Method and system for improving magnetic particle imaging resolution based on frequency domain information filtering

Abstract

The present invention proposes a method and system for improving the resolution of magnetic particle imaging based on frequency domain information filtering, which relates to the field of magnetic particle imaging technology. The method includes inputting the original magnetic particle imaging data into a Transformer model, selectively filtering the frequency domain information in the magnetic particle imaging data through a fusion frequency domain discrimination feedforward network module in the encoder, and obtaining the output of the encoder; inputting the output of the encoder into a bottleneck layer, aggregating global and local information through a multi-scale attention mechanism module, and then selectively filtering the frequency domain information in the magnetic particle imaging data again to obtain the output of the bottleneck layer; inputting the output of the bottleneck layer into a decoder with the same structure as the bottleneck layer to obtain deep features; inputting the deep features into a convolutional layer to obtain a residual image; and calculating the sum of the original magnetic particle imaging data and the residual image to obtain a reconstructed image. The present invention improves the overall imaging resolution of MPI images under low gradient field acquisition conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of magnetic particle imaging, and in particular relates to a method and system for improving the resolution of magnetic particle imaging based on frequency domain information filtering. Background Art

[0002] Magnetic nanoparticle imaging (MPI) is an emerging molecular imaging technology that can noninvasively track the in vivo distribution of superparamagnetic iron oxide nanoparticles (SPIONs) in real time. This technology has significant advantages, including high sensitivity, lack of background signal interference, rapid dynamic imaging, unrestricted tissue depth, and the absence of ionizing radiation. It has garnered widespread attention in the fields of medical imaging and clinical translation in recent years.

[0003] Spatial resolution is one of the key indicators for evaluating MPI imaging quality. Traditional methods for improving spatial resolution rely primarily on increasing the gradient field strength of the MPI system during scanning. However, while higher gradient field settings can improve image resolution, they significantly prolong scanning time, increase system energy consumption, and may even reduce imaging sensitivity or image contrast. Therefore, to improve image quality while maintaining imaging efficiency, data is typically acquired under low gradient field conditions, which also limits the spatial resolution of the reconstructed image.

[0004] In recent years, deep learning methods have demonstrated powerful feature extraction and nonlinear modeling capabilities in image reconstruction tasks. By learning spatial structure and frequency domain features from a large number of samples, deep neural networks can effectively suppress image artifacts and enhance detailed information. Encoder-decoder architectures, exemplified by UNet and its variants, have been widely used in medical image super-resolution reconstruction by leveraging multi-scale feature fusion and skip connections. They also hold broad application prospects in MPI low-resolution image restoration.

[0005] Existing technologies still face multiple challenges in improving MPI image resolution. First, as mentioned above, obtaining high-resolution images typically requires setting the MPI system to a higher gradient field strength to refine the magnetic field degrees of freedom and improve spatial positioning accuracy. However, this significantly prolongs scan time, increases device power consumption, and may affect dynamic imaging performance due to limited sampling rates, hindering real-time, efficient clinical applications.

[0006] Secondly, while deep learning methods have made some progress in the field of image restoration, they still have key shortcomings. Current mainstream models (such as UNet and its various improved structures) rely primarily on convolution operations to extract local spatial features. While they can effectively restore some image details, they generally lack explicit modeling and utilization of frequency components, making it impossible to fully recover high-frequency detail information, which can easily lead to blurred image edges or structural distortion. Although the UNet architecture introduces skip connections to alleviate information loss, its encoder-decoder structure is still limited in capturing long-range spatial dependencies in images, making it difficult to effectively integrate global contextual information. In MPI images in particular, the signal energy distribution is complex, low frequencies dominate, and high frequencies are weak. Existing networks often model frequency components through rough superposition, lacking dynamic decoupling and discrimination mechanisms. This limits the model's ability to adaptively process information in different frequency bands, ultimately affecting image restoration quality. Summary of the Invention

[0007] To overcome the shortcomings of the above-mentioned existing technologies, the present invention provides a method and system for improving the resolution of magnetic particle imaging based on frequency domain information filtering, and proposes an image reconstruction algorithm that integrates a frequency domain discriminant feedforward network and a multi-scale attention mechanism to improve the spatial resolution and structural restoration capability of MPI images, aiming to solve the problem of low resolution of reconstructed images obtained by existing magnetic particle imaging systems under low gradient field conditions.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0009] A first aspect of the present invention provides a method for improving the resolution of magnetic particle imaging based on frequency domain information filtering.

[0010] The method for improving the resolution of magnetic particle imaging based on frequency domain information filtering includes the following steps:

[0011] Acquire raw magnetic particle imaging data;

[0012] The raw magnetic particle imaging data is input into a Transformer model consisting of an encoder, a bottleneck layer, and a decoder. The frequency domain information in the magnetic particle imaging data is selectively filtered through the fused frequency domain discriminant feedforward network module in the encoder to obtain the encoder output.

[0013] The output of the encoder is input into the bottleneck layer, and the global and local information is aggregated through the multi-scale attention mechanism module. Then, the frequency domain information in the magnetic particle imaging data is selectively filtered using the fusion frequency domain discriminant feedforward network module to obtain the output of the bottleneck layer.

[0014] The output of the bottleneck layer is input into the decoder with the same structure as the bottleneck layer, the feature resolution is restored step by step, and the deep features are obtained by fusing it with the features corresponding to the encoder stage through skip connections;

[0015] Input the deep features into the convolution layer to obtain the residual image;

[0016] The sum of the original magnetic particle imaging data and the residual image is calculated to obtain the reconstructed image.

[0017] A second aspect of the present invention provides a magnetic particle imaging resolution enhancement system based on frequency domain information filtering.

[0018] The magnetic particle imaging resolution enhancement system based on frequency domain information filtering includes:

[0019] The data acquisition module is configured to: acquire raw magnetic particle imaging data;

[0020] The encoder data processing module is configured to: input the raw magnetic particle imaging data into a Transformer model including an encoder, a bottleneck layer, and a decoder, selectively filter the frequency domain information in the magnetic particle imaging data through the fused frequency domain discriminant feedforward network module in the encoder, and obtain the encoder output;

[0021] The bottleneck layer data processing module is configured to: input the encoder output to the bottleneck layer, aggregate global and local information through the multi-scale attention mechanism module, and then use the fused frequency domain discriminant feedforward network module to selectively filter the frequency domain information in the magnetic particle imaging data to obtain the output of the bottleneck layer;

[0022] The decoder data processing module is configured to: input the output of the bottleneck layer into the decoder with the same structure as the bottleneck layer, restore the feature resolution level by level, and fuse it with the features corresponding to the encoder stage through skip connections to obtain deep features;

[0023] The residual image generation module is configured to: input the deep features into the convolution layer to obtain the residual image;

[0024] The reconstructed image calculation module is configured to calculate the sum of the original magnetic particle imaging data and the residual image to obtain a reconstructed image.

[0025] One or more of the above technical solutions have the following beneficial effects:

[0026] This paper designs an image reconstruction algorithm that integrates a frequency-domain discriminant feedforward network with a multi-scale attention mechanism to improve the spatial resolution and structural restoration capabilities of MPI images. By introducing a frequency-domain discriminant feedforward network, this method adaptively analyzes and selects the frequency components of the input image, effectively identifying and retaining key frequency information, thereby enhancing the image's ability to express detailed information. Furthermore, combined with a multi-scale attention mechanism, it captures the dependencies between spatial features at different scales, enhancing the ability to model subtle structures and long-range dependencies in the image.

[0027] Through the deep fusion of frequency domain and spatial information, the present invention significantly improves the restoration quality of edge structures and low-contrast areas in the image, effectively suppresses reconstruction artifacts, and improves the overall imaging resolution of MPI images under low gradient field acquisition conditions.

[0028] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0030] Figure 1 This is a flow chart of the method of embodiment 1.

[0031] Figure 2 This is a flow chart of model training in Example 1.

[0032] Figure 3 This is a diagram of the model data processing architecture of Example 1.

[0033] Figure 4 This is a diagram showing the effect of improving the resolution of magnetic particle imaging in Example 1. DETAILED DESCRIPTION

[0034] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0035] It should be noted that the terms used herein are for describing particular embodiments only and are not intended to limit the exemplary embodiments according to the present invention.

[0036] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0037] Example 1

[0038] This embodiment discloses a method for improving the resolution of magnetic particle imaging based on frequency domain information filtering, which aims to solve the problem of low resolution of reconstructed images obtained by existing magnetic particle imaging systems under low gradient field conditions.

[0039] like Figure 1 As shown, the method for improving the resolution of magnetic particle imaging based on frequency domain information filtering includes the following steps:

[0040] Acquire raw magnetic particle imaging data;

[0041] The raw magnetic particle imaging data is input into a Transformer model consisting of an encoder, a bottleneck layer, and a decoder. The frequency domain information in the magnetic particle imaging data is selectively filtered through the fused frequency domain discriminant feedforward network module in the encoder to obtain the encoder output.

[0042] The output of the encoder is input into the bottleneck layer, and the global and local information is aggregated through the multi-scale attention mechanism module. Then, the frequency domain information in the magnetic particle imaging data is selectively filtered using the fusion frequency domain discriminant feedforward network module to obtain the output of the bottleneck layer.

[0043] The output of the bottleneck layer is input into the decoder with the same structure as the bottleneck layer, the feature resolution is restored step by step, and the deep features are obtained by fusing it with the features corresponding to the encoder stage through skip connections;

[0044] Input the deep features into the convolution layer to obtain the residual image;

[0045] The sum of the original magnetic particle imaging data and the residual image is calculated to obtain the reconstructed image.

[0046] To address the low resolution of reconstructed magnetic particle imaging images in low-gradient fields, this paper proposes a magnetic particle imaging resolution enhancement algorithm that integrates a frequency-domain discriminant feedforward network module with a multi-scale attention mechanism. The algorithm comprises a fused frequency-domain discriminant feedforward network module and a multi-scale attention mechanism module. The proposed algorithm effectively removes artifacts and improves the resolution of reconstructed images.

[0047] The embodiment of the present invention provides a Transformer architecture that integrates a frequency domain discriminant feedforward network and a multi-scale attention mechanism, which is an encoder-decoder structure composed of a feedforward network module and an attention mechanism module. Figure 3 shown.

[0048] Next, the Transformer model structure, training process, and data processing process designed in this embodiment will be explained with reference to the accompanying drawings.

[0049] When training the Transformer model, such as Figure 2 As shown in Figure 2, the overall training process includes:

[0050] (1) Obtain the signal data of magnetic particle imaging under high gradient field and low gradient field to be reconstructed through code simulation;

[0051] (2) Reconstruct high-resolution and low-resolution magnetic particle imaging images using the X-space algorithm;

[0052] (3) Construct a magnetic particle imaging image dataset and divide the dataset into a training set and a test set;

[0053] (4) Construct a Transformer model based on the fusion of frequency domain discriminant feedforward network and multi-scale attention mechanism, and train the model through the training set;

[0054] (5) The test set is input into the Transformer model based on the fusion of frequency domain discriminant feedforward network and multi-scale attention mechanism to obtain a high-resolution magnetic particle reconstructed image after removing artifacts and blurring from the low-resolution reconstructed image.

[0055] like Figure 3 As shown, the data processing flow will be described in detail in combination with the model structure:

[0056] Step S1:

[0057] In this embodiment, the signal data of magnetic particle imaging under high gradient field and low gradient field to be reconstructed are obtained by program code simulation. The high gradient field amplitude is set to 6 T / m and the low gradient field amplitude is set to 2 T / m, where T represents the magnetic field unit Tesla and m represents the length unit meter. The sampling frequency is set to The Lissajous scanning trajectory was set to scan the simulation image to obtain the reconstruction data, and the high-resolution and low-resolution magnetic particle imaging reconstruction images were obtained by X-space algorithm reconstruction.

[0058] Step S2:

[0059] Reconstructed image obtained under high gradient field As the target output, the reconstructed image obtained under the corresponding low gradient field As input. For a given input image, the convolution layer is first used to generate low-level feature representation , and then the low-level features pass The encoder and decoder networks are symmetrical and embedded into the deep features of the decoder output. In this embodiment The encoder and decoder network consists of The encoder basic unit, the decoder includes A basic unit of the decoder.

[0060] Each level of the encoder basic unit is composed of The network consists of an encoder basic block and a convolutional layer for downsampling. Each encoder basic block contains a fused frequency domain discriminant feedforward network module.

[0061] Each level of the decoder basic unit in the decoder consists of The network consists of a decoder basic block and a single convolutional layer. Each decoder basic block contains a fusion frequency domain discriminant feedforward network module and a multi-scale attention mechanism module.

[0062] A bottleneck layer is introduced before the decoder, and the bottleneck layer and the decoder use the same module. Figure 3 As shown in the figure, the bottleneck layer includes N3 bottleneck layer basic units and a convolutional layer. Each bottleneck layer basic unit includes a fused frequency domain discriminant feedforward network module and a multi-scale attention mechanism module.

[0063] It can be understood that the above N1, N2 and N3 all represent quantities.

[0064] Step S3:

[0065] The encoder, bottleneck layer and decoder all contain a fusion frequency domain discriminant feedforward network module. The fusion frequency domain discriminant feedforward network module introduces a learnable quantization matrix inspired by the JPEG image compression algorithm. , and learn the matrix by the inverse method of JPEG compression , to determine which frequency domain information should be retained.

[0066] Specifically, the learnable matrix It is not predefined, but consists of parameters that are updated and optimized during training along with the rest of the neural network. is defined as a four-dimensional tensor, ,in is the batch size of training images in the network, is the number of channels in the network, is the matrix size. This structure provides a unique The quantization matrix achieves highly specific and adaptive filtering. It is implemented by element-by-element multiplication with the frequency domain representation of the image features. Values ​​close to 1 in the product matrix are used to retain certain frequencies, and values ​​close to 0 are used to suppress certain frequencies.

[0067] Specifically, a value of 0.8 of the element value in the product matrix is ​​used to retain certain frequencies, and a value less than 0.2 is used to suppress certain frequencies.

[0068] Inspired by the JPEG image compression algorithm, the fusion frequency domain discriminant feedforward network module imitates the JPEG algorithm to decompose the feature map into non-overlapping Image blocks are converted to the frequency domain using fast Fourier transform and a learnable quantization matrix Element-wise multiplication is performed with the frequency domain image block to learn the frequency domain information and determine the frequency components suitable for retention. Finally, the filtered frequency domain image block is converted back to the spatial domain using the inverse fast Fourier transform.

[0069] After processing The image blocks are reassembled into a complete feature map, which is then activated through a GEGLU function and connected to the original input through a residual connection. Added together, this produces the final output of the module.

[0070] For a given input feature , first perform a preliminary mapping of the features,

[0071] ,

[0072] Among them, LN represents layer normalization; Conv represents convolution processing.

[0073] Then the frequency conversion process is performed:

[0074] ,

[0075] Utilizing a learnable quantization matrix Discriminatively filter frequency domain information:

[0076] ,

[0077] Finally, reconstruct the spatial features:

[0078] ,

[0079] in, represents the Fourier transform, represents the inverse Fourier transform, Represents the element-wise product operation, GEGLU represents the gated activation function; P represents the expansion processing operation in the JEPJ algorithm; Represents the input features after frequency domain processing; Represents the input features after frequency domain filtering; Represents the output features of the encoder; f represents the operation in the frequency domain.

[0080] Step S4:

[0081] Both the bottleneck layer and the decoder include a multi-scale attention mechanism module. At the bottleneck layer, the information first passes through the multi-scale attention mechanism module and then through the fused frequency-domain discriminant feedforward network module. The multi-scale attention mechanism module modifies the large-core attention using a multi-scale and gating mechanism to obtain attention maps at different scales, thereby aggregating global and local information. By combining the gating mechanism with spatial attention, unnecessary linear layers are removed, and information-rich spatial context is aggregated.

[0082] First, use the large kernel attention to the output features after the encoder Capturing long-range dependency information at different scales:

[0083] ,

[0084] Then through the gating mechanism Adjust the attention response to avoid artifacts:

[0085] ,

[0086] Finally, the output features of the multi-scale attention mechanism module are integrated:

[0087] ,

[0088] in represents a learnable scaling factor, represents point-wise convolution, represents the depth-wise dilated convolution, represents depthwise convolution; represents a large-core attention operation; MLKA(X) represents a large-core attention operation with a gating mechanism; represents the final output of the multi-scale attention mechanism module; X represents the feature output after the large-core attention operation.

[0089] Step S5:

[0090] The decoder restores feature resolution step by step. Each stage consists of a fused frequency-domain discriminative feedforward network module and a multi-scale attention mechanism module, and gradually restores the image spatial dimensions through upsampling. Features are fused with corresponding features from the encoder stage via skip connections.

[0091] Step S6:

[0092] Finally, the convolutional layer is The network generates a residual image, and the reconstructed image is obtained by summing the input image and the residual image. The model is trained using the Charbonnier loss function for optimization.

[0093]

[0094] in represents the reconstructed image obtained by the model, is the reconstructed image obtained under high gradient field; is an infinitesimal quantity, whose size is set to ; Represents the loss value.

[0095] The initial learning rate and batch size were set to 1e-5 and 1 respectively, and the model was trained for 100 epochs. It was trained with PyTorch on an Ubuntu system equipped with an NVIDIA A100 (40 GB) GPU and 128 GB RAM.

[0096] To address the issue of insufficient image resolution in magnetic particle imaging reconstruction under low-gradient field conditions, this paper proposes an image resolution enhancement algorithm that integrates a frequency-domain discriminant feedforward network with a multi-scale attention mechanism. This method incorporates a fused frequency-domain discriminant feedforward network module to adaptively identify and retain frequency components critical for image detail recovery. Simultaneously, it incorporates a multi-scale attention mechanism module to effectively extract and aggregate spatial features at different scales. This attention mechanism, combined with a gating strategy and a spatial attention map, not only improves the modeling of global and local dependencies but also removes redundant linear computation layers, enhancing the model's expressive efficiency and generalization capabilities.

[0097] Based on the above mechanism, the present invention designs a Transformer architecture that integrates frequency domain discrimination and multi-scale attention, which can effectively suppress image blur and artifact problems caused by low-gradient field sampling, and significantly improve the spatial resolution and visual quality of magnetic particle images.

[0098] Experimental part

[0099] The proposed algorithm has been experimentally validated on the MNIST handwriting simulation dataset. Code simulations were used to obtain MPI reconstructed images under high and low gradient field conditions. A total of 520 256×256 image data were generated, 500 of which were used for training and 20 for testing.

[0100] To verify the effectiveness of the proposed method, a comparative analysis was conducted using mainstream image reconstruction methods, including UNet, FAD former, FDS-MPI, TransUnet, and Uformer. Experimental results show that, under the same test conditions, the proposed algorithm achieves superior performance in metrics such as peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM), significantly outperforming existing methods in terms of reconstructed image resolution and detail preservation.

[0101] Figure 4This figure demonstrates the resolution improvement of magnetic particle imaging based on the fusion of frequency domain discrimination and multi-scale attention. The first column represents the input image, and the last column represents the ground-truth image. The columns between the first and last columns represent the reconstruction results using UNet, FAD former, FDS-MPI, TransUnet, Uformer, and the proposed method, respectively. The rows represent the reconstructed image and the error plot between the reconstructed image and the ground-truth image.

[0102] Table 1 shows the comparison of quantitative results (peak signal-to-noise ratio PSNR and structural similarity SSIM) on the simulated handwriting dataset.

[0103] Table 1

[0104]

[0105] Example 2

[0106] This embodiment discloses a magnetic particle imaging resolution enhancement system based on frequency domain information filtering.

[0107] The magnetic particle imaging resolution enhancement system based on frequency domain information filtering includes:

[0108] The data acquisition module is configured to: acquire raw magnetic particle imaging data;

[0109] The encoder data processing module is configured to: input the raw magnetic particle imaging data into a Transformer model including an encoder, a bottleneck layer, and a decoder, selectively filter the frequency domain information in the magnetic particle imaging data through the fused frequency domain discriminant feedforward network module in the encoder, and obtain the encoder output;

[0110] The bottleneck layer data processing module is configured to: input the encoder output to the bottleneck layer, aggregate global and local information through the multi-scale attention mechanism module, and then use the fused frequency domain discriminant feedforward network module to selectively filter the frequency domain information in the magnetic particle imaging data to obtain the output of the bottleneck layer;

[0111] The decoder data processing module is configured to: input the output of the bottleneck layer into the decoder with the same structure as the bottleneck layer, restore the feature resolution level by level, and fuse it with the features corresponding to the encoder stage through skip connections to obtain deep features;

[0112] The residual image generation module is configured to: input the deep features into the convolution layer to obtain the residual image;

[0113] The reconstructed image calculation module is configured to calculate the sum of the original magnetic particle imaging data and the residual image to obtain a reconstructed image.

[0114] Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using a general-purpose computer device. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0115] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A method for improving the resolution of magnetic particle imaging based on frequency domain information filtering, characterized in that: The following steps are involved: Acquire raw magnetic particle imaging data; The raw magnetic particle imaging data is input into a Transformer model consisting of an encoder, a bottleneck layer, and a decoder. The frequency domain information in the magnetic particle imaging data is selectively filtered through the fused frequency domain discriminant feedforward network module in the encoder to obtain the encoder output. The output of the encoder is input into the bottleneck layer, and the global and local information is aggregated through the multi-scale attention mechanism module. Then, the frequency domain information in the magnetic particle imaging data is selectively filtered using the fusion frequency domain discriminant feedforward network module to obtain the output of the bottleneck layer. The output of the bottleneck layer is input into the decoder with the same structure as the bottleneck layer, the feature resolution is restored step by step, and the deep features are obtained by fusing it with the features corresponding to the encoder stage through skip connections; Input the deep features into the convolution layer to obtain the residual image; Calculate the sum of the original magnetic particle imaging data and the residual image to obtain a reconstructed image; In the encoder, the frequency domain information in the magnetic particle imaging data is selectively filtered by the fusion frequency domain discrimination feedforward network module in the encoder to obtain the output of the encoder, which specifically includes: , ， ， in represents the Fourier transform, represents the inverse Fourier transform, represents the element-wise product operation, GEGLU represents the gated activation function; P represents the expansion processing operation in the JEPJ algorithm; X in represents the raw magnetic particle imaging data; is a learnable quantization matrix; Represents the input features after frequency domain processing; Represents the input features after frequency domain filtering; Represents the output features of the encoder; f represents the operation in the frequency domain.

2. The method for improving the resolution of magnetic particle imaging based on frequency domain information filtering according to claim 1, characterized in that: In the bottleneck layer, the multi-scale attention mechanism module modifies the large core attention by using multi-scale and gating mechanisms to obtain attention maps of different scales, thereby aggregating global and local information. Specifically: ， ， , in, represents a learnable scaling factor, represents point-wise convolution, represents the depth-wise dilated convolution, represents depthwise convolution; represents a large-core attention operation; MLKA(X) represents a large-core attention operation with a gating mechanism; represents the final output of the multi-scale attention mechanism module; X represents the feature output after the large-core attention operation; LN represents layer normalization; Conv represents convolution processing; represents the gating mechanism; Represents the output features of the encoder.

3. The method for improving the resolution of magnetic particle imaging based on frequency domain information filtering according to claim 1, characterized in that: It also includes training the Transformer model: Simulate and obtain signal data of magnetic particle imaging under high gradient field and low gradient field to be reconstructed; The obtained simulation image is scanned to obtain reconstruction data, and the magnetic particle imaging reconstruction images under high gradient field and low gradient field are reconstructed by X-space algorithm; The reconstructed image obtained under high gradient field As the target output, the reconstructed image obtained under the corresponding low gradient field As input to the Transformer model; The Charbonneau loss function is used for optimization training to obtain the trained Transformer model.

4. The method for improving the resolution of magnetic particle imaging based on frequency domain information filtering according to claim 3, characterized in that: The Chabenny loss function is specifically: ， in represents the reconstructed image obtained by the Transformer model, is an infinitesimal quantity; Indicates the loss value; is the reconstructed image obtained under high gradient field.

5. The method for improving the resolution of magnetic particle imaging based on frequency domain information filtering according to claim 1, characterized in that: The specific structures of the encoder, bottleneck layer and decoder include: The encoder includes encoder basic units, each of which includes An encoder basic block and a single convolutional layer, wherein the encoder basic block includes a fused frequency domain discriminant feedforward network module; The decoder comprises decoder basic units, each of which includes A decoder basic block and a single convolutional layer, wherein the decoder basic block includes a multi-scale attention mechanism module and a fused frequency domain discriminant feedforward network module; The bottleneck layer includes N3 bottleneck layer basic units and a convolutional layer, and each bottleneck layer basic unit includes a multi-scale attention mechanism module and a fusion frequency domain discrimination feedforward network module.

6. The method for improving the resolution of magnetic particle imaging based on frequency domain information filtering according to claim 1, wherein: Before the raw magnetic particle imaging data is fed into the Transformer model, it also includes: The raw magnetic particle imaging data is input into the convolutional layer to extract low-level features, and the low-level features are input into the Transformer model for processing.

7. The method for improving the resolution of magnetic particle imaging based on frequency domain information filtering according to claim 1, characterized in that: The learnable quantization matrix It is learned through the inverse method of JPEG compression, including: The learnable quantization matrix Defined as a four-dimensional tensor; The fused frequency domain discriminant feedforward network module imitates the JPEG algorithm to decompose the feature map into non-overlapping image blocks, and uses fast Fourier transform to convert the image blocks into the frequency domain to obtain the frequency domain representation of multiple image blocks; Learnable quantization matrix Perform element-by-element multiplication with the frequency domain representation of the image block to learn the frequency domain information, determine the frequency components suitable for retention, and obtain the frequency domain information of the filtered image block; Use inverse fast Fourier transform to convert the frequency domain information of the filtered image block back to the spatial domain to obtain the filtered image block; The filtered image blocks are reassembled into a complete feature map; The completed feature map is passed through a GEGLU activation function and connected to the original input through a residual connection. Added together, this produces the final output of the module.

8. The method for improving magnetic particle imaging resolution based on frequency domain information filtering according to claim 7, characterized in that: Learnable quantization matrix Perform element-by-element multiplication with the frequency domain representation of the image block to obtain a product matrix; The values ​​in the product matrix that are greater than the first set threshold are used to retain certain frequencies, and the values ​​that are less than the second set threshold are used to suppress certain frequencies, thereby determining the frequency components suitable for retention and obtaining the frequency domain information of the filtered image block.

9. A magnetic particle imaging resolution enhancement system based on frequency domain information filtering, characterized in that: include: The data acquisition module is configured to: acquire raw magnetic particle imaging data; The encoder data processing module is configured to: input the raw magnetic particle imaging data into a Transformer model including an encoder, a bottleneck layer, and a decoder, selectively filter the frequency domain information in the magnetic particle imaging data through the fused frequency domain discriminant feedforward network module in the encoder, and obtain the encoder output; The bottleneck layer data processing module is configured to: input the encoder output to the bottleneck layer, aggregate global and local information through the multi-scale attention mechanism module, and then use the fused frequency domain discriminant feedforward network module to selectively filter the frequency domain information in the magnetic particle imaging data to obtain the output of the bottleneck layer; The decoder data processing module is configured to: input the output of the bottleneck layer into the decoder with the same structure as the bottleneck layer, restore the feature resolution level by level, and fuse it with the features corresponding to the encoder stage through skip connections to obtain deep features; The residual image generation module is configured to: input the deep features into the convolution layer to obtain the residual image; The reconstructed image calculation module is configured to: calculate the sum of the original magnetic particle imaging data and the residual image to obtain a reconstructed image; In the encoder, the frequency domain information in the magnetic particle imaging data is selectively filtered by the fusion frequency domain discrimination feedforward network module in the encoder to obtain the output of the encoder, which specifically includes: , ， ， in represents the Fourier transform, represents the inverse Fourier transform, represents the element-wise product operation, GEGLU represents the gated activation function; P represents the expansion processing operation in the JEPJ algorithm; X in represents the raw magnetic particle imaging data; is a learnable quantization matrix; Represents the input features after frequency domain processing; Represents the input features after frequency domain filtering; Represents the output features of the encoder; f represents the operation in the frequency domain.

Images