A method and system for fast calibration of MPI system matrix based on multi-level wavelet transform and double-branch attention network

Abstract

The present application relates to the field of magnetic particle imaging, and particularly relates to a kind of fast calibration method and system of MPI system matrix based on multi-stage wavelet transform and double branch attention network, to solve the problem of time-consuming, complex of existing calibration-based system matrix acquisition method.The method of the present application comprises: obtaining a low-resolution system matrix according to a pre-set undersampling grid;The system matrix is encoded by row RGB to obtain an RGB image;The above-mentioned RGB image is input into the trained MWaveDAN network model to obtain a high-resolution RGB image;The high-resolution RGB image is decoded into complex form to obtain a complex-based system matrix.The present application effectively accelerates the high-resolution acquisition of system matrix in MPI, and provides important support for the future development of MPI technology.

Description

Technical Field

[0001] This invention relates to the field of magnetic particle imaging, specifically to a method and system for fast matrix calibration of MPI systems based on multi-level wavelet transform and bi-branch attention network. Background Technology

[0002] Magnetic particle imaging (MPI) is a radiation-free medical imaging technique that utilizes superparamagnetic iron oxide nanoparticles (SPIONs) as tracers to precisely quantify the distribution of tracers within living organisms with high sensitivity and spatiotemporal resolution. MPI technology has shown promise in biomedical fields such as cell tracing, vascular imaging, brain imaging, and magnetothermal therapy.

[0003] In MPI technology, image reconstruction based on the System Matrix (SM) is one of the core research areas. The System Matrix establishes a precise mapping between the magnetic particle signal at each location in the imaging space and the induced voltage received by the measurement device, serving as the cornerstone for achieving high-quality image reconstruction. However, traditional physical calibration methods for the System Matrix are extremely time-consuming. This process requires using a robotic arm to scan a tiny magnetic particle sample point by point across the entire Field of View (FOV) and acquire the signal at each point. For example, for a 30mm × 30mm × 30mm imaging field, a complete calibration process can take approximately 15 hours. When magnetic field conditions, equipment parameters, or magnetic particle characteristics change, recalibration is necessary, resulting in significant time and labor costs.

[0004] To shorten calibration time, the academic community has proposed various acceleration strategies, mainly divided into two categories: compressed sensing (CS) and deep learning. Compressed sensing methods utilize the sparsity of the system matrix in a certain transform domain (such as the discrete cosine transform domain) for reconstruction, but the reconstruction process is still relatively time-consuming, and the sparsity assumption needs further optimization. Therefore, this field urgently needs a system matrix calibration method that can efficiently fuse multi-scale features and enhance high-frequency details, significantly reducing calibration time while maintaining accuracy. In recent years, deep learning, especially based on the idea of ​​super-resolution technology, has been introduced into the fast calibration of system matrices. This method typically treats each row of the system matrix as an image, acquiring a low-resolution undersampled system matrix and then using a deep learning model to reconstruct it into a high-resolution matrix. However, existing super-resolution models face multiple challenges when processing MPI system matrices. First, compared to natural images, system matrix images are smaller and have higher noise levels, making feature extraction difficult and prone to information loss. Second, many models generate images with over-smoothing, leading to the loss of high-frequency details and blurred edges, affecting the final imaging accuracy. In addition, some complex models (such as Transformer) may have superior performance, but they are difficult to train and their generalization ability may be limited.

[0005] Therefore, there is an urgent need in this field for a new technical solution that can efficiently integrate multi-scale features and enhance the ability to recover high-frequency detail information, thereby significantly shortening calibration time and reducing application costs while ensuring reconstruction accuracy. Summary of the Invention

[0006] This invention aims to address the problems of long calibration time, low accuracy, and insufficient detail recovery capability in existing MPI system matrix calibration technologies. This invention provides a fast MPI system matrix calibration method and system based on multi-level wavelet transform and a dual-branch attention network. By introducing multi-level discrete wavelet transform (DWT) to decompose and process features at different scales, and combining specially designed low-frequency and high-frequency feature processing modules with an attention fusion mechanism, it achieves accurate and rapid reconstruction from a low-resolution undersampled system matrix to a high-resolution system matrix.

[0007] The technical solution adopted in this invention is: a fast matrix calibration method for MPI systems based on multi-level wavelet transform and dual-branch attention network, which includes the following steps:

[0008] Step S11: Obtain the low-resolution system matrix, based on the pre-defined undersampling grid, whose length and width are respectively... Obtain the undersampled system matrix M represents the number of rows in the system matrix. Indicates the number of columns in the system matrix;

[0009] Step S12: Image encoding, for the above undersampling system matrix The images are encoded in RGB format line by line to obtain M low-resolution RGB images. ;

[0010] Step S13: Image Super-Resolution Reconstruction: Construct a multi-branch attention network model based on multi-level wavelet (MWaveDAN), with the low-resolution RGB image mentioned above as the input. The model outputs a high-resolution image. Where r is the upsampling factor;

[0011] Step S14: Image decoding, converting the above high-resolution image... Decoded into complex form, thus obtaining the high-resolution system matrix in complex form;

[0012] This is a further optimization of the fast matrix calibration method for MPI systems based on multi-level wavelet transform and bi-branch attention network.

[0013] Preferably, in step S11, a low-resolution system matrix is ​​obtained. The method is as follows:

[0014] Prepare a standard sample containing magnetic particles, with dimensions of 1mm × 1mm × 1mm;

[0015] Set the length and width parameters h and w of the undersampled grid;

[0016] The robot or robotic arm automatically places a standard small sample containing magnetic particles at each undersampled grid location, and then collects the MPI voltage signal at each location. The response signal contains M frequency points.

[0017] The voltage signals acquired at each location are used as column vectors of the system matrix. After all grid locations have been acquired, the corresponding low-resolution system matrix is ​​obtained. N = h × w.

[0018] Preferably, in step S12, the system matrix is ​​encoded in RGB format, and the method is as follows:

[0019] For low-resolution system matrix Process the data line by line, mapping the complex data in each line to a 3-channel RGB image;

[0020] Each row corresponds to one RGB image. A total of M low-resolution RGB images were obtained.

[0021] Preferably, the data preprocessing module in step S13 processes the input image as follows:

[0022] Input image First, the image size is changed from 3×h×w to 3×rh×rw through Bicubic interpolation, where r is the upsampling factor.

[0023] The interpolated image is processed by two 3×3 convolutional layers, which increases the number of channels from 3 to 64, making it easier for subsequent feature extraction and processing.

[0024] Preferably, the feature processing flow of the MWaveDAN network in step S13 is as follows:

[0025] The preprocessed image is processed by three levels of discrete wavelet transform (DWT), and each level of DWT yields one low-frequency sub-image and three high-frequency sub-images;

[0026] The low-frequency sub-image enters the low-frequency feature module (LFB), which contains two 3×3 convolutional layers. The first convolutional layer captures coarse-grained features, and the second convolutional layer performs fine-grained feature fusion.

[0027] The high-frequency sub-image enters the high-frequency feature module (HFB), which includes three sub-networks in three directions. The horizontal sub-network consists of a 1×1 convolutional layer and group convolutional layers (3×3 convolutional kernels, with the number of groups equal to the number of input channels). The vertical sub-network consists of max pooling and a 1×1 convolutional layer. The diagonal sub-network integrates the structures of the horizontal and vertical sub-networks. All three sub-networks use GELU as the activation function.

[0028] During the decoding stage, the feature-enhanced sub-maps at each level are processed by inverse discrete wavelet transform (IWT) to obtain the reconstructed sub-maps;

[0029] The reconstructed subgraph and the low-frequency feature enhancement subgraph from the previous level are input into the attention feature fusion module (AFF). The inputs are first added together, and then channel attention weights of different scales are extracted through two branches. One branch uses global average pooling combined with two 3×3 convolutional layers, and the other branch directly uses two 3×3 convolutional layers. The global weights and local weights are added together and then adjusted by the sigmoid function to achieve feature fusion.

[0030] Finally, the enhanced image is processed by a 1×1 convolutional block to obtain a high-resolution reconstructed image. .

[0031] Preferably, in step S14, the high-resolution image is decoded into a complex form of a high-resolution system matrix, and the method is as follows:

[0032] High-resolution RGB images output from the network Reverse mapping is performed on each row of the corresponding low-resolution system matrix;

[0033] The 3-channel RGB image data is converted into complex form to obtain high-resolution complex data for each row;

[0034] After all rows are processed, they are combined to form a high-resolution system matrix in complex form, which improves the resolution by a factor of r in both the h and w directions.

[0035] The beneficial effects of this invention are:

[0036] (1) High efficiency and accuracy: By converting the time-consuming physical sampling process into an efficient image super-resolution reconstruction task, the calibration cycle of the system matrix is ​​greatly shortened. At the same time, by using multi-level wavelet transform to perform multi-scale feature decomposition on the image and combining a dual-branch attention mechanism to specifically enhance low-frequency structural information and high-frequency details, the problems of image blurring, artifacts and loss of details caused by traditional methods are effectively solved, and the calibration accuracy of the system matrix is ​​significantly improved.

[0037] (2) Structural Innovation: This invention embeds multi-level wavelet transform into a deep learning network and designs dedicated LFB and HFB modules for processing low-frequency and high-frequency wavelet coefficients, as well as an AFF module for cross-level information integration. This design can better preserve the global structure and local texture details of the image, and is especially suitable for image data with small size and high noise, such as MPI system matrices.

[0038] (3) Resolution enhancement: This method can enhance the spatial resolution by a factor of r by setting the upsampling factor r and the super-resolution processing mechanism of the network while completing the system matrix calibration, thus meeting the MPI technology’s requirement for high-resolution system matrices in one step and simplifying the processing flow. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating a fast matrix calibration method for MPI systems based on multi-level wavelet transform and dual-branch attention network according to the present invention.

[0040] Figure 2 This is a schematic diagram of the fast matrix calibration system framework of an MPI system based on multi-level wavelet transform and dual-branch attention network according to the present invention.

[0041] Figure 3 This is a schematic diagram of the overall architecture of the MWaveDAN network of this invention;

[0042] Figure 4 This is a schematic diagram of the low-frequency and high-frequency sub-images after a first-order discrete wavelet transform.

[0043] Figure 5This is a visual comparison diagram of the reconstruction results of the method of this invention and other reconstruction methods on a simulation dataset using a 4x downsampling method;

[0044] Figure 6 This diagram illustrates the comparison of image reconstruction results using the high-resolution system matrix reconstructed by the method of this invention and other methods on the OpenMPI public dataset. Detailed Implementation

[0045] 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.

[0046] 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.

[0047] A fast matrix calibration method for MPI systems based on multi-level wavelet transform and dual-branch attention network according to the first embodiment of the present invention, such as... Figure 1 As shown, steps S11-S14 are included, and each step is described in detail below:

[0048] Step S11, Obtaining the low-resolution system matrix: First, obtain a low-resolution undersampled system matrix. The process includes the following steps:

[0049] 1. Fabricate a standard physical sample containing magnetic nanoparticles (Perimag), with a sample size of 1mm × 1mm × 1mm;

[0050] 2. Set an undersampled two-dimensional grid within the imaging field of view (FOV), with size parameters h and w;

[0051] 3. Use a three-axis robot or robotic arm to automatically place standard samples sequentially on each undersampled grid point;

[0052] 4. At each location, acquire the voltage signal generated by the induction coil of the MPI device, which contains M frequency components;

[0053] 5. The voltage signal collected at each grid location, containing M frequency points, is stored as a column vector in the system matrix. After traversing all N = h × w grid points, the low-resolution system matrix is ​​constructed. .

[0054] Step S12, RGB encoding of the system matrix: For the system matrix in complex form Encode it and convert it into RGB format suitable for image processing.

[0055] Step S13, High-resolution reconstruction based on the MWaveDAN network: The encoded low-resolution RGB image is input into the MWaveDAN network model for super-resolution reconstruction. Figure 3 The detailed architecture of the MWaveDAN network is shown, including the following steps:

[0056] 1. Data preprocessing: Input low-resolution image First, a Bicubic interpolation is performed to enlarge the image size by a factor of r to 3×rh×rw. Then, two 3×3 convolutional layers are used to expand the number of image channels from 3 to 64, which facilitates deeper feature extraction in the subsequent process.

[0057] 2. Multi-level wavelet decomposition and feature extraction: Preprocessed feature map Entering a three-level wavelet decomposition-feature enhancement process:

[0058] (1) At each level (n=1, 2, 3), use DWT to convert the input low-frequency feature map Decomposed into a low-frequency subgraph and three high-frequency subgraphs , , Figure 4 A schematic diagram of the low-frequency and high-frequency sub-images after the first-order discrete wavelet transform is shown;

[0059] (2) Low-frequency subgraph The data is fed into the LFB, which contains two cascaded 3×3 convolutional layers. The first convolutional layer is used to capture coarse-grained features, while the second convolutional layer fuses fine-grained features to achieve hierarchical abstraction of features.

[0060] (3) The three high-frequency sub-images are fed into the HFB, which contains three parallel sub-networks to process detailed information in the horizontal, vertical and diagonal directions respectively: the horizontal sub-network consists of a 1×1 convolutional layer and group convolutional layers (the number of groups equals the number of input channels) to capture horizontal spatial patterns; the vertical sub-network consists of a max pooling layer and a 1×1 convolutional layer to preserve significant edge features in the vertical direction; the diagonal sub-network integrates the above two structures to capture more complex diagonal textures; all sub-networks in the HFB use GELU as the activation function.

[0061] 3. Fusion of decoding and attention features:

[0062] (1) In the decoding stage, starting from the deepest layer (n=3), IWT is used to reconstruct a single feature map from the low-frequency and high-frequency feature maps processed by LFB and HFB. ;

[0063] (2) The reconstructed feature map Enhanced low-frequency feature map compared to the previous level Both are sent to the AFF module;

[0064] (3) The AFF module first adds the two input feature maps, and then extracts channel attention weights through a dual-branch structure. One branch uses global average pooling to obtain global context attention; the other branch directly obtains local attention through convolutional layers. After fusing the two weights, they are adjusted by the Sigmoid function and applied to the original input to enhance the structured information.

[0065] (4) This fusion process is repeated step by step upwards until the final feature map is output. .

[0066] 4. Final Output: Finally, a 1×1 convolutional block is used to process the feature map. The number of channels is restored to 3 to obtain the final high-resolution reconstructed image. .

[0067] Step S14, Decoding the high-resolution system matrix: Decoding the high-resolution RGB image output by the network. Decode and restore the system matrix to its complex form. .

[0068] Figure 5 The reconstruction results of the proposed method are compared with those of other methods. The results show that the proposed method is optimal in simulation experiments.

[0069] Figure 6 The results demonstrate the comparative performance of MPI image reconstruction based on the high-resolution system matrix reconstructed by the method of this invention and other comparative methods. The results are based on the OpenMPI dataset and show that the reconstruction results of the method proposed in this invention are also the best.

[0070] A fast calibration system for an MPI system matrix based on multi-level wavelet transform and a dual-branch attention network, as described in the second embodiment of the present invention, is as follows: Figure 2 As shown, it includes: a system matrix acquisition module, an image encoding module, an image reconstruction module, and an image decoding module;

[0071] The system matrix acquisition module is configured to obtain the matrix based on the undersampled grid size, i.e., the length and width are respectively... MPI signals are collected at each grid location to construct the system matrix. ,in, Indicates the total number of grid points;

[0072] The image encoding module is configured to encode the system matrix. Perform RGB encoding to obtain M RGB images. ;

[0073] The image reconstruction module is configured to reconstruct low-resolution images. As network input, it is fed into a pre-built MWaveDAN neural network, and the network calculates to obtain the reconstructed high-resolution image. The original high-resolution image As ground truth labels, the loss is calculated and back propagated in each iteration, and a high-resolution reconstructed image is output after the iteration is completed. ;

[0074] The image decoding module is configured to decode a high-resolution RGB image into a complex number form to obtain a system matrix based on complex numbers.

Claims

1. A fast matrix calibration method for MPI systems based on multi-level wavelet transform and dual-branch attention network, characterized in that, The method includes the following steps: Step S11: Obtain the low-resolution system matrix, based on the pre-defined undersampling grid, whose length and width are respectively... Obtain the undersampled system matrix M represents the number of rows in the system matrix. Indicates the number of columns in the system matrix; Step S12: Image encoding, for the above undersampling system matrix Each line is encoded in RGB format to obtain M low-resolution RGB images. ; Step S13: Image Super-Resolution Reconstruction: Construct the MWaveDAN network training model, with the aforementioned low-resolution RGB image as the model input. The model outputs a high-resolution image. Where r is the upsampling factor; Step S14: Image decoding, converting the above high-resolution image... Decoded into complex form, thus obtaining the high-resolution system matrix in complex form; The MWaveDAN network training model includes a data preprocessing module, a low-frequency feature module, a high-frequency feature module, and an attention feature fusion module. The input image first undergoes image size and channel number transformation in the data preprocessing module. The processed image then undergoes three levels of DWT processing, each resulting in one low-frequency sub-image and three high-frequency sub-images. These four sub-images are then processed by the low-frequency and high-frequency feature modules for feature enhancement. In the decoding stage, each enhanced feature sub-image undergoes IWT processing to obtain a reconstructed sub-image. This reconstructed sub-image, along with the enhanced low-frequency feature sub-image from the previous level, is input to the attention feature fusion module. This module includes a dual-branch attention network for extracting channel attention weights. One branch uses global average pooling to obtain global contextual attention, while the other branch obtains local attention through convolutional layers. The fused weights are applied to the original input to further enhance the image's structural information. Finally, the enhanced image is processed... Convolutional block processing yields a high-resolution reconstructed image.

2. The fast matrix calibration method for MPI systems based on multi-level wavelet transform and dual-branch attention network according to claim 1, characterized in that, In step S11, the low-resolution system matrix is ​​obtained. The method is as follows: Prepare a standard sample containing magnetic particles, the size of which is ; Set the length and width parameters of the undersampled grid ; The robot or robotic arm automatically places a standard sample containing magnetic particles at each undersampled grid location, and then collects the MPI voltage signal at each location; The voltage signals acquired at each location are used as column vectors of the system matrix. After all grid locations have been acquired, the corresponding low-resolution system matrix is ​​obtained. .

3. The fast matrix calibration method for MPI systems based on multi-level wavelet transform and dual-branch attention network according to claim 1, characterized in that, The data preprocessing module transforms the input image in terms of image size and number of channels, and the method is as follows: Input image First, the image size is reduced using Bicubic interpolation. ; The interpolated image is processed through two convolutional layers to increase the number of channels to 64, with a kernel size of [missing value]. .

4. The fast matrix calibration method for MPI systems based on multi-level wavelet transform and dual-branch attention network according to claim 1, characterized in that, The low-frequency feature module includes two The convolutional layers consist of two stages: the first convolutional layer captures coarse-grained features, while the second convolutional layer fuses fine-grained features. This two-stage convolutional process creates a hierarchical progression of feature abstraction.

5. The fast matrix calibration method for MPI systems based on multi-level wavelet transform and dual-branch attention network according to claim 1, characterized in that, The high-frequency feature module includes sub-networks in three directions: horizontal, vertical, and diagonal. The horizontal sub-network consists of one... It consists of convolutional layers and group convolutional layers, with the group convolutional kernel size being [missing value]. The number of groups equals the number of input channels, and GELU is used as the activation function; the vertical subnetwork consists of max pooling and one... The system consists of convolutional layers, using GELU as the activation function; the diagonal subnetworks integrate the horizontal and vertical subnetwork structures, also using GELU as the activation function.

6. The fast matrix calibration method for MPI systems based on multi-level wavelet transform and dual-branch attention network according to claim 1, characterized in that, The MWaveDAN network training method is as follows: A11, Obtain the low-resolution undersampling system matrix. ; A12, to RGB encoding is performed line by line to obtain M RGB images. ; A13, low-resolution image The pre-built super-resolution reconstruction model MWaveDAN is input for image super-resolution processing to obtain the reconstructed high-resolution image. ; A14, based on the reconstructed image High-resolution images of true values The mean squared error loss function is calculated to obtain the loss value, and then backpropagation is performed to update the model parameters of MWaveDAN; A15, loop through A13-A14 until the trained MWaveDAN model is obtained.

7. A fast matrix calibration system for MPI systems based on multi-level wavelet transform and a dual-branch attention network, characterized in that, The system includes the following modules: The system matrix acquisition module is configured to acquire the matrix based on the undersampled grid size, i.e., the length and width are respectively... MPI signals are collected at each grid location to construct the system matrix. ,in, Indicates the total number of grid points; The image encoding module is configured to encode the system matrix. Perform RGB encoding to obtain M RGB images. ; The image reconstruction module is configured to reconstruct low-resolution images. As network input, it is fed into a pre-built MWaveDAN neural network, and the network calculates to obtain the reconstructed high-resolution image. ; The MWaveDAN neural network comprises a data preprocessing module, a low-frequency feature module, a high-frequency feature module, and an attention feature fusion module. The input image first undergoes image size and channel number transformation in the data preprocessing module. The processed image then undergoes three levels of Direct WT (DWT), each resulting in one low-frequency sub-image and three high-frequency sub-images. These four sub-images are then processed by the low-frequency and high-frequency feature modules for feature enhancement. In the decoding stage, each enhanced feature sub-image undergoes Indirect WT (IWT) to obtain a reconstructed sub-image. This reconstructed sub-image, along with the enhanced low-frequency feature sub-image from the previous level, is input to the attention feature fusion module. This module includes a dual-branch attention network for extracting channel attention weights. One branch uses global average pooling to obtain global contextual attention, while the other branch obtains local attention through convolutional layers. The fused weights are applied to the original input to further enhance the image's structural information. Finally, the enhanced image is processed by a 1×1 convolutional block to obtain a high-resolution reconstructed image. The image decoding module is configured to decode a high-resolution RGB image into a complex number form to obtain a system matrix based on complex numbers.

Images