Nuclear magnetic resonance shimming coil topology optimization method and device based on depth prior

Abstract

The invention discloses a nuclear magnetic resonance shimming coil topology optimization method and device based on depth prior, and relates to the field of nuclear magnetic resonance coil design, and the method comprises the steps: S1, determining a design domain, a design variable and a target detection region; s2, establishing an incidence relation between the magnetic field intensity of the target detection area and a design variable; s3, performing finite element grid discrete division on the design domain, determining boundary conditions and constraint conditions, and constructing a target function; and S4, inputting the random variable into a preset deep neural network model, using an objective function as a loss function, performing parameter iteration of the deep neural network model by using an optimization algorithm, and outputting an optimized shim coil configuration. The depth prior technology is introduced into shimming coil topological optimization, performance optimization of a coil structure can be effectively improved, sensitivity does not need to be analyzed, and therefore the magnetic field compensation accuracy and efficiency of a complex morphology detection area are improved; and a coil structure data set does not need to be constructed in advance for training, so that the applicability is improved.

Description

Technical Field

[0001] The present invention relates to the field of nuclear magnetic resonance coil design, and in particular to a method and device for topology optimization of nuclear magnetic resonance shim coils based on deep priori. Background Art

[0002] The shim coil is the core component of the nuclear magnetic resonance instrument. It is responsible for compensating the uniformity of the main magnetic field in the target detection area, thereby ensuring high precision and high resolution, and directly affecting the signal and image quality. Traditional coil optimization design methods mainly include the target field method and the stream function method. The target field method solves the current density distribution of the design surface through Fourier-Searle series expansion and Fourier transform; the stream function method uses the stream function as the design variable, discretizes the coil current-carrying surface through the finite element grid, calculates the magnetic field generated by each current element in the target area, and approximates the actual magnetic field distribution through linear superposition.

[0003] As the application of nuclear magnetic resonance technology expands, traditional methods have low efficiency and poor linearity when dealing with complex, irregular or micro-scale detection areas, making it difficult to meet the increasing complexity and miniaturization needs. Topology optimization has attracted much attention because it can design high-performance structures without relying on experience. Through different optimization models and design variable update criteria, topology optimization significantly improves design efficiency and accuracy. The researchers applied density-based topology optimization to coil design, optimized the distribution of conductive materials, improved the space utilization and efficiency of complex detection areas, and reduced post-processing errors.

[0004] In recent years, deep learning technology has developed rapidly in the field of topology optimization. Topology optimization methods based on deep learning have overcome the problems of slow calculation speed and difficulty in handling complex constraints of traditional methods. They have shown significant advantages in handling complex nonlinear problems and calculation speed, and have promoted the development of intelligent structural design. However, conventional deep learning relies on a large number of samples to construct data sets, and physical and design constraints are often relaxed during training, resulting in the lack of generalization ability of the model and difficulty in dealing with new boundary conditions. In addition, using pixel differences as loss functions may cause structural fractures or lack of practical significance in the design, making the predicted structural performance inferior to traditional methods. To this end, researchers combined a new neural network with topology optimization, converted the design variable update into a network parameter update, thereby optimizing the topology structure, achieving efficient optimization, and adapting to a variety of boundary conditions without pre-building a data set. The topology optimization of the shim coil requires comprehensive consideration of performance such as power consumption, inductance, and torque. However, existing topology optimization technologies have few optimization design methods for the shim coil configuration. Summary of the invention

[0005] In response to the above problems, the present invention proposes a method and device for topology optimization of nuclear magnetic resonance shim coils based on deep priors, adds prior knowledge as constraints, constructs the objective function with magnetic field linearity and resistance as auxiliary objective functions, and optimizes the topology of shim coils through learning of a deep neural network model, thereby improving the accuracy and efficiency of magnetic field compensation in complex morphology detection areas; and there is no need to pre-build a data set, which improves adaptability.

[0006] On the one hand, a topology optimization method for nuclear magnetic resonance shim coils based on deep priors, the specific steps are as follows:

[0007] S1, taking the current-carrying surface of the shim coil as the design domain and the density of the conductive material as the design variable of the shim coil configuration, determine the parameters of the target detection area;

[0008] S2, using the Biot-Savart law to establish the correlation between the magnetic field intensity of the target detection area and the design variables;

[0009] S3, discretize the design domain by finite element mesh, determine the boundary conditions and constraints, and construct the objective function that meets the boundary conditions and constraints based on the magnetic field strength of the target detection area;

[0010] S4, input the random variables into the pre-designed deep neural network model, take the objective function as the loss function, and use the optimization algorithm to iterate the parameters of the deep neural network model. During the iteration process, the output of the deep neural network model is filtered and projected as the input of the loss function; after the number of iterations is reached, the output of the deep neural network model is filtered and projected as the optimized uniform field coil configuration.

[0011] Preferably, in S3, the boundary conditions and constraint conditions are expressed as:

[0012]

[0013] Where V represents the node potential value V j The vector composed of; P represents the load vector corresponding to the potential boundary condition; K represents the conductive matrix; represents the material volume constraint; ρ j Represents the relative density of the finite element mesh node j. Each unit takes a continuous density value between 0 and 1, where 0 represents an empty unit and 1 represents a solid unit. Vol j represents the volume fraction of finite element mesh node j; represents the target volume fraction; v represents ρ j The vector formed; j = 1, 2,…, n, where n represents the number of finite element mesh nodes.

[0014] Preferably, in S3, the objective function is expressed as:

[0015] f=φ(V;ρ)+λR(V;ρ)

[0016] Among them, φ(V; ρ) represents the magnetic field linearity objective function; R(V; ρ) represents the auxiliary objective function; λ represents the weight coefficient of the auxiliary objective function; the magnetic field linearity objective function is the least squares form of the residual of the actual magnetic field intensity and the target magnetic field intensity in the target detection area; the auxiliary objective function is constructed based on the resistance of the shim coil.

[0017] Preferably, the actual magnetic field strength of the target detection area is expressed as:

[0018]

[0019] Where t represents the thickness of the current-carrying surface; μ0 represents the magnetic permeability of vacuum; σ(ρ) represents the conductivity of the shim coil material; l represents the circumference of the bottom of the cylinder; r0 represents the radius of the current-carrying surface; V j represents the potential value of the finite element node j; N j represents the shape function of the finite element node j; RC = (r0cos(l / r0)-x i ), RS=(r0sin(l / r0)-y i ); x i ,y i and z i represents the field point coordinate value of the i-th sampling point in the target detection area; z c represents the coordinate value of the current-carrying surface; Ω represents the design domain.

[0020] Preferably, the electrical conductivity of the shim coil material is constructed using a density method and is expressed as:

[0021] σ(ρ)=σ Air +(σ Copper -σ Air )(ρ(α)) p

[0022] Where ρ(α)∈[0,1] represents the density of the conductive material, α represents the coordinates of the conductive material; σ(ρ) where ρ represents the conductivity of the shim coil material with the conductive material density as a parameter; p represents the exponential penalty term; σ Copper Represents the electrical conductivity of copper wire material; σ Air Represents vacuum conductivity.

[0023] Preferably, the auxiliary objective function is constructed based on power consumption, inductance or torque.

[0024] Preferably, in S4, the deep neural network model includes a convolutional layer, an upsampling layer, an activation function and batch normalization.

[0025] Preferably, in S4, the filtering and projection operations are specifically as follows:

[0026] First, the output of the deep neural network model is filtered, expressed as:

[0027]

[0028] Where s represents the preliminary logical image of the output of the deep neural network model; represents the coil density after filtering; represents the Laplace operator; R filt Indicates the filter radius;

[0029] Afterwards, the hyperbolic tangent projection scheme is used to Projection is performed, expressed as:

[0030]

[0031] in, represents the clear topological configuration after projection, β represents the projection slope, and 0.5 represents the threshold.

[0032] Preferably, in S4, the optimization algorithm includes a BFGS optimization algorithm or an Adam algorithm.

[0033] On the other hand, a topology optimization device for nuclear magnetic resonance shim coils based on deep priors includes the following:

[0034] A design parameter determination module is used to determine the target detection area parameters by taking the current-carrying surface of the shim coil as the design domain and the density of the conductive material as the design variable of the shim coil configuration;

[0035] A correlation establishing module, used for establishing the correlation between the magnetic field intensity of the target detection area and the design variables by using the Biot-Savart law;

[0036] Finite element discretization module, used to perform finite element mesh discretization on the design domain, determine boundary conditions and constraints, and construct an objective function that satisfies boundary conditions and constraints based on the magnetic field strength of the target detection area;

[0037] The deep neural network model learning and output module is used to input random variables into a pre-designed deep neural network model, use the objective function as the loss function, use the optimization algorithm to iterate the parameters of the deep neural network model, and output the optimized shim coil configuration after reaching the number of iterations. During the iteration process, the output of the deep neural network model is filtered and projected as the input of the loss function.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] (1) The present invention adopts deep prior technology and takes the density of conductive materials as the design variable. It replaces the traditional conductive material density update method by updating the parameters of the deep neural network model, and directly outputs the required shim coil density distribution through the neural network, thereby effectively optimizing the conductive material density of the shim coil and improving the compensation performance of the magnetic field deviation in the complex detection area. In addition, this method does not need to construct a coil structure data set and use a large amount of data for training. It is suitable for the design of shim coils under various boundary conditions and has higher applicability and efficiency.

[0040] (2) The present invention automatically solves the sensitivity through the automatic differentiation technology of deep neural networks, eliminating the need for manual sensitivity analysis and significantly accelerating the research process of topology optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The present invention is further described in detail below in conjunction with the accompanying drawings;

[0042] Figure 1 It is a flow chart of a method for topology optimization of nuclear magnetic resonance shim coils based on deep priors according to an embodiment of the present invention;

[0043] Figure 2 It is a schematic flow chart of a method for topology optimization of nuclear magnetic resonance shim coils based on deep priors according to an embodiment of the present invention;

[0044] Figure 3 It is a typical optimization method based on sensitivity analysis and iterative solution optimization algorithm flow chart;

[0045] Figure 4 A schematic diagram of a cylindrical current-carrying surface and its development surface of a shim coil in a topology optimization method for a nuclear magnetic resonance shim coil based on deep priori according to an embodiment of the present invention;

[0046] Figure 5 A schematic diagram of a deep neural network architecture of a nuclear magnetic resonance shim coil topology optimization method based on deep priors according to an embodiment of the present invention;

[0047] Figure 6 A schematic diagram of coil distribution in a design domain of a topology optimization method for nuclear magnetic resonance shim coils based on deep priors according to an embodiment of the present invention;

[0048] Figure 7 It is a structural block diagram of a nuclear magnetic resonance shim coil topology optimization device based on deep prior according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The present invention is further described below through specific implementation modes.

[0050] The function of the shim coil is to generate a specific magnetic field distribution in the target detection area of ​​the nuclear magnetic resonance instrument, compensate for the non-uniformity of the magnetic field distribution in the area, and construct a uniform magnetic field to improve the resolution of the detection signal. The so-called topological optimization is to seek the optimal material distribution problem in a given design domain according to certain boundary conditions. In this embodiment, the establishment of the optimization problem specifically includes the following contents: the model is discretely meshed by the finite element method, divided into multiple quadrilateral meshes, and the topological distribution of the conductive material in the finite element model of the nuclear magnetic resonance instrument detection area is used as the design domain, and the other parts of the finite element model of the target area are used as non-design areas. Set the potential boundary conditions and the insulation boundary conditions, establish the objective function by minimizing the magnetic field linearity, use power consumption, resistance or inductance as auxiliary objective functions, and use material density as a constraint condition.

[0051] See also Figure 1 Shown and Figure 2 As shown, the specific steps of the topology optimization method of nuclear magnetic resonance shim coil based on deep prior in this embodiment are as follows:

[0052] S1, taking the current-carrying surface of the shim coil as the design domain and the density of the conductive material as the design variable of the shim coil configuration, determine the parameters of the target detection area.

[0053] See also Figure 3 As shown in FIG. 1 , a typical iterative optimization algorithm flow chart of an optimization method based on sensitivity analysis is shown. This method requires manual sensitivity analysis and has problems such as slow calculation speed and difficulty in handling complex constraints.

[0054] See also Figure 4 As shown in the figure, the cylindrical current-carrying surface of the shim coil (i.e., the design domain surface) and its unfolded surface schematic diagram, as well as the target detection area, are shown. The density of the conductive material is used as the design variable of the shim coil configuration. First, the boundary conditions are established, and the cylindrical current-carrying surface of the shim coil is used as the design domain Ω. In the design domain, solving the electric field problem requires combining the coil boundary Γ=Γ D ∪Γ N Boundary conditions on Γ D represents the first type (Dirichlet) boundary condition, that is, the electric potential boundary condition; Γ N represents the second type (Neumann) boundary condition, that is, the insulating boundary condition.

[0055] S2, using the Biot-Savart law to establish the correlation between the magnetic field intensity of the target detection area and the design variables.

[0056] In order to ensure that the conductive material density design variable eventually converges to 0 or 1, the density interpolation method is used to represent the material properties of conductivity, which can be expressed as:

[0057] σ(ρ)=σ Air+(σ Copper -σ Air )(ρ(α)) p

[0058] Where σ(ρ) represents the conductivity of the shim coil material, ρ in σ(ρ) represents the conductivity of the shim coil material with the conductive material density as a parameter, ρ(α)∈[0,1] is the conductive material density, α represents the conductive material coordinate, p is the exponential penalty term, and σ Copper is the electrical conductivity of the copper wire material, σ Air is the vacuum conductivity.

[0059] The current continuity equation is used to establish the relationship between the surface current density and the conductivity of the shim coil material. The x and y components of the current density on the surface of the shim coil cylinder are expressed in Cartesian coordinates as:

[0060]

[0061] Among them, J x (l,z) represents the component of the current density on the surface of the shim coil cylinder in the x-axis direction in the Cartesian coordinate system; J y (l,z) represents the component of the current density on the surface of the shim coil cylinder in the y-axis direction in the Cartesian coordinate system; z represents the direction of the main magnetic field; Vol represents the volume of the ROI (Region of Interest, i.e., the target detection area); t represents the thickness of the current-carrying surface; l represents the circumference of the bottom surface of the cylinder; and r0 represents the radius of the current-carrying surface.

[0062] According to the magnetic field distribution in the ROI area of ​​the nuclear magnetic resonance instrument system, a three-dimensional magnetic field model B is established. zobj , using the Biot–Savart law to establish a relationship between the magnetic field distribution, the corresponding current carrying surface current, and the material conductivity.

[0063] In order to obtain the target magnetic field distribution, the optimization target of the magnetic field linearity is the actual magnetic field intensity in the ROI area. With the target magnetic field strength B zobj The least squares form of the residual is as follows:

[0064]

[0065] Among them, Min means Minimize; it means the number of sampling points in the ROI area, represents the actual magnetic field strength at the i-th sampling point, Represents the target magnetic field strength at the i-th sampling point.

[0066] S3, discretize the design domain by finite element mesh, determine the boundary conditions and constraints, and construct an objective function that meets the boundary conditions and constraints based on the magnetic field strength of the target detection area.

[0067] The design area is discretized by finite element mesh units and divided into multiple quadrilateral meshes. The density of the discretized conductive material is expressed as:

[0068]

[0069] Where N(α) is the finite element node shape function N j (α)(j=1,…,n), ρ is the density of the conductive material at the finite element node ρ j (j=1,…,n), where n is the number of finite element mesh nodes.

[0070] The finite element method is used to transform the current continuity equation into a discrete form KV=P, where V is the potential value V of the finite element grid node. j (j=1,…,n); P is the load vector corresponding to the potential boundary condition, and K is the conductive matrix.

[0071] According to the Biot-Savart law, the z-direction component of the magnetic field intensity at the i-th point in the ROI area after grid discretization is It is expressed as:

[0072]

[0073] Where μ0 is the vacuum magnetic permeability, RC = (r0cos(l / r0)-x i ), RS=(r0sin(l / r0)-y i ), x i ,y i and z i is the field point coordinate of the i-th point, z c Indicates the coordinate value of the current-carrying surface.

[0074] Construct an objective function. Since it is difficult to ensure that the shim coil configuration converges to an ideal black-and-white clear state by using magnetic field linearity alone as the optimization objective, other objectives need to be introduced in the design process to assist in the optimization. In this embodiment, the optimized resistance is used as an auxiliary objective function. Under the unit voltage boundary condition, the shim coil resistance is expressed as:

[0075]

[0076] in, is the coil energy consumption, and its finite element discrete form is expressed as:

[0077]

[0078] Therefore, the optimization model for the topology optimization problem of the shim coil in this embodiment is expressed as:

[0079]

[0080] Among them, Min means Minimize; f represents the objective function, φ(V; ρ) is the magnetic field linearity objective function; R(V; ρ) is the optimization model of the resistor (i.e., the resistor-assisted objective function); λ is the weight coefficient of the resistor-assisted objective function; is the material volume constraint; ρ j is the relative density of finite element j, and each element takes a continuous density value between 0 and 1, where 0 represents an empty element and 1 represents a solid element; Vol j =∫ Ω N j (α)dΩ is the volume fraction of finite element j; is the target volume fraction.

[0081] It should be noted that, in the design of shim coils, the selection of the auxiliary objective function may include optimizing other parameters such as energy consumption, inductance and torque, which is not limited in this embodiment.

[0082] S4, input the random variables into the pre-designed deep neural network model, take the objective function as the loss function, and use the optimization algorithm to iterate the parameters of the deep neural network model. During the iteration process, the output of the deep neural network model is filtered and projected as the input of the loss function; after the number of iterations is reached, the output of the deep neural network model is filtered and projected as the optimized uniform field coil configuration.

[0083] Initialize a neural network f with random parameters θ θ , for a given random input m, the initial output matrix s, that is, the preliminary logical image, can be obtained through the forward calculation process of the deep neural network.

[0084] See also Figure 5 As shown, the deep neural network model architecture used in this embodiment includes a network structure accumulated by multiple network sub-modules consisting of convolutional layers, upsampling layers, activation functions and batch normalization.

[0085] The preliminary logical image s output by the deep neural network model is converted into a clear structural topology that satisfies the volume and design variable constraints through preprocessing The physical density is used as the subsequent objective function for iterative optimization.

[0086] Specifically, in order to transform the preliminary logical image s in the grid into a clear binary density and avoid the checkerboard pattern, the preliminary logical image is first filtered and then processed by projection so that the design variables satisfy the 0 or 1 constraint, and finally a specific structural topological configuration without intermediate materials is obtained. The calculation steps are as follows:

[0087] First, by solving the partial differential equation Filtering is performed; among them, is the Laplace operator, R filt is the filter radius.

[0088] Then, the filtered variables For projection, this embodiment uses a hyperbolic tangent projection scheme to obtain a clearer specific structural topological configuration. (or physical density distribution diagram), the calculation process is as follows:

[0089]

[0090] Among them, β controls the projection slope and the threshold is set to 0.5.

[0091] After obtaining the specific structural topology, finite element calculation is performed according to the corresponding material interpolation model. The objective function of the topology optimization problem is regarded as a loss function, and the topology optimization process is embedded in the loss function of deep learning to drive the back propagation process of deep learning. When the upper limit of the number of iterations is reached, the optimized shim coil configuration is output.

[0092] Specifically, the parameter iterative training method for the deep neural network model can select the BFGS (BroydenFletcher Goldfarb Shanno) optimization algorithm or the gradient descent method (such as the Adam algorithm).

[0093] The deep prior-based NMR shim coil topology optimization method of this embodiment utilizes deep prior technology, takes the density of conductive materials as the design variable, and replaces the traditional conductive material density update method by updating the deep neural network parameters. The required shim coil density distribution is directly output through the neural network, thereby effectively optimizing the conductive material density of the shim coil and improving the compensation performance of the magnetic field deviation in the complex detection area. In addition, this method does not require a large amount of data training, is suitable for the design of shim coils under various boundary conditions, and has higher applicability and efficiency.

[0094] The deep prior-based NMR shim coil topology optimization method of this embodiment automatically solves the sensitivity through the automatic differentiation technology of deep neural networks, eliminating the need for manual sensitivity analysis and significantly accelerating the research process of topology optimization.

[0095] In this embodiment, during the optimization process of the shim coil topology configuration, the constraints on the shim coil topology configuration include ROI area size, current carrying surface size and shim coil minimum radius information; copper wire material conductivity and air conductivity; potential boundary conditions and insulation boundary conditions; the deviation between the magnetic field actually generated by the coil and the target magnetic field, and adding resistance, power consumption, inductance or torque constraints to obtain a smoother line shape; optimization target and auxiliary optimization target; volume fraction of conductive material; and the number of finite element division networks.

[0096] See also Figure 6 As shown, this embodiment takes the design of a nuclear magnetic resonance shim coil on a cylindrical surface as an example, and uses a nuclear magnetic resonance shim coil topology optimization design method based on deep prior to perform topology optimization prediction on the shim coil configuration. It should be noted that this method can be extended to the optimization design of gradient coils, and can also be optimized for coil configurations under other geometric surfaces or physical boundary conditions. The specific settings are made according to needs, and this embodiment does not limit it.

[0097] See also Figure 7 As shown, the embodiment of the present invention also discloses a topology optimization device for nuclear magnetic resonance shim coils based on deep prior, comprising:

[0098] A design parameter determination module 701 is used to determine the target detection area parameters by taking the current-carrying surface of the shim coil as the design domain and the density of the conductive material as the design variable of the shim coil configuration;

[0099] The association establishing module 702 is used to establish the association relationship between the magnetic field intensity of the target detection area and the design variables using the Biot-Savart law;

[0100] The finite element discretization module 703 is used to perform finite element mesh discretization on the design domain, determine boundary conditions and constraint conditions, and construct an objective function that satisfies the boundary conditions and constraint conditions based on the magnetic field strength of the target detection area;

[0101] The deep neural network model learning and output module 704 is used to input random variables into a pre-designed deep neural network model, use the objective function as the loss function, and use an optimization algorithm to iterate the parameters of the deep neural network model. During the iteration process, the output of the deep neural network model is filtered and projected as the input of the loss function; after the number of iterations is reached, the output of the deep neural network model is filtered and projected as the optimized uniform field coil configuration.

[0102] The specific implementation of the apparatus for topology optimization of nuclear magnetic resonance shim coils based on deep prior is the same as the method for topology optimization of nuclear magnetic resonance shim coils based on deep prior, and will not be repeated in this embodiment.

[0103] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A method for topology optimization of nuclear magnetic resonance shim coils based on deep priors, characterized in that: The steps include: S1, taking the current-carrying surface of the shim coil as the design domain and the density of the conductive material as the design variable of the shim coil configuration, determine the parameters of the target detection area; S2, using the Biot-Savart law to establish the correlation between the magnetic field intensity of the target detection area and the design variables; S3, discretize the design domain by finite element mesh, determine the boundary conditions and constraints, and construct the objective function that meets the boundary conditions and constraints based on the magnetic field strength of the target detection area; S4, input the random variables into the pre-designed deep neural network model, take the objective function as the loss function, and use the optimization algorithm to iterate the parameters of the deep neural network model. During the iteration process, the output of the deep neural network model is filtered and projected as the input of the loss function; after the number of iterations is reached, the output of the deep neural network model is filtered and projected as the optimized uniform field coil configuration.

2. The method for topology optimization of nuclear magnetic resonance shim coils based on deep prior according to claim 1, characterized in that: In S3, the boundary conditions and constraints are expressed as: KV=P 0≤ρ≤1 Where V represents the node potential value V j The vector composed of; P represents the load vector corresponding to the potential boundary condition; K represents the conductive matrix; represents the material volume constraint; ρ j Represents the relative density of the finite element mesh node j. Each unit takes a continuous density value between 0 and 1, where 0 represents an empty unit and 1 represents a solid unit. Vol j represents the volume fraction of finite element mesh node j; represents the target volume fraction; ρ represents ρ j The vector formed; j = 1, 2,…, n, where n represents the number of finite element mesh nodes.

3. The method for topology optimization of nuclear magnetic resonance shim coils based on deep prior according to claim 2, characterized in that: In S3, the objective function is expressed as: f=φ(V;ρ)+λR(V;ρ) Among them, φ(V; ρ) represents the magnetic field linearity objective function; R(V; v) represents the auxiliary objective function; λ represents the weight coefficient of the auxiliary objective function; the magnetic field linearity objective function is the least squares form of the residual of the actual magnetic field intensity and the target magnetic field intensity in the target detection area; the auxiliary objective function is constructed based on the resistance of the shim coil.

4. The method for topology optimization of nuclear magnetic resonance shim coils based on deep prior according to claim 3, characterized in that: The actual magnetic field strength of the target detection area is expressed as: Where t represents the thickness of the current-carrying surface; μ0 represents the magnetic permeability of vacuum; σ(ρ) represents the conductivity of the shim coil material; l represents the circumference of the bottom of the cylinder; r0 represents the radius of the current-carrying surface; V j represents the potential value of the finite element node j; N j represents the shape function of the finite element node j; RC = (r0 cos(l / r0)-x i ), RS=(r0 sin(l / r0)-y i ); x i ,y i and z i represents the field point coordinate value of the i-th sampling point in the target detection area; z c represents the coordinate value of the current-carrying surface; Ω represents the design domain.

5. The method for topology optimization of nuclear magnetic resonance shim coils based on deep prior according to claim 4, characterized in that: The electrical conductivity of the shim coil material is constructed using the density method and is expressed as: σ(ρ)=σ Air +(s Copper -s Air (p(a)) p Where ρ(α)∈[0,1] represents the density of the conductive material, α represents the coordinates of the conductive material; σ(ρ) where ρ represents the conductivity of the shim coil material with the conductive material density as a parameter; p represents the exponential penalty term; σ Copper Represents the electrical conductivity of copper wire material; σ Air Represents vacuum conductivity.

6. The method for topology optimization of nuclear magnetic resonance shim coils based on deep prior according to claim 2, characterized in that: The auxiliary objective function is constructed based on power consumption, inductance or torque.

7. The method for topology optimization of nuclear magnetic resonance shim coils based on deep prior according to claim 1, characterized in that: In S4, the deep neural network model includes a convolutional layer, an upsampling layer, an activation function and batch normalization.

8. The method for topology optimization of nuclear magnetic resonance shim coils based on deep prior according to claim 1, characterized in that: In S4, the filtering and projection operations are as follows: First, the output of the deep neural network model is filtered, expressed as: Where s represents the preliminary logical image of the output of the deep neural network model; represents the coil density after filtering; represents the Laplace operator; R filt Indicates the filter radius; Afterwards, the hyperbolic tangent projection scheme is used to Projection is performed, expressed as: in, represents the clear topological configuration after projection, β represents the projection slope, and 0.5 represents the threshold.

9. The method for topology optimization of nuclear magnetic resonance shim coils based on deep prior according to claim 1, characterized in that: In S4, the optimization algorithm includes a BFGS optimization algorithm or an Adam algorithm.

10. A topology optimization device for nuclear magnetic resonance shim coils based on deep priors, comprising: A design parameter determination module is used to determine the target detection area parameters by taking the current-carrying surface of the shim coil as the design domain and the density of the conductive material as the design variable of the shim coil configuration; A correlation establishing module, used for establishing a correlation relationship between the magnetic field intensity of the target detection area and the design variables by using the Biot-Savart law; Finite element discretization module, used to perform finite element mesh discretization on the design domain, determine boundary conditions and constraints, and construct an objective function that satisfies boundary conditions and constraints based on the magnetic field strength of the target detection area; The deep neural network model learning and output module is used to input random variables into the pre-designed deep neural network model, use the objective function as the loss function, and use the optimization algorithm to iterate the parameters of the deep neural network model. During the iteration process, the output of the deep neural network model is filtered and projected as the input of the loss function; After reaching the number of iterations, the output of the deep neural network model is filtered and projected as the optimized shim coil configuration.

Images