MPI magnetic particle imaging method and system for stage detection of PD patient

Abstract

The invention belongs to the technical field of magnetic nanoparticle imaging, particularly relates to an MPI magnetic particle imaging method and system for stage detection of a PD patient, and aims to solve the problem that the traditional reconstruction technology is difficult to meet the requirements of precise imaging and precise disease stage determination of the PD patient. The method comprises the following steps: constructing and training a generative network model; the tracer agent is subjected to surface modification, and the modified tracer agent is applied; mPI time sequence one-dimensional signals are collected, and image reconstruction is carried out based on the generative network model to obtain MPI time sequence images; dividing a brain region to obtain an average voxel signal; obtaining a root mean square value and extracting an STFT spectrogram; wherein the generated network model is a GRU-3D Vision Mama network, and the generated network model is a GRU-3D Vision Mama network. According to the method, the image sequence is generated by adopting an end-to-end method, and global-local-potential multi-level optimization is formed without generating two-dimensional image transition and network structure multi-layer collaborative arrangement, so that accurate data is provided for subsequent application.

Description

Background Art

[0002] Parkinson's disease (PD), as the second most common neurodegenerative disease of the central nervous system, is mainly characterized pathologically by the selective loss of dopaminergic neurons in the substantia nigra pars compacta of the midbrain. Iron deposition in the substantia nigra is a key factor leading to the degeneration of dopaminergic neurons. Under normal physiological conditions, iron gradually deposits in brain regions such as the substantia nigra, caudate nucleus, and globus pallidus with age, but this deposition phenomenon is more severe in PD patients. According to autopsy results of PD patients, the iron content in the substantia nigra increases by more than 30% on average, and the change in substantia nigra iron content is stage-dependent and changes with the development of PD symptoms.

[0003] With the development of modern imaging technologies, clinical experiments and diagnoses are often carried out using techniques such as MRI, PET, and CT to explore the pathogenesis and disease progression of PD. However, clinically, the CT and MRI images of the brains of Parkinson's disease patients are often indistinguishable from those of normal people, and it is difficult to accurately identify iron deposition in the substantia nigra, making it difficult to use these two techniques for accurate diagnosis of PD. Many PD patients are already in the late pathological stage when they are diagnosed. Although the PET technique has high contrast, the injected tracer is radioactive and has low spatial resolution, which limits its application in clinical diagnosis.

[0004] Magnetic particle imaging (MPI), as an emerging imaging modality, can be used to reconstruct the concentration distribution of superparamagnetic nanoparticles in the object to be measured. This technique usually uses superparamagnetic iron oxide nanoparticles (SPION) as tracers and can modify the surface of the tracers according to different imaging tasks, with strong flexibility. MPI devices are not interfered by the body's own background signals, produce images with high contrast and good sensitivity, and the signal intensity is proportional to the tracer concentration, enabling the acquisition of quantitative data, showing great potential in fields such as stem cell tracking, angiography, and targeted drug delivery. Traditional MPI reconstruction techniques include methods that rely on the pre-characterization of the SPION signal response by the system matrix and the X-space algorithm. However, there are also certain limitations. Compared with the advanced deep learning methods, traditional MPI reconstruction techniques are limited by the computational amount and imaging effect and are not suitable for imaging the pathological regions of PD patients. When using deep learning methods to reconstruct MPI image sequences, common generative models such as GANs have problems such as difficult training fitting and insufficient diversity of generated data; diffusion models may have a mismatch between the input signal and the semantic information of the output reconstructed image; if based on the Transformer architecture, there are also common problems such as slow training speed and large memory consumption.

[0005] In summary, the existing MPI reconstruction techniques are difficult to meet the requirements of accurately and sensitively imaging the changes in iron deposition in the pathological regions of PD patients and accurately diagnosing the disease stage. Summary of the Invention

[0006] To solve the above problems in the prior art, that is, the traditional MPI reconstruction technology is difficult to meet the requirements of accurate, sensitive imaging of iron deposition changes in the pathological regions of PD patients and accurate diagnosis of disease stages, in the first aspect of the present invention, an MPI magnetic particle imaging method for staging detection of PD patients is proposed, which combines MPI technology and the Mamba model in deep learning to achieve accurate and sensitive temporal imaging of the pathological regions of PD patients. The method includes the following steps: Step S100: Construct a generative network model and train it; Step S200: Modify the surface of the tracer and apply the modified tracer to the target to be modeled; Step S300: Collect the MPI temporal one-dimensional signals of the target to be modeled at rest, input the MPI temporal one-dimensional signals into the trained generative network model for image reconstruction, and obtain MPI time series images; Step S400: Preprocess the MPI time series images, divide the preprocessed MPI time series images into brain regions, and extract the average voxel signals; Step S500: Calculate the root mean square value and the STFT spectrogram based on the average voxel signals; Wherein, the generative network model is a GRU-3D Vision Mamba network, which includes a bidirectional double-layer GRU network and a 3D Vision Mamba network connected in sequence. The 3D Vision Mamba network is composed of L 3D-Vim Blocks and L-1 upsampling blocks arranged alternately. A dimension rearrangement layer is provided in front of the first 3D-Vim Block for dimension rearrangement of the signal sequence output by the bidirectional double-layer GRU network; The 3D-Vim Block divides the dimension-rearranged signal sequence into multiple paths for processing after passing through a normalization layer; wherein, one path outputs forward data and backward data, and the remaining paths generate multiple derivative data and perform specific association operations with the forward data / backward data respectively to obtain the output result; the path refers to a logical channel for performing specific processing on the input data and outputting the result.

[0007] In some preferred embodiments, the number of the paths is four, including a first path, a second path, a third path, and a fourth path; Among them, the first path processes the input data and outputs the forward data and the backward data; the second path, the third path, and the fourth path perform internal processing on their respective input data in different ways, and sum the respective output data after gating operations to obtain a first intermediate result; the first intermediate result performs gating operations with the forward data and the backward data respectively to obtain a second intermediate result and a third intermediate result; the second intermediate result and the third intermediate result are added to obtain an output tensor, and the output tensor is fused with the input tensor of the 3D-Vim Block normalization layer after passing through a linear layer as the output of the 3D-Vim Block.

[0008] In some preferred embodiments, the first path is composed of a forward convolution, a backward convolution, a forward state space model, and a backward state space model, and the second path is composed of an activation layer; the third path is composed of an S-T Cross Attention network; the fourth path is composed of an S-T 3D VAE network.

[0009] In some preferred embodiments, the S-T Cross Attention network includes a spatial attention branch and a temporal attention branch. The input tensor sequence is decoupled in space and time and then undergoes attention learning and latent space feature extraction through the spatial attention branch and the temporal attention branch respectively, and spatial and temporal feature fusion is performed after the learning is completed.

[0010] In some preferred embodiments, the S-T Cross Attention network processes the input temporal feature tensor, and the method is as follows: In the spatial attention branch, the temporal feature tensor is used to obtain a query matrix, a key matrix, and a value matrix through 3D convolution, the spatial dimension is flattened, so that the attention calculation is limited to the spatial scale independent of each time step, and the spatial dimension is restored after the calculation is completed; In the temporal attention branch, the temporal feature tensor is used to obtain a query matrix, a key matrix, and a value matrix through 3D convolution, the temporal dimension is flattened, so that the attention calculation is limited to the temporal scale independent of each spatial position, and the temporal dimension is restored after the calculation is completed; Fuse the temporal feature tensors processed by the spatial attention branch and the temporal attention branch to obtain a fused temporal feature tensor; The fused temporal feature tensor is fused with the input temporal feature tensor after passing through a linear layer as the output of the S-T Cross Attention network.

[0011] In some preferred embodiments, the S-T 3D VAE network includes a spatial encoder, a temporal encoder, a temporal decoder, and a spatial decoder; The S-T 3D VAE network processes the input temporal feature tensor, and the method is as follows: Perform spatial scale downsampling and temporal scale downsampling (spatiotemporal separation downsampling) on the input temporal feature tensor to obtain the extracted latent space features, and then perform temporal scale upsampling and spatial scale upsampling (spatiotemporal separation upsampling) on the latent space features to obtain the output tensor of the S-T 3D VAE.

[0012] In the second aspect of the present invention, an MPI magnetic particle imaging system for staging detection of PD patients is proposed, and the system includes a signal acquisition device and a central processing device; The signal acquisition device includes an MPI magnetic particle imaging device; the signal acquisition device is configured to administer a modified tracer to the target to be modeled, set scanning parameters, and obtain the MPI one-dimensional voltage signal of the target to be modeled in the resting state based on the scanning parameters; The central processing device includes a CPU and a GPU; the central processing device includes a model construction module, a three-dimensional reconstruction module, a brain region division module, and a spectrogram generation module; The model construction module is configured to construct and train a generative network model; The three-dimensional reconstruction module is configured to perform image reconstruction based on the MPI one-dimensional voltage signal through the generative network model to obtain an MPI time series image; The brain region division module is configured to divide brain regions based on the MPI time series image to obtain average voxel signals; The spectrogram generation module is configured to calculate the root mean square value RMS based on the average voxel signal, and further draw an STFT spectrogram.

[0013] Advantages of the present invention: (1) The temporal reconstruction network combines the GRU, a type of RNN temporal model, and the 3D Vision Mamba, a type of SSM sequence model, which further improves the model's learning ability for temporal sequence information. At the same time, it combines the upsampling operation and innovatively uses it in the field of image sequence generation in an end-to-end manner; only through a simple dimension rearrangement mechanism, it can be aligned with the parameterization idea in Mamba, and directly and quickly generate MPI time series images without going through the transition of generating two-dimensional images; (2) By surface modification of the existing SPION tracer, its stability is increased and the tendency of immune response is reduced, enabling the SPION tracer to penetrate the blood-brain barrier and enter specific cells to play a role, providing a method for MPI technology to image the substantia nigra region of PD patients, and enabling accurate and sensitive imaging of the iron deposition changes in the pathological regions of PD patients; (3) By means of the improved 3D Vision Mamba network, while learning the internal relationships (spatial relationships) of single-frame elements in the tensor sequence, it also learns the relationships between elements of different frames in the tensor sequence, can efficiently capture long-range temporal dependencies between video frames and short-range spatial dependencies within frames, and the resulting state space and feature tensors have spatio-temporal consistency, which can improve the generation effect of this solution for the scenario; (4) Different branches are set in the 3D Vision Mamba network. Different branches are responsible for processing tensor information with the same source in different ways, and perform gating operations and combination on the results of each branch. The final result can effectively absorb the information provided by each branch. And if there are situations such as learning deviation or fitting difficulties in some branches during the training process, the gating operation and combination operation can effectively ensure that the results of the remaining branches correct and limit the results of the problematic branches, thereby increasing the robustness during the model training process. This closed-loop collaboration enables the model to not only process long videos with the linear computational complexity of 3D Vision Mamba, but also enhance spatial detail perception through the attention mechanism, and at the same time improve the generation controllability by means of the latent space normalization of S-T 3D VAE. Finally, a multi-level optimization of "global-local-latent" is formed in terms of computational efficiency, feature discriminability and generation quality, thereby breaking through the performance bottleneck of single or dual-module combinations and significantly improving the quality of the reconstructed temporal image sequence. Description of the Drawings

[0014] Other features, objectives and advantages of this application will become more obvious by reading the detailed description of the non-restrictive embodiments with reference to the following drawings: Figure 1 is a flowchart of the MPI magnetic particle imaging method for staging detection of PD patients in the present invention; Figure 2 is a schematic structural diagram of the GRU-3D Vision Mamba network in the embodiments of the present invention; Figure 3 is a schematic structural diagram of the 3D-Vim Block in the embodiments of the present invention; Figure 4 is a schematic structural diagram of the S-T Cross Attention network in the embodiments of the present invention; Figure 5 is a schematic structural diagram of the S-T 3D VAE in the embodiments of the present invention. Detailed Embodiments

[0001] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the related invention, rather than limiting the invention. Additionally, it should be noted that for ease of description, only parts related to the relevant invention are shown in the drawings.

[0002] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and embodiments.

[0003] The present application combines a class of RNN time series model and a class of SSM sequence model, and through a simple dimension rearrangement mechanism, realizes the alignment of parametric ideas, directly and quickly generates MPI time series images and STFT spectrograms without going through the transition of generating two-dimensional images, providing stronger technical support for subsequent research and prediction of PD diseases.

[0004] To more clearly illustrate the MPI magnetic particle imaging method of the present invention for PD patient staging detection, the following combines Figures 1-5 to elaborate on each step in the embodiments of the present invention.

[0005] A method for MPI magnetic particle imaging for PD patient staging detection according to the first embodiment of the present invention is used to reconstruct the MPI time series image of the target to be modeled, and then draw the STFT spectrogram, including steps S100 - step S600. Each step is described in detail as follows:

[0006] Step S100: Construct a generative network model and train it; The generative network model is a GRU-3D Vision Mamba network, including a bidirectional double-layer GRU network and a 3D Vision Mamba network connected in sequence. The 3D Vision Mamba network is composed of L 3D-Vim Blocks and L - 1 upsampling blocks arranged alternately. A dimension rearrangement layer is provided before the first 3D-Vim Block to perform dimension rearrangement on the signal sequence output by the bidirectional double-layer GRU network; The 3D-Vim Block is divided into multiple paths after passing through a normalization layer; among them, one path is used to process the output data, output forward data and backward data, and the remaining paths generate multiple derivative data and perform specific association operations with the forward data / backward data respectively to obtain the output result; the path refers to a logical channel for performing specific processing on the input data and outputting the result.

[0007] Preferably, a linear layer is provided before the bidirectional double-layer GRU network to adjust the feature dimension of the input tensor to adapt to the subsequent network and enhance the generalization ability of the model.

[0008] Inside the bidirectional double-layer GRU network, it will perform bidirectional cyclic operations for the length (T) of the sequence to be processed. The forward network polls the one-dimensional signal sequence input in the forward order, while the reverse network polls it in the reverse order. Each time, it receives the state information output from the previous time step as the input state for this time step, and takes the tensor polled this time as the input to learn the temporal information relationship. Finally, it saves the output of this time step and submits the state information of this time step to the next time step. The bidirectional double-layer GRU network is initialized using a normal distribution.

[0009] Preferably, there are four paths. The first path outputs the forward data and the backward data; the second path, the third path, and the fourth path perform internal processing on their respective input data in different ways, and sum the output data after gating operations to obtain a first intermediate result; the first intermediate result performs gating operations with the forward data and the backward data respectively to obtain a second intermediate result and a third intermediate result; the second intermediate result and the third intermediate result are added to obtain an output tensor, and the output tensor passes through a linear layer and is fused with the input tensor of the 3D-Vim Block normalization layer as the output of the 3D-Vim Block.

[0010] The 3D Vision Mamba network is composed of L 3D-Vim Blocks and L - 1 upsampling blocks arranged alternately. The first 3D-Vim Block first rearranges the dimension of the signal sequence output by the bidirectional double-layer GRU network, changing it from (Batchsize, Time, embed_dim) to (Batchsize*Time, Height*width, Channel), and then further splits and rearranges it to (Batchsize, Channel, Time, Height, width); then it standardizes it through a normalization layer; then it linearly maps the standardized tensor into a tensor of size D in m, n, p, q, corresponding to the first path, the second path, the third path, and the fourth path respectively.

[0011] It should be noted that only the first 3D-Vim Block is equipped with a dimension rearrangement layer.

[0012] Further preferably, in this embodiment, the first path is composed of a forward convolution, a backward convolution, a forward state space model, and a backward state space model. The first path processes m in the forward and reverse directions respectively. For each direction, first apply a 3D convolution to m to obtain m s , then, map m s to B s , C s , Δ s . Then convert Δ s into and Finally, calculate y through 3D SSM forward and y backward

[0013] Further preferably, in this embodiment, the second path is composed of an activation layer; the second path activates the tensor n for later use.

[0014] Further preferably, in this embodiment, the third path is composed of an S-T Cross Attention network. The S-T Cross Attention network includes a spatial attention branch and a temporal attention branch. The input tensor sequence p is decoupled spatio-temporally and then undergoes attention learning and latent space feature extraction through the spatial attention branch and the temporal attention branch respectively, and spatio-temporal feature fusion is performed after the learning is completed.

[0015] The S-T Cross Attention network processes the input temporal feature tensor, and the method is as follows: A. In the spatial attention branch, the temporal feature tensor is convolved through 3D convolution to obtain a query matrix Q, a key matrix K, and a value matrix V. The spatial dimension is flattened to confine the attention calculation to the spatial scale that is independent of each time step, and the spatial dimension is restored after the calculation is completed; B. In the temporal attention branch, the temporal feature tensor is convolved through 3D convolution to obtain a query matrix Q, a key matrix K, and a value matrix V. The temporal dimension is flattened to confine the attention calculation to the temporal scale that is independent of each spatial position, and the temporal dimension is restored after the calculation is completed; C. Fuse the temporal feature tensors processed by the spatial attention branch and the temporal attention branch to obtain a fused temporal feature tensor; D. The fused temporal feature tensor is fused with the input temporal feature tensor after passing through a linear layer as the output of the S-T Cross Attention network.

[0016] Further preferably, in this embodiment, the fourth path is composed of an S-T 3D VAE network. In the S-T 3D VAE network, both the downsampling Encoder and the upsampling Decoder are divided into two steps: The first-step Encoder is the Spatio Encoder, which only performs spatial downsampling on the tensor, keeping the time scale and dimension unchanged; the second-step Encoder is the Time Encoder, which only performs temporal downsampling on the tensor, keeping the spatial scale and dimension unchanged; after two steps of spatio-temporal separated Encoders, the latent space features extracted by the S-T 3D VAE are obtained. After obtaining the latent space features, first, the first-step Decoder is the symmetric Time Decoder, which only performs temporal upsampling on the latent space feature tensor, keeping the spatial scale and dimension unchanged; the second-step Decoder is the Spatio Decoder, which only performs spatial upsampling on the tensor, keeping the time scale and dimension unchanged; finally, the output of the S-T 3D VAE is obtained.

[0015] When the activation layer, S-T Cross Attention, and S-T 3D VAE network complete the internal processing of their respective input data, their respective output data are gated and then summed to obtain the first intermediate result; the first intermediate result and y obtained from the first path forward and y backward are then gated and added to obtain the output, and the dimension arrangement of the output tensor is (Batchsize, Channel, Time, Height, width).

[0017] The S-T Cross Attention and S-T 3D VAE structures adopt the idea of spatio-temporal separation (separation of motion and content), decouple spatio-temporal information and perform attention learning and latent space feature extraction separately, and then perform spatio-temporal feature fusion after learning. This reduces the model learning difficulty while improving the model learning effect, and can generate more realistic MPI time-series image sequences. Different branches are responsible for processing tensor information with the same source in different ways, and the results of each branch are subjected to gated operations and combination. The final result can effectively absorb the information provided by each branch. And if there are situations such as learning deviation or fitting difficulties in some branches during the training process, the gated operation and combination operation can effectively ensure that the results of the remaining branches correct and limit the results of the problem branches, thus increasing the robustness during the model training process. The long sequence processing ability of 3D Vision Mamba provides global temporal context for spatio-temporal attention, enabling it to more accurately locate key frames and local regions; S-T Cross Attention makes up for the deficiency of 3D Vision Mamba in local feature modeling through refined interaction in the spatial and temporal dimensions; while the latent space separation of S-T 3DVAE not only provides a structured feature representation for the former two, but also avoids the confusion of spatio-temporal features through decoupling constraints.

[0018] Preferably, each 3D-Vim Block has a Skip Connection mechanism to fuse the feature tensors obtained from the front and rear of the 3D-Vim Block, ensuring the information concentration while preventing the model from experiencing gradient vanishing or gradient explosion, providing the possibility for the establishment of a deep model.

[0019] After the 3D-Vim Block, a 3D-Patch Expanding operation is connected. The 3D-Patch Expanding is responsible for gradually increasing the resolution of each frame of the image without changing the number of features in the time dimension (i.e., the number of frames), and inputting it to the next-level 3D-Vim Block. There is no convolution operation inside the 3D-Patch Expanding, only linear transformation and tensor rearrangement, so the "checkerboard effect" brought by the convolution operation can be avoided, and the upsampling quality can be improved. The subsequent 3D-Vim Block receives the tensor output by the previous upsampling block and conducts learning, and the steps are the same as those of the first 3D-Vim Block (the subsequent 3D-Vim Block does not contain the dimension rearrangement operation). There are a total of L 3D-Vim Blocks and L - 1 3D-Patch Expanding blocks. The tensor with the output dimension arrangement of (Batchsize, Channel, Time, Height, width) output by the last 3D-Vim Block is the reconstructed time-series image sequence.

[0020]

[0020] The application of 3D convolution and 3D SSM, compared with the original 1D convolution and 1D SSM, not only learns the internal relationships (spatial relationships) of single-frame elements in the tensor sequence, but also learns the relationships between elements of different frames in the tensor sequence. It can efficiently capture the long-range temporal dependencies between video frames and the short-range spatial dependencies within frames. The resulting state space and feature tensors have spatio-temporal consistency, which can improve the generation effect of this solution for the scenario. It can not only process long videos with the linear computational complexity of 3D Vision Mamba, enhance the perception of spatial details through the attention mechanism, and improve the generation controllability by leveraging the latent space normalization of S-T 3D VAE. Ultimately, it forms a multi-level optimization of "global-local-latent" in terms of computational efficiency, feature discriminability, and generation quality, thereby breaking through the performance bottleneck of single or dual-module combinations and significantly improving the quality of reconstructed temporal image sequences.

[0021]

[0021] 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 as follows: Use simulation software to generate three-dimensional temporal image sequences and one-dimensional signal sequences. The temporal image sequences and the signal sequences are paired data in one-to-one correspondence; 70% of the paired data is used as the training set, and 30% is used as the test set, and real-time acquired MPI temporal one-dimensional signals of pre-made phantoms are added to the test set. Specifically as follows: Use MATLAB to generate 10,000 temporal grayscale image sequences of three different models, namely geometric models, letter models, and resolution models. After obtaining the temporal grayscale image sequences, use MPIRF simulation software to generate the corresponding one-dimensional signal sequences; add a time dimension to all the temporal binary image sequences, and stack the two-dimensional images in the time dimension in their corresponding sequence order to generate 10,000 corresponding three-dimensional temporal image sequences, and the one-dimensional signal sequences and the three-dimensional temporal image sequences are in one-to-one correspondence; 70% of the data of 10,000 three-dimensional temporal image sequences and their corresponding one-dimensional signals is used as the training set, and 30% is used as the test set; among them, the one-dimensional signal sequence in the training set is denoted as sequence S, the three-dimensional temporal image sequence is denoted as sequence I, the one-dimensional signal sequence in the test set is denoted as sequence S*, and the three-dimensional temporal image sequence is denoted as sequence I*; and 100 groups of MPI temporal one-dimensional signals collected in real time by MPI devices are added to the test set.

[0022]

[0022] More preferably, collect the data of the phantom as the MPI temporal one-dimensional signal collected in real time during the training stage of the above model construction. Inject magnetic nanoparticle reagents into different pre-made phantom models, use MPI devices to scan the phantom models, and actually collect 100 groups of signal sequence data of the phantom models as another part of the test set data.

[0023] Preferably, during the training process of the generation network model, the frame-averaged L1 Loss between the MPI time series image and the time series image sequence is used as the loss function of the GRU-3D Vision Mamba network: where T represents the number of frames in the time series image sequence in the training set, x represents the time series image sequence, and y represents the MPI time series image output by the GRU-3D Vision Mamba; the subscript i represents the i-th frame in the image sequence; represents the L1 distance, which is used to measure the frame-averaged pixel-level difference between the MPI time series image y and the time series image sequence x; α and β represent weight coefficients; y i represents the i-th frame of the MPI time series image output by the GRU-3D Vision Mamba; x i represents the i-th frame of the time series image sequence in the training set; φ(y i / x i ) h,w represents the pixel value at the h-th row and w-th column of the image; H l represents the height of the image; W l represents the width of the image.

[0024] Preferably, after obtaining the time series image sequence, verify the time series image sequence: The test set includes sequence S*, sequence I*, and the MPI time series one-dimensional signal. First, observe the similarity between the MPI time series image generated after the input of sequence S* and sequence I*, and use three indicators: the frame-averaged root mean square error (RMSE), the frame-averaged peak signal-to-noise ratio (PSNR), and the frame-averaged structural similarity index (SSIM) to measure the performance of the model on the test set. If the RMSE error is less than the pre-set threshold and the SSIM and PSNR indicators are higher than the pre-set threshold, the accuracy meets the requirements; then observe the effect of the MPI time series image generated after the input of the MPI time series one-dimensional signal. If the generated image is clear and the boundary is clear, the generation effect meets the requirements; otherwise, retrain.

[0025] In the model training stage, after the MPI time series image sequence output by the generation network model passes the verification of other indicators such as sensitivity and no problems are found during the experiment, the MPI time series one-dimensional signal of the subsequent target to be modeled in the resting state can be obtained.

[0026] Preferably, in this embodiment, the GRU-3D VisionMamba network is constructed using the Pytorch 2.5.1 and cuda 12.1 frameworks, and it is trained using 4 NVIDIA GeForce RTX 4090s. The Adam algorithm is used for optimization during the training process, and the learning rate is set to 0.00001.

[0027] Step S200: Surface-modify the tracer and administer the modified tracer to the target to be modeled. In this embodiment, SPIONs with surface-modified iron chelators (or high-iron affinity ligands) are used as NTBI recognition probes, making them SPION probes with the ability to specifically bind or specifically retain in iron-rich microenvironments (with significant aggregation or distribution differences at higher iron deposition sites in the substantia nigra).

[0028] Preferably, in this embodiment, the SPIONs are first coated with a PEG or SiO2 shell to increase stability and reduce the tendency of immune responses. Then, carboxyl (-COOH) active groups are introduced onto the shell to provide chemical anchor points for coupling iron chelators in the following steps. The DFO molecule is selected as the iron chelator, and its covalent coupling with the surface functional groups of SPIONs can be achieved through simple EDC / NHS activation. After the reaction is completed, centrifugation is carried out to remove the excess uncoupled chelator. Dynamic light scattering (DLS) and transmission electron microscopy (TEM) are used to characterize the particle size and dispersibility to ensure that the size and monodispersity of the modified SPIONs are maintained within an appropriate range.

[0029] It should be noted that PD patient detection can only be carried out after experimental verification. Before scanning, the modified tracer is injected into PD patients.

[0030] During the traditional MPI imaging process, SPION tracers exist in the blood and do not enter specific cells. Therefore, unmodified SPIONs cannot be used to image the substantia nigra region of PD patients. This patent surface-modifies existing SPION tracers, including increasing stability and reducing the tendency of immune responses, coupling iron chelators, and using techniques such as DLS (dynamic light scattering) and TEM (transmission electron microscopy) to ensure that the size and monodispersity of the modified SPIONs are maintained within an appropriate range, enabling the SPION tracer to penetrate the blood-brain barrier and enter specific cells to play a role, providing a method for using MPI technology to image the substantia nigra region of PD patients.

[0031] Step S300: Set the scanning parameters, and the MPI device collects the one-dimensional MPI time-series signal of the target to be modeled in the resting state. The one-dimensional MPI time-series signal is input into the trained generation network model for image reconstruction to obtain an MPI time-series image, which is a time-series brain functional image distributed on the time axis. In this embodiment, scanning parameters are set. The excitation magnetic field gradient is 1.5 T, the excitation field frequency is 25 kHz, and the magnetic field gradient is set as follows: in the Z direction: 1 T / m, in the X direction: 2 T / m, and in the Y direction: 1 T / m.

[0032] Step S400: Preprocess the MPI time series image, divide the preprocessed MPI time series image into brain regions, and extract the average voxel signal of the brain region corresponding to the target basal ganglia to be modeled.

[0033] Preferably, the method for obtaining the average voxel signal is as follows: After dividing the brain regions, for the brain region corresponding to the basal ganglia, it is divided into a number of voxel values with a size of 2 mm * 2 mm * 2 mm. Each voxel value is arranged according to the time series to obtain n voxel signals, and each voxel signal has t features over time. Average these n voxel signals to obtain the average voxel signal x of the brain region corresponding to the basal ganglia. t .

[0034] Preferably, the preprocessing includes operations such as position correction, registration, smoothing, and filtering; based on the AAL human brain standard template, the registered MPI time series image is divided into brain regions.

[0035] Step S500: Calculate the root mean square value and the STFT spectrogram based on the average voxel signal. The method is as follows: Calculate the root mean square value RMS of the average voxel signal x t : Use the short-time Fourier transform to detect the consistency of frequency distribution: According to X(t,f), use Matlab software to draw the STFT spectrogram; where X(t,f) represents the energy density of the signal at time t and frequency f; t represents the current analysis time center point; τ represents the integration variable, which is restricted by the window function ω(τ - t); x(τ) represents the original signal; j represents the imaginary unit; e -j2πfτ represents the complex exponential basis function Using the root mean square value of the signal as the amplitude measure can, to a certain extent, offset the influence of random noise.

[0036] Although the various steps are described in the above order in the above embodiments, those skilled in the art can understand that in order to achieve the effects 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 reversed order, and these simple changes are all within the protection scope of the present invention.

[0037] The MPI magnetic particle imaging system for staging detection of PD patients in the second embodiment of the present invention, the system includes a signal acquisition device and a central processing device; The signal acquisition device includes an MPI magnetic particle imaging device; the signal acquisition device is configured to administer a modified tracer to the target to be modeled, and set scanning parameters, and acquire the MPI one-dimensional voltage signal of the target to be modeled in the resting state based on the scanning parameters; The central processing device includes a CPU and a GPU; the central processing device includes a model construction module, a three-dimensional reconstruction module, a brain region division module, and a spectrogram generation module; The model construction module is configured to construct and train a generation network model; The three-dimensional reconstruction module is configured to perform image reconstruction based on the MPI one-dimensional voltage signal through the generation network model to obtain an MPI time series image; The brain region division module is configured to divide brain regions based on the MPI time series image to obtain an average voxel signal x t ; The spectrogram generation module is configured to calculate the RMS and the X(t,f) based on the average voxel signal, and then draw an STFT spectrogram using Matlab software according to the X(t,f).

[0038] Further, the root mean square value RMS and the STFT spectrogram of the average voxel signal x t are analyzed and compared with the root mean square value RMS and the STFT spectrogram of the average voxel signal x t of the normal aging population in the database to realize the diagnosis of the specific stage of the PD disease of PD patients.

[0039] The analysis method is as follows: First, confirm that the obtained average voxel signal x t is statistically valid. After obtaining the STFT spectrogram of x t , for each time slice m, find the peak position near the third harmonic frequency from the STFT spectrogram, and record it as Then calculate within the entire time slice mean and standard deviation. For the STFT spectrogram of the average voxel signal x t of the normal aging population in the database, use the same calculation method, and compare the calculated mean and standard deviation of of PD patients with the mean and standard deviation of of the normal aging population. Then the average voxel signal x tThe root mean square value RMS and the average voxel signal x of the normal aging population in the database t are compared with the root mean square value RMS of

[0040] PD patients and the normal aging population only differ in the NTBI content in the substantia nigra region. Therefore, the modified SPION tracer can combine with local NTBI in the substantia nigra region of PD patients to form a stable SPION-iron complex, resulting in an enhanced non-linear response signal of magnetic nanoparticles in the substantia nigra region during imaging, that is, the average voxel signal x of PD patients t will have a higher root mean square value RMS than the average voxel signal x of the normal aging population t However, theoretically, the difference in NTBI content does not affect the frequency distribution of the SPION magnetic response signal. Therefore, the mean and standard deviation of of PD patients and those of the normal aging population should be roughly equal to ensure that the obtained average voxel signal x t is statistically valid. If the root mean square value RMS of the average voxel signal x of PD patients t slightly increases relative to the normal aging population but is relatively limited, it can generally be determined that the PD symptoms are in the early stage. If the root mean square value RMS of the average voxel signal x of PD patients t significantly increases, the signal distribution in the substantia nigra region expands, and the signal amplitude is significantly higher than the normal value range, it can generally be determined that the PD symptoms are in the mid-stage. If the root mean square value RMS of the average voxel signal x of PD patients t is extremely high, and even the root mean square value RMS of the average voxel signal x of the related nuclei around the basal ganglia may also increase, and the signal in the substantia nigra region is not only strong but also shows a more diffuse or irregular enhancement pattern in the substantia nigra and even adjacent nucleus regions, it can generally be determined that the PD symptoms are in the late stage. t

[0041] Those skilled in the art of the present technology can clearly understand that for the convenience and simplicity of description, the specific working process and related descriptions of the above-described system can refer to the corresponding process in the foregoing method embodiments and will not be elaborated herein.

[0042] ​It should be noted that the MPI magnetic particle imaging system for staging detection of PD patients provided in the above embodiments is only illustrated by dividing the above functional modules. In actual applications, the above functions can be allocated to different functional modules according to needs, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiments can be combined into one module, or further split 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 each module or step, and are not regarded as improper limitations of the present invention.

[0043] An electronic device according to a third embodiment of the present invention includes: At least one processor; and A memory communicatively connected to at least one of the processors; wherein, The memory stores instructions executable 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 staging detection of PD patients.

[0044] A computer-readable storage medium according to a fourth embodiment of the present invention 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 staging detection of PD patients.

[0045] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes and related descriptions of the above-described electronic device and computer-readable storage medium can refer to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0046] Those skilled in the art should be able to realize that the modules and method steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. The programs corresponding to the software modules and method steps can be placed in a random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium well-known in the art. To clearly illustrate the interchangeability of electronic hardware and software, the components and steps of each example have been generally described according to functions in the above description. Whether these functions are executed in the form of 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.

[0047] Computer program code for performing the operations of this application can be written in one or more programming languages or combinations thereof. The above-mentioned programming languages include object-oriented programming languages such as Java, Smalltalk, C++, and also include conventional procedural programming languages such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, executed as an independent software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (for example, by using an Internet service provider to connect through the Internet).

[0048] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in the flowchart or block diagram can represent a module, a program segment, or a part of code that 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 blocks may occur in a different order than that marked in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and the combination of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0049] The terms "first", "second", etc. are used to distinguish similar objects, rather than to describe or represent a specific order or sequence.

[0050] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, so that a process, method, article, or device / equipment that includes a series of elements not only includes those elements, but also includes other elements not explicitly listed, or also includes elements inherent in these processes, methods, articles, or devices / equipment.

[0051] So far, the technical solution of the present invention has been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easily understood by those skilled in the art 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 staged detection of PD patients, which is used to reconstruct the MPI time series images of the target to be modeled and then draw the STFT spectrogram, and is characterized in that, It includes the following steps: Step S100: Construct and train a generation network model; Step S200: Perform surface modification on the tracer and apply the modified tracer to the target to be modeled; Step S300: Collect the one-dimensional MPI time series signal of the resting state of the target to be modeled, input the one-dimensional MPI time series signal into the trained generation network model for image reconstruction, and obtain the MPI time series image; Step S400: Preprocess the MPI time series image, divide the preprocessed MPI time series image into brain regions, and extract the average voxel signal; Step S500: Calculate the root mean square value and the STFT spectrogram based on the average voxel signal; Among them, the generation network model is a GRU-3D Vision Mamba network, which includes a bidirectional double-layer GRU network and a 3D Vision Mamba network connected in sequence. The 3D Vision Mamba network is composed of L 3D-Vim Blocks and L-1 upsampling blocks arranged alternately. A dimension rearrangement layer is provided before the first 3D-Vim Block for dimension rearrangement of the signal sequence output by the bidirectional double-layer GRU network; The 3D-Vim Block divides the dimension-rearranged signal sequence into multiple paths for processing after passing through a normalization layer; among them, one path outputs forward data and backward data, and the remaining paths generate multiple derivative data and perform specific correlation operations with the forward data / backward data respectively to obtain the output result.

2. The MPI magnetic particle imaging method for staging detection of PD patients according to claim 1, wherein There are four paths, including the first path, the second path, the third path, and the fourth path; the path refers to the logical channel for performing specific processing on the input data and outputting the result. Among them, the first path processes the input data and outputs the forward data and the backward data; the second path, the third path, and the fourth path perform internal processing on their respective input data in different ways, and sum the output data after gating operation to obtain the first intermediate result; the first intermediate result performs gating operations with the forward data and the backward data respectively to obtain the second intermediate result and the third intermediate result; the second intermediate result and the third intermediate result are added to obtain the output tensor, and the output tensor is fused with the input tensor of the normalization layer of the 3D-Vim Block after passing through a linear layer as the output of the 3D-Vim Block.

3. A MPI magnetic particle imaging method for staging detection of PD patients according to claim 2, wherein, The first path is composed of a forward convolution, a backward convolution, a forward state space model, and a backward state space model. The second path is composed of an activation layer; the third path is composed of an S-T Cross Attention network; the fourth path is composed of an S-T 3D VAE network.

4. A MPI magnetic particle imaging method for staging detection of PD patients according to claim 3, characterized in that, The S-T Cross Attention network includes a spatial attention branch and a temporal attention branch. The input tensor sequence is decoupled in space and time and then undergoes attention learning and latent space feature extraction through the spatial attention branch and the temporal attention branch respectively, and spatio-temporal feature fusion is performed after the learning ends.

5. A MPI magnetic particle imaging method for staging detection of PD patients according to claim 4, characterized in that, The S-T Cross Attention network processes the input temporal feature tensor in the following way: In the spatial attention branch, the temporal feature tensor is used to obtain a query matrix, a key matrix, and a value matrix through 3D convolution. The spatial dimension is flattened, so that the attention calculation is limited to the spatial scale that is independent for each time step. After the calculation is completed, the spatial dimension is restored; In the temporal attention branch, the temporal feature tensor is used to obtain a query matrix, a key matrix, and a value matrix through 3D convolution. The temporal dimension is flattened, so that the attention calculation is limited to the temporal scale that is independent for each spatial position. After the calculation is completed, the temporal dimension is restored; The temporal feature tensors processed by the spatial attention branch and the temporal attention branch are fused to obtain a fused temporal feature tensor; The fused temporal feature tensor is linearly processed and then fused with the input temporal feature tensor as the output of the S-T Cross Attention network.

6. A MPI magnetic particle imaging method for staging detection of PD patients according to claim 3, wherein The S-T 3D VAE network includes a spatial encoder, a temporal encoder, a temporal decoder, and a spatial decoder; The S-T 3D VAE network processes the input temporal feature tensor in the following way: The input temporal feature tensor is downsampled in the spatial scale and the temporal scale to obtain the extracted latent space features, and then the latent space features are upsampled in the temporal scale and the spatial scale to obtain the output tensor of the S-T 3D VAE.

7. A MPI magnetic particle imaging method for staging detection of PD patients according to claim 1, characterized in that, When the generation network model is trained, a three-dimensional time-series image sequence and a one-dimensional signal sequence are generated by simulation software. The time-series image sequence and the signal sequence are paired data that correspond one-to-one; 70% of the paired data is used as the training set, and 30% is used as the test set. The MPI time-series one-dimensional signal of a pre-made phantom collected in real time is added to the test set.

8. A MPI magnetic particle imaging method for staging detection of PD patients according to claim 7, characterized in that, During the training process of the generation network model, the loss function L of the GRU-3DVision Mamba network GRU-3DVisionMamba is as follows: Among them, T represents the number of frames in the temporal image sequence of the training set; represents the L1 distance, which is used to measure the average pixel-level difference per frame between the MPI time series image output by GRU-3DVision Mamba and the temporal image sequence in the training set; α and β represent weight coefficients; y i represents the i-th frame of the MPI time series image output by GRU-3DVision Mamba; x i represents the i-th frame of the temporal image sequence in the training set; φ(y i / x i ) h,w represents the pixel value at the h-th row and w-th column of the image; H l represents the height of the image; W l represents the width of the image.

9. A MPI magnetic particle imaging method for staging detection of PD patients according to claim 1, characterized in that, The root mean square value and the STFT spectrogram are calculated based on the average voxel signal in the following way: Calculate the root mean square value RMS of the average voxel signal x t : The short-time Fourier transform is used to detect the consistency of the frequency distribution: Draw the STFT spectrogram according to X(t,f); where X(t,f) represents the energy density of the signal at time t and frequency f; t represents the central time point of the current analysis; τ represents the integration variable, restricted by the window function ω(τ - t); x(τ) represents the original signal; j represents the imaginary unit; e -j2πfτ represents the complex exponential basis function.

10. An MPI magnetic particle imaging system for staging detection of PD patients, based on an MPI magnetic particle imaging method for staging detection of PD patients according to any one of claims 1-9. The system includes a signal acquisition device and a central processing device; The signal acquisition device includes an MPI magnetic particle imaging device; the signal acquisition device is configured to apply a modified tracer to the target to be modeled, set scanning parameters, and obtain the MPI one-dimensional voltage signal of the target to be modeled in the resting state based on the scanning parameters; The central processing device includes a CPU and a GPU; the central processing device includes a model construction module, a three-dimensional reconstruction module, a brain region division module, and a spectrogram generation module; The model construction module is configured to construct and train a generation network model; The three-dimensional reconstruction module is configured to perform image reconstruction based on the MPI one-dimensional voltage signal through the generation network model to obtain an MPI time-series image; The brain region division module is configured to divide brain regions based on the MPI time-series image to obtain an average voxel signal; The spectrogram generation module is configured to calculate the root mean square value RMS and the energy density X(t,f) based on the average voxel signal, and then draw the STFT spectrogram.

Images