MPI acceleration calibration method based on system matrix super-resolution network

Abstract

The application belongs to the field of magnetic particle imaging, and particularly relates to an MPI acceleration calibration method based on a system matrix super-resolution network, aiming to solve the problems of long calibration time and poor accuracy of reconstructed MPI images of the existing magnetic particle imaging system matrix reconstruction method. The method comprises the following steps: placing a unit MNP sample in the imaging field of view of a magnetic particle imaging device for sparse calibration scanning, and constructing a system matrix as a low-resolution system matrix; inputting the low-resolution system matrix into a system matrix super-resolution network model after pretreatment, to obtain a super-resolved system matrix row set; performing post-processing on the super-resolved system matrix row set to obtain a high-resolution system matrix; and combining the high-resolution system matrix to solve the magnetic particle concentration distribution of a target object to be imaged, and reconstruct an MPI image of the target object to be imaged. The application can greatly reduce the calibration time while accurately reconstructing the MPI image.

Description

Technical Field

[0001] This invention belongs to the field of magnetic particle imaging, specifically relating to an MPI accelerated calibration method, system, and electronic device based on a system matrix super-resolution network. Background Technology

[0002] Magnetic particle imaging (MPI) is a novel tracer-based imaging technique in the field of medical imaging. It is a tracer method for detecting the spatial distribution of superparamagnetic iron oxide nanoparticles (SPIONs). MPI does not display the anatomical structure of the subject, has no background signal interference, and the signal intensity is proportional to the tracer concentration. It is a functional imaging technique that can obtain quantitative data and has broad clinical application prospects.

[0003] MPI image reconstruction methods are divided into two main categories: x-space based reconstruction methods and system matrix based reconstruction methods. x-space based reconstruction methods directly image the time-domain signal, resulting in low spatial resolution and artifacts. System matrix based reconstruction methods perform a Fourier transform on the time-domain signal acquired by MPI and reconstruct an image with higher spatial resolution using the transformed frequency-domain signal.

[0004] There are generally two methods for obtaining the system matrix: the mathematical model-based method and the measurement-based method. The mathematical model-based method requires system parameters of the MPI (Magnetic Process Interference Model), such as the magnetic field vector, magnetization curve, and transfer function of the receiver chain, to construct the system matrix. However, due to the relaxation effect, the dynamic behavior of particles is difficult to model accurately, and the particle models considered so far are only approximations; therefore, it is impossible to accurately obtain the MPI system matrix.

[0005] Measurement-based system matrix acquisition methods employ a grid calibration approach. A calibration robot moves a unit sample containing magnetic nanoparticles (MNPs) of a specific volume and concentration. At all grid locations within the imaging field of view, the induced voltage signals on all receiving coils are measured. These voltage signals undergo time-frequency domain transformation, and the system matrix is ​​constructed by combining the frequency domain signals acquired from all grid points in a specific order. This measurement-based method does not require information on MPI system parameters or particle modeling, thus improving the accuracy of the system matrix. However, it is extremely time-consuming. For example, calibrating a relatively dense 66*66*27 grid area in a 3D MPI imaging field requires 117,612 scans. If the sum of the data acquisition time and sample movement time for each grid is 1.2 seconds, the total calibration time would be as long as 39 hours. Using a sparsely grid-calibrated system matrix for concentration reconstruction would reduce the resolution of the reconstructed image. Furthermore, changing the magnetic particles or scanning trajectory also requires recalibrating the system matrix, limiting the use of MPI.

[0006] Traditional compressed sensing methods can reduce the calibration scanning time of MPI system matrices by performing high-resolution reconstruction of randomly sampled sparse system matrices. However, random sparse sampling is difficult to implement in hardware, and compressed sensing methods cannot learn the general features and prior knowledge of the system matrix, resulting in inaccurate reconstructed MPI images. Therefore, using deep learning methods to perform super-resolution on low-resolution system matrices can significantly reduce calibration time while acquiring dense, high-resolution system matrices and accurately reconstructing MPI particle concentration distribution images. Based on this, this invention proposes an accelerated MPI calibration method based on a system matrix super-resolution network. Summary of the Invention

[0007] To address the aforementioned problems in existing technologies, namely the long calibration time and poor reconstruction accuracy of traditional methods in existing magnetic particle imaging system matrix reconstruction methods, this invention provides an MPI-accelerated calibration method based on a system matrix super-resolution network. This method is used to recover the high-resolution system matrix and then accurately reconstruct the image of the target object. The method includes:

[0008] Step S100: Place a unit MNP sample with a certain volume and concentration in the imaging field of the magnetic particle imaging device for sparse calibration scanning, and construct a system matrix as a low-resolution system matrix.

[0009] Step S200: Preprocess the low-resolution system matrix, and input the preprocessed system matrix into the trained system matrix super-resolution network model to obtain the super-resolution system matrix row set; Postprocess the super-resolution system matrix row set to obtain the high-resolution system matrix.

[0010] Step S300: Combine the high-resolution system matrix to solve for the magnetic particle concentration distribution of the target object to be imaged, and reconstruct the MPI image of the target object to be imaged.

[0011] In some preferred embodiments, the preprocessing includes background reduction processing, signal-to-noise ratio threshold filtering, noise whitening processing, real and imaginary part splitting, and matrix transformation operations; the postprocessing includes matrix dimension transformation and real and imaginary part combination.

[0012] In some preferred embodiments, the system matrix super-resolution network model includes a first convolutional module, multiple residual groups, a scaling module, a bicubic interpolation residual connection module, and a second convolutional module;

[0013] The input to the first convolution module is the preprocessed low-resolution system matrix rows;

[0014] Multiple residual groups are connected in series. The first convolution module is connected to the first residual group among the multiple residual groups, and the output of the first convolution module is added element by element to the output of the last residual group and used as the input of the scale-up module.

[0015] The scaling module is connected to the second convolution module;

[0016] The preprocessed low-resolution system matrix rows are processed by the bicubic interpolation residual connection module and then added element by element to the output of the second convolution module to obtain the super-resolution system matrix row set.

[0017] The bicubic interpolation residual connection module is used to upsample the super-resolution scaling coefficients of the preprocessed low-resolution system matrix rows using the bicubic interpolation method.

[0018] In some preferred embodiments, the residual group includes at least one wide activation residual block and one dynamic convolutional residual block; the wide activation residual block is connected to the dynamic convolutional residual block;

[0019] The wide activation residual block is constructed based on a first convolutional layer, an activation function layer, a second convolutional layer, and a third convolutional layer connected in sequence; the first convolutional layer and the second convolutional layer have the same size; the second convolutional layer and the third convolutional layer have different sizes;

[0020] The dynamic convolutional residual block includes two context-modulated dynamic convolutional layers; there is an activation function layer between the two dynamic convolutional layers;

[0021] The dynamic convolutional layer is constructed based on a pooling operation unit, a context encoding operation unit, a channel interaction operation unit, a context decoding operation unit, and a modulation operation unit connected in sequence.

[0022] The pooling operation unit is constructed based on a pooling layer; the context encoding operation unit is constructed based on a sequentially connected linear layer, batch normalization layer, and activation function layer; the channel interaction operation unit is constructed based on a sequentially connected grouped linear layer, batch normalization layer, and activation function layer; the context decoding operation unit includes two linear layers, wherein the input of one linear layer is the output of the channel interaction operation unit, and the input of the other linear layer is the output of the context encoding operation unit. The outputs of the two linear layers are added element-wise to obtain the output of the context decoding operation unit, thus obtaining a context modulation mask; the modulation operation unit is used to modulate the convolution kernel using the context modulation mask.

[0023] In some preferred embodiments, the scaling module includes at least one upsampling module and a bicubic interpolation module, which are connected in sequence.

[0024] The upsampling module is built based on a convolutional layer and a pixelshuffle module connected in sequence.

[0025] In some preferred embodiments, the training method for the system matrix super-resolution network model is as follows:

[0026] Step A100: Obtain the high-resolution system matrix and downsample it to obtain the low-resolution system matrix; split the high-resolution system matrix and low-resolution system matrix pair into high-resolution system matrix rows and low-resolution system matrix row pairs, and then perform preprocessing; use the preprocessed high-resolution system matrix rows and low-resolution system matrix row pairs as training samples, and use the high-resolution system matrix rows corresponding to the training samples as ground truth labels to construct a training set; the high-resolution system matrix includes high-resolution system matrices from the same or different devices, the same or different magnetic particles from the same device, and the same or different system parameters from the same device; the downsampling includes step downsampling and average downsampling.

[0027] Step A200: Input the training samples in the training set into the pre-constructed system matrix super-resolution network model to obtain the set of rows of the super-resolution system matrix, and then obtain the high-resolution system matrix as the reconstruction matrix;

[0028] Step A300: Based on the reconstruction matrix and its corresponding truth labels, calculate the loss value and update the model parameters of the system matrix super-resolution network model;

[0029] Step A400, repeat steps A200-A300 until the trained system matrix super-resolution network model is obtained.

[0030] In some preferred embodiments, the loss function of the system matrix super-resolution network model during training is:

[0031]

[0032] Where Loss represents the loss function, k represents the number of rows in the system matrix participating in training, and ||·||1 represents the first-order norm. This is the set of rows of the preprocessed low-resolution system matrix. θ represents the label corresponding to the set of rows of the high-resolution system matrix, θ represents the parameters of the system matrix super-resolution network model, f represents the mapping function of the system matrix super-resolution network model, and i represents the sequence number of the system matrix row.

[0033] A second aspect of the present invention proposes an MPI accelerated calibration system based on a system matrix super-resolution network for recovering a high-resolution system matrix, thereby performing MPI image reconstruction on the target object to be imaged. The system includes:

[0034] The sparse system matrix acquisition module is configured to place a unit MNP sample with a certain volume and concentration in the imaging field of the magnetic particle imaging device for sparse calibration scanning and construct a system matrix as a low-resolution system matrix.

[0035] The dense system matrix recovery module is configured to preprocess the low-resolution system matrix, input the preprocessed matrix into a trained system matrix super-resolution network model to obtain a set of super-resolution system matrix rows; and post-process the set of super-resolution system matrix rows to obtain a high-resolution system matrix.

[0036] The image reconstruction module is configured to combine the high-resolution system matrix to solve for the magnetic particle concentration distribution of the target object to be imaged, and reconstruct the MPI image of the target object to be imaged.

[0037] A third aspect of the present invention provides an electronic device comprising: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor for implementing the above-described MPI accelerated calibration method based on a system matrix super-resolution network.

[0038] In a fourth aspect, the present invention provides a computer-readable storage medium storing computer instructions for execution by a computer to implement the above-described MPI accelerated calibration method based on a system matrix super-resolution network.

[0039] The beneficial effects of this invention are:

[0040] This invention can accurately reconstruct MPI images while greatly reducing calibration time.

[0041] 1) This invention performs super-resolution on the MPI system matrix, requiring only the acquisition of the sparse system matrix during the MPI calibration scan, which greatly reduces the calibration scan time of the MPI system matrix and improves the efficiency of MPI utilization; and compared with the traditional compressed sensing method, the deep learning method can accurately recover the high-resolution system matrix by utilizing the general features and prior knowledge of the system matrix.

[0042] 2) This invention designs a system matrix super-resolution network model tailored to the characteristics of the system matrix, improving the super-resolution effect of the system matrix: The bicubic interpolation residual connection proposed in this invention fully utilizes the low-frequency structural information of the system matrix rows; the context-modulated dynamic convolution proposed in this invention fully utilizes the contextual relationships between the system matrix rows, extracting global contextual features through dynamic modulation of the convolution module, which is superior to ordinary convolution for extracting local features; the wide activation residual block mechanism proposed in this invention can retain more information before nonlinear activation, improving the super-resolution effect;

[0043] 3) This invention improves the reconstruction quality of MPI particle concentration distribution images. Compared with images reconstructed using low-resolution system matrices without super-resolution, images reconstructed using high-resolution system matrices after super-resolution have higher quality and higher spatial resolution. Attached Figure Description

[0044] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0045] Figure 1 This is a flowchart illustrating an embodiment of the MPI accelerated calibration method based on a system matrix super-resolution network according to the present invention.

[0046] Figure 2 This is a schematic diagram of the framework of an MPI accelerated calibration system based on a system matrix super-resolution network according to an embodiment of the present invention;

[0047] Figure 3This is a simplified flowchart illustrating an embodiment of the MPI accelerated calibration method based on a system matrix super-resolution network according to the present invention.

[0048] Figure 4 This is a schematic diagram of the structure of a system matrix super-resolution network model according to an embodiment of the present invention;

[0049] Figure 5 This is a schematic diagram of the structure of the residual group according to an embodiment of the present invention;

[0050] Figure 6 This is a schematic diagram of the structure of a dynamic convolutional layer according to an embodiment of the present invention;

[0051] Figure 7 This is a schematic diagram of the scale magnification module according to an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0054] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0055] This invention discloses an MPI-accelerated calibration method based on a system matrix super-resolution network, used to recover a high-resolution system matrix, thereby enabling accurate image reconstruction of the target object to be imaged. Figure 1 As shown, it includes the following steps:

[0056] Step S100: Place a unit MNP sample with a certain volume and concentration in the imaging field of the magnetic particle imaging device for sparse calibration scanning, and construct a system matrix as a low-resolution system matrix.

[0057] Step S200: Preprocess the low-resolution system matrix, and input the preprocessed system matrix into the trained system matrix super-resolution network model to obtain the super-resolution system matrix row set; Postprocess the super-resolution system matrix row set to obtain the high-resolution system matrix.

[0058] Step S300: Combine the high-resolution system matrix to solve for the magnetic particle concentration distribution of the target object to be imaged, and reconstruct the MPI image of the target object to be imaged.

[0059] To more clearly explain the MPI accelerated calibration method based on system matrix super-resolution networks of this invention, the following description is in conjunction with the appendix. Figure 3 The steps in the embodiments of the method of the present invention are described in detail below.

[0060] In the following embodiments, the generation process of the system matrix super-resolution network model is first explained, and then the process of image reconstruction using the MPI accelerated calibration method based on the system matrix super-resolution network is described in detail.

[0061] 1. Training process of the system matrix super-resolution network model

[0062] Step A100: Obtain the high-resolution system matrix and downsample it to obtain the low-resolution system matrix; split the high-resolution system matrix and low-resolution system matrix pair into high-resolution system matrix rows and low-resolution system matrix row pairs, and then perform preprocessing; use the preprocessed high-resolution system matrix rows and low-resolution system matrix row pairs as training samples, and use the high-resolution system matrix rows corresponding to the training samples as ground truth labels to construct a training set; the high-resolution system matrix includes high-resolution system matrices from the same or different devices, the same or different magnetic particles from the same device, and the same or different system parameters from the same device; the downsampling includes step downsampling and average downsampling.

[0063] In this embodiment, high-resolution system matrices are obtained from the same or different devices, from the same or different magnetic particles within the same device, and from the same or different system parameters within the same device. These high-resolution system matrices are then processed using either step downsampling or average downsampling to obtain low-resolution system matrices according to different super-resolution scale requirements. Step downsampling simulates the calibration scan of system matrices for small-volume unit MNP samples, while average downsampling simulates the calibration scan of large-volume unit MNP samples. The two downsampling methods are as follows:

[0064] S LR =Downsample strided (S HR (1)

[0065] S LR=Downsample averaged (S HR (2)

[0066] Among them, S HR S represents the high-resolution system matrix. LR Represents a low-resolution system matrix, Downsample strided (·) indicates step downsampling. averaged (·) indicates average downsampling.

[0067] The complete system matrix Split into system matrix rows The super-resolution process is carried out with the system matrix row as the smallest basic unit, where p represents the number of system matrices; m represents the number of system matrix rows, i.e. the number of imaging frequency points; n represents the number of system matrix columns, i.e. the number of imaging voxels; j represents the sequence number of the system matrix row, and j = 1, 2, ..., q; q represents the total number of system matrix rows, and q = p × m;

[0068] Preprocessing operations are performed on the split system matrix rows, including:

[0069] Background reduction processing is applied to the rows of the system matrix;

[0070] The system matrix rows are filtered based on the signal-to-noise ratio (SNR) at each frequency point, and the system matrix rows with high SNR (e.g., SNR>5) are selected, as shown below:

[0071] S i＝1，2，...，l =S j＝1，2，...，q (SNR j >5) (3)

[0072] Among them, SNR j denoted by , i represents the signal-to-noise ratio at the frequency point where the j-th system matrix row is located, i represents the sequence number of the filtered system matrix row, and l represents the total number of filtered system matrix rows.

[0073] The system matrix rows are whitened based on the standard deviation of the system matrix rows and the signal-to-noise ratio of the corresponding frequency points, as shown below:

[0074]

[0075] in, std(S) represents the whitened system matrix rows. i ) represents the standard deviation of the i-th row of the system matrix.

[0076] The whitened system matrix rows are split into real and imaginary parts and processed as dual-channel data, as shown below:

[0077]

[0078] in, This represents the set of rows in a dual-channel system matrix. Represents the system matrix row S i The set of real parts, Represents the system matrix row S i The set of imaginary parts.

[0079] Matrix transformations can convert a one-dimensional system matrix into a two-dimensional or three-dimensional matrix. Taking two-dimensional matrix as an example, it is represented as follows:

[0080]

[0081] Where Reshape2D(·) represents a matrix transformation from one dimension to two dimensions. Let n represent the set of rows (one-dimensional vector form) of the system matrix before transformation, where n represents the length of the row vectors of the system matrix before transformation. Let h represent the set of rows (two-dimensional matrix form) of the transformed system matrix, where h represents the height of the transformed two-dimensional matrix and w represents the width of the transformed two-dimensional matrix.

[0082] Step A200: Input the training samples in the training set into the pre-constructed system matrix super-resolution network model to obtain the set of rows of the super-resolution system matrix, and then obtain the high-resolution system matrix as the reconstruction matrix;

[0083] In this embodiment, the pre-built dataset is divided into training, validation, and test sets in a 6:2:2 ratio. A system matrix super-resolution network model is designed and built, and the dataset is used to train the network and determine the parameters of the network model.

[0084] The system matrix super-resolution network model includes a first convolutional module (i.e. Figure 4 The convolutional layer located before the residual group, multiple residual groups, a scaling module, a bicubic interpolation residual connection module, and a second convolutional module (i.e., Figure 4 (The convolutional layer following the mesoscale magnification module);

[0085] The input to the first convolutional module is the preprocessed low-resolution system matrix row; after the preferred dual-channel low-resolution system matrix row is input into the system matrix super-resolution network model, the features are first extracted through a 3x3 convolutional layer (i.e., the first convolutional module) with 2 input channels and 64 output channels to form a feature map with 64 channels, wherein the stride of the convolutional layer is 1 and the padding is 1.

[0086] Multiple residual groups are connected in series. The first convolution module is connected to the first residual group among the multiple residual groups, and the output of the first convolution module is added element by element to the output of the last residual group and used as the input of the scale-up module.

[0087] The scaling module is connected to the second convolution module;

[0088] After the low-resolution system matrix rows are processed by the bicubic interpolation residual connection module, they are added element-wise to the output of the second convolution module to obtain the set of super-resolution system matrix rows.

[0089] The bicubic interpolation residual connection module (i.e. Figure 4 The bicubic interpolation residual connection is used to upsample the super-resolution scaling coefficients of the preprocessed low-resolution system matrix rows using the bicubic interpolation method; the global bicubic interpolation residual connection is realized by making long hop connections with the entire neural network to further make full use of the low-frequency structural information of the system matrix rows.

[0090] The residual group includes at least one wide activation residual block and one dynamic convolutional residual block; the wide activation residual block is connected to the dynamic convolutional residual block, as shown below. Figure 5 As shown;

[0091] The wide activation residual block is constructed based on a first convolutional layer, an activation function layer, a second convolutional layer, and a third convolutional layer connected in sequence. The first and second convolutional layers are of the same size, preferably 1x1 convolutional layers. The 1x1 convolution augments the feature map, expanding the activation space before the nonlinear operation of the activation function, thus retaining more information with almost the same computational cost. The second and third convolutional layers are of different sizes, with the third convolutional layer preferably being a 3x3 convolutional layer. Specifically:

[0092] The input feature map has 64 channels. After the first 1x1 convolutional layer, the number of feature map channels is increased to 64 × 6 = 384. For example, the activation function is LeakyReLU to expand the non-linear activation range. After the activation function, a 1x1 convolutional layer is connected, which initially reduces the number of feature map channels to floor(384 × 0.8) = 307, where floor(x) represents rounding down x. After this, a 3x3 ordinary convolutional layer is connected, which further reduces the number of feature map channels to 64, restoring the original number of feature map channels. The stride of the 1x1 convolutional layer is 1 and the padding is 0. The stride of the 3x3 convolutional layer is 1 and the padding is 1.

[0093] The dynamic convolutional residual block includes two context-modulated dynamic convolutional layers; there is an activation function layer between the two dynamic convolutional layers; the kernel size of the dynamic convolutional layer is 3x3, the stride is 1, and the padding is 1; the activation function is LeakyReLU.

[0094] The dynamic convolutional layer is constructed based on a series of sequentially connected pooling operation units, context encoding operation units, channel interaction operation units, context decoding operation units, and modulation operation units; such as Figure 6 As shown, global context information is extracted by performing average pooling on the feature maps of the system matrix rows; the context encoding unit encodes the context information; the encoded context information is divided into two paths, one of which undergoes channel interaction; the context information after channel interaction is integrated with the context information without channel interaction, and then decoded through a linear layer to obtain a context modulation mask. This mask is used to modulate ordinary convolutional kernels to obtain convolutional modules that can dynamically modulate according to context information. The dynamic convolutional residual blocks composed of these dynamic convolutional modules can enhance the network's attention to global features and improve the super-resolution effect of the system matrix.

[0095] Specifically: the pooling operation unit is constructed based on a pooling layer, with the output size preferably set to 3x3, and is converted into a vector of length 9; the context encoding operation unit is constructed based on a linear layer (9 input channels, 5 output channels), a batch normalization layer, and an activation function layer (using LeakyReLU) connected in sequence; the channel interaction operation unit is constructed based on a grouped linear layer, a batch normalization layer, and an activation function layer connected in sequence (i.e., first grouping, dividing the feature map into 4 groups, each with 16 channels, and then inputting them into a linear layer, where the number of input and output channels of the linear layer is 16, and the activation function is Softmax); the context decoding operation unit includes two linear layers (two linear layers with 5 input channels and 9 output channels), where the input of one linear layer is the output of the channel interaction operation unit, and the input of the other linear layer is the output of the context encoding operation unit. The outputs of the two linear layers are added element-wise as the output of the context decoding operation unit to obtain a context modulation mask of size 3x3. The modulation operation unit is used to modulate a 3x3 convolutional kernel using the context modulation mask. Specifically, it multiplies a mask containing contextual and channel attention information element-wise with a regular convolutional kernel (a static convolutional kernel, as generally defined in convolutional neural networks) to obtain a dynamic convolutional module containing contextual information. The modulation operation unit is built upon convolutional layers. The context modulation mask is a coefficient of the same scale as the convolutional kernel; multiplying this coefficient element-wise with the convolutional kernel completes the modulation of the convolutional kernel.

[0096] The scaling module includes at least one upsampling module and a bicubic interpolation module, which are connected sequentially; as shown in the example Figure 7 As shown;

[0097] The upsampling module is based on sequentially connected convolutional layers (kernel size 3x3, stride 1, padding 1, input channels are the same as the feature map channels, 64, output channels are 64×2). 2 =256), Pixelshuffle module construction.

[0098] The Pixelshuffle module has an upsampling factor of 2, achieving a 2x upsampling and restoring the feature map channel count to 64; the number of upsampling modules is num. upsampler Its maximum value depends on the specific super-resolution task. It is related to the scale factor r of the super-resolution, and is expressed as For example, in an 8x super-resolution task, a maximum of 3 upsampling modules are used; in order to enhance the low-frequency structural information of the system matrix rows, a bicubic interpolation module is placed at the end of the scaling module, and the upsampling factor of the bicubic interpolation module is always equal to the super-resolution scaling factor.

[0099] The second convolutional module is preferably set to have 64 input channels and 2 output channels in a 3x3 convolutional layer, restoring the original dual-channel system matrix row data.

[0100] The mapping function of the system matrix super-resolution network model is:

[0101]

[0102] This is the set of rows of the preprocessed low-resolution system matrix. θ represents the set of rows of the high-resolution system matrix output by the network, and θ represents the network model parameters.

[0103] Step A300: Based on the reconstruction matrix and its corresponding truth labels, calculate the loss value and update the model parameters of the system matrix super-resolution network model;

[0104] In this embodiment, the loss function of the system matrix super-resolution network model during training is:

[0105]

[0106] Where Loss represents the loss function, k represents the number of rows in the system matrix participating in training, and ||•|1 represents the first-order norm. This is the set of rows of the preprocessed low-resolution system matrix. θ represents the label corresponding to the set of rows of the high-resolution system matrix, θ represents the parameters of the system matrix super-resolution network model, f represents the mapping function of the system matrix super-resolution network model, and i represents the sequence number of the system matrix row.

[0107] Step A400, repeat steps A200-A300 until the trained system matrix super-resolution network model is obtained.

[0108] 2. A High-Resolution Reconstruction Method for Magnetic Particle Images Based on System Matrix Super-Resolution Networks

[0109] Step S100: Place a unit MNP sample with a certain volume and concentration in the imaging field of the magnetic particle imaging device for sparse calibration scanning, and construct a system matrix as a low-resolution system matrix.

[0110] In this embodiment, it is preferable to use small-volume or large-volume unit MNP samples to perform sparse system matrix calibration scanning to obtain a low-resolution system matrix.

[0111] Step S200: Preprocess the low-resolution system matrix, and input the preprocessed system matrix into the trained system matrix super-resolution network model to obtain the super-resolution system matrix row set; Postprocess the super-resolution system matrix row set to obtain the high-resolution system matrix.

[0112] In this embodiment, the preprocessing of the low-resolution system matrix includes background reduction, signal-to-noise ratio threshold filtering, noise whitening, separation of real and imaginary parts, and matrix dimension transformation to construct a set of low-resolution system matrix rows. This set of low-resolution system matrix rows is then input into a trained system matrix super-resolution network model to obtain the super-resolution system matrix row set, represented as:

[0113]

[0114] in, This represents the set of rows of the system matrix after super-resolution. Let f represent the set of rows of the low-resolution system matrix, f represent the mapping function of the system matrix super-resolution network model, and θ represent the set of rows of the low-resolution system matrix. trained The parameters of the trained network model are represented by , i represents the sequence number of the system matrix row, and l represents the total number of system matrix rows. The post-processing steps, including matrix dimension transformation and real / imaginary part combination, are performed on the super-resolution system matrix row set to construct the high-resolution system matrix.

[0115] Step S300: Combine the high-resolution system matrix to solve for the magnetic particle concentration distribution of the target object to be imaged, and reconstruct the MPI image of the target object to be imaged.

[0116] In this embodiment, image reconstruction can be represented as an inverse problem. The high-resolution system matrix obtained from the super-resolution module is used to solve for the magnetic particle concentration distribution of the target object. The solution formula is as follows:

[0117]

[0118] Where u represents the frequency domain signal of the induced voltage measurement value of the target object to be imaged. Let be the dense system matrix after super-resolution, i.e., the high-resolution system matrix, and c represent the magnetic particle concentration distribution of the target object to be imaged. For example, the solution method can use methods well-known in the art, such as the Kaczmarz method or the Alternating Direction Method of Multipliers (ADMM).

[0119] A second embodiment of the present invention provides an MPI accelerated calibration system based on a system matrix super-resolution network, used to recover a high-resolution system matrix, and then perform MPI image reconstruction on the target object to be imaged, such as... Figure 2 As shown, it includes:

[0120] The sparse system matrix acquisition module 100 is configured to place a unit MNP sample with a certain volume and a certain concentration in the imaging field of the magnetic particle imaging device for sparse calibration scanning and to construct a system matrix as a low-resolution system matrix.

[0121] The dense system matrix recovery module 200 is configured to preprocess the low-resolution system matrix, input the preprocessed matrix into a trained system matrix super-resolution network model to obtain a set of super-resolution system matrix rows; and post-process the set of super-resolution system matrix rows to obtain a high-resolution system matrix.

[0122] The image reconstruction module 300 is configured to combine the high-resolution system matrix to solve the magnetic particle concentration distribution of the target object to be imaged, and reconstruct the MPI image of the target object to be imaged.

[0123] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the system described above can be found in the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0124] It should be noted that the MPI accelerated calibration system based on a system matrix super-resolution network provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiments can be merged into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the various modules or steps and are not considered as an improper limitation of the present invention.

[0125] An electronic device according to a third embodiment of the present invention includes: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to implement the above-described MPI accelerated calibration method based on a system matrix super-resolution network.

[0126] A computer-readable storage medium according to a fourth embodiment of the present invention stores computer instructions, which are executed by the computer to implement the above-described MPI accelerated calibration method based on a system matrix super-resolution network.

[0127] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the electronic devices and computer-readable storage media described above can be referred to the corresponding processes in the foregoing method examples, and will not be repeated here.

[0128] Those skilled in the art will recognize that the modules and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. The programs corresponding to the software modules and method steps can be placed in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. To clearly illustrate the interchangeability of electronic hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the invention.

[0129] The terms “first,” “second,” “third,” etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0130] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. An MPI accelerated calibration method based on a system matrix super-resolution network, used to recover a high-resolution system matrix and then perform MPI image reconstruction of the target object to be imaged, characterized in that, The method includes the following steps: Step S100: Place a unit MNP sample with a certain volume and concentration in the imaging field of the magnetic particle imaging device for sparse calibration scanning, and construct a system matrix as a low-resolution system matrix. Step S200: Preprocess the low-resolution system matrix, and input the preprocessed system matrix into the trained system matrix super-resolution network model to obtain the super-resolution system matrix row set; Postprocess the super-resolution system matrix row set to obtain the high-resolution system matrix. Step S300: Combine the high-resolution system matrix to solve the magnetic particle concentration distribution of the target object to be imaged, and reconstruct the MPI image of the target object to be imaged; The system matrix super-resolution network model includes a first convolution module, multiple residual groups, a scaling module, a bicubic interpolation residual connection module, and a second convolution module. The input to the first convolution module is the preprocessed low-resolution system matrix rows; Multiple residual groups are connected in series. The first convolution module is connected to the first residual group among the multiple residual groups, and the output of the first convolution module is added element by element to the output of the last residual group and used as the input of the scale-up module. The scaling module is connected to the second convolution module; The preprocessed low-resolution system matrix rows are processed by the bicubic interpolation residual connection module and then added element by element to the output of the second convolution module to obtain the super-resolution system matrix row set. The bicubic interpolation residual connection module is used to upsample the super-resolution scaling coefficients of the preprocessed low-resolution system matrix rows using the bicubic interpolation method. The residual group includes at least one wide activation residual block and one dynamic convolutional residual block; the wide activation residual block is connected to the dynamic convolutional residual block; The wide activation residual block is constructed based on a first convolutional layer, an activation function layer, a second convolutional layer, and a third convolutional layer connected in sequence; the first convolutional layer and the second convolutional layer have the same size; the second convolutional layer and the third convolutional layer have different sizes; The dynamic convolutional residual block includes two context-modulated dynamic convolutional layers; there is an activation function layer between the two dynamic convolutional layers; The dynamic convolutional layer is constructed based on a pooling operation unit, a context encoding operation unit, a channel interaction operation unit, a context decoding operation unit, and a modulation operation unit connected in sequence. The scaling module includes at least one upsampling module and a bicubic interpolation module, which are connected in sequence. The upsampling module is built based on a convolutional layer and a pixelshuffle module connected in sequence.

2. The MPI accelerated calibration method based on a system matrix super-resolution network according to claim 1, characterized in that, The preprocessing includes background reduction, signal-to-noise ratio threshold filtering, noise whitening, real-to-virtual part splitting, and matrix transformation operations; the postprocessing includes matrix dimension transformation and real-to-virtual part combination.

3. The MPI accelerated calibration method based on a system matrix super-resolution network according to claim 2, characterized in that, The pooling operation unit is constructed based on the pooling layer; the context encoding operation unit is constructed based on a linear layer, a batch normalization layer, and an activation function layer connected in sequence; the channel interaction operation unit is constructed based on a grouped linear layer, a batch normalization layer, and an activation function layer connected in sequence. The context decoding operation unit includes two linear layers. The input of one linear layer is the output of the channel interaction operation unit, and the input of the other linear layer is the output of the context coding operation unit. The outputs of the two linear layers are added element by element to obtain the output of the context decoding operation unit, thus obtaining the context modulation mask. The modulation operation unit is used to modulate the convolution kernel using the context modulation mask.

4. The MPI accelerated calibration method based on a system matrix super-resolution network according to claim 2, characterized in that, The training method for the system matrix super-resolution network model is as follows: Step A100: Obtain the high-resolution system matrix and downsample it to obtain the low-resolution system matrix; split the high-resolution system matrix and low-resolution system matrix pair into high-resolution system matrix rows and low-resolution system matrix row pairs, and then perform preprocessing. The high-resolution system matrix rows and low-resolution system matrix rows obtained after preprocessing are used as training samples, and the high-resolution system matrix rows corresponding to the training samples are used as ground value labels to construct a training set; the high-resolution system matrix includes high-resolution system matrices of the same or different devices, the same or different magnetic particles of the same device, and the same or different system parameters of the same device; the downsampling includes step downsampling and average downsampling. Step A200: Input the training samples in the training set into the pre-constructed system matrix super-resolution network model to obtain the set of rows of the super-resolution system matrix, and then obtain the high-resolution system matrix as the reconstruction matrix; Step A300: Based on the reconstruction matrix and its corresponding truth labels, calculate the loss value and update the model parameters of the system matrix super-resolution network model; Step A400, repeat steps A200-A300 until the trained system matrix super-resolution network model is obtained.

5. The MPI accelerated calibration method based on a system matrix super-resolution network according to claim 4, characterized in that, The loss function of the system matrix super-resolution network model during training is: in, Represents the loss function. This indicates the number of rows in the system matrix that participated in the training. Denotes the first-order norm, This is the set of rows of the preprocessed low-resolution system matrix. The labels are the row sets of the high-resolution system matrix. The parameters of the system matrix super-resolution network model are... The mapping function represents the system matrix super-resolution network model. This indicates the sequence number of the row in the system matrix.

6. An MPI accelerated calibration system based on a system matrix super-resolution network, used to construct a high-resolution system matrix and then perform MPI image reconstruction of the target object to be imaged, characterized in that, The system includes: The sparse system matrix acquisition module is configured to place a unit MNP sample with a certain volume and concentration in the imaging field of the magnetic particle imaging device for sparse calibration scanning and construct a system matrix as a low-resolution system matrix. The dense system matrix recovery module is configured to preprocess the low-resolution system matrix, input the preprocessed matrix into a trained system matrix super-resolution network model to obtain a set of super-resolution system matrix rows; and post-process the set of super-resolution system matrix rows to obtain a high-resolution system matrix. The image reconstruction module is configured to combine the high-resolution system matrix to solve for the magnetic particle concentration distribution of the target object to be imaged, and reconstruct the MPI image of the target object to be imaged; The system matrix super-resolution network model includes a first convolution module, multiple residual groups, a scaling module, a bicubic interpolation residual connection module, and a second convolution module. The input to the first convolution module is the preprocessed low-resolution system matrix rows; Multiple residual groups are connected in series. The first convolution module is connected to the first residual group among the multiple residual groups, and the output of the first convolution module is added element by element to the output of the last residual group and used as the input of the scale-up module. The scaling module is connected to the second convolution module; The preprocessed low-resolution system matrix rows are processed by the bicubic interpolation residual connection module and then added element by element to the output of the second convolution module to obtain the super-resolution system matrix row set. The bicubic interpolation residual connection module is used to upsample the super-resolution scaling coefficients of the preprocessed low-resolution system matrix rows using the bicubic interpolation method. The residual group includes at least one wide activation residual block and one dynamic convolutional residual block; the wide activation residual block is connected to the dynamic convolutional residual block; The wide activation residual block is constructed based on a first convolutional layer, an activation function layer, a second convolutional layer, and a third convolutional layer connected in sequence; the first convolutional layer and the second convolutional layer have the same size; the second convolutional layer and the third convolutional layer have different sizes; The dynamic convolutional residual block includes two context-modulated dynamic convolutional layers; there is an activation function layer between the two dynamic convolutional layers; The dynamic convolutional layer is constructed based on a pooling operation unit, a context encoding operation unit, a channel interaction operation unit, a context decoding operation unit, and a modulation operation unit connected in sequence. The scaling module includes at least one upsampling module and a bicubic interpolation module, which are connected in sequence. The upsampling module is built based on a convolutional layer and a pixelshuffle module connected in sequence.

7. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to at least one of the processors; The memory stores instructions that can be executed by the processor to implement the MPI accelerated calibration method based on a system matrix super-resolution network as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that are executed by the computer to implement the MPI accelerated calibration method based on a system matrix super-resolution network as described in any one of claims 1-5.

Images