Gradient coil model correction method for ultralow field magnetic resonance

Abstract

The invention provides a gradient coil model correction method for ultra-low field magnetic resonance, and relates to the technical field of gradient coil control in medical imaging equipment, and the method comprises the steps: collecting a real voltage signal d (n) and a real current signal x (n) of a gradient coil in real time, and n represents a sampling moment; a gradient coil model is constructed, the gradient coil model comprises a linear model constructed based on an adaptive filtering algorithm and a nonlinear model constructed based on a neural network algorithm, and a first prediction voltage signal # imgabs0 # output by the linear model and a second prediction voltage signal # imgabs1 # output by the nonlinear model are determined according to the real current signal x (n); determining a mixed prediction voltage signal # imgabs4 # of the gradient coil model according to the first prediction voltage signal # imgabs2 # and the second prediction voltage signal # imgabs3 #; and updating the weight of the linear model and the weight of the nonlinear model according to the real voltage signal d (n) and the mixed predicted voltage signal # imgabs5 # so as to adapt to gradient coil model correction under complex working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of gradient coil control in medical imaging devices, and more particularly to a method for correcting a gradient coil model for ultra-low field magnetic resonance. Background Art

[0002] Ultra-low field magnetic resonance is an important branch in the field of magnetic resonance imaging, generally referring to a magnetic resonance imaging system with a magnetic field strength lower than 0.5T. In ultra-low field magnetic resonance, higher performance requirements are imposed on the gradient power amplifier. During the magnetic resonance scanning process, the gradient coil generates a gradient magnetic field by rapidly switching the current. However, under extreme working conditions (such as high load, insufficient heat dissipation, and coil aging), it may cause an arc due to insulation failure or mechanical deformation, threatening the safety of patients. Existing technologies preset a static load model in the gradient controller, predict the theoretical output voltage and compare it with the actual value. If the threshold is exceeded, the scan is terminated. However, the traditional solution has the following limitations: 1) The fixed-parameter model cannot dynamically track the load changes, and the error increases with the drift of the working conditions. The load characteristics of the gradient coil can be expressed as a combination of resistance, inductance, and mutual inductance transformer. Among them, the series resistance of the coil and the contact resistance at the connection terminal are both affected by the change in the coil temperature; at the same time, along with the aging of the copper wire of the coil and the oxidation of the insulating material, the inductance and mutual inductance transformer parameters are also affected. However, the default linear model at the factory cannot adapt to the changes in the load characteristics caused by factors such as temperature changes and coil aging, which further leads to an increase in the error between the model prediction result and the actual measured voltage, resulting in false positives and false negatives in the abnormal arc identification and affecting the normal scan imaging; 2) Nonlinearity neglect: The traditional linear model cannot characterize the nonlinear response caused by coil aging and deformation, and the risk of false negatives is high. Summary of the Invention

[0003] In view of this, an object of the present invention is to provide a method for correcting a gradient coil model for ultra-low field magnetic resonance, so as to adapt to the correction of the gradient coil model under complex working conditions by designing a dynamic correction and compensation gradient coil model based on a linear model constructed by an adaptive filtering algorithm and a nonlinear model constructed by a neural network algorithm through hybrid load modeling.

[0004] An embodiment of the present invention provides a method for correcting a gradient coil model for ultra-low field magnetic resonance, including: real-time collecting a true voltage signal d(n) and a true current signal x(n) of the gradient coil, where n represents the sampling time; constructing a gradient coil model, the gradient coil model including a linear model constructed based on an adaptive filtering algorithm and a nonlinear model constructed based on a neural network algorithm, wherein a first predicted voltage signal output by the linear model is determined according to the true current signal x(n) and a second predicted voltage signal output by the nonlinear model ; according to the first predicted voltage signal and the second predicted voltage signal determine the hybrid predicted voltage signal of the gradient coil model ; according to the true voltage signal d(n) and the hybrid predicted voltage signal , update the weights of the linear model and the weights of the non - linear model to achieve the correction of the gradient coil model.

[0005] In some embodiments, the linear model constructed based on the adaptive filtering algorithm and the non - linear model constructed based on the neural network algorithm include: the expression of the linear model constructed based on the LMS algorithm is , where k is the order, is the first weight; the expression of the non - linear model constructed based on the neural network algorithm is , where k is the order, is the second weight.

[0006] In some embodiments, according to the first predicted voltage signal and the second predicted voltage signal determine the hybrid predicted voltage signal of the gradient coil model , including: the expression of the hybrid predicted voltage signal of the gradient coil model is , where is and 's combined weight, and the combined weight is the third weight.

[0007] In some embodiments, according to the true voltage signal d(n) and the hybrid predicted voltage signal , update the weights of the linear model and the weights of the non - linear model to achieve the correction of the gradient coil model, including: constructing a correction parameter model, the expression of the correction parameter model is , where is the step - size factor; is the instantaneous voltage error; correct the gradient coil model according to the correction parameter model.

[0008] In some embodiments, the order m of the non - linear model satisfies m≥4 and is a positive integer.

[0009] In some embodiments, after real - time collecting the true voltage signal d(n) and the true current signal x(n) of the gradient coil, it includes: inputting the true voltage signal d(n) and the true current signal x(n) into the trained noise reduction model, and outputting the noise - reduced voltage signal and the noise - reduced current signal, and the noise - reduced voltage signal and the noise - reduced current signal are used for the training of the gradient coil model.

[0010] In some embodiments, the step factor is determined according to the autocorrelation matrix of the true current signal x(n).

[0011] In some embodiments, based on the instantaneous voltage errors obtained at different sampling moments, the state data at the current sampling moment is determined, and the Q-learning algorithm is used to determine the step factor according to the state data at the current sampling moment , and the current state data includes the mean value of the instantaneous voltage error, the variance of the instantaneous voltage error, and the load power, and the optimal step is selected by maximizing the long-term reward.

[0012] In some embodiments, when the absolute value of the instantaneous voltage error is greater than a preset error threshold within T sampling periods, a warning message is sent, where T≥5 and T is a positive integer.

[0013] In some embodiments, after the absolute value of the instantaneous voltage error is greater than a preset error threshold within T sampling periods, it further includes: inputting the instantaneous voltage error into a trained classification model to output a fault probability; and sending the warning message according to the fault probability and a preset probability threshold.

[0014] The embodiments of the present invention bring the following beneficial effects: 1) There is a gradient coil model with dynamic correction and compensation of a linear model constructed based on an adaptive filtering algorithm and a nonlinear model constructed based on a neural network algorithm, which has the advantage of adapting to the correction of the gradient coil model under complex working conditions (such as temperature drift and coil aging); 2) The compensation factor is dynamically adjusted through Q-learning reinforcement learning to balance the convergence speed and stability; 3) The false alarm rate is reduced through the joint decision of the primary threshold detection and the logistic regression classification model.

[0015] Other features and advantages of the present disclosure will be described in the following description, or some features and advantages can be inferred from the description or determined without doubt, or can be learned by implementing the above technologies of the present disclosure.

[0016] To make the above objects, features, and advantages of the present disclosure more obvious and understandable, the following specific preferred embodiments are given below, and in conjunction with the accompanying drawings, the detailed description is as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0018] Figure 1 This is a block diagram of an implementation of a gradient coil model correction method for ultra-low field magnetic resonance provided by an embodiment of the present invention; Figure 2 This is a flowchart of a gradient coil model correction method for ultra-low field magnetic resonance provided by an embodiment of the present invention. Specific implementation manner

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0020] To facilitate the understanding of this embodiment, first, a magnetic resonance image analysis system disclosed in the embodiments of the present invention will be introduced in detail.

[0021] Combined with Figure 1 and Figure 2 The embodiments of the technical solution of the present invention will be described in detail. Specifically, the present invention constructs a gradient coil model, which includes a linear model constructed based on an adaptive filtering algorithm and a non-linear model constructed based on a neural network algorithm.

[0022] Among them, the linear model constructed based on the adaptive filtering algorithm can be simplified to a series combination of an equivalent inductance L(n) and an equivalent resistance R(n), and a weight vector is defined , where n represents the sampling time, k is the order of the LMS (Least Mean Squares) model, and the LMS model refers to a linear adaptive filter based on the least mean square algorithm. The process of calculating the predicted voltage of the linear model is as follows: The ordinary differential expression of the linear part of the gradient coil model is: Where: t is a continuous time variable (unit: second), representing the actual time when the gradient coil works; v(t) is the coil terminal voltage at time t; i(t) is the coil current at time t; L is the equivalent inductance; R is the equivalent resistance.

[0023] Discretizing it, we can get: The discrete time variable n represents the nth sampling time (such as n = 0, 1, 2,...), corresponding to the time t = n Ts, where Ts is the sampling period; Denotes the predicted voltage output of the linear model at time n.

[0024] Parametrize the time-varying inductance L(n) and resistance R(n), and denote the discrete current signal as x(n), obtaining Finally, represent the parameters using the weight vector to obtain: where, mainly reflects the resistance component, reflects the inductance component; Rewrite the above expression into the form of the FIR model expression as follows: Through the above modeling process, the expression of the linear model is obtained.

[0025] Exemplarily, the recursive least squares (RLS) algorithm can also be used to replace the LMS algorithm for linear model construction, which has fast convergence but high computational complexity.

[0026] In addition, the non-linear part of the gradient coil exhibits characteristics such as eddy currents and c-cross-talk between different axes (X / Y / Z). Generally, a mutual inductance transformer is used to model it, and in mathematics, an IIR filter is usually used instead of an FIR filter. However, considering that the B0 field of the low-field magnetic resonance device is much smaller than that of the conventional magnetic resonance device, and the gradient coil voltage is also much smaller, the contribution of the non-linear part can be further characterized by a high-order FIR (finite impulse response) filter. Specifically, the expression of the non-linear model is , where k is the order, is the weight, where the order m ≥ 4 and is a positive integer. In this embodiment, m = 4 is taken as an example for illustration. The non-linear model of this embodiment adopts a 4th-order FIR-structured micro neural network, and learns the non-linear characteristics (such as eddy currents and crosstalk) of the gradient coil through offline pre-training (using real current data monitored historically) and online fine-tuning (through the instantaneous voltage error ), and updates the weights through backpropagation.

[0027] Both the linear model and the non-linear model are constructed through the FIR model. Therefore, the weights can be combined, is and 's combination. Therefore, the expression of the hybrid predicted voltage signal of the gradient coil model constructed by the present invention is .

[0028] In this embodiment, the hybrid predicted voltage signal The expression of is

[0029] An embodiment of the present invention provides a method for accelerating the calculation of a hybrid model (a linear model constructed based on the LMS algorithm and a nonlinear model constructed based on the neural network algorithm) using an FPGA coprocessor, achieving an effect of a calculation delay < 50 μs. At the same time, an 8-bit fixed-point quantization technique is used to compress the neural network weights from 32-bit floating-point to 8-bit integer, and the memory occupancy is reduced from 128 KB to < 50 KB. Through hardware acceleration and resource optimization, intelligent deployment is realized, so that even in a low-computing-power controller of an ultra-low-field magnetic resonance, the gradient coil model provided by the present invention can respond and operate quickly.

[0030] The present invention provides an embodiment, which inputs a real voltage signal d(n) and a real current signal x(n) into a trained noise reduction model, and outputs a noise-reduced voltage signal and a noise-reduced current signal for the training of the nonlinear model. The noise-reduced voltage signal and the noise-reduced current signal are used for the training of the gradient coil model, adaptively denoise the collected signal, retain the transient characteristics of the signal, and improve the signal-to-noise ratio.

[0031] In the embodiment provided by the present invention, the training process of the nonlinear model constructed based on the neural network includes: using the real current signal x(n) collected at different sampling times as the input of the model, and using the real nonlinear voltage distortion component = d(n) - as the label, and using and to construct an MSE (mean square error) loss function. The training of the model is affected by the step factor, which affects the stability of the model training.

[0032] According to the real voltage signal d(n) and the hybrid predicted voltage signal , update the weights of the linear model and the nonlinear model to realize the correction of the gradient coil model, including: constructing a correction parameter model, and the expression of the correction parameter model is , where is the step factor; is the instantaneous voltage error; correct the gradient coil model according to the correction parameter model. Specifically, use the Q-learning algorithm to determine the step factor according to the state data at the current sampling time, and the state data includes the mean value of the instantaneous voltage error, the variance of the instantaneous voltage error, and the load power. Exemplarily, the affine projection algorithm (APA) can also be used to replace Q-learning to simplify the step adjustment, but additional storage space is required.

[0033] It can also be determined according to the autocorrelation matrix of the real current signal x(n) collected in real time, where n represents the sampling time, n≥1 and is a positive integer. Calculate the maximum eigenvalue of the autocorrelation matrix , and the step factor needs to satisfy to ensure the balance between the convergence speed and stability of the algorithm.

[0034] This embodiment also provides an early warning judgment scheme. When the absolute value of the instantaneous voltage error is greater than the preset error threshold within T sampling periods, an early warning message is sent. T≥5 and is a positive integer, realizing primary threshold detection. At the same time, a logistic regression model is constructed to substitute and calculate the input error sequence to obtain the probability of an arc (fault) occurrence. The logistic regression model is trained with a large amount of instantaneous voltage error data during arcing to conduct joint early warning of risks through the above multi-level arc determination, improving the accuracy of early warning.

[0035] Unless otherwise specifically stated, the relative steps, numerical expressions, and values of the components and steps set forth in these embodiments do not limit the scope of the present invention.

[0036] In all the examples shown and described here, any specific value should be construed as merely exemplary and not as a limitation. Therefore, other examples of the exemplary embodiments may have different values.

[0037] Finally, it should be noted that the above-described embodiments are only specific embodiments of the present invention, used to illustrate the technical solutions of the present invention, and are not intended to limit it. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: any person skilled in the art within the technical scope disclosed by the present invention can still modify or easily conceive of changes to the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications, changes, or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for correcting a gradient coil model for ultra-low field magnetic resonance, characterized in that, Including: Real-time collecting the true voltage signal d(n) and the true current signal x(n) of the gradient coil, where n represents the sampling moment; Construct a gradient coil model, where the gradient coil model includes a linear model constructed based on an adaptive filtering algorithm and a nonlinear model constructed based on a neural network algorithm, and the first predicted voltage signal output by the linear model is determined according to the true current signal x(n). and the second predicted voltage signal output by the nonlinear model ; Based on the first predicted voltage signal and the second predicted voltage signal determine the hybrid predicted voltage signal of the gradient coil model ; Based on the real voltage signal d(n) and the hybrid predicted voltage signal , update the weights of the linear model and the weights of the non-linear model to achieve the correction of the gradient coil model.

2. The method according to claim 1, characterized in that, After the real-time collecting the true voltage signal d(n) and the true current signal x(n) of the gradient coil, including: Inputting the true voltage signal d(n) and the true current signal x(n) into the trained noise reduction model to output the noise-reduced voltage signal and the noise-reduced current signal, and the noise-reduced voltage signal and the noise-reduced current signal are used for the training of the gradient coil model.

3. The method according to claim 1, characterized in that, The linear model constructed based on the adaptive filtering algorithm and the nonlinear model constructed based on the neural network algorithm, including: The expression of the linear model constructed based on the LMS algorithm is , where k is the order, is the first weight; The expression of the non-linear model constructed based on the neural network algorithm is , where k is the order, is the second weight.

4. The method according to claim 3, wherein Including: The order m of the nonlinear model satisfies m≥4 and is a positive integer.

5. The method according to claim 3, characterized in that, determining a hybrid predicted voltage signal of the gradient coil model based on the first predicted voltage signal and the second predicted voltage signal comprises: ​ The hybrid prediction voltage signal of the gradient coil model is expressed as , where is and The combined weight, and the combined weight is the third weight.

6. The method according to claim 3, wherein Said according to the real voltage signal d(n) and the hybrid predicted voltage signal , updating the weights of the linear model and the weights of the nonlinear model to achieve correction of the gradient coil model, including: Construct a calibration parameter model, and the expression of the calibration parameter model is , where is the step factor; is the instantaneous voltage error; Correcting the gradient coil model according to the correction parameter model.

7. The method according to claim 6, wherein Including: The step factor is determined according to the autocorrelation matrix of the true current signal x(n).

8. The method according to claim 6, wherein Including: Based on the instantaneous voltage error obtained at different sampling moments, determining the state data at the current sampling moment, where the state data includes the mean value of the instantaneous voltage error, the variance of the instantaneous voltage error, and the load power; The step size factor is determined according to the state data at the current sampling moment by using the Q-learning algorithm .

9. The method according to claim 6, wherein Also including: When the absolute value of the instantaneous voltage error is greater than a preset error threshold within T sampling periods, sending out a warning message, where T≥5 and is a positive integer.

10. The method according to claim 9, characterized in that, After the absolute value of the instantaneous voltage error is greater than a preset error threshold within T sampling periods, further including: Inputting the instantaneous voltage error into the trained classification model to output the failure probability; Sending out the warning message according to the failure probability and the preset probability threshold.

Images