Magnetic resonance signal denoising method and device based on physical prior gating mechanism

Abstract

The invention discloses a magnetic resonance signal de-noising method and device based on a physical prior gating mechanism, and the method employs the magnetic resonance physical principle, i.e., the real signal of a biological tissue maintains the physical consistency (the amplitude and phase are basically unchanged) in two times of collection, and the transient EMI noise in the environment has the characteristics of randomness and non-stationarity. A self-adaptive dead zone gating mechanism based on amplitude energy and phase perception is provided, namely, the identification of transient high-amplitude electromagnetic interference in two magnetic resonance signals collected at the same k space position is realized by utilizing the physical consistency among multiple collection of the magnetic resonance signals, and the fusion weight of transient EMI interference signal points is dynamically adjusted, so that the transient high-amplitude electromagnetic interference in the two magnetic resonance signals collected at the same k space position is identified. Therefore, non-linear elimination of non-stationary transient electromagnetic interference is realized. Therefore, the method can accurately recognize and eliminate accidental transient high-amplitude electromagnetic interference, avoids the problem of denoising failure, and can remarkably improve the purity of the image.

Description

Technical Field

[0001] This invention belongs to the field of magnetic resonance imaging technology, specifically relating to a method and apparatus for denoising magnetic resonance signals based on a physical prior gating mechanism. Background Technology

[0002] Traditional MRI requires placement in a shielded room to isolate electromagnetic interference noise from the environment. One purpose of portable (low-field) MRI instruments is to enable the equipment to be moved, which requires no shielded room. However, MRI in an open environment faces the problem of electromagnetic interference (EMI), which leads to severe EMI noise in the resulting images, thus seriously affecting image quality.

[0003] EMI noise is an additive noise, meaning that data affected by EMI noise consists of the signal plus the EMI noise. Existing methods for removing EMI noise generally require multiple external electromagnetic interference induction coils (EMI coils) to receive environmental EMI noise while acquiring magnetic resonance signals (MRI coils), thus using this as prior knowledge to remove EMI noise within the imaging coil. At the same time, in order to fit the mapping relationship between EMI noise in the MRI coil and the EMI coils, a signal needs to be sampled in the absence of MRI signals (when only environmental EMI noise is present) as a calibration signal for fitting the mapping relationship.

[0004] Currently, Deep-DSP is commonly used for EMI denoising (such as the existing technology with application number CN202310504030.5, and the reference article: Robust EMI elimination for RF shielding-free MRI through deep learning direct MR signal prediction). This method is a deep learning approach that maps EMI-noisy MRI signals and signals from EMI coils into clean, noise-free MRI signals. Its training input comes from calibration data and additional EMI-free MRI signals, constructed as follows: one channel superimposes the EMI-free, high signal-to-noise ratio MRI signal onto the environmental EMI noise acquired by the MRI coil in the calibration data. This simulates the actual MRI signal + EMI noise situation; other training inputs come from the calibration signal of the EMI coil; and the training label is the aforementioned EMI-free, high signal-to-noise ratio MRI signal.

[0005] Specifically, the training process is as follows: Figure 1As shown in Figure (C), the training of this network essentially involves mapping all channel data (including MRI and EMI coils) to a clean, EMI-free MRI signal. A training sample is a line acquired by the MRI and EMI coils; that is, it's a one-dimensional network. The testing process involves inputting a k-space line and all data acquired simultaneously from the EMI coils into the network to obtain an EMI-free MRI k-space line, as shown in Figure (C). Figure 1 As shown in Figure (D).

[0006] However, the aforementioned prior art has the following shortcomings: (1) This method has a good effect on suppressing stable broadband background noise, but it is not good at suppressing transient strong EMI noise (occasional, high amplitude, such as spike noise, sweep frequency noise, etc.), and cannot effectively suppress this type of noise, such as Figure 2 As shown, Figure 2 Figure (a) shows the k-space (signal received by the MRI coil) containing EMI noise. It can be seen that in addition to the broadband EMI noise, there is also a very strong short-duration segment of EMI noise. Figure 2 Figure (b) shows the k-space after Deep-DSP denoising. It can be seen that this strong EMI noise still has residue. Figure 2 Figure (c) in the middle is the reconstructed image. It can be seen that this small segment of EMI noise has a serious interference to the image, thus seriously affecting the imaging quality; (2) In conventional magnetic resonance scanning, the signal-to-noise ratio is usually improved by repeatedly acquiring (NEX>1) and averaging (especially portable low-field magnetic resonance, which is more necessary due to its low signal-to-noise ratio); However, based on the above technical defects 1, since the amplitude of the above transient interference is much higher than that of the normal magnetic resonance signal, simple linear averaging cannot eliminate the high-energy anomaly value, but will instead spread the interference energy into the final image, thus causing the denoising failure.

[0007] Therefore, given the aforementioned shortcomings, how to provide a magnetic resonance signal denoising method based on a physical prior gating mechanism that can remove transient strong EMI noise and thus improve the quality of magnetic resonance imaging has become an urgent problem to be solved. Summary of the Invention

[0008] The purpose of this invention is to provide a magnetic resonance signal denoising method and apparatus based on a physical prior gating mechanism, in order to solve the problem that the existing technology cannot suppress transient strong EMI noise, which leads to denoising failure and thus affects the imaging quality.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a magnetic resonance signal denoising method based on a physical prior gating mechanism is provided, including: Acquire a first magnetic resonance signal and a second magnetic resonance signal obtained by at least two repeated acquisitions at the same k-space location in an unshielded environment, wherein both the first magnetic resonance signal and the second magnetic resonance signal include signals from the MRI coil and the EMI coil; A signal denoising network is constructed, which includes a twin residual network and an adaptive dead-time heuristic gating. The first magnetic resonance signal and the second magnetic resonance signal are respectively input into the twin residual network in the signal denoising network for initial denoising processing to obtain the first initial denoised signal and the second initial denoised signal; The first initial denoised signal and the second initial denoised signal are input to the adaptive dead-zone heuristic gating system for local energy difference calculation and phase conflict detection. Based on the local energy difference and phase conflict detection results, the fusion weight of each signal point in the first initial denoised signal and the second initial denoised signal is determined. When the local energy difference is greater than or equal to the energy threshold, or when the phase conflict detection result indicates the presence of a phase conflict, the fusion weight of the signal point with higher energy in the two initial denoised signals is less than the fusion weight of the signal point with lower energy. Based on the fusion weights of each signal point in the first and second initial denoised signals, the first and second initial denoised signals are fused to obtain a denoised magnetic resonance signal.

[0010] Based on the above-disclosed content, this invention first acquires a first magnetic resonance signal and a second magnetic resonance signal obtained by at least two repeated acquisitions of the same k-space location under an unshielded environment; then, it constructs a signal denoising network including a twin residual network and an adaptive dead-zone heuristic gating to perform signal denoising, that is, using a weight-sharing twin residual network to perform preliminary denoising on the aforementioned two magnetic resonance signals; then, it uses adaptive dead-zone heuristic gating to remove transient strong interference that cannot be handled by existing technologies; specifically, it performs local energy difference calculation and phase conflict detection on the two initial denoised signals, thereby adaptively determining the two... The initial denoised signals are weighted according to the fusion weights of each signal point. When the local energy difference is greater than or equal to the energy threshold, or when the phase conflict detection result indicates the presence of a phase conflict, it indicates that the higher energy signal is contaminated by transient EMI (because the energy of a real MRI signal is stable, and a sudden increase in energy must originate from additive noise). Therefore, a high fusion weight is assigned to the low energy signal, and a low fusion weight is assigned to the high energy signal. At this time, the high energy signal receives a very small weight, which is equivalent to completing the removal of high-energy noise. Finally, the signals are fused according to the fusion weights of each signal point in the two initial denoised signals to obtain the denoised magnetic resonance signal.

[0011] Through the above design, this invention utilizes the physical principle of magnetic resonance, namely that the real signals of biological tissue maintain physical consistency (amplitude and phase remain basically unchanged) in two acquisitions, while transient EMI noise in the environment has randomness and non-stationarity (it may only appear once in two acquisitions, or the intensity may differ greatly). It proposes an adaptive dead-zone gating mechanism based on amplitude energy and phase perception. This mechanism uses the physical consistency between multiple acquisitions of the aforementioned magnetic resonance signal to identify transient high-amplitude electromagnetic interference in two magnetic resonance signals acquired at the same k-space location, and dynamically adjusts the fusion weight of transient EMI interference signal points, thereby achieving nonlinear elimination of non-stationary transient electromagnetic interference. Therefore, this invention can accurately identify and eliminate occasional transient high-amplitude electromagnetic interference, avoiding the problem of noise reduction failure, and thus significantly improving image purity.

[0012] In one possible design, the Siamese residual network includes: a first deep fully convolutional residual subnetwork and a second deep fully convolutional residual subnetwork, wherein the first deep fully convolutional residual subnetwork and the second deep fully convolutional residual subnetwork share the same network structure and weights; The first and second magnetic resonance signals are respectively input into the twin residual network in the signal denoising network for initial denoising processing, including: Obtain the EMI prior signals corresponding to the first and second magnetic resonance signals, respectively. The first magnetic resonance signal and the corresponding EMI prior signal are input into a first depth fully convolutional residual subnetwork for initial denoising processing, and the second magnetic resonance signal and the corresponding EMI prior signal are input into a second depth fully convolutional residual subnetwork for initial denoising processing, so as to obtain the first initial denoised signal and the second initial denoised signal respectively.

[0013] In one possible design, the first deep fully convolutional residual subnetwork includes: a head convolutional layer, a residual convolutional layer and a tail convolutional layer connected in sequence, wherein the residual convolutional layer includes a plurality of residual convolutional blocks, and any residual convolutional block includes a first convolutional layer, a ReLU layer and a second convolutional layer connected in sequence.

[0014] In one possible design, the first and second initial denoised signals are input to the adaptive dead-time heuristic gating system for local energy difference calculation and phase conflict detection, including: The first initial denoised signal and the second initial denoised signal are subjected to modulus taking and smoothing processing to obtain the first envelope energy curve and the second envelope energy curve, respectively. Based on the first envelope energy curve and the second envelope energy curve, the local energy difference between signal points at the same position in the first initial denoised signal and the second initial denoised signal is calculated. The complex dot product of signal points at the same position in the first and second initial denoised signals is calculated, and the phase conflict detection result of signal points at the same position in the first and second initial denoised signals is obtained based on the complex dot product of signal points at the same position.

[0015] In one possible design, based on the first envelope energy curve and the second envelope energy curve, the local energy difference between signal points at the same locations in the first and second initial denoised signals is calculated, including: Calculate the energy difference between the i-th signal point in the second envelope energy curve and the i-th signal point in the first envelope energy curve, and calculate the sum of the energies of the i-th signal point in the second envelope energy curve and the i-th signal point in the first envelope energy curve, where i is a positive integer; The ratio between the energy difference and the energy sum is used as the local energy difference between the i-th signal point in the first and second initial denoised signals. Increment i by 1 and recalculate the energy difference between the i-th signal point in the second envelope energy curve and the i-th signal point in the first envelope energy curve until i equals n. This yields the local energy difference between signal points at the same positions in the first and second initial denoised signals. Here, the initial value of i is 1, and n is the signal length of the first initial denoised signal. Accordingly, based on the complex dot product of signal points at the same locations, the phase conflict detection results of signal points at the same locations in the first and second initial denoised signals are obtained, including: For any two signal points at the same position in the first and second initial denoised signals, determine whether the complex dot product of the two signal points is negative; If so, it is determined that there is a phase conflict between the two signal points at any of the same positions.

[0016] In one possible design, based on the local energy difference and phase conflict detection results, the fusion weights of each signal point in the first and second initial denoised signals are determined, including: For any two signal points at the same position in the first initial denoised signal and the second initial denoised signal, determine whether the local energy difference between the two signal points is greater than or equal to the energy threshold, or whether the phase conflict detection result indicates the presence of a phase conflict. If so, the first initial weight is calculated based on the local energy difference between the two signal points, and the difference between 1 and the first initial weight is used as the second initial weight; The largest of the first and second initial weights is used as the fusion weight of the signal point with the lowest energy value among the two signal points, and the smallest of the first and second initial weights is used as the fusion weight of the signal point with the highest energy value among the two signal points. After all signal points in the first and second initial denoised signals have been polled, the fusion weight of each signal point in the first and second initial denoised signals is obtained.

[0017] In one possible design, if the local energy difference between two signal points is less than an energy threshold, and the phase conflict detection result indicates no phase conflict exists, then the method further includes: Set the same fusion weight for both signal points; The first initial weight is calculated based on the local energy difference between the two signal points, including: The first initial weight is calculated based on the local energy difference between the two signal points and using the Sigmoid function.

[0018] In one possible design, a signal denoising network is trained using magnetic resonance training data as input and denoised magnetic resonance signals as output. Each set of magnetic resonance training data includes a first sample magnetic resonance signal and a second sample magnetic resonance signal, and the loss function of the signal denoising network is: ; In the formula, For loss function, The first sample magnetic resonance signal is the initial denoised signal output after being input into the twin residual network in the signal denoising network. The second sample magnetic resonance signal is the initial denoised signal output after being input into the twin residual network in the signal denoising network. This is the denoised magnetic resonance signal corresponding to the magnetic resonance training data. The label data corresponding to the magnetic resonance training data. It is an L1 norm.

[0019] Secondly, a magnetic resonance signal denoising device based on a physical prior gating mechanism is provided, comprising: The acquisition unit is used to acquire a first magnetic resonance signal and a second magnetic resonance signal obtained by at least two repeated acquisitions at the same k-space location in an unshielded environment, wherein the first magnetic resonance signal and the second magnetic resonance signal both include signals in the MRI coil and the EMI coil. The network building unit is used to construct a signal denoising network, which includes a twin residual network and an adaptive dead-time heuristic gating. The denoising unit is used to input the first magnetic resonance signal and the second magnetic resonance signal into the twin residual network in the signal denoising network for initial denoising processing to obtain the first initial denoised signal and the second initial denoised signal. The denoising unit is used to input the first initial denoised signal and the second initial denoised signal into the adaptive dead-zone heuristic gating to calculate the local energy difference and detect the phase conflict. Based on the local energy difference and the phase conflict detection results, the fusion weight of each signal point in the first initial denoised signal and the second initial denoised signal is determined. When the local energy difference is greater than or equal to the energy threshold, or when the phase conflict detection result indicates the presence of a phase conflict, the fusion weight of the signal point with higher energy in the two initial denoised signals is less than the fusion weight of the signal point with lower energy. The denoising unit is also used to perform fusion processing on the first initial denoised signal and the second initial denoised signal according to the fusion weight of each signal point in the first initial denoised signal and the second initial denoised signal, so as to obtain the denoised magnetic resonance signal after the fusion processing.

[0020] Thirdly, another magnetic resonance signal denoising device based on a physical prior gating mechanism is provided. Taking the device as an electronic device as an example, it includes a memory, a processor, and a transceiver that are connected in sequence. The memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the magnetic resonance signal denoising method based on a physical prior gating mechanism as described in the first aspect or any possible design of the first aspect.

[0021] Fourthly, a storage medium is provided, on which instructions are stored, which, when executed on a computer, perform the magnetic resonance signal denoising method based on a physical prior gating mechanism as described in the first aspect or any possible design of the first aspect.

[0022] Fifthly, a computer program product containing instructions is provided, which, when executed on a computer, causes the computer to perform the magnetic resonance signal denoising method based on a physical prior gating mechanism as described in the first aspect or any possible design of the first aspect.

[0023] Beneficial effects: (1) This invention proposes an adaptive dead-zone gating mechanism based on amplitude energy and phase perception. That is, by utilizing the physical consistency prior between multiple acquisitions of magnetic resonance signals, the transient high-amplitude electromagnetic interference in two magnetic resonance signals acquired at the same k spatial position is identified, and the fusion weight of transient EMI interference signal points is dynamically adjusted, thereby realizing the nonlinear elimination of non-stationary transient electromagnetic interference. Thus, this invention can accurately identify and eliminate occasional transient high-amplitude electromagnetic interference, avoiding the problem of noise reduction failure, thereby significantly improving the purity of the image. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of training and denoising corresponding to traditional Deep-DSP denoising provided by the present invention; Figure 2 This is a comparison chart of the effects of traditional Deep-DSP noise reduction provided in this embodiment of the invention; Figure 3 A flowchart illustrating the steps of a magnetic resonance signal denoising method based on a physical prior gating mechanism provided in an embodiment of the present invention; Figure 4 A flowchart for magnetic resonance signal denoising provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the first depthwise fully convolutional residual subnetwork provided in an embodiment of the present invention; Figure 6 A logic curve diagram of adaptive dead-time heuristic gating provided in an embodiment of the present invention; Figure 7 This is a schematic diagram comparing the noise reduction of traditional technologies with those provided in the embodiments of the present invention; Figure 8 This is a schematic diagram comparing the imaging of the present invention with that of conventional techniques, provided in an embodiment of the invention. Figure 9 A structural diagram of a magnetic resonance signal denoising device based on a physical prior gating mechanism provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0026] It should be understood that although the terms first, second, etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit, without departing from the scope of the exemplary embodiments of the invention.

[0027] It should be understood that the term "and / or" that may appear in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" that may appear in this document describes another relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " that may appear in this document generally indicates that the related objects before and after it are in an "or" relationship.

[0028] Example: See Figure 3 As shown, the magnetic resonance signal denoising method based on physical prior gating mechanism provided in this embodiment designs an adaptive dead-zone gating mechanism based on amplitude and phase perception. That is, it uses the physical consistency prior between multiple magnetic resonance acquisitions to achieve nonlinear elimination of non-stationary transient electromagnetic interference in the signal. In this way, occasional transient high-amplitude electromagnetic interference can be accurately identified and eliminated, thereby significantly improving the purity of the imaging image (the denoising referred to in this embodiment refers to the removal of EMI noise). For example, this method can be run on the signal denoising end side. Optionally, the signal denoising end can be, but is not limited to, a personal computer (PC) or a server. It is understood that the aforementioned execution subject does not constitute a limitation on the embodiment of this application. Accordingly, the operation steps of this method can be, but are not limited to, the steps S1 to S5 below.

[0029] S1. Acquire a first magnetic resonance signal and a second magnetic resonance signal obtained by at least two repeated acquisitions of the same k-space location in an unshielded environment, wherein both the first and second magnetic resonance signals include signals from the MRI coil and the EMI coil; in specific implementation, for example, in an unshielded environment, two acquisitions are performed on the same k-space location, and the time interval between the two acquisitions is less than a preset duration (the preset duration is set to a small value, such as 0.2s, of course, it can be set according to actual use); at this time, according to the physical principle of magnetic resonance, that is, the real signal of biological tissue maintains physical consistency in the two acquisitions (amplitude and phase remain basically unchanged), while the transient EMI noise in the environment has the characteristics of randomness and non-stationarity (it may only appear once in the two acquisitions, or the intensity difference is huge), to compare the amplitude and sense the phase conflict of the two magnetic resonance signals acquired above, thereby identifying and eliminating transient high-amplitude electromagnetic interference in the signal based on amplitude comparison and phase conflict sensing, and thus obtaining the final denoised magnetic resonance signal.

[0030] The specific noise reduction process is shown in steps S2 to S5 below.

[0031] S2. A signal denoising network is constructed, which includes a twin residual network and an adaptive dead-time heuristic gating. In specific implementation, the twin residual network performs preliminary denoising on the two acquired magnetic resonance signals through convolution processing, thereby outputting the preliminary denoised signals (i.e., the first initial denoised signal and the second initial denoised signal). The adaptive dead-time heuristic gating, based on the prior knowledge of physical consistency between multiple magnetic resonance acquisitions, performs amplitude comparison and phase conflict detection on the two initial denoised signals to identify transient high-amplitude electromagnetic interference in the signal. Then, the fusion weight of the corresponding signal points is adaptively adjusted, that is, a larger fusion weight is given to the low-energy signal points, and a smaller fusion weight is given to the high-energy signal points (i.e., transient high-amplitude electromagnetic interference), so that the high-energy signal points receive minimal weighting. In this way, high-energy noise is eliminated, thereby achieving nonlinear elimination of non-stationary transient electromagnetic interference.

[0032] Optionally, the initial denoising process is as shown in step S3 below.

[0033] S3. The first magnetic resonance signal and the second magnetic resonance signal are respectively input into the twin residual network in the signal denoising network for initial denoising processing to obtain the first initial denoised signal and the second initial denoised signal. In specific applications, the core of the twin residual network is a deep fully convolutional residual structure. It uses the deep fully convolutional residual structure and the trained weights to perform preliminary feature extraction and background noise suppression on the two signals (i.e., the first magnetic resonance signal and the second magnetic resonance signal) to output the feature signals after preliminary denoising (i.e., the first initial denoised signal and the second initial denoised signal).

[0034] For example, the twin residual network may include, but is not limited to, a first deep fully convolutional residual subnetwork and a second deep fully convolutional residual subnetwork, and the first deep fully convolutional residual subnetwork and the second deep fully convolutional residual subnetwork share the same network structure and weights; thus, when performing preliminary denoising of the two magnetic resonance signals, the EMI prior signals (i.e., the EMI signals received by the electromagnetic interference induction coil) corresponding to the first magnetic resonance signal and the second magnetic resonance signal are first obtained; then, the first magnetic resonance signal and the EMI prior signal corresponding to the first magnetic resonance signal are input into the first deep fully convolutional residual subnetwork for initial denoising processing, and the second magnetic resonance signal and the EMI prior signal corresponding to the second magnetic resonance signal are input into the second deep fully convolutional residual subnetwork for initial denoising processing (see...). Figure 4 As shown in the figure, the first initial denoised signal and the second initial denoised signal can be obtained respectively.

[0035] Furthermore, in order to avoid information loss that may result from using the traditional encoder-decoder (UNet) structure, this embodiment improves the traditional Deep-DSP network structure by using an improved deep fully convolutional residual subnetwork for initial denoising. Specifically, since the first and second deep fully convolutional residual subnetworks have the same structure, the following description uses the first deep fully convolutional residual subnetwork as an example to illustrate the network structure.

[0036] For specific applications, see Figure 5 As shown, the first deep fully convolutional residual subnetwork may include, but is not limited to, a head convolutional layer, a residual convolutional layer (i.e., stacked residuals), and a tail convolutional layer connected in sequence. The residual convolutional layer includes several residual convolutional blocks, and any residual convolutional block includes a first convolutional layer, a ReLU layer, and a second convolutional layer connected in sequence. At the same time, the input of any residual convolutional block is fused with the output of the second convolutional layer and then output to the next residual convolutional block. In this way, the network can learn the residual mapping between the input and the output, thereby solving the network degradation problem, alleviating gradient vanishing, and achieving the goal of accelerating training.

[0037] Furthermore, for example, both the head and tail convolutional layers use 9×1 convolutional kernels for convolution operations, and the first and second convolutional layers in any residual convolutional block use 3×1 convolutional kernels for convolution operations; thus, when the aforementioned twin residual network performs preliminary denoising, it keeps the feature map size unchanged during transmission, thereby preserving the high-frequency details of the magnetic resonance signal to the greatest extent.

[0038] After the initial denoising of the two magnetic resonance signals is completed, the aforementioned adaptive dead-zone heuristic gating and the use of physical consistency priors can be used to calculate the fusion weight of each signal point in the two initial denoised signals. Based on this, the intelligent fusion of the two initial denoised signals can be realized, thereby removing transient strong interference that cannot be handled by existing technologies. The calculation process of the fusion weight based on the adaptive dead-zone heuristic gating is shown in step S4 below.

[0039] S4. Input the first initial denoised signal and the second initial denoised signal into the adaptive dead-zone heuristic gating to calculate the local energy difference and detect the phase conflict. Based on the local energy difference and the phase conflict detection results, determine the fusion weight of each signal point in the first initial denoised signal and the second initial denoised signal. Wherein, when the local energy difference is greater than or equal to the energy threshold, or when the phase conflict detection result indicates the presence of a phase conflict, the fusion weight of the signal point with higher energy in the two initial denoised signals is less than the fusion weight of the signal point with lower energy.

[0040] For specific implementation, see Figure 4 As shown, two initial denoised signals are input to an adaptive dead-time heuristic gating system to remove transient strong interference noise. The adaptive dead-time heuristic gating system identifies transient high-amplitude electromagnetic interference in the signal by performing amplitude comparison (i.e., local energy difference calculation) and phase conflict detection, and then dynamically assigns fusion weights to the transient high-amplitude electromagnetic interference in the signal. Specifically, the local energy difference and phase conflict detection process can be, but is not limited to, the steps S41 to S43 below.

[0041] S41. The first initial denoised signal and the second initial denoised signal are subjected to modulo-taking and smoothing processing to obtain the first envelope energy curve and the second envelope energy curve, respectively. In this embodiment, the nuclear magnetic resonance signal is often in complex form, containing real and imaginary parts. Therefore, by taking the modulus of the complex signal, the amplitude information of the signal can be obtained. This step helps to remove phase information and retain only the intensity change of the signal. The smoothing processing (such as sliding window averaging) reduces noise and unnecessary details in the signal while retaining the main features of the signal. In this way, the signal quality can be improved, making the subsequent analysis more accurate.

[0042] Thus, after performing modulus extraction and smoothing on the two initial denoised signals, their respective envelope energy curves can be obtained. At this point, the local energy difference of signal points at the same positions in the two initial denoised signals can be calculated based on the two envelope energy curves, as shown in step S42 below.

[0043] S42. Based on the first envelope energy curve and the second envelope energy curve, calculate the local energy difference between signal points at the same positions in the first initial denoised signal and the second initial denoised signal; in this embodiment, the length of the two initial denoised signals is the same as the length of the two envelope energy curves. Therefore, the local energy difference between signal points at the same positions on the two envelope energy curves is the local energy difference between signal points at the same positions in the two initial denoised signals.

[0044] Based on this, the specific calculation process of local energy difference is as follows: First, calculate the energy difference between the i-th signal point in the second envelope energy curve and the i-th signal point in the first envelope energy curve, and calculate the sum of the energies of the i-th signal point in the second envelope energy curve and the i-th signal point in the first envelope energy curve (i is a positive integer); then, take the ratio between the energy difference and the energy sum as the local energy difference between the i-th signal point in the first initial denoised signal and the second initial denoised signal; finally, increment i by 1 and recalculate the energy difference between the i-th signal point in the second envelope energy curve and the i-th signal point in the first envelope energy curve, until i equals n, then the local energy difference between the signal points at the same positions in the first initial denoised signal and the second initial denoised signal can be obtained, where the initial value of i is 1, and n is the signal length of the first initial denoised signal.

[0045] Optionally, the local energy difference of signal points at the same location in the two initial denoised signals can be expressed by the formula: ; In the formula, This indicates a modulus extraction and smoothing operation. These represent the first initial denoised signal and the second initial denoised signal, respectively. Based on this, and These are the first envelope energy curve and the second envelope energy curve, respectively. Substituting the energy (i.e., amplitude) of the signal points at corresponding positions on the two envelope energy curves into the aforementioned formula yields the local energy differences of the signal points at the same positions in the two initial denoised signals. In this embodiment, to distinguish it from the initial denoised signal output by the twin residual network during subsequent training, in practical applications, [the following is used]. This represents the first and second initial denoised signals used in actual applications.

[0046] Thus, based on the aforementioned method, after calculating the local energy difference between signal points at the same locations in the two initial denoised signals, phase conflict detection can be performed, as shown in step S43 below.

[0047] S43. Calculate the complex dot product of signal points at the same positions in the first and second initial denoised signals, and obtain the phase conflict detection result of signal points at the same positions in the first and second initial denoised signals based on the complex dot product of signal points at the same positions. In this embodiment, for any two signal points at the same position in the first and second initial denoised signals, determine whether the complex dot product of the two signal points is negative. If it is, it is determined that there is a phase conflict between the two signal points at the same position (i.e., the phase conflict detection result is that there is a phase conflict), which is determined to be a phase conflict area. If this area is directly averaged, it will cause signal cancellation. Therefore, the area composed of signal points with phase conflicts is marked as a high-risk area, and the fusion weight needs to be dynamically adjusted. Of course, if the complex dot product of the two signal points at the same position is not negative, it is a normal signal point, which means that there is no phase conflict.

[0048] Thus, based on the aforementioned steps S41 to S43, after calculating the local energy difference and detecting the phase conflict of each signal point in the two initial denoised signals, the fusion weight can be calculated, as shown in steps S44 to S46 below.

[0049] S44. For any two signal points at the same position in the first initial denoised signal and the second initial denoised signal, determine whether the local energy difference between the two signal points is greater than or equal to the energy threshold, or whether the phase conflict detection result is that a phase conflict exists; in this embodiment, the energy threshold can be set to 0.15, but is not limited to it. Of course, it can be set according to actual use. This embodiment is not limited to the above example.

[0050] When the local energy difference between two signal points is greater than or equal to the energy threshold, or when there is a phase conflict, it can be determined that the one with higher energy has been contaminated by transient EMI (because the energy of the real MRI signal is stable, and the sudden increase in energy must be due to additive noise). Based on this, different fusion weights need to be assigned to the two signal points. The calculation process is shown in steps S45 and S46 below.

[0051] S45. If so, calculate the first initial weight based on the local energy difference between the two signal points, and use the difference between 1 and the first initial weight as the second initial weight. In specific implementation, for example, but not limited to, the first initial weight can be calculated based on the local energy difference between the two signal points and using the Sigmoid function. Specifically, the local energy difference between the two signal points is used as the independent variable of the Sigmoid function and substituted into the function to calculate the first initial weight. Thus, after calculating the two initial weights based on the aforementioned step S45, the weights can be allocated according to the energy magnitude of the two signal points, as shown in step S46 below.

[0052] S46. The largest weight among the first initial weight and the second initial weight is used as the fusion weight of the signal point with the lowest energy value among the two signal points, and the smallest weight among the first initial weight and the second initial weight is used as the fusion weight of the signal point with the highest energy value among the two signal points. After polling all signal points in the first initial denoised signal and the second initial denoised signal, the fusion weight of each signal point in the first initial denoised signal and the second initial denoised signal is obtained.

[0053] In practical implementation, as explained above, high-energy signal points are subject to transient EMI contamination. Therefore, they need to be assigned a smaller weight, even a minimal weighting, to weaken this part of the signal and thus eliminate high-energy noise. Similarly, low-energy signal points are normal signal points, and they need to be assigned a larger weight to preserve the normal signal. The logic curve for adaptive dead-time heuristic gating can be found, but is not limited to, in [reference needed]. Figure 6 As shown.

[0054] Of course, if the local energy difference between the two signal points is less than the energy threshold and the phase conflict detection result is that there is no phase conflict, then it is determined that the difference between the two signal points is only caused by thermal noise and the signals have physical consistency. Therefore, the same fusion weight (i.e., 0.5) can be set for the two signal points to perform a weighted average operation to improve the signal-to-noise ratio.

[0055] Based on the aforementioned steps S44 to S46, after calculating the fusion weights of signal points at the same positions in the two initial denoised signals, signal fusion can be performed based on these fusion weights, as shown in step S5 below.

[0056] S5. Based on the fusion weights of each signal point in the first and second initial denoised signals, the first and second initial denoised signals are fused to obtain a denoised magnetic resonance signal. In this embodiment, for two signal points at any identical position in the two initial denoised signals, the two signal points are multiplied by their respective fusion weights and then summed to obtain the denoised signal points. In this way, after iterating through all the signal points in the two initial denoised signals in the aforementioned manner, a denoised magnetic resonance signal can be formed. Finally, the image can be reconstructed based on the denoised magnetic resonance signal.

[0057] Thus, based on the aforementioned integrated twin residual network and adaptive dead-zone heuristic gating signal denoising network, nonlinear elimination of non-stationary transient electromagnetic interference can be achieved.

[0058] Optionally, for example, but not limited to, using magnetic resonance training data as input and denoised magnetic resonance signal output, the signal denoising network can be trained. One set of magnetic resonance training data includes a first sample magnetic resonance signal and a second sample magnetic resonance signal (in this embodiment, the first sample magnetic resonance signal is constructed by superimposing a calibration signal obtained from empty sampling with a magnetic resonance signal free from EMI interference; the construction process of the second sample magnetic resonance signal is similar and will not be elaborated here). Furthermore, to simultaneously optimize the denoising capability of a single acquisition and the effect of dual fusion, this embodiment constructs a hybrid loss function, i.e., the loss function of the signal denoising network is: ; In the formula, For loss function, The first sample magnetic resonance signal is the initial denoised signal output after being input into the twin residual network in the signal denoising network. The second sample magnetic resonance signal is the initial denoised signal output after being input into the twin residual network in the signal denoising network. This is the denoised magnetic resonance signal corresponding to the magnetic resonance training data. The label data (i.e., the real signal, that is, the magnetic resonance signal without EMI interference) is the corresponding magnetic resonance training data. It is an L1 norm.

[0059] As can be seen from the aforementioned loss function calculation formula, in order to preserve signal details, the L1 norm is used to constrain the consistency between the network output and the label. Simultaneously, constraints are also applied. and Each branch approximates the true label (i.e., single-branch constraint), thus ensuring that the network has basic denoising capabilities, and the final output after constraint-gated fusion... By approximating the real labels, and thus implementing fusion constraints, the network is forced to learn how to distinguish between transient noise and real signals.

[0060] Additionally, this embodiment shows a comparison chart of the denoised signal obtained using the method provided in this embodiment and the denoised signal obtained using a conventional method; see also Figure 7 As shown; where, Figure 7 Figure (a) shows the denoised signal obtained after denoising using the method provided in this embodiment, while Figure 7 Figure (b) shows the denoised signal obtained by averaging multiple data acquisitions; through Figure 7 As can be seen from the comparison, this embodiment can remove transient strong EMI noise.

[0061] This embodiment also provides a comparative illustration of the reconstructed final image, such as... Figure 8 As shown; where, Figure 8 Figure (a) shows the reconstructed image based on the Deep-DSP method, which involves directly averaging multiple data acquisitions for signal denoising. Figure 8 Figure (b) shows the image reconstructed after signal denoising using the method provided in this embodiment. By comparing Figure (8), it can be seen that the image reconstructed by the denoised signal in this embodiment is clearer and significantly improves the purity of the image.

[0062] Therefore, the magnetic resonance signal denoising method based on physical prior gating mechanism described in detail in steps S1 to S6 above has the following beneficial effects: (1) Breaking through the limitations of traditional linear averaging: Compared with the traditional multiple acquisition averaging which "spreads out" noise and causes image artifacts, this invention uses a "comparison-selection" mechanism, namely an adaptive dead zone gating mechanism based on amplitude and phase perception, to achieve nonlinear elimination of non-stationary transient electromagnetic interference, thereby significantly improving the purity of the image.

[0063] (2) Strong physical interpretability: The fusion mechanism is not a black box operation, but is based on the physical principle of "conservation of real signal energy and superposition of noise energy", which ensures that the real MRI signal will not be damaged while removing noise.

[0064] like Figure 9 As shown, the second aspect of this embodiment provides a hardware device for implementing the magnetic resonance signal denoising method based on a physical prior gating mechanism described in the first aspect of the embodiment, comprising: The acquisition unit is used to acquire a first magnetic resonance signal and a second magnetic resonance signal obtained by at least two repeated acquisitions at the same k-space location in an unshielded environment, wherein both the first magnetic resonance signal and the second magnetic resonance signal include signals from the MRI coil and the EMI coil.

[0065] The network building unit is used to construct a signal denoising network, which includes a twin residual network and an adaptive dead-time heuristic gating.

[0066] The denoising unit is used to input the first magnetic resonance signal and the second magnetic resonance signal into the twin residual network in the signal denoising network for initial denoising processing, so as to obtain the first initial denoised signal and the second initial denoised signal.

[0067] The denoising unit is used to input the first initial denoised signal and the second initial denoised signal into the adaptive dead-zone heuristic gating system to calculate the local energy difference and detect the phase conflict. Based on the local energy difference and the phase conflict detection results, the fusion weight of each signal point in the first initial denoised signal and the second initial denoised signal is determined. When the local energy difference is greater than or equal to the energy threshold, or when the phase conflict detection result indicates the presence of a phase conflict, the fusion weight of the signal point with higher energy in the two initial denoised signals is less than the fusion weight of the signal point with lower energy.

[0068] The denoising unit is also used to perform fusion processing on the first initial denoised signal and the second initial denoised signal according to the fusion weight of each signal point in the first initial denoised signal and the second initial denoised signal, so as to obtain the denoised magnetic resonance signal after the fusion processing.

[0069] The working process, working details and technical effects of the device provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0070] like Figure 10 As shown, the third aspect of this embodiment provides another magnetic resonance signal denoising device based on a physical prior gating mechanism. Taking the device as an electronic device as an example, it includes: a memory, a processor, and a transceiver that are connected in sequence. The memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program and execute the magnetic resonance signal denoising method based on a physical prior gating mechanism as described in the first aspect of the embodiment.

[0071] For specific examples, the memory may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, first-in-first-out (FIFO) memory, and / or first-in-last-out (FILO) memory, etc.; specifically, the processor may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor may be implemented using at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), PLA (Programmable Logic Array). The processor may also include a main processor and a coprocessor. The main processor, also known as the CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state.

[0072] In some embodiments, the processor may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. For example, the processor may not be limited to microprocessors of the STM32F105 series, reduced instruction set computer (RISC) microprocessors, x86 architecture processors, or processors with integrated neural network processing units (NPUs). The transceiver may be, but is not limited to, a Wi-Fi transceiver, a Bluetooth transceiver, a General Packet Radio Service (GPRS) transceiver, a ZigBee (a low-power LAN protocol based on the IEEE 802.15.4 standard) transceiver, a 3G transceiver, a 4G transceiver, and / or a 5G transceiver. Furthermore, the device may also include, but is not limited to, a power module, a display screen, and other necessary components.

[0073] The working process, working details and technical effects of the electronic device provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0074] The fourth aspect of this embodiment provides a storage medium that stores instructions for a magnetic resonance signal denoising method based on a physical prior gating mechanism as described in the first aspect of the embodiment. That is, the storage medium stores instructions that, when executed on a computer, perform the magnetic resonance signal denoising method based on a physical prior gating mechanism as described in the first aspect of the embodiment.

[0075] The storage medium refers to a carrier for storing data, which may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or memory sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0076] The working process, working details and technical effects of the storage medium provided in this embodiment can be found in the first aspect of the embodiment, and will not be repeated here.

[0077] The fifth aspect of this embodiment provides a computer program product containing instructions that, when executed on a computer, cause the computer to perform the magnetic resonance signal denoising method based on a physical prior gating mechanism as described in the first aspect of the embodiment, wherein the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0078] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for denoising magnetic resonance signals based on a physical prior gating mechanism, characterized in that, include: Acquire a first magnetic resonance signal and a second magnetic resonance signal obtained by at least two repeated acquisitions at the same k-space location in an unshielded environment, wherein both the first magnetic resonance signal and the second magnetic resonance signal include signals from the MRI coil and the EMI coil; A signal denoising network is constructed, which includes a twin residual network and an adaptive dead-time heuristic gating. The first magnetic resonance signal and the second magnetic resonance signal are respectively input into the twin residual network in the signal denoising network for initial denoising processing to obtain the first initial denoised signal and the second initial denoised signal; The first initial denoised signal and the second initial denoised signal are input to the adaptive dead-zone heuristic gating system for local energy difference calculation and phase conflict detection. Based on the local energy difference and phase conflict detection results, the fusion weight of each signal point in the first initial denoised signal and the second initial denoised signal is determined. When the local energy difference is greater than or equal to the energy threshold, or when the phase conflict detection result indicates the presence of a phase conflict, the fusion weight of the signal point with higher energy in the two initial denoised signals is less than the fusion weight of the signal point with lower energy. Based on the fusion weights of each signal point in the first and second initial denoised signals, the first and second initial denoised signals are fused to obtain a denoised magnetic resonance signal.

2. The method according to claim 1, characterized in that, The twin residual network includes: a first deep fully convolutional residual subnetwork and a second deep fully convolutional residual subnetwork, and the first deep fully convolutional residual subnetwork and the second deep fully convolutional residual subnetwork share the same network structure and weights; The first and second magnetic resonance signals are respectively input into the twin residual network in the signal denoising network for initial denoising processing, including: Obtain the EMI prior signals corresponding to the first and second magnetic resonance signals, respectively. The first magnetic resonance signal and the corresponding EMI prior signal are input into a first depth fully convolutional residual subnetwork for initial denoising processing, and the second magnetic resonance signal and the corresponding EMI prior signal are input into a second depth fully convolutional residual subnetwork for initial denoising processing, so as to obtain the first initial denoised signal and the second initial denoised signal respectively.

3. The method according to claim 2, characterized in that, The first deep fully convolutional residual subnetwork includes a head convolutional layer, a residual convolutional layer and a tail convolutional layer connected in sequence, wherein the residual convolutional layer includes a plurality of residual convolutional blocks, and any residual convolutional block includes a first convolutional layer, a ReLU layer and a second convolutional layer connected in sequence.

4. The method according to claim 1, characterized in that, The first and second initial denoised signals are input into the adaptive dead-time heuristic gating system for local energy difference calculation and phase conflict detection, including: The first initial denoised signal and the second initial denoised signal are subjected to modulus taking and smoothing processing to obtain the first envelope energy curve and the second envelope energy curve, respectively. Based on the first envelope energy curve and the second envelope energy curve, the local energy difference between signal points at the same position in the first initial denoised signal and the second initial denoised signal is calculated. The complex dot product of signal points at the same position in the first and second initial denoised signals is calculated, and the phase conflict detection result of signal points at the same position in the first and second initial denoised signals is obtained based on the complex dot product of signal points at the same position.

5. The method according to claim 4, characterized in that, Based on the first envelope energy curve and the second envelope energy curve, the local energy difference between signal points at the same location in the first and second initial denoised signals is calculated, including: Calculate the energy difference between the i-th signal point in the second envelope energy curve and the i-th signal point in the first envelope energy curve, and calculate the sum of the energies of the i-th signal point in the second envelope energy curve and the i-th signal point in the first envelope energy curve, where i is a positive integer; The ratio between the energy difference and the energy sum is used as the local energy difference between the i-th signal point in the first and second initial denoised signals. Increment i by 1 and recalculate the energy difference between the i-th signal point in the second envelope energy curve and the i-th signal point in the first envelope energy curve until i equals n. This yields the local energy difference between signal points at the same positions in the first and second initial denoised signals. Here, the initial value of i is 1, and n is the signal length of the first initial denoised signal. Accordingly, based on the complex dot product of signal points at the same locations, the phase conflict detection results of signal points at the same locations in the first and second initial denoised signals are obtained, including: For any two signal points at the same position in the first and second initial denoised signals, determine whether the complex dot product of the two signal points is negative; If so, it is determined that there is a phase conflict between the two signal points at any of the same positions.

6. The method according to claim 4, characterized in that, Based on the results of local energy difference and phase conflict detection, the fusion weights of each signal point in the first and second initial denoised signals are determined, including: For any two signal points at the same position in the first initial denoised signal and the second initial denoised signal, determine whether the local energy difference between the two signal points is greater than or equal to the energy threshold, or whether the phase conflict detection result indicates the presence of a phase conflict. If so, the first initial weight is calculated based on the local energy difference between the two signal points, and the difference between 1 and the first initial weight is used as the second initial weight; The largest of the first and second initial weights is used as the fusion weight of the signal point with the lowest energy value among the two signal points, and the smallest of the first and second initial weights is used as the fusion weight of the signal point with the highest energy value among the two signal points. After all signal points in the first and second initial denoised signals have been polled, the fusion weight of each signal point in the first and second initial denoised signals is obtained.

7. The method according to claim 6, characterized in that, If the local energy difference between the two signal points is less than the energy threshold, and the phase conflict detection result is that there is no phase conflict, then the method further includes: Set the same fusion weight for both signal points; The first initial weight is calculated based on the local energy difference between the two signal points, including: The first initial weight is calculated based on the local energy difference between the two signal points and using the Sigmoid function.

8. The method according to claim 1, characterized in that, Using magnetic resonance training data as input and denoised magnetic resonance signals as output, a signal denoising network is trained. Each set of magnetic resonance training data includes a first sample magnetic resonance signal and a second sample magnetic resonance signal. The loss function of the signal denoising network is: ； In the formula, For loss function, The first sample magnetic resonance signal is the initial denoised signal output after being input into the twin residual network in the signal denoising network. The second sample magnetic resonance signal is the initial denoised signal output after being input into the twin residual network in the signal denoising network. This is the denoised magnetic resonance signal corresponding to the magnetic resonance training data. The label data corresponding to the magnetic resonance training data. It is an L1 norm.

9. A magnetic resonance signal denoising device based on a physical prior gating mechanism, characterized in that, include: The acquisition unit is used to acquire a first magnetic resonance signal and a second magnetic resonance signal obtained by at least two repeated acquisitions at the same k-space location in an unshielded environment, wherein the first magnetic resonance signal and the second magnetic resonance signal both include signals in the MRI coil and the EMI coil. The network building unit is used to construct a signal denoising network, which includes a twin residual network and an adaptive dead-time heuristic gating. The denoising unit is used to input the first magnetic resonance signal and the second magnetic resonance signal into the twin residual network in the signal denoising network for initial denoising processing to obtain the first initial denoised signal and the second initial denoised signal. The denoising unit is used to input the first initial denoised signal and the second initial denoised signal into the adaptive dead-zone heuristic gating to calculate the local energy difference and detect the phase conflict. Based on the local energy difference and the phase conflict detection results, the fusion weight of each signal point in the first initial denoised signal and the second initial denoised signal is determined. When the local energy difference is greater than or equal to the energy threshold, or when the phase conflict detection result indicates the presence of a phase conflict, the fusion weight of the signal point with higher energy in the two initial denoised signals is less than the fusion weight of the signal point with lower energy. The denoising unit is also used to perform fusion processing on the first initial denoised signal and the second initial denoised signal according to the fusion weight of each signal point in the first initial denoised signal and the second initial denoised signal, so as to obtain the denoised magnetic resonance signal after the fusion processing.

10. An electronic device, characterized in that, include: A memory, a processor, and a transceiver are sequentially connected in communication, wherein the memory is used to store computer programs, the transceiver is used to send and receive messages, and the processor is used to read the computer programs and execute the magnetic resonance signal denoising method based on physical prior gating mechanism as described in any one of claims 1 to 8.

Images