A method for assessing the quality of magnetoencephalography (MEG) based on multi-band features

Abstract

This invention discloses a method for assessing the quality of magnetoencephalography (MEG) based on multi-band features, belonging to the field of MEG processing technology. The invention acquires a reference signal and the patient's MEG signal, dividing them into five frequency bands in the frequency domain. After merging the signals from each band and performing iFFT transformation, a frequency band fused time-domain signal is obtained. This fused signal is then linearly mixed with the patient's MEG signal to generate a mixed signal. Next, each frame of the mixed signal is subjected to FFT transformation to extract five sets of depth and spectral features, and wavelet features are extracted through wavelet transform. These three types of features are then concatenated and dimensionally aligned to form a feature map. Finally, the feature map is input into a U-Net-structured deep learning network to obtain the signal quality assessment result of the mixed signal. This method effectively improves the accuracy of MEG quality assessment through the fusion of multi-band features and processing by a deep learning network.

Description

Technical Field

[0001] This invention relates to the field of magnetoencephalography (MEG) processing technology, and more specifically to a method for assessing MEG quality based on multi-band features. Background Technology

[0002] Magnetoencephalography (MEG) is a non-invasive neurological imaging technique based on superconducting quantum interference devices (SQUIDs). It achieves millisecond-level temporal resolution monitoring of neural activity by detecting the weak magnetic field signals generated by the synchronous firing of cortical pyramidal cells. The biomagnetic field signals recorded by this technique are extremely weak, with intensities ranging from 10^-15 T (fefetesla) to 10^-12 T (pictesla), which is 9-12 orders of magnitude lower than the Earth's static magnetic field (approximately 50 × 10^-6 T).

[0003] Despite employing multi-layer magnetically shielded rooms (MSR) and active noise cancellation systems, MEG signals still face multiple interference challenges: 1) environmental electromagnetic interference (including power frequency harmonics and high-frequency electromagnetic radiation); 2) bioelectric artifacts (such as the 50-100 pT magnetic field generated by the ECG R wave and the 100-300 pT magnetic field fluctuations caused by eye movements); 3) intrinsic device noise (SQUID device noise density is approximately 2-5 fT / √Hz). Signals for clinical analysis require high signal quality; therefore, establishing a systematic quality assessment system is of significant clinical importance.

[0004] The limitation of existing technologies lies in the single dimension of quality assessment. Current research focuses on specific interference suppression (e.g., CN112220482A, which focuses on eye movement artifacts), and lacks accurate assessment of signal quality. Summary of the Invention

[0005] In view of the above-mentioned shortcomings in the prior art, the present invention provides a magnetoencephalogram (MEG) quality assessment method based on multi-band features, which solves the problem of low accuracy in MEG quality assessment in the prior art.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a method for assessing the quality of magnetoencephalography (MEG) based on multi-band features, comprising the following steps:

[0007] S1. Acquire reference signal and patient's magnetoencephalogram (MEG) signal. In the frequency domain, divide the two signals into 5 frequency bands and perform signal merging and iFFT transformation in each frequency band to obtain the frequency band fused time domain signal.

[0008] S2. Linearly mix the frequency band fused time-domain signal with the patient's magnetoencephalogram (MEG) signal to obtain a mixed signal;

[0009] S3. Perform FFT transformation on each frame of the mixed signal and divide it into 5 frequency bands, extracting 5 sets of depth features and spectral features;

[0010] S4. Perform wavelet transform on each frame of the mixed signal and extract wavelet features;

[0011] S5. Concatenate and align the wavelet features, depth features, and spectral features to obtain the feature map;

[0012] S6. Input the feature map into a deep learning network with a U-Net structure to obtain the signal quality evaluation results of the mixed signal.

[0013] Furthermore, S1 includes the following sub-steps:

[0014] S11. Collect the signal from the empty room as a reference signal;

[0015] S12. Preprocess the reference signal and the patient's magnetoencephalogram (MEG) signal respectively to obtain the reference preprocessed signal and the MEG preprocessed signal.

[0016] S13. Convert the reference preprocessed signal and the magnetoencephalogram preprocessed signal to the frequency domain respectively, and divide them into 5 frequency bands;

[0017] S14. Scale the amplitude of all frequency points in each frequency band of the reference preprocessed signal;

[0018] S15. Merge each scaled frequency band of the reference preprocessed signal with the same frequency band of the magnetoencephalogram preprocessed signal to obtain merged frequency bands. Perform windowing compensation, overlap addition and iFFT transformation on the 5 merged frequency bands to obtain the frequency band fused time domain signal.

[0019] Furthermore, S13 includes the following sub-steps:

[0020] S131. The reference preprocessed signal is divided into frames, and each frame is windowed and transformed using a Hamming window to obtain the spectrum of the reference preprocessed signal.

[0021] S132. The magnetoencephalogram (MEG) preprocessing signal is divided into frames, and each frame is windowed and transformed using a Hamming window to obtain the spectrum of the MEG preprocessing signal.

[0022] S133. Divide the spectrum of the reference preprocessed signal and the spectrum of the magnetoencephalogram preprocessed signal into 5 frequency bands respectively.

[0023] Furthermore, S3 includes the following sub-steps:

[0024] S31. Divide the mixed signal into frames to obtain multi-frame mixed sub-signals;

[0025] S32. Perform windowing operation on each frame of mixed sub-signal;

[0026] S33. Perform FFT transformation on the signal segment below the window to obtain the spectrum of the mixed sub-signal for each frame;

[0027] S34. Divide the spectrum of each frame of mixed sub-signal into 5 frequency bands, extract depth features and spectral features for each frequency band, and obtain 5 sets of depth features and spectral features.

[0028] Furthermore, the process of obtaining depth features in S34 includes: using convolution kernels to extract depth features from the same frequency band of multi-frame mixed sub-signals;

[0029] The spectral features in S34 include amplitude energy and the standard deviation of amplitude energy. The specific acquisition process includes: extracting the amplitude energy of the same frequency band of each frame of mixed sub-signal, and extracting the standard deviation of the amplitude energy of the same frequency band of each frame of mixed sub-signal.

[0030] Furthermore, S4 includes the following sub-steps:

[0031] S41. Divide the mixed signal into frames to obtain multi-frame mixed sub-signals;

[0032] S42. Divide the mixed sub-signal of each frame into blocks, and perform discrete wavelet transform on each block to obtain the wavelet transform domain signal.

[0033] S43. Divide each sub-band in the wavelet transform domain signal into 5 frequency bands;

[0034] S44. Extract wavelet features based on the five frequency bands corresponding to the wavelet transform domain signal.

[0035] Furthermore, the specific process of S43 includes: allocating subband 0 to the Delta band, subband 1 to the Theta band, subband 2 to the Alpha band, subbands 3 to 6 to the Beta band, and subbands 7 to 24 to the Gamma band.

[0036] Furthermore, the wavelet features in S44 include: energy features, average power, relative power, variance, and entropy.

[0037] Furthermore, the energy characteristics include Delta energy, Theta energy, Alpha energy, Beta energy, and Gamma energy. The specific acquisition process includes: squaring each coefficient in subband 0 and summing the squared values ​​to obtain the Delta energy; squaring each coefficient in subband 1 and summing the squared values ​​to obtain the Theta energy; squaring each coefficient in subband 2 and summing the squared values ​​to obtain the Alpha energy; squaring each coefficient in subbands 3-6 and summing the squared values ​​to obtain the Beta energy; and squaring each coefficient in subbands 7-24 and summing the squared values ​​to obtain the Gamma energy.

[0038] Average power is the ratio of the energy of a frequency band to the total number of sampling points in the corresponding frame.

[0039] Furthermore, the relative power is the ratio of the power of each frame to the total power, the power of each frame is the sum of the powers of all sub-bands corresponding to the mixed sub-signal of each frame, and the total power is the sum of the powers of all sub-bands corresponding to the mixed sub-signal of all frames.

[0040] The variance is the variance of the coefficients of the sub-bands in the corresponding frequency bands of all frame mixed sub-signals;

[0041] The process of obtaining entropy includes: taking the logarithm of each coefficient in the subband of the corresponding frequency band of all frame mixed sub-signals, multiplying the logarithmic coefficient with the corresponding coefficient, adding all the multiplication results in the frequency band, and taking the negative number to obtain the entropy.

[0042] The beneficial effects of this invention are as follows: Existing technologies mostly focus on specific interference suppression, resulting in a single dimension of quality assessment and making it difficult to comprehensively reflect the true quality of magnetoencephalography (MEG) signals. This invention mixes the reference signal from an empty room with the patient's MEG signal to obtain a hybrid signal. Then, through multi-band division (dividing the signal into 5 frequency bands), features are obtained from both FFT and wavelet transform directions, covering depth features, spectral features, and wavelet features, achieving a comprehensive and multi-dimensional assessment of MEG signal quality. This multi-feature fusion approach can more accurately capture subtle quality differences in the signal, effectively solving the problem of low assessment accuracy in existing technologies. Attached Figure Description

[0043] Figure 1 This is a flowchart of a magnetoencephalography (MEG) quality assessment method based on multi-band features.

[0044] Figure 2 This is an illustration to show to doctors. Detailed Implementation

[0045] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0046] like Figure 1 As shown, a magnetoencephalography (MEG) quality assessment method based on multi-band features includes the following steps:

[0047] S1. Acquire reference signal and patient's magnetoencephalogram (MEG) signal. In the frequency domain, divide the two signals into 5 frequency bands and perform signal merging and iFFT transformation in each frequency band to obtain the frequency band fused time domain signal.

[0048] S2. Linearly mix the frequency band fused time-domain signal with the patient's magnetoencephalogram (MEG) signal to obtain a mixed signal;

[0049] S3. Perform FFT transformation on each frame of the mixed signal and divide it into 5 frequency bands, extracting 5 sets of depth features and spectral features;

[0050] S4. Perform wavelet transform on each frame of the mixed signal and extract wavelet features;

[0051] S5. Concatenate and align the wavelet features, depth features, and spectral features to obtain the feature map;

[0052] S6. Input the feature map into a deep learning network with a U-Net structure to obtain the signal quality evaluation results of the mixed signal.

[0053] The technical solution of this invention includes two processes: the training process of the deep learning network with U-Net structure and the process of using the deep learning network with U-Net structure after training.

[0054] The training process of the U-Net deep learning network includes: obtaining a mixed signal through S1 and S2, manually scoring the mixed signal to obtain the human quality score result, which is used as the training label, and then executing S3~S6. Using the label and the output result of the U-Net deep learning network in S6, the loss function is calculated, and the gradient backpropagation algorithm is used to train the U-Net deep learning network.

[0055] The process of using the U-Net deep learning network after training includes: using the actual magnetoencephalogram (MEG) signal collected by the machine as a mixed signal, and executing S3, S4, S5, and S6 in sequence to obtain the quality assessment result of the actual MEG signal.

[0056] The signal from the empty room was collected as a reference signal, namely the background noise signal recorded by the magnetoencephalography (MEG) device.

[0057] In this embodiment, S1 includes the following sub-steps:

[0058] S11. Collect the signal from the empty room as a reference signal;

[0059] S12. Preprocess the reference signal and the patient's magnetoencephalogram (MEG) signal respectively to obtain the reference preprocessed signal and the MEG preprocessed signal.

[0060] S13. Convert the reference preprocessed signal and the magnetoencephalogram preprocessed signal to the frequency domain respectively, and divide them into 5 frequency bands;

[0061] S14. Scale the amplitude of all frequency points in each frequency band of the reference preprocessed signal;

[0062] S15. Merge each scaled frequency band of the reference preprocessed signal with the same frequency band of the magnetoencephalogram preprocessed signal to obtain merged frequency bands. Perform windowing compensation, overlap addition and iFFT transformation on the 5 merged frequency bands to obtain the frequency band fused time domain signal.

[0063] In this embodiment, S13 includes the following sub-steps:

[0064] S131. The reference preprocessed signal is divided into frames, and each frame is windowed and transformed using a Hamming window to obtain the spectrum of the reference preprocessed signal.

[0065] S132. The magnetoencephalogram (MEG) preprocessing signal is divided into frames, and each frame is windowed and transformed using a Hamming window to obtain the spectrum of the MEG preprocessing signal.

[0066] S133. Divide the spectrum of the reference preprocessed signal and the spectrum of the magnetoencephalogram preprocessed signal into 5 frequency bands respectively.

[0067] In this embodiment, the preprocessing steps of the reference signal in S12 include: temporal signal space separation (tSSS) and power frequency (power line) noise filtering.

[0068] The preprocessing steps for the S12 magnetoencephalogram (MEG) signal include: spatial separation of time-domain signals, independent component analysis (ICA) to remove eye movement and heartbeat artifacts, and power line noise filtering.

[0069] The process of converting the reference preprocessed signal to the frequency domain in S13 includes: dividing the signal into frames according to the set frame length and step size, then performing a windowing operation using a Hamming window, and finally performing an FFT transformation to obtain the spectrum of the reference preprocessed signal.

[0070] The process of converting the magnetoencephalogram (MEG) preprocessing signal to the frequency domain in S13 includes: framing according to the set frame length and step size, then performing windowing operation using a Hamming window, and finally performing FFT transformation to obtain the spectrum of the MEG signal. The parameters of framing, windowing, and FFT are consistent with the parameters of the reference preprocessing signal converted to the frequency domain.

[0071] The five frequency bands of S13 are: Delta band, Theta band, Alpha band, Beta band, and Gamma band.

[0072] The specific process of S14 includes: independently scaling the amplitude values ​​of each frequency band (Delta, Theta, Alpha, Beta, Gamma) of the reference signal spectrum within a preset scaling range. This scaling range needs to be calibrated and determined according to the actual operating conditions of each magnetoencephalography (MEG) device, and the typical value range is [1 / 8, 4].

[0073] The specific process of S15 includes: merging the frequency band of the scaled reference signal with the same frequency band of the preprocessed magnetoencephalogram (MEG) signal to obtain a frequency band fused signal. Then, windowing compensation (de-windowing effect), overlap-add (OLA) processing to merge overlapping parts, and inverse fast Fourier transform (iFFT) are performed to obtain the frequency band fused time-domain signal.

[0074] In S3, the length of the mixed signal is 5-20s.

[0075] In this embodiment, S3 includes the following sub-steps:

[0076] S31. Divide the mixed signal into frames to obtain multi-frame mixed sub-signals;

[0077] S32. Perform windowing operation on each frame of mixed sub-signal;

[0078] S33. Perform FFT transformation on the signal segment below the window to obtain the spectrum of the mixed sub-signal for each frame;

[0079] S34. Divide the spectrum of each frame of mixed sub-signal into 5 frequency bands, extract depth features and spectral features for each frequency band, and obtain 5 sets of depth features and spectral features.

[0080] In this embodiment, the specific process of S3 includes: dividing the mixed signal into frames to obtain K frames of mixed sub-signals; performing a windowing operation on each frame of mixed sub-signals; performing a Fast Fourier Transform (FFT) on the signal segments under the window to obtain the spectrum of the corresponding mixed sub-signals; and dividing the spectrum of each frame of mixed sub-signals into 5 frequency bands. A set of features is extracted from each frequency band, resulting in a total of 5 sets of features.

[0081] Each set of features includes:

[0082] Spectral characteristics (based on Fourier transform): Calculate the amplitude energy of K frames in this frequency band (dimension: K×5). Calculate the standard deviation of the energy of the K frames in this frequency band (dimension: 1×5).

[0083] Deep features: Five independent convolutional kernels (each corresponding to a frequency band) are used to process the spectral data through deep learning convolutional layers to extract deep features.

[0084] In this embodiment, the process of obtaining depth features in S34 includes: using convolution kernels to extract depth features from the same frequency band of multi-frame mixed sub-signals. The convolution kernels are divided into 5 groups, corresponding to 5 frequency bands, and the weights of each group of convolution kernels are independent of each other.

[0085] The spectral features in S34 include amplitude energy and the standard deviation of amplitude energy. The specific acquisition process includes: extracting the amplitude energy of the same frequency band of each frame of mixed sub-signal, and extracting the standard deviation of the amplitude energy of the same frequency band of each frame of mixed sub-signal.

[0086] In this embodiment, S4 includes the following sub-steps:

[0087] S41. Divide the mixed signal into frames to obtain multi-frame mixed sub-signals;

[0088] S42. Divide the mixed sub-signal of each frame into blocks, and perform discrete wavelet transform on each block to obtain the wavelet transform domain signal.

[0089] S43. Divide each sub-band in the wavelet transform domain signal into 5 frequency bands;

[0090] S44. Extract wavelet features based on the five frequency bands corresponding to the wavelet transform domain signal.

[0091] In this embodiment, the discrete wavelet transform adopts the wavelet packet transform (WPT), and the mother wavelet can be selected from: Daubechies series (e.g., 'db4', 'db8'), Symlets series (e.g., 'sym4', 'sym8'), Coiflets series (e.g., 'coif2').

[0092] The number of decomposition levels L satisfies the following formula: Where L is the number of decomposition layers. To round up, log2 represents the logarithm to the base 2, Fs is the sampling rate of the mixed sub-signal, and BW is the bandwidth of the frequency band.

[0093] For non-stationary signals in magnetoencephalography (MEG), the mother wavelet 'db4' from the Daubechies series is selected. MEGIN MEG sampling rates are typically 1000Hz or 2000Hz, and are initially downsampled to 256Hz. Considering the Delta, Threta, and Alpha bands, with a bandwidth BW of approximately 4Hz, the calculated decomposition layer L is set to 5, resulting in 2^5 = 32 subbands. Delta (0-4Hz) corresponds to subband 0. Theta (4-8Hz) corresponds to subband 1. Alpha (8-12Hz) corresponds to subband 2. Beta (12-28Hz) is approximated as subbands 3 to 6. Gamma (28-100Hz) is approximated as subbands 7 to 24.

[0094] In this embodiment, the specific process of S43 includes: assigning sub-band 0 to the Delta band, assigning sub-band 1 to the Theta band, assigning sub-band 2 to the Alpha band, assigning sub-band 3 to sub-band 6 to the Beta band, and assigning sub-band 7 to sub-band 24 to the Gamma band.

[0095] In this embodiment, the wavelet features in S44 include: energy features, average power, relative power, variance, and entropy.

[0096] The dimensions of energy characteristics are K×5; the dimensions of average power are K×5; the dimensions of relative power are K×1; the dimensions of variance are 1×5; and the dimensions of entropy are 1×5.

[0097] In this embodiment, the energy characteristics include Delta energy, Theta energy, Alpha energy, Beta energy, and Gamma energy. The specific acquisition process includes: squaring each coefficient in subband 0 and summing the squared values ​​to obtain Delta energy; squaring each coefficient in subband 1 and summing the squared values ​​to obtain Theta energy; squaring each coefficient in subband 2 and summing the squared values ​​to obtain Alpha energy; squaring each coefficient in subbands 3 to 6 and summing the squared values ​​to obtain Beta energy; and squaring each coefficient in subbands 7 to 24 and summing the squared values ​​to obtain Gamma energy.

[0098] Delta energy: Take the coefficients of sub-band 0 and substitute them into the formula E_Delta=Σ(coefficients_0[i]²), where E_Delta is the Delta energy, and coefficients_0[i] is the i-th coefficient in sub-band 0. The coefficients include approximation coefficients (cA) or detail coefficients (cD), and the approximation coefficients (cA) or detail coefficients (cD) are the lengths of the mixed sub-signal. .

[0099] Theta energy: Take the coefficients of subband 1 and substitute them into the formula E_Theta=Σ(coefficients_1[i]²), where E_Theta is the Theta energy and coefficients_1[i] is the i-th coefficient in subband 1;

[0100] Alpha energy: Take the coefficients of subband 2 and substitute them into the formula E_Alpha=Σ(coefficients_2[i]²), where E_Alpha is the Alpha energy and coefficients_2[i] is the i-th coefficient in subband 2;

[0101] Beta energy: Take the coefficients of subbands 3 to 6, calculate their energy separately, and then sum them up, that is, E_Beta=Σ(coefficients_3[i]²)+Σ(coefficients_4[i]²)+Σ(coefficients_5[i]²)+Σ(coefficients_6[i]²), where E_Beta is the Beta energy, coefficients_3[i] is the i-th coefficient in subband 3, coefficients_4[i] is the i-th coefficient in subband 4, coefficients_5[i] is the i-th coefficient in subband 5, and coefficients_6[i] is the i-th coefficient in subband 6;

[0102] Gamma energy: Take the coefficients of subbands 7 to 24, calculate their energies separately, and then sum them up, that is, E_Gamma=Σ(coefficients_7[i]²)+...+Σ(coefficients_24[i]²), where E_Gamma is the Gamma energy, coefficients_7[i] is the i-th coefficient in subband 7, and coefficients_24[i] is the i-th coefficient in subband 24.

[0103] Average power is the ratio of the energy of a frequency band to the total number of sampling points in the corresponding frame, specifically including: the ratio of Delta energy to the total number of sampling points in the corresponding frame, the ratio of Theta energy to the total number of sampling points in the corresponding frame, the ratio of Alpha energy to the total number of sampling points in the corresponding frame, the ratio of Beta energy to the total number of sampling points in the corresponding frame, and the ratio of Gamma energy to the total number of sampling points in the corresponding frame.

[0104] In this embodiment, the relative power is the ratio of the power of each frame to the total power. The power of each frame is the sum of the powers of all sub-bands corresponding to the mixed sub-signal in each frame. The total power is the sum of the powers of all sub-bands corresponding to the mixed sub-signal in all frames. In this embodiment, there are K frames of mixed sub-signal, therefore, there are K relative powers.

[0105] The variance is the variance of the coefficients of the sub-bands in the corresponding frequency bands of all frame mixed sub-signals. Specifically, it includes: calculating the variance of the coefficients of the sub-bands in the Delta frequency band of all frame mixed sub-signals, calculating the variance of the coefficients of the sub-bands in the Theta frequency band of all frame mixed sub-signals, calculating the variance of the coefficients of the sub-bands in the Alpha frequency band of all frame mixed sub-signals, calculating the variance of the coefficients of the sub-bands in the Beta frequency band of all frame mixed sub-signals, and calculating the variance of the coefficients of the sub-bands in the Gamma frequency band of all frame mixed sub-signals.

[0106] The entropy acquisition process includes: taking the logarithm of each coefficient in the sub-band of the corresponding frequency band of all frame mixed sub-signals, multiplying the logarithmic coefficient by its corresponding coefficient, summing all the multiplication results in the frequency band, and taking the negative number to obtain the entropy. Specifically, this includes: taking the logarithm of each coefficient in the Delta frequency band of all frame mixed sub-signals, multiplying the logarithmic coefficient by its corresponding coefficient, summing all the multiplication results in the Delta frequency band, and taking the negative number to obtain the entropy of the Delta frequency band; and taking the logarithm of each coefficient in the Theta frequency band of all frame mixed sub-signals, multiplying the logarithmic coefficient by its corresponding coefficient, summing all the multiplication results in the Theta frequency band, and taking the negative number to obtain the entropy of the Theta frequency band. For each coefficient in the sub-band of the Alpha band of all frame mixed sub-signals, take the logarithm, multiply the logarithmic coefficient by the corresponding coefficient, sum all the multiplication results in the Alpha band, and take the negative number to obtain the entropy of the Alpha band; for each coefficient in the sub-band of the Beta band of all frame mixed sub-signals, take the logarithm, multiply the logarithmic coefficient by the corresponding coefficient, sum all the multiplication results in the Beta band, and take the negative number to obtain the entropy of the Beta band; for each coefficient in the sub-band of the Gamma band of all frame mixed sub-signals, take the logarithm, multiply the logarithmic coefficient by the corresponding coefficient, sum all the multiplication results in the Gamma band, and take the negative number to obtain the entropy of the Gamma band.

[0107] In this embodiment, the formula for calculating entropy is: H k =-Σcoefficients_k[i]•log 10 coefficients_k[i], where H k Let be the entropy of the k-th frequency band, and coefficients_k[i] be the i-th coefficient of the sub-band in the k-th frequency band, where k is a positive integer from 1 to 5.

[0108] In this embodiment, the extracted spectral features, depth features, and wavelet features are concatenated to form five sets of comprehensive features. To address the issue of inconsistencies in different feature dimensions, one-hot encoding and zero-padding methods are used to align the feature dimensions, ultimately generating a feature map with unified dimensions.

[0109] This invention uses a mixed signal acquired via S2, along with the corresponding magnetoencephalogram (MEG) signal, to present to the interpreter. Both signal segments are T seconds long, with T typically ranging from 5 to 20 seconds. The interpreter assigns a score based on subjective judgment, from 0 to 100, with lower scores indicating poorer mixed signal quality. A subjective score is then calculated by combining scores from at least three interpreters (segments with significant errors are discarded). The evaluation criteria for the subjective score are: signal integrity (0-50 points, 10 points per frequency band), noise pollution (0-25 points, 5 points per frequency band), and clinical usability (0-25 points).

[0110] Verification of subjective rating criteria:

[0111] 1) Feasibility: When presenting the unmixed raw magnetoencephalogram (MEG) signal to the evaluator, the physician's scoring standard needs to be >= 80; when using only the signal from an empty room with noise, the physician's scoring standard needs to be <= 20.

[0112] 2) Regularity: Randomly select duration ratios of 0%, 30%, and 60%, set the amplitude value of a specific frequency band of the magnetoencephalogram (MEG) signal to zero, then splice in other spectra and convert to the time domain. Subsequently, linearly superimpose the frequency band fusion time-domain signal. The evaluators' scores are negatively correlated with the duration ratio, meaning 0% received the highest score and 60% received the lowest. Similarly, swap the scaled reference signal and the MEG signal, and linearly superimpose the frequency band fusion time-domain signal with the MEG signal according to the 0%, 30%, and 60% duration ratios. The evaluators' scores are positively correlated with the duration ratio, meaning 0% received the lowest score and 60% received the highest.

[0113] 3) Correlation: Considering the subjective sensitivity of manual evaluation, there are certain differences for signals in different frequency bands. Therefore, for a specific frequency band, based on the original signal-to-noise ratio, a mixed signal with a new signal-to-noise ratio is generated according to scaling factors of 1 / 4, 1, and 4, while the signal-to-noise ratios of the other four frequency bands remain unchanged. The evaluator's score needs to be positively correlated with the change in signal-to-noise ratio.

[0114] 4) Repeatability: For the same mixed signal, the same doctor should repeat the evaluation at intervals. If the relative error of the score is less than 10%, it is considered to pass. The pass rate needs to be >85% to confirm the objectivity of the score.

[0115] Training the quality assessment model:

[0116] 1) The T-second mixed signal is transformed to the frequency domain using Fourier transform, and divided into 5 frequency bands according to Delta, Theta, Alpha, Beta, and Gamma bands. 5 sets of depth features and spectral features are extracted.

[0117] 2) Perform wavelet transform on each frame of the mixed signal and extract wavelet features;

[0118] 3) The wavelet features, depth features, and spectral features are concatenated and dimensionally aligned to obtain the feature map;

[0119] 4) Use the U-Net structure to map feature maps to prediction results;

[0120] 5) Normalize the doctor's subjective rating to 0-1, calculate the loss value with the prediction result in 4), and train a deep learning network with U-Net structure through backpropagation algorithm.

[0121] Existing technologies mostly focus on specific interference suppression, resulting in a single dimension of quality assessment and making it difficult to comprehensively reflect the true quality of magnetoencephalography (MEG) signals. This invention mixes the reference signal from an empty room with the patient's MEG signal to obtain a hybrid signal. Then, through multi-band segmentation (dividing the signal into 5 frequency bands), features are extracted from both FFT and wavelet transform directions, encompassing depth features, spectral features, and wavelet features. This achieves a comprehensive, multi-dimensional assessment of MEG signal quality. This multi-feature fusion approach can more accurately capture subtle quality differences in the signal, effectively solving the problem of low assessment accuracy in existing technologies.

[0122] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. 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 assessing the quality of magnetoencephalography (MEG) based on multi-band features, characterized in that, Includes the following steps: S1. Acquire reference signal and patient's magnetoencephalogram (MEG) signal. In the frequency domain, divide the two signals into 5 frequency bands and perform signal merging and iFFT transformation in each frequency band to obtain the frequency band fused time domain signal. S2. Linearly mix the frequency band fused time-domain signal with the patient's magnetoencephalogram (MEG) signal to obtain a mixed signal; S3. Perform FFT transformation on each frame of the mixed signal and divide it into 5 frequency bands, extracting 5 sets of depth features and spectral features; S4. Perform wavelet transform on each frame of the mixed signal and extract wavelet features; S5. Concatenate and align the wavelet features, depth features, and spectral features to obtain the feature map; S6. Input the feature map into a deep learning network with a U-Net structure to obtain the signal quality evaluation results of the mixed signal; S1 includes the following steps: S11. Collect the signal from the empty room as a reference signal; S12. Preprocess the reference signal and the patient's magnetoencephalogram (MEG) signal respectively to obtain the reference preprocessed signal and the MEG preprocessed signal. S13. Convert the reference preprocessed signal and the magnetoencephalogram preprocessed signal to the frequency domain respectively, and divide them into 5 frequency bands; S14. Scale the amplitude of all frequency points in each frequency band of the reference preprocessed signal; S15. Merge each scaled frequency band of the reference preprocessed signal with the same frequency band of the magnetoencephalogram preprocessed signal to obtain merged frequency bands. Perform windowing compensation, overlap addition and iFFT transformation on the 5 merged frequency bands to obtain the frequency band fused time domain signal. The process of obtaining depth features includes: using convolution kernels to extract depth features from the same frequency band of multi-frame mixed sub-signals; The spectral features include amplitude energy and the standard deviation of amplitude energy. The specific acquisition process includes: extracting the amplitude energy of the same frequency band of each frame of mixed sub-signal, and extracting the standard deviation of the amplitude energy of the same frequency band of each frame of mixed sub-signal. S4 includes the following steps: S41. Divide the mixed signal into frames to obtain multi-frame mixed sub-signals; S42. Divide each frame of mixed sub-signals into blocks, and perform discrete wavelet transform on each block of signal to obtain the wavelet transform domain signal; S43. Divide each sub-band in the wavelet transform domain signal into 5 frequency bands; S44. Extract wavelet features based on the five frequency bands corresponding to the wavelet transform domain signal; The specific process of S43 includes: assigning sub-band 0 to the Delta band, sub-band 1 to the Theta band, sub-band 2 to the Alpha band, sub-bands 3 to 6 to the Beta band, and sub-bands 7 to 24 to the Gamma band.

2. The magnetoencephalography (MEG) quality assessment method based on multi-band features according to claim 1, characterized in that, S13 includes the following steps: S131. The reference preprocessed signal is divided into frames, and each frame is windowed and transformed using a Hamming window to obtain the spectrum of the reference preprocessed signal. S132. The magnetoencephalogram (MEG) preprocessing signal is divided into frames, and each frame is windowed and transformed using a Hamming window to obtain the spectrum of the MEG preprocessing signal. S133. Divide the spectrum of the reference preprocessed signal and the spectrum of the magnetoencephalogram preprocessed signal into 5 frequency bands respectively.

3. The magnetoencephalography (MEG) quality assessment method based on multi-band features according to claim 1, characterized in that, S3 includes the following steps: S31. Divide the mixed signal into frames to obtain multi-frame mixed sub-signals; S32. Perform windowing operation on each frame of mixed sub-signal; S33. Perform FFT transformation on the signal segment below the window to obtain the spectrum of the mixed sub-signal for each frame; S34. Divide the spectrum of each frame of mixed sub-signal into 5 frequency bands, extract depth features and spectral features for each frequency band, and obtain 5 sets of depth features and spectral features.

4. The magnetoencephalography (MEG) quality assessment method based on multi-band features according to claim 1, characterized in that, The wavelet features in S44 include: energy features, average power, relative power, variance, and entropy.

5. The magnetoencephalography (MEG) quality assessment method based on multi-band features according to claim 4, characterized in that, The energy characteristics include Delta energy, Theta energy, Alpha energy, Beta energy, and Gamma energy. The specific acquisition process includes: squaring each coefficient in subband 0 and summing the squared values ​​to obtain the Delta energy; squaring each coefficient in subband 1 and summing the squared values ​​to obtain the Theta energy; squaring each coefficient in subband 2 and summing the squared values ​​to obtain the Alpha energy; squaring each coefficient in subbands 3-6 and summing the squared values ​​to obtain the Beta energy; and squaring each coefficient in subbands 7-24 and summing the squared values ​​to obtain the Gamma energy. Average power is the ratio of the energy of a frequency band to the total number of sampling points in the corresponding frame.

6. The magnetoencephalography (MEG) quality assessment method based on multi-band features according to claim 4, characterized in that, The relative power is the ratio of the power of each frame to the total power. The power of each frame is the sum of the power of all sub-bands corresponding to the mixed sub-signal in each frame. The total power is the sum of the power of all sub-bands corresponding to the mixed sub-signal in all frames. The variance is the variance of the coefficients of the sub-bands in the corresponding frequency bands of all frame mixed sub-signals; The process of obtaining entropy includes: taking the logarithm of each coefficient in the subband of the corresponding frequency band of all frame mixed sub-signals, multiplying the logarithmic coefficient with the corresponding coefficient, adding all the multiplication results in the frequency band, and taking the negative number to obtain the entropy.

Images