MMPI mixed signal separation method based on self-attention mechanism

Abstract

The present invention belongs to the field of biomedical molecular imaging, and specifically relates to a MMPI mixed signal separation method, system, and device based on a self-attention mechanism, which aims to solve the problem that the separation error of the mixed SPIOs separation model constructed based on physical characteristics in the existing MMPI technology is large and difficult to quantify. The method comprises: constructing a biological lesion model of an organism to be imaged by MMPI, and injecting a targeted molecular probe wrapped with SPIOs into the biological lesion model; after injection, using an MPI device to collect signals from the biological lesion model to obtain a time domain mixed signal containing n kinds of SPIOs; separating the time domain mixed signal through a trained self-attention mechanism neural network to obtain n kinds of separate SPIOs time domain voltage signals. The present invention realizes the precise separation of the time domain mixed signal and improves the detection sensitivity of MMPI.

Description

Technical Field

[0001] The present invention belongs to the field of biomedical molecular imaging, and specifically relates to a MMPI mixed signal separation method, system and device based on a self-attention mechanism. Background Art

[0002] In recent years, magnetic particle imaging (MPI) has developed rapidly as a new type of molecular imaging technology and has become an international research hotspot. Its basic principle is to use the nonlinear response remagnetization of superparamagnetic iron oxide particles (SPIOs) in the non-magnetic field region in a high gradient magnetic field to quantitatively obtain the three-dimensional concentration distribution of magnetic particles in the body with high sensitivity and specificity. The intensity of the response voltage signal collected by MPI is proportional to the concentration of SPIOs. Through modeling analysis, the concentration of SPIOs in the body can be inferred from the collected voltage signal. At present, MPI has been successfully applied in tumor cell tracing, cardiovascular and cerebrovascular imaging, and neuroimaging. Multi-color magnetic particle imaging (MMPI) is a further extension of the traditional MPI imaging function. Inspired by multispectral optical imaging technology, MMPI can simultaneously perform quantitative visualization analysis of multiple SPIOs in the same field of view based on the different physical properties of different types of SPIOs under alternating magnetic fields, which is expected to provide a powerful technical means for the realization of multi-molecule accurate detection of tumors.

[0003] In order to perform multi-color imaging of multiple SPIOs, it is necessary to first accurately distinguish multiple SPIOs, and then reconstruct the image to obtain a multi-color imaging display representing different SPIOs. Therefore, the accuracy of distinguishing multiple SPIOs is crucial for the final reconstructed image. At present, MMPI technology is based on the physical properties of SPIOs for distinction, including spectral characteristics and relaxation characteristics. The former draws on the principle of multi-spectral imaging technology and directly distinguishes the received mixed magnetic particle response voltage in the frequency domain according to the different spectral information of each SPIOs. The mixed SPIOs separation model based on relaxation characteristics performs feature mapping according to the different relaxation effects of different SPIOs and directly generates multi-color relaxation maps. However, in practical applications, the separation model based on spectral characteristics will be affected by the mutual interference between multiple SPIOs, resulting in spectral aliasing problems, which leads to large distinction errors; the separation model based on relaxation mapping can only reflect the differences in relaxation effects between different SPIOs, but cannot achieve quantitative research on mixed magnetic particles. This problem seriously affects the detection sensitivity of MMPI and will greatly limit the application prospects of MMPI. Therefore, in order to further expand the application scope of MMPI imaging in biomedical research, it is necessary to study an accurate and robust hybrid SPIOs separation model to improve the detection sensitivity of MMPI. Summary of the invention

[0004] In order to solve the above problems in the prior art, that is, to solve the problem that the separation error of the mixed SPIOs separation model constructed based on physical characteristics in the existing MMPI technology is large and difficult to quantify, the first aspect of the present invention proposes an MMPI mixed signal separation method based on a self-attention mechanism, which is used to separate a time domain mixed signal containing n kinds of SPIOs, and then perform MMPI imaging based on the separated n kinds of separate SPIOs time domain voltage signals, and the method comprises the following steps:

[0005] S100, constructing a biological lesion model of a living body to be imaged by MMPI, and injecting a targeting molecular probe encapsulating SPIOs into the biological lesion model;

[0006] S200, after the injection, using an MPI device to collect signals from the biological lesion model to obtain a time-domain mixed signal containing n types of SPIOs; wherein n represents a preset number;

[0007] S300, separating the time domain mixed signal through a trained self-attention mechanism neural network to obtain n separate SPIOs time domain voltage signals;

[0008] Wherein, the self-attention mechanism neural network includes a time domain encoder, a signal separator and a time domain decoder;

[0009] The time domain encoder is configured to divide the time domain mixed signal into r sequence segments u of length L i , and then encode it into a D-dimensional vector representation through a one-dimensional convolution i = 1, 2, ... r, i.e., eigenvector;

[0010] The signal separator is configured to concatenate the encoded feature vector with its position code to obtain a concatenated vector; in combination with a preset mask vector, the concatenated vector is separated by a double-layer self-attention unit to obtain source signal vectors of n types of magnetic particles;

[0011] The self-attention unit includes a plurality of attention blocks; each attention block is constructed based on a self-attention layer, a residual connection, a layer normalization, a feedforward neural network, a residual connection and a layer normalization connected in sequence;

[0012] The time domain decoder is configured to convert the source signal vector into a feature vector w after the time domain mixed signal is encoded. 1 Multiply and decode the multiplied vector to obtain n separate SPIOs time-domain voltage signals; among them,

[0013] In some preferred embodiments, the position code is obtained by:

[0014] PE(i,2j)=sin(i / 10000 2j / D )

[0015] PE(i,2j+1)=cos(i / 10000 2j / D )

[0016] Where i represents u i In the time domain mixed signal u mul The relative position in, j represents u i in U T The corresponding index in PE() stands for position encoding.

[0017] In some preferred embodiments, the attention block processes the input vector as follows:

[0018] The correlation between different parts of the input vector is learned through the self-attention layer. The learning process is:

[0019]

[0020] Q=w 2 W Q

[0021] K=w 2 W K

[0022] V=w 2 W V

[0023] Among them, W Q , W K , W V represents the randomly initialized matrix, w 2 Denotes the feature vector w 1 The vector concatenated with the position code, d k Represents the vector dimensions of Q, K, and V;

[0024] Based on the vector output after learning from the self-attention layer, combined with the concatenated vector, residual connection and layer normalization are performed, and the processed vector is used as the first vector;

[0025] Input the first vector into a feedforward neural network to obtain a second vector;

[0026] Based on the second vector and combined with the first vector, a residual connection is performed, and layer normalization is performed, and the processed vector is used as the output of the self-attention block.

[0027] In some preferred embodiments, the concatenated vectors are separated by a double-layer self-attention unit in combination with a preset mask vector to obtain source signal vectors of n types of magnetic particles, and the method is as follows:

[0028] The concatenated vectors are processed by a two-layer self-attention unit, and the processed r vectors are concatenated in sequence to obtain a one-dimensional mixed signal vector representation;

[0029] The preset n mask vectors are multiplied by the mixed signal vector representation to obtain the source signal vectors of n magnetic particles, that is,

[0030] b q =w 3 ·m q

[0031] in, q=1,2,…,n,m q represents the mask vector, w 3 represents the mixed signal vector representation, b q Represents the source signal vector.

[0032] In some preferred embodiments, the self-attention mechanism neural network is trained by:

[0033] A100, constructing a training set; the training samples in the training set include a time domain mixed signal obtained by adding multiple response voltage signals and its corresponding original true value source signal label; the response voltage signal is a time domain voltage signal;

[0034] A200, separates the time-domain mixed signal through a pre-built self-attention mechanism neural network to obtain a separate SPIOs time-domain voltage signal as a prediction signal;

[0035] A300, based on the predicted signal and the original true source signal label, calculates the loss value through the pre-built loss function and updates the network parameters of the self-attention mechanism neural network;

[0036] A400, loop A100-A300 until a trained self-attention mechanism neural network is obtained.

[0037] In some preferred embodiments, the response voltage signal in the training sample is obtained by:

[0038] The digital mouse was used as a standard imaging space, in which the distribution of SPIOs of different positions, types, and concentrations was set;

[0039] The Debye simulation algorithm is used to simulate the SPIOs distribution and obtain its corresponding response voltage signal.

[0040] In some preferred embodiments, the loss function used in the training of the self-attention mechanism neural network is a function that maximizes the scale-invariant signal-to-noise ratio, specifically:

[0041]

[0042]

[0043]

[0044] in, and u are the predicted signal and the original true source signal label respectively, and SI-SNR represents the signal-to-noise ratio function that maximizes the scale-invariant.

[0045] In a second aspect of the present invention, a MMPI mixed signal separation system based on a self-attention mechanism is proposed, which is used to separate a time-domain mixed signal containing n types of SPIOs, and then perform MMPI imaging based on the separated n types of individual SPIOs time-domain voltage signals. The system includes: a lesion model construction module, a signal acquisition module, and a signal separation module;

[0046] The lesion model construction module is configured to construct a biological lesion model of an organism to be imaged by MMPI, and inject the targeted molecular probes encapsulating SPIOs into the biological lesion model;

[0047] The signal acquisition module is configured to collect signals from the biological lesion model using an MPI device after injection to obtain a time-domain mixed signal containing n types of SPIOs; wherein n represents a preset number;

[0048] The signal separation module is configured to separate the time domain mixed signal through a trained self-attention mechanism neural network to obtain n separate SPIOs time domain voltage signals;

[0049] Wherein, the self-attention mechanism neural network includes a time domain encoder, a signal separator and a time domain decoder;

[0050] The time domain encoder is configured to divide the time domain mixed signal into r sequence segments u of length L i , and then encode it into a D-dimensional vector representation through a one-dimensional convolution i = 1, 2, ... r, i.e., eigenvector;

[0051] The signal separator is configured to concatenate the encoded feature vector with its position code to obtain a concatenated vector; in combination with a preset mask vector, the concatenated vector is separated by a double-layer self-attention unit to obtain source signal vectors of n types of magnetic particles;

[0052] The self-attention unit includes a plurality of attention blocks; each attention block is constructed based on a self-attention layer, a residual connection, a layer normalization, a feedforward neural network, a residual connection and a layer normalization connected in sequence;

[0053] The time domain decoder is configured to convert the source signal vector into a feature vector w after the time domain mixed signal is encoded. 1 Multiply and decode the multiplied vector to obtain n separate SPIOs time-domain voltage signals; among them,

[0054] According to a third aspect of the present invention, an electronic device is provided, comprising at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the above-mentioned MMPI mixed signal separation method based on the self-attention mechanism.

[0055] In a fourth aspect of the present invention, a computer-readable storage medium is proposed, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to be executed by the computer to implement the above-mentioned MMPI mixed signal separation method based on the self-attention mechanism.

[0056] Beneficial effects of the present invention:

[0057] The present invention realizes accurate separation of time domain mixed signals and improves the detection sensitivity of MMPI.

[0058] 1) The present invention is based on the machine learning theory of statistical learning and uses neural networks to directly extract and analyze the features of mixed signals, ultimately achieving accurate separation of mixed signals.

[0059] 2) The present invention constructs an algorithm model from the perspective of signal processing, avoids the influence of the physical properties of SPIOs on the separation accuracy, and improves the separation accuracy of mixed SPIOs;

[0060] 3) The present invention uses Debye simulation to construct a training and verification sample set, and uses sample combination to construct a variety of SPIOs mixed samples, thereby improving the generalization ability of the self-attention network. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings.

[0062] Figure 1 It is a flowchart of a MMPI mixed signal separation method based on a self-attention mechanism according to an embodiment of the present invention;

[0063] Figure 2 It is a schematic diagram of a framework of an MMPI mixed signal separation system based on a self-attention mechanism according to an embodiment of the present invention;

[0064] Figure 3 It is a flowchart of the training process and application process of the self-attention mechanism neural network according to an embodiment of the present invention;

[0065] Figure 4 is a detailed schematic diagram of a neural network structure of a self-attention mechanism according to an embodiment of the present invention;

[0066] Figure 5 It is a structural diagram of a computer system of an embodiment of the present invention that is suitable for implementing an electronic device of an embodiment of the present application. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.

[0069] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application may be combined with each other.

[0070] The present invention provides a MMPI mixed signal separation method based on self-attention mechanism, such as Figure 1 As shown, for separating a time-domain mixed signal containing n kinds of SPIOs, and then performing MMPI imaging based on the separated n kinds of individual SPIOs time-domain voltage signals, the method comprises the following steps:

[0071] S100, constructing a biological lesion model of a living body to be imaged by MMPI, and injecting a targeting molecular probe encapsulating SPIOs into the biological lesion model;

[0072] S200, after the injection, using an MPI device to collect signals from the biological lesion model to obtain a time-domain mixed signal containing n types of SPIOs; wherein n represents a preset number;

[0073] S300, separating the time domain mixed signal through a trained self-attention mechanism neural network to obtain n separate SPIOs time domain voltage signals;

[0074] Wherein, the self-attention mechanism neural network includes a time domain encoder, a signal separator and a time domain decoder;

[0075] The time domain encoder is configured to divide the time domain mixed signal into r sequence segments u of length L i , and then encode it into a D-dimensional vector representation through a one-dimensional convolution i = 1, 2, ... r, i.e., eigenvector;

[0076] The signal separator is configured to concatenate the encoded feature vector with its position code to obtain a concatenated vector; in combination with a preset mask vector, the concatenated vector is separated by a double-layer self-attention unit to obtain source signal vectors of n types of magnetic particles;

[0077] The self-attention unit includes a plurality of attention blocks; each attention block is constructed based on a self-attention layer, a residual connection, a layer normalization, a feedforward neural network, a residual connection and a layer normalization connected in sequence;

[0078] The time domain decoder is configured to convert the source signal vector into a feature vector w after the time domain mixed signal is encoded. 1 Multiply and decode the multiplied vector to obtain n separate SPIOs time-domain voltage signals; among them,

[0079] In order to more clearly illustrate the MMPI mixed signal separation method based on the self-attention mechanism of the present invention, each step in an embodiment of the method of the present invention is described in detail below with reference to the accompanying drawings.

[0080] In the following embodiments, the training process of the self-attention mechanism neural network is first described, and then the process of obtaining a separate SPIOs time-domain voltage signal by the MMPI mixed signal separation method based on the self-attention mechanism is described in detail. Figure 3 As shown, the details are as follows:

[0081] 1. The training process of the self-attention mechanism neural network

[0082] A100, constructing a training set; the training samples in the training set include a time domain mixed signal obtained by adding multiple response voltage signals and its corresponding original true value source signal label; the response voltage signal is a time domain voltage signal;

[0083] In this embodiment, the specific process of obtaining training samples is as follows:

[0084] A110, using the digital mouse as the standard imaging space, set up SPIOs distributions of different positions, types, and concentrations in it; using the Debye simulation algorithm to simulate the SPIOs distribution to obtain its corresponding response voltage signal (i.e. Figure 3 A single type of SPIOs sample set); the response voltage signal is a time domain voltage signal;

[0085] A120, randomly select n response voltage signals (preferably 2, 3 or 4 in the present invention) from the generated response voltage signals and add them to obtain a time domain mixed signal, and add the actual SPIOs distribution corresponding to the response voltage signals as the original true value source signal label; specifically:

[0086]

[0087]

[0088] Among them, x mul and u mul are the true distributions of the added SPIOs, i.e., the original true source signal labels and the time domain mixed signals, s is the sample number, S n is the set of n selected samples, x s and u s are the true SPIO distribution information and response voltage signal of the sth sample respectively.

[0089] A130, constructing training samples based on the time domain mixed signal and its corresponding original true value source signal label, thereby obtaining a training set (i.e. Figure 3 The training sample set of multiple SPIOs mixed in );

[0090] A200, separates the time-domain mixed signal through a pre-built self-attention mechanism neural network to obtain a separate SPIOs time-domain voltage signal as a prediction signal;

[0091] In this embodiment, the self-attention mechanism neural network includes a time domain encoder, a signal separator (i.e. Figure 4 Signal separation module in) and time domain decoder;

[0092] The calculation process of the time domain encoder is:

[0093] First, the mixed time domain signal is converted into a corresponding representation in the feature space through a linear encoder. The time domain encoder is essentially a one-dimensional convolution layer. The application of convolution allows the time domain encoder to process consecutive frames or fragments in parallel, thereby speeding up the separation process and greatly compressing the model parameters. Specifically:

[0094] The input time domain mixed signal u mul is divided into r sequence segments of length L, using i=1,2,...,r. Then u is converted into i Convert to D-dimensional vector representation As shown in formula (3):

[0095]

[0096] in, Contains D vectors (encoder basis functions), each of which has a length of L. In this way, the time domain signal u mul is encoded into a D-dimensional vector with r Features

[0097] The signal separator realizes the separation of sequence signals through a double-layer self-attention unit and a mask, and its calculation process is:

[0098] First, the encoded feature vector is concatenated with its positional encoding to obtain the concatenated vector, which is input into the self-attention layer, where the positional encoding is as follows:

[0099] PE(i,2j)=sin(i / 10000 2j / D ) (4)

[0100] PE(i,2j+1)=cos(i / 10000 2j / D ) (5)

[0101] Where i represents u i In the time domain mixed signal u mul The relative position in, j represents u i in U T The corresponding index in PE() stands for position encoding.

[0102] By concatenating the encoded feature vector with its positional encoding, the two-layer self-attention layer (i.e., the self-attention model) can fully consider the relative positional relationship within the signal sequence during the learning process. Another task of the self-attention unit is to learn the correlation between different parts of the entire input. Figure 4 As shown, the self-attention unit includes multiple attention blocks; each attention block is constructed based on the sequentially connected self-attention layer, residual connection, layer normalization, feedforward neural network, residual connection and layer normalization.

[0103] The attention block processes the input vector as follows:

[0104] The correlation between different parts of the input vector is learned through the self-attention layer. The learning process is:

[0105] Q=w 2 W Q (6)

[0106] K=w 2 W K (7)

[0107] V=w 2 W V (8)

[0108]

[0109] Among them, W Q , W K , W Vrepresents the random initialization matrix, the purpose of which is to make each position contain the information of all positions in the current signal, w 2 Represents the concatenated vector, i.e., the eigenvector w 1 The result after concatenation with the position code, d k Represents the vector dimensions of Q, K, and V.

[0110] Based on the vector output after learning from the self-attention layer, combined with the concatenated vector, residual connection and layer normalization are performed, and the processed vector is used as the first vector to prevent gradient disappearance and speed up the model convergence.

[0111] The first vector is input into a feedforward neural network to obtain a second vector; wherein the feedforward neural network introduces nonlinearity to enhance the expressive power of the model.

[0112] Based on the second vector, combined with the first vector, residual connection is performed, and layer normalization is performed (i.e. Figure 4 Add and Norm in , and use the processed vector as the output of the self-attention block.

[0113] The concatenated vector is processed based on the above-mentioned two-layer self-attention unit, and the processed r vectors are concatenated in sequence to obtain a one-dimensional mixed signal vector representation.

[0114] Finally, the signal separation is achieved by estimating n mask vectors (in addition, before and after the combined mask separation, two activation functions ReLu and Sigmoid are also used for processing). q=1,2,…,n,m q ∈[0,1]. q The mixed signal vector representation w of the self-attention layer output 3 Multiply them to obtain the source signal vector representation of n kinds of magnetic particles:

[0115] b q =w 3 ·m q (10)

[0116] Among them, m q represents the mask vector, b q Represents the source signal vector.

[0117] The time domain decoder module restores the source signal vector obtained by the separator to the source time domain signal, that is, the source signal vector and the time domain mixed signal are encoded into the feature vector w 1 Multiply and decode the multiplied vector to obtain n separate SPIOs time-domain voltage signals; expressed by matrix multiplication:

[0118]

[0119] in, Refers to the separated single source timing signal, that is, the separate SPIOs time domain voltage signal, It consists of the basis functions of the decoder.

[0120] A300, based on the predicted signal and the original true source signal label, calculates the loss value through the pre-built loss function and updates the network parameters of the self-attention mechanism neural network;

[0121] In this embodiment, the optimization function of the self-attention mechanism neural network is the Adam optimization method, and the loss function of the network training adopts the maximum scale-invariant signal-to-noise ratio function (SI-SNR). This function is defined as:

[0122]

[0123]

[0124]

[0125] in, and u are the predicted signal and the original true source signal labels respectively, and SI-SNR represents the signal-to-noise ratio function that maximizes the scale invariance. Before calculation, and u are normalized to zero mean to ensure scale invariance.

[0126] A400, loop A100-A300 until a trained self-attention mechanism neural network is obtained.

[0127] In this embodiment, the self-attention mechanism neural network is trained cyclically until a trained self-attention mechanism neural network is obtained.

[0128] 2. MMPI mixed signal separation method based on self-attention mechanism

[0129] S100, constructing a biological lesion model of a living body to be imaged by MMPI, and injecting a targeting molecular probe encapsulating SPIOs into the biological lesion model;

[0130] S200, after the injection, using an MPI device to collect signals from the biological lesion model to obtain a time-domain mixed signal containing n types of SPIOs; wherein n represents a preset number;

[0131] S300, separating the time domain mixed signal through a trained self-attention mechanism neural network to obtain n separate SPIOs time domain voltage signals.

[0132] In this embodiment, an in vivo experiment is carried out, firstly a biological lesion model of the organism is constructed, and then a targeted molecular probe wrapped with SPIOs is injected. After the injection, MPI data of a mixed variety of SPIOs are collected based on an MPI device to obtain a time-domain mixed signal containing n kinds of SPIOs; the MPI device in the present invention preferably uses a small animal three-dimensional MPI imaging device based on magnetic field-free lines or a small animal three-dimensional MPI imaging device based on magnetic field-free points. In other embodiments, it can be selected according to actual needs.

[0133] The time-domain mixed signal is input into the above-mentioned trained self-attention mechanism neural network to obtain n separate SPIOs time-domain voltage signals.

[0134] The present invention is based on a data-driven neural network method, and from the perspective of signal processing, performs feature extraction and analysis on mixed signals containing multiple SPIOs to improve the detection sensitivity of MMPI.

[0135] The second embodiment of the present invention is a MMPI mixed signal separation system based on a self-attention mechanism, which is used to separate a time-domain mixed signal containing n types of SPIOs, and then perform MMPI imaging based on the separated n types of separate SPIOs time-domain voltage signals, such as Figure 2 As shown, the system includes: a lesion model building module 100, a signal acquisition module 200, and a signal separation module 300;

[0136] The lesion model construction module 100 is configured to construct a biological lesion model of an organism to be imaged by MMPI, and inject the targeted molecular probes encapsulating SPIOs into the biological lesion model;

[0137] The signal acquisition module 200 is configured to collect signals from the biological lesion model using an MPI device after injection to obtain a time-domain mixed signal containing n types of SPIOs; wherein n represents a preset number;

[0138] The signal separation module 300 is configured to separate the time domain mixed signal through a trained self-attention mechanism neural network to obtain n separate SPIOs time domain voltage signals;

[0139] Wherein, the self-attention mechanism neural network includes a time domain encoder, a signal separator and a time domain decoder;

[0140] The time domain encoder is configured to divide the time domain mixed signal into r sequence segments u of length L i , and then encode it into a D-dimensional vector representation through a one-dimensional convolution i = 1, 2, ... r, i.e., eigenvector;

[0141] The signal separator is configured to concatenate the encoded feature vector with its position code to obtain a concatenated vector; in combination with a preset mask vector, the concatenated vector is separated by a double-layer self-attention unit to obtain source signal vectors of n types of magnetic particles;

[0142] The self-attention unit includes a plurality of attention blocks; each attention block is constructed based on a self-attention layer, a residual connection, a layer normalization, a feedforward neural network, a residual connection and a layer normalization connected in sequence;

[0143] The time domain decoder is configured to decode the source signal vector to obtain n separate SPIOs time domain voltage signals.

[0144] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0145] It should be noted that the MMPI mixed signal separation system based on the self-attention mechanism provided in the above embodiment is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be decomposed or combined. For example, the modules in the above embodiment can be combined into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the modules or steps, and are not regarded as improper limitations of the present invention.

[0146] An electronic device according to the third embodiment of the present invention comprises at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the above-mentioned MMPI mixed signal separation method based on the self-attention mechanism.

[0147] A computer-readable storage medium according to a fourth embodiment of the present invention stores computer instructions, and the computer instructions are used to be executed by the computer to implement the above-mentioned MMPI mixed signal separation method based on the self-attention mechanism.

[0148] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the above-described devices and computer-readable storage media can refer to the corresponding processes in the aforementioned method examples and will not be repeated here.

[0149] Reference below Figure 5, which shows a schematic diagram of the structure of a computer system of a server suitable for implementing the method, system, and device embodiments of the present application. Figure 5 The server shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0150] like Figure 5 As shown, the computer system includes a central processing unit (CPU) 501, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 502 or the program loaded from the storage part 508 to the random access memory (RAM) 503. Various programs and data required for system operation are also stored in the RAM 503. The CPU 501, ROM 502 and RAM 503 are connected to each other through a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0151] The following components are connected to the I / O interface 505: an input section 506 including a keyboard, a mouse, etc.; an output section 507 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN (Local Area Network) card, a modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to the I / O interface 505 as needed. A removable medium 511, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 510 as needed so that a computer program read therefrom is installed into the storage section 508 as needed.

[0152] In particular, according to an embodiment of the present disclosure, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a computer readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through the communication part 509, and / or installed from a removable medium 511. When the computer program is executed by the central processing unit (CPU501), the above functions defined in the method of the present application are executed. It should be noted that the computer-readable medium mentioned in the present application may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application In the present application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate, or transmit a program for use by or in combination with an instruction execution system, device or device. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0153] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0154] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0155] The terms "first", "second", etc. are used to distinguish similar objects rather than to describe or indicate a particular order or sequence.

[0156] The term "comprise" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that includes a list of elements includes not only those elements but also other elements not expressly listed, or also includes elements inherent to such process, method, article, or apparatus / device.

[0157] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

Claims

1. A MMPI mixed signal separation method based on a self-attention mechanism is used to separate a time-domain mixed signal containing n types of SPIOs, and then perform MMPI imaging based on the separated n types of individual SPIOs time-domain voltage signals, characterized in that: The method comprises the following steps: S100, constructing a biological lesion model of a living body to be imaged by MMPI, and injecting a targeting molecular probe encapsulating SPIOs into the biological lesion model; S200, after the injection, using an MPI device to collect signals from the biological lesion model to obtain a time-domain mixed signal containing n types of SPIOs; wherein n represents a preset number; S300, separating the time domain mixed signal through a trained self-attention mechanism neural network to obtain n separate SPIOs time domain voltage signals; Wherein, the self-attention mechanism neural network includes a time domain encoder, a signal separator and a time domain decoder; The time domain encoder is configured to divide the time domain mixed signal into r sequence segments u of length L i , and then encode it into a D-dimensional vector representation through a one-dimensional convolution i = 1, 2, ... r, i.e., eigenvector; The signal separator is configured to concatenate the encoded feature vector with its position code to obtain a concatenated vector; in combination with a preset mask vector, the concatenated vector is separated by a double-layer self-attention unit to obtain source signal vectors of n types of magnetic particles; The self-attention unit includes a plurality of attention blocks; each attention block is constructed based on a self-attention layer, a residual connection, a layer normalization, a feedforward neural network, a residual connection and a layer normalization connected in sequence; The time domain decoder is configured to convert the source signal vector into a feature vector w after the time domain mixed signal is encoded. 1 Multiply them, and decode the multiplied vector to obtain n separate SPIOs time-domain voltage signals; among them, 2. The MMPI mixed signal separation method based on the self-attention mechanism according to claim 1, characterized in that: The position code is obtained by: PE(i,2j)=sin(i / 10000 2j / D ) PE(i,2j+1)=cos(i / 10000 2j / D ) Where i represents u i In the time domain mixed signal u mul The relative position in, j represents u i in U T The corresponding index in PE() stands for position encoding.

3. The MMPI mixed signal separation method based on the self-attention mechanism according to claim 2 is characterized in that: The attention block processes the input vector as follows: The correlation between different parts of the input vector is learned through the self-attention layer. The learning process is: Q=w 2 W Q K=w 2 W K V=w 2 W Among them, W Q , W K , W V represents the randomly initialized matrix, w 2 Denotes the feature vector w 1 The vector concatenated with the position code, d k Represents the vector dimensions of Q, K, and V; Based on the vector output after learning from the self-attention layer, combined with the concatenated vector, residual connection and layer normalization are performed, and the processed vector is used as the first vector; Input the first vector into a feedforward neural network to obtain a second vector; Based on the second vector and combined with the first vector, a residual connection is performed, and layer normalization is performed, and the processed vector is used as the output of the self-attention block.

4. The MMPI mixed signal separation method based on the self-attention mechanism according to claim 3 is characterized in that: Combined with the preset mask vector, the concatenated vector is separated by a double-layer self-attention unit to obtain the source signal vectors of n kinds of magnetic particles. The method is as follows: The concatenated vectors are processed by a two-layer self-attention unit, and the processed r vectors are concatenated in sequence to obtain a one-dimensional mixed signal vector representation; The preset n mask vectors are multiplied by the mixed signal vector representation to obtain the source signal vectors of n magnetic particles, that is, b q =w 3 ·m q in, q=1,2,...,n,m q represents the mask vector, w 3 represents the mixed signal vector representation, b q Represents the source signal vector.

5. The MMPI mixed signal separation method based on the self-attention mechanism according to claim 1, characterized in that: The self-attention mechanism neural network, its training method is: A100, constructing a training set; the training samples in the training set include a time domain mixed signal obtained by adding multiple response voltage signals and its corresponding original true value source signal label; the response voltage signal is a time domain voltage signal; A200, separates the time-domain mixed signal through a pre-built self-attention mechanism neural network to obtain a separate SPIOs time-domain voltage signal as a prediction signal; A300, based on the predicted signal and the original true source signal label, calculates the loss value through the pre-built loss function and updates the network parameters of the self-attention mechanism neural network; A400, loop A100-A300 until a trained self-attention mechanism neural network is obtained.

6. The MMPI mixed signal separation method based on the self-attention mechanism according to claim 5, characterized in that: The response voltage signal in the training sample is obtained by: The digital mouse was used as a standard imaging space, in which the distribution of SPIOs of different positions, types, and concentrations was set; The Debye simulation algorithm is used to simulate the SPIOs distribution and obtain its corresponding response voltage signal.

7. The MMPI mixed signal separation method based on the self-attention mechanism according to claim 6, characterized in that: The loss function used by the self-attention mechanism neural network during training is to maximize the scale-invariant signal-to-noise ratio function, specifically: in, and u are the predicted signal and the original true source signal label respectively, and SI-SNR represents the signal-to-noise ratio function that maximizes the scale-invariant.

8. A MMPI mixed signal separation system based on a self-attention mechanism, used to separate a time-domain mixed signal containing n types of SPIOs, and then perform MMPI imaging based on the separated n types of individual SPIOs time-domain voltage signals, characterized in that: The system includes: a lesion model building module, a signal acquisition module, and a signal separation module; The lesion model construction module is configured to construct a biological lesion model of an organism to be imaged by MMPI, and inject the targeted molecular probes encapsulating SPIOs into the biological lesion model; The signal acquisition module is configured to collect signals from the biological lesion model using an MPI device after injection to obtain a time-domain mixed signal containing n types of SPIOs; wherein n represents a preset number; The signal separation module is configured to separate the time domain mixed signal through a trained self-attention mechanism neural network to obtain n separate SPIOs time domain voltage signals; Wherein, the self-attention mechanism neural network includes a time domain encoder, a signal separator and a time domain decoder; The time domain encoder is configured to divide the time domain mixed signal into r sequence segments u of length L i , and then encode it into a D-dimensional vector representation through a one-dimensional convolution i = 1, 2, ... r, i.e., eigenvector; The signal separator is configured to concatenate the encoded feature vector with its position code to obtain a concatenated vector; in combination with a preset mask vector, the concatenated vector is separated by a double-layer self-attention unit to obtain source signal vectors of n types of magnetic particles; The self-attention unit includes a plurality of attention blocks; each attention block is constructed based on a self-attention layer, a residual connection, a layer normalization, a feedforward neural network, a residual connection and a layer normalization connected in sequence; The time domain decoder is configured to convert the source signal vector into a feature vector w after the time domain mixed signal is encoded. 1 Multiply them, and decode the multiplied vector to obtain n separate SPIOs time-domain voltage signals; among them, 9. An electronic device, characterized in that: include: at least one processor; as well as a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the MMPI mixed signal separation method based on the self-attention mechanism as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to be executed by the computer to implement the MMPI mixed signal separation method based on the self-attention mechanism according to any one of claims 1 to 7.

Images