A method and device for analyzing neurological images

Abstract

The present invention discloses a method and device for analyzing nervous system images, and relates to the technical field of diagnosis and treatment of nervous system diseases. The specific steps of the analysis method are: S100, multimodal data fusion acquisition: collecting patients' functional magnetic resonance imaging (fMRI), magnetic resonance spectroscopy (MRS), diffusion tensor imaging (DTI), and positron emission tomography (PET) nervous system imaging data through the hospital's internal system. The present invention integrates multimodal data of functional magnetic resonance imaging, magnetic resonance spectroscopy (MRS), diffusion tensor imaging (DTI), and positron emission tomography (PET), and performs quality assessment and integration processes in a central data processing unit, thereby significantly enhancing the accuracy and reliability of the data. At the same time, a spatiotemporal joint feature analysis model is introduced to connect the time and space dimension information in the image data, and to explore the changing patterns and intrinsic correlations of the data in the spatiotemporal and temporal dimensions, thereby improving the diagnostic accuracy of nervous system diseases.

Description

Technical Field

[0001] The present invention relates to the technical field of diagnosis and treatment of nervous system diseases, and in particular to a nervous system image analysis method and device. Background Art

[0002] With the continuous advancement of medical technology, neurological imaging analysis plays an increasingly important role in the diagnosis, treatment and research of neurological diseases. In recent years, multimodal imaging technologies, such as functional magnetic resonance imaging (fMRI), magnetic resonance spectroscopy (MRS), diffusion tensor imaging (DTI) and positron emission tomography (PET), have been increasingly used in the diagnosis of neurological diseases. These technologies can reveal the structural and functional changes of the nervous system from different perspectives, providing doctors with rich diagnostic information. At the same time, with the rapid development of big data, artificial intelligence and machine learning technologies, data processing and analysis capabilities have been significantly improved, providing new means and methods for neurological imaging analysis.

[0003] Although traditional nervous system imaging analysis technology has achieved certain results, it still has some shortcomings. First, traditional methods only rely on single-modality imaging data for analysis, which limits the comprehensiveness and accuracy of diagnostic information. Imaging data of different modalities reflect information on different aspects of the nervous system. Single-modality analysis tends to ignore other important information, thus affecting the accuracy of diagnosis. Second, when processing and analyzing imaging data, traditional methods often ignore the changing patterns and correlations of data in the temporal and spatial dimensions, resulting in insufficient sensitivity and specificity of diagnostic results. In addition, traditional methods lack personalized treatment recommendations and cannot meet the specific needs of different patients, affecting the treatment effect and the quality of life of patients.

[0004] Therefore, the development of a neurological imaging analysis method and device is expected to not only address the limitations of traditional technologies, but also provide more effective, safe and personalized treatment services for patients with neurological diseases. Summary of the Invention

[0005] The purpose of the present invention is to make up for the shortcomings of the existing technology and provide a method and device for analyzing nervous system images. By fusing multimodal imaging data and utilizing spatiotemporal joint feature analysis technology, the diagnostic accuracy of nervous system diseases is improved. At the same time, combined with the treatment effect prediction model, personalized treatment plans are recommended according to the patient's specific situation, thus achieving the goal of precision medicine.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: On the one hand, a method for analyzing nervous system images, the specific steps of the analysis method are:

[0007] S100, multimodal data fusion acquisition: The hospital collects patients' multimodal neurological imaging data, including functional magnetic resonance imaging (fMRI), magnetic resonance spectroscopy (MRS), diffusion tensor imaging (DTI), and positron emission tomography (PET), through its internal system. The collected raw data is transmitted to the central data processing unit in real time. The integrated multimodal data is then quality-assessed and any problematic data is repaired. After repair, the data is integrated using a multimodal data fusion formula.

[0008] S200, spatiotemporal joint feature analysis: pre-process the fused dynamic nervous system image data, remove artifacts and noise, perform spatiotemporal segmentation, extract spatiotemporal features, and fuse the spatiotemporal feature vector with the temporal feature vector to construct the spatiotemporal feature vector. Use the spatiotemporal correlation formula to mine the correlation between the spatiotemporal feature vector and its surrounding neighborhood in the spatiotemporal dimension;

[0009] S300, Construction of Multi-Disease Joint Diagnosis Model: Collect the patient's clinical information data and combine it with the data from the spatiotemporal joint feature analysis step, divide it into a test set and a validation set, and use the disease discrimination formula to build a diagnosis model. The formula is: Where d represents the disease type, m is the number of disease-related feature patterns, θ k 、ω k (x,y,z,t) and μ k (d) is the model parameter, φ k (t) is the time characteristic function, Z is the normalization constant, C ST (x, y, z) represents the spatiotemporal correlation feature value at the spatial position (x, y, z). The model is trained using the training set to determine the parameters θ in the model. k 、ω k (x,y,z,t),μ k (d) and φ k (t), then evaluate the model performance through the test set and optimize the model based on the evaluation results;

[0010] S400, Treatment Effect Prediction: Extract features related to treatment effects from multimodal imaging data and clinical information, construct feature vectors, build a prediction model using the treatment effect prediction formula, and use grid search to tune the model's hyperparameters. After independent data verification, input the corresponding features of new patients into the model to obtain treatment effect prediction results for different treatment methods;

[0011] S500, personalized treatment plan recommendation: Construct a knowledge graph for the treatment of neurological diseases, integrate diagnosis, prediction results and patient clinical information to form a personalized treatment information network, use intelligent decision-making algorithms to generate preliminary personalized treatment plans based on the network, continuously collect feedback information to update the network during the treatment process, and update the network and optimize the treatment plan based on the treatment feedback information.

[0012] Furthermore, in the multimodal data fusion acquisition, the multimodal data fusion formula is used to integrate the data. Assume that the signal strength of the t-th mode in the multimodal image data at the spatial position (x, y, z) and time t is S i (x,y,z,t), the formula is: Where n is the number of modes, α i , β, γ i and λ are fusion coefficients, τ is the variable in time integration, and dτ indicates that the integrated variable is τ.

[0013] Furthermore, in the S100, the signal strength S in the multimodal data fusion acquisition i Calculation of (x,y,z,t):

[0014] The calculation of the functional magnetic resonance imaging signal intensity is as follows: assuming that the signal intensity at the spatial position (x, y, z) and time t is S1(x, y, z, t), the expression thereof is S1(x, y, z, t) = M(x, y, z, t) × f(BOLD(x, y, z, t)), wherein M(x, y, z, t) represents a physical quantity related to magnetic resonance, and f(BOLD(x, y, z, t)) represents a function related to the BOLD effect;

[0015] The calculation of the magnetic resonance spectroscopy imaging signal intensity is as follows: S2(x, y, z, t) is the signal intensity of MRS at the spatial position (x, y, z) and time t. S2(x, y, z, t) = C metabolite (x,y,z,t)×g(MRS p arameters(x,y,z,t)), where C metubolite (x, y, z, t) is the concentration of a specific metabolite, g(MRS p arameters(x,y,z,t)) is a function related to MRS acquisition parameters;

[0016] The calculation of the diffusion tensor imaging signal intensity is as follows: S3(x,y,z,t) is the signal intensity of DTI at the spatial position (x,y,z) and time t. S3(x,y,z,t)=A(x,y,z,t)×exp(-bD(x,y,z,t)), where A(x,y,z,t) is a coefficient related to the tissue intrinsic signal intensity and other imaging parameters, b is a diffusion weighting factor, and D(x,y,z,t) is a diffusion tensor matrix whose elements D ij (x, y, z, t) (i, j = x, y, z) describes the diffusion characteristics of water molecules in different directions;

[0017] The positron emission tomography signal intensity is calculated by assuming that S4(x,y,z,t) is the signal intensity of PET at the spatial position (x,y,z) and time t. S4(x,y,z,t)=λ(x,y,z,t)×∫ V ρ r (x′,y′,z′,t)×R(xx′,yy′,zz′,t)dx′dy′dz′, where λ(x,y,z,t) is a coefficient related to detector efficiency and photon attenuation factors, and ρ r (x′,y′,z′,t) is the concentration of the radioactive tracer at the spatial location (x′,y′,z′) and time t, and R(xx′,yy′,zz′,t) is a function that describes the physical processes of attenuation and scattering of photons from the emission point (x′,y′,z′) to the detection point (x,y,z).

[0018] Furthermore, in the S200, temporal feature extraction in the spatiotemporal joint feature analysis, for each ROI time window subsequence Use discrete wavelet transform to extract time features, and set the wavelet function as ψ j (t)(j is the wavelet scale), then the wavelet coefficient Constructing time feature vector Where J is the number of selected wavelet scales.

[0019] Furthermore, in the spatial feature extraction in the spatiotemporal joint feature analysis in S200, Geary's C index is used to calculate the spatial correlation feature between voxels in each ROI, and the voxel i in the ROI is set to be (x i ,y i ,z i ) and voxel j = (x j ,y j ,z j ) is characterized by the value of F i and F j , the calculation formula is: Where N is the number of voxels in the ROI, wij is the spatial weight matrix, when voxels i and j are adjacent, w ij =1, otherwise w ij =0, The dynamic time warping algorithm is used to calculate the functional connectivity features between different ROIs. The time series features of the two ROIs are and The calculation formula of DTW distance is: Where ω is the time series matching path, d is the feature distance measurement function, and the spatial feature vector is constructed.

[0020] Furthermore, in the S200, the spatiotemporal joint feature analysis utilizes the spatiotemporal correlation formula to mine the correlation between the spatiotemporal feature vector and the surrounding neighborhood in the spatiotemporal dimension, and the formula is: Where: C ST (x,y,z) represents the spatiotemporal correlation eigenvalue at the spatial position (x,y,z), N(x,y,z) is the neighborhood of the spatial position (x,y,z), and F ST,k,n (x, y, z) is the spatiotemporal feature value of the kth region of interest at the spatial position (x, y, z) in the nth time window, is the mean of the spatiotemporal features of the kth ROI at the spatial position (x, y, z), and is calculated as follows:

[0021] Furthermore, in the S300, the training of the diagnosis model in the construction of the multi-disease joint diagnosis model is to input the training set data into the constructed model and use the back propagation algorithm to train the model. F i is the characteristic data of the i-th sample, d i is the corresponding disease type label, and the model prediction output is Calculate the cross entropy loss function, the formula is: Calculate the loss function for θ k The gradient of is: The specific gradient expression is obtained by gradually expanding the calculation through the chain rule, and the model parameters are updated using stochastic gradient descent. The formula is: η is the learning rate, the number of training rounds is set to E, and after each round of training, the model performance is evaluated on the validation set.

[0022] Furthermore, in the S400, a prediction model is constructed by using a treatment effect prediction formula in the treatment effect prediction, and the formula is: Where: E(r|F,d) represents the probability value of the treatment outcome indicator r predicted based on the image feature F(x,y,z,t) and the disease type d, β0, β1, β2 are coefficients determined by clinical data training, δ r is the adjustment coefficient related to the characteristics of the treatment outcome indicator r itself.

[0023] Furthermore, in the above S500, the personalized treatment plan recommendation uses an intelligent decision-making algorithm to generate a preliminary personalized treatment plan based on this network, and the formula is: Where: T p represents the preliminary personalized treatment plan, n is the number of treatment plans in the knowledge graph that are related to the current patient's disease and meet the screening conditions, and w i is the initial weight of the i-th treatment plan in the knowledge graph, R i (d) is the adjustment factor based on disease matching, calculated as: in is the patient's disease feature vector, It is treatment plan T i The characteristic vector of the applicable disease, α is the adjustment parameter, C i (P) is an adjustment factor based on individual patient characteristics, K i (E) is an adjustment factor based on the predicted treatment effect, calculated as: Where β is the adjustment parameter, E i Predicted value of treatment effect, T i is the i-th treatment plan in the knowledge graph.

[0024] On the other hand, a nervous system imaging analysis device includes: a multimodal data fusion acquisition module, a spatiotemporal joint feature analysis module, a multi-disease joint diagnosis model construction module, a treatment effect prediction module, and a personalized treatment plan recommendation module.

[0025] The multimodal data fusion acquisition module collects the patient's fMRI, MRS, DTI, and PET modal neurological imaging data through the hospital's internal system, transmits them to the central data processing unit in real time, repairs problematic data, and integrates the data using a data fusion formula;

[0026] The spatiotemporal joint feature analysis module pre-processes the fused dynamic nervous system image data to remove artifacts and noise, extracts the temporal and spatial features of the data, fuses them to construct a spatiotemporal feature vector, and uses the spatiotemporal correlation formula to mine its correlation with the surrounding neighborhood;

[0027] The multi-disease joint diagnosis model construction module collects clinical information data and spatiotemporal joint feature analysis data, divides them into test sets and validation sets, constructs a model using a disease discrimination formula, trains the model using a backpropagation algorithm using the training set, and evaluates the model performance using the validation set;

[0028] The treatment effect prediction module extracts treatment effect-related features from multimodal imaging data and clinical information to construct a vector, builds a model based on the treatment effect prediction formula, adjusts hyperparameters using grid search, and verifies with independent data. It inputs new patient characteristics to obtain prediction results for the effects of different treatment methods.

[0029] The personalized treatment plan recommendation module: constructs a knowledge graph for the treatment of neurological diseases, integrates diagnosis, prediction results and patient clinical information to form a personalized treatment information network, uses an intelligent decision-making algorithm to generate a preliminary plan, and continuously collects feedback during treatment to update the network optimization plan.

[0030] Compared with the existing technology, this method and device for analyzing nervous system images have the following beneficial effects:

[0031] 1. The present invention integrates multimodal data from functional magnetic resonance imaging, magnetic resonance spectroscopy, diffusion tensor imaging, and positron emission tomography, and performs quality assessment and integration processes in the central data processing unit, significantly enhancing the accuracy and reliability of the data. At the same time, it introduces a joint spatiotemporal feature analysis model to connect the temporal and spatial dimensional information in the imaging data, and to explore the changing patterns and intrinsic correlations of the data in both the temporal and spatial dimensions, thereby improving the diagnostic accuracy of neurological diseases. This not only breaks through the limitations of traditional neurological imaging analysis, but also provides a foundation for early diagnosis and personalized precision treatment of neurological diseases, effectively promoting technological progress and clinical application expansion in this field.

[0032] 2. The present invention realizes accurate diagnosis of neurological diseases and personalized treatment plan recommendation by constructing a multi-disease joint diagnosis model and a knowledge graph for the treatment of neurological diseases. The multi-disease joint diagnosis model can comprehensively consider the patient's clinical information and the results of spatiotemporal joint feature analysis to perform differential diagnosis of various neurological diseases. The knowledge graph for the treatment of neurological diseases integrates diagnosis, prediction results and patient clinical information, and can generate personalized treatment plans using intelligent decision-making algorithms. This not only improves the diagnostic accuracy of the disease, but also provides patients with personalized treatment plans, which helps to improve the treatment effect and the patient's quality of life.

[0033] Other advantages, objects and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be learned from the practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort.

[0035] Figure 1 This is a flowchart of a method for analyzing nervous system images;

[0036] Figure 2 This is a schematic diagram of the structure of a nervous system image analysis device. DETAILED DESCRIPTION

[0037] The following is a clear and complete description of the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0038] Example 1:

[0039] Recommended diagnosis and treatment options for Alzheimer's disease

[0040] Multimodal data fusion acquisition: Patient A went to the hospital for treatment due to symptoms of memory loss and cognitive decline. The hospital's internal system collected functional magnetic resonance imaging (fMRI), magnetic resonance spectroscopy (MRS), diffusion tensor imaging (DTI), and positron emission tomography (PET) neurological imaging data for patient A within a week. Let the signal intensity of the t-th mode in the multimodal imaging data at spatial position (x, y, z) and time t be S i (x, y, z, t), the collected raw data is integrated through the multimodal data fusion formula, the formula is: Where n is the number of modes, α i , β, and λ are fusion coefficients. The integrated original data are transmitted to the central data processing unit in real time. The quality of the integrated multimodal data is evaluated and data with problems are repaired.

[0041] Spatiotemporal joint feature analysis, pre-processing of the fused dynamic nervous system imaging data, removal of artifacts and noise, and spatiotemporal segmentation, for temporal feature extraction, for each ROI time window subsequence Use discrete wavelet transform to extract time features, and set the wavelet function as ψ j(t) (j is the wavelet scale), then the wavelet coefficients are: Constructing time feature vector Where J is the number of wavelet scales selected. In terms of spatial feature extraction, Geary's C index is used to calculate the spatial correlation characteristics between voxels in each ROI. Let the voxel i in ROI = (x i ,y i ,z i ) and voxel j = (x j ,y j ,z j ) is characterized by the value of F i and F j , the calculation formula is: Where N is the number of voxels in the ROI, w i j is the spatial weight matrix, when voxels i and j are adjacent, w ij =1, otherwise w ij =0, The dynamic time warping algorithm is used to calculate the functional connectivity features between different ROIs. Assuming that the time series features of the two ROIs are F1(t) and F2(t), the calculation formula of the DTW distance is: DTW(F1, F2) = min ω ∑ (i,j)∈ω d(F1(i),F2(j)), where ω is the time series matching path, d is the feature distance metric function, and the spatial feature vector is constructed Finally, the time feature vector and the space feature vector are fused to construct the space-time feature vector F ST,k,n =F time,k,n ,F space,k , and use the spatiotemporal correlation formula to mine its correlation with the surrounding neighborhood in the spatiotemporal dimension. The formula is: Among them C S T(x,y,z) represents the spatiotemporal correlation eigenvalue at the spatial position (x,y,z), N(x,y,z) is the neighborhood of the spatial position (x,y,z), and F ST,k,n (x, y, z) is the spatiotemporal feature value of the kth region of interest at the spatial position (x, y, z) in the nth time window, is the mean of the spatiotemporal features of the kth ROI at the spatial position (x, y, z).

[0042] To build a multi-disease joint diagnosis model, we collect clinical information data of patient A, such as age, family medical history, and lifestyle habits, and combine it with the data from the spatiotemporal joint feature analysis step. The data is divided into a test set and a validation set, and the model is built using the disease discrimination formula (formula: Where d represents the disease type, m is the number of disease-related feature patterns, θ k 、ω k (x,y,z,t) and μ k (d) is the model parameter, φ k (t) is the time characteristic function, Z is the normalization constant, the training set data is input into the constructed model, and the model is trained using the back propagation algorithm. Suppose the training set data is F i is the characteristic data of the i-th sample, d i is the corresponding disease type label, and the model prediction output is d i , calculate the cross entropy loss function: Calculate the loss function for θ k Ladder: Degrees The specific gradient expression is obtained by gradually expanding the calculation through the chain rule, and the model parameters are updated using stochastic gradient descent: η is the learning rate, and the number of training rounds is set to E. After each round of training, the model performance is evaluated on the validation set, and the model is optimized and adjusted based on the evaluation results.

[0043] Treatment effect prediction: extract features related to treatment effect from multimodal imaging data and clinical information, construct feature vectors, and build a prediction model using the treatment effect prediction formula. The formula is: Where E(r|F,d) represents the probability value of the treatment outcome indicator r predicted based on the image feature F(x,y,z,t) and the disease type d, β0, β1, and β2 are coefficients determined by clinical data training, and δ r It is an adjustment coefficient related to the characteristics of the treatment outcome indicator r itself. Grid search is used to tune the model's hyperparameters. After independent data verification, the corresponding characteristics of patient A are input into the model to obtain the predicted results of the treatment effect of different treatment methods (such as drug therapy and cognitive training).

[0044] To recommend personalized treatment plans, we build a knowledge graph for the treatment of neurological diseases, integrate diagnosis and prediction results, and the clinical information of patient A to form a personalized treatment information network. We then use an intelligent decision-making algorithm to generate a preliminary personalized treatment plan based on this network. The formula is: Where T p represents the preliminary personalized treatment plan, n is the number of treatment plans in the knowledge graph that are related to the current patient's disease and meet the screening conditions, and w i is the initial weight of the i-th treatment plan in the knowledge graph, R i(d) is the adjustment factor based on disease matching, calculated as: Among them F d is the patient's disease feature vector, It is treatment plan T i The characteristic vector of the applicable disease, α is the adjustment parameter, C i (P) is an adjustment factor based on individual patient characteristics, K i (E) is an adjustment factor based on the prediction of treatment effect: Where β is the adjustment parameter, E i Predicted value of treatment effect, T i It is the i-th treatment plan in the knowledge graph. During the treatment process, feedback information is continuously collected to update the network, and the network is updated according to the treatment feedback information to optimize the treatment plan.

[0045] In summary, this embodiment conducts a comprehensive diagnosis and treatment analysis of Alzheimer's disease patient A through multimodal data fusion acquisition, spatiotemporal feature analysis, multi-disease joint diagnosis model construction, treatment effect prediction, and personalized treatment plan recommendation steps. It obtains rich information from multiple imaging modalities, accurately assesses the condition and predicts the treatment effect through complex calculations and model construction, and ultimately provides patient A with a personalized treatment plan, which can be continuously optimized during treatment, demonstrating the effectiveness and practicality of this method in the diagnosis and treatment of neurological diseases.

[0046] Example 2:

[0047] Parkinson's disease diagnosis and treatment process

[0048] Multimodal data fusion acquisition: Patient B has symptoms of hand tremor and slow movement. The hospital collected neurological imaging data of fMRI, MRS, DTI, and PET modalities within a week. Similarly, the signal intensity of the tth modality in the multimodal imaging data at spatial position (x, y, z) and time t is S i (x, y, z, t), the original data is integrated through the multimodal data fusion formula, the formula is: n is the number of modes, α i , β, and λ are fusion coefficients, and the integrated raw data are then transmitted to the central data processing unit in real time to perform quality assessment on the integrated multimodal data to ensure data quality.

[0049] Spatiotemporal joint feature analysis, pre-processing the fused dynamic nervous system image data, removing artifacts and noise, and performing spatiotemporal segmentation. When extracting temporal features, for each ROI time window subsequence Using discrete wavelet transform, let the wavelet function be ψ j(t) (j is the wavelet scale), wavelet coefficients: Constructing time feature vector (J is the number of wavelet scales selected), spatial feature extraction uses Geary's C index to calculate the spatial correlation characteristics between voxels in each ROI, assuming that the voxel i in ROI is i ,y i ,z i ) and voxel j = (x j ,y j ,z j ) is characterized by the value of F i and F j , the calculation formula is: N is the number of voxels in the ROI, w ij is the spatial weight matrix, the adjacent voxels w ij =1, otherwise w ij =0, The dynamic time warping algorithm is used to calculate the functional connectivity features between different ROIs. Assume that the time series features of the two ROIs are F1(t) and F2(t), and the DTW distance calculation formula is: DTW(F1, F2) = min ω ∑ (i,j)∈ω d(F1(i),F2(j)), ω is the time series matching path, d is the feature distance metric function, and the spatial feature vector is constructed Then the time feature vector and the space feature vector are fused into the space-time feature vector F ST,k,n =F time,k,n ,F space,k , use the spatiotemporal correlation formula to mine its spatiotemporal correlation, the formula is: represents the spatiotemporal correlation eigenvalue at the spatial position (x, y, z), N(x, y, z) is the neighborhood of the spatial position (x, y, z), and F ST,k,n (x, y, z) is the spatiotemporal feature value of the kth region of interest at the spatial position (x, y, z) in the nth time window, is the mean of the spatiotemporal features of the kth ROI at the spatial position (x, y, z).

[0050] The multi-disease joint diagnosis model is constructed by combining the clinical information of patient B (such as onset time and whether there are other underlying diseases) with the data from the spatiotemporal joint feature analysis step, dividing it into a test set and a validation set, and using the disease discrimination formula to build the model. The formula is: d represents the disease type, m is the number of disease-related feature patterns, θ k 、ω k (x,y,z,t) and μ k (d) is the model parameter, φ k (t) is the time characteristic function, Z is the normalization constant, and the training set data (F i is the sample characteristic data, d i is the disease type label, and the model prediction output is d i ) Input the model, use the back propagation algorithm to train, and calculate the cross entropy loss function: Calculate the loss function for θ k Gradient: The gradient expression is calculated by chain rule expansion, and the model parameters are updated using stochastic gradient descent. η is the learning rate, the number of training rounds is set to E, and the model performance is evaluated and optimized on the validation set after each round of training.

[0051] To predict the therapeutic effect, we extract features related to the therapeutic effect of Parkinson's disease, such as the imaging features of the substantia nigra and dopamine metabolism-related indicators to construct a feature vector. We then build a prediction model using the therapeutic effect prediction formula: Where E(r|F,d) represents the probability value of the treatment outcome indicator r predicted based on the image feature F(x,y,z,t) and the disease type d, β0, β1, and β2 are coefficients determined by clinical data training, and δ r It is an adjustment coefficient related to the characteristics of the treatment outcome indicator r itself. Grid search is used to tune the hyperparameters. After independent data verification, the corresponding characteristics of patient B are input into the model to predict the effects of different treatment methods (such as drug treatment and surgical treatment).

[0052] Personalized treatment plan recommendation, build a knowledge graph for the treatment of neurological diseases, integrate diagnosis, prediction results and clinical information of patient B, form a personalized treatment information network, and use intelligent decision-making algorithms to generate preliminary personalized treatment plans based on this network. Suppose there are n treatment plans related to Parkinson's disease and meeting the screening conditions in the knowledge graph (the value of n is determined by the number of plans that meet the conditions in the actual knowledge graph). For each treatment plan T i , calculate its initial weight w in the knowledge graph i , calculate the adjustment factor based on disease matching: Among them F d is the disease feature vector of patient B, It is treatment plan T i The characteristic vector of the applicable disease, α is the adjustment parameter, and the adjustment factor C is determined based on the individual characteristics of the patienti (P), calculate the adjustment factor based on the prediction of treatment effect: Where β is the adjustment parameter, E i The predicted value of the treatment effect of the treatment plan on patient B obtained in the treatment effect prediction step is used to calculate the preliminary personalized treatment plan T according to the formula. p , the formula is: For example, if there are three related treatment options in the knowledge graph, namely drug treatment option T1, surgical treatment option T2 and rehabilitation training option T3, after the above calculations to determine the weights and adjustment factors of each option, the personalized treatment option suitable for patient B may be drug treatment combined with a certain degree of rehabilitation training. During the treatment process, feedback information is continuously collected to update the network, such as the improvement of patient B's symptoms after a period of treatment and information on drug side effects. Based on this treatment feedback information, the treatment plan is further optimized, such as adjusting the drug dosage and increasing the intensity of rehabilitation training.

[0053] In summary, for Parkinson's disease patient B, this embodiment uses a series of processes, including collecting multimodal imaging data, analyzing spatiotemporal characteristics, building a diagnostic model, predicting treatment effects, and generating a personalized plan. In each link, data is processed according to relevant formulas and algorithms, and multiple factors are comprehensively considered to ultimately determine a treatment plan suitable for patient B. Continuous improvement is made through feedback, which demonstrates the scientific nature and reliability of this method in the diagnosis and treatment of Parkinson's disease and provides strong support for clinical treatment.

[0054] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A method for analyzing nervous system images, characterized in that: The specific steps of this analysis method are: S100, multimodal data fusion acquisition: The hospital collects patients' multimodal neurological imaging data, including functional magnetic resonance imaging (fMRI), magnetic resonance spectroscopy (MRS), diffusion tensor imaging (DTI), and positron emission tomography (PET), through its internal system. The collected raw data is transmitted to the central data processing unit in real time. The integrated multimodal data is then quality-assessed and any problematic data is repaired. After repair, the data is integrated using a multimodal data fusion formula. S200, spatiotemporal joint feature analysis: pre-process the fused dynamic nervous system image data, remove artifacts and noise, perform spatiotemporal segmentation, extract spatiotemporal features, and fuse the spatiotemporal feature vector with the temporal feature vector to construct the spatiotemporal feature vector. Use the spatiotemporal correlation formula to mine the correlation between the spatiotemporal feature vector and its surrounding neighborhood in the spatiotemporal dimension; S300, Construction of Multi-Disease Joint Diagnosis Model: Collect the patient's clinical information data and combine it with the data from the spatiotemporal joint feature analysis step, divide it into a test set and a validation set, and use the disease discrimination formula to build a diagnosis model. The formula is: Where d represents the disease type, m is the number of disease-related feature patterns, θ k 、ω k (x,y,z,t) and μ k (d) is the model parameter, φ k (t) is the time characteristic function, Z is the normalization constant, C ST (x, y, z) represents the spatiotemporal correlation feature value at the spatial position (x, y, z). The model is trained using the training set to determine the parameters θ in the model. k 、ω k (x,y,z,t),μ k (d) and φ k (t), then evaluate the model performance through the test set and optimize the model based on the evaluation results; S400, Treatment Effect Prediction: Extract features related to treatment effects from multimodal imaging data and clinical information, construct feature vectors, build a prediction model using the treatment effect prediction formula, and use grid search to tune the model's hyperparameters. After independent data verification, input the corresponding features of new patients into the model to obtain treatment effect prediction results for different treatment methods; S500, personalized treatment plan recommendation: Construct a knowledge graph for the treatment of neurological diseases, integrate diagnosis, prediction results and patient clinical information to form a personalized treatment information network, use intelligent decision-making algorithms to generate preliminary personalized treatment plans based on the network, continuously collect feedback information to update the network during the treatment process, and update the network and optimize the treatment plan based on the treatment feedback information.

2. A method for analyzing nervous system images according to claim 1, characterized in that: In the above S100, data integration is performed using a multimodal data fusion formula in the multimodal data fusion acquisition. Suppose the signal strength of the t-th mode in the multimodal image data at the spatial position (x, y, z) and time t is S i (x,y,z,t), the formula is: Where n is the number of modes, α i , β, γ i and λ are fusion coefficients, τ is the variable in time integration, and dτ indicates that the integrated variable is τ.

3. A method for analyzing nervous system images according to claim 1, characterized in that: S100, signal strength S in multimodal data fusion acquisition i Calculation of (x,y,z,t): The calculation of the functional magnetic resonance imaging signal intensity is as follows: assuming that the signal intensity at the spatial position (x, y, z) and time t is S1(x, y, z, t), the expression thereof is S1(x, y, z, t) = M(x, y, z, t) × f(BOLD(x, y, z, t)), wherein M(x, y, z, t) represents a physical quantity related to magnetic resonance, and f(BOLD(x, y, z, t)) represents a function related to the BOLD effect; The calculation of the magnetic resonance spectroscopy imaging signal intensity is as follows: S2(x, y, z, t) is the signal intensity of MRS at the spatial position (x, y, z) and time t. S2(x, y, z, t) = C metabolite (x,y,z,t)×g(MRS p arameters(x,y,z,t)), where C metubolite (x, y, z, t) is the concentration of a specific metabolite, g(MRS p arameters(x,y,z,t)) is a function related to MRS acquisition parameters; The calculation of the diffusion tensor imaging signal intensity is as follows: S3(x,y,z,t) is the signal intensity of DTI at the spatial position (x,y,z) and time t. S3(x,y,z,t)=A(x,y,z,t)×exp(-bD(x,y,z,t)), where A(x,y,z,t) is a coefficient related to the tissue intrinsic signal intensity and other imaging parameters, b is a diffusion weighting factor, and D(x,y,z,t) is a diffusion tensor matrix whose elements D ij (x, y, z, t) (i, j = x, y, z) describes the diffusion characteristics of water molecules in different directions; The positron emission tomography signal intensity is calculated by assuming that S4(x,y,z,t) is the signal intensity of PET at the spatial position (x,y,z) and time t. S4(x,y,z,t)=λ(x,y,z,t)×∫ V ρ r (x′,y′,z′,t)×R(xx′,yy′,zz′,t)dx′dy′dz′, where λ(x,y,z,t) is a coefficient related to detector efficiency and photon attenuation factors, and ρ r (x′,y′,z′,t) is the concentration of the radioactive tracer at the spatial location (x′,y′,z′) and time t, and R(xx′,yy′,zz′,t) is a function that describes the physical processes of attenuation and scattering of photons from the emission point (x′,y′,z′) to the detection point (x,y,z).

4. A method for analyzing nervous system images according to claim 1, characterized in that: S200, temporal feature extraction in spatiotemporal joint feature analysis, for each ROI time window subsequence Use discrete wavelet transform to extract time features, and set the wavelet function as ψ j (t), j is the wavelet scale, then the wavelet coefficient Constructing time feature vector Where J is the number of selected wavelet scales.

5. The method for analyzing nervous system images according to claim 1, wherein: In the S200, spatial feature extraction in the spatiotemporal joint feature analysis, Geary's C index is used to calculate the spatial correlation feature between voxels in each ROI, assuming that the voxel i in the ROI is equal to (x i ,y i ,z i ) and voxel j = (x j ,y j ,z j ) is characterized by the value of F i and F j , the calculation formula is: Where N is the number of voxels in the ROI, w ij is the spatial weight matrix, when voxels i and j are adjacent, w ij =1, otherwise w ij =0, The dynamic time warping algorithm is used to calculate the functional connectivity features between different ROIs. The time series features of the two ROIs are and The calculation formula of DTW distance is: Where ω is the time series matching path, d is the feature distance measurement function, and the spatial feature vector is constructed.

6. A method for analyzing nervous system images according to claim 1, characterized in that: In the above S200, the spatiotemporal joint feature analysis utilizes the spatiotemporal correlation formula to mine the correlation between the spatiotemporal feature vector and its surrounding neighborhood in the spatiotemporal dimension. The formula is: Where: C ST (x,y,z) represents the spatiotemporal correlation eigenvalue at the spatial position (x,y,z), N(x,y,z) is the neighborhood of the spatial position (x,y,z), and F ST,k,n (x, y, z) is the spatiotemporal feature value of the kth region of interest at the spatial position (x, y, z) in the nth time window, is the mean of the spatiotemporal features of the kth ROI at the spatial position (x, y, z), and is calculated as follows:

7. A method for analyzing nervous system images according to claim 1, characterized in that: In the S300, the training of the diagnostic model in the construction of the multi-disease joint diagnostic model is to input the training set data into the constructed model and use the back propagation algorithm to train the model. Suppose the training set data is F i is the characteristic data of the i-th sample, d i is the corresponding disease type label, and the model prediction output is Calculate the cross entropy loss function, the formula is: Calculate the loss function for θ k The gradient of is: The specific gradient expression is obtained by gradually expanding the calculation through the chain rule, and the model parameters are updated using stochastic gradient descent. The formula is: η is the learning rate, the number of training rounds is set to E, and after each round of training, the model performance is evaluated on the validation set.

8. The method for analyzing nervous system images according to claim 1, wherein: In the step S400, a prediction model is constructed by using a treatment effect prediction formula in the treatment effect prediction, and the formula is: Where: E(r|F,d) represents the probability value of the treatment outcome indicator r predicted based on the image feature F(x,y,z,t) and the disease type d, β0, β1, β2 are coefficients determined by clinical data training, δ r is the adjustment coefficient related to the characteristics of the treatment outcome indicator r itself.

9. A method for analyzing nervous system images according to claim 1, characterized in that: In the above-mentioned S500, the intelligent decision-making algorithm is used to generate a preliminary personalized treatment plan based on the network in the personalized treatment plan recommendation. The formula is: Where: T p represents the preliminary personalized treatment plan, n is the number of treatment plans in the knowledge graph that are related to the current patient's disease and meet the screening conditions, and w i is the initial weight of the i-th treatment plan in the knowledge graph, R i (d) is the adjustment factor based on disease matching, calculated as: in is the patient's disease feature vector, It is treatment plan T i The characteristic vector of the applicable disease, α is the adjustment parameter, C i (P) is an adjustment factor based on individual patient characteristics, K i (E) is an adjustment factor based on the predicted treatment effect, calculated as: Where β is the adjustment parameter, E i Predicted value of treatment effect, T i is the i-th treatment plan in the knowledge graph.

10. A nervous system image analysis device, applicable to a nervous system image analysis method according to any one of claims 1 to 9, characterized in that: The device includes: a multimodal data fusion acquisition module, a spatiotemporal joint feature analysis module, a multi-disease joint diagnosis model construction module, a treatment effect prediction module, and a personalized treatment plan recommendation module. The multimodal data fusion acquisition module collects the patient's fMRI, MRS, DTI, and PET modal neurological imaging data through the hospital's internal system, transmits them to the central data processing unit in real time, repairs problematic data, and integrates the data using a data fusion formula; The spatiotemporal joint feature analysis module pre-processes the fused dynamic nervous system image data to remove artifacts and noise, extracts the temporal and spatial features of the data, fuses them to construct a spatiotemporal feature vector, and uses the spatiotemporal correlation formula to mine its correlation with the surrounding neighborhood; The multi-disease joint diagnosis model construction module collects clinical information data and spatiotemporal joint feature analysis data, divides them into test sets and validation sets, constructs a model using a disease discrimination formula, trains the model using a backpropagation algorithm using the training set, and evaluates the model performance using the validation set; The treatment effect prediction module extracts treatment effect-related features from multimodal imaging data and clinical information to construct a vector, builds a model based on the treatment effect prediction formula, adjusts hyperparameters using grid search, and verifies with independent data. It inputs new patient characteristics to obtain prediction results for the effects of different treatment methods. The personalized treatment plan recommendation module: constructs a knowledge graph for the treatment of neurological diseases, integrates diagnosis, prediction results and patient clinical information to form a personalized treatment information network, uses an intelligent decision-making algorithm to generate a preliminary plan, and continuously collects feedback during treatment to update the network optimization plan.

Images