Magnetic resonance random noise intelligent suppression method and system based on dynamic weighting

Abstract

The invention belongs to the technical field of nuclear magnetic resonance sounding, and relates to a magnetic resonance random noise intelligent suppression method and system based on dynamic weighting, and the method comprises the steps: employing a forward conversion path to enable a Y-domain noisy signal to generate a pseudo-X-domain pure signal, or enable a pseudo-Y-domain noisy signal to generate a reconstructed X-domain pure signal; generating the pure signal of the pseudo X domain into a reconstructed noisy signal of the Y domain by adopting a reverse conversion path, or generating the pure signal of the X domain into a noisy signal of the pseudo Y domain; performing forward feature authenticity evaluation calculation on the pure signal of the pseudo X domain; performing reverse feature authenticity evaluation calculation on the noisy signal of the pseudo Y domain; alternately optimizing the two conversion paths; and de-noising processing is carried out on the actually measured noisy data by adopting the optimized forward conversion path. The suppression effect on random interference in actual field detection signals with more complex noise conditions and lower signal-to-noise ratios is enhanced.

Description

Technical Field

[0001] This application belongs to the technical field of Magnetic Resonance Sounding (MRS), and specifically relates to a method and system for intelligently suppressing magnetic resonance random noise based on dynamic weighting. Background Art

[0002] With the increasingly severe global fresh water crisis and frequent occurrence of water-source geological disasters, the development of efficient groundwater detection technologies has become an urgent need to ensure water resource security and prevent and control geological disasters. As a non-invasive geophysical method for directly detecting groundwater, ground nuclear magnetic resonance has significant advantages such as fast speed, high efficiency, and rich information, and has received extensive attention and application in recent years.

[0003] However, the magnetic resonance signal during actual detection is very weak, usually at the nanovolt level, and is extremely vulnerable to various electromagnetic noise interferences in the surrounding environment, especially a large amount of random noise. For random noise, due to its randomness in statistical characteristics, there is no unified law, and its frequency distribution range is wide, often overlapping with the frequency band of the magnetic resonance signal, making it difficult to effectively extract the signal and being unfavorable for the subsequent inversion and interpretation of hydrogeological parameters. Therefore, effectively suppressing random noise is of great significance for improving the efficiency of magnetic resonance groundwater detection.

[0004] Patent CN116148935A discloses "a method for suppressing magnetic resonance random noise based on an adaptive autoencoder". This method constructs a noise suppression model to suppress the random noise of magnetic resonance groundwater detection through the noise suppression model, including: adding random noise to multiple groups of ideal magnetic resonance signals to obtain a data set; building a model, the model includes an encoder and a decoder, and initializing network parameters; training the model using the training set S, extracting features of the data through the encoder to obtain a latent variable z, and reconstructing an effective signal from the latent variable z through the decoder; enabling the autoencoder to establish a probability distribution model of the training samples to learn the distribution law of the signals, determining a loss function based on the deviation between the input signal and the reconstructed signal, introducing a parameter estimation error to constrain the loss function, and updating the network model parameters until the trend of the loss function is stable to obtain an adaptive autoencoder denoising model; using the test set T to test the denoising effect of the model. It solves the problem of limited denoising effect and improves the denoising efficiency. However, this method depends on the quality of the training data and is difficult to have large generalization. When dealing with field detection signals with more complex actual random noise interference and lower signal-to-noise ratio, the denoising effect of the model is poor. Summary of the Invention

[0005] The embodiments of this application provide a method for intelligently suppressing magnetic resonance random noise based on dynamic weighting, which solves the problem of poor denoising effect when dealing with actual field detection signals with more complex random noise interference and lower signal-to-noise ratio.

[0006] Another aspect of the embodiments of the present application further provides a magnetic resonance random noise intelligent suppression system based on dynamic weighting.

[0007] The technical solution of the embodiments of the present application includes: A magnetic resonance random noise intelligent suppression method based on dynamic weighting, the method includes: Using a forward conversion path to generate a pure signal in the pseudo-X domain from the noisy signal in the Y domain, or generating a pure signal in the reconstructed X domain from the noisy signal in the pseudo-Y domain; Using a reverse conversion path to generate a noisy signal in the reconstructed Y domain from the pure signal in the pseudo-X domain, or generating a noisy signal in the pseudo-Y domain from the pure signal in the X domain; Performing a forward feature authenticity evaluation calculation on the pure signal in the pseudo-X domain; Performing a reverse feature authenticity evaluation calculation on the noisy signal in the pseudo-Y domain; Alternately optimizing the two conversion paths until the signal generated by the conversion path approaches the signal in the real target domain, and when the feature difference between the reconstructed signal and the original input signal reaches stable convergence, the optimization ends; Using the optimized forward conversion path to perform denoising processing on the measured noisy data.

[0008] Further, both the forward conversion path and the reverse conversion path perform the following processing on the input signal: Extracting the feature detail information of the signal through multiple downsampling convolution operations; Using multiple residual transformation operations to gradually map the feature detail information to the target domain; Parallelly performing the first weighting process and the second weighting process to dynamically weight the mapped feature detail data; Restoring to the input signal dimension through a gradual upsampling convolution operation.

[0009] Further, the first weighting process includes: performing a max pooling operation and an average pooling operation on each channel of the input feature detail information respectively to obtain max pooling channel feature data and average pooling channel feature data; adding the max pooling channel feature data and the average pooling channel feature data after parallel processing through a shared weight transformation, and then using a Sigmoid activation function to obtain a channel weight coefficient; multiplying the weight channel coefficient element-wise with the input feature detail information to obtain the output after the first weighting process.

[0010] Further, the second weighting process includes: Perform max-pooling operations and average-pooling operations on the input feature detail information along the channel dimension respectively to obtain max-pooling spatial feature data and average-pooling spatial feature data; concatenate the max-pooling spatial feature data and the average-pooling spatial feature data along the channel dimension and then fuse them through a convolution operation, and then use the Sigmoid activation function to obtain spatial weight coefficients; multiply the spatial weight coefficients element-wise with the input feature detail information to obtain the output after the second weighted processing.

[0011] Further, the alternating optimization of the forward conversion path and the backward conversion path includes: Input the noisy signal in the Y domain, process it through the forward conversion path to obtain the pure signal in the pseudo-X domain, perform forward feature authenticity evaluation calculation on the pure signal in the pseudo-X domain and the pure signal in the real X domain, and update the parameters of the forward conversion path according to the evaluation result, so that the noisy signal in the Y domain can obtain a pure signal closer to the real X domain after being processed through the forward conversion path; Input the pure signal in the X domain, generate a noisy signal in the pseudo-Y domain through the backward conversion path, perform backward feature authenticity evaluation calculation on the noisy signal in the pseudo-Y domain and the original noisy signal in the Y domain, and update the parameters of the backward conversion path according to the evaluation result; Input the pure signal in the pseudo-X domain obtained by processing through the forward conversion path into the backward conversion path to obtain the reconstructed noisy signal in the Y domain, compare the reconstructed noisy signal in the Y domain with the original noisy signal in the Y domain, and update the parameters of the two conversion paths according to the feature difference; input the noisy signal in the pseudo-Y domain obtained by processing through the backward conversion path into the forward conversion path to obtain the reconstructed pure signal in the X domain, and compare it with the real pure signal in the X domain, and update the parameters of the two conversion paths according to the feature difference.

[0012] Further, both the forward feature authenticity evaluation calculation and the backward feature authenticity evaluation calculation include: gradually extract feature information through multiple downsampling convolution operations, perform feature normalization processing and non-linear activation processing after each convolution operation, obtain a probability matrix through the last convolution operation, and take the mean value of the probability matrix as the feature authenticity evaluation result.

[0013] Further, the calculation of the feature difference includes: Calculate the norm of the difference between each real pure signal in the X domain and the corresponding reconstructed pure signal in the X domain, and take the average of all norms to obtain the first mean value; Calculate the norm of the difference between each original noisy signal in the Y domain and the corresponding reconstructed noisy signal in the Y domain, and take the average of all norms to obtain the second mean value; Sum the first mean value and the second mean value to obtain the feature difference value.

[0014] Another aspect of the embodiments of the present application provides a dynamic weighted-based intelligent magnetic resonance random noise suppression system, which includes: A forward conversion module, configured to generate a pure signal in the pseudo-X domain from a noisy signal in the Y domain, or generate a pure signal in the reconstructed X domain from a noisy signal in the pseudo-Y domain; A reverse conversion module, configured to generate a noisy signal in the reconstructed Y domain from the pure signal in the pseudo-X domain, or generate a noisy signal in the pseudo-Y domain from the pure signal in the X domain; A forward evaluation module, configured to perform forward feature authenticity evaluation calculation on the pure signal in the pseudo-X domain; A reverse evaluation module, configured to perform reverse feature authenticity evaluation calculation on the noisy signal in the pseudo-Y domain; An optimization module, alternately optimizing the two conversion paths until the signals generated by the conversion paths approximate the signals in the real target domain, and when the feature difference between the reconstructed signal and the original input signal reaches stable convergence, the optimization ends; The optimized forward conversion module performs noise reduction processing on the measured noisy data.

[0015] Compared with the prior art, the present application has at least the following beneficial effects: The present application can dynamically focus on key features and adversarial training for multi-scenario field detection data with complex noise characteristics, automatically adjust the feature extraction strategy for the actual noise environment, generate a more accurate noise model for blind denoising, and enhance the suppression effect on random interference in actual field detection signals with more complex noise conditions and lower signal-to-noise ratios. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic flow chart of a dynamic weighted-based intelligent magnetic resonance random noise suppression method provided by an embodiment of the present application; Figure 2 It is a structural block diagram of a dynamic weighted-based intelligent magnetic resonance random noise suppression system provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0018] See Figure 1 As shown, a dynamic weighted-based intelligent magnetic resonance random noise suppression method includes: Generate a clean signal in the pseudo-X domain from the noisy signal in the Y domain using a forward conversion path, or generate a clean signal in the reconstructed X domain from the noisy signal in the pseudo-Y domain; Here, the space where the noisy signal is located is regarded as the Y domain, the space where the clean signal is located is regarded as the X domain, and the noisy signal in the Y domain is converted to the X domain. Since it is not the clean signal in the real X domain, a clean signal in the pseudo-X domain is used to represent it here. The noisy signal in the pseudo-Y domain is not the original noisy signal in the Y domain, but the noisy signal converted to the Y domain through other means.

[0019] S102 Generate a reconstructed noisy signal in the Y domain from the clean signal in the pseudo-X domain using a reverse conversion path, or generate a noisy signal in the pseudo-Y domain from the clean signal in the X domain; It can be understood that the forward conversion path is used to implement the mapping conversion process from the Y domain to the X domain, and compare the generated clean signal in the pseudo-X domain with the clean signal in the real X domain to perform the feature authenticity evaluation calculation; the reverse conversion path is used to implement the reverse mapping conversion process from the X domain to the Y domain, and compare the generated reconstructed noisy signal in the Y domain with the original noisy signal in the Y domain to perform the feature authenticity evaluation calculation.

[0020] Perform a forward feature authenticity evaluation calculation on the clean signal in the pseudo-X domain; By generating a clean signal in the pseudo-X domain from the noisy signal in the Y domain using a forward conversion path, it is necessary to verify whether the clean signal in the pseudo-X domain is close to the clean signal in the real X domain. Here, the clean signal in the pseudo-X domain is evaluated, and the forward conversion path is adjusted according to the evaluation result.

[0021] Perform a reverse feature authenticity evaluation calculation on the noisy signal in the pseudo-Y domain; It can be understood that by generating a noisy signal in the pseudo-Y domain from the clean signal in the X domain using a reverse conversion path, it is necessary to verify whether the noisy signal in the pseudo-Y domain is close to the original noisy signal in the Y domain. Here, the noisy signal in the pseudo-Y domain is evaluated, and the noisy signal in the pseudo-Y domain is adjusted according to the evaluation result.

[0022] Alternately optimize the two conversion paths until the signal generated by the conversion path approaches the signal in the real target domain, and when the feature difference between the reconstructed signal and the original input signal reaches stable convergence, end the optimization; The target domain here refers to the domain obtained through the conversion path, which can be the Y domain or the X domain. For example, by using the forward conversion path, a noisy signal in the Y domain is used to generate a clean signal in the pseudo-X domain. The domain where the clean signal in the pseudo-X domain is located is the target domain. In this example, the signal in the true target domain refers to the original clean signal or the real clean signal in the X domain. Another example is that by using the reverse conversion path, a clean signal in the X domain is used to generate a noisy signal in the pseudo-Y domain. The domain where the noisy signal in the pseudo-Y domain is located is the target domain. In this example, the signal in the true target domain refers to the real noisy signal or the original noisy signal in the Y domain.

[0023] The reconstructed signals include the clean signal in the reconstructed X domain obtained by reconstructing the noisy signal in the pseudo-Y domain into a clean signal in the X domain, and the noisy signal in the reconstructed Y domain obtained by reconstructing the clean signal in the pseudo-X domain into a noisy signal in the Y domain; the original input signals refer to the original (real) noisy signal in the Y domain and the real (original) clean signal in the X domain.

[0024] Alternately optimizing the two conversion paths includes: taking the original (real) noisy signal in the Y domain as the input, using the forward conversion path to generate a clean signal in the pseudo-X domain from the noisy signal in the Y domain, and then using the reverse conversion path to generate a reconstructed noisy signal in the Y domain from the clean signal in the pseudo-X domain; Taking the real (original) clean signal in the X domain as the input to generate a noisy signal in the pseudo-Y domain from the clean signal in the X domain, and then using the forward conversion path to generate a reconstructed clean signal in the X domain from the noisy signal in the pseudo-Y domain.

[0025] And performing a forward feature authenticity evaluation calculation on the clean signal in the pseudo-X domain; Performing a reverse feature authenticity evaluation calculation on the noisy signal in the pseudo-Y domain.

[0026] Continuously repeat the above process until two conditions are met. One condition is that the signal generated by the conversion path approximates the signal in the true target domain, that is, it is impossible to distinguish the signal in the true target domain from the signal generated by the conversion path. For example, the signal in the true target domain is the real clean signal in the X domain, and the signal generated by the conversion path is the clean signal in the pseudo-X domain; the second condition is that when the feature difference between the reconstructed signal and the original input signal reaches stable convergence. For example, the reconstructed signal is the reconstructed noisy signal in the Y domain, and the original input signal is the original noisy signal in the Y domain.

[0027] Use the optimized forward conversion path to perform denoising processing on the measured noisy data.

[0028] In the embodiments of the present application, for multi-scenario field detection data with complex noise characteristics, unsupervised training is achieved through the forward conversion path and the reverse conversion path of dynamic weighted bidirectional mapping, as well as the forward feature authenticity evaluation calculation and the reverse feature authenticity evaluation calculation. It can dynamically focus on key features and adversarial training, automatically adjust the feature extraction strategy for the actual noise environment, generate a more accurate forward conversion path for blind denoising, and enhance the suppression effect on random interference in actual field detection signals with more complex noise conditions and lower signal-to-noise ratios.

[0029] In one embodiment, the noisy signal in the Y domain is a pure signal containing noise, and the dataset of the noisy signal in the Y domain and the dataset of the pure signal in the X domain , meaning, , the noisy signal in the Y domain is derived from the field magnetic resonance noisy signal actually collected by a nuclear magnetic resonance groundwater detector, and the pure signal in the X domain is derived from the magnetic resonance signal simulated according to the expression . The parameters are selected according to the actual exploration situation. For example, the initial amplitude is 200 nV, the relaxation time is 150 ms, the Larmor frequency is 2320 Hz, the initial phase is 0, t = (0:N - 1) / , the sampling rate is 50 kHz, and the data length is 10000.

[0030] In one embodiment, both the forward conversion path and the reverse conversion path perform the following processing on the input signal: Extract the feature detail information of the signal through multiple downsampling convolution operations; Use multiple residual transformation operations to gradually map the feature detail information to the target domain; Dynamically weight the mapped feature detail data by parallelly adopting the first weighted processing and the second weighted processing; Restore to the input signal dimension through gradually upsampling convolution operations.

[0031] Specifically: The input data first undergoes multiple downsampling operations. In one embodiment, convolution operations, data normalization processing, and non-linear activation processing can be sequentially performed to obtain feature detail information, which can be deep feature data. Then, through multiple residual transformation operations, such as convolution operations, non-linear activation, adding the result after convolution operations to the input, and then performing non-linear activation processing to gradually map the feature detail information to the target domain. Then, it undergoes the first weighting process and the second weighting process in parallel to dynamically weight the transformed feature detail data. Finally, through multiple upsampling operations, such as sequentially performing inverse convolution operations, data normalization, and non-linear activation processing to restore the input signal dimension or the original signal dimension, and output the data after the transformation process.

[0032] In one embodiment, the first weighting process includes: performing max pooling operations and average pooling operations on each channel of the input feature detail information respectively to obtain max pooling channel feature data and average pooling channel feature data; adding the max pooling channel feature data and the average pooling channel feature data after parallel processing through a shared weight transformation, and then using the Sigmoid activation function to obtain channel weight coefficients; multiplying the weight channel coefficients element-wise with the input feature detail information to obtain the output after the first weighting process.

[0033] The second weighting process includes: performing max pooling operations and average pooling operations on the input feature detail information along the channel dimension respectively to obtain max pooling spatial feature data and average pooling spatial feature data; concatenating the max pooling spatial feature data and the average pooling spatial feature data along the channel dimension and then fusing them through convolution operations, and then using the Sigmoid activation function to obtain spatial weight coefficients; multiplying the spatial weight coefficients element-wise with the input feature detail information to obtain the output after the second weighting process.

[0034] In one embodiment, the first weighting process and the second weighting process perform different dynamic weightings on the transformed feature data through two parallel independent paths, that is, according to the importance differences of each part in the transformed feature data, corresponding weight coefficients are assigned and then weighted and fused to obtain the output after focusing on the key features. Concatenate the two weighted outputs obtained along the channel dimension, and then compress the number of channels after concatenation back to the original number of channels through a 1x1 convolution operation. Thus, it can independently and comprehensively capture the key information of the input features in the spatial and channel dimensions, avoid the loss of sequential concatenated feature information, and has higher computational efficiency through two independent paths; and the gradients can be backpropagated in the opposite direction to the two independent paths simultaneously, avoiding the problem of gradient attenuation caused by deep concatenation.

[0035] In one embodiment, alternately optimizing the forward transformation path and the reverse transformation path includes: The noisy signal in the input Y domain is processed through the forward conversion path to obtain a clean signal in the pseudo-X domain. The clean signal in the pseudo-X domain is subjected to a forward feature authenticity evaluation calculation with the clean signal in the real X domain. According to the evaluation result, the parameters of the forward conversion path are updated so that the noisy signal in the Y domain can obtain a clean signal closer to the real X domain after being processed by the forward conversion path; The clean signal in the input X domain is processed through the reverse conversion path to generate a noisy signal in the pseudo-Y domain. The noisy signal in the pseudo-Y domain is subjected to a reverse feature authenticity evaluation calculation with the original noisy signal in the Y domain. According to the evaluation result, the parameters of the reverse conversion path are updated; The clean signal in the pseudo-X domain obtained by processing through the forward conversion path is input into the reverse conversion path to obtain a reconstructed noisy signal in the Y domain. The reconstructed noisy signal in the Y domain is compared with the original noisy signal in the Y domain, and the parameters of the two conversion paths are updated according to the feature difference; The noisy signal in the pseudo-Y domain obtained by processing through the reverse conversion path is input into the forward conversion path to obtain a reconstructed clean signal in the X domain, and it is compared with the real clean signal in the X domain, and the parameters of the two conversion paths are updated according to the feature difference.

[0036] In one embodiment, the feature difference is used to measure whether the parameters of the two conversion paths are optimized. The calculation of the feature difference includes: calculating the norm of the difference between each real clean signal in the X domain and the corresponding reconstructed clean signal in the X domain, and taking the average of all norms to obtain the first mean value; Calculating the norm of the difference between each original noisy signal in the Y domain and the corresponding reconstructed noisy signal in the Y domain, and taking the mean value of all norms to obtain the second mean value; The sum of the first mean value and the second mean value is used to obtain the feature difference value.

[0037] It can be understood that the feature difference value is an overall difference value, that is, the feature difference of the data in the forward conversion path and the reverse conversion path. When the overall feature difference value is the smallest and the signal generated by the conversion path approximates the signal in the real target domain and it is impossible to distinguish the original input signal from the generated signal, the obtained forward conversion path is the final required forward conversion path.

[0038] In one embodiment of the present application, refer to Figure 2The structural block diagram of an intelligent magnetic resonance random noise suppression system based on dynamic weighting using the noisy signal in the Y domain and the pure signal in the X domain is shown. An intelligent magnetic resonance random noise suppression system based on dynamic weighting can be correspondingly explained with an intelligent magnetic resonance random noise suppression method based on dynamic weighting. It includes: a forward conversion module for generating a pure signal in the pseudo-X domain from the noisy signal in the Y domain, or generating a reconstructed pure signal in the X domain from the noisy signal in the pseudo-Y domain; a reverse conversion module for generating a reconstructed noisy signal in the Y domain from the pure signal in the pseudo-X domain, or generating a noisy signal in the pseudo-Y domain from the pure signal in the X domain; a forward evaluation module for performing forward feature authenticity evaluation calculations on the pure signal in the pseudo-X domain; a reverse evaluation module for performing reverse feature authenticity evaluation calculations on the noisy signal in the pseudo-Y domain; an optimization module that alternately optimizes the two conversion paths until the signals generated by the conversion paths approximate the signals in the true target domain and the feature difference between the reconstructed signal and the original input signal reaches stable convergence, at which point the optimization ends; The optimized forward conversion module performs noise reduction processing on the measured noisy data.

[0039] In the embodiments of the present application, for multi-scenario field detection data with complex noise characteristics, an unsupervised training system is implemented through the forward conversion module and reverse conversion module of dynamic weighted bidirectional mapping, as well as the forward evaluation module and reverse evaluation module. It can dynamically focus on key features and adversarial training, automatically adjust the feature extraction strategy for the actual noise environment, generate a more accurate forward conversion module for blind denoising, and enhance the suppression effect on random interference in actual field detection signals with more complex noise conditions and lower signal-to-noise ratios.

[0040] The above are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present application shall be included within the protection scope of the present application.

Claims

1. A magnetic resonance random noise intelligent suppression method based on dynamic weighting, characterized in that: The method includes: A forward conversion path is used to generate a noisy signal in the Y domain into a clean signal in the pseudo X domain, or to generate a noisy signal in the pseudo Y domain into a reconstructed clean signal in the X domain; A reverse conversion path is used to generate a pseudo X-domain clean signal into a reconstructed Y-domain noisy signal, or to generate a pseudo Y-domain noisy signal from an X-domain clean signal; Perform positive feature authenticity evaluation calculation on the pure signal in the pseudo-X domain; Perform reverse feature authenticity evaluation calculation on the noisy signal in the pseudo-Y domain; Alternately optimize the two conversion paths until the signal generated by the conversion path is close to the signal in the real target domain, and the characteristic difference between the reconstructed signal and the original input signal reaches stable convergence, then end the optimization; The optimized forward conversion path is used to denoise the measured noisy data.

2. The method for intelligent suppression of magnetic resonance random noise based on dynamic weighting according to claim 1, characterized in that: Both the forward and reverse conversion paths perform the following processing on the input signal: After multiple downsampling convolution operations, the characteristic details of the signal are extracted; Multiple residual transformation operations are used to gradually map feature detail information to the target domain; Dynamically weighting the mapped feature detail data using the first weighting process and the second weighting process in parallel; The step-wise upsampling convolution operation restores the input signal dimension.

3. The magnetic resonance random noise intelligent suppression method based on dynamic weighting according to claim 2 is characterized in that: The first weighted processing includes: performing maximum pooling operations and average pooling operations on each channel of the input feature detail information to obtain maximum pooling channel feature data and average pooling channel feature data; adding the maximum pooling channel feature data and the average pooling channel feature data after parallel processing through a shared weight transformation, and then using the Sigmoid activation function to obtain the channel weight coefficient; multiplying the weight channel coefficient and the input feature detail information element by element to obtain the output after the first weighted processing.

4. The method for intelligent suppression of magnetic resonance random noise based on dynamic weighting according to claim 2, characterized in that: The second weighting process includes: The maximum pooling operation and the average pooling operation are respectively performed on the input feature detail information along the channel dimension to obtain the maximum pooling spatial feature data and the average pooling spatial feature data; the maximum pooling spatial feature data and the average pooling spatial feature data are concatenated along the channel dimension and fused through a convolution operation, and then the Sigmoid activation function is used to obtain the spatial weight coefficient; the spatial weight coefficient is multiplied element by element with the input feature detail information to obtain the output after the second weighted processing.

5. The method for intelligent suppression of magnetic resonance random noise based on dynamic weighting according to claim 1, characterized in that: The alternate optimization of the forward conversion path and the reverse conversion path includes: A noisy signal in the Y domain is input, and a pure signal in the pseudo X domain is obtained through a forward conversion path. The pure signal in the pseudo X domain and the pure signal in the real X domain are evaluated and calculated for forward feature authenticity, and the parameters of the forward conversion path are updated according to the evaluation result, so that the noisy signal in the Y domain can be processed by the forward conversion path to obtain a pure signal closer to the real X domain; Input the pure signal of the X domain, generate the pseudo Y domain noisy signal through the reverse conversion path processing, perform reverse feature authenticity evaluation calculation on the pseudo Y domain noisy signal and the original Y domain noisy signal, and update the parameters of the reverse conversion path according to the evaluation results; The pseudo-X domain pure signal obtained by processing the forward conversion path is input into the reverse conversion path to obtain the reconstructed Y domain noisy signal, the reconstructed Y domain noisy signal is compared with the original Y domain noisy signal, and the parameters of the two conversion paths are updated according to the feature difference; the pseudo-Y domain noisy signal obtained by processing the reverse conversion path is input into the forward conversion path to obtain the reconstructed X domain pure signal, and compared with the real X domain pure signal, and the parameters of the two conversion paths are updated according to the feature difference.

6. The magnetic resonance random noise intelligent suppression method based on dynamic weighting according to claim 1 is characterized in that: The forward feature authenticity evaluation calculation and the reverse feature authenticity evaluation calculation both include: gradually extracting feature information through multiple downsampling convolution operations, performing feature normalization processing and nonlinear activation processing after each convolution operation, obtaining a probability matrix through the final convolution operation, and taking the mean of the probability matrix as the feature authenticity evaluation result.

7. The method for intelligent suppression of magnetic resonance random noise based on dynamic weighting according to claim 6, characterized in that: The calculation of the feature difference includes: Calculate the difference between each real X-domain clean signal and the corresponding reconstructed X-domain clean signal norm, for all The norms are averaged to obtain the first mean; Calculate the difference between each original Y-domain noise signal and the corresponding reconstructed Y-domain noisy signal norm, for all The norm is averaged to obtain the second mean; The first mean and the second mean are summed to obtain the feature difference value.

8. A magnetic resonance random noise intelligent suppression system based on dynamic weighting, characterized in that: The system includes: A forward conversion module, used to generate a pure signal in the pseudo X domain from a noisy signal in the Y domain, or to generate a pure signal in the reconstructed X domain from a noisy signal in the pseudo Y domain; A reverse conversion module, used to generate a pseudo X-domain clean signal into a reconstructed Y-domain noisy signal, or to generate a pseudo Y-domain noisy signal from an X-domain clean signal; A forward evaluation module is used to perform forward feature authenticity evaluation calculation on the pure signal in the pseudo-X domain; A reverse evaluation module is used to perform reverse feature authenticity evaluation calculation on the noisy signal in the pseudo Y domain; The optimization module optimizes the two conversion paths alternately until the signal generated by the conversion path approaches the signal of the real target domain and the characteristic difference between the reconstructed signal and the original input signal reaches stable convergence, and then the optimization ends; The optimized forward conversion module performs denoising on the measured noisy data.

Images