MPI system matrix rapid calibration method and system based on multistage wavelet transform and double-branch attention network

Abstract

The invention relates to the field of magnetic particle imaging, in particular to an MPI system matrix rapid calibration method and system based on multistage wavelet transform and a double-branch attention network, and aims to solve the problem that an existing calibration-based system matrix acquisition method is time-consuming and complicated. The method comprises the following steps: acquiring a low-resolution system matrix according to a preset under-sampling grid; carrying out RGB coding on the system matrix according to rows to obtain an RGB image; the RGB image is input into a trained MWaveDAN network model, and a high-resolution RGB image is obtained; a high-resolution RGB image is decoded into a complex number form, and a complex number-based system matrix is obtained. According to the method, high-resolution acquisition of the system matrix in the MPI is effectively accelerated, and important support is provided for future development of the MPI technology.

Description

TECHNICAL FIELD

[0001] The application relates to the field of magnetic particle imaging, and in particular to an MPI system matrix fast calibration method and system based on a multi-level wavelet transform and a double-branch attention network. BACKGROUND

[0002] Magnetic particle imaging (MPI) is a non-radiation medical imaging technology that uses superparamagnetic iron oxide nanoparticles (SPIONs) as tracers to accurately quantify the distribution of tracers in the body with high sensitivity and high spatial-temporal resolution. MPI technology has application potential in biomedical fields such as cell tracking, blood vessel imaging, brain imaging and magnetic hyperthermia.

[0003] In MPI technology, image reconstruction based on the system matrix (SM) is one of the most core research links. The system matrix establishes the precise mapping relationship between the magnetic particle signal at each position in the imaging space and the induced voltage received by the measurement device, and is the cornerstone of high-quality image reconstruction. However, the traditional physical calibration method of the system matrix is extremely time-consuming. This process requires the use of a mechanical arm to scan a tiny magnetic particle sample point by point through the entire imaging field of view (FOV) and collect the signal at each point. For example, for an imaging field of view of 30mm x 30mm x 30mm, the complete calibration process may take about 15 hours. When the magnetic field conditions, device parameters or magnetic particle characteristics change, calibration must be performed again, which brings a large amount of time and labor costs.

[0004] To shorten the calibration time, the academia has proposed various acceleration strategies, mainly divided into two categories: compressed sensing (CS) and deep learning. The compressed sensing method utilizes the sparse characteristics 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, there is an urgent need for a system matrix calibration method that can efficiently fuse multi-scale features and enhance high-frequency details, while significantly shortening the calibration time while ensuring accuracy. In recent years, deep learning, especially the idea based on super-resolution (Super-Resolution) technology, has been introduced into the rapid calibration of system matrices. This method usually treats each row of the system matrix as an image, acquires a low-resolution undersampled system matrix, and then uses 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 with natural images, system matrix images are smaller in size and have higher noise levels, making feature extraction difficult and prone to information loss. Second, many models generate images that are oversmoothed, resulting in the loss of high-frequency details and blurred edges, which affects the final imaging accuracy. In addition, some complex models (such as Transformers) have superior performance but are difficult to train and may have limited generalization capabilities.

[0005] Therefore, there is an urgent need in the art for a new technical solution that can efficiently fuse multi-scale features and enhance the recovery of high-frequency detail information, thereby significantly shortening the calibration time while ensuring reconstruction accuracy and reducing application costs. SUMMARY

[0006] The present application aims to solve the problems of long calibration time, low accuracy, and insufficient detail recovery in the prior art. The present application provides a fast calibration method and system for MPI system matrices based on multi-level wavelet transform and dual-branch attention network. By introducing multi-level discrete wavelet transform (DWT) to decompose and process features of different scales, and combining specially designed low-frequency and high-frequency feature processing modules and attention fusion mechanisms, accurate and rapid reconstruction from a low-resolution undersampled system matrix to a high-resolution system matrix is achieved.

[0007] The technical solution adopted by the present application is: a fast calibration method for MPI system matrices based on multi-level wavelet transform and dual-branch attention network, which comprises the following steps:

[0008] Step S11: Obtain a low-resolution system matrix. According to a pre-set undersampling grid, the length and width are , and obtain an undersampled system matrix , where M represents the number of rows of the system matrix, The number of columns of the system matrix is represented by M;

[0009] Step S12: image encoding, the undersampled system matrix is encoded in RGB format row by row, and M low-resolution RGB images are obtained.

[0010] Step S13: image super-resolution reconstruction: a multi-level wavelet-based dual-branch attention (MWaveDAN) network model is constructed, the input of the model is the low-resolution RGB image , and the output of the model is a high-resolution image , where r is the up-sampling multiple.

[0011] Step S14: image decoding, the high-resolution image is decoded into a complex form, and a high-resolution system matrix in complex form is obtained.

[0012] As a further optimization of the MPI system matrix fast calibration method based on multi-level wavelet transform and dual-branch attention network.

[0013] Preferably, the low-resolution system matrix is obtained in step S11 by the following method:

[0014] A standard sample containing magnetic particles is prepared, with a size of 1mm x 1mm x 1mm.

[0015] The length and width parameters h and w of the undersampling grid are set.

[0016] A robot or a mechanical arm automatically places the standard sample containing magnetic particles at each undersampling grid position, and then collects the MPI voltage signals at each position. The response signal contains M frequency points.

[0017] The voltage signals collected at each position are used as column vectors of the system matrix. After all grid positions are collected, the corresponding low-resolution system matrix is obtained, where N = h x w.

[0018] Preferably, the system matrix is encoded in RGB format in step S12 by the following method:

[0019] For the low-resolution system matrix , the complex data in each row is mapped to a 3-channel RGB image row by row.

[0020] Each row corresponds to generate an RGB image , and M low-resolution RGB images are obtained.

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

[0022] The input image First, the Bicubic interpolation processing is performed to change the image size from 3xh xw to 3xrh xrw, where r is the up-sampling multiple;

[0023] The interpolated image is processed by two 3x3 convolution layers to expand the channel number from 3 to 64, facilitating 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 after each level of DWT, a low-frequency subgraph and three high-frequency subgraphs are obtained;

[0026] The low-frequency subgraph enters the low-frequency feature module (LFB), which includes two 3x3 convolution layers. The first convolution layer captures coarse-grained features, and the second convolution layer performs fine-grained feature fusion.

[0027] The high-frequency subgraph enters the high-frequency feature module (HFB), which includes three directional subnetworks. The horizontal directional subnetwork is composed of a 1x1 convolution layer and a group convolution layer (3x3 convolution kernel, the number of groups is equal to the number of input channels). The vertical directional subnetwork is composed of a max pooling layer and a 1x1 convolution layer. The diagonal directional subnetwork integrates the horizontal and vertical directional subnetwork structures. The three subnetworks use GELU as the activation function.

[0028] In the decoding stage, the feature enhancement subgraph of each level is processed by inverse discrete wavelet transform (IWT) to obtain a reconstructed subgraph.

[0029] The reconstructed subgraph and the low-frequency feature enhancement subgraph of the previous level are jointly input into the attention feature fusion module (AFF). First, the input is added, and then two branches are used to extract channel attention weights of different scales. One branch uses global average pooling combined with two layers of 3x3 convolution layers, and the other branch directly uses two layers of 3x3 convolution layers. The global weight and the local weight are added and adjusted by a sigmoid function to realize feature fusion.

[0030] Finally, the enhanced image is processed by a 1x1 convolution block to obtain a high-resolution reconstructed image .

[0031] Preferably, in step S14, the high-resolution image is decoded into a high-resolution system matrix in complex form by the following method:

[0032] The high-resolution RGB image output by the network , each row of the low-resolution system matrix is inversely mapped according to the rows;

[0033] The 3-channel RGB image data is converted into a complex form, obtaining high-resolution complex data corresponding to each row;

[0034] After all the rows are processed, the high-resolution system matrix in complex form is combined, and the resolution is improved by r times in the h and w directions.

[0035] The beneficial effects of the present application are:

[0036] (1) High efficiency and precision: By converting the time-consuming physical sampling process into an efficient image super-resolution reconstruction task, the calibration period of the system matrix is greatly shortened. At the same time, with the help of multi-level wavelet transform for multi-scale feature decomposition of the image, and combined with the double-branch attention mechanism to enhance the low-frequency structure information and high-frequency details, the image blurring, artifacts and detail loss problems caused by traditional methods are effectively solved, and the calibration accuracy of the system matrix is significantly improved.

[0037] (2) Structural innovation: The present application embeds multi-level wavelet transform into the deep learning network, and designs LFB and HFB modules dedicated to processing low-frequency and high-frequency wavelet coefficients, as well as 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 MPI system matrix image data with small size and high noise.

[0038] (3) Resolution improvement: By setting the up-sampling multiple r and the super-resolution processing mechanism of the network, the spatial resolution of the system matrix can be improved by r times while completing the calibration, which meets the demand of MPI technology for high-resolution system matrix in one step and simplifies the processing flow. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a flowchart of the present application, which is a fast calibration method for MPI system matrix based on multi-level wavelet transform and double-branch attention network;

[0040] Figure 2 is a schematic diagram of the system framework of the present application, which is a fast calibration system for MPI system matrix based on multi-level wavelet transform and double-branch attention network;

[0041] Figure 3 is a schematic diagram of the overall architecture of the MWaveDAN network of the present application;

[0042] Figure 4 is a schematic diagram of the low-frequency and high-frequency sub-images after one-level discrete wavelet transform;

[0043] Figure 5is a visual comparison diagram of the 4-fold down-sampling reconstruction of the simulation data set by the method of the application and other reconstruction methods;

[0044] Figure 6 is a visual comparison diagram of the image reconstruction of the OpenMPI public data set by the high-resolution system matrix reconstructed by the method of the application and other methods. DETAILED DESCRIPTION

[0045] The application will be further described below in detail with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only intended to explain the related application, and not to limit the application. In addition, it should be noted that only the parts related to the application are shown in the drawings for ease of description.

[0046] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the drawings and embodiments.

[0047] A MPI system matrix fast calibration method based on multi-level wavelet transform and double-branch attention network according to the first embodiment of the application, as shown in Figure 1 The steps S11-S14 are described in detail as follows:

[0048] Step S11, obtaining a low-resolution system matrix: first, obtaining a low-resolution under-sampling system matrix The process includes the following steps:

[0049] 1. A standard physical sample containing magnetic nanoparticles (Perimag) is made, and the sample size is 1mmx1mmx1mm;

[0050] 2. An under-sampling two-dimensional grid is set in the imaging field of view (FOV), and the size parameters are h and w;

[0051] 3. A three-axis robot or mechanical arm is used to automatically place the standard sample on each under-sampling grid point in turn;

[0052] 4. At each position, the voltage signal generated by the MPI device induction coil is collected, which contains M frequency components;

[0053] 5. The voltage signal containing M frequency points collected at each grid position is stored as a column vector in the system matrix, and after traversing all N=hxw grid points, the low-resolution system matrix is constructed.

[0054] Step S12, RGB encoding of the system matrix: the complex form of the system matrix Encoding, converting it into an RGB format suitable for image processing.

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

[0056] 1. Data preprocessing: input low-resolution image First, it is enlarged by r times to 3xrhxrw through Bicubic interpolation. Then, through two 3x3 convolution layers, the number of image channels is expanded from 3 to 64 to facilitate subsequent deeper feature extraction.

[0057] 2. Multilevel wavelet decomposition and feature extraction: preprocessed feature map Enter a three-level wavelet decomposition-feature enhancement process:

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

[0059] (2) The low-frequency subgraph is sent to the LFB, which contains two 3x3 convolution layers in series. The first convolution layer is used to capture coarse-grained features, and the second convolution layer is used to fuse fine-grained features to achieve hierarchical abstraction of features.

[0060] (3) The three high-frequency subgraphs are sent to the HFB, which contains three parallel subnetworks that process horizontal, vertical, and diagonal direction details respectively: the horizontal direction subnetwork is composed of a 1x1 convolution layer and a group convolution layer (the number of groups is equal to the number of input channels), which is used to capture horizontal spatial patterns; the vertical direction subnetwork: composed of a max pooling layer and a 1x1 convolution layer, used to retain significant edge features in the vertical direction; the diagonal direction subnetwork: integrates the above two structures to capture more complex diagonal line textures; all subnetworks in HFB use GELU as the activation function.

[0061] 3. Decoding and attention feature fusion:

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

[0063] (2) The reconstructed feature map is sent to the AFF module together with the enhanced low-frequency feature map of the previous level;

[0064] (3) The AFF module first adds the two input feature maps, and then extracts channel attention weights through a double-branch structure, one branch uses global average pooling to obtain global context attention; the other branch directly obtains local attention through a convolution layer. After fusing the two weights, the Sigmoid function is adjusted and acts on the original input to realize the enhancement of structured information;

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

[0066] 4. Final output: Finally, a 1x1 convolution block is used to process the feature map , restore its channel number to 3, and obtain the final high-resolution reconstructed image .

[0067] Step S14, decoding of the high-resolution system matrix: the high-resolution RGB image output by the network is decoded and restored to a complex-valued system matrix .

[0068] Figure 5 The comparative reconstruction results of the method proposed in the application and other methods are shown. The results show that in the simulation experiment, the method proposed in the application is the best.

[0069] Figure 6 The comparative effects of MPI image reconstruction based on the high-resolution system matrix reconstructed by the method of the application and other comparative methods are shown, based on the OpenMPI dataset, and the results show that the reconstruction result of the method proposed in the application is also the best.

[0070] A kind of MPI system matrix fast calibration system based on multi-level wavelet transform and double-branch attention network according to the second embodiment of the application, as Figure 2 shown, comprising: 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 collect MPI signals at each grid position according to the undersampling grid size, i.e. the length and width are , to construct a system matrix , wherein represents the total number of grid points; ​

[0072] an image encoding module configured to encode the system matrix into RGB, obtaining M RGB images ;

[0073] an image reconstruction module configured to reconstruct the low-resolution image as network input, input into a pre-constructed MWaveDAN neural network, and obtain a reconstructed high-resolution image through network calculation , the original high-resolution image as a true value label, calculate the loss and return each time, and output the high-resolution reconstructed image after iteration ;

[0074] an image decoding module configured to decode the high-resolution RGB image into a complex form, obtaining a complex-based system matrix.

Claims

1. A method for fast calibration of MPI system matrix based on multi-level wavelet transform and dual-branch attention network, characterized in that, The method comprises the following steps: Step S11: obtaining a low-resolution system matrix, according to a pre-set under-sampling grid, the length and width of which are , obtaining an under-sampling system matrix , M represents the number of rows of the system matrix, represents the number of columns of the system matrix; Step S12: image encoding, encoding the above-mentioned under-sampling system matrix , in RGB format row by row, and then obtaining M low-resolution RGB images ; Step S13: image super-resolution reconstruction: a MWaveDAN network training model is constructed, the model input is the low-resolution RGB image described above , and the model output is a high-resolution image , wherein r is the up-sampling multiple; Step S14: image decoding, decoding the above high-resolution image into a complex form, thereby obtaining a high-resolution system matrix in a complex form; The MWaveDAN network training model comprises a data preprocessing module, a low-frequency feature module, a high-frequency feature module, and an attention feature fusion module. The model input image is first subjected to image size and channel number transformation by the data preprocessing module. The processed image is subjected to three-level DWT processing. After each level of DWT, a low-frequency subgraph and three high-frequency subgraphs are obtained. The four subgraphs are then respectively subjected to feature enhancement in the low-frequency feature module and the high-frequency feature module. In the decoding stage, the feature enhancement subgraph of each level is subjected to IWT processing to obtain a reconstructed subgraph. The reconstructed subgraph and the low-frequency feature enhancement subgraph of the previous level are jointly input into the attention feature fusion module to further enhance the structural information of the image. Finally, the enhanced image is subjected to convolution block processing to obtain a high-resolution reconstructed image.

2. The MPI system matrix fast calibration method based on multi-level wavelet transform and double-branch attention network according to claim 1, characterized in that, The low-resolution system matrix is obtained in step S11 The method is: Standard samples containing magnetic particles were made with dimensions of ; Setting long, width parameters of the undersampling grid ; The robot or mechanical arm automatically places a standard sample containing magnetic particles at each under-sampling grid position, and then collects the MPI voltage signals of each position; The voltage signals collected at each position are respectively taken as column vectors of the system matrix, and after all the grid positions are completed, the corresponding low-resolution system matrix is obtained .

3. The method of claim 1, wherein, The data preprocessing module performs image size and channel number conversion on the input image, and the method is as follows: Input image First, the image size is changed to ; The interpolated image is processed by two convolutional layers, which expand the number of channels to 64, with a kernel size of .

4. The method of claim 1, wherein, The low-frequency feature module includes two convolution layers, the first convolution layer captures coarse-grained features, and the second convolution layer performs fine-grained feature fusion, and the two-stage convolution process forms a hierarchical progression of feature abstraction.

5. The method of claim 1, wherein, The high-frequency feature module includes three directional sub-networks, namely a horizontal direction, a vertical direction and a diagonal direction; the horizontal direction sub-network is composed of one convolution layer and a group convolution layer, the group convolution kernel size is , the number of groups is equal to the number of input channels, and GELU is used as an activation function; the vertical direction sub-network is composed of a maximum pooling and one convolution layer, and GELU is used as an activation function; and the diagonal direction sub-network integrates the horizontal and vertical direction sub-network structures, and GELU is also used as an activation function.

6. The method of claim 1, wherein, The attention feature fusion module is configured to receive the reconstructed feature map of the current decoding layer and the low-frequency feature map of the previous level as input, and calculate the fusion weight through a double-branch attention mechanism containing global and local branches to enhance the structural information of the image.

7. The method of claim 1, wherein, The method for training the MWaveDAN network is as follows: A11, obtaining a low resolution under-sampled system matrix ; A12, to RGB-encode the rows to obtain M RGB images ; A13, the low-resolution image inputting the pre-constructed super-resolution reconstruction model MWaveDAN to perform image super-resolution processing, to obtain a reconstructed high-resolution image ; A14, based on the reconstructed image and the ground truth high resolution image The loss value is calculated by computing the mean square error loss function, and then the model parameters of the MWaveDAN are updated by back propagation. A15, cycle A13-A14, until a trained MWaveDAN model is obtained.

8. A fast calibration system for MPI system matrix based on multi-level wavelet transform and dual-branch attention network, characterized in that, The system comprises the following modules: The system matrix obtaining module is configured to obtain the system matrix according to the undersampling grid size, that is, the length and the width are respectively Collecting the MPI signal at each grid position to construct the system matrix Wherein, Indicates the total number of grid points; an image encoding module configured to encode the system matrix perform RGB encoding to obtain M RGB images ; An image reconstruction module is configured to reconstruct a low-resolution image As a network input, input to the pre-constructed MWaveDAN neural network, and through network calculation, a reconstructed high-resolution image is obtained , the original high-resolution image As a true value label, the loss is calculated and returned each time the iteration is calculated, and the high-resolution reconstructed image is output after the iteration is completed ; An image decoding module configured to decode the high-resolution RGB image into a complex form to obtain a complex-based system matrix.

Images