A multi-slab magnetic particle imaging method and system

Abstract

The application provides a kind of multi-block magnetic particle imaging method and system, which can be applied to medical image processing technical field.The method comprises the following steps: collecting a plurality of magnetic particle system matrices corresponding to a plurality of sub-imaging regions in the region to be imaged respectively; using a trained boundary prediction model to perform extrapolation on the plurality of magnetic particle system matrices to obtain a plurality of matrix extrapolated magnetic particle system matrices, the trained boundary prediction model being trained using simulation-generated magnetic particle system matrix samples and a loss function based on spatial continuity constraints and edge sparsity constraints; using the plurality of matrix extrapolated magnetic particle system matrices to perform image reconstruction on the plurality of sub-imaging regions to obtain a plurality of field-of-view expanded magnetic particle image blocks; performing cropping on the plurality of field-of-view expanded magnetic particle image blocks to obtain a plurality of boundary artifact-eliminated magnetic particle image blocks; and splicing the plurality of boundary artifact-eliminated magnetic particle image blocks to obtain a magnetic particle imaging result of the region to be imaged.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, in particular to a multi-block magnetic particle imaging method and system. BACKGROUND

[0002] MPI (Magnetic Particle Image) is a new molecular imaging technology, which is used for visualizing the spatial distribution of superparamagnetic iron oxide nanoparticles in a living body, has the advantages of no imaging depth limit, high sensitivity, linear quantification, no background signal interference, no ionizing radiation hazard, etc., and has gradually become a research frontier of a new generation of in vivo tomography, and has broad clinical application prospects.

[0003] The imaging field of view (FOV) of the existing MPI device can only meet the imaging needs of small targets, and it is difficult to adapt to the imaging needs of large targets; and the difficulty of improving the FOV lies in how to ensure biological safety and controllability and imaging quality without degradation while increasing the size. The current MPI imaging method for improving the FOV has boundary artifacts in image reconstruction, and there are also technical problems such as long calibration time of the system matrix, high computational complexity, and difficulty in capturing complex boundary features of the system matrix. SUMMARY

[0004] In view of the above problems, the present application provides a multi-block magnetic particle imaging method and system for at least solving one of the above technical problems.

[0005] According to a first aspect of the present application, a multi-block magnetic particle imaging method is provided, comprising: acquiring a plurality of magnetic particle system matrices each corresponding to a plurality of sub-imaging regions in a to-be-imaged region; performing extrapolation operation on the plurality of magnetic particle system matrices by using a trained boundary prediction model to obtain a plurality of matrix extrapolated magnetic particle system matrices, wherein the trained boundary prediction model is trained by using simulation generated magnetic particle system matrix samples and a loss function based on spatial continuity constraint and edge sparsity constraint; performing image reconstruction on the plurality of sub-imaging regions by using the plurality of matrix extrapolated magnetic particle system matrices to obtain a plurality of field of view expanded magnetic particle image blocks; performing cropping on the plurality of field of view expanded magnetic particle image blocks to obtain a plurality of boundary artifact eliminated magnetic particle image blocks; and splicing the plurality of boundary artifact eliminated magnetic particle image blocks to obtain a magnetic particle imaging result of the to-be-imaged region.

[0006] According to the embodiment of the present application, the boundary prediction model trained above utilizes the simulation generated magnetic particle system matrix sample and the loss function based on the spatial continuity constraint and the edge sparsity constraint to train the boundary prediction model, including: generating the magnetic particle system matrix sample based on the parameter setting information, wherein the magnetic particle system matrix sample includes the specified field of view sample and the field of view extrapolation sample; processing the specified field of view sample by using the boundary prediction model to obtain the predicted system matrix of the field of view extrapolation; processing the field of view extrapolation sample and the predicted system matrix of the field of view extrapolation by using the loss function based on the spatial continuity constraint and the edge sparsity constraint, and performing the parameter update of the boundary prediction model by using the obtained loss value; iteratively performing the data sample processing operation of the model, the loss value calculation operation and the parameter update operation of the model until the preset training condition is met, and obtaining the trained boundary prediction model.

[0007] According to the embodiment of the present application, the simulation generation of the magnetic particle system matrix sample based on the parameter setting information includes: performing the superposition operation of the static field and the driving field in the two-dimensional imaging space based on the parameter setting information and the specified size field of view to obtain the magnetic field distribution model; calculating the harmonic vector and the mixed frequency vector in the sampling frequency range by using the magnetic field distribution model according to the sampling parameter in the parameter setting information to obtain the acquisition frequency, determining the position information of the harmonic vector by analyzing the relationship between the sampling frequency and the acquisition frequency, and performing the frequency feature analysis by using the position information to obtain the number of effective frequencies; generating the initial specified field of view sample by using the initial magnetic particle concentration distribution matrix, the magnetic field distribution model, the starting point coordinate information of the two-dimensional imaging space, the number of effective frequencies and the calibration additional point number, and performing the system matrix calibration operation on the initial specified field of view sample to obtain the specified field of view sample.

[0008] According to the embodiment of the present application, the superposition operation of the static field and the driving field in the two-dimensional imaging space based on the parameter setting information and the specified size field of view to obtain the magnetic field distribution model includes: generating the coordinate grid of the two-dimensional imaging space by calculating the scanning time parameter and the sampling parameter in the parameter setting information based on the specified size field of view, and deriving the static component of the static field in different directions by using the coordinate grid; performing the magnetic field direction setting constraint on the static component, and calculating the scaling ratio of the specific imaging area, the driving field gradient and the magnetic field component after the imaging target position is moved by using the constrained static component and the specific imaging area parameter in the parameter setting information, and then completing the superposition of the static field and the driving field to obtain the magnetic field distribution model.

[0009] According to the embodiment of the present application, the system matrix calibration operation is performed on the initial designated field of view sample to obtain a designated field of view sample, including: for the pixel point with non-zero concentration in the two-dimensional imaging space, the magnetic moment component in different directions is calculated by a preset algorithm using a magnetic field distribution model; the magnetic moment component is mapped to the position of the corresponding target frequency vector according to the effective frequency index, and the mapped magnetic moment component is integrated to complete the system matrix calibration operation to obtain the designated field of view sample, wherein the target frequency vector corresponds to the initial designated field of view sample.

[0010] According to the embodiment of the present application, the system matrix sample of the magnetic particle is simulated and generated based on the parameter setting information, and the method further includes: performing data simulation on the field of view after the size extrapolation according to the parameter setting information through the magnetic field distribution modeling operation, the frequency characteristic analysis operation and the system matrix calibration operation to obtain the field of view extrapolation sample in the system matrix sample of the magnetic particle.

[0011] According to the embodiment of the present application, the method for processing the designated field of view sample by using the boundary prediction model to obtain the predicted system matrix of the field of view extrapolation includes: performing a multi-level spatial down-sampling operation on the designated field of view sample by using a hierarchical encoder of the boundary prediction model to obtain multi-level semantic features; performing global dependency coding modeling on the multi-level semantic features by using a global context modeling module of the boundary prediction model to obtain multi-level coding features; performing a multi-level up-sampling operation on the multi-level coding features by using a structure reconstruction decoding module of the boundary prediction model to complete the same level skip connection with the multi-level semantic features, thereby obtaining an initial predicted system matrix of the field of view extrapolation; performing size alignment on the initial predicted system matrix of the field of view extrapolation by using an output alignment module of the boundary prediction model to obtain an aligned predicted system matrix; and mapping the feature map dimension of the aligned predicted system matrix to a target dimension by using an output mapping generation module of the boundary prediction model to obtain the predicted system matrix of the field of view extrapolation.

[0012] According to the embodiment of the present application, each sub-encoder of the hierarchical encoder includes a first convolutional layer, a first normalization layer, a pooling layer and an activation function; the global context modeling module includes a plurality of self-attention encoders, each of which corresponds to a sub-encoder; each sub-decoder of the structure reconstruction decoding module includes an up-sampling layer, a second convolutional layer and a second normalization layer, and the sub-decoder and the sub-encoder of the same level are connected through the self-attention encoder of the same level.

[0013] According to the embodiment of the present application, the image reconstruction on the plurality of sub-imaging regions by using the plurality of matrix extrapolated magnetic particle system matrices comprises: performing orthogonal projection on the initial solution vector corresponding to each sub-imaging region by using the matrix extrapolated magnetic particle system matrix corresponding to each sub-imaging region to obtain an updated solution vector, and performing regularization constraint on the updated solution vector; performing step optimization on the matrix extrapolated magnetic particle system matrix, and performing the next round of solution vector updating operation by using the step optimized magnetic particle system matrix; iteratively performing the solution vector updating operation, the regularization constraint operation and the step optimization operation until a preset condition is met, and obtaining the field-of-view expanded magnetic particle image block corresponding to each sub-imaging region.

[0014] The second aspect of the present application provides a multi-block magnetic particle imaging system, comprising: a system matrix acquisition module configured to acquire a plurality of magnetic particle system matrices corresponding to a plurality of sub-imaging regions in a to-be-imaged region; a model data processing module configured to perform extrapolation operation on the plurality of magnetic particle system matrices by using a trained boundary prediction model to obtain a plurality of matrix extrapolated magnetic particle system matrices, wherein the trained boundary prediction model is trained by using a simulation generated magnetic particle system matrix sample and a loss function based on spatial continuity constraint and edge sparsity constraint; an image reconstruction module configured to perform image reconstruction on the plurality of sub-imaging regions by using the plurality of matrix extrapolated magnetic particle system matrices to obtain a plurality of field-of-view expanded magnetic particle image blocks; an image cropping module configured to crop the plurality of field-of-view expanded magnetic particle image blocks to obtain a plurality of boundary artifact eliminated magnetic particle image blocks; and an image stitching module configured to stitch the plurality of boundary artifact eliminated magnetic particle image blocks to obtain a magnetic particle imaging result of the to-be-imaged region.

[0015] The multi-block magnetic particle imaging method provided by the application can significantly weaken the image boundary artifacts caused by the splicing of multiple local system matrices, improve the structural consistency and reconstruction quality of the overall image, and effectively solve the boundary artifact problem in multi-block splicing magnetic particle imaging; meanwhile, the boundary prediction model provided by the application is trained by the simulation-generated magnetic particle system matrix sample and the loss function based on the spatial continuity constraint and the edge sparsity constraint, can fully learn the long-distance dependence relationship between the boundary and the internal region of the system matrix, thereby more accurately inferring the response characteristics of the boundary region, extrapolating the boundary region, avoiding repeated measurement of the overlapping region in the experiment, reducing the system matrix acquisition cost and time length, and providing efficient data support for large-scale and high-resolution imaging; in addition, the multi-block magnetic particle imaging method provided by the application can adapt to system matrix inputs of different sizes and different resolutions, support two-dimensional and three-dimensional system matrix modeling, has strong universality and scalability, and is convenient to deploy and use in various MPI imaging systems. BRIEF DESCRIPTION OF DRAWINGS

[0016] The above and other objects, features and advantages of the present application will become more apparent from the following description of embodiments of the present application, taken in conjunction with the accompanying drawings, in which:

[0017] Figure 1 is an application scenario diagram of the multi-block magnetic particle imaging method according to the embodiment of the application;

[0018] Figure 2 is a flowchart of the multi-block magnetic particle imaging method according to the embodiment of the application;

[0019] Figure 3 is a digital phantom schematic diagram used in simulation according to the embodiment of the application;

[0020] Figure 4 is a schematic diagram of the data simulation process according to the embodiment of the application;

[0021] Figure 5 is a comparison schematic diagram of the MPI images reconstructed by the system matrix obtained by simulation and the system matrix obtained by additional over-scan of 10 units according to the embodiment of the application;

[0022] Figure 6 is a comparison schematic diagram of the MPI images in X, Y and Z three dimensions reconstructed by the calibration 6 system matrix in the public data set collected by the application and the system matrix obtained by additional over-scan of 6 units according to the embodiment of the application;

[0023] Figure 7 is a structural schematic diagram of the EP-Uformer model according to the embodiment of the application;

[0024] Figure 8 is a schematic diagram of image reconstruction after system matrix extrapolation on a simulated dataset based on an EP-Uformer model according to an embodiment of the present application;

[0025] Figure 9 is a schematic diagram of image reconstruction after system matrix extrapolation on a real public dataset based on an EP-Uformer model according to an embodiment of the present application;

[0026] Figure 10 is a structural block diagram of a multi-block magnetic particle imaging system according to an embodiment of the present application;

[0027] Figure 11 is a block diagram of an electronic device suitable for implementing a multi-block magnetic particle imaging method according to an embodiment of the present application. DETAILED DESCRIPTION

[0028] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. It is to be understood, however, that these descriptions are merely exemplary and are intended to provide a thorough and complete disclosure of the embodiments of the present application, as copies are filed in a country or region in which the present application is filed. Therefore, many modifications, variations and changes in addition to what is described here can be possible without departing from the scope of the present application. Further, in the following description, descriptions of well-known structures and techniques have been omitted to avoid unnecessarily obscuring the concept of the present application.

[0029] The terms used herein are used only to describe specific embodiments, and are not intended to limit the present application. The terms "include", "comprise" and the like used herein indicate the presence of the described features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0030] All terms used herein, including technical and scientific terms, have meanings commonly understood by one of ordinary skill in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having meanings consistent with the context of the specification, and should not be interpreted in an idealized or overly formal manner.

[0031] In the case of using expressions similar to "at least one of A, B, and C, etc.", in general, it should be interpreted as having a meaning that includes one or more of the corresponding categories recited (e.g., "a system having at least one of A, B, and C" should include but not be limited to a system having A alone, a system having B alone, a system having C alone, a system having both A and B, a system having both A and C, a system having both B and C, and / or a system having A, B, and C together, etc.).

[0032] The magnetic field free region in the MPI device moves in the target region along a trajectory under the action of the driving field, and the size of the region covered by the trajectory is referred to as the field of view (FOV). Currently, the FOV of the MPI device is mostly 3cm 3 to 12cm 3 , which can only meet the needs of small animal imaging. Studies have found that increasing the FOV can be achieved by increasing the driving field amplitude or reducing the gradient field intensity, but this will cause peripheral nerve stimulation in the human body and a decrease in imaging resolution, respectively. The difficulty in increasing the FOV lies in how to increase the size while ensuring biological safety and controllability and not reducing the imaging quality. In recent years, researchers have carried out extensive exploration on the FOV increasing method, and proposed a multi-block MPI based on the system matrix reconstruction method. In order to obtain a larger FOV without increasing the additional system matrix calibration time, the FOV is often divided into multiple non-overlapping FOV blocks, and the image of each block is reconstructed based on the system matrix reconstruction algorithm, and then the image blocks are spliced into a whole image. However, due to the contribution of particles outside the block, the information at the boundary of the FOV is lost, causing boundary artifacts during image reconstruction. The traditional method can partially reduce the boundary artifacts by expanding the system matrix calibration range of each FOV block through overscanning, but this method will significantly increase the calibration time and computational complexity of the system matrix. In order to avoid the generation of boundary artifacts at the block boundary, a method based on interpolation and regression is proposed to realize the extrapolation of the system matrix, but its extrapolation ability is limited and depends on the traditional interpolation strategy, which can be used for fast system matrix estimation, but it is difficult to capture the complex boundary characteristics of the system matrix.

[0033] To at least solve one of the above technical problems, the present application provides a multi-block magnetic particle imaging method and system, which uses a trained boundary prediction model to process a smaller field of view, i.e. an MPI system matrix lacking a boundary, to extrapolate it to obtain a larger field of view, i.e. an MPI system matrix with a generated boundary, to reduce the boundary artifacts of block splicing, improve the consistency and accuracy of MPI in large-scale imaging, and expand its application value in tumor imaging and organ imaging. Compared with the method of improving the FOV size of MPI by optimizing the hardware, the multi-block magnetic particle imaging method and system provided by the present application will not cause an increase in cost and will not cause a loss of other imaging indicators; and compared with the method of overscanning the MPI system matrix, the multi-block magnetic particle imaging method and system provided by the present application does not need to measure the system matrix of the overscanning area, and does not need additional calibration time.

[0034] Figure 1 is an application scenario diagram of the multi-block magnetic particle imaging method according to an embodiment of the present application.

[0035] As Figure 1As shown, the application scenario 100 according to this embodiment can include medical image processing, medical diagnosis, and the like application scenarios. The network 104 is a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 can include various connection types, such as wired, wireless communication links, or fiber optic cables, and the like.

[0036] The user can use the first terminal device 101, the second terminal device 102, the third terminal device 103 to interact with the server 105 through the network 104 to receive or send messages, and the like. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, and the like (only as examples).

[0037] The first terminal device 101, the second terminal device 102, the third terminal device 103 can be various electronic devices with a display screen and supporting web browsing, including but not limited to smartphones, tablet computers, laptop computers, desktop computers, and the like.

[0038] The server 105 can be a server providing various services, such as a background management server supporting a website browsed by the user using the first terminal device 101, the second terminal device 102, the third terminal device 103 (only as an example). The background management server can analyze and process received user requests and the like data, and feed back the processing results (such as web pages, information, or data generated or obtained according to user requests) to the terminal device.

[0039] It should be noted that the multi-piece magnetic particle imaging method provided by the embodiment of the present application can generally be executed by the server 105. Correspondingly, the multi-piece magnetic particle imaging system provided by the embodiment of the present application can generally be arranged in the server 105. The multi-piece magnetic particle imaging method provided by the embodiment of the present application can also be executed by a server or a server cluster different from the server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or the server 105. Correspondingly, the multi-piece magnetic particle imaging system provided by the embodiment of the present application can also be arranged in a server or a server cluster different from the server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or the server 105.

[0040] It should be understood that Figure 1 The number of terminal devices, networks, and servers in the above-mentioned application scenario is only illustrative. According to the implementation needs, there can be any number of terminal devices, networks, and servers.

[0041] The following will be based on Figure 1 The described scenario, by Figures 2-9 The multi-block magnetic particle imaging method of the disclosed embodiment is described in detail.

[0042] Figure 2 is a flowchart of the multi-block magnetic particle imaging method according to the embodiment of the application.

[0043] As Figure 2 shown, the multi-block magnetic particle imaging of this embodiment includes operations S210-S250.

[0044] In operation S210, a magnetic particle system matrix corresponding to each of a plurality of sub-imaging regions in a region to be imaged is collected.

[0045] In operation S220, an extrapolation operation is performed on the plurality of magnetic particle system matrices using a trained boundary prediction model to obtain a plurality of matrix extrapolated magnetic particle system matrices, wherein the trained boundary prediction model is trained using simulation generated magnetic particle system matrix samples and a loss function based on spatial continuity constraints and edge sparsity constraints.

[0046] The above boundary prediction model is a multi-dimensional boundary prediction model (Edge Prediction Uformer, EP-Uformer model) that fuses the U-Net structure of the Transformer.

[0047] The extrapolation operation is performed on the magnetic particle system matrix, i.e., the magnetic particle system matrix of a smaller field of view is extrapolated to a certain size to obtain a magnetic particle system matrix of a larger field of view.

[0048] In operation S230, image reconstruction is performed on the plurality of sub-imaging regions using the plurality of matrix extrapolated magnetic particle system matrices to obtain a plurality of field of view expanded magnetic particle image blocks.

[0049] In operation S240, cropping is performed on the plurality of field of view expanded magnetic particle image blocks to obtain a plurality of boundary artifact eliminated magnetic particle image blocks.

[0050] In operation S250, the plurality of boundary artifact eliminated magnetic particle image blocks are spliced to obtain a magnetic particle imaging result of the region to be imaged.

[0051] The multi-block magnetic particle imaging method provided by the application can significantly weaken image boundary artifacts caused by splicing of multiple local system matrices, improve the structural consistency and reconstruction quality of the overall image, and effectively solve the boundary artifact problem in multi-block splicing magnetic particle imaging. Meanwhile, the boundary prediction model provided by the application is trained by using the simulation-generated magnetic particle system matrix sample and the loss function based on the spatial continuity constraint and the edge sparsity constraint, can fully learn the long-distance dependence relationship between the boundary and the internal region of the system matrix, thereby more accurately inferring the response characteristics of the boundary region, extrapolating the boundary region, avoiding repeated measurement of the overlapping region in the experiment, reducing the system matrix acquisition cost and time length, and providing efficient data support for large-scale and high-resolution imaging. In addition, the multi-block magnetic particle imaging method provided by the application can adapt to system matrix inputs of different sizes and different resolutions, support two-dimensional and three-dimensional system matrix modeling, has strong universality and scalability, and is convenient to deploy and use in various MPI imaging systems.

[0052] According to the embodiment of the application, the training of the completed boundary prediction model uses the simulation-generated magnetic particle system matrix sample and the loss function based on the spatial continuity constraint and the edge sparsity constraint, and includes: simulating the magnetic particle system matrix sample based on the parameter setting information, wherein the magnetic particle system matrix sample includes a specified field sample and a field extrapolation sample; processing the specified field sample using the boundary prediction model to obtain a predicted system matrix for field extrapolation; processing the field extrapolation sample and the predicted system matrix for field extrapolation using the loss function based on the spatial continuity constraint and the edge sparsity constraint, and performing parameter update on the boundary prediction model using the obtained loss value; iteratively performing the data sample processing operation of the model, the loss value calculation operation and the parameter update operation of the model until the preset training condition is met, and obtaining the trained boundary prediction model.

[0053] The specified field sample is a magnetic particle system matrix sample of a specified size field; the field extrapolation sample is a magnetic particle system matrix sample after field extrapolation of a specified size; in the simulation data generation process, the specified field sample and the field extrapolation sample are generated in pairs to train the boundary prediction model.

[0054] The training process of the boundary prediction model provided by the application will be further described in detail below through specific embodiments.

[0055] In the boundary prediction model training process, the simulation-generated magnetic particle system matrix samples are divided into training set, validation set and test set according to a certain proportion, wherein the training set is used for model parameter update, the validation set is used for monitoring overfitting and model optimization, and the test set is used for final performance evaluation. Each set of training samples includes: input item - part of the system matrix in a small field of view range, as the model input; output item - the complete extrapolation region system matrix corresponding to the small field of view system matrix, as the supervision label. The extrapolation region of the sample in the training set is generated by simulation, ensuring that the boundary information is complete and accurate, forming a high-quality training supervision pair.

[0056] The boundary prediction model not only can process system matrix data, but also supports two-dimensional data or three-dimensional data input, and in the training process of the boundary prediction model, the boundary prediction model structure (such as the number of input / output channels, the width of intermediate feature channels, the number of blocks, etc.), the optimizer configuration (such as the learning rate of Adam optimizer, the weight decay parameter, etc.), the learning rate scheduling strategy (StepLR, supporting setting step and decay factor) and the total number of training rounds can be determined through parameter setting.

[0057] Training data set loading: support two types of data input of multi-scene simulation system matrix and real magnetic particle system, load multiple groups of data through experiment ID, divide training set and validation set, and complete normalization processing, mask annotation and batch organization (Batch Size can be set) through custom data loader.

[0058] Training process: in each training round, forward propagation and back propagation are performed on batch input in turn, loss function is calculated and model parameters are updated. The model is scheduled once after each training round to control the stability and convergence speed of the optimization process.

[0059] Model verification and index evaluation: every preset number of rounds (for example, every 5 rounds), the model evaluates the prediction effect on the validation set, and adopts normalized root mean square error (nRMSE, Normalized Root Mean Square Error) as the performance index. The verification process includes output result restoration (mean variance inverse normalization), complex synthesis, mask extraction, vectorization transformation and error calculation. If the current nRMSE is lower than the historical best value, the current model is saved as the best model.

[0060] Model testing: after the boundary prediction model is trained, the trained boundary prediction model (i.e., the trained EP-Uformer model, the same below) is deployed and applied to realize extrapolation prediction of the actual small field of view system matrix, and to expand the imaging field of view range of the magnetic particle imaging system. First, input test data, wherein the test data includes (1) the test set small field of view system matrix reserved in the simulation data set; or (2) the experimental system matrix collected by the real MPI device (usually only covering part of the field of view). The input system matrix needs to be preprocessed and standardized in the same way as in the training stage to ensure that the model input dimension and scale are consistent with the training stage.

[0061] System matrix boundary extrapolation: input the processed small field of view system matrix into the trained EP-Uformer model, and output the extrapolated large field of view system matrix. Use an image reconstruction algorithm (such as the Kaczmarz regularization reconstruction algorithm) to reconstruct the extrapolated system matrix, and then splice each block of the image after image reconstruction to obtain a complete MPI image, and further evaluate the improvement effect of boundary prediction on image quality.

[0062] According to the embodiment of the application, the above-mentioned simulation of the magnetic particle system matrix sample based on the parameter setting information comprises: based on the parameter setting information and the specified size of the field of view, performing superposition operation of the static field and the driving field in the two-dimensional imaging space to obtain a magnetic field distribution model; according to the sampling parameters in the parameter setting information, calculating the harmonic vector and the mixed frequency vector in the sampling frequency range by using the magnetic field distribution model to obtain the acquisition frequency, determining the position information of the harmonic vector by analyzing the relationship between the sampling frequency and the acquisition frequency, and performing frequency feature analysis by using the position information to obtain the number of effective frequencies; generating an initial specified field of view sample by using the initial magnetic particle concentration distribution matrix, the magnetic field distribution model, the starting point coordinate information of the two-dimensional imaging space, the number of effective frequencies and the calibration additional point number, and performing system matrix calibration operation on the initial specified field of view sample to obtain the specified field of view sample.

[0063] The magnetic particle system matrix sample simulation process mainly includes data simulation parameter setting, magnetic field distribution modeling, frequency feature analysis, and system matrix calibration operation.

[0064] Among them, the parameter setting information, that is, by analyzing and setting the key parameters used in the data simulation process: comprehensively configuring the scanning parameters, the magnetic field parameters, the two-dimensional imaging space parameters, the imaging target parameters and the system matrix calibration parameters. By flexibly adjusting the scanning time, the sampling frequency, the static field gradient, the driving field frequency, the two-dimensional imaging space size, the target shape and size, etc., the precise simulation of diversified imaging scenarios is realized, and the data simulation requirements under different application requirements are met.

[0065] According to the embodiment of the present application, the above-mentioned field distribution model is obtained by superimposing the static field and the driving field in the two-dimensional imaging space based on the field of view of the specified size and the parameter setting information, and includes: generating a coordinate grid of the two-dimensional imaging space by calculating the scanning time parameter and the sampling parameter in the parameter setting information based on the field of view of the specified size, and deriving the static components of the static field in different directions by using the coordinate grid; performing magnetic field direction setting constraints on the static components, and calculating the scaling ratio, the driving field gradient and the magnetic field component after the imaging target position is moved of the specific imaging area by using the constrained static components and the specific imaging area parameter in the parameter setting information, and then completing the superimposition of the static field and the driving field to obtain the field distribution model.

[0066] According to the embodiment of the present application, the above-mentioned system matrix calibration operation is performed on the initial specified field of view sample to obtain the specified field of view sample, and includes: calculating the magnetic moment components of the magnetic moment in different directions by using the field distribution model through a preset algorithm for the pixel points with non-zero concentration in the two-dimensional imaging space; mapping the magnetic moment components to the positions of the corresponding target frequency vectors according to the effective frequency index, and integrating the mapped magnetic moment components to complete the system matrix calibration operation to obtain the sample of the specified field of view, wherein the target frequency vector corresponds to the initial specified field of view sample.

[0067] According to the embodiment of the present application, the above-mentioned simulation of generating the magnetic particle system matrix sample based on the parameter setting information further includes: performing data simulation on the field of view after the specified size extrapolation by the field distribution modeling operation, the frequency characteristic analysis operation and the system matrix calibration operation according to the parameter setting information to obtain the field of view extrapolation sample in the magnetic particle system matrix sample.

[0068] The above-mentioned embodiments relate to the process of simulating the generation of the magnetic particle system matrix sample, and the simulation process of the magnetic particle system matrix sample will be further described in detail through specific embodiments.

[0069] The field distribution modeling is based on the parameter configuration information used for data simulation to construct the static field and driving field distribution model of the two-dimensional imaging space. The scanning related time and sampling parameters are calculated first to generate the coordinate grid of the two-dimensional imaging space, and the components of the static field in different directions are derived based on this. At the same time, part of the components are processed to meet the magnetic field direction setting. Then, the scaling ratio, the driving field gradient and the magnetic field component after the target position is moved of the specific imaging area are calculated in combination with the specific imaging area parameter, the static field and the driving field components are superimposed to obtain the complete magnetic field amplitude and the magnetic field component in each direction, and the construction of the two-dimensional imaging space field distribution model is completed to create conditions for simulating the response of the magnetic particle in the magnetic field.

[0070] Frequency characteristic analysis, frequency mixing processing and harmonic position determination: based on the driving field frequency and the sampling parameters, the harmonic vectors and the mixed frequency vectors in the sampling frequency range are calculated to determine the acquisition frequency. By analyzing the relationship between the sampling frequency and the acquisition frequency, the harmonic positions are accurately determined. These harmonic position information is crucial for analyzing the frequency characteristics of the magnetic particle signal and subsequent system matrix calculation, and helps to accurately extract the effective frequency components in the signal.

[0071] System matrix calibration operation: first, an initial concentration distribution matrix is created, and the concentration values of specific regions in the matrix are set according to the starting point coordinates of the specific imaging area and the number of additional calibration points, thereby simulating the distribution of magnetic particles. For pixel points with non-zero concentration, the components of the magnetic moment in different directions are calculated using relevant algorithms, and then the magnetic moment components are converted into elements of the system matrix at the corresponding harmonic positions. The complete system matrix is obtained by integration, which clearly defines the relationship between the input (magnetic particle distribution) and the output (measurement signal) of the imaging system, and is the key basis for subsequent image reconstruction.

[0072] After the system matrix correction operation, target scanning and signal acquisition are carried out: according to the pre-set imaging target type, an imaging target image of the corresponding shape is generated and placed at the center of the two-dimensional imaging space. Based on the generated magnetic field and the imaging target image, a specific algorithm is used to calculate the components of the magnetic moment in different directions within the target area. Through integration, differentiation and Fourier transform operations, the signal vector at the specific harmonic position is extracted, which simulates the signal data containing effective information of the imaging target collected by the actual imaging system.

[0073] Finally, image reconstruction and result saving: based on the calculated system matrix and the acquired signal, a regularization algorithm is used for image reconstruction. During the reconstruction process, the regularization parameter can be adjusted according to the requirements, and the reconstructed image can be normalized and processed. At the same time, the quality of the reconstructed image is evaluated by calculating specific indicators, and the reconstructed image is finally saved to the specified path for subsequent analysis and verification. Through the above series of steps, the data simulation module completely simulates the whole process of multi-block magnetic particle imaging from magnetic field generation, signal acquisition to image reconstruction, providing a solid data foundation for the matrix extrapolation method based on the boundary prediction model.

[0074] According to the embodiment of the present application, the step of superimposing the static field and the driving field component to obtain the complete magnetic field distribution is: calculating the components of the static field at each point in the two-dimensional imaging space using the pre-configured gradient parameters, calculating the change components of the driving field at different times and different positions based on the position and size parameters of the imaging area, and generating the magnetic field amplitude and direction distribution data covering the entire two-dimensional imaging space and changing with time through point-by-point superposition.

[0075] According to the embodiment of the present application, the step of calculating the system matrix further comprises: performing a dimension adaptation process on the calculated system matrix, adjusting the matrix structure according to the dimension (two-dimensional or three-dimensional) of the imaging data, and ensuring the compatibility of the system matrix with the subsequent image reconstruction algorithm; at the same time, an optimized numerical calculation method is used to improve the efficiency and accuracy of the system matrix calculation.

[0076] According to the embodiment of the present application, the step of processing the specified field of view sample by using the boundary prediction model to obtain the predicted system matrix of the field of view extrapolation comprises: performing a multi-level spatial downsampling operation on the specified field of view sample by using the hierarchical encoder of the boundary prediction model to obtain multi-level semantic features; performing global dependency coding modeling on the multi-level semantic features by using the global context modeling module of the boundary prediction model to obtain multi-level coding features; performing a multi-level upsampling operation on the multi-level coding features by using the structure reconstruction decoding module of the boundary prediction model to complete the same level jump connection with the multi-level semantic features, and obtaining an initial predicted system matrix of the field of view extrapolation; performing size alignment on the initial predicted system matrix of the field of view extrapolation by using the output alignment module of the boundary prediction model to obtain an aligned predicted system matrix; and mapping the feature map dimension of the aligned predicted system matrix to a target dimension by using the output mapping generation module of the boundary prediction model to obtain the predicted system matrix of the field of view extrapolation.

[0077] The boundary prediction model provided by the present application aims to accurately and efficiently extrapolate the boundary region in the system matrix of the magnetic particle imaging system. The module fuses the local modeling capability of the U-Net architecture and the global modeling capability of the Transformer architecture by constructing an EP-Uformer model, effectively captures the global dependency relationship between distant positions in the system matrix while maintaining the integrity of image details.

[0078] According to the embodiment of the present application, each sub-encoder of the hierarchical encoder comprises a first convolutional layer, a first normalization layer, a pooling layer and an activation function; wherein the global context modeling module comprises a plurality of self-attention encoders, each self-attention encoder corresponding to a sub-encoder; wherein each sub-decoder of the structure reconstruction decoding module comprises an upsampling layer, a second convolutional layer and a second normalization layer, and the sub-decoder and the sub-encoder of the same level are connected through the self-attention encoder of the same level.

[0079] The EP-Uformer model adopts a symmetrical U-Net backbone architecture as the basic network framework, mainly including five parts: a hierarchical encoder module, a global context modeling module, a structure reconstruction decoding module, an output size padding module, and an output mapping generation module.

[0080] The hierarchical encoder module includes a plurality of autoencoders, each of which is composed of a plurality of first convolutional layers, first normalization layers, pooling layers and activation functions (ReLU), and extracts local spatial features from the input system matrix (or magnetic particle system matrix sample). After each level of convolution, a downsampling operation (such as AvgPool: average pooling, or Strided Conv: stride convolution) is performed, and after extracting the preliminary features, multi-level spatial downsampling is performed to gradually reduce the spatial resolution and extract multi-level semantic features, enabling the network to understand larger-scale context information. The structure is similar to the U-Net encoder part, and a skip connection is used. Each layer of features is cached during the encoding process for subsequent decoder skip connection.

[0081] The global context modeling module introduces a Transformer encoder module (i.e., self-attention encoder) based on the high-order features output by the encoder to model the global dependence of the system matrix tokens in space and capture long-range spatial interactions. The Transformer is embedded in the U-Net structure and placed in the feature bottleneck layer between the encoder and the decoder. The module includes a plurality of Transformer blocks, each of which includes a normalization processing unit, a self-attention calculation unit and a feedforward neural network unit, which can model the global dependence of the input system matrix features. The normalization processing unit normalizes the input features to improve the numerical stability of the subsequent feature modeling process. The self-attention calculation unit models the feature relationship of each spatial position based on the normalized features using a multi-head self-attention mechanism to realize long-distance dependence calculation within the spatial region. The feedforward neural network unit includes at least two linear transformations and a nonlinear activation unit to further enhance the expression ability of the features output by the self-attention. The self-attention calculation unit and the feedforward neural network unit are both connected through a residual connection and fused with the input features to maintain the continuity and consistency of the features and avoid the gradient vanishing problem in the training process of the deep model.

[0082] The structure reconstruction decoding module symmetrically recovers the spatial size with the encoder, gradually enlarges the spatial size by layer-by-layer upsampling (ConvTranspose), and fuses the features from the encoder and the current decoding features at the corresponding positions in the spatial dimension. In particular, in order to balance the modeling capability and computational efficiency, the EP-Uformer model only introduces the Transformer block in the jump connection of the deepest layer and the next deepest layer to fuse the high semantic level encoder and decoder features and improve the global consistency modeling capability; while the element-wise addition is directly used in the shallow layer jump connection to quickly transfer the local detail information of the encoder to the decoder, effectively recovering the low-level features such as edges and textures, and avoiding unnecessary computational redundancy. The module uses the multi-head self-attention mechanism to establish global dependence between skip features and decoder features, improving semantic consistency and boundary recovery capability in the decoding process. Through the step-by-step transposed convolution upsampling (Transposed Conv) of high-level semantic features, combined with the features from the same level of the encoder, the output matrix of the target size is gradually restored through Transformer fusion, and the key boundary information is effectively preserved.

[0083] The output alignment module fills the features to the target size through interpolation when the size of the decoder output does not match the target size, realizing size alignment.

[0084] The output mapping generation module maps the final multi-channel feature map to the target output dimension through a 3x3 final convolution layer, completing the final prediction of the system matrix boundary completion.

[0085] The EP-Uformer model introduces a multi-dimensional input and output mechanism to adapt to the high-dimensional characteristics of the magnetic particle imaging system matrix. The system matrix data usually contains multiple dimensions (such as frequency dimension, channel dimension, spatial dimension, etc.), and the EP-Uformer model ensures the effective fusion and propagation of features in each dimension in the network through the design of multi-dimensional convolution and position encoding modules, thereby improving the understanding ability of the boundary region structure information.

[0086] In addition, considering the physical characteristics of the MPI system matrix, the EP-Uformer model introduces continuity and sparsity constraints based on spatial physical laws in the loss function design. Specifically, the model constructs a joint optimization target based on L1 / L2 loss, combines two types of physical priors of spatial continuity and boundary sparsity, and enhances the smoothness and physical reasonableness of the boundary region by differentiating the system matrix in the spatial coordinate dimension. At the same time, with the help of the boundary mask to limit the constraint range, the model only applies physical priors to the boundary region, effectively improving the accuracy of boundary prediction and reducing the artifacts and discontinuity phenomena in the extrapolation region.

[0087] Ultimately, the model employs standard neural network initialization and weight normalization techniques (LayerNorm and weight decay) to ensure stable training and enhance generalization capabilities. The overall structure is highly modular and scalable, adaptable to system matrix extrapolation tasks of varying resolutions and dimensions.

[0088] Through the above design, the model building module constructs an EP-Uformer model that integrates global modeling capabilities and local detail preservation capabilities. This model possesses excellent boundary learning capabilities and high-dimensional data adaptation capabilities, significantly improving the prediction accuracy of the system matrix in unmeasured edge regions. It provides key algorithmic support for the stitching of multi-block magnetic particle imaging system matrices and large-field imaging.

[0089] According to an embodiment of the present invention, the above-mentioned method of performing image reconstruction on multiple sub-imaging regions using a magnetic particle system matrix extrapolated from multiple matrices to obtain multiple magnetic particle image blocks with expanded field of view includes: orthogonally projecting the initial solution vector corresponding to each sub-imaging region using the magnetic particle system matrix extrapolated from the matrix corresponding to each sub-imaging region to obtain an updated solution vector, and performing regularization constraints on the updated solution vector; performing step size optimization on the magnetic particle system matrix extrapolated from the matrix, and performing the next round of solution vector update operation using the magnetic particle system matrix with optimized step size; iteratively performing the solution vector update operation, regularization constraint operation, and step size optimization operation until a preset condition is met to obtain a magnetic particle image block with expanded field of view corresponding to each sub-imaging region.

[0090] Based on the embodiments disclosed in this invention, those skilled in the art can use other image reconstruction regularization algorithms according to actual needs, such as the Kaczmarz regularization algorithm.

[0091] The following detailed explanation, through specific experiments and in conjunction with the accompanying drawings, further illustrates the multi-particle magnetic particle imaging method, EP-Uformer model, and magnetic particle image simulation process provided by this invention.

[0092] Figure 3 This is a schematic diagram of a digital phantom used in simulation according to an embodiment of the present invention. Figure 3 In the text, (a) represents a vascular phantom. Figure 3 In the diagram, (b) represents a nested rectangular phantom. Figure 3 (c) in the figure represents a uniform blocky phantom.

[0093] Figure 4 This is a schematic diagram of the data simulation process according to an embodiment of the present invention.

[0094] Figure 5is a comparison of the MPI images reconstructed from the simulated system matrix according to an embodiment of the present application and the system matrix reconstructed with an additional overscanning of 10 units.

[0095] Figure 6 is a comparison of the MPI images in three dimensions of X, Y and Z axes reconstructed from the calibration 6 system matrix in the public dataset acquired in real and the system matrix reconstructed with an additional overscanning of 6 units according to an embodiment of the present application. Among them, Figure 6 The public dataset used is the OpenMPI dataset.

[0096] The key parameters are set as follows: the driving frequency in the x direction is 24.5 kHz, the driving frequency in the y direction is 26 kHz, and the sampling frequency is 2.5 MHz. The training set includes two system matrices, and the selected field gradient is 0.5 and 1 T / m, and the particle diameter is 12.5 nm. The validation set includes one system matrix, and the selected field gradient is 0.5 T / m, and the particle diameter is 15 nm. The system matrices of the training set and the validation set are divided into 4 blocks, and the size of the smaller field of view is 20x20, and the extrapolation length is 10, so the size of the larger field of view after extrapolation is 40x40, as shown in Figure 4 , 20x20 is the system matrix of the designated size field of view, which is taken as the input, and the system matrix of the larger field of view of 40x40 after extrapolation is taken as the label. The test set includes three system matrices, and the selected field gradient is 0.5, 1 and 3 T / m, and the corresponding particle diameter is 13.5 nm, 15 nm and 12.5 nm, respectively. Among them, the imaging phantom corresponding to the selected field gradient of 0.5 T / m and the particle diameter of 13.5 nm is a blood vessel phantom, and the system matrix is divided into 9 blocks; the imaging phantom corresponding to the selected field gradient of 1 T / m and the particle diameter of 15 nm is a uniform block phantom, and the system matrix is divided into 4 blocks; the imaging phantom corresponding to the selected field gradient of 3 T / m and the particle diameter of 12.5 nm is a nested rectangular phantom, and the system matrix is divided into 4 blocks. The schematic diagram of the three phantoms is shown in Figure 5 . After obtaining the simulated system matrix, a data pair needs to be constructed, and the specific method is shown in Figure 5 In this embodiment, the system matrix of the larger field of view (40x40) is taken as the label, and the input data (20x20) is obtained after cutting the extrapolation length (10). The comparison of the MPI images reconstructed from the simulated system matrix and the system matrix reconstructed with an additional overscanning of 10 units in the specific embodiment of the present application is shown in Figure 5 , wherein, Figure 5 (a) and (d) in Figure 5 (a) is a schematic diagram of the splicing of the 9 blocks of the blood vessel phantom without overscanning, Figure 5 (d) is a schematic diagram of the splicing of the 9 blocks of the blood vessel phantom with an overscanning of 10 units; Figure 5(b) and (e) in the diagram are schematic diagrams of the four pieces of a nested rectangular phasor. Figure 5 (b) in the diagram is a schematic diagram of the stitching of four nested rectangular phantoms that have not been scanned. Figure 5 (e) in the diagram is a schematic diagram of four nested rectangular phantoms stitched together after 10 units of overscanning; Figure 5 (c) and (f) in the diagram are schematic diagrams of the four pieces of a uniform block phantom being assembled. Figure 5 (c) in the diagram is a schematic diagram of the stitching together of four uniform block-shaped phantoms that have not been scanned. Figure 6 (f) is a schematic diagram of four uniform block-shaped phantoms stitched together after 10 units of overscanning; a comparative schematic diagram of the MPI image reconstructed using the system matrix acquired by OpenMPI with the system matrix after an additional 6 units of overscanning in a specific embodiment of the present invention in the X, Y, and Z axes is shown below. Figure 6 As shown, where, Figure 6 (a) is a schematic diagram of the MPI reconstructed image of the unscanned 3D phantom on the X-axis. Figure 6 (d) is a schematic diagram of the MPI reconstructed image of the 3D phantom overscanned by 6 units on the X-axis; Figure 6 (b) is a schematic diagram of the unscanned MPI reconstructed image on the Y-axis. Figure 6 (e) is a schematic diagram of the MPI reconstructed image of the 3D phantom overscanned by 6 units on the Y-axis; Figure 6 (c) is a schematic diagram of the unscanned MPI reconstructed image on the Z-axis. Figure 7 (f) is a schematic diagram of the MPI reconstructed image of the 3D phantom with 6 overscan units on the Z-axis; it can be seen that the boundary artifacts in the reconstructed image after overscan are significantly reduced. The model constructed in this invention predicts the boundary of the system matrix through deep learning methods, reducing the boundary artifacts after multiple blocks are stitched together without overscanning, and without increasing the scanning cost.

[0097] Figure 7 This is a schematic diagram of the EP-Uformer model according to an embodiment of the present invention. Figure 7 (a) in the diagram represents the overall structure of the EP-Uformer model. Figure 7 (b) in the diagram represents the structural diagram of the Transformer block. Figure 7 (c) in the diagram represents the structure of the upsample block.

[0098] like Figure 7 As shown, the EP-Uformer model includes a hierarchical encoder module, a global context modeling module, a structural reconstruction encoder module, an output size alignment module, and an output mapping generation module.

[0099] The hierarchical encoder module aims to first extract local spatial features from the input using a series of convolutional layers (Conv), batch normalization layers (batchnorm), average pooling layers, and activation functions. Then, it performs multi-level spatial downsampling, with each layer compressing the spatial dimension and enhancing the semantic level through AvgPool (average pooling), enabling the network to understand a wider range of contextual information. Features from each layer are cached during the encoding process for subsequent skip connections in the decoder. In a specific embodiment of this invention, N=3 is set, meaning the hierarchical encoder includes 3 encoding blocks, such as... Figure 7 As shown in (a) of the diagram.

[0100] The purpose of the global context modeling module is to introduce a Transformer encoder module (multi-layer Multi-head Attention + Feed Forward network) based on the high-order features output by the encoder. This Transformer encoder module includes multiple Transformer blocks; it performs global dependency modeling on tokens in the space to capture long-range spatial interactions. In a specific embodiment of this invention, M=13, meaning it includes 13 Transformer blocks, such as... Figure 7 As shown in (a), the structure of the Transformer block is as follows: Figure 7 As shown in (b) in the figure, the GELU (Gaussian Error Linear Unit) activation function represents the Gaussian error linear activation function.

[0101] The purpose of the structural reconstruction decoder module is to recover the spatial dimensions symmetrically with the encoder. It contains the same number of decoder blocks as the encoder blocks. Each decoder block includes an upsampling block, a convolutional layer, and a normalization layer. The specific structure is as follows: Figure 8 As shown in (c), the RELU (Linear Rectification Function) activation function represents the linear rectified activation function. The spatial size is gradually enlarged through layer-by-layer upsampling (ConvTranspose), and the skip features from the corresponding encoder are added to the current decoded features. Simultaneously, Transformer blocks are added only to skip connections in the deepest / second-deepest layers (the penultimate and penultimate layers) to fuse encoder and decoder features. Shallow features are directly added, achieving detail restoration and high-resolution reconstruction.

[0102] The purpose of the output size alignment module is to ensure that the output size and the target size are the same. Normally, the output size of the decoder is consistent with the target. If there is a discrepancy, an interpolation method is used, such as bilinear (2D) or trilinear (3D) interpolation, to perform smooth amplification.

[0103] The output mapping module maps the final multi-channel feature map to the target output dimension through a 3×3 convolution, completing the final prediction of system matrix boundary completion.

[0104] The input to the EP-Uformer model is the simulated / real MPI system matrix with a smaller field of view, and the label is the corresponding simulated / real MPI system matrix with a larger field of view. The optimization algorithm used for supervised training of the model is the Adam optimizer. The loss function is to effectively improve the extrapolation accuracy of the boundary region of the MPI system matrix. This invention combines the basic pixel-level error and the spatial physical prior constraint to construct the polynomial joint loss as shown in formula (1). :

[0105] (1).

[0106] The element-wise difference between the predicted system matrix and the true matrix is ​​measured only within the boundary mask region, as shown in formula (2):

[0107] (2),

[0108] in and The first The complex value of the prediction system matrix and the first prediction system matrix The complex value of a matrix predicting the true system. This represents the set of spatial coordinates within the mask region; this directly optimizes reconstruction accuracy. This represents the L2 norm.

[0109] This is a spatial continuity constraint. Physically, the MPI system matrix has a certain continuity in spatial coordinates, and adjacent pixels or voxels should not exhibit drastic fluctuations. The smoothness term, based on the differences in each spatial dimension, constrains the spatial variation amplitude of the system matrix, as shown in formula (3):

[0110] (3),

[0111] in Indicates the number of spatial dimensions (2D / 3D). Indicates the first The complex value of the prediction system matrix in the th... The difference value of each spatial dimension, It is the set of valid coordinates of the mask region along the difference direction. This represents the square of the L2 norm.

[0112] is the edge sparsity constraint, the real system matrix changes sharply at the boundary but the region is small, it can be assumed that the change region has sparsity, and the TV constraint is used to keep the boundary change sparse and suppress unnecessary high-frequency artifacts. Unlike smoothness, TV uses the L1 norm ), which is more conducive to retaining the abrupt characteristics of the edge, as shown in equation (4):

[0113] (4),

[0114] wherein, and respectively represent the influence degree of controlling spatial continuity and the influence degree of sparsity, in the specific experiments of the present application , .

[0115] In the specific experiment process, the training process of the EP-Uformer model is as follows:

[0116] (1) Load the training set and validation set data, including input data, labels and boundary mask information. At the same time, record the mean and standard deviation parameters of the training data for result de-normalization after training.

[0117] (2) Based on the given parameters, construct the model to be trained (i.e. EP-Uformer model), and according to the configuration, choose whether to resume training from the breakpoint state or train from scratch. If it is resumed from the breakpoint, load the saved model parameters and optimizer state.

[0118] (3) Use the learning rate scheduling strategy of step decay (StepLR, step learning rate scheduler) to regularly reduce the learning rate of the optimizer.

[0119] (4) In each training epoch (training period), for each batch of data, forward propagation is used to calculate the prediction result, backward propagation is used to optimize the model parameters, and the training loss is accumulated. At the same time, inference evaluation is carried out on the validation set, the error between the prediction result and the high-resolution label is calculated, and the validation loss is counted. Adjust the learning rate after each epoch.

[0120] (5) In the verification process, the model output is de-normalized from the standardized space back to the original physical space, and the reconstruction error in the complex space is calculated, and the model performance is further quantitatively evaluated based on the normalized root mean square error (nRMSE) index.

[0121] (6) Evaluate the model performance every 5 epochs, and save the current model parameters as the optimal model when the nRMSE index reaches the best.

[0122] In the case where the number of training times does not reach the preset number of training times, operations (4) to (6) are repeatedly performed, and after the training reaches the preset number of training times, the loss, learning rate, nRMSE and other indicators in the training process are completely saved for subsequent analysis and visualization.

[0123] Figure 9 is a schematic diagram of image reconstruction after system matrix extrapolation of EP-Uformer model-based MPI on a simulation dataset according to an embodiment of the present application.

[0124] Figure 8 is a schematic diagram of image reconstruction after system matrix extrapolation of EP-Uformer model-based MPI on a real public dataset according to an embodiment of the present application.

[0125] The EP-Uformer model trained is used to process the system matrix of a smaller field of view in the simulation / real MPI dataset test set to obtain a system matrix of a larger field of view corresponding thereto, and the input (smaller field of view) and output (larger field of view) system matrices are used to reconstruct the MPI image and splice the result as shown in Figure 9 and Figure 8 ; wherein, Figure 8 (a) in (a) of FIG. 8 represents a schematic diagram of a blood vessel phantom, Figure 8 (b) in (b) of FIG. 8 represents an MPI reconstruction image of the blood vessel phantom without overscanning, Figure 8 (c) in (c) of FIG. 8 represents an MPI reconstruction image of the blood vessel phantom predicted by the EP-Uformer model, Figure 8 (d) in (d) of FIG. 8 represents an MPI reconstruction image of the blood vessel phantom overscanned by 10 units; Figure 8 (e) in (e) of FIG. 8 represents a schematic diagram of a nested rectangular phantom, Figure 8 (f) in (f) of FIG. 8 represents an MPI reconstruction image of the nested rectangular phantom without overscanning, Figure 8 (g) in (g) of FIG. 8 represents an MPI reconstruction image of the nested rectangular phantom predicted by the EP-Uformer model, Figure 8 (h) in (h) of FIG. 8 represents an MPI reconstruction image of the nested rectangular phantom overscanned by 10 units; Figure 8 (i) in (i) of FIG. 8 represents a schematic diagram of a uniform block phantom, Figure 8 (j) in (j) of FIG. 8 represents an MPI reconstruction image of the uniform block phantom without overscanning, Figure 9 (k) in (k) of FIG. 8 represents an MPI reconstruction image of the uniform block phantom predicted by the EP-Uformer model, Figure 9 (l) in (l) of FIG. 8 represents an MPI reconstruction image of the uniform block phantom overscanned by 10 units; wherein, Figure 9 the public dataset used is an OpenMPI dataset, wherein, Figure 9 (a) in (a) of FIG. 9 represents a schematic diagram of a 3D phantom in the X-axis, Figure 9(b) in (a) represents a schematic diagram of the 3D phantom MPI image reconstructed by the system matrix without overscanning in the X axis, Figure 9 (c) in (a) represents a schematic diagram of the 3D phantom MPI image reconstructed by the system matrix predicted by the EP-Uformer model in the X axis, Figure 9 (d) in (a) represents a schematic diagram of the 3D phantom MPI image reconstructed by the system matrix with overscanning of 6 units in the X axis; Figure 9 (e) in (a) represents a schematic diagram of the 3D phantom in the Y axis, Figure 9 (f) in (a) represents a schematic diagram of the 3D phantom MPI image reconstructed by the system matrix without overscanning in the Y axis, Figure 9 (g) in (a) represents a schematic diagram of the 3D phantom MPI image reconstructed by the system matrix predicted by the EP-Uformer model in the Y axis, Figure 9 (h) in (a) represents a schematic diagram of the 3D phantom MPI image reconstructed by the system matrix with overscanning of 6 units in the Y axis; Figure 9 (i) in (a) represents a schematic diagram of the 3D phantom in the Z axis, Figure 8 (j) in (a) represents a schematic diagram of the 3D phantom MPI image reconstructed by the system matrix without overscanning in the Z axis, Figure 9 (k) in (a) represents a schematic diagram of the 3D phantom MPI image reconstructed by the system matrix predicted by the EP-Uformer model in the Z axis, Figure 10 (l) in (a) represents a schematic diagram of the 3D phantom MPI image reconstructed by the system matrix with overscanning of 6 units in the Z axis; according to Figure 10 and Figure 10 As shown in the image reconstruction schematic diagram shown in (a) and (b), it can be seen that the boundary artifact is significantly suppressed by using the EP-Uformer model provided by the present application for image reconstruction, which proves the effectiveness of the present application.

[0126] Based on the above multi-block magnetic particle imaging method, the present application further provides a multi-block magnetic particle imaging system. The device will be described in detail below. Figure 11

[0127] Figure 11 is a structural block diagram of the multi-block magnetic particle imaging system according to the embodiment of the present application.

[0128] As shown in (a) and (b), the multi-block magnetic particle imaging system 1000 of this embodiment includes a system matrix acquisition module 1010, a model data processing module 1020, an image reconstruction module 1030, an image cropping module 1040, and an image splicing module 1050. ​

[0129] ​​The system matrix acquisition module 1010 is configured to acquire a plurality of magnetic particle system matrices corresponding to a plurality of sub-imaging regions in the to-be-imaged region respectively; in an embodiment, the system matrix acquisition module 1010 can be configured to perform the operation S210 described above, and details are not repeated herein.

[0130] The model data processing module 1020 is configured to perform extrapolation on the plurality of magnetic particle system matrices by using a trained boundary prediction model to obtain a plurality of extrapolated magnetic particle system matrices, wherein the trained boundary prediction model is trained by using the simulation-generated magnetic particle system matrix samples and the loss function based on the spatial continuity constraint and the edge sparsity constraint; in an embodiment, the model data processing module 1020 can be configured to perform the operation S220 described above, and details are not repeated herein.

[0131] The image reconstruction module 1030 is configured to perform image reconstruction on the plurality of sub-imaging regions by using the plurality of extrapolated magnetic particle system matrices to obtain a plurality of field-of-view expanded magnetic particle image blocks; in an embodiment, the image reconstruction module 1030 can be configured to perform the operation S230 described above, and details are not repeated herein.

[0132] The image cropping module 1040 is configured to perform cropping on the plurality of field-of-view expanded magnetic particle image blocks to obtain a plurality of boundary artifact-eliminated magnetic particle image blocks; in an embodiment, the image cropping module 1040 can be configured to perform the operation S240 described above, and details are not repeated herein.

[0133] The image stitching module 1050 is configured to stitch the plurality of boundary artifact-eliminated magnetic particle image blocks to obtain a magnetic particle imaging result of the to-be-imaged region; in an embodiment, the image stitching module 1050 can be configured to perform the operation S250 described above, and details are not repeated herein.

[0134] The multi-block magnetic particle imaging method and system provided by the application simulate the system matrix of the magnetic particle imaging process under different field-of-view sizes through data simulation, and construct a boundary prediction model combining the U-Net and the Transformer structure. Through the model training process, the spatial continuity, the boundary sparsity and other physical constraints are introduced, so that the boundary prediction model can accurately predict the boundary region of the system matrix, and the system matrix without additional hardware cost and additional scanning is realized. Finally, the trained boundary prediction model is used to extrapolate the small field-of-view system matrix to obtain a complete system matrix of a large field-of-view, effectively improving the imaging field-of-view range of the magnetic particle imaging and the reconstruction quality of the multi-block magnetic particle imaging. The application has the advantages of reducing the complexity of the imaging system, improving the boundary prediction accuracy, reducing the artifacts of the multi-block stitched reconstruction image, and is suitable for system matrix extrapolation of the multi-block magnetic particle imaging system and large field-of-view multi-block high-precision reconstruction application.

[0135] According to an embodiment of the present application, any of the system matrix acquisition module 1010, the model data processing module 1020, the image reconstruction module 1030, the image cropping module 1040 and the image stitching module 1050 can be combined in one module, or any of them can be split into multiple modules. Alternatively, at least part of the function of one or more of these modules can be combined with at least part of the function of other modules and implemented in one module. According to an embodiment of the present application, at least one of the system matrix acquisition module 1010, the model data processing module 1020, the image reconstruction module 1030, the image cropping module 1040 and the image stitching module 1050 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on chip, a system on board, a system on package, an application specific integrated circuit (ASIC), or any other reasonable way of integrating or packaging a circuit, etc. or implemented by hardware or firmware, or implemented in any one of software, hardware and firmware or in a proper combination of any of them. Alternatively, at least one of the system matrix acquisition module 1010, the model data processing module 1020, the image reconstruction module 1030, the image cropping module 1040 and the image stitching module 1050 can be at least partially implemented as a computer program module which, when executed, can perform the corresponding function.

[0136] ​ is a block diagram of an electronic device suitable for implementing a multi-block magnetic particle imaging method according to an embodiment of the present application.

[0137] As shown in ​ , the electronic device 1100 according to an embodiment of the present application includes a processor 1101 which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1102 or a program loaded from a storage portion 1108 into a random access memory (RAM) 1103. The processor 1101 can include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or a related chipset and / or a special-purpose microprocessor (e.g., an application specific integrated circuit (ASIC)), etc. The processor 1101 can also include an on-board memory for cache use. The processor 1101 can include a single processing unit or multiple processing units for performing different actions of the method processes according to an embodiment of the present application.

[0138] In the RAM 1103, various programs and data required for the operation of the electronic device 1100 are stored. The processor 1101, the ROM 1102, and the RAM 1103 are connected to each other via the bus 1104. The processor 1101 performs various operations of the method flow according to the embodiments of the present application by executing the programs in the ROM 1102 and / or the RAM 1103. It is to be noted that the programs can also be stored in one or more memories other than the ROM 1102 and the RAM 1103. The processor 1101 can also perform various operations of the method flow according to the embodiments of the present application by executing the programs stored in the one or more memories.

[0139] According to the embodiments of the present application, the electronic device 1100 can further include an input / output (I / O) interface 1105, which is also connected to the bus 1104. The electronic device 1100 can further include one or more of the following components connected to the input / output (I / O) interface 1105: an input part 1106 including a keyboard, a mouse, etc.; an output part 1107 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage part 1108 including a hard disk, etc.; and a communication part 1109 including a network interface card such as a LAN card, a modem, etc. The communication part 1109 performs communication processing via a network such as the Internet. A drive 1110 is also connected to the input / output (I / O) interface 1105 as necessary. A removable recording medium 1111 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is attached to the drive 1110 as necessary, so that a computer program read out therefrom is installed in the storage part 1108 as necessary.

[0140] The present application also provides a computer readable storage medium, which can be included in the device / apparatus / system described in the above embodiments; or can exist separately without being assembled into the device / apparatus / system. The above computer readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of the present application.

[0141] According to embodiments of the present application, the computer readable storage medium can be a non-transitory computer readable storage medium, such as, for example, without limitation, a portable computer diskette, a hard disk, random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the present application, a computer readable storage medium can be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. For example, according to embodiments of the present application, the computer readable storage medium can include one or more of the above-described ROM 1102 and / or RAM 1103 and / or one or more memory devices other than the ROM 1102 and the RAM 1103.

[0142] The flow diagrams and the block diagrams in the drawings are illustrations of possible architectures, functions, and operations for systems, methods, and computer program products according to various embodiments of the present application. In this regard, each block in the flow diagrams or block diagrams can represent a module, a segment, or a portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may

[0143] Those skilled in the art will understand that features recited in various embodiments of the present application can be combined and / or integrated in various ways, even if such combinations or integrations are not expressly noted in the present application. In particular, features recited in various embodiments of the present application can be combined and / or integrated in ways that are not expressly noted in the present application, without departing from the spirit and teachings of the present application. All such combinations and / or integrations are within the scope of the present application.

[0144] The embodiments of the present application have been described above. However, these embodiments are merely meant to be illustrative, and are not meant to limit the scope of the present application. Although each of the embodiments has been described above separately, this does not mean that the measures in each of the embodiments cannot be used advantageously in combination. Various alternatives and modifications can be made to the embodiments of the present application by those skilled in the art without departing from the scope of the present application, and all such alternatives and modifications are intended to fall within the scope of the present application.

Claims

1. A multi-slab magnetic particle imaging method, characterized by, The method comprises: Collecting a plurality of magnetic particle system matrices respectively corresponding to a plurality of sub-imaging regions in a region to be imaged; Performing extrapolation operation on the plurality of magnetic particle system matrices by using the trained boundary prediction model to obtain a plurality of matrix extrapolated magnetic particle system matrices, wherein the trained boundary prediction model is trained by using simulation generated magnetic particle system matrix samples and a loss function based on spatial continuity constraint and edge sparsity constraint; Performing image reconstruction on the plurality of sub-imaging regions by using the plurality of matrix extrapolated magnetic particle system matrices to obtain a plurality of field of view expanded magnetic particle image blocks; Performing cropping on the plurality of field of view expanded magnetic particle image blocks to obtain a plurality of boundary artifact eliminated magnetic particle image blocks; Splicing the plurality of boundary artifact eliminated magnetic particle image blocks to obtain a magnetic particle imaging result of the region to be imaged.

2. The method of claim 1, wherein, The trained boundary prediction model is trained by using simulation generated magnetic particle system matrix samples and a loss function based on spatial continuity constraint and edge sparsity constraint, comprising: Simulate the magnetic particle system matrix samples based on parameter setting information, wherein the magnetic particle system matrix samples include specified field of view samples and field of view extrapolation samples; Process the specified field of view samples by using the boundary prediction model to obtain a field of view extrapolated predicted system matrix; Process the field of view extrapolation samples and the field of view extrapolated predicted system matrix by using the loss function based on spatial continuity constraint and edge sparsity constraint, and perform parameter update on the boundary prediction model by using the obtained loss value; Iteratively perform data sample processing operation of the model, calculation operation of the loss value and parameter update operation of the model until a preset training condition is met to obtain the trained boundary prediction model.

3. The method of claim 2, wherein, Simulate the magnetic particle system matrix samples based on parameter setting information, comprising: Based on the parameter setting information and the specified size of the field of view, perform superposition operation of static field and driving field in the two-dimensional imaging space to obtain a magnetic field distribution model; According to the sampling parameters in the parameter setting information, calculate the harmonic vector and the mixed frequency vector in the sampling frequency range by using the magnetic field distribution model to obtain the acquisition frequency, determine the position information of the harmonic vector by analyzing the relationship between the sampling frequency and the acquisition frequency, and perform frequency feature analysis by using the position information to obtain the number of effective frequencies; Generate an initial specified field of view sample by using an initial magnetic particle concentration distribution matrix, the magnetic field distribution model, the starting point coordinate information of the two-dimensional imaging space, the number of effective frequencies and the number of calibration additional points, and perform system matrix calibration operation on the initial specified field of view sample to obtain the specified field of view sample.

4. The method of claim 3, wherein, Based on the parameter setting information and the specified size of the field of view, perform superposition operation of static field and driving field in the two-dimensional imaging space to obtain a magnetic field distribution model, comprising: Based on the specified size of the field of view, the sampling parameters and the scanning time parameters in the parameter setting information are calculated to generate a coordinate grid of the two-dimensional imaging space, and the coordinate grid is used to derive static components of the static field in different directions; The magnetic field direction setting constraint is performed on the static components, and the scaling ratio of the specific imaging region, the magnetic field component after the imaging target position is moved, and the drive field gradient are calculated using the constrained static components and the specific imaging region parameters in the parameter setting information, and then the superposition of the static field and the drive field is completed to obtain the magnetic field distribution model.

5. The method of claim 3, wherein, The system matrix calibration operation is performed on the initial specified field of view sample to obtain the specified field of view sample, including: For the pixel points with non-zero concentration in the two-dimensional imaging space, the magnetic moment components in different directions are calculated by a preset algorithm using the magnetic field distribution model; The magnetic moment components are mapped to the positions of the corresponding target frequency vectors according to the effective frequency index, and the mapped magnetic moment components are integrated to complete the magnetic particle system matrix calibration operation to obtain the field of view extrapolation sample in the specified field of view, wherein the target frequency vector corresponds to the initial specified field of view sample.

6. The method of claim 3, wherein, Further comprising: According to the parameter setting information, data simulation is performed on the field of view after the size extrapolation by the magnetic field distribution modeling operation, the frequency feature analysis operation and the system matrix calibration operation to obtain the field of view extrapolation sample in the magnetic particle system matrix sample.

7. The method of claim 2, wherein, The specified field of view sample is processed using the boundary prediction model to obtain the predicted system matrix of the field of view extrapolation, including: A multi-level spatial downsampling operation is performed on the specified field of view sample using the hierarchical encoder of the boundary prediction model to obtain multi-level semantic features; A global dependency coding modeling is performed on the multi-level semantic features using the global context modeling module of the boundary prediction model to obtain multi-level coding features; A multi-level upsampling operation is performed on the multi-level coding features using the structure reconstruction decoding module of the boundary prediction model to complete the same level skip connection with the multi-level semantic features to obtain the initial predicted system matrix of the field of view extrapolation; A size alignment is performed on the initial predicted system matrix of the field of view extrapolation using the output alignment module of the boundary prediction model to obtain the aligned predicted system matrix; The feature map dimension of the aligned predicted system matrix is mapped to the target dimension using the output mapping generation module of the boundary prediction model to obtain the predicted system matrix of the field of view extrapolation.

8. The method of claim 7, wherein, Each sub-encoder of the hierarchical encoder includes a first convolutional layer, a first normalization layer, a pooling layer and an activation function; The global context modeling module includes a plurality of self-attention encoders, each of which corresponds to a sub-encoder; Each sub-decoder of the structure reconstruction decoding module includes an upsampling layer, a second convolutional layer and a second normalization layer, and the sub-decoder is connected to the sub-encoder of the same level through the self-attention encoder of the same level.

9. The method of claim 1, wherein, The image reconstruction is performed on the plurality of sub-imaging areas by using the plurality of matrix extrapolated magnetic particle system matrices, and a plurality of field-of-view expanded magnetic particle image blocks are obtained. The initial solution vector corresponding to each sub-imaging area is orthogonally projected by using the matrix extrapolated magnetic particle system matrix corresponding to each sub-imaging area, and an updated solution vector is obtained, and the regularization constraint is performed on the updated solution vector. The step optimization is performed on the matrix extrapolated magnetic particle system matrix, and the step optimized magnetic particle system matrix is used to perform the next round of solution vector updating operation. The solution vector updating operation, the regularization constraint operation and the step optimization operation are iteratively performed until a preset condition is met, and the field-of-view expanded magnetic particle image block corresponding to each sub-imaging area is obtained.

10. A multi-slab magnetic particle imaging system, characterized by, The system comprises: A system matrix acquisition module is configured to acquire magnetic particle system matrices corresponding to a plurality of sub-imaging areas in a to-be-imaged region respectively; A model data processing module is configured to perform extrapolation operation on the plurality of magnetic particle system matrices by using a trained boundary prediction model, and obtain a plurality of matrix extrapolated magnetic particle system matrices, wherein the trained boundary prediction model is trained by using a simulation generated magnetic particle system matrix sample and a loss function based on spatial continuity constraint and edge sparsity constraint; An image reconstruction module is configured to perform image reconstruction on the plurality of sub-imaging areas by using the plurality of matrix extrapolated magnetic particle system matrices, and obtain a plurality of field-of-view expanded magnetic particle image blocks; An image cropping module is configured to perform cropping on the plurality of field-of-view expanded magnetic particle image blocks, and obtain a plurality of boundary artifact eliminated magnetic particle image blocks; An image stitching module is configured to stitch the plurality of boundary artifact eliminated magnetic particle image blocks, and obtain a magnetic particle imaging result of the to-be-imaged region.

Images