Magnetic particle imaging image processing method and device

Abstract

The present invention provides a magnetic particle imaging image processing method and device, which can be applied to the field of image processing technology. The method includes: using a Gaussian noise network to perform noise enhancement processing on a magnetic particle imaging noise image to obtain a target noise image; using N cascaded downsampling attention layers to perform multi-level feature extraction on the target noise image to obtain N image features, wherein the downsampling attention layer is constructed based on L cascaded downsampling attention excitation sublayers; using a grayscale adaptive learning network layer to perform feature fusion on the Nth image feature to obtain an image grayscale enhancement feature, wherein the grayscale adaptive learning network layer is obtained based on the adjustment of adaptive grayscale enhancement learning parameters; using N cascaded upsampling attention layers to process N image features and image grayscale enhancement features to obtain a magnetic particle imaging denoised image, wherein the upsampling attention layer is constructed based on L cascaded upsampling attention excitation sublayers.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a magnetic particle imaging image processing method and device. Background Art

[0002] Magnetic particle imaging is a tracer-based tomography technique that can detect the concentration distribution of superparamagnetic iron oxide nanoparticles (SPIONs). When there are multiple samples with large concentration differences in the field of view of magnetic particle imaging, the signal of the high-concentration sample will interfere with the signal of the low-concentration sample, causing the low-concentration sample signal to be obscured, resulting in loss or damage of its sample information. Current reconstruction algorithms are difficult to accurately restore tissue information of low-concentration samples.

[0003] In the field of medical image analysis, existing reconstruction algorithms based on hardware optimization increase the cost and complexity of magnetic particle imaging systems. At the same time, they fail to consider the impact of different dynamic ranges, have low robustness, and low accuracy in recovering low-concentration sample tissue information. Summary of the invention

[0004] In view of the above problems, the present invention provides a magnetic particle imaging image processing method and device.

[0005] According to a first aspect of the present invention, a magnetic particle imaging image processing method is provided, comprising: performing noise enhancement processing on a magnetic particle imaging noise image using a Gaussian noise network to obtain a target noise image, wherein the magnetic particle imaging noise image is obtained by scanning biological tissues by a magnetic particle imaging system; performing multi-level feature extraction on the target noise image using N cascaded downsampling attention layers to obtain N image features, wherein the downsampling attention layer is constructed based on L cascaded downsampling attention excitation sublayers, and N and L are both positive integers; performing feature fusion on the Nth image feature using a grayscale adaptive learning network layer to obtain an image grayscale enhancement feature, wherein the grayscale adaptive learning network layer is obtained based on the adjustment of adaptive grayscale enhancement learning parameters; processing N image features and image grayscale enhancement features using N cascaded upsampling attention layers to obtain a magnetic particle imaging denoised image, wherein the upsampling attention layer is constructed based on L cascaded upsampling attention excitation sublayers.

[0006] Optionally, the grayscale adaptive learning network layer includes a first grayscale feature extraction sublayer, M cascaded intermediate attention excitation sublayers, and a second grayscale feature extraction sublayer, where M < L and M is a positive integer; using the grayscale adaptive learning network layer to perform feature fusion on the Nth image feature, the obtained image grayscale enhancement feature includes: using the first grayscale feature extraction sublayer to process the Nth image feature to obtain an initial image grayscale feature; using M cascaded intermediate attention excitation sublayers to process the initial image grayscale feature to obtain an image excitation feature, where the image excitation feature represents the fused deep image grayscale information and image position information; using the second grayscale feature extraction sublayer to process the image excitation feature to obtain an image grayscale enhancement feature.

[0007] Optionally, the intermediate attention excitation sublayer includes an intermediate residual excitation module and an intermediate self-attention module; where using M cascaded intermediate attention excitation sublayers to process the initial image grayscale feature to obtain an image excitation feature includes: using the intermediate residual excitation module in the first intermediate attention excitation sublayer to process the initial image grayscale feature to obtain a first residual feature; using the intermediate self-attention module in the first intermediate attention excitation sublayer to process the first residual feature to obtain a first intermediate image excitation feature; using M - 1 cascaded intermediate attention excitation sublayers to process the first intermediate image excitation feature and output an image excitation feature.

[0008] Optionally, the intermediate residual excitation module includes a first intermediate convolution unit, an intermediate channel squeeze excitation unit, an intermediate spatial excitation unit, and a second intermediate convolution unit; where using the intermediate residual excitation module in the first intermediate attention excitation sublayer to process the initial image grayscale feature to obtain a first residual feature includes: using the first intermediate convolution unit to process the initial image grayscale feature to obtain a first convolution feature; using the intermediate channel squeeze excitation unit to process the first convolution feature to obtain a channel feature, where the intermediate channel squeeze excitation unit is constructed based on the channel attention mechanism; using the intermediate spatial excitation unit to process the first convolution feature to obtain a spatial feature, where the intermediate spatial excitation unit is constructed based on the spatial attention mechanism; using the second intermediate convolution unit to process the channel feature and the spatial feature to obtain a first residual feature.

[0009] Optionally, the first grayscale feature extraction sublayer includes a convolution module, a grayscale attention module, and a batch normalization module; using the first grayscale feature extraction sublayer to process the Nth image feature to obtain an initial image grayscale feature includes: using the convolution module to process the Nth image feature to obtain a second convolution feature; using the grayscale attention module to process the second convolution feature to obtain an image grayscale attention feature; using the batch normalization module to process the image grayscale attention feature to obtain an initial image grayscale feature.

[0010] Optionally, using a grayscale attention module to process the second convolution feature to obtain an image grayscale attention feature includes: performing vector conversion on the second convolution feature to obtain a query vector, a key vector and a value vector; using a normalization function to process the energy vector to obtain a grayscale attention weight, wherein the energy vector is determined based on the query vector and the key vector; and obtaining an image grayscale attention feature based on the weighted vector, the second convolution feature and an adaptive grayscale enhancement learning parameter, wherein the weighted vector is determined based on the grayscale attention weight and the value vector.

[0011] Optionally, using N cascaded upsampling attention layers to process N image features and image grayscale enhancement features to obtain a magnetic particle imaging denoised image includes: inputting the image grayscale enhancement feature and the Nth image feature into the first upsampling attention layer to obtain the first upsampling feature; inputting the nth upsampling feature and the n-1th image feature into the n-1th upsampling attention layer to obtain the n-1th upsampling feature, wherein n N; the second up-sampled feature and the first image feature are input into the Nth up-sampled attention layer to obtain the magnetic particle imaging denoised image.

[0012] Optionally, the denoising model includes a downsampling attention layer, a grayscale adaptive learning network layer, and an upsampling attention layer, and the denoising model is trained based on the following operations: obtaining training samples, the training samples include a sample noise image and a sample noise-free simulation image, wherein the sample noise-free simulation image is obtained by simulating the sample biological tissue image based on a magnetic particle reconstruction system, and the grayscale value variation range between pixels in different areas in the sample noise-free simulation image satisfies a preset simulation dynamic concentration range; based on a preset noise range, superimposing the noise in the sample noise image to the sample noise-free simulation image to obtain a sample noise simulation image; performing noise enhancement processing on the sample noise simulation image using a Gaussian noise network to obtain a sample target noise image; processing the sample target noise image using a denoising model to obtain a sample predicted magnetic particle imaging denoised image; training a denoising model based on the sample predicted magnetic particle imaging denoised image and the sample noise-free simulation image to obtain a trained denoising model.

[0013] Optionally, based on a preset noise range, the noise in the sample noise image is superimposed on the sample noise-free simulation image to obtain the sample noise simulation image, including: based on a single noise superposition amount parameter, superimposing the noise in different areas of the sample noise image on the sample noise-free simulation image to obtain an initial sample noise simulation image; obtaining the noise level of the initial sample noise simulation image according to the respective grayscale values ​​of the sample noise-free simulation image and the initial sample noise simulation image; when the noise level does not meet the preset noise range, based on the single noise superposition amount parameter, continuing to superimpose the noise in different areas of the sample noise image on the sample noise-free simulation image; when the noise level meets the preset noise range, determining the initial sample noise simulation image as the sample noise simulation image.

[0014] The second aspect of the present invention provides a magnetic particle imaging image processing device, including: an enhancement module, which is used to use a Gaussian noise network to perform noise enhancement processing on a magnetic particle imaging noise image to obtain a target noise image, wherein the magnetic particle imaging noise image is obtained by scanning biological tissue by a magnetic particle imaging system; an extraction module, which is used to use N cascaded downsampling attention layers to perform multi-level feature extraction on the target noise image to obtain N image features, wherein the downsampling attention layer is constructed based on L cascaded downsampling attention excitation sublayers, and N and L are both positive integers; a fusion module, which is used to use a grayscale adaptive learning network layer to perform feature fusion on the Nth image feature to obtain an image grayscale enhancement feature, wherein the grayscale adaptive learning network layer is obtained based on the adjustment of adaptive grayscale enhancement learning parameters; a denoising module, which is used to use N cascaded upsampling attention layers to process N image features and image grayscale enhancement features to obtain a magnetic particle imaging denoised image, wherein the upsampling attention layer is constructed based on L cascaded upsampling attention excitation sublayers.

[0015] According to the magnetic particle imaging image processing method and device provided by the present invention, N image features are obtained by performing multi-level feature extraction on the target noise image using N cascaded downsampling attention layers; the Nth image feature is subjected to feature fusion using a grayscale adaptive learning network layer to obtain an image grayscale enhancement feature; and the N image features and the image grayscale enhancement feature are processed using N cascaded upsampling attention layers. Since the downsampling attention excitation sublayer and the upsampling attention excitation sublayer are both constructed based on the channel attention mechanism and the spatial attention mechanism, the feature excitation fusion is performed by the channel attention mechanism and the spatial attention mechanism in the upsampling stage and the downsampling stage, so that the pixel position, shape, and contour information of the low grayscale value area in the target noise image can be better paid attention to; in addition, the pixel grayscale information in the image feature is adaptively enhanced using the adaptive grayscale enhancement learning parameters, so that the grayscale features of the low grayscale value area are better paid attention to and extracted, and after enhancing the learning of the pixel position information, contour information, grayscale information, etc. of the low grayscale value area, the accuracy of the denoising model for positioning and denoising low grayscale value pixels is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The above and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings.

[0017] Figure 1 A flow chart of a magnetic particle imaging image processing method according to an embodiment of the present invention is shown.

[0018] Figure 2A An example schematic diagram of an intermediate attention incentive sublayer according to an embodiment of the present invention is shown.

[0019] Figure 2B An example schematic diagram of the first grayscale feature extraction sublayer according to an embodiment of the present invention is shown.

[0020] Figure 2C An example schematic diagram of a denoising model according to an embodiment of the present invention is shown;

[0021] Figure 3 An example schematic diagram of generating a sample noise simulation image according to an embodiment of the present invention is shown.

[0022] Figure 4 A structural block diagram of a magnetic particle imaging image processing device according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0023] Below, optional will be described with reference to the accompanying drawings.But it should be understood that these descriptions are exemplary only, and are not intended to limit the scope of the present invention.In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of embodiments of the present invention.However, it is obvious that one or more embodiments may also be implemented without these specific details.In addition, in the following description, the description of known structures and technologies is omitted to avoid unnecessary confusion of concepts of the present invention.

[0024] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise", "include", etc. used herein indicate the existence of features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0025] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0026] When using expressions such as "at least one of A, B, and C, etc.", they should generally be interpreted according to the meaning of the expression commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0027] In the process of realizing the present invention, it was found that magnetic particle imaging is a tracer-based tomography technology that can detect the concentration distribution of superparamagnetic iron oxide nanoparticles (SPIONs). When there are multiple samples with large concentration differences in the field of view of magnetic particle imaging, the signal of the high-concentration sample will interfere with the signal of the low-concentration sample, causing the low-concentration sample signal to be obscured, resulting in loss or damage of its sample information. The current reconstruction algorithm is difficult to accurately restore the tissue information of low-concentration samples. In the field of medical image analysis, the existing reconstruction algorithm based on hardware optimization will increase the cost and complexity of the magnetic particle imaging system, and at the same time cannot consider the impact of different dynamic ranges, and has low robustness and low accuracy in restoring the tissue information of low-concentration samples.

[0028] In view of this, a magnetic particle imaging image processing method and device are optionally provided. The method includes: using a Gaussian noise network to perform noise enhancement processing on a magnetic particle imaging noise image to obtain a target noise image; using N cascaded downsampling attention layers to perform multi-level feature extraction on the target noise image to obtain N image features, wherein the downsampling attention layer is constructed based on L cascaded downsampling attention excitation sublayers; using a grayscale adaptive learning network layer to perform feature fusion on the Nth image feature to obtain an image grayscale enhancement feature, wherein the grayscale adaptive learning network layer is obtained based on the adjustment of adaptive grayscale enhancement learning parameters; using N cascaded upsampling attention layers to process N image features and image grayscale enhancement features to obtain a magnetic particle imaging denoised image, wherein the upsampling attention layer is constructed based on L cascaded upsampling attention excitation sublayers.

[0029] In the technical solution of the present invention, the user information (including but not limited to user personal information, user image information, user device information, such as location information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved are all information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data comply with relevant laws, regulations and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0030] It should be noted that the sequence numbers of the operations in the following method are only used as representations of the operations for the purpose of description, and should not be regarded as representing the execution order of the operations. Unless explicitly stated, the method does not need to be executed completely in the order shown.

[0031] Figure 1 A flow chart of a magnetic particle imaging image processing method according to an embodiment of the present invention is shown.

[0032] like Figure 1 As shown, the method includes operations S110 to S140.

[0033] In operation S110, a Gaussian noise network is used to perform noise enhancement processing on the magnetic particle imaging noise image to obtain a target noise image.

[0034] Optionally, the magnetic particle imaging noise image is obtained by scanning biological tissue using a magnetic particle imaging (MPI) system.

[0035] Optionally, the sample tissue concentration is determined according to the degree of absorption of the particles by the biological tissue. The higher the absorption degree, the greater the sample tissue concentration, and thus the stronger the particle signal, and the higher the brightness reflected in the corresponding area of ​​the magnetic particle imaging image. In the case where there are multiple samples with large concentration differences (such as more than 50 times) in the imaging field of view of the magnetic particle imaging system, the high particle signal will interfere with the surrounding low particle signal, resulting in the high particle signal of the high concentration sample covering the low particle signal of the surrounding low concentration sample, so that the low concentration sample signal is obscured, and the low concentration sample tissue information in the magnetic particle imaging image is lost or damaged.

[0036] Optionally, the magnetic particle imaging noise image represents a magnetic particle imaging image in which low-concentration sample tissue information is lost or damaged due to signal noise. The difference in pixel grayscale values ​​between different regions in the magnetic particle imaging noise image is more than 50 times.

[0037] Optionally, sampling the target noise image from a uniform distribution may include: stepwise sampling the magnetic particle imaging noise image x at each time step t. t Add a Gaussian noise layer to get the noise-enhanced image x t+1 The parameters of the Gaussian noise layer are predicted using the triple-enhanced conditional information and the magnetic particle imaging noise image xt, and the parameters of the Gaussian noise layer are learned in the denoising of the target noise image to achieve inverse denoising and low-particle signal recovery.

[0038] Optionally, the target noise image represents the noise image of the magnetic particle imaging after noise enhancement.

[0039] In operation S120, multi-level feature extraction is performed on the target noisy image using N cascaded down-sampling attention layers to obtain N image features.

[0040] Optionally, the downsampling attention layer is constructed based on L cascaded downsampling attention excitation sublayers, where N and L are both positive integers.

[0041] Optionally, the denoising model includes N cascaded downsampling attention layers, each downsampling attention layer includes L cascaded downsampling attention excitation sublayers, and the target noisy image is input into the first downsampling attention excitation sublayer in the first downsampling attention layer for multi-level feature extraction, and an image feature is obtained in each downsampling attention layer.

[0042] Optionally, the image feature representation incorporates pixel position information, shape information, key point information, etc.

[0043] In operation S130, the Nth image feature is fused using a grayscale adaptive learning network layer to obtain an image grayscale enhanced feature.

[0044] Optionally, the grayscale adaptive learning network layer is obtained based on the adaptive grayscale enhancement learning parameter adjustment.

[0045] Optionally, the Nth image feature is output in the Nth downsampling attention layer, and the Nth image feature is input into the grayscale adaptive learning network layer to obtain the image grayscale enhanced feature.

[0046] Optionally, the grayscale adaptive learning network layer can be constructed based on a grayscale attention mechanism. In the grayscale attention mechanism, the grayscale information of pixels in the Nth image feature is adaptively adjusted and enhanced by using adaptive grayscale enhancement learning parameters to obtain an image grayscale enhancement feature. The image grayscale enhancement feature combines the Nth image feature and the grayscale information enhanced by the grayscale attention mechanism.

[0047] In operation S140, N image features and image grayscale enhancement features are processed using N cascaded upsampling attention layers to obtain a magnetic particle imaging denoised image.

[0048] Optionally, the up-sampling attention layer is constructed based on L cascaded up-sampling attention-excitation sub-layers. The network structure of the up-sampling attention-excitation sub-layer is the same as that of the down-sampling attention-excitation sub-layer.

[0049] Optionally, the structures of N cascaded downsampling attention layers and N cascaded upsampling attention layers are symmetrical, and the image features and image grayscale enhancement features output by the upsampling attention layers are input into the corresponding upsampling attention layers for multi-level feature extraction, and the magnetic particle imaging denoised image is obtained in the Nth upsampling attention layer.

[0050] For example, N=2, the second image feature and the image grayscale enhancement feature are fused and input into the first upsampling attention layer, the output result of the first upsampling attention layer is fused with the first image feature and input into the second upsampling attention layer to obtain the magnetic particle imaging denoised image.

[0051] For example, taking the above-mentioned first upsampling attention layer as an example, L=4, inputting the features after fusing the second image features and the image grayscale enhancement features into the L cascaded upsampling attention excitation sublayers in the first upsampling attention layer may include: inputting the fused features into the first upsampling attention excitation sublayer to obtain the first output result; then inputting the first output result into the second upsampling attention excitation sublayer to obtain the second output result; then inputting the second output result into the third upsampling attention excitation sublayer to obtain the third output result; then inputting the third output result into the fourth upsampling attention excitation sublayer to obtain the fourth output result, and the fourth output result is the output result of the first upsampling attention layer.

[0052] Optionally, the denoised image of magnetic particle imaging represents the magnetic particle imaging image that restores the tissue information of low-concentration samples after removing signal noise.

[0053] Optionally, since both the downsampling attention excitation sublayer and the upsampling attention excitation sublayer are constructed based on the channel attention mechanism and the spatial attention mechanism, feature excitation fusion can be performed through the channel attention mechanism and the spatial attention mechanism in the upsampling stage and the downsampling stage, which can better focus on the pixel positions, shapes, and contour information in the low-gray-value regions of the target noise image. In addition, the adaptive gray-scale enhancement learning parameter is used to adaptively enhance the gray-scale information of the pixel points in the image features, so that the gray-scale features in the low-gray-value regions can be better focused on and extracted. After enhancing the learning of the pixel position information, contour information, and gray-scale information in the low-gray-value regions, the accuracy of the denoising model for locating and denoising low-gray-value pixel points is improved.

[0054] Optionally, the gray-scale adaptive learning network layer includes a first gray-scale feature extraction sublayer, M cascaded intermediate attention excitation sublayers, and a second gray-scale feature extraction sublayer, where M < L and M is a positive integer. Feature fusion is performed on the Nth image feature using the gray-scale adaptive learning network layer to obtain the image gray-scale enhancement feature, including: processing the Nth image feature using the first gray-scale feature extraction sublayer to obtain the initial image gray-scale feature; processing the initial image gray-scale feature using M cascaded intermediate attention excitation sublayers to obtain the image excitation feature, where the image excitation feature represents the fused deep-level image gray-scale information and image position information; processing the image excitation feature using the second gray-scale feature extraction sublayer to obtain the image gray-scale enhancement feature.

[0055] Optionally, the structures of the first gray-scale feature extraction sublayer and the second gray-scale feature extraction sublayer are the same. Both the first gray-scale feature extraction sublayer and the second gray-scale feature extraction sublayer contain adaptive gray-scale enhancement learning parameters, and the intermediate attention excitation sublayer is constructed based on the spatial attention mechanism and the channel attention mechanism.

[0056] Optionally, the Nth image feature is input into the first gray-scale feature extraction sublayer, and the initial image gray-scale feature is output. The initial image gray-scale feature represents the initial pixel point gray-scale information adaptively adjusted and enhanced based on the gray-scale attention mechanism.

[0057] Optionally, the initial image gray-scale feature is input into M cascaded intermediate attention excitation sublayers for multi-level feature extraction to obtain the image excitation feature. The image excitation feature fuses the image information with enhanced position information and key point information based on the channel attention mechanism and the spatial attention mechanism.

[0058] For example, M=2, the initial grayscale features of the image are input into the first intermediate attention excitation sublayer for feature extraction, and then the extracted features are input into the second intermediate attention excitation sublayer to output the image excitation features.

[0059] Optionally, the image excitation feature is input into the second grayscale feature extraction sublayer to output the image grayscale enhancement feature. The image grayscale enhancement feature integrates the initial pixel grayscale information of the Nth image feature, the grayscale information enhanced based on the grayscale attention mechanism, the position information enhanced based on the channel attention mechanism and the spatial attention mechanism, and the key point information.

[0060] Optionally, adaptive enhancement learning is performed on the grayscale information corresponding to different concentration areas in the Nth image feature in the grayscale adaptive learning network layer, so that the grayscale information in the low-concentration area is better focused on and extracted, thereby achieving the removal and restoration of the high-concentration dynamic range magnetic particle imaging noise image.

[0061] Optionally, the intermediate attention excitation sublayer includes an intermediate residual excitation module and an intermediate self-attention module; wherein, using M cascaded intermediate attention excitation sublayers to process the initial grayscale features of the image to obtain the image excitation features includes: using the intermediate residual excitation module in the first intermediate attention excitation sublayer to process the initial grayscale features of the image to obtain the first residual feature; using the intermediate self-attention module in the first intermediate attention excitation sublayer to process the first residual feature to obtain the first intermediate image excitation feature; using M-1 cascaded intermediate attention excitation sublayers to process the first intermediate image excitation feature, and outputting the image excitation feature.

[0062] Optionally, the intermediate self-attention module is constructed based on the self-attention mechanism. The intermediate residual excitation module is constructed based on the channel attention mechanism and the spatial attention mechanism.

[0063] Optionally, the initial grayscale feature of the image is input into the intermediate residual excitation module in the first intermediate attention excitation sublayer to obtain the first residual feature; and then the first residual feature is input into the intermediate self-attention module to obtain the first intermediate image excitation feature.

[0064] Optionally, the first intermediate image excitation feature is input into M-1 cascaded intermediate attention excitation sublayers, the processing process in the intermediate attention excitation sublayer is the same as the processing process in the first intermediate attention excitation sublayer, and the image excitation feature is output in the Mth intermediate attention excitation sublayer.

[0065] For example, M=4, the first intermediate image excitation feature is input into the intermediate attention excitation sublayer of the second cascade to obtain the second intermediate image excitation feature; the second intermediate image excitation feature is input into the intermediate attention excitation sublayer of the third cascade to obtain the third intermediate image excitation feature; the third intermediate image excitation feature is input into the intermediate attention excitation sublayer of the fourth cascade to obtain the image excitation feature.

[0066] Optionally, the intermediate residual excitation module includes a first intermediate convolution unit, an intermediate channel squeezing excitation unit, an intermediate spatial excitation unit, and a second intermediate convolution unit; wherein, using the intermediate residual excitation module in the first intermediate attention excitation sublayer to process the initial grayscale features of the image to obtain the first residual feature includes: using the first intermediate convolution unit to process the initial grayscale features of the image to obtain the first convolution feature; using the intermediate channel squeezing excitation unit to process the first convolution feature to obtain the channel feature, wherein the intermediate channel squeezing excitation unit is constructed based on the channel attention mechanism; using the intermediate spatial excitation unit to process the first convolution feature to obtain the spatial feature, wherein the intermediate spatial excitation unit is constructed based on the spatial attention mechanism; using the second intermediate convolution unit to process the channel feature and the spatial feature to obtain the first residual feature.

[0067] Optionally, the first intermediate convolution unit and the second intermediate convolution unit are both constructed based on a 3×3 convolution kernel.

[0068] Optionally, the intermediate channel squeezing excitation unit is constructed based on the channel attention mechanism, and the intermediate spatial excitation unit is constructed based on the spatial attention mechanism.

[0069] Figure 2A An example schematic diagram of an intermediate attention incentive sublayer according to an embodiment of the present invention is shown.

[0070] like Figure 2A As shown, the intermediate attention excitation sublayer includes an intermediate residual excitation module and an intermediate self-attention module. The intermediate residual excitation module includes a first intermediate convolution unit, an intermediate channel squeezing excitation unit, an intermediate spatial excitation unit, and a second intermediate convolution unit. The image grayscale initial feature is input into the first intermediate convolution unit to obtain the first convolution feature; the first convolution feature is respectively input into the intermediate channel squeezing excitation unit and the intermediate spatial excitation unit to obtain the channel feature and the spatial feature respectively. The channel feature and the spatial feature are spliced, and then the spliced ​​features are input into the second intermediate convolution unit to obtain the first residual feature, and the first residual feature is input into the intermediate self-attention module to obtain the first intermediate image excitation feature.

[0071] Optionally, in the middle channel squeeze excitation unit, the dimension of the first convolution feature is changed from (C, H, WC, H, W) to (C, 1, 1C, 1, 1) by global average pooling; two 1×1 convolution layers are used for dimensionality reduction and dimensionality increase; the sigmoid function is used for normalization to obtain the channel attention weight; the channel attention weight is multiplied by the first convolution feature channel by channel to enhance the feature information such as the contour and shape of the image, and then the channel feature is obtained. The whole process can suppress useless features and speed up learning. The channel feature integrates key point information, pixel texture information, pixel contour information, etc.

[0072] Optionally, the spatial feature is pixel position information. In the intermediate spatial excitation unit, the spatial dimension of the first convolution feature is operated, and the channel information of each spatial position of the first convolution feature is integrated into a weight value through a 1×1 convolution layer; the weight value is normalized by a sigmoid function and weighted to the corresponding spatial position of the first convolution feature, thereby enhancing the spatially important features and suppressing the unimportant features, thereby obtaining the spatial features.

[0073] Optionally, feature excitation fusion is performed through the channel attention mechanism and the spatial attention mechanism to better focus on the location and pixel detail information of low-concentration samples in the image, thereby improving the accuracy and other performance of the denoising model.

[0074] Optionally, the first grayscale feature extraction sublayer includes a convolution module, a grayscale attention module and a batch normalization module; using the first grayscale feature extraction sublayer to process the Nth image feature to obtain the image grayscale initial feature includes: using the convolution module to process the Nth image feature to obtain a second convolution feature; using the grayscale attention module to process the second convolution feature to obtain the image grayscale attention feature; using the batch normalization module to process the image grayscale attention feature to obtain the image grayscale initial feature.

[0075] Optionally, the convolution module is constructed based on a 3×3 convolution kernel.

[0076] Optionally, the grayscale attention module is constructed based on the grayscale attention mechanism.

[0077] Figure 2B An example schematic diagram of the first grayscale feature extraction sublayer according to an embodiment of the present invention is shown.

[0078] like Figure 2B As shown, the first grayscale feature extraction sublayer includes a convolution module, a grayscale attention module and a batch normalization module. The Nth image feature is input into the convolution module to obtain the second convolution feature; the second convolution feature is input into the grayscale attention module to obtain the image grayscale attention feature.

[0079] Optionally, the image grayscale attention feature is grayscale information enhanced based on the grayscale attention mechanism.

[0080] Optionally, the image grayscale attention features are input into a batch normalization module to obtain the image grayscale initial features.

[0081] Optionally, using a grayscale attention module to process the second convolution feature to obtain an image grayscale attention feature includes: performing vector conversion on the second convolution feature to obtain a query vector, a key vector and a value vector; using a normalization function to process the energy vector to obtain a grayscale attention weight, wherein the energy vector is determined based on the query vector and the key vector; and obtaining an image grayscale attention feature based on the weighted vector, the second convolution feature and an adaptive grayscale enhancement learning parameter, wherein the weighted vector is determined based on the grayscale attention weight and the value vector.

[0082] Optionally, the second convolutional feature is vectorized to obtain a query vector (Q), a key vector (K), and a value vector (V).

[0083] Optionally, a dot product is calculated between the query vector and the key vector to obtain an energy vector.

[0084] In one embodiment, the energy vector E is as shown in formula (1):

[0085] (1);

[0086] in, Represents the query vector, Represents the transpose of the key vector.

[0087] In one embodiment, the grayscale attention weight A is as shown in formula (2):

[0088] (2);

[0089] in, Represents the energy vector, Characterize the normalization function.

[0090] In one embodiment, the weight vector As shown in formula (3):

[0091] (3);

[0092] in, Characterize the value vector, Represents the transpose of the grayscale attention weights.

[0093] In one embodiment, the image grayscale attention feature As shown in formula (4):

[0094] (4);

[0095] in, Characterize the adaptive grayscale reinforcement learning parameters, represents the weight vector, Characterize the second convolutional feature.

[0096] Optionally, the second convolution feature is adaptively adjusted according to the adaptive grayscale enhancement learning parameters, and the grayscale information corresponding to different concentration areas is adaptively enhanced learning, so that the grayscale information of the low-concentration area is better focused and extracted, thereby achieving the removal and restoration of the high-concentration dynamic range magnetic particle imaging noise image.

[0097] Optionally, using N cascaded upsampling attention layers to process N image features and image grayscale enhancement features to obtain a magnetic particle imaging denoised image includes: inputting the image grayscale enhancement feature and the Nth image feature into the first upsampling attention layer to obtain the first upsampling feature; inputting the nth upsampling feature and the n-1th image feature into the n-1th upsampling attention layer to obtain the n-1th upsampling feature, wherein n N; the second up-sampled feature and the first image feature are input into the Nth up-sampled attention layer to obtain the magnetic particle imaging denoised image.

[0098] Optionally, the N image features and image grayscale enhancement features output by N upsampling attention layers are input into the corresponding upsampling attention layer for multi-level feature extraction, and the magnetic particle imaging denoised image is obtained in the Nth upsampling attention layer.

[0099] Figure 2C An example schematic diagram of a denoising model according to an embodiment of the present invention is shown.

[0100] like Figure 2C As shown, taking N=4, L=4, and M=2 as an example, the upsampling attention layer includes 4 cascaded upsampling attention excitation sublayers, the downsampling attention layer includes 4 cascaded downsampling attention excitation sublayers, and the grayscale adaptive learning network layer includes a first grayscale feature extraction sublayer, 2 cascaded intermediate attention excitation sublayers, and a second grayscale feature extraction sublayer. The intermediate attention excitation sublayer has the same network structure as the upsampling attention excitation sublayer and the downsampling attention excitation sublayer. The target noise image is input into the 4 cascaded downsampling attention layers for multi-level feature extraction, and 4 image features are output respectively; the 4th image feature is input into the grayscale adaptive learning network layer to obtain the image grayscale enhancement feature.

[0101] Optionally, the fourth image feature and the image grayscale enhancement feature are fused and input into the first upsampling attention layer to obtain the first upsampling feature; the first upsampling feature and the third image feature are fused and input into the second upsampling attention layer to obtain the second upsampling feature; the second upsampling feature and the second image feature are fused and input into the third upsampling attention layer to obtain the third upsampling feature; the third upsampling feature and the first image feature are fused and input into the fourth upsampling attention layer to obtain a magnetic particle imaging denoised image.

[0102] Optionally, the denoising model includes a downsampling attention layer, a grayscale adaptive learning network layer, and an upsampling attention layer, and the denoising model is trained based on the following operations: obtaining training samples, the training samples include a sample noise image and a sample noise-free simulation image, wherein the sample noise-free simulation image is obtained by simulating the sample biological tissue image based on a magnetic particle reconstruction system, and the grayscale value variation range between pixels in different areas in the sample noise-free simulation image satisfies a preset simulation dynamic concentration range; based on a preset noise range, superimposing the noise in the sample noise image to the sample noise-free simulation image to obtain a sample noise simulation image; performing noise enhancement processing on the sample noise simulation image using a Gaussian noise network to obtain a sample target noise image; processing the sample target noise image using a denoising model to obtain a sample predicted magnetic particle imaging denoised image; training a denoising model based on the sample predicted magnetic particle imaging denoised image and the sample noise-free simulation image to obtain a trained denoising model.

[0103] Optionally, the preset simulated dynamic concentration range is that the particle absorption concentration range in different regions of the sample biological tissue differs by more than 50 times.

[0104] Optionally, the sample biological tissue image can be a binary image in the shape of a point or tumor generated by an algorithm. Since the training process requires a large number of training samples for training, it is necessary to use a magnetic particle reconstruction system to simulate the sample biological tissue image to obtain 100,000 sample noise-free simulation images. The sample noise-free simulation image is the imaging result of the simulated particles penetrating the tumor tissue.

[0105] Optionally, the sample noise-free simulated image is a grayscale image with a grayscale value distribution of 0-255. The size, position, grayscale value and number of tumors and pixels in the sample noise-free simulated image are randomly distributed to simulate the results of magnetic particle imaging in real situations. The grayscale values ​​between pixels in different areas of the sample noise-free simulated image differ by more than 50 times to meet the preset simulated dynamic concentration range.

[0106] Optionally, the sample noise image is a sample magnetic particle imaging image in which low-concentration sample tissue information is lost or damaged based on the influence of signal noise. The noise in the sample noise image is repeatedly superimposed on the sample noise-free simulation image. When the noise of the sample noise-free simulation image meets the preset noise range, the superposition is stopped to obtain the sample noise simulation image.

[0107] Figure 3 An example schematic diagram of generating a sample noise simulation image according to an embodiment of the present invention is shown.

[0108] like Figure 3 As shown, the sample biological tissue image is simulated by a magnetic particle reconstruction system to obtain a sample noise-free simulation image, and the tumor tissue is scanned by a magnetic particle imaging device to obtain a sample noise image. The noise in the sample noise image is repeatedly superimposed on the sample noise-free simulation image to obtain a sample noise simulation image.

[0109] Optionally, in the process of forward noise enhancement, a Gaussian noise network is used to perform noise enhancement processing on the sample noise simulation image to obtain a sample target noise image. In the process of reverse denoising, a denoising model is used to process the sample target noise image to obtain a sample predicted magnetic particle imaging denoised image.

[0110] Optionally, during the training process, the number of training epochs is set to 400. The loss function can be used to calculate the loss value between the sample predicted magnetic particle imaging denoised image and the sample noise-free simulation image, and the parameters of the denoising model are adjusted according to the loss value to train the denoising model until the model is optimal.

[0111] Optionally, the downsampling attention layer, the grayscale adaptive learning network layer, and the upsampling attention layer in the denoising model are used to better focus on the location and detail information of low-concentration tissues in the sample target noise image, and the grayscale features of different concentration areas in the sample target noise image are adaptively enhanced to learn, so that the grayscale features of the low-concentration area can be better focused on and extracted. Therefore, it has a certain generalization performance for different dynamic ranges and noise levels, and ultimately the low-concentration particle distribution is accurately restored, thereby realizing the reconstruction of high dynamic range magnetic particle imaging images.

[0112] Optionally, based on a preset noise range, the noise in the sample noise image is superimposed on the sample noise-free simulation image to obtain the sample noise simulation image, including: based on a single noise superposition amount parameter, superimposing the noise in different areas of the sample noise image on the sample noise-free simulation image to obtain an initial sample noise simulation image; obtaining the noise level of the initial sample noise simulation image according to the respective grayscale values ​​of the sample noise-free simulation image and the initial sample noise simulation image; when the noise level does not meet the preset noise range, based on the single noise superposition amount parameter, continuing to superimpose the noise in different areas of the sample noise image on the sample noise-free simulation image; when the noise level meets the preset noise range, determining the initial sample noise simulation image as the sample noise simulation image.

[0113] Optionally, during each noise superposition process, noise in different regions of the sample noise image is superimposed on the sample noise-free simulation image of the corresponding region according to a noise superposition amount parameter to obtain an initial sample noise simulation image.

[0114] In one embodiment, the noise level of the initial sample noise simulation image is As shown in formula (5):

[0115] (5);

[0116] in, Characterizes the maximum pixel gray value in the sample noise-free simulation image, Characterizes the mean square error of the pixel grayscale values ​​of the sample noise-free simulated image and the initial sample noise simulated image.

[0117] Optionally, the preset noise range is a numerical interval. When the noise level of the initial sample noise simulation image is not within the numerical interval of the preset noise range, based on the single noise superposition amount parameter, the noise in different areas of the sample noise image is again superimposed on the sample noise-free simulation image.

[0118] Optionally, when the noise level is within a numerical interval of a preset noise range, the noise superposition is stopped, and the current initial sample noise simulation image is determined as the sample noise simulation image.

[0119] Based on the above magnetic particle imaging image processing method, the present invention also provides a magnetic particle imaging image processing device. Figure 4 The device is described in detail.

[0120] Figure 4 A structural block diagram of a magnetic particle imaging image processing device according to an embodiment of the present invention is shown.

[0121] like Figure 4As shown, the magnetic particle imaging image processing device 400 of this embodiment includes an enhancement module 410 , an extraction module 420 , a fusion module 430 and a denoising module 440 .

[0122] The enhancement module 410 is used to perform noise enhancement processing on the magnetic particle imaging noise image using a Gaussian noise network to obtain a target noise image, wherein the magnetic particle imaging noise image is obtained by scanning biological tissues by a magnetic particle imaging system. In one embodiment, the enhancement module 410 can be used to perform the operation S110 described above, which will not be described in detail here.

[0123] The extraction module 420 is used to perform multi-level feature extraction on the target noise image using N cascaded down-sampling attention layers to obtain N image features, wherein the down-sampling attention layer is constructed based on L cascaded down-sampling attention excitation sub-layers, and N and L are both positive integers. In one embodiment, the extraction module 420 can be used to perform the operation S120 described above, which will not be repeated here.

[0124] The fusion module 430 is used to perform feature fusion on the Nth image feature using the grayscale adaptive learning network layer to obtain the image grayscale enhanced feature, wherein the grayscale adaptive learning network layer is obtained based on the adaptive grayscale enhanced learning parameter adjustment. In one embodiment, the fusion module 430 can be used to perform the operation S130 described above, which will not be repeated here.

[0125] The denoising module 440 is used to process N image features and image grayscale enhancement features using N cascaded upsampling attention layers to obtain a magnetic particle imaging denoised image, wherein the upsampling attention layer is constructed based on L cascaded upsampling attention excitation sublayers. In one embodiment, the denoising module 440 can be used to perform the operation S140 described above, which will not be repeated here.

[0126] Optionally, the fusion module 430 includes a first fusion submodule, a second fusion submodule and a third fusion submodule.

[0127] The first fusion submodule is used to process the Nth image feature using the first grayscale feature extraction sublayer to obtain an initial image grayscale feature.

[0128] The second fusion submodule is used to use M cascaded intermediate attention excitation sublayers to process the initial grayscale features of the image to obtain image excitation features, wherein the image excitation feature representation integrates deep-level image grayscale information and image position information.

[0129] The third fusion submodule is used to process the image excitation feature using the second grayscale feature extraction sublayer to obtain the image grayscale enhancement feature.

[0130] Optionally, the second fusion submodule includes a first fusion unit, a second fusion unit and a third fusion unit.

[0131] The first fusion unit is used to process the initial grayscale features of the image using the intermediate residual excitation module in the first intermediate attention excitation sublayer to obtain the first residual feature.

[0132] The second fusion unit is used to process the first residual feature using the intermediate self-attention module in the first intermediate attention excitation sublayer to obtain the first intermediate image excitation feature.

[0133] The third fusion unit is used to process the first intermediate image excitation feature using M-1 cascaded intermediate attention excitation sublayers and output the image excitation feature.

[0134] Optionally, the first fusion unit includes a first fusion subunit, a second fusion subunit, a third fusion subunit and a fourth fusion subunit.

[0135] The first fusion subunit is used to process the initial grayscale features of the image using the first intermediate convolution unit to obtain the first convolution features.

[0136] The second fusion subunit is used to process the first convolution feature by using the intermediate channel squeezing and excitation unit to obtain the channel feature, wherein the intermediate channel squeezing and excitation unit is constructed based on the channel attention mechanism.

[0137] The third fusion subunit is used to process the first convolution feature by using the intermediate spatial excitation unit to obtain the spatial feature, wherein the intermediate spatial excitation unit is constructed based on the spatial attention mechanism.

[0138] The fourth fusion subunit is used to process the channel feature and the spatial feature using the second intermediate convolution unit to obtain the first residual feature.

[0139] Optionally, the first fusion submodule includes a fourth fusion unit, a fifth fusion unit and a sixth fusion unit.

[0140] The fourth fusion unit is used to process the Nth image feature by using the convolution module to obtain a second convolution feature.

[0141] The fifth fusion unit is used to process the second convolution feature by using the grayscale attention module to obtain the image grayscale attention feature.

[0142] The sixth fusion unit is used to process the image grayscale attention features using a batch normalization module to obtain the image grayscale initial features.

[0143] Optionally, the fifth fusion unit includes a fifth fusion subunit, a sixth fusion subunit and a seventh fusion subunit.

[0144] The fifth fusion subunit is used to perform vector conversion on the second convolution feature to obtain a query vector, a key vector and a value vector.

[0145] The sixth fusion subunit is used to process the energy vector using a normalization function to obtain a grayscale attention weight, wherein the energy vector is determined based on the query vector and the key vector.

[0146] The seventh fusion subunit is used to obtain the image grayscale attention feature according to the weighted vector, the second convolution feature and the adaptive grayscale enhancement learning parameter, wherein the weighted vector is determined based on the grayscale attention weight and the value vector.

[0147] Optionally, the denoising module 440 includes a first denoising submodule, a second denoising submodule and a third denoising submodule.

[0148] The first denoising submodule is used to input the image grayscale enhancement feature and the Nth image feature into the first upsampling attention layer to obtain the first upsampling feature.

[0149] The second denoising submodule is used to input the nth upsampled feature and the n-1th image feature into the n-1th upsampled attention layer to obtain the n-1th upsampled feature, where n N.

[0150] The third denoising submodule is used to input the second up-sampled feature and the first image feature into the Nth up-sampled attention layer to obtain a magnetic particle imaging denoised image.

[0151] Optionally, any multiple modules among the modules, submodules, units, and subunits can be combined into one module for implementation, or any one of the modules can be split into multiple modules. Alternatively, at least part of the functions of one or more of these modules can be combined with at least part of the functions of other modules and implemented in one module. Optionally, at least one of the modules, submodules, units, and subunits can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or can be implemented by hardware or firmware such as any other reasonable way of integrating or packaging the circuit, or implemented in any one of the three implementation methods of software, hardware, and firmware, or in any appropriate combination of any of them. Alternatively, at least one of the modules, submodules, units, and subunits can be at least partially implemented as a computer program module, and when the computer program module is run, the corresponding function can be executed.

[0152] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. In this regard, each box in the flowchart or block diagram may represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box may also occur in an order different from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions. It can be understood by those skilled in the art that the features recorded in the various embodiments of the present invention can be combined and / or combined in various ways, even if such a combination or combination is not explicitly recorded in the present invention. In particular, without departing from the spirit and teachings of the present invention, the features recorded in the various embodiments of the present invention can be combined and / or combined in various ways. All these combinations and / or combinations fall within the scope of the present invention.

[0153] The above is described as optional. However, these embodiments are only for the purpose of illustration, and are not intended to limit the scope of the present invention. Although each embodiment is described above, this does not mean that the measures in each embodiment cannot be used in combination advantageously. Without departing from the scope of the present invention, those skilled in the art may make multiple substitutions and modifications, which should all fall within the scope of the present invention.

Claims

1. A magnetic particle imaging image processing method, characterized in that: The method includes: Performing noise enhancement processing on the magnetic particle imaging noise image by using a Gaussian noise network to obtain a target noise image, where the magnetic particle imaging noise image is obtained by scanning biological tissue with a magnetic particle imaging system; Performing multi-level feature extraction on the target noise image by using N cascaded downsampling attention layers to obtain N image features, where the downsampling attention layer is constructed based on L cascaded downsampling attention excitation sub-layers, and both N and L are positive integers; Performing feature fusion on the Nth image feature by using a gray-scale adaptive learning network layer to obtain an image gray-scale enhancement feature, including: Processing the Nth image feature by using a first gray-scale feature extraction sub-layer to obtain an initial image gray-scale feature; Processing the initial image gray-scale feature by using M cascaded intermediate attention excitation sub-layers to obtain an image excitation feature, where the image excitation feature represents the fusion of deep-level image gray-scale information and image position information; Processing the image excitation feature by using a second gray-scale feature extraction sub-layer to obtain the image gray-scale enhancement feature, where the gray-scale adaptive learning network layer is adjusted based on adaptive gray-scale enhancement learning parameters, and the gray-scale adaptive learning network layer includes the first gray-scale feature extraction sub-layer, M cascaded intermediate attention excitation sub-layers, and the second gray-scale feature extraction sub-layer, M < L, and M is a positive integer; Processing the N image features and the image gray-scale enhancement feature by using N cascaded upsampling attention layers to obtain a magnetic particle imaging denoised image, where the upsampling attention layer is constructed based on L cascaded upsampling attention excitation sub-layers.

2. The method according to claim 1, characterized in that The intermediate attention excitation sub-layer includes an intermediate residual excitation module and an intermediate self-attention module; Among them, the processing of the initial image gray-scale feature by using M cascaded intermediate attention excitation sub-layers to obtain an image excitation feature includes: Processing the initial image gray-scale feature by using the intermediate residual excitation module in the first intermediate attention excitation sub-layer to obtain a first residual feature; Processing the first residual feature by using the intermediate self-attention module in the first intermediate attention excitation sub-layer to obtain a first intermediate image excitation feature; Processing the first intermediate image excitation feature by using M - 1 cascaded intermediate attention excitation sub-layers and outputting the image excitation feature.

3. The method according to claim 2, characterized in that The intermediate residual excitation module includes a first intermediate convolution unit, an intermediate channel squeeze excitation unit, an intermediate spatial excitation unit, and a second intermediate convolution unit; Among them, the processing of the initial image gray-scale feature by using the intermediate residual excitation module in the first intermediate attention excitation sub-layer to obtain a first residual feature includes: Processing the initial image gray-scale feature by using the first intermediate convolution unit to obtain a first convolution feature; Processing the first convolution feature by using the intermediate channel squeeze excitation unit to obtain a channel feature, where the intermediate channel squeeze excitation unit is constructed based on a channel attention mechanism; Processing the first convolution feature using the intermediate spatial excitation unit to obtain a spatial feature, wherein the intermediate spatial excitation unit is constructed based on a spatial attention mechanism; The second intermediate convolution unit is used to process the channel feature and the spatial feature to obtain the first residual feature.

4. The method according to claim 1, characterized in that The first grayscale feature extraction sublayer includes a convolution module, a grayscale attention module and a batch normalization module; The using the first grayscale feature extraction sublayer to process the Nth image feature to obtain the image grayscale initial feature comprises: Processing the Nth image feature using the convolution module to obtain a second convolution feature; Processing the second convolution feature using the grayscale attention module to obtain an image grayscale attention feature; The batch normalization module is used to process the image grayscale attention feature to obtain the image grayscale initial feature.

5. The method according to claim 4, characterized in that The step of processing the second convolution feature using the grayscale attention module to obtain the image grayscale attention feature comprises: Performing vector conversion on the second convolutional feature to obtain a query vector, a key vector, and a value vector; Processing an energy vector using a normalization function to obtain a grayscale attention weight, wherein the energy vector is determined based on the query vector and the key vector; The image grayscale attention feature is obtained according to the weighted vector, the second convolution feature and the adaptive grayscale enhancement learning parameter, wherein the weighted vector is determined based on the grayscale attention weight and the value vector.

6. The method according to claim 1, characterized in that The method of using N cascaded upsampling attention layers to process the N image features and the image grayscale enhancement features to obtain a magnetic particle imaging denoised image comprises: Inputting the image grayscale enhancement feature and the Nth image feature into the first upsampling attention layer to obtain the first upsampling feature; Input the nth upsampled feature and the n-1th image feature into the n-1th upsampled attention layer to obtain the n-1th upsampled feature, where n≤N; The second up-sampled feature and the first image feature are input into the Nth up-sampled attention layer to obtain the magnetic particle imaging denoised image.

7. The method according to claim 1, characterized in that The denoising model includes the downsampling attention layer, the grayscale adaptive learning network layer, and the upsampling attention layer. The denoising model is trained based on the following operations: Acquire training samples, wherein the training samples include sample noise images and sample noise-free simulated images, wherein the sample noise-free simulated images are obtained by simulating sample biological tissue images based on a magnetic particle reconstruction system, and the grayscale value variation range between pixel points in different areas in the sample noise-free simulated images meets a preset simulated dynamic concentration range; Based on a preset noise range, the noise in the sample noise image is superimposed on the sample noise-free simulation image to obtain a sample noise simulation image; Using the Gaussian noise network to perform noise enhancement processing on the sample noise simulation image to obtain a sample target noise image; Processing the sample target noise image using the denoising model to obtain a sample predicted magnetic particle imaging denoised image; Predict the denoised magnetic particle imaging image and the noise-free simulation image of the sample according to the sample, and train the denoising model to obtain the trained denoising model.

8. The method according to claim 7, characterized in that The step of adding the noise in the sample noise image to the noise-free simulation image of the sample based on a preset noise range to obtain a sample noise simulation image includes: Based on the single noise addition amount parameter, add the noise in different regions of the sample noise image to the noise-free simulation image of the sample to obtain an initial sample noise simulation image; According to the gray values of the noise-free simulation image of the sample and the initial sample noise simulation image respectively, obtain the noise degree of the initial sample noise simulation image; In the case that the noise degree does not meet the preset noise range, based on the single noise addition amount parameter, continue to add the noise in different regions of the sample noise image to the noise-free simulation image of the sample; In the case that the noise degree meets the preset noise range, determine the initial sample noise simulation image as the sample noise simulation image.

9. A magnetic particle imaging image processing device, the device includes: An enhancement module, configured to perform noise enhancement processing on a magnetic particle imaging noise image by using a Gaussian noise network to obtain a target noise image, where the magnetic particle imaging noise image is obtained by scanning a biological tissue by a magnetic particle imaging system; An extraction module, configured to perform multi-level feature extraction on the target noise image by using N cascaded downsampling attention layers to obtain N image features, where the downsampling attention layer is constructed based on L cascaded downsampling attention excitation sub-layers, and both N and L are positive integers; A fusion module, configured to perform feature fusion on the Nth image feature by using a gray-scale adaptive learning network layer to obtain an image gray-scale enhancement feature, and the fusion module includes a first fusion sub-module, a second fusion sub-module, and a third fusion sub-module; The first fusion sub-module is configured to process the Nth image feature by using a first gray-scale feature extraction sub-layer to obtain an initial image gray-scale feature; The second fusion sub-module is configured to process the initial image gray-scale feature by using M cascaded intermediate attention excitation sub-layers to obtain an image excitation feature, where the image excitation feature represents the fusion of deep image gray-scale information and image position information; The third fusion sub-module is configured to process the image excitation feature by using a second gray-scale feature extraction sub-layer to obtain the image gray-scale enhancement feature, where the gray-scale adaptive learning network layer is adjusted based on an adaptive gray-scale enhancement learning parameter, and the gray-scale adaptive learning network layer includes the first gray-scale feature extraction sub-layer, M cascaded intermediate attention excitation sub-layers, and the second gray-scale feature extraction sub-layer, M < L, and M is a positive integer; A denoising module, configured to process the N image features and the image gray-scale enhancement feature by using N cascaded upsampling attention layers to obtain a magnetic particle imaging denoised image, where the upsampling attention layer is constructed based on L cascaded upsampling attention excitation sub-layers.

Images