Speech enhancement method and device in magnetic resonance imaging and magnetic resonance imaging system

Abstract

The embodiment of the invention discloses a speech enhancement method and device in magnetic resonance imaging and a magnetic resonance imaging system. The method comprises the following steps: acquiring a time domain sound signal in an MR scanning room in real time; converting the time domain sound signal into a frequency domain to obtain a frequency domain sound signal; inputting the frequency domain sound signal into a band-pass filter of which the filtering bandwidth is equal to the filtering bandwidth of human auditory sense, and inputting the frequency domain sound signal after band-pass filtering into a first artificial intelligence module for noise elimination to obtain a first frequency domain clean sound signal; and converting the frequency domain first clean sound signal from the frequency domain to the time domain to obtain a time domain clean sound signal. According to the embodiment of the invention, the voice quality of internal communication in MR imaging is improved.

Description

Technical Field

[0001] This invention relates to the field of MR (Magnetic Resonance) technology, and in particular to speech enhancement methods, devices, and MRI (Magnetic Resonance Imaging) systems in MR imaging. Background Technology

[0002] Smooth communication between the patient and the operator is crucial during MR imaging, as it improves patient comfort, increases operator efficiency, and shortens scan time. Figure 1 This diagram illustrates the status of the patient and operator during MRI system operation. 11 represents the patient undergoing an MRI scan in the MRI scanning room, and 12 represents the operator controlling the MRI scan process and receiving MRI images in the MRI control room. In some situations, such as when a patient feels unwell, they may need to communicate with the operator to calm down; in other cases, the operator may also want to know the patient's status. Current MRI systems typically have an internal communication system that allows communication between the patient in the MRI scanning room and the operator in the MRI control room.

[0003] However, the acoustic noise generated during MRI system operation can interfere with the use of internal communication systems. This noise is picked up by the microphones in the MRI system and can mask the patient's voice in the audio signal transmitted to the operator in the MRI control room, making it difficult for the operator to understand what the patient is saying. Communication difficulties between the patient and operator can have a negative psychological impact on the patient. For patients who are uncomfortable in a confined space, the overall experience of MRI imaging can be challenging, and their discomfort can be exacerbated when they realize that communication with the operator is difficult. In practical applications, this is a significant problem, leading some patients to find MRI imaging completely unbearable.

[0004] Current solutions in MR imaging typically involve the operator pausing the MR scan and communicating with the patient via an internal communication system when the patient wishes to communicate, or the operator going to the MR scan room to speak with the patient. This obviously reduces the efficiency of MR scans. Furthermore, there is currently no solution for speech enhancement in extremely noisy environments within MRI systems; noise levels in some MRI sequences can exceed 100 dB, making speech enhancement very challenging. Existing multi-microphone solutions offer a method for signal denoising, but this requires multiple microphones for beamforming, which is impractical in MRI systems due to space constraints and potential RF (Radio Frequency) interference. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide, on the one hand, a method and apparatus for speech enhancement in MR imaging to improve the speech quality of internal communication in MR imaging; on the other hand, an MRI system to improve the speech quality of internal communication in MR imaging; and on yet another aspect, computer program products, computer-readable storage media, and electronic devices to improve the speech quality of internal communication in MR imaging.

[0006] A speech enhancement method in magnetic resonance imaging, the method comprising:

[0007] Real-time acquisition of time-domain sound signals within the MR scanning chamber;

[0008] The time-domain sound signal is converted from the time domain to the frequency domain to obtain a frequency-domain sound signal;

[0009] The frequency domain sound signal is input into a bandpass filter with a filtering bandwidth equal to that of human hearing to obtain a bandpass filtered frequency domain sound signal.

[0010] The bandpass filtered frequency domain audio signal is input into a pre-trained first artificial intelligence module for noise removal to obtain a first clean frequency domain audio signal.

[0011] The first clean audio signal in the frequency domain is converted from the frequency domain to the time domain to obtain the clean audio signal in the time domain.

[0012] The first artificial intelligence module includes multiple encoding layers and multiple decoding layers, and the multiple encoding layers and multiple decoding layers are connected in a skip connection.

[0013] The first artificial intelligence module includes: m encoding layers, m-1 decoding layers, and 1 gain prediction layer, where m ≥ 2;

[0014] The first to m-1 encoding layers and the first to m-1 decoding layers of the first artificial intelligence module are symmetrically arranged with respect to the m-th encoding layer, and the first to m-1 encoding layers and the m-1 to first decoding layers are respectively connected in a skip connection manner.

[0015] After the real-time acquisition of the time-domain sound signal inside the MR scanning room and before the conversion of the time-domain sound signal from the time domain to the frequency domain, the process further includes:

[0016] Calculate the signal-to-noise ratio (SNR) of the time-domain audio signal; if the SNR is less than a preset SNR threshold, input the time-domain audio signal into a low-pass filter to obtain a low-pass filtered time-domain audio signal.

[0017] The step of converting the time-domain audio signal from the time domain to the frequency domain includes:

[0018] The low-pass filtered time-domain audio signal is converted from the time domain to the frequency domain.

[0019] The bandpass filter whose filtering bandwidth is equivalent to that of human hearing is called an Equivalent Rectangular Bandwidth (ERB) filter.

[0020] The first artificial intelligence module is obtained through the following training process:

[0021] A1. Collect multiple training samples, wherein each training sample is obtained by mixing a clean speech signal collected with a segment of environmental noise collected during the operation of the MRI system in the MR scanning room, or by directly collecting a mixed sound signal of speech and environmental noise collected during the operation of the MRI system in the MR scanning room; the target sound signal corresponding to each training sample is: the clean speech signal corresponding to that training sample.

[0022] B1. Select a training sample and convert it from the time domain to the frequency domain to obtain a frequency domain sound signal. Input the frequency domain sound signal into a bandpass filter with a filtering bandwidth equal to that of human hearing to obtain a bandpass-filtered frequency domain sound signal. Input the bandpass-filtered frequency domain sound signal into the first artificial intelligence module for noise removal to obtain a predicted first clean frequency domain sound signal. Convert the predicted first clean frequency domain sound signal from the frequency domain to the time domain to obtain a predicted clean time domain sound signal. Compare the predicted clean time domain sound signal with the target sound signal corresponding to the input training sample and adjust the parameters of the first artificial intelligence module according to the comparison result.

[0023] Repeat step B1 until the first AI module converges, and use the converged first AI module as the final first AI module.

[0024] After obtaining the frequency domain sound signal, and before inputting the frequency domain sound signal into a bandpass filter with a filtering bandwidth equal to that of human hearing, the process further includes:

[0025] Make a copy of the frequency domain sound signal;

[0026] Furthermore, the step of inputting the frequency domain sound signal into a bandpass filter with a filtering bandwidth equal to that of human hearing includes:

[0027] A portion of the frequency domain sound signal is input into a bandpass filter with a filtering bandwidth equal to that of human hearing.

[0028] Furthermore, the step of inputting the frequency domain sound signal into a bandpass filter with a filtering bandwidth equal to that of human hearing further includes:

[0029] Another frequency domain sound signal is input into a pre-trained second artificial intelligence module for noise removal to obtain a second clean frequency domain sound signal.

[0030] The step of converting the first clean audio signal in the frequency domain from the frequency domain to the time domain to obtain a clean audio signal in the time domain includes:

[0031] The first clean audio signal in the frequency domain and the second clean audio signal in the frequency domain are mixed to obtain the third clean audio signal in the frequency domain. The third clean audio signal in the frequency domain is then converted from the frequency domain to the time domain to obtain the clean audio signal in the time domain.

[0032] The second artificial intelligence module includes multiple encoding layers and multiple decoding layers, with skip connections between the multiple encoding layers and multiple decoding layers.

[0033] The second artificial intelligence module includes: n encoding layers and n-1 decoding layers, where n≥2;

[0034] Furthermore, the first to n-1 encoding layers and the first to n-1 decoding layers of the second artificial intelligence module are symmetrically arranged with respect to the nth encoding layer, and the first to n-1 encoding layers and the n-1 to first decoding layers are respectively connected in a skip connection manner.

[0035] The first artificial intelligence module and the second artificial intelligence module are obtained through the following training process:

[0036] A2. Collect multiple training samples, wherein each training sample is obtained by mixing a clean speech signal collected with a segment of environmental noise collected during the operation of the MRI system in the MR scanning room, or by directly collecting a mixed sound signal of speech and environmental noise collected during the operation of the MRI system in the MR scanning room; the target sound signal corresponding to each training sample is: the clean speech signal corresponding to that training sample.

[0037] B2. Select a training sample and convert it from the time domain to the frequency domain to obtain a frequency domain sound signal. Make a copy of the frequency domain sound signal. Input one copy of the frequency domain sound signal into a bandpass filter with a filtering bandwidth equal to that of human hearing to obtain a bandpass-filtered frequency domain sound signal. Input the bandpass-filtered frequency domain sound signal into the first artificial intelligence module for noise removal to obtain a predicted first clean frequency domain sound signal. Input another copy of the frequency domain sound signal into the second artificial intelligence module for noise removal to obtain a predicted second clean frequency domain sound signal. Mix the predicted first and second clean frequency domain sound signals to obtain a predicted third clean frequency domain sound signal. Convert the predicted third clean frequency domain sound signal from the frequency domain to the time domain to obtain a predicted clean time domain sound signal. Compare the predicted clean time domain sound signal with the target sound signal corresponding to the input training sample, and adjust the parameters of the first and second artificial intelligence modules according to the comparison result.

[0038] Repeat step B2 until the first AI module and the second AI module converge. The converged first AI module and the second AI module are then used as the final first AI module and the second AI module.

[0039] A speech enhancement device for magnetic resonance imaging, the device comprising:

[0040] The time-frequency domain conversion module is used for: real-time acquisition of time-domain sound signals in the magnetic resonance MR scanning chamber; conversion of the time-domain sound signals from the time domain to the frequency domain to obtain frequency-domain sound signals; and outputting the frequency-domain sound signals to a bandpass filter.

[0041] A bandpass filter is used to perform bandpass filtering on frequency domain audio signals and output the bandpass-filtered frequency domain audio signals to the first artificial intelligence module; wherein, the filtering bandwidth of the bandpass filter is equal to the filtering bandwidth of human hearing;

[0042] The first artificial intelligence module is used to: perform noise removal on the bandpass filtered frequency domain audio signal to obtain the first clean frequency domain audio signal, and output the first clean frequency domain audio signal to the frequency-time domain conversion module;

[0043] The frequency-time domain conversion module is used to convert the first clean audio signal in the frequency domain to the time domain, thereby obtaining a clean audio signal in the time domain.

[0044] The first artificial intelligence module includes multiple encoding layers and multiple decoding layers, and the multiple encoding layers and multiple decoding layers are connected in a skip connection.

[0045] The first artificial intelligence module includes: m encoding layers, m-1 decoding layers, and 1 gain prediction layer, where m ≥ 2;

[0046] The first to m-1 encoding layers and the first to m-1 decoding layers of the first artificial intelligence module are symmetrically arranged with respect to the m-th encoding layer, and the first to m-1 encoding layers and the m-1 to first decoding layers are respectively connected in a skip connection manner.

[0047] The time-frequency domain conversion module includes: a signal-to-noise ratio calculation and judgment module, a low-pass filter, and a conversion module, wherein:

[0048] The signal-to-noise ratio (SNR) calculation and judgment module is used to: acquire time-domain sound signals in the MR scanning room in real time, calculate the SNR of the time-domain sound signals, and if the SNR is less than a preset SNR threshold, input the time-domain sound signals into a low-pass filter; otherwise, directly output the time-domain sound signals acquired by the microphone to the conversion module.

[0049] A low-pass filter is used to: perform low-pass filtering on the time-domain audio signal and output the low-pass filtered time-domain audio signal to the conversion module;

[0050] The conversion module is used to: convert the time-domain audio signal from the future information-to-noise ratio calculation and judgment module from the time domain to the frequency domain to obtain a frequency-domain audio signal, or convert the low-pass filtered time-domain audio signal from the low-pass filter from the time domain to the frequency domain to obtain a frequency-domain audio signal; and output the frequency-domain audio signal to the band-pass filter.

[0051] The device further includes: a second artificial intelligence module;

[0052] After the time-frequency domain conversion module obtains the frequency domain sound signal and before outputting the frequency domain sound signal to the bandpass filter, it further includes: making a copy of the frequency domain sound signal;

[0053] Furthermore, the time-frequency domain conversion module outputs the frequency domain audio signal to the bandpass filter, including: outputting a copy of the frequency domain audio signal to the bandpass filter;

[0054] Furthermore, while the time-frequency domain conversion module outputs the frequency domain audio signal to the bandpass filter, it further includes: outputting another copy of the frequency domain audio signal to the second artificial intelligence module;

[0055] The second artificial intelligence module is used to: perform noise removal on the frequency domain sound signal to obtain a second clean sound signal in the frequency domain, and output the second clean sound signal in the frequency domain to the frequency-time domain conversion module;

[0056] The frequency-time domain conversion module converts the first clean sound signal in the frequency domain from the frequency domain to the time domain to obtain a clean sound signal in the time domain. This includes: mixing the first clean sound signal in the frequency domain and the second clean sound signal in the frequency domain to obtain a third clean sound signal in the frequency domain, and converting the third clean sound signal in the frequency domain from the frequency domain to the time domain to obtain a clean sound signal in the time domain.

[0057] A magnetic resonance imaging system, the magnetic resonance imaging system including a speech enhancement device as described in any of the above magnetic resonance imaging methods.

[0058] In this embodiment of the invention, considering that noise in the MR imaging environment is mainly caused by the periodic frequency of the MR sequence, the time-domain sound signal acquired in the MR scanning room is first converted to the frequency domain. Then, using the psychoacoustic characteristics of the frequency-domain sound signal, such as logarithmic response and frequency perception, a bandpass filter with a filtering bandwidth equivalent to the human hearing bandwidth is used to filter the frequency-domain sound signal to obtain a frequency-domain sound signal with a receiving bandwidth equivalent to that of the human ear. Then, noise cancellation is performed on the frequency-domain sound signal with a receiving bandwidth equivalent to that of the human ear to obtain a first clean sound signal in the frequency domain. Finally, the first clean sound signal in the frequency domain is converted to the time domain to obtain a clean sound signal in the time domain. Without adding additional system components and hardware costs, the voice quality of internal communication in MR imaging is improved, MR scanning efficiency is improved, and the experience of patients and operators is improved. In addition, since noise cancellation is performed in the frequency domain, and noise in the MR imaging environment is mainly caused by the periodic frequency of the MR sequence, the voice enhancement method provided by this embodiment of the invention can filter out noise in the sound signal even when facing an MR imaging environment using an unknown MR sequence, and is not limited by the application scenario. Attached Figure Description

[0059] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which will make the above and other features and advantages of the present invention more apparent to those skilled in the art. In the drawings:

[0060] Figure 1 A schematic diagram illustrating the status of the patient and operator during MRI system operation;

[0061] Figure 2 This is a flowchart of a speech enhancement method in MR imaging provided by an embodiment of the present invention;

[0062] Figure 3 This is a schematic diagram of the structure of the first artificial intelligence module in an application example of the present invention;

[0063] Figure 4 A flowchart of a speech enhancement method in MR imaging provided in another embodiment of the present invention;

[0064] Figure 5 This is a schematic diagram of the structure of the second artificial intelligence module in an application example of the present invention;

[0065] Figure 6 A flowchart illustrating a training method for a first artificial intelligence module and a second artificial intelligence module according to an embodiment of the present invention;

[0066] Figure 7 This is a schematic diagram of the structure of a speech enhancement device in MR imaging according to an embodiment of the present invention;

[0067] Figure 8 This is a schematic diagram of the structure of a speech enhancement device in MR imaging provided in another embodiment of the present invention;

[0068] Figure 9 This is a schematic diagram of the structure of a speech enhancement device in MR imaging provided in another embodiment of the present invention.

[0069] The reference numerals in the attached figures are as follows:

[0070]

[0071] Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the present invention clearer, the following embodiments are provided to further illustrate the present invention in detail.

[0073] Figure 2 Here is a flowchart of a speech enhancement method in MR imaging according to an embodiment of the present invention, and its specific steps are as follows:

[0074] Step 201: Real-time acquisition of time-domain sound signals in the MR scanning room.

[0075] Specifically, the time-domain sound signal is acquired in real time from the microphone inside the MR scanning room. To achieve the best acquisition effect of the patient's voice, the microphone is usually placed as close as possible to the patient's mouth.

[0076] Considering that when an MRI system uses high-frequency MR sequences, the SNR (Signal Noise Ratio) of ambient noise is typically low, and most voiced speech frequencies are below 5kHz, a low-pass filter can be used to filter out ambient noise first. In an optional embodiment, after step 201 and before step 202, the method further includes: calculating the SNR of the time-domain sound signal; if the calculated SNR is less than a preset SNR threshold, then the time-domain sound signal is input to a low-pass filter to obtain a low-pass filtered time-domain sound signal.

[0077] Step 202: Convert the time-domain sound signal from the time domain to the frequency domain to obtain the frequency-domain sound signal.

[0078] In one optional embodiment, the STFT (Short-Time Fourier Transform) algorithm is used to convert the time-domain audio signal from the time domain to the frequency domain.

[0079] If the time-domain audio signal acquired by the microphone is low-pass filtered after step 201, then in this step 202, the low-pass filtered time-domain audio signal is converted from the time domain to the frequency domain.

[0080] Step 203: Input the frequency domain sound signal into a bandpass filter with a filtering bandwidth equal to that of human hearing to obtain a bandpass filtered frequency domain sound signal.

[0081] In one optional embodiment, the bandpass filter with a filtering bandwidth equivalent to that of human hearing is an ERB (Equivalent Rectangular Bandwidth) filter.

[0082] Step 204: Input the bandpass filtered frequency domain audio signal into the pre-trained first artificial intelligence module for noise removal to obtain the first clean frequency domain audio signal.

[0083] In one optional embodiment, the first artificial intelligence module includes: multiple encoding layers and multiple decoding layers, and the multiple encoding layers and multiple decoding layers are connected in a skip connection.

[0084] In one optional embodiment, the first artificial intelligence module includes: m encoding layers, m-1 decoding layers and 1 gain prediction layer, where m ≥ 2.

[0085] To improve speech enhancement by transmitting as many encoded features as possible to the decoding layer, an optional embodiment proposes that the first to m-1 encoding layers and the first to m-1 decoding layers of the first artificial intelligence module are symmetrically arranged with respect to the m-th encoding layer, and the first to m-1 encoding layers and the m-1 to 1 decoding layers are connected in a skip connection manner. For example, with the m-th encoding layer as the axis of symmetry, for each pair of encoding and decoding layers that are symmetrical on the left and right sides of the first artificial intelligence module with respect to the m-th encoding layer, the encoded features output by the encoding layer in the pair are input to the decoding layer as input features of the decoding layer. For example: the (m-1)th encoding layer is paired with the 1st decoding layer, the (m-2)th encoding layer is paired with the 2nd decoding layer, ..., the mp-th encoding layer is paired with the p-th decoding layer, ..., the 2nd encoding layer is paired with the (m-2)th decoding layer, and the 1st encoding layer is paired with the (m-1)th decoding layer; where, for the mp-th encoding layer and the p-th decoding layer, 1≤p≤m-1, the encoded features output by the mp-th encoding layer are output to the p-th decoding layer as the input features of the p-th decoding layer.

[0086] Figure 3 This is a schematic diagram of the structure of the first artificial intelligence module in an application example of the present invention. Figure 3 As shown, the first artificial intelligence module includes: m coding layers C1-Cm, m-1 decoding layers D1-Dm-1, and one gain prediction layer G1, where m ≥ 2. The size of the coding layers decreases sequentially from C1 to Cm, while the size of the decoding layers increases sequentially from D1 to Dm-1. The size of the decoding layer Dm-1 is the same as that of the coding layer C1, ensuring that the length of the noise-cancelled audio signal is equal to the length of the audio signal before noise cancellation. Furthermore, the encoded features output from the m-th coding layer Cm are output to the first decoding layer D1, and the encoded features output from the mp-th coding layer Cm-p are output to the p-th decoding layer Dp, where 1 ≤ p ≤ m-1.

[0087] The first artificial intelligence module can adopt a DNN (Deep Neural Network), and all encoding layers, decoding layers and gain prediction layers of the first artificial intelligence module are convolutional layers.

[0088] In one optional embodiment, the first artificial intelligence module is obtained through the following training process:

[0089] A1. Collect multiple training samples, wherein each training sample is obtained by mixing a clean speech signal collected with a segment of environmental noise collected during the operation of the MRI system in the MR scanning room, or by directly collecting a mixed sound signal of speech and environmental noise collected during the operation of the MRI system in the MR scanning room; the target sound signal corresponding to each training sample is: the clean speech signal corresponding to that training sample.

[0090] B1. Select a training sample and convert it from the time domain to the frequency domain to obtain a frequency domain sound signal. Input the frequency domain sound signal into a bandpass filter with a filtering bandwidth equal to that of human hearing to obtain a bandpass-filtered frequency domain sound signal. Input the bandpass-filtered frequency domain sound signal into the first artificial intelligence module for noise removal to obtain a predicted first clean frequency domain sound signal. Convert the predicted first clean frequency domain sound signal from the frequency domain to the time domain to obtain a predicted clean time domain sound signal. Compare the predicted clean time domain sound signal with the target sound signal corresponding to the input training sample and adjust the parameters of the first artificial intelligence module based on the comparison result.

[0091] Repeat step B1 until the first AI module converges, and use the converged first AI module as the final first AI module.

[0092] Step 205: Convert the first clean audio signal in the frequency domain to the time domain to obtain a clean audio signal in the time domain.

[0093] In one optional embodiment, the ISTFT (Inverse Short-Time Fourier Transform) algorithm is used to convert the first clean audio signal in the frequency domain from the frequency domain to the time domain.

[0094] After obtaining the clean time-domain sound signal, the clean time-domain sound signal is transmitted to the MRI control room through the internal communication system of the MRI system, such as a walkie-talkie.

[0095] In the above embodiments, considering that noise in the MR imaging environment is mainly caused by the periodic frequency of the MR sequence, the time-domain sound signal acquired in the MR scanning room is first converted to the frequency domain. Then, utilizing the psychoacoustic characteristics of the frequency-domain sound signal, such as logarithmic response and frequency perception, a bandpass filter with a filtering bandwidth equal to that of human hearing is used to filter the frequency-domain sound signal, obtaining a frequency-domain sound signal with a bandwidth equal to that of human hearing. Then, noise removal is performed on the frequency-domain sound signal with the same bandwidth as that of human hearing to obtain the first clean frequency-domain sound signal. Finally, the noise is... The first clean audio signal in the frequency domain is converted to the time domain to obtain a clean audio signal in the time domain. This improves the voice quality of internal communication in MR imaging, increases MR scanning efficiency, and enhances the experience for patients and operators without adding extra system components or hardware costs. Furthermore, since noise cancellation is performed in the frequency domain, and noise in the MR imaging environment is mainly caused by the periodic frequency of the MR sequence, the voice enhancement method provided in this embodiment can filter out noise in the audio signal even when facing an MR imaging environment using an unknown MR sequence, and is not limited by the application scenario.

[0096] It should be noted that if the microphone is a multi-channel microphone, that is, the microphone can simultaneously collect multiple time-domain audio signals, then each time-domain audio signal can be enhanced by the above steps 202-205.

[0097] Figure 4 A flowchart of a speech enhancement method in MR imaging provided in another embodiment of the present invention is shown, and its specific steps are as follows:

[0098] Step 401: Real-time acquisition of time-domain sound signals in the MR scanning room.

[0099] Specifically, the time-domain sound signal is acquired in real time from the microphone inside the MR scanning room. To achieve the best acquisition effect of the patient's voice, the microphone is usually placed as close as possible to the patient's mouth.

[0100] Considering that when an MRI system uses high-frequency MR sequences, the SNR (Signal Noise Ratio) of ambient noise is typically low, and most voiced speech frequencies are below 5kHz, a low-pass filter can be used to filter out ambient noise first. In an optional embodiment, after step 401 and before step 402, the method further includes: calculating the SNR of the time-domain sound signal; if the calculated SNR is less than a preset SNR threshold, then the time-domain sound signal is input to a low-pass filter to obtain a low-pass filtered time-domain sound signal.

[0101] Step 402: Convert the time-domain sound signal from the time domain to the frequency domain to obtain the frequency-domain sound signal.

[0102] In one optional embodiment, the STFT algorithm is used to convert the time-domain audio signal from the time domain to the frequency domain.

[0103] If the time-domain audio signal acquired by the microphone is low-pass filtered after step 401, then in this step 402, the low-pass filtered time-domain audio signal is converted from the time domain to the frequency domain.

[0104] Step 403: Make a copy of the frequency domain audio signal.

[0105] Step 404: Input a frequency domain sound signal into a bandpass filter with a filtering bandwidth equal to that of human hearing to obtain a bandpass filtered frequency domain sound signal; input the bandpass filtered frequency domain sound signal into a pre-trained first artificial intelligence module for noise removal to obtain a first clean frequency domain sound signal.

[0106] In one optional embodiment, the bandpass filter with a filtering bandwidth equivalent to that of human hearing is an ERB filter.

[0107] In one optional embodiment, the first artificial intelligence module includes: multiple encoding layers and multiple decoding layers, and the multiple encoding layers and multiple decoding layers are connected in a skip connection.

[0108] In one optional embodiment, the first artificial intelligence module includes: m encoding layers, m-1 decoding layers and 1 gain prediction layer, where m ≥ 2.

[0109] To improve speech enhancement by transmitting as many encoded features as possible to the decoding layer, an optional embodiment proposes that the first to m-1 encoding layers and the first to m-1 decoding layers of the first artificial intelligence module are symmetrically arranged with respect to the m-th encoding layer, and the first to m-1 encoding layers and the m-1 to 1 decoding layers are connected in a skip connection manner. For example, with the m-th encoding layer as the axis of symmetry, for each pair of encoding and decoding layers that are symmetrical on the left and right sides of the first artificial intelligence module with respect to the m-th encoding layer, the encoded features output by the encoding layer in the pair are input to the decoding layer as input features of the decoding layer. For example: the (m-1)th encoding layer is paired with the 1st decoding layer, the (m-2)th encoding layer is paired with the 2nd decoding layer, ..., the mp-th encoding layer is paired with the p-th decoding layer, ..., the 2nd encoding layer is paired with the (m-2)th decoding layer, and the 1st encoding layer is paired with the (m-1)th decoding layer; where, for the mp-th encoding layer and the p-th decoding layer, 1≤p≤m-1, the encoded features output by the mp-th encoding layer are output to the p-th decoding layer as the input features of the p-th decoding layer.

[0110] The structure of the first artificial intelligence module can be as follows: Figure 3 As shown.

[0111] Step 405: Input another frequency domain sound signal into the pre-trained second artificial intelligence module for noise removal to obtain a second clean frequency domain sound signal.

[0112] In one optional embodiment, the second artificial intelligence module includes: multiple encoding layers and multiple decoding layers, and the multiple encoding layers and multiple decoding layers are connected in a skip connection.

[0113] In one optional embodiment, the second artificial intelligence module includes: n encoding layers and n-1 decoding layers, where n≥2.

[0114] To improve speech enhancement by transmitting as many coded features as possible to the decoding layer, an optional embodiment proposes that the first to (n-1)th coding layers and the first to (n-1)th decoding layers of the second artificial intelligence module are symmetrically arranged with respect to the nth coding layer, and the first to (n-1)th coding layers and the (n-1)th to (1)th decoding layers are connected in a skip connection manner. For example, with the nth coding layer as the axis of symmetry, for each coding layer and decoding layer pair symmetrically positioned on the left and right sides of the second artificial intelligence module with respect to the nth coding layer, the coded features output by the coding layer in the coding layer and decoding layer pair are input to the decoding layer in the coding layer and decoding layer pair as input features for the decoding layer. For example: the (n-1)th encoding layer is paired with the 1st decoding layer, the (n-2)th encoding layer is paired with the 2nd decoding layer, ..., the nqth encoding layer is paired with the qth decoding layer, ..., the 2nd encoding layer is paired with the (n-2)th decoding layer, and the 1st encoding layer is paired with the (n-1)th decoding layer; where, for the nqth encoding layer and the qth decoding layer, 1≤q≤n-1, the encoded features output by the nqth encoding layer are output to the qth decoding layer as input features of the qth decoding layer. Skip connections, i.e., for two non-adjacent layers a and b, the output features F of layer a are... a The input is fed into layer b, so that the output feature F of layer a becomes... a The output feature F of layer b-1 will be compared with the layer preceding layer b. b-1 It also serves as the final input feature for layer b.

[0115] Figure 5 This is a schematic diagram of the structure of the second artificial intelligence module in an application example of the present invention. For example... Figure 5As shown, the second artificial intelligence module includes n coding layers C1-Cn and n-1 decoding layers D1-Dn-1, where n ≥ 2. The size of the coding layers decreases sequentially from C1 to Cn, while the size of the decoding layers increases sequentially from D1 to Dn-1. The size of the decoding layer Dn-1 is the same as that of the coding layer C1, ensuring that the length of the noise-cancelled audio signal is equal to the length of the audio signal before noise cancellation. Furthermore, the encoded features output from the nth coding layer Cn are output to the first decoding layer D1, and the encoded features output from the nqth coding layer Cn-q are output to the qth decoding layer Dq, where 1 ≤ q ≤ n-1.

[0116] It should be noted that the first and second artificial intelligence modules can also use the same artificial intelligence module. In this case, the input signal of the artificial intelligence module is two parallel signals: one is the frequency domain sound signal in step 402, and the other is the bandpass filtered frequency domain sound signal in step 404. The output signal of the artificial intelligence module can be two signals: a first clean frequency domain sound signal and a second clean frequency domain sound signal. Then, steps 406-407 are executed to finally obtain a clean time domain sound signal. Alternatively, the output signal of the artificial intelligence module can be only one clean frequency domain sound signal. This clean frequency domain sound signal can be directly converted from the frequency domain to the time domain to obtain a clean time domain sound signal. The artificial intelligence module can include: l coding layers, l-1 decoding layers, and one gain prediction layer, where l ≥ 2. Among them, the l-th coding layer and the o-th decoding layer are skipped, that is, the coding features output by the l-th coding layer can be input to the o-th decoding layer as input to the o-th decoding layer, where 1 ≤ o ≤ l-1.

[0117] Step 406: Mix the first clean audio signal in the frequency domain and the second clean audio signal in the frequency domain to obtain the third clean audio signal in the frequency domain.

[0118] Step 407: Convert the third clean sound signal in the frequency domain to the time domain to obtain a clean sound signal in the time domain.

[0119] In one optional embodiment, the ISTFT algorithm is used to convert the third clean sound signal in the frequency domain from the frequency domain to the time domain.

[0120] After obtaining the clean time-domain sound signal, the clean time-domain sound signal is transmitted to the MRI control room through the internal communication system of the MRI system, such as a walkie-talkie.

[0121] In the above embodiments, considering that noise in the MR imaging environment is mainly caused by the periodic frequency of the MR sequence, the time-domain sound signal acquired in the MR scanning room is first converted to the frequency domain. Then, a copy of the frequency-domain sound signal is made. One copy is filtered using a bandpass filter with a bandwidth equal to that of human hearing, utilizing the psychoacoustic characteristics of the frequency-domain sound signal, such as logarithmic response and frequency perception, to obtain a frequency-domain sound signal with a bandwidth equal to that of human hearing. Then, noise removal is performed on the frequency-domain sound signal with the bandwidth equal to that of human hearing to obtain the first clean frequency-domain sound signal. The other copy is directly input into the second artificial intelligence module for noise removal to obtain the second clean frequency-domain sound signal. The audio signal is processed to eliminate noise in the low and high frequency components of the frequency domain audio signal. Finally, the audio signals output from the two modules are mixed and converted to the time domain to obtain a clean time domain audio signal. This improves the voice quality of internal communication in MR imaging, increases MR scanning efficiency, and enhances the experience for patients and operators without adding extra system components or hardware costs. In addition, since noise cancellation is performed in the frequency domain, and noise in the MR imaging environment is mainly caused by the periodic frequency of the MR sequence, the voice enhancement method provided by this invention can filter out noise in the audio signal even when facing MR imaging environments using unknown MR sequences, and is not limited by the application scenario.

[0122] It should be noted that if the microphone is a multi-channel microphone, that is, the microphone can simultaneously acquire multiple time-domain audio signals, then each time-domain audio signal can be enhanced by the above steps 402-407.

[0123] Figure 6 A flowchart illustrating the training method for the first and second artificial intelligence modules provided in an embodiment of the present invention is shown below. Figure 6 As shown, the specific steps are as follows:

[0124] Step 601: Collect multiple training samples, wherein each training sample is obtained by mixing a clean speech signal collected with a segment of environmental noise collected during the operation of the MRI system in the MR scanning room, or by directly collecting a mixed sound signal of speech and environmental noise collected during the operation of the MRI system in the MR scanning room; the target sound signal corresponding to each training sample is: the clean speech signal corresponding to that training sample.

[0125] Step 602: Select a training sample and convert the selected training sample from the time domain to the frequency domain to obtain the frequency domain sound signal.

[0126] In one optional embodiment, after selecting a training sample and before converting the selected training sample from the time domain to the frequency domain, the method further includes: calculating the SNR of the time-domain audio signal; if the calculated SNR is less than a preset SNR threshold, then inputting the time-domain audio signal into a low-pass filter to obtain a low-pass filtered time-domain audio signal.

[0127] Step 603: Make a copy of the frequency domain audio signal.

[0128] Step 604: Input a frequency domain sound signal into a bandpass filter with a filtering bandwidth equal to that of human hearing to obtain a bandpass filtered frequency domain sound signal. Input the bandpass filtered frequency domain sound signal into the first artificial intelligence module for noise removal to obtain the predicted first clean frequency domain sound signal.

[0129] In one optional embodiment, the first artificial intelligence module includes: multiple encoding layers and multiple decoding layers, and the multiple encoding layers and multiple decoding layers are connected in a skip connection.

[0130] In one optional embodiment, the first artificial intelligence module includes: m encoding layers, m-1 decoding layers and 1 gain prediction layer, where m ≥ 2.

[0131] In one optional embodiment, the first to m-1 encoding layers and the first to m-1 decoding layers of the first artificial intelligence module are symmetrically arranged with respect to the m-th encoding layer, and the first to m-1 encoding layers and the (m-1)th to first decoding layers are respectively connected in a skip connection manner. For example, the mp-th encoding layer and the p-th decoding layer are connected in a skip connection manner, that is, the encoded features output by the mp-th encoding layer are output to the p-th decoding layer as the input features of the p-th decoding layer, 1≤p≤m-1.

[0132] In one optional embodiment, the bandpass filter with a filtering bandwidth equivalent to that of human hearing is an ERB filter.

[0133] Step 605: Input another frequency domain sound signal into the second artificial intelligence module for noise removal to obtain the predicted second clean frequency domain sound signal.

[0134] In one optional embodiment, the second artificial intelligence module includes: multiple encoding layers and multiple decoding layers, and the multiple encoding layers and multiple decoding layers are connected in a skip connection.

[0135] In one optional embodiment, the second artificial intelligence module includes: n encoding layers and n-1 decoding layers, where n≥2.

[0136] In one optional embodiment, the first to (n-1)th encoding layers and the first to (n-1)th decoding layers of the second artificial intelligence module are symmetrically arranged with respect to the nth encoding layer, and the first to (n-1)th encoding layers and the (n-1)th to (n-1)th decoding layers are respectively connected in a skip connection manner. For example, the nqth encoding layer and the qth decoding layer are connected in a skip connection, that is, the encoded features output by the nqth encoding layer are output to the qth decoding layer as the input features of the qth decoding layer, where 1≤q≤n-1.

[0137] Step 606: Mix the predicted first clean sound signal in the frequency domain and the predicted second clean sound signal in the frequency domain to obtain the predicted third clean sound signal in the frequency domain.

[0138] Step 607: Convert the predicted frequency domain third clean sound signal from the frequency domain to the time domain to obtain the predicted time domain clean sound signal.

[0139] Step 608: Compare the predicted clean sound signal in the time domain with the target sound signal corresponding to the input training sample, and adjust the parameters of the first artificial intelligence module and the second artificial intelligence module according to the comparison result.

[0140] Step 609: Repeat steps 602 to 608 until the first AI module and the second AI module converge. The converged first AI module and the second AI module are used as the final first AI module and the second AI module.

[0141] Figure 7 This is a schematic diagram of the structure of a speech enhancement device 70 in MR imaging according to an embodiment of the present invention, as shown below. Figure 7 As shown, the device 70 mainly includes: a time-frequency domain conversion module 71, a bandpass filter 72, a first artificial intelligence module 73, and a frequency-time domain conversion module 74, wherein:

[0142] The time-frequency domain conversion module 71 is used to: acquire time-domain sound signals in the MR scanning room in real time; convert the time-domain sound signals from the time domain to the frequency domain to obtain frequency-domain sound signals; and output the frequency-domain sound signals to the bandpass filter 72.

[0143] The bandpass filter 72 is used to perform bandpass filtering on the frequency domain sound signal and output the bandpass-filtered frequency domain sound signal to the first artificial intelligence module 73; wherein, the filtering bandwidth of the bandpass filter is equal to the filtering bandwidth of human hearing.

[0144] The first artificial intelligence module 73 is used to: eliminate noise from the bandpass filtered frequency domain sound signal to obtain a first clean frequency domain sound signal, and output the first clean frequency domain sound signal to the frequency-time domain conversion module 74.

[0145] The frequency-time domain conversion module 74 is used to convert the first clean audio signal in the frequency domain from the frequency domain to the time domain to obtain a clean audio signal in the time domain.

[0146] In one optional embodiment, the bandpass filter 72 is an ERB filter.

[0147] In one optional embodiment, the first artificial intelligence module includes: multiple encoding layers and multiple decoding layers, and the multiple encoding layers and multiple decoding layers are connected in a skip connection.

[0148] In one optional embodiment, the first artificial intelligence module 73 includes: m encoding layers, m-1 decoding layers and 1 gain prediction layer, where m ≥ 2. The first to m-1 encoding layers and the first to m-1 decoding layers of the first artificial intelligence module 73 are symmetrically arranged with respect to the m encoding layer, and the first to m-1 encoding layers and the m-1 to first decoding layers are respectively connected in a skip connection manner.

[0149] Figure 8 This is a schematic diagram of the structure of a speech enhancement device 80 in MR imaging according to another embodiment of the present invention, as shown below. Figure 8 As shown, with Figure 7 Compared to the device 70 shown, Figure 8 The device 80 shown further includes: a second artificial intelligence module 75;

[0150] Furthermore, after the time-frequency domain conversion module 71 obtains the frequency domain sound signal and before outputting the frequency domain sound signal to the bandpass filter 72, it further includes: making a copy of the frequency domain sound signal;

[0151] Furthermore, the time-frequency domain conversion module 71 outputs the frequency domain audio signal to the bandpass filter 72, including: outputting a frequency domain audio signal to the bandpass filter 72;

[0152] Furthermore, while the time-frequency domain conversion module 71 outputs the frequency domain audio signal to the bandpass filter 72, it also includes: outputting another frequency domain audio signal to the second artificial intelligence module 75;

[0153] The second artificial intelligence module 75 is used to: perform noise removal on the frequency domain sound signal to obtain a second clean sound signal in the frequency domain, and output the second clean sound signal in the frequency domain to the frequency-time domain conversion module 74.

[0154] The frequency-time domain conversion module 74 converts the first clean sound signal in the frequency domain from the frequency domain to the time domain to obtain a clean sound signal in the time domain, including: mixing the first clean sound signal in the frequency domain and the second clean sound signal in the frequency domain to obtain a third clean sound signal in the frequency domain, and converting the third clean sound signal in the frequency domain from the frequency domain to the time domain to obtain a clean sound signal in the time domain.

[0155] In one optional embodiment, the second artificial intelligence module 75 includes: multiple encoding layers and multiple decoding layers, and the multiple encoding layers and multiple decoding layers are connected in a skip connection.

[0156] In one optional embodiment, the second artificial intelligence module 75 includes: n encoding layers and n-1 decoding layers, where n≥2. The first to n-1 encoding layers and the first to n-1 decoding layers of the second artificial intelligence module 75 are symmetrically arranged with respect to the nth encoding layer, and the first to n-1 encoding layers and the n-1 to n-1 decoding layers are respectively connected in a skip connection manner.

[0157] Figure 9 This is a schematic diagram of the structure of a speech enhancement device 90 in MR imaging according to another embodiment of the present invention, as shown below. Figure 9 As shown, with Figure 8 Compared to the device 80 shown, the time-frequency domain conversion module 71 mainly includes: a signal-to-noise ratio calculation and judgment module 711, a low-pass filter 712, and a conversion module 713, wherein:

[0158] The signal-to-noise ratio (SNR) calculation and judgment module 711 is used to: acquire time-domain sound signals in the MR scanning room in real time, calculate the SNR of the time-domain sound signals, and if the calculated SNR is less than the preset SNR threshold, input the time-domain sound signals to the low-pass filter 712; otherwise, directly output the time-domain sound signals acquired by the microphone to the conversion module 713.

[0159] The low-pass filter 712 is used to perform low-pass filtering on the time-domain audio signal and output the low-pass filtered time-domain audio signal to the conversion module 713.

[0160] The conversion module 713 is used to: convert the time-domain audio signal from the confidence-to-noise ratio calculation and judgment module 711 from the time domain to the frequency domain to obtain a frequency-domain audio signal, or convert the low-pass filtered time-domain audio signal from the low-pass filter 712 from the time domain to the frequency domain to obtain a frequency-domain audio signal; copy the frequency-domain audio signal, output one copy of the frequency-domain audio signal to the band-pass filter 72, and output the other copy of the frequency-domain audio signal to the second artificial intelligence module 75.

[0161] In an optional embodiment, the above-described apparatus 80 or 90 further includes: a training module, used for:

[0162] A2. Collect multiple training samples, wherein each training sample is obtained by mixing a clean speech signal collected with a segment of environmental noise collected during the operation of the MRI system in the MR scanning room, or by directly collecting a mixed sound signal of speech and environmental noise collected during the operation of the MRI system in the MR scanning room; the target sound signal corresponding to each training sample is: the clean speech signal corresponding to that training sample.

[0163] B2. Select a training sample and convert it from the time domain to the frequency domain to obtain a frequency domain sound signal. Make a copy of the frequency domain sound signal. Input the copy of the frequency domain sound signal into a bandpass filter with a filtering bandwidth equal to that of human hearing to obtain a bandpass-filtered frequency domain sound signal. Input the bandpass-filtered frequency domain sound signal into the first artificial intelligence module for noise removal to obtain a predicted first clean frequency domain sound signal. Input another copy of the frequency domain sound signal into the second artificial intelligence module for noise removal to obtain a predicted second clean frequency domain sound signal. Mix the predicted first and second clean frequency domain sound signals to obtain a predicted third clean frequency domain sound signal. Convert the predicted third clean frequency domain sound signal from the frequency domain to the time domain to obtain a predicted clean time domain sound signal. Compare the predicted clean time domain sound signal with the target sound signal corresponding to the input training sample, and adjust the parameters of the first and second artificial intelligence modules based on the comparison results.

[0164] Repeat step B2 until the first AI module and the second AI module converge. The converged first AI module and the second AI module are then used as the final first AI module and the second AI module.

[0165] This invention also provides an MRI system that includes a speech enhancement device 70, 80, or 90 for MR imaging as described above.

[0166] It should be noted that the speech enhancement method and apparatus in MR imaging and the MRI system provided in the embodiments of the present invention can be methods, apparatus and systems applied in medical imaging.

[0167] This invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the speech enhancement method in MR imaging as described in any of the above embodiments.

[0168] This invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, can perform the steps of the speech enhancement method in MR imaging as described above. In practical applications, the computer-readable medium may be included in the devices / apparatus / systems described above, or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium stores instructions that, when executed by a processor, can perform the steps of the speech enhancement method in MR imaging as described above.

[0169] This invention also provides an electronic device. The electronic device may include a processor with one or more processing cores, a memory of one or more computer-readable storage media, and a computer program stored in the memory and executable on the processor. When the program in the memory is executed, the aforementioned speech enhancement method in MR imaging can be implemented.

[0170] Those skilled in the art will understand that the features described in the various embodiments and / or claims disclosed in this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, without departing from the spirit and teachings of this application, the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways, and all such combinations and / or combinations fall within the scope of this application.

[0171] This document uses specific embodiments to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and core ideas of this application, and are not intended to limit this application. For those skilled in the art, changes can be made to the specific implementation methods and application scope based on the ideas, spirit and principles of this application. Any modifications, equivalent substitutions, improvements, etc., made should be included within the scope of protection of this application.

Claims

1. A speech enhancement method in magnetic resonance imaging, characterized in that, The method includes: Real-time acquisition of time-domain sound signals within the MR scanning chamber; The time-domain sound signal is converted from the time domain to the frequency domain to obtain a frequency-domain sound signal; The frequency domain sound signal is input into a bandpass filter with a filtering bandwidth equal to that of human hearing to obtain a bandpass filtered frequency domain sound signal. The bandpass filtered frequency domain audio signal is input into a pre-trained first artificial intelligence module for noise removal to obtain a first clean frequency domain audio signal. The first clean audio signal in the frequency domain is converted from the frequency domain to the time domain to obtain the clean audio signal in the time domain.

2. The method according to claim 1, characterized in that, The first artificial intelligence module includes multiple encoding layers and multiple decoding layers, and the multiple encoding layers and multiple decoding layers are connected in a skip connection.

3. The method according to claim 1, characterized in that, The first artificial intelligence module includes: m encoding layers, m-1 decoding layers, and 1 gain prediction layer, where m ≥ 2; The first to m-1 encoding layers and the first to m-1 decoding layers of the first artificial intelligence module are symmetrically arranged with respect to the m-th encoding layer, and the first to m-1 encoding layers and the m-1 to first decoding layers are respectively connected in a skip connection manner.

4. The method according to claim 1, characterized in that, After the real-time acquisition of the time-domain sound signal inside the MR scanning room and before the conversion of the time-domain sound signal from the time domain to the frequency domain, the process further includes: Calculate the signal-to-noise ratio (SNR) of the time-domain audio signal; if the SNR is less than a preset SNR threshold, input the time-domain audio signal into a low-pass filter to obtain a low-pass filtered time-domain audio signal. The step of converting the time-domain audio signal from the time domain to the frequency domain includes: The low-pass filtered time-domain audio signal is converted from the time domain to the frequency domain.

5. The method according to claim 1, characterized in that, The bandpass filter whose filtering bandwidth is equivalent to that of human hearing is called an Equivalent Rectangular Bandwidth (ERB) filter.

6. The method according to any one of claims 1 to 5, characterized in that, The first artificial intelligence module is obtained through the following training process: A1. Collect multiple training samples, wherein each training sample is obtained by mixing a clean speech signal collected with a segment of environmental noise collected during the operation of the MRI system in the MR scanning room, or by directly collecting a mixed sound signal of speech and environmental noise collected during the operation of the MRI system in the MR scanning room; the target sound signal corresponding to each training sample is: the clean speech signal corresponding to that training sample. B1. Select a training sample and convert it from the time domain to the frequency domain to obtain a frequency domain sound signal. Input the frequency domain sound signal into a bandpass filter with a filtering bandwidth equal to that of human hearing to obtain a bandpass-filtered frequency domain sound signal. Input the bandpass-filtered frequency domain sound signal into the first artificial intelligence module for noise removal to obtain a predicted first clean frequency domain sound signal. Convert the predicted first clean frequency domain sound signal from the frequency domain to the time domain to obtain a predicted clean time domain sound signal. Compare the predicted clean time domain sound signal with the target sound signal corresponding to the input training sample and adjust the parameters of the first artificial intelligence module according to the comparison result. Repeat step B1 until the first AI module converges, and use the converged first AI module as the final first AI module.

7. The method according to claim 1, characterized in that, After obtaining the frequency domain sound signal, and before inputting the frequency domain sound signal into a bandpass filter with a filtering bandwidth equal to that of human hearing, the process further includes: Make a copy of the frequency domain sound signal; Furthermore, the step of inputting the frequency domain sound signal into a bandpass filter with a filtering bandwidth equal to that of human hearing includes: A portion of the frequency domain sound signal is input into a bandpass filter with a filtering bandwidth equal to that of human hearing. Furthermore, the step of inputting the frequency domain sound signal into a bandpass filter with a filtering bandwidth equal to that of human hearing further includes: Another frequency domain sound signal is input into a pre-trained second artificial intelligence module for noise removal to obtain a second clean frequency domain sound signal. The step of converting the first clean audio signal in the frequency domain from the frequency domain to the time domain to obtain a clean audio signal in the time domain includes: The first clean audio signal in the frequency domain and the second clean audio signal in the frequency domain are mixed to obtain the third clean audio signal in the frequency domain. The third clean audio signal in the frequency domain is then converted from the frequency domain to the time domain to obtain the clean audio signal in the time domain.

8. The method according to claim 7, characterized in that, The second artificial intelligence module includes multiple encoding layers and multiple decoding layers, with skip connections between the multiple encoding layers and multiple decoding layers.

9. The method according to claim 7, characterized in that, The second artificial intelligence module includes: n encoding layers and n-1 decoding layers, where n≥2; Furthermore, the first to n-1 encoding layers and the first to n-1 decoding layers of the second artificial intelligence module are symmetrically arranged with respect to the nth encoding layer, and the first to n-1 encoding layers and the n-1 to first decoding layers are respectively connected in a skip connection manner.

10. The method according to any one of claims 7 to 9, characterized in that, The first artificial intelligence module and the second artificial intelligence module are obtained through the following training process: A2. Collect multiple training samples, wherein each training sample is obtained by mixing a clean speech signal collected with a segment of environmental noise collected during the operation of the MRI system in the MR scanning room, or by directly collecting a mixed sound signal of speech and environmental noise collected during the operation of the MRI system in the MR scanning room; the target sound signal corresponding to each training sample is: the clean speech signal corresponding to that training sample. B2. Select a training sample and convert it from the time domain to the frequency domain to obtain a frequency domain sound signal. Make a copy of the frequency domain sound signal. Input one copy of the frequency domain sound signal into a bandpass filter with a filtering bandwidth equal to that of human hearing to obtain a bandpass-filtered frequency domain sound signal. Input the bandpass-filtered frequency domain sound signal into the first artificial intelligence module for noise removal to obtain a predicted first clean frequency domain sound signal. Input another copy of the frequency domain sound signal into the second artificial intelligence module for noise removal to obtain a predicted second clean frequency domain sound signal. Mix the predicted first and second clean frequency domain sound signals to obtain a predicted third clean frequency domain sound signal. Convert the predicted third clean frequency domain sound signal from the frequency domain to the time domain to obtain a predicted clean time domain sound signal. Compare the predicted clean time domain sound signal with the target sound signal corresponding to the input training sample, and adjust the parameters of the first and second artificial intelligence modules according to the comparison result. Repeat step B2 until the first AI module and the second AI module converge. The converged first AI module and the second AI module are then used as the final first AI module and the second AI module.

11. A speech enhancement device for magnetic resonance imaging, characterized in that, The device includes: The time-frequency domain conversion module is used for: real-time acquisition of time-domain sound signals in the magnetic resonance MR scanning chamber; conversion of the time-domain sound signals from the time domain to the frequency domain to obtain frequency-domain sound signals; and outputting the frequency-domain sound signals to a bandpass filter. A bandpass filter is used to perform bandpass filtering on frequency domain audio signals and output the bandpass-filtered frequency domain audio signals to the first artificial intelligence module; wherein, the filtering bandwidth of the bandpass filter is equal to the filtering bandwidth of human hearing; The first artificial intelligence module is used to: perform noise removal on the bandpass filtered frequency domain audio signal to obtain the first clean frequency domain audio signal, and output the first clean frequency domain audio signal to the frequency-time domain conversion module; The frequency-time domain conversion module is used to convert the first clean audio signal in the frequency domain to the time domain, thereby obtaining a clean audio signal in the time domain.

12. The apparatus according to claim 11, characterized in that, The first artificial intelligence module includes multiple encoding layers and multiple decoding layers, and the multiple encoding layers and multiple decoding layers are connected in a skip connection.

13. The apparatus according to claim 11, characterized in that, The first artificial intelligence module includes: m encoding layers, m-1 decoding layers, and 1 gain prediction layer, where m ≥ 2; The first to m-1 encoding layers and the first to m-1 decoding layers of the first artificial intelligence module are symmetrically arranged with respect to the m-th encoding layer, and the first to m-1 encoding layers and the m-1 to first decoding layers are respectively connected in a skip connection manner.

14. The apparatus according to claim 11, characterized in that, The time-frequency domain conversion module includes: a signal-to-noise ratio calculation and judgment module, a low-pass filter, and a conversion module, wherein: The signal-to-noise ratio (SNR) calculation and judgment module is used to: acquire time-domain sound signals in the MR scanning room in real time, calculate the SNR of the time-domain sound signals, and if the SNR is less than a preset SNR threshold, input the time-domain sound signals into a low-pass filter; otherwise, directly output the time-domain sound signals acquired by the microphone to the conversion module. A low-pass filter is used to: perform low-pass filtering on the time-domain audio signal and output the low-pass filtered time-domain audio signal to the conversion module; The conversion module is used to: convert the time-domain audio signal from the future information-to-noise ratio calculation and judgment module from the time domain to the frequency domain to obtain a frequency-domain audio signal, or convert the low-pass filtered time-domain audio signal from the low-pass filter from the time domain to the frequency domain to obtain a frequency-domain audio signal; and output the frequency-domain audio signal to the band-pass filter.

15. The apparatus according to claim 11, characterized in that, The device further includes: a second artificial intelligence module; After the time-frequency domain conversion module obtains the frequency domain sound signal and before outputting the frequency domain sound signal to the bandpass filter, it further includes: making a copy of the frequency domain sound signal; Furthermore, the time-frequency domain conversion module outputs the frequency domain audio signal to the bandpass filter, including: outputting a copy of the frequency domain audio signal to the bandpass filter; Furthermore, while the time-frequency domain conversion module outputs the frequency domain audio signal to the bandpass filter, it further includes: outputting another copy of the frequency domain audio signal to the second artificial intelligence module; The second artificial intelligence module is used to: perform noise removal on the frequency domain sound signal to obtain a second clean sound signal in the frequency domain, and output the second clean sound signal in the frequency domain to the frequency-time domain conversion module; The frequency-time domain conversion module converts the first clean sound signal in the frequency domain from the frequency domain to the time domain to obtain a clean sound signal in the time domain. This includes: mixing the first clean sound signal in the frequency domain and the second clean sound signal in the frequency domain to obtain a third clean sound signal in the frequency domain, and converting the third clean sound signal in the frequency domain from the frequency domain to the time domain to obtain a clean sound signal in the time domain.

16. A magnetic resonance imaging system, characterized in that, The magnetic resonance imaging system includes a speech enhancement device for magnetic resonance imaging as described in any one of claims 11 to 15.

Images