MPI magnetic particle imaging method and system for early detection of AD patients

Abstract

The present invention belongs to the field of magnetic nanoparticle imaging technology, and specifically relates to an MPI magnetic particle imaging method and system for early detection of AD patients, which aims to complete MPI time-series imaging using a diffusion model, and solve the problem that the characteristic dimension of the input signal in the diffusion model does not match the resolution of the expected generated image. The method includes: constructing a generative network model and training in stages; collecting MPI time-series one-dimensional signals, reconstructing images based on the generative network model to obtain MPI time-series images; performing registration and dividing brain regions; and constructing a functional connection map and a functional connection matrix using RV coefficients to draw brain networks. The present invention constructs a Seq‑VAE / Regression / Sig‑LDM step-by-step network and trains in stages to achieve free input of signal sequences with fixed characteristic dimensions, improve the image sequence reconstruction effect, and draw a default mode network based on the reconstructed image sequence to provide accurate data for subsequent applications.

Description

Background Art

[0002] In the field of neuroscience, the traditional view has long been that different areas of the human brain have relatively independent and different functions. However, with the deepening of research in recent years, the field has gradually come to believe that the relationship between cognitive function and brain regions is not a simple linear relationship, but a many-to-many complex topological relationship, that is, the realization of a certain function requires the cooperation of multiple brain regions, and a certain brain region also has the ability to realize multiple functions. Experiments have shown that the human brain also conforms to the above understanding when handling daily tasks, mainly through the joint operation of multiple interacting systems to form a network, thereby realizing various functions.

[0003] In order to better construct and explore the internal structure and network properties of brain networks, neuroscientists used tools and concepts derived from network science, such as nodes (representing brain regions), edges (representing functional connections between brain regions), degree (the number of edges connected to reference nodes), clustering coefficient (a measure of the degree of local clustering of the network), characteristic path length (a measure of the network's information transmission ability) and global efficiency (a measure of how fast information is transmitted in the network), etc., to provide a detailed and quantifiable description of the brain's topological structure and the interactions between multiple neural systems in the network, laying a theoretical foundation for analyzing experimental results.

[0004] In the brain network, the default mode network (DMN) is activated when in a resting state and not performing external tasks. Studies on the default mode network have shown that it is associated with a variety of neurological diseases, such as Alzheimer's disease (AD). AD is a degenerative disease of the central nervous system, and the main feature of patients with AD is decreased cognitive function. Therefore, researchers often use functional neuroimaging technology to construct brain networks and verify the topological property connections within and between brain networks, so as to achieve the feasibility and practical value of diagnosing abnormal brain networks in patients with Alzheimer's disease or other related diseases in the early stages.

[0005] In the prior art, magnetic nanoparticle imaging (MPI) is used to reconstruct the concentration distribution of super-sequential nanoparticles in the object to be tested. There are two traditional MPI reconstruction technologies. One relies on the system matrix to pre-characterize the signal response of SPION, and the other reconstruction technology is the X-space algorithm. However, the above two methods are limited by the amount of calculation and the calculation method, and are not suitable for reconstructing three-dimensional time-series image sequences. If a deep learning method is used to reconstruct the MPI image sequence, the generation model has technical routes such as GAN or diffusion model to choose from, but GAN generally has problems such as difficulty in fitting the training process or insufficient diversity of generated data.

[0006] In summary, it is necessary to study the feasibility of combining MPI technology and diffusion model to construct DMN brain network and functional connection matrix, and analyze the topological property connection between DMN brain network and functional connection matrix. When constructing brain network or functional connection matrix, traditional methods often use methods such as Pearson correlation and linear regression analysis; although it is simple to use, since functional brain areas are usually clustered together and show regional homogeneity, this highly detailed spatial structural information will be lost in the univariate analysis process; and the obtained functional connection map often has the effect of "salt and pepper noise", which hinders the analysis of the results. At the same time, the traditional diffusion model has the problem that it cannot freely input signals of fixed feature dimensions and generate them as fixed resolution images. Therefore, the traditional diffusion model method is not suitable for reconstructing time series image sequences. Summary of the invention

[0007] In order to solve the above problems in the prior art, namely, the traditional diffusion model is not suitable for reconstructing time series image sequences, and the traditional method of constructing functional connection maps is prone to spatial structural information loss or salt and pepper noise, the first aspect of the present invention proposes an MPI magnetic particle imaging method for early detection of AD patients, which is used to reconstruct the MPI time series image of the target to be modeled, and then draw the default mode network. The method comprises the following steps:

[0008] Step S100: constructing a generative network model and training the generative network model;

[0009] Step S200, the MPI device collects the MPI time series one-dimensional signal of the target to be modeled in a resting state;

[0010] Step S300, inputting the MPI time series one-dimensional signal into the trained generative network model for image reconstruction to obtain an MPI time series image;

[0011] Step S400, preprocessing the MPI time series images, and dividing the preprocessed MPI time series images into brain regions to obtain partition time series signals;

[0012] Step S500, constructing a functional connectivity map and / or a functional connectivity matrix based on the partitioned time series signal using the RV coefficient;

[0013] Step S600, converting the RV values ​​in the functional connectivity map and / or the functional connectivity matrix into Z values, performing statistical analysis on each Z value to obtain the significance of the functional connectivity of the target to be modeled, and drawing the default mode network of the target to be modeled;

[0014] The generative network model includes a sequence variational auto-encoder (Seq-VAE), a regression network and a latent diffusion model (Signal Latent Diffusion Model, Sig-LDM), which is a Seq-VAE / Regression / Sig-LDM step-by-step network; the Seq-VAE network includes a first adaptive signal linear layer, a GRUcell (gated recurrent unit) network, a second adaptive signal linear layer, an Encoder layer, a Decoder layer and a splicing layer connected in series in sequence; the Regression network includes multiple superimposed linear layers; the Sig-LDM network includes a forward noise addition module, a prediction noise module and a reverse denoising module, and the prediction noise module is a prediction noise model Time-condition Swin-Unet;

[0015] When training the generative network model, the Seq-VAE network, the Regression network and the Sig-LDM network are trained in three stages respectively.

[0016] In some preferred embodiments, when training the generative network model, simulation software is used to generate a three-dimensional simulation image sequence and a one-dimensional signal sequence as a training set, and the simulation image sequence and the signal sequence correspond one to one; and a pre-made MPI time-series one-dimensional signal of a phantom is collected in real time to form a test set.

[0017] In some preferred embodiments, the generative network model is trained in three stages, and the method is as follows:

[0018] Step S121, using the simulated image sequence in the training set as the input of the Seq-VAE network to train the Seq-VAE network;

[0019] Step S122, inputting the simulated image sequence in the training set into the first adapted signal linear layer of the trained Seq-VAE network, using the output tensor of the first adapted signal linear layer as the regression target output by the Regression network, using the signal sequence in the training set as the input of the Regression network, and training the Regression network;

[0020] Step S123: input the simulated image sequence in the training set into the trained Seq-VAE network, use the latent space features output by the Encoder layer as the input of the Sig-LDM network, and train the Sig-LDM network.

[0021] In some preferred embodiments, during the training of the generative network model, the loss function of the regression network is :

[0022] ;

[0023] Where T represents the number of frames of the signal sequence in the training set, Represents the feature sequence output by the Regression network; It represents the output sequence of the first adaptation signal linear layer obtained by inputting the simulated image sequence in the training set into the trained Seq-VAE, that is, the regression target of the Regression network; Represents the i-th sequence item in the feature sequence output by the Regression network; Represents the i-th sequence term in the regression target.

[0024] In some preferred embodiments, the generated network model is tested by:

[0025] The signal sequence in the test set is input into the trained Regression network for regression, and the regressed signal sequence is directly input into the GRUcell network of the trained Seq-VAE network; and the output of the Encoder layer of the Seq-VAE network is input into the trained Sig-LDM model for cyclic denoising, and after denoising, it is input into the Decoder layer and splicing layer of the trained Seq-VAE network to obtain a time series image sequence.

[0026] In some preferred embodiments, the cyclic denoising method is:

[0027] A. The forward noise addition module obtains the noise randomly sampled from the standard normal distribution according to the preset noise intensity coefficient sequence and the randomly sampled vector time value t and adds it to the output of the Encoder layer of the Seq-VAE network to obtain the latent space feature after noise addition. ;

[0028] B. Calculate and obtain time code based on the vector time value t;

[0029] C. The noise prediction module is based on the time coding and the Infer the noise at the current time step and output the predicted noise;

[0030] D. The reverse denoising module converts the Z t Subtract the prediction noise to obtain a new latent space feature ;

[0031] E. Let t = t-1;

[0032] F. If t=0, the reasoning is complete and the output is Otherwise, the Z t-1 As new Return to step B.

[0033] In some preferred embodiments, the MPI time series images are obtained by:

[0034] Inputting the MPI time series one-dimensional signal into the Regression network for regression to obtain a regressed MPI time series one-dimensional signal with unchanged feature dimension;

[0035] Skipping the first adaptation signal linear layer, inputting the regressed MPI time series one-dimensional signal into the GRUcell network of the Seq-VAE network, and outputting the latent space features after passing through the GRUcell network, the second adaptation signal linear layer, and the Encoder layer in sequence;

[0036] The latent space features are input into the trained Sig-LDM model for cyclic denoising, and the denoised latent space features are returned to the Decoder layer of the Seq-VAE network;

[0037] The decoder layer obtains T-frame MPI two-dimensional images, and the MPI two-dimensional images are processed by the splicing layer to obtain MPI time series images.

[0038] In some preferred embodiments, a functional connection diagram is constructed by:

[0039] Each brain region is defined as a seed region, and the voxel values ​​in each seed region are arranged in time series to obtain a voxel signal matrix;

[0040] defining a search cube centered on a specific voxel, also extracting a voxel signal matrix from the cube, and traversing each of the seed regions by moving the search cube voxel by voxel;

[0041] Based on the RV coefficient, the multivariate similarity between the voxel signal matrix falling into the search cube and the voxel signal matrix in the seed region is measured, and the functional connection map of the seed region represented by the RV value is obtained;

[0042] The RV coefficient is:

[0043] ;

[0044] in, and They are the n*p matrix from the seed region and the n*q matrix from the search cube, T represents the transpose of the matrix, and tr is the trace operator of the matrix;

[0045] After all seed regions are traversed, the functional connectivity maps of all seed regions are arithmetic averaged to obtain the functional connectivity map of the MPI time series image.

[0046] In some preferred embodiments, a functional connectivity matrix is ​​constructed by:

[0047] Each brain region is defined as a ROI, and the voxel values ​​of each ROI are arranged in time series to obtain a voxel signal matrix;

[0048] The RV coefficients of the voxel signal matrices of each ROI were calculated pairwise to obtain the functional connection matrix between the ROIs represented by the RV value;

[0049] The RV coefficient is:

[0050] ;

[0051] in, and They are n*p matrices from different ROIs, T represents the transpose of the matrix, and tr is the trace operator of the matrix.

[0052] In a second aspect of the present invention, an MPI magnetic particle imaging system for early detection of AD patients is proposed, the system comprising a signal acquisition device and a central processing device;

[0053] The signal acquisition device comprises an MPI magnetic particle imaging device; the signal acquisition device is configured to apply a tracer to the target to be modeled, set scanning parameters, and acquire an MPI one-dimensional voltage signal of the target to be modeled in a resting state based on the scanning parameters;

[0054] The central processing device includes a CPU and a GPU; the central processing device includes a model building module, a three-dimensional reconstruction module, a brain region division module, a saliency authentication module, and a brain network generation module;

[0055] A model building module, configured to build a generative network model and train it in stages;

[0056] A three-dimensional reconstruction module, configured to perform image reconstruction by generating a network model based on the MPI one-dimensional voltage signal to obtain a three-dimensional reconstructed image;

[0057] A brain region division module, configured to divide the brain regions based on the three-dimensional reconstructed image, thereby obtaining a time series of partition signals;

[0058] A significance authentication module is configured to construct a functional connection map or a functional connection matrix based on the partition signal time series using the RV coefficient, convert the RV value in the functional connection map or the functional connection matrix into a Z value, and perform statistical analysis on the Z value to obtain the significance of the target functional connection to be modeled

[0059] The brain network generation module is configured to draw a default mode network of the target to be modeled on the three-dimensional reconstructed image according to the Z value.

[0060] Beneficial effects of the present invention:

[0061] (1) The present invention constructs a generative network model, which is implemented by Seq-VAE / Regression / Sig-LDM step-by-step network. It freely inputs a signal sequence with a fixed feature dimension and reconstructs an MPI time series image, thereby solving the problem that the feature dimension of the desired input signal does not match the resolution of the expected generated image. It constructs a functional connectivity map or a functional connectivity matrix based on the RV coefficient to solve the functional connectivity information between brain regions, thereby avoiding the "salt and pepper noise" effect that often appears in the functional connectivity map.

[0062] (2) By setting up two different paths in the Seq-VAE network, the problem of the feature dimension of the desired input signal not matching the expected resolution of the generated image can be solved, and the dimensions can be aligned. In addition, it can ensure that all modules in path 2 used during inference have completed pre-training, thereby improving the model accuracy. If the feature dimension of the input signal changes when used for other tasks, it is only necessary to change the dimension hyperparameters in the first adaptive signal linear layer and the second adaptive signal linear layer in the Seq-VAE network, which has good scalability.

[0063] (3) The mapping relationship between the paired signal sequence in the training set (denoted as sequence S) and the signal sequence output by the first adaptive signal linear layer (denoted as sequence F) is established through the regression network, so that the information distribution of sequence S is transformed into an information distribution close to sequence F. The Seq-VAE network and Sig-LDM network have lower difficulty and better effect in identifying the temporal information of the distribution close to sequence F, extracting latent space features, denoising and reconstructing;

[0064] (4) The noise prediction model is based on Swin-Unet with time coding. Compared with the traditional denoising model based on convolution or replacing the U-Net structure with pure Transformer, its ability to predict noise is better. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0066] Figure 1 is a flow chart of the MPI magnetic particle imaging method for early detection of AD patients of the present invention;

[0067] Figure 2 is a schematic diagram of a step-by-step training process for generating a network model in an embodiment of the present invention;

[0068] Figure 3 is a flow chart of Sig-LDM network training in an embodiment of the present invention;

[0069] Figure 4 is a flow chart of Sig-LDM network reasoning in an embodiment of the present invention;

[0070] Figure 5 is a schematic diagram of the structure of a noise prediction model in an embodiment of the present invention;

[0071] Figure 6 Schematic diagram of a functional connection matrix in an embodiment of the present invention. DETAILED DESCRIPTION

[0072] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.

[0073] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0074] This application constructs a Seq-VAE / Regression / Sig-LDM step-by-step network to achieve free input of signal sequences with fixed feature dimensions and reconstruct MPI time series images, and solves the functional connection information between brain regions based on the RV coefficient to draw the default mode network of the target to be modeled; improves the accuracy and precision of three-dimensional reconstructed images, and the drawn brain network provides stronger technical support for subsequent research and prediction of AD diseases.

[0075] In order to more clearly explain the MPI magnetic particle imaging method for early detection of AD patients of the present invention, the following is combined with Figure 1 Each step in the embodiment of the present invention is described in detail.

[0076] The first embodiment of the present invention is an MPI magnetic particle imaging method for early detection of AD patients, which is used to reconstruct the MPI time series image of the target to be modeled, and then draw the default mode network, including steps S100 to S600, and each step is described in detail as follows:

[0077] Step S100, constructing a generation network model and training it in stages;

[0078] The generative network model is a Seq-VAE / Regression / Sig-LDM step-by-step network, and the specific network structure is as follows:

[0079] The Seq-VAE network includes a first adaptive signal linear layer, a GRU cell network, a second adaptive signal linear layer, an Encoder layer, a Decoder layer and a concatenation layer connected in series in sequence, wherein:

[0080] The first adaptation signal linear layer performs a linear transformation on each image in the input image sequence, and the number of feature dimensions of the image becomes the number of feature dimensions of the one-dimensional signal to be used in the test set; it is only used in the training process of the Seq-VAE network in the following step S120, that is, only the training set image sequence is received;

[0081] The GRUcell network is used to loop the length of the sequence to be processed (T) times, poll the input tensor sequence, receive the output of the previous time step as the input state of this time each time, and use the tensor polled this time as input to learn the time sequence information relationship therein; finally, save the output of this time step and submit it to the next time step; in this embodiment, the GRUcell network is initialized using a normal distribution;

[0082] The second adaptive signal linear layer changes the number of feature dimensions of each tensor in the sequence output by the GRU cell network back to the number of feature dimensions of the initial training set image;

[0083] The Encoder layer performs step-by-step feature learning and downsampling on each tensor in the sequence output by the second adaptive signal linear layer, and obtains the latent space feature Zt after feature concentration; the Decoder layer receives the latent space feature output by the Encoder layer in step S120, and step S120 receives the denoised latent space feature output by the Sig-LDM layer. The Encoder and Decoder adopt the Swin-Unet network structure, and realize feature learning, downsampling and upsampling through the Swin transformer Block (shift window transformer block), Patch Merging (patch merging) and Patch Expanding (patch expansion).

[0084] The Regression network is mainly composed of 4 internal linear layers. The first two linear layers multiply the input tensor feature dimensions, and the last two linear layers multiply the tensor feature dimensions, so that the feature dimensions of the final output tensor of the Regression network are consistent with the received input tensor feature dimensions. The linear layers are connected using nonlinear activation functions. The mapping relationship between the signal sequence in the input training set and the output of the first adaptation signal linear layer is learned and established through the Regression network, and the output of the Regression network is made as close to the output of the first adaptation signal linear layer as possible.

[0085] Preferably, the Sig-LDM network includes a forward noise addition module, a prediction noise module and a reverse noise removal module, and the specific structure is:

[0086] The forward denoising module does not contain any trainable parameters and is composed of a pre-set denoising function. The denoising function pre-sets a sequence of noise intensity coefficients that increases over time and performs denoising operations on the data according to this sequence. During the training process, the latent space features output by the encoder of the Seq-VAE are received. , and then randomly sample the vector time value t from the preset hyperparameter 0-T and combine it with the preset noise intensity coefficient sequence to obtain the noise randomly sampled from the standard normal distribution and After adding noise And time value t is input to the prediction noise module;

[0087] The prediction noise module is a prediction noise model Time-condition Swin-Unet, which is mainly composed of PatchPartition, Time-condition Linear Embedding, Time-condition Swin transformer Block, PatchMerging and Patch Expanding (data processing direction is as follows Figure 5 As shown by the arrow in the middle, a time coding design is introduced; the prediction noise module contains trainable parameters;

[0088] The reverse denoising module does not contain any trainable parameters and is composed of a pre-set denoising function to achieve the subtraction of the noise inferred by the predicted noise model from the tensor of the current time step and multiplication by the set coefficient; the reverse denoising module is not used in the following step S120.

[0089] In this embodiment, the method for generating the network model by staged training is as follows:

[0090] Step S110, preparing training set and test set data;

[0091] The simulation software is used to generate a simulation image sequence and a signal sequence as a training set, wherein the simulation image sequence and the signal sequence correspond to each other one by one; and the phantom signal data is collected in real time to form a test set.

[0092] Preferably, in this embodiment, in order to improve the model performance, the accuracy of the model during testing is further monitored and the test set is adjusted:

[0093] MATLAB was used to generate 10,000 time-series grayscale image sequences of three different models, namely, geometric model, letter model and resolution model. After obtaining the time-series grayscale image sequences, the MPIRF simulation software was used to generate corresponding one-dimensional signal sequences; a time dimension was added to all time-series binary image sequences, and the two-dimensional images were superimposed in the time dimension according to their corresponding sequence order to generate corresponding 10,000 three-dimensional time-series image sequences, and the one-dimensional signal sequence and the three-dimensional time-series image sequence were in one-to-one correspondence; 80% of the data of the 10,000 three-dimensional time-series image sequences and the corresponding one-dimensional signals were used as training sets, and 20% as test sets; the one-dimensional signal sequence in the training set was recorded as sequence S, the three-dimensional time-series image sequence was recorded as sequence I, the one-dimensional signal sequence in the test set was recorded as sequence S*, and the three-dimensional time-series image sequence was recorded as sequence I*; and 100 groups of MPI time-series one-dimensional signals collected in real time by MPI devices were added to the test set.

[0094] Further preferably, the data of the phantom is collected as the real-time MPI time-series one-dimensional signal collected in the above-mentioned model construction training stage, the magnetic nanoparticle reagent is injected into different phantom models made in advance, the phantom model is scanned using an MPI device, and 100 groups of signal sequence data of the phantom model are actually collected as another part of the test set data.

[0095] Step S120: Seq-VAE network, Regression network and Sig-LDM network are trained in three stages respectively. Figure 2 As shown, specifically:

[0096] Step S121, the simulated image sequence in the training set is used as the input of the Seq-VAE network, input into the first adaptation signal linear layer (path 1), and the Seq-VAE network is trained;

[0097] When training the Seq-VAE network, the output of the Encoder layer is The KL divergence between the latent space representation and the standard normal distribution and the frame average L1 loss between the reconstructed image sequence and the simulated image sequence are combined as the loss function of Seq-VAE :

[0098] ;

[0099] Where T represents the number of frames in the simulation image sequence, The i-th frame representing the input simulated image sequence data sample; represents the i-th frame of the reconstructed image sequence; represents the latent variable, which is the latent space distribution extracted by the Encoder layer from the i-th frame simulation image sequence; Represents the posterior distribution, indicating that given the input data Then, the latent variable The probability distribution of Decide; Represents the prior distribution, which is usually set to the standard normal distribution and is used to constrain the distribution of the latent variable; represents the KL divergence, It is used to measure the posterior distribution With prior distribution The average difference between the frames Represents L1 distance, which is used to measure the reconstructed image sequence The frame-average pixel-level difference between the simulated image sequence x; Represents the weight coefficient, which is used to balance the relative importance of the KL divergence term and the L1 reconstruction error term;

[0100] Step S122, the simulated image sequence in the training set is input into the trained Seq-VAE network to obtain the output tensor of the first adaptation signal linear layer, the output tensor is used as the regression target output by the Regression network, the signal sequence in the training set is used as the input of the Regression network, and the Regression network is trained;

[0101] When training the Regression network, the average MSE Loss of the sequence items between the regression target and the feature sequence output by the Regression network is used as the loss function of the Regression network. :

[0102] ;

[0103] Where T represents the number of frames of the signal sequence in the training set, Represents the feature sequence output by the Regression network; It represents the output sequence of the first adaptation signal linear layer obtained by inputting the simulated image sequence in the training set into the trained Seq-VAE, that is, the regression target of the Regression network; Represents the i-th sequence item in the feature sequence output by the Regression network; Represents the i-th sequence item in the regression target;

[0104] Step S123: The simulated image sequence in the training set is input into the trained Seq-VAE network, such as Figure 3 As shown, the latent space features output by the Encoder layer are Z t As the input of Sig-LDM network, train the Sig-LDM network;

[0105] When training the Sig-LDM network, the L1Loss between the predicted noise output by the Sig-LDM network and the actual added noise is used as the loss function of the Sig-LDM network:

[0106] ;

[0107] in, a latent space representation representing the latent space features; Represents random noise Add to The latent representation obtained in represents the real added noise; represents the prediction noise; It stands for L1 distance, which is used to measure the difference between the predicted noise and the real noise.

[0108] Step S130: Test and generate the network model, the method is as follows:

[0109] Step S131, the signal sequence in the test set is input into the trained Regression network for regression to obtain a regressed signal sequence with unchanged feature dimension (referred to as sequence O), sequence O is input into the trained Seq-VAE network, and the first adaptation signal linear layer is skipped to the GRU cell network (path 2);

[0110] Step S132: After obtaining the latent space features output by the Encoder layer, the latent space features are input into the trained Sig-LDM model for cyclic denoising, and the denoised latent space features are returned to the Decoder and concatenation layers of the Seq-VAE network for subsequent reconstruction steps;

[0111] Step S133, the Seq-VAE network outputs the calculation results to obtain a time-series image sequence.

[0112] Preferably, the cyclic denoising mentioned in the above method is as follows: Figure 4 As shown, the method is:

[0113] A. The noise adding function obtains the noise randomly sampled from the standard normal distribution according to the preset noise intensity coefficient sequence and the randomly sampled vector time value t and adds it to the input latent space feature to obtain the latent space feature after noise addition. ;

[0114] B. Calculate and obtain time code based on the vector time value t;

[0115] C. The noise prediction module is based on the latent space features after temporal coding and noise addition. Infer the noise at the current time step and output the predicted noise;

[0116] D. The reverse denoising module converts the latent space features after denoising Subtract the prediction noise to obtain a new latent space feature ;

[0117] E. Let t = t-1;

[0118] F. If t=0, the reasoning is completed and the latent space features are output Otherwise, the latent space features As new Return to step B.

[0119] Preferably, after obtaining the time-series image sequence, the time-series image sequence is verified:

[0120] The test set includes sequence S*, sequence I* and MPI time series one-dimensional signals. First, observe the similarity between the time series image sequence generated after the sequence S* is input and the sequence I*. If the RMSE error is less than the preset threshold and the SSIM and PSNR indicators are higher than the preset threshold, the accuracy meets the requirements.

[0121] Then observe the time series image sequence effect generated after the MPI time series one-dimensional signal is input. If the generated image is clear and has clear boundaries, the generated effect meets the requirements;

[0122] Otherwise, return to step S120 to perform training again.

[0123] During the model training phase, only if the MPI time-series image sequence generated by the network model output passes the verification of other indicators such as sensitivity and no problems are found during the experiment, can the subsequent acquisition of the MPI time-series one-dimensional signal of the resting state of the target to be modeled be carried out.

[0124] Preferably, in this embodiment, the Pytorch 2.5.1, cuda 12.1 framework is used to build the Seq-VAE / Regression / Sig-LDM network, and four NVIDIA GeForce RTX 4090s are used to train it. The training process is optimized using the Adam algorithm, the Seq-VAE learning rate is set to 0.00001, the Regression network learning rate is set to 0.00002, and the Sig-LDM learning rate is set to 0.00001.

[0125] The Seq-VAE / Regression / Sig-LDM network of the present invention can solve the problem that the characteristic dimension of the signal to be input does not match the expected resolution of the generated image by setting two different paths in Seq-VAE (i.e., whether to pass through the first adaptive signal linear layer, using path 1 during training and path 2 during reasoning), and can ensure that all modules in path 2 used during reasoning have completed pre-training (path 2 is a subset of path 1). In addition, the design of this solution also has good scalability. If the characteristic dimension of the signal to be input changes when this solution is used for other tasks, it is only necessary to change the dimension hyperparameters in the first adaptive signal linear layer and the second adaptive signal linear layer in Seq-VAE.

[0126] The paired signal sequence data (denoted as sequence S) and image sequence data (denoted as sequence I) in the training set correspond to the same magnetic nanoparticle distribution during simulation generation, and are different representation forms of the same magnetic nanoparticle distribution, and there is a mapping relationship between the two. The present invention uses a regression network to establish a mapping relationship between sequence S and the output signal of the first adaptive signal linear layer (denoted as sequence F) through training. This mapping relationship can transform the distribution of sequence S in the training set into a distribution close to sequence F. Since the time series information recognition, extraction of latent space features and reconstruction after the first adaptive signal linear layer are all based on sequence F when training Seq-VAE in the first stage, similarly, when training Sig-LDM in the third stage, its prediction noise is also based on the latent space features obtained by transforming sequence F. Therefore, during inference, the sequence S* is first input into the trained Regression network to obtain the regressed signal sequence O with unchanged feature dimension, and the distribution information alignment between sequence S* and sequence F is completed through sequence O; then the sequence O is input into path 2 of the trained Seq-VAE model and subsequent denoising is performed. The difficulty of Seq-VAE and Sig-LDM models in identifying the temporal information of the distribution (sequence O) close to sequence F, extracting latent space features, denoising and reconstruction will be significantly reduced, and the effect will be better, so that the subsequent reconstruction and inference effects will also be better, and a higher quality MPI temporal image sequence can be obtained than GAN or other traditional reconstruction methods.

[0127] Step S200, the MPI device collects the MPI time series one-dimensional signal of the target to be modeled in a resting state;

[0128] In this embodiment, the scanning parameters are set, the excitation magnetic field gradient is 1.5T, the excitation field frequency is 25KHz, and the magnetic field gradient is set as follows: Z direction: 1T / m, X direction: 2T / m, Y direction: 1T / m.

[0129] Compared with the EEG and MEG technologies in the prior art, the MPI imaging device does not have the volume conduction effect. The MPI device is used to obtain the resting state brain network map of the target to be modeled based on MPI. Because of the above advantages, the specific location of the magnetic particle signal generated in the brain can be clearly located.

[0130] Step S300: input the MPI time series one-dimensional signal into the trained generative network model for image reconstruction to obtain an MPI time series image, wherein the method is as follows:

[0131] Inputting the MPI time series one-dimensional signal into the Regression network for regression to obtain a regressed MPI time series one-dimensional signal with unchanged feature dimension;

[0132] The first adaptive signal linear layer is skipped to input the regressed MPI time series one-dimensional signal into the GRU cell network of the Seq-VAE network, and the latent space features are output after passing through the GRU cell network, the second adaptive signal linear layer, and the Encoder layer in sequence;

[0133] The latent space features are input into the trained Sig-LDM model for cyclic denoising, and the denoised latent space features are returned to the Decoder layer of the Seq-VAE network;

[0134] After being processed by the Decoder layer, a T-frame MPI two-dimensional image is obtained, and after being processed by the splicing layer, the MPI two-dimensional image is obtained as an MPI time series image.

[0135] Step S400: preprocess the MPI time series images, and divide the preprocessed MPI time series images into brain regions to obtain partition time series signals.

[0136] Preferably, the preprocessing includes operations such as position correction, registration, smoothing, filtering, etc.; the registered MPI time series images are divided into brain regions based on the AAL human brain standard template.

[0137] Step S500: construct a functional connectivity map and / or a functional connectivity matrix based on the partitioned time series signal using the RV coefficient.

[0138] Preferably, the RV coefficient is used to construct a functional connectivity map, and the method is:

[0139] Each brain region is defined as a seed region, and the voxel values ​​in each seed region are arranged in time series to obtain a voxel signal matrix;

[0140] defining a search cube centered on a specific voxel, also extracting a voxel signal matrix from the cube, and traversing each of the seed regions by moving the search cube voxel by voxel;

[0141] The multivariate similarity between the voxel signal matrix falling into the search cube and the voxel signal matrix in the seed region is measured based on the RV coefficient, and the functional connection map of the seed region represented by the RV value is obtained;

[0142] The RV coefficient is:

[0143] ;

[0144] in, and are the n*p matrix from the seed region and the n*q matrix from the search cube, T represents the transpose of the matrix, and tr is the trace operator of the matrix;

[0145] After all seed regions are traversed, the functional connectivity maps of all seed regions are arithmetic averaged to obtain the functional connectivity map of the MPI time series image.

[0146] Furthermore, the voxel arrangement method is:

[0147] For each seed region, it is divided into several voxel values ​​of 2mm*2mm*2mm, and each voxel value is arranged in time series to obtain an n*p voxel signal matrix, where n represents that there are n voxels in the seed region, and p represents that each voxel signal has p features over time.

[0148] Furthermore, the search cube contains a plurality of adjacent voxels of specific size and shape, and an n*q voxel signal matrix can be obtained, where n represents that there are n voxels in the cube, and q represents that each voxel signal has q features over time.

[0149] Preferably, the RV coefficient is used to construct a functional connectivity matrix, and the method is:

[0150] Each brain region is defined as ROI. For each ROI, it is divided into several voxel values ​​of 2mm*2mm*2mm. Each voxel value is arranged in time series to obtain an n*p voxel signal matrix, where n represents that there are n voxels in the seed region, and p represents that each voxel signal has p features.

[0151] The RV coefficients are calculated for each voxel signal matrix of each ROI, and the RV coefficients are:

[0152] ;

[0153] in, and They are n*p matrices from different ROIs, T represents the transpose of the matrix, and tr is the trace operator of the matrix.

[0154] Get the functional connection matrix between ROIs represented by RV values ​​(such as Figure 6 The matrix reflects the functional connectivity of different brain regions in the default mode network. The larger the value, the stronger the functional connectivity.

[0155] Based on comprehensive measurement and analysis of seed region (voxel level) correlation and ROI level analysis, the brain network map and functional connection matrix can be calculated more comprehensively and precisely, thereby achieving more comprehensive and accurate brain network analysis and detection of early AD patients in subsequent use, reducing the possibility of misdiagnosis.

[0156] Step S600: convert the RV values ​​in the functional connectivity map and / or the functional connectivity matrix into Z values, perform statistical analysis on the Z values ​​to obtain the significance of the functional connections of the target to be modeled, and draw the default mode network of the target to be modeled.

[0157] Preferably, a significance test is performed on the whole-brain functional connectivity map\functional connectivity matrix, and the functional connectivity map\functional connectivity matrix is ​​converted into a normal distribution map by using the following formula:

[0158] ;

[0159] This Z statistic is used to approximate the functional connectivity map and test its significance, and then the brain network in the resting state of the target to be modeled is drawn.

[0160] Although the various steps in the above embodiment are described in the above-mentioned order, those skilled in the art can understand that in order to achieve the effect of this embodiment, different steps do not have to be executed in such an order. They can be executed simultaneously (in parallel) or in a reverse order. These simple changes are within the scope of protection of the present invention.

[0161] The MPI magnetic particle imaging system for early detection of AD patients according to the second embodiment of the present invention comprises a signal acquisition device and a central processing device;

[0162] The signal acquisition device comprises an MPI magnetic particle imaging device; the signal acquisition device is configured to apply a tracer to the target to be modeled, set scanning parameters, and acquire an MPI one-dimensional voltage signal of the target to be modeled in a resting state based on the scanning parameters;

[0163] The central processing device includes a CPU and a GPU; the central processing device includes a model building module, a three-dimensional reconstruction module, a brain region division module, a saliency authentication module, and a brain network generation module;

[0164] The model building module is configured to build a generation network model and train it in stages;

[0165] The three-dimensional reconstruction module is configured to perform image reconstruction by generating a network model based on the MPI one-dimensional voltage signal to obtain a three-dimensional reconstructed image;

[0166] The brain region division module is configured to divide the brain regions based on the three-dimensional reconstructed image, thereby obtaining a time series of partition signals;

[0167] The significance authentication module is configured to construct a functional connection map or a functional connection matrix based on the partition signal time series using the RV coefficient, convert the RV value in the functional connection map or the functional connection matrix into a Z value, and perform statistical analysis on the Z value to obtain the significance of the target functional connection to be modeled;

[0168] The brain network generation module is configured to draw a default mode network of the target to be modeled on the three-dimensional reconstructed image according to the Z value.

[0169] Furthermore, the default mode network and functional connection matrix of the target to be modeled are compared and analyzed with the DMN brain network and DMN internal functional connection matrix of the normal aging population in the database. And by analyzing the whole-brain functional connection map of the target to be modeled obtained in the above steps, the connection between the DMN brain network of the target to be modeled and the other brain networks is calculated and compared with the connection between the DMN brain network of the normal aging population and the other brain networks.

[0170] The analysis method is as follows:

[0171] The GRETNA software was used to calculate and analyze the DMN brain network and whole-brain functional connectivity map of the modeling target, and the following three network parameters were used to describe the global properties of the DMN brain network and the whole-brain functional connectivity map: global efficiency, clustering coefficient, and characteristic path length. Compared with the DMN brain network of the normal aging population, the global efficiency, clustering coefficient, and characteristic path length of the DMN brain network of AD patients were significantly reduced. In addition, compared with the normal aging population, the connection between the DMN brain network of AD patients and the rest of the brain networks was significantly abnormal, which was reflected in the RV value or Z value of the whole-brain functional connectivity map, showing that the values ​​of specific areas were significantly different from those of the normal aging population. For the DMN internal functional connection matrix of the target to be modeled, if the Z value or RV coefficient between typical areas such as the hippocampus and parahippocampal gyrus, between the cingulate cortex and the precuneus, between the parahippocampal gyrus and the cingulate cortex, between the hippocampus and the cingulate cortex, or between other areas is significantly reduced, then the analysis results of the DMN brain network, the whole-brain functional connection map and the DMN internal functional connection matrix can be combined to judge that the target to be modeled has AD symptoms, thereby realizing the detection of early AD patients, and then realizing clinical intervention for AD patients in the early stage to slow down the progression of their disease, which also lays the foundation for the future use of fMPI in the diagnosis of Alzheimer's disease or other related diseases.

[0172] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0173] It should be noted that the MPI magnetic particle imaging system for early detection of AD patients provided in the above embodiment is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be decomposed or combined. For example, the modules in the above embodiment can be combined into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the modules or steps, and are not regarded as improper limitations of the present invention.

[0174] An electronic device according to a third embodiment of the present invention includes:

[0175] at least one processor; and

[0176] a memory communicatively connected to at least one of the processors; wherein,

[0177] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the above-mentioned MPI magnetic particle imaging method for early detection of AD patients.

[0178] A fourth embodiment of the present invention is a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to be executed by the computer to implement the above-mentioned MPI magnetic particle imaging method for early detection of AD patients.

[0179] Technicians in the relevant technical field can clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the electronic device and computer-readable storage medium described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0180] Those skilled in the art should be able to appreciate that the modules and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented with electronic hardware, computer software, or a combination of the two, and the programs corresponding to the software modules and method steps can be placed in random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the technical field. In order to clearly illustrate the interchangeability of electronic hardware and software, the composition and steps of each example have been generally described in the above description according to the function. Whether these functions are performed in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0181] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0182] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square 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 square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0183] The terms "first", "second", etc. are used to distinguish similar objects rather than to describe or indicate a particular order or sequence.

[0184] The term "comprise" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that includes a list of elements includes not only those elements but also other elements not expressly listed, or also includes elements inherent to such process, method, article, or apparatus / device.

[0185] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

Claims

1. An MPI magnetic particle imaging method for early detection of AD patients, used to reconstruct the MPI time series images of the target to be modeled, and then draw the default mode network, characterized in that: The following steps are involved: Step S100: constructing a generative network model and training the generative network model; Step S200, the MPI device collects the MPI time series one-dimensional signal of the target to be modeled in a resting state; Step S300, inputting the MPI time series one-dimensional signal into the trained generative network model for image reconstruction to obtain an MPI time series image; Step S400, preprocessing the MPI time series images, and dividing the preprocessed MPI time series images into brain regions to obtain partition time series signals; Step S500, constructing a functional connectivity map and / or a functional connectivity matrix based on the partitioned time series signal using the RV coefficient; Step S600, converting the RV values ​​in the functional connectivity map and / or the functional connectivity matrix into Z values, performing statistical analysis on each Z value to obtain the significance of the functional connectivity of the target to be modeled, and drawing the default mode network of the target to be modeled; The generative network model is a Seq-VAE / Regression / Sig-LDM step-by-step network; the Seq-VAE network includes a first adaptive signal linear layer, a GRU cell network, a second adaptive signal linear layer, an Encoder layer, a Decoder layer and a splicing layer connected in series in sequence; the Sig-LDM network includes a forward noise addition module, a prediction noise module and a reverse denoising module; The generative network model is trained in three stages, and the method is as follows: Step S121, using the simulated image sequence in the training set as the input of the Seq-VAE network to train the Seq-VAE network; Step S122, inputting the simulated image sequence in the training set into the first adapted signal linear layer of the trained Seq-VAE network, taking the output tensor of the first adapted signal linear layer as the regression target, taking the signal sequence in the training set as input, and training the Regression network; Step S123: input the simulated image sequence in the training set into the trained Seq-VAE network, use the latent space features output by the Encoder layer as the input of the Sig-LDM network, and train the Sig-LDM network.

2. The MPI magnetic particle imaging method for early detection of AD patients according to claim 1, characterized in that: When training the generative network model, simulation software is used to generate a three-dimensional simulation image sequence and a one-dimensional signal sequence as a training set, wherein the simulation image sequence and the signal sequence correspond one to one; and a test set is formed by real-time acquisition of a pre-made MPI time-series one-dimensional signal of the phantom.

3. The MPI magnetic particle imaging method for early detection of AD patients according to claim 1, characterized in that: During the training of the generative network model, the loss function of the regression network is for: ； Where T represents the number of frames of the signal sequence in the training set, Represents the feature sequence output by the Regression network; Represents the output sequence of the first adaptation signal linear layer obtained by inputting the simulated image sequence in the training set into the trained Seq-VAE network, that is, the regression target of the Regression network; Represents the i-th sequence item in the feature sequence output by the Regression network; Represents the i-th sequence term in the regression target.

4. The MPI magnetic particle imaging method for early detection of AD patients according to claim 2, characterized in that: The generated network model is tested by: The signal sequence in the test set is input into the trained Regression network for regression, and the regressed signal sequence is directly input into the GRUcell network of the trained Seq-VAE network; and the output of the Encoder layer of the Seq-VAE network is input into the trained Sig-LDM model for cyclic denoising, and after denoising, it is input into the Decoder layer and splicing layer of the trained Seq-VAE network to obtain a time series image sequence.

5. The MPI magnetic particle imaging method for early detection of AD patients according to claim 4, characterized in that: The method of the cyclic denoising is as follows: A. The forward noise addition module obtains the noise randomly sampled from the standard normal distribution according to the preset noise intensity coefficient sequence and the randomly sampled vector time value t and adds it to the output of the Encoder layer of the Seq-VAE network to obtain the latent space feature after noise addition. ; B. Calculate and obtain time code based on the vector time value t; C. The noise prediction module is based on the time coding and the Infer the noise at the current time step and output the predicted noise; D. The reverse denoising module converts the Subtract the prediction noise to obtain a new latent space feature ; E. Let t = t-1; F. If t=0, the reasoning is complete and the output is Otherwise, the As new Return to step B.

6. The MPI magnetic particle imaging method for early detection of AD patients according to claim 1, characterized in that: The method to obtain the MPI time series image is as follows: Inputting the MPI time series one-dimensional signal into the Regression network for regression to obtain a regressed MPI time series one-dimensional signal with unchanged feature dimension; Skipping the first adaptation signal linear layer, inputting the regressed MPI time series one-dimensional signal into the GRUcell network of the Seq-VAE network, and outputting the latent space features after passing through the GRUcell network, the second adaptation signal linear layer, and the Encoder layer in sequence; The latent space features are input into the trained Sig-LDM model for cyclic denoising, and the denoised latent space features are returned to the Decoder layer of the Seq-VAE network; The decoder layer obtains T-frame MPI two-dimensional images, and the MPI two-dimensional images are processed by the splicing layer to obtain MPI time series images.

7. The MPI magnetic particle imaging method for early detection of AD patients according to claim 1, characterized in that: Construct a functional connection map using: Each brain region is defined as a seed region, and the voxel values ​​in each seed region are arranged in time series to obtain a voxel signal matrix; defining a search cube centered at a specific voxel, also extracting a voxel signal matrix from the cube, and traversing each of the seed regions by moving the search cube voxel by voxel; The multivariate similarity between the voxel signal matrix falling into the search cube and the voxel signal matrix in the seed region is measured based on the RV coefficient, and the functional connection map of the seed region represented by the RV value is obtained; The RV coefficient is: ； in, and They are the n*p matrix from the seed region and the n*q matrix from the search cube, T represents the transpose of the matrix, is the trace operator of the matrix; After all seed regions are traversed, the functional connectivity maps of all seed regions are arithmetic averaged to obtain the functional connectivity map based on the MPI time series image.

8. The MPI magnetic particle imaging method for early detection of AD patients according to claim 1, characterized in that: Construct the functional connectivity matrix as follows: Each brain region is defined as a ROI, and the voxel values ​​of each ROI are arranged in time series to obtain a voxel signal matrix; The RV coefficients of the voxel signal matrices of each ROI were calculated pairwise to obtain the functional connection matrix between the ROIs represented by the RV value; The RV coefficient is: ； in, and They are n*p matrices from different ROIs, T represents the transpose of the matrix. is the trace operator of the matrix.

9. An MPI magnetic particle imaging system for early detection of AD patients, according to the MPI magnetic particle imaging method for early detection of AD patients according to any one of claims 1 to 8, characterized in that: The system includes a signal acquisition device and a central processing device; The signal acquisition device comprises an MPI magnetic particle imaging device; the signal acquisition device is configured to apply a tracer to the target to be modeled, set scanning parameters, and acquire an MPI one-dimensional voltage signal of the target to be modeled in a resting state based on the scanning parameters; The central processing device includes a CPU and a GPU; the central processing device includes a model building module, a three-dimensional reconstruction module, a brain region division module, a saliency authentication module, and a brain network generation module; The model building module is configured to build a generation network model and train it in stages; The three-dimensional reconstruction module is configured to perform image reconstruction by generating a network model based on the MPI one-dimensional voltage signal to obtain a three-dimensional reconstructed image; The brain region division module is configured to divide the brain regions based on the three-dimensional reconstructed image, thereby obtaining a time series of partition signals; The significance authentication module is configured to construct a functional connection map or a functional connection matrix based on the partition signal time series using the RV coefficient, convert the RV value in the functional connection map or the functional connection matrix into a Z value, and perform statistical analysis on the Z value to obtain the significance of the target functional connection to be modeled The brain network generation module is configured to draw a default mode network of the target to be modeled on the three-dimensional reconstructed image according to the Z value.

Images