss function, wherein the weighted mean absolute error loss function is:

[0111]

[0112] wherein, N s is the number of cycles of the recurrent neural network, x i is the magnetic resonance parameter image calculated and output by the recurrent neural network in the i-th cycle, D is the denoising unit, b i-1 is the state parameter output by the recurrent neural network in the i-1-th cycle, and x gt is the real parameter image.

[0113] Please refer to Figure 6 , Figure 6 is a structural schematic block diagram of an embodiment of the electronic device. An embodiment of the present application further provides an electronic device 1001, which comprises a memory 1003 and a processor 1002, the memory 1003 stores a computer program 1004, and the processor 1002 implements the steps of any one of the above-mentioned magnetic resonance parameter quantification methods based on a recurrent network when executing the computer program 1004, which comprises: S1, acquiring a plurality of magnetic resonance weighted images of different contrasts; S2, inputting the plurality of magnetic resonance weighted images into the constructed recurrent neural network, and obtaining a magnetic resonance parameter image after multiple cycles; wherein the recurrent neural network is constructed based on a physical model and a convolutional neural network, for each cycle of the recurrent neural network, the corresponding input is a plurality of magnetic resonance weighted images of different contrasts and a state parameter output by the recurrent neural network in the last cycle, and the corresponding output is a magnetic resonance parameter image calculated by the recurrent neural network in the current cycle and a state parameter; the physical model is determined by the acquisition process of the plurality of magnetic resonance weighted images of different contrasts

[0114] Please refer to Figure 7 , Figure 7 is a structural schematic block diagram of an embodiment of the computer readable storage medium of the present application. The embodiment of the present application also provides a computer readable storage medium 2001, which stores a computer program 1004, and the computer program 1004 is executed by a processor 1002 to realize the steps of any one of the above-mentioned magnetic resonance parameter quantification methods based on the recurrent network, including: S1, acquiring a plurality of magnetic resonance weighted images of different contrasts; S2, inputting the plurality of magnetic resonance weighted images into the constructed recurrent neural network, and after multiple cycles, obtaining a magnetic resonance parameter image; wherein the recurrent neural network is constructed based on a physical model and a convolutional neural network, and for each cycle of the recurrent neural network, the corresponding input is a plurality of magnetic resonance weighted images of different contrasts and a state parameter output by the recurrent neural network in the last cycle, and the corresponding output is a magnetic resonance parameter image and a state parameter calculated by the recurrent neural network in the current cycle; and the physical model is determined by the acquisition process of the plurality of magnetic resonance weighted images of different contrasts.

[0115] Compared with the prior art, the beneficial effects of the present application are: by combining the physical model with the convolutional neural network to obtain the recurrent neural network, compared with the existing deep learning network, the recurrent neural network can make full use of the known physical model relied on by the magnetic resonance signal acquisition to cope with the influence brought by the acquisition parameter change of the input magnetic resonance weighted image, so that the constructed recurrent neural network is more robust to the change of the acquisition parameter of the magnetic resonance weighted image, and good parameter quantification quality can be obtained under different acquisition parameters, and at the same time, the convolutional neural network in the recurrent neural network can more effectively constrain the generated magnetic resonance parameter image, further ensuring high-quality magnetic resonance parameter quantification.

[0116] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiment methods can be included. Any reference to memory, storage, database, or other medium provided by the present application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0117] The present application is not limited to the above-described embodiments, and various modifications or changes can be made to the present application without departing from the spirit and scope of the present application, and the present application is intended to include such modifications and changes within the scope of the claims and equivalent technology.

Claims

1. A magnetic resonance parameter quantization method based on a recurrent network, characterized in that, Includes the following steps: Acquire multiple magnetic resonance weighted images with different contrasts; Multiple weighted magnetic resonance images are input into a pre-constructed recurrent neural network (RNN). After multiple iterations, a magnetic resonance parameter image is obtained. The RNN is constructed based on a physical model and a convolutional neural network. For each iteration of the RNN, the input consists of multiple weighted magnetic resonance images of different contrasts and the state parameters output from the previous iteration. The output consists of the magnetic resonance parameter image and state parameters calculated by the RNN in the current iteration. The physical model is determined by the acquisition process of multiple weighted magnetic resonance images of different contrasts, and the state parameters include… , and ; The process of constructing the recurrent neural network includes: The recurrent neural network consists of gradient units, Hessian units, denoising units, and two scaling factors. and It is composed of combinations and connections; For the first recurrent neural network In the nth iteration, the input of the gradient unit is the nth iteration of the recurrent neural network. The status parameters output in the next loop , , and multiple magnetic resonance weighted images with different contrasts acquired. Its output is the objective function. Relative to state parameters exist gradient value at ,in, , The objective function is defined as the number of input magnetic resonance weighted images with different contrasts. for: ； In the formula, For the first The physical model is determined during the acquisition of individual input magnetic resonance weighted images. The inputs to the physical model are state parameters. Its output is a magnetic resonance weighted image calculated using a physical model; The input of the Hessian unit is the first unit of the recurrent neural network. The status parameters output in the next loop Its output is the objective function. Relative to state parameters exist Hessian matrix at the location ; The input to the denoising unit is the first recurrent neural network. The status parameters output in the next loop Its output is a statement of the state parameters. State parameters after noise removal .

2. The magnetic resonance parameter quantization method based on cyclic networks according to claim 1, characterized in that, The denoising unit is a convolutional neural network with a residual structure, which consists of multiple convolutional layers connected in series. Except for the last convolutional layer, each convolutional layer is followed by a linear rectified activation function. The output of the last convolutional layer is added to the input of the convolutional neural network to obtain the output of the convolutional neural network.

3. The magnetic resonance parameter quantization method based on cyclic networks according to claim 1, characterized in that, The recurrent neural network is calculated using the following formula to obtain its first... The status parameters output in the next loop , and : ； ； ； In the formula, The output Hessian matrix of the Hessian unit The inverse matrix, The output of the noise reduction unit. It is an integer greater than or equal to 1.

4. The magnetic resonance parameter quantization method based on cyclic networks according to claim 3, characterized in that, The recurrent neural network is calculated using the following formula to obtain its first... Magnetic resonance parameter images output by the next cycle : 。 5. The magnetic resonance parameter quantization method based on cyclic networks according to claim 1, characterized in that, The proportionality coefficient and Both are learned using a weighted average absolute error loss function, where the weighted average absolute error loss function is: ； In the formula, This represents the number of iterations in the recurrent neural network. For the recurrent neural network The magnetic resonance parameter image output by the next iteration. For noise reduction unit, For the recurrent neural network The status parameters output in the next loop. This is a true parameter image.

6. The magnetic resonance parameter quantization method based on cyclic networks according to claim 1, characterized in that, The multiple magnetic resonance weighted images with different contrasts are magnetic resonance weighted images acquired using different sequence parameters.

7. A magnetic resonance parameter quantization device based on a recurrent network, characterized in that, include: The acquisition module is used to acquire multiple magnetic resonance weighted images with different contrasts; An input module is used to input multiple weighted magnetic resonance images into a pre-constructed recurrent neural network (RNN). After multiple iterations, a magnetic resonance parameter image is obtained. The RNN is constructed based on a physical model and a convolutional neural network. For each iteration of the RNN, the corresponding inputs are multiple weighted magnetic resonance images of different contrasts and the state parameters output from the previous iteration of the RNN. The corresponding outputs are the magnetic resonance parameter image and state parameters calculated by the RNN in the current iteration. The physical model is determined by the acquisition process of multiple weighted magnetic resonance images of different contrasts, and the state parameters include… , and ; The process of constructing the recurrent neural network includes: the recurrent neural network consists of gradient units, Hessian units, denoising units, and two scaling factors. and It is composed of combinations and connections; For the first recurrent neural network In the nth iteration, the input of the gradient unit is the nth iteration of the recurrent neural network. The status parameters output in the next loop , , and multiple magnetic resonance weighted images with different contrasts acquired. Its output is the objective function. Relative to state parameters exist gradient value at ,in, , The objective function is defined as the number of input magnetic resonance weighted images with different contrasts. for: ； In the formula, For the first The physical model is determined during the acquisition of individual input magnetic resonance weighted images. The inputs to the physical model are state parameters. Its output is a magnetic resonance weighted image calculated using a physical model; The input of the Hessian unit is the first unit of the recurrent neural network. The status parameters output in the next loop Its output is the objective function. Relative to state parameters exist Hessian matrix at the location ; The input to the denoising unit is the first recurrent neural network. The status parameters output in the next loop Its output is a statement of the state parameters. State parameters after noise removal .

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-6.

Images