Physical-embedded mr birdcage coil tuning and matching prediction method and apparatus

Abstract

This invention discloses a method and device for predicting tuning and matching of a birdcage coil in magnetic resonance imaging based on physical embedding, belonging to the field of magnetic resonance imaging technology. In the optimization stage of the differential evolution algorithm, this invention does not simply rely on mathematical S-parameter approximation, but also constructs the value range of the components of the L-Match matching network on the actual radio frequency test bench as the matching network sensitivity penalty term and out-of-bounds penalty function, thereby eliminating capacitor combinations that meet the S-parameters but cannot be physically welded in advance. This invention converts the capacitance value into the capacitive reactance value, and the feature transformation mechanism of first performing an inverse proportional nonlinear mapping between capacitance and reactance and then performing model loss calculation completely eliminates the numerical nonlinear singularity of capacitance near the high-frequency resonance point, improves the stability of network convergence, and also significantly reduces the computational complexity of the model.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance imaging technology, and in particular to a method and device for predicting tuning and matching of a physically embedded magnetic resonance birdcage coil. Background Technology

[0002] Birdcage coils are known for their excellent radio frequency field (B1). + Uniformity is crucial for magnetic resonance imaging (MRI) systems. In actual engineering assembly, engineers need to weld tuning capacitors to the coil to achieve resonance at a specific frequency (e.g., 55.8 MHz) and introduce impedance matching networks (e.g., L-Match networks) at the feed port to achieve an impedance of fifty ohms. However, existing coil tuning methods face the following technical barriers: The load effect is complex: when the coil is loaded into the human body or water model, the dielectric effect and eddy current loss generated by polar molecules and ions will introduce "parasitic capacitance", causing the theoretically calculated no-load capacitance to completely fail; existing methods rely heavily on engineers to use a network analyzer (VNA) on a bench for extremely time-consuming trial and error debugging.

[0003] Purely data-driven neural networks lack common sense about physics: If existing neural networks are used to fit simulation data, the neural networks are very likely to give extreme solutions that violate physical topology in order to cater to the mathematical fifty-ohm matching (such as a 0.1pF matching capacitor that approaches an open circuit), which is completely unusable in radio frequency engineering. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned problems in the prior art by providing a method and device for predicting the tuning matching of a magnetic resonance birdcage coil based on physical embedding.

[0005] The above-mentioned objectives of the present invention are achieved by the following technical means: The method for predicting tuning matching based on physically embedded magnetic resonance birdcage coils includes the following steps: Step 1: Set the scanning range of the input parameters, perform simulation scanning on different combinations of input parameters of the birdcage coil model, obtain the S-parameter matrix and load state parameters under different combinations of input parameters, and construct the physical state vector of the birdcage coil model based on the coil physical dimensions and load state parameters in the input parameters.

[0006] Step 2: Obtain the Y-parameter matrix of the L-Match matching network under the combination of each input parameter and the corresponding load state parameters; A multi-objective fitness function is set up, and the multi-objective fitness function is minimized using the differential evolution algorithm to obtain the optimal capacitor combination that makes the Y parameter matrix satisfy the impedance matching condition. The optimal capacitor combination includes the optimal capacitance value of each capacitor in the capacitor combination. Then, the capacitance values ​​in the optimal capacitor combination are converted into capacitive reactance values, and the capacitive reactance values ​​corresponding to the optimal capacitor combination are used to construct the capacitive reactance vector of the optimal capacitor combination.

[0007] Step 3: Train the neural network model using the physical state vector as the model input. Calculate the loss function between the predicted capacitive reactance value in the predicted capacitive reactance vector of the capacitor combination and the capacitive reactance value in the capacitive reactance vector of the optimal capacitor combination output by the neural network model. After training, the trained neural network model is obtained.

[0008] Step 4: Construct a physical state vector based on the physical dimensions and load state parameters of the birdcage coil to be processed. Input the physical state vector into the trained neural network model to obtain the predicted capacitive reactance vector of the capacitor combination. Finally, convert the predicted capacitive reactance value in the predicted capacitive reactance vector of the capacitor combination into the predicted capacitance value.

[0009] As mentioned above, the input parameters also include the equivalent dielectric constant and equivalent conductivity of the water model load; the physical dimensions of the coil include the coil diameter. and coil length The load state parameters include the target resonant frequency. and equivalent load loss resistance Among them, the equivalent dielectric constant and equivalent conductivity are used to calculate the target resonant frequency and the equivalent load loss resistance; physical state vector .

[0010] The capacitor assembly described above includes a tuning capacitor. Parallel matching capacitors and series matching capacitor The capacitance values ​​in the optimal capacitor combination are converted into capacitive reactance values, and the tuning capacitive reactance values ​​are obtained respectively. Parallel matching capacitive reactance value and series matching capacitive reactance value The capacitive reactance vector of the capacitor combination .

[0011] As mentioned above, the capacitance value in the optimal capacitor combination is converted into capacitive reactance based on the following formula: ; In the formula, This is the capacitance value. This is the capacitance value.

[0012] As mentioned above, the multi-objective fitness function is: ; In the formula, For multi-objective fitness functions; To match the network sensitivity penalty, For out-of-bounds penalty functions, , as well as These are the return losses at the feed port. Matching network sensitivity penalty item and out-of-bounds penalty function The weight, For modulo operation; The matching network sensitivity penalty term is set to the absolute value of the difference between the tuning capacitor and the parallel matching capacitor; When the tuning capacitor Or parallel matching capacitor Or series matching capacitor When the value exceeds the corresponding range, =1, otherwise =0.

[0013] Feed port return loss weight Set to 1; Match network sensitivity penalty item weight Set to 0.1~0.5; weight of the out-of-bounds penalty function. Set to 10000 .

[0014] As mentioned above, the impedance matching condition is ≤-15dB.

[0015] As described above, the neural network model is a multilayer perceptron model, which includes an input layer, an output layer, and 2 to 4 fully connected hidden layers. Each fully connected hidden layer includes 64 to 256 neurons and uses the GELU activation function. In the initialization phase of the multilayer perceptron model, the weight matrix of the fully connected hidden layer is initialized using the Kaiming normal distribution, and the bias vector of each fully connected hidden layer is uniformly initialized to zero; the Adam optimizer is used.

[0016] As mentioned above, the loss function of the multilayer perceptron model adopts the MSE loss function.

[0017] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the physically embedded magnetic resonance birdcage coil tuning matching prediction method as described above.

[0018] Compared with the prior art, the present invention has the following advantages: (1) Unlike the conventional approach of hard-coding partial differential equations into the loss function in classical physical embedded neural networks (PINN), this invention converts the capacitance value into the capacitive reactance value. The feature transformation mechanism of first performing the inverse proportional nonlinear mapping between capacitance and reactance and then performing the model loss calculation completely eliminates the numerical nonlinear singularity of capacitance near the high-frequency resonance point, improves the stability of network convergence, and also greatly reduces the computational complexity of the model.

[0019] (2) In the optimization stage of the differential evolution algorithm, this invention does not rely solely on mathematical S-parameter approximation, but also constructs the value range of the components of the L-Match matching network on the actual RF test bench as a matching network sensitivity penalty term and an out-of-bounds penalty function, thereby eliminating capacitor combinations that meet the S-parameters but cannot be physically soldered in advance. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 A schematic diagram of the physical topology of the L-Match impedance matching network at the birdcage coil and the feed cut-out; Figure 3 This is a schematic diagram of the architecture and input / output nodes of the Physical Embedded Neural Network (PINN) of the present invention. Detailed Implementation

[0021] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to embodiments. The embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] Example 1: like Figure 1 As shown, the tuning matching prediction method based on physically embedded magnetic resonance birdcage coils includes the following steps: Step 1: Construct a birdcage coil model with a dielectric load using full-wave electromagnetic simulation software (such as HFSS). Simulate and scan the birdcage coil model using different combinations of input parameters in the software to obtain the S-parameter matrix and load state parameters under different combinations of input parameters. Construct the physical state vector of the birdcage coil model. This includes the following steps: The combination of input parameters for scanning includes the physical dimensions of the coil, the equivalent dielectric constant of the water model load, and the equivalent conductivity of the water model load. The physical dimensions of the coil include the coil diameter. and coil length .

[0023] Set the scan range of the input parameters, including the coil diameter. and coil length The scanning range (e.g., coil diameter) =80~600mm, coil length The equivalent dielectric constant and equivalent conductivity scanning ranges for the water model load (100~700mm) were determined using full-wave electromagnetic simulation software. The S-parameter matrices for different combinations of coil physical dimensions were derived (e.g., exporting an S16P file, a standardized data file used to store and exchange S-parameter matrices). Load state parameters, including the target resonant frequency, were then calculated. and equivalent load loss resistance .

[0024] Among them, the equivalent load loss resistance It is calculated in the following way: The dielectric loss power volume integral was performed on the three-dimensional load region using the field calculator of the full-wave electromagnetic simulation software. (1) In the formula, This refers to dielectric loss power; The integral volume of the three-dimensional load region; The equivalent conductivity of the water model load; Let be the electric field intensity vector within the three-dimensional load region. For modulo operation; after integration, extract the intrinsic current amplitude of the coil branch. Then, based on the principle of energy equivalence, the corresponding lumped load loss resistance is calculated. .

[0025] Step 2: Extract the L-Match matching network (e.g., based on the full-wave electromagnetic field simulation model) for each combination of input parameters and the corresponding load state parameters. Figure 2 The Y-parameter matrix (admittance matrix) is shown in the figure; a multi-objective fitness function is set, and the differential evolution (DE) algorithm is used to minimize the multi-objective fitness function, thereby iteratively optimizing the Y-parameter matrix to satisfy the impedance matching condition. The optimal capacitor combination is ≤-15dB, which includes the optimal capacitance value of each capacitor in the combination; then the capacitance values ​​in the optimal capacitor combination are converted into capacitive reactance values.

[0026] The capacitor assembly includes a tuning capacitor. Parallel matching capacitors and series matching capacitor The capacitance values ​​in the optimal capacitor combination are determined using physical equations. The tunable capacitive reactance is obtained by converting the capacitive reactance into the complex impedance space. Parallel matching capacitive reactance value and series matching capacitive reactance value ,in, This is the capacitance value. This is the capacitance value.

[0027] Differential evolution algorithm uses three basic operations—mutation, crossover, and selection—to iterate the population within a predefined component parameter space, driven by minimizing the multi-objective fitness function, and gradually approaches the global optimum.

[0028] In the extraction process of the differential evolution algorithm, a multi-objective fitness function that combines the actual engineering physical boundary is constructed for the Y parameter matrix. This is used to filter out usable solutions that satisfy the 50-ohm impedance matching requirement at the port. Multi-objective fitness function. Specifically: (2) In the formula, The return loss at the feed port (in decibels) is calculated based on the S-parameter matrix obtained from the current capacitor combination. For modulo operation, For the return loss of the feed port The model.

[0029] To match the network sensitivity penalty term, in this embodiment, since the target object is a volume transmitter coil, a network sensitivity penalty term is matched to ensure the circular polarization symmetry of the radio frequency field inside the coil. It is constructed directly based on the numerical difference between the tuning capacitor and the parallel matching capacitor, and the expression is: In mathematical solutions, although an extreme ratio of capacitor combinations (i.e., when the difference is extremely large) can achieve a 50-ohm impedance, it will result in an equivalent matching circuit with an extremely high node quality factor (Q value), making it extremely sensitive to manufacturing tolerances. The matching network sensitivity penalty term effectively eliminates fragile solutions that are extremely unstable in engineering by dynamically constraining the gradient of the value ratio between related RF components, thus ensuring the robustness of bench assembly.

[0030] The out-of-bounds penalty function is used when the tuning capacitor... Or parallel matching capacitor Or series matching capacitor The value range of physical components exceeds that available in actual engineering (in this embodiment, to conform to the general physical specifications of high-voltage non-magnetic ceramic trimmer capacitors that can be purchased for actual RF test benches, the tuning capacitor...). Parallel matching capacitors and series matching capacitor When the value range is uniformly set to 1 pF (picofarad) to 200 pF, =1, otherwise =0.

[0031] , as well as These are the return losses at the feed port. Matching network sensitivity penalty item and out-of-bounds penalty function The weights are determined by the core objective of the matching network to achieve 50-ohm impedance matching. Set to 1.0; to exclude highly sensitive and fragile solutions, the secondary weights... Based on prior experience with bench machining tolerances, the tolerance is set between 0.1 and 0.5. The extreme value penalty weight is set to 10000 to directly eliminate out-of-bounds solutions.

[0032] By minimizing the above multi-objective fitness function and satisfy The constraint condition of ≤-15dB ultimately yields the result that makes the Y-parameter matrix satisfy the impedance matching condition. Optimal capacitor combination with a voltage rating of ≤-15dB.

[0033] Step 3: Train the neural network model using the physical state vector as the model input. Calculate the loss function by comparing the predicted capacitive reactance value in the predicted capacitive reactance vector of the optimal capacitor combination with the capacitive reactance value in the capacitive reactance vector of the optimal capacitor combination output by the neural network model. After training, the trained neural network model is obtained. This specifically includes the following steps: Step 3.1: Construct a sample set: Construct a physical state vector from the coil physical dimensions and load state parameters. , in physical state vector The initial input samples are used as the basis for normalization, and then all the initial input samples are normalized to obtain the final input samples.

[0034] Normalization of all initial input samples can be achieved using Z-score standardization based on the mean and standard deviation, or Min-Max normalization.

[0035] The capacitive reactance values ​​corresponding to the optimal capacitor combination are used to construct the capacitive reactance vector of the optimal capacitor combination. The capacitive reactance vector of the optimal capacitor combination, formed by the capacitive reactance values ​​corresponding to the optimal capacitor combinations of each input sample. For tags.

[0036] Each input sample and its corresponding label constitute a sample. All samples constitute a sample set, which is then divided into a training set and a test set.

[0037] Step 3.2, as follows Figure 3 As shown, a multilayer perceptron (MLP) model is constructed. The input of the multilayer perceptron model is a normalized physical state vector. The output of the multilayer perceptron model is the predicted capacitive reactance value of the capacitor combination; the training set is input into the multilayer perceptron model for training.

[0038] The multilayer perceptron model of this invention does not directly use pure capacitance values ​​as the model output, but instead utilizes the capacitive reactance vector of the capacitor combination in the complex impedance space. For the model output, a loss function is calculated between the predicted capacitive reactance of the capacitor combination output by the multilayer perceptron model and the capacitive reactance of the capacitor combination in the label. Training is completed by minimizing the loss function. After training, the model parameters are saved to obtain the trained multilayer perceptron model.

[0039] The feature transformation mechanism, which converts capacitance values ​​into capacitive reactance values ​​before calculating model loss, completely eliminates numerical nonlinear singularities near high-frequency resonant points of the capacitor, thus improving the stability of network convergence.

[0040] As one possible implementation, the multilayer perceptron model includes an input layer, an output layer, and 2 to 4 fully connected hidden layers. Each fully connected hidden layer contains 64 to 256 neurons and uses the GELU activation function. During network initialization, to accommodate the nonlinear characteristics of the GELU activation function and avoid gradient vanishing, the weight matrix of the fully connected hidden layers is initialized using the Kaiming normal distribution (He initialization), and the bias vector of each fully connected hidden layer is uniformly initialized to zero. The Adam optimizer is used.

[0041] The multilayer perceptron model trained with capacitive reactance value of the present invention can accurately sense the coupling between physical size and frequency and predict the optimal capacitor combination under load. In practical applications, the theoretical unloaded capacitance value is obtained by using the built-in classical theoretical unloaded capacitance calculation formula. The difference between the unloaded theoretical capacitance value and the predicted capacitance value under load of the present invention can be calculated to further accurately extract the parasitic capacitance deviation introduced by the dielectric load.

[0042] Existing conventional neural networks, lacking microwave physics equations and topological boundary constraints (i.e., without steps 2 and 3), are prone to getting trapped in mathematical local optima, outputting non-physical extrema such as 0.1 pF (approximately open circuit) or solutions exceeding 1000 pF (approximately short circuit). This leads to bench assembly relative errors frequently exceeding 100% or even complete failure. This invention converts capacitance values ​​into capacitive reactance values. The feature transformation mechanism, which first performs an inverse proportional nonlinear mapping between capacitance and reactance before calculating model loss, completely eliminates numerical nonlinear singularities near high-frequency resonant points of the capacitor, improving network convergence stability and significantly reducing model computational complexity.

[0043] Step 4: Construct a physical state vector based on the physical dimensions of the birdcage coil to be processed, the target resonant frequency, and the equivalent load loss resistance. Input the physical state vector into the trained multilayer perceptron model to obtain the predicted capacitive reactance vector of the capacitor combination. Finally, use the inverse physical equation to convert the predicted capacitive reactance value in the predicted capacitive reactance vector of the capacitor combination into the predicted capacitive reactance value. This is converted into a predicted capacitance value, where, To predict the capacitive reactance value, To predict the capacitance value, and thus restore it to the capacitance command value that can be directly assembled on the bench.

[0044] Table 1 compares the predicted results of three randomly selected groups of coil parameters of different sizes and frequencies with the actual measurement results from a virtual network analyzer (VNA) bench in the RF laboratory. As shown in Table 1, the capacitance values ​​of the capacitor combination obtained by the method of the present invention are highly consistent with the bench measurement results, and the relative error of each channel capacitance is stably controlled between 4.6% and 18.9%. In actual RF bench assembly, due to the influence of the inherent 1~3pF distributed parasitic capacitance of PCB traces and shielding cylinder, the prediction error range is within the excellent engineering tolerance range, and can be quickly and finally compensated by connecting conventional 1~12pF ceramic trimmer capacitors in series and parallel.

[0045] This invention eliminates the physically invalid values ​​that violate electromagnetic topology and are blindly output by existing neural networks. While ensuring 50-ohm port matching, it outputs capacitance parameter values ​​that can be directly used for bench assembly. Compared to existing technologies that rely on engineers using handheld network analyzers (VNAs) for blind tuning, which typically takes 3 to 5 working days to tune and match a single coil in the face of impedance disturbances from complex water model loads, this invention bypasses the trial-and-error process, compressing the prediction delay to less than 50 milliseconds, thus significantly improving efficiency.

[0046] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0047] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0048] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0049] It should be noted that the embodiments described in this invention are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains can make various modifications or additions to the described embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for predicting tuning matching of a physically embedded magnetic resonance birdcage coil, characterized in that, Includes the following steps: Step 1: Set the scanning range of the input parameters, perform simulation scanning on different combinations of input parameters of the birdcage coil model, obtain the S-parameter matrix and load state parameters under different combinations of input parameters, and construct the physical state vector of the birdcage coil model based on the coil physical dimensions and load state parameters in the input parameters. Step 2: Obtain the Y-parameter matrix of the L-Match matching network under the combination of each input parameter and the corresponding load state parameters; A multi-objective fitness function is set up, and the multi-objective fitness function is minimized using the differential evolution algorithm to obtain the optimal capacitor combination that makes the Y parameter matrix satisfy the impedance matching condition. The optimal capacitor combination includes the optimal capacitance value of each capacitor in the capacitor combination. Then, the capacitance values ​​in the optimal capacitor combination are converted into capacitive reactance values, and the capacitive reactance values ​​corresponding to the optimal capacitor combination are used to construct the capacitive reactance vector of the optimal capacitor combination. Step 3: Train the neural network model using the physical state vector as the model input. Calculate the loss function for the predicted capacitive reactance value in the predicted capacitive reactance vector of the capacitor combination and the capacitive reactance value in the capacitive reactance vector of the optimal capacitor combination output by the neural network model. After training, the trained neural network model is obtained. Step 4: Construct a physical state vector based on the physical dimensions and load state parameters of the birdcage coil to be processed. Input the physical state vector into the trained neural network model to obtain the predicted capacitive reactance vector of the capacitor combination. Finally, convert the predicted capacitive reactance value in the predicted capacitive reactance vector of the capacitor combination into the predicted capacitance value.

2. The method for predicting tuning matching of a physically embedded magnetic resonance birdcage coil according to claim 1, characterized in that, The input parameters also include the equivalent dielectric constant and equivalent conductivity of the water model load; the physical dimensions of the coil include the coil diameter. and coil length The load state parameters include the target resonant frequency. and equivalent load loss resistance Among them, the equivalent dielectric constant and equivalent conductivity are used to calculate the target resonant frequency and the equivalent load loss resistance; physical state vector .

3. The method for predicting tuning and matching of a physically embedded magnetic resonance birdcage coil according to claim 2, characterized in that, The capacitor assembly includes tuning capacitors. Parallel matching capacitors and series matching capacitor The capacitance values ​​in the optimal capacitor combination are converted into capacitive reactance values, and the tuning capacitive reactance values ​​are obtained respectively. Parallel matching capacitive reactance value and series matching capacitive reactance value The capacitive reactance vector of the capacitor combination .

4. The method for predicting tuning and matching of a magnetic resonance birdcage coil based on physical embedding as described in claim 3, characterized in that, The conversion of capacitance values ​​in the optimal capacitor combination into capacitive reactance values ​​is based on the following formula: ； In the formula, This is the capacitance value. This is the capacitance value.

5. The method for predicting tuning matching of a physically embedded magnetic resonance birdcage coil according to claim 1, characterized in that, The multi-objective fitness function is: ； In the formula, For multi-objective fitness functions; To match the network sensitivity penalty, For out-of-bounds penalty functions, , as well as These are the return losses at the feed port. Matching network sensitivity penalty item and out-of-bounds penalty function The weight, For modulo operation; The matching network sensitivity penalty term is set to the absolute value of the difference between the tuning capacitor and the parallel matching capacitor; When the tuning capacitor Or parallel matching capacitor Or series matching capacitor When the value exceeds the corresponding range, =1, otherwise =0.

6. The method for predicting tuning matching of a physically embedded magnetic resonance birdcage coil according to claim 5, characterized in that, Feed port return loss weight Set to 1; Match network sensitivity penalty item weight Set the value to 0.1~0.5; set the weight of the out-of-bounds penalty function to 10000. .

7. The method for predicting tuning matching of a magnetic resonance birdcage coil based on physical embedding as described in claim 1, characterized in that, The impedance matching condition is as follows: ≤-15dB.

8. The method for predicting tuning matching of a physically embedded magnetic resonance birdcage coil according to claim 1, characterized in that, The neural network model is a multilayer perceptron model, which includes an input layer, an output layer, and 2 to 4 fully connected hidden layers. Each fully connected hidden layer includes 64 to 256 neurons and uses the GELU activation function. In the initialization phase of the multilayer perceptron model, the weight matrix of the fully connected hidden layer is initialized using the Kaiming normal distribution, and the bias vector of each fully connected hidden layer is uniformly initialized to zero; the Adam optimizer is used.

9. The method for predicting tuning matching of a magnetic resonance birdcage coil based on physical embedding as described in claim 8, characterized in that, The loss function of the multilayer perceptron model is the MSE loss function.

10. A computer 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 physically embedded magnetic resonance birdcage coil tuning matching prediction method according to any one of claims 1 to 9.

Images