Parkinson's disease motion symptom auxiliary diagnosis system based on artificial intelligence

Abstract

The invention relates to a Parkinson's disease motion symptom auxiliary diagnosis system based on multi-modal artificial intelligence, and belongs to the field of medical AI auxiliary diagnosis. The system integrates multi-modal image data, and realizes common symptom classification and severity degree scoring of four motion symptoms of Parkinson's disease through preprocessing, multi-modal feature extraction combined with a 3D CNN branch and a Vision Transform branch, causal feature screening based on an IAMB algorithm, and a multi-modal fusion and multi-task learning module. The system can effectively fuse structure and functional image information, improves diagnosis accuracy and interpretability, is especially suitable for early symptom identification and quantitative evaluation, and has high clinical practical value.

Description

Technical Field

[0001] This invention belongs to the field of medical artificial intelligence-assisted diagnostic devices, specifically relating to an artificial intelligence-based auxiliary diagnostic system for motor symptoms of Parkinson's disease. Background Technology

[0002] Parkinson's disease (PD) is a common neurodegenerative disorder characterized by motor symptoms such as resting tremor, bradykinesia, rigidity, and postural instability, severely impacting patients' motor function and quality of life. With the increasing aging of the global population, the prevalence of PD is rising annually; statistics show that there are nearly 8.5 million patients worldwide, and the disease is showing a trend towards affecting younger people. Early and accurate diagnosis and assessment of the severity of motor symptoms are crucial for slowing disease progression and developing individualized treatment plans.

[0003] Currently, the clinical diagnosis of Parkinson's disease mainly relies on physicians' subjective assessment of patients' motor symptoms (such as the Unified Parkinson's Disease Rating Scale, UPDRS) and neuroimaging examinations (such as MRI and SPECT / PET-CT). However, existing diagnostic methods have significant limitations: insufficient utilization of multimodal data; while MRI can reflect changes in brain structure (such as substantia nigra atrophy and white matter fiber tract damage), and SPECT / PET-CT can show dopamine neurotransmitter dysfunction, traditional methods often independently analyze single-modal data, failing to fully explore the complementary information between structural and functional images, resulting in limited diagnostic sensitivity and specificity; symptom assessment is highly subjective; clinical scores of motor symptoms (such as UPDRS sub-items) depend on physician experience, exhibiting inter-rater variability, and making it difficult to quantify the dynamic changes of subtle early symptoms; and poor model interpretability; while existing deep learning models (such as 3D CNN and Transformer) have improved diagnostic accuracy, they remain "black boxes." The model struggles to locate key brain regions associated with symptoms (such as the basal ganglia and substantia nigra), leading to insufficient trust from clinicians and patients in the model's decision-making. Furthermore, the lack of causal relationship mining results in complex causal relationships between motor symptoms of Parkinson's disease and brain region structure and neurotransmitter function. Traditional feature screening methods (such as correlation analysis) cannot effectively identify biomarkers with causal effects (such as the association between dopamine transporter density and tremor), thus limiting the model's generalization ability.

[0004] In recent years, artificial intelligence technology has been gradually applied in the diagnosis of Parkinson's disease. For example, deep learning models based on monomodal MRI can help distinguish Parkinson's disease from other movement disorders. However, there are problems such as the immaturity of multimodal fusion technology. Existing fusion methods (such as early feature stitching and late decision fusion) do not fully consider the spatiotemporal characteristics of different modal data (such as the structural stability of MRI and the metabolic dynamics of PET), resulting in information redundancy or loss. The multi-task prediction ability is insufficient. Most models can only determine the presence of symptoms and cannot simultaneously output symptom severity scores, which is difficult to meet the clinical needs for quantitative assessment of the condition. The performance deteriorates in small sample scenarios. The cost of labeling Parkinson's disease image data is high and the sample size is limited. Traditional models are prone to overfitting in small sample scenarios and have insufficient generalization ability for rare symptom combinations.

[0005] Therefore, there is an urgent need for an auxiliary diagnostic system that can integrate multimodal image data, mine causal relationship features, and simultaneously achieve symptom comorbidity classification and quantitative scoring. Summary of the Invention

[0006] [Technical Issues] The technical problem to be solved by this invention is: how to overcome the limitations of existing Parkinson's disease diagnostic methods, such as insufficient utilization of multimodal data, strong subjectivity of assessment, poor model interpretability, and lack of causal relationship mining, so as to provide an auxiliary diagnostic system that can integrate multimodal images, mine causal features, and simultaneously realize symptom coexistence classification and severity quantification scoring.

[0007] [Technical Solution] To address the above problems, this invention provides an artificial intelligence-based auxiliary diagnostic system for motor symptoms of Parkinson's disease.

[0008] In a first aspect, the present invention provides an artificial intelligence-based auxiliary diagnostic system for motor symptoms of Parkinson's disease, the system comprising: The data acquisition module is configured to acquire multimodal image data of Parkinson's disease patients, including magnetic resonance imaging (MRI) data and nuclear medicine imaging (SPECT) or PET-CT data; the MRI data includes T1-weighted images, T2-weighted images, diffusion tensor imaging (DTI), and functional magnetic resonance imaging (fMRI). The data preprocessing module is configured to perform standardization and spatial registration processing on the multimodal image data; The Parkinson's disease motor symptom prediction module is configured to output the comorbidity classification results of four motor symptoms of Parkinson's disease patients and the severity score of each motor symptom based on preprocessed multimodal image data. The four motor symptoms include resting tremor, bradykinesia, rigidity and postural balance disorder; wherein the severity score is matched with the score of the corresponding item in the Unified Parkinson's Disease Rating Scale. The Parkinson's disease motor symptom prediction module includes a multimodal feature extraction module, a causal feature screening module, and a multimodal fusion and multi-task learning module; The multimodal feature extraction module is configured to extract features from preprocessed multimodal image data; the causal feature filtering module is configured to filter the extracted features and retain features that have a strong predictive association with the four motor symptoms; the multimodal fusion and multi-task learning module is configured to integrate the filtered features and output comorbidity classification results and severity scores through multi-task learning.

[0009] Optionally, the multimodal feature extraction module includes a 3D CNN branch and a Vision Transformer branch; the 3D CNN branch includes a DenseNet-3D network and a ResNet-3D network combined with an SE module; the DenseNet-3D network extracts structural features from MRI data through stacked dense blocks, and the ResNet-3D network combined with an SE module captures deep structural features from MRI data through residual connections, and enhances the feature weights of key brain regions through the SE module; the key brain regions are determined based on medical prior knowledge, including the substantia nigra and basal ganglia, and the model further adaptively learns refined brain region subregions and feature weights based on the key brain regions through the SE module; the Vision Transformer branch is configured to segment 3D slices of SPECT or PET-CT data and fMRI data into non-overlapping 3D slice blocks, perform linear projection on each 3D slice block and add position encoding, and then capture the whole-brain metabolic patterns of SPECT or PET-CT data and the whole-brain functional connectivity patterns of fMRI data through a multi-head attention mechanism; The calculation method for feature extraction by the DenseNet-3D network is as follows: Let the input feature map be... After the first The output feature map is obtained after convolution operations on dense blocks. ,Right now: ,in, Indicates the first Convolution operation on dense blocks Indicates the previous The feature maps of the layers are stitched together; In the ResNet-3D network incorporating the SE module, the residual blocks are calculated as follows: ,in Represents residual mapping, As weight, For the input of the residual block, The output is the residual block; specifically, the SE module performs global average pooling to process the feature map. Convert to channel descriptor ,Right now: ,in , , These are the height, width, and depth of the feature map, respectively. Representation of feature map In coordinates The pixel value at that location is then used to generate channel attention weights through two fully connected layers and an activation function. The first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the Sigmoid activation function; Compared with the original feature map Element-wise multiplication yields a feature map with enhanced weights for key brain regions. The calculation process of the Vision Transformer branch is as follows: assuming the image is segmented into... Non-overlapping 3D slices Each slice block is obtained after linear projection. ,in It's a projection matrix; then position encoding is added. get In multi-head attention mechanisms, queries ,key Sum Through respectively , , The attention score was calculated. for: ,in, For key The dimension; the final output ,in This is the output of the Vision Transformer branch. Optionally, the causal feature selection module uses the IAMB algorithm for feature selection, specifically: Initialize Markov boundary set Let the set of features be empty. Target variable It is any one of the four motor symptoms; For each feature ,calculate and A subset of the currently determined Markov boundaries Conditional mutual information:

[0010] in, for , and The joint probability distribution, for under conditions and The conditional joint probability distribution, and They are respectively under conditions and The conditional probability distribution; like Greater than the preset threshold Then Add Markov boundary set This continues until all features are traversed, yielding the result that matches the target variable. A set of features that have local causal relationships.

[0011] Optionally, the preset threshold pass The cross-validation method was determined, among which... For integers greater than or equal to 2, by evaluating different The optimal threshold is selected based on the predictive performance of the model after feature filtering.

[0012] Optionally, the multimodal fusion and multi-task learning module includes a joint co-learning fusion submodule and a multi-task output layer submodule; The calculation process of the joint co-learning fusion submodule is as follows: the 3D CNN branch feature vectors output by the multimodal feature extraction module are processed... With Vision Transformer branch feature vectors The concatenation is performed to obtain the concatenated feature vector. ; through fully connected layers Perform a transformation to obtain a shared representation. ,in This is the weight matrix of the fully connected layer. The bias vector of the fully connected layer is used; an Excitation Network is introduced to... Channel attention weighting is performed to obtain weighted feature vectors. The Excitation Network comprises two fully connected layers: the first fully connected layer has neurons that share representations. One-quarter of the dimension, using the ReLU activation function; the number of neurons in the second fully connected layer is... Since the dimensions are the same, the Sigmoid activation function is used. The multi-task output layer submodule is configured to receive weighted features and perform symptom co-occurrence classification and symptom severity score regression tasks in parallel. The symptom co-occurrence classification task uses a fully connected layer with a Sigmoid activation function to output a four-dimensional binary vector to determine the presence of four symptoms. The symptom severity score regression task uses a fully connected layer with a linear activation function to output a four-dimensional continuous vector, corresponding to the severity scores of four types of symptoms.

[0013] Optionally, the Excitation Network is calculated as follows: Shared representation Perform global average pooling to obtain the channel descriptors, as shown in the following formula:

[0014] in for The characteristic length, For channel indexing; Will Input two fully connected layers, the output of the first fully connected layer is The output of the second fully connected layer is ,in , This is the weight matrix of the fully connected layer. , This is the bias vector of the fully connected layer; Will and Element-by-element multiplication yields .

[0015] Optionally, the symptom comorbidity classification task is implemented by: […]. The input is a fully connected layer, which is activated by the Sigmoid function and outputs a four-dimensional binary vector. ,Right now: ,in, The fully connected layer for the classification task corresponds to the first... The weight vector of the various movement symptoms, For the corresponding bias term of the fully connected layer in the classification task, It is the Sigmoid activation function. Indicates the first The motor symptoms are present. Indicates the first The motor symptoms are absent; The symptom severity score regression task is implemented as follows: The input is a fully connected layer, which is then processed by a linear activation function to output a four-dimensional continuous vector. ,Right now: ,in This is the weight matrix of the fully connected layer for the regression task. This represents the bias vector corresponding to the fully connected layer in the regression task. For the first The severity score of the motor symptoms; The total loss function for the multi-task learning is:

[0016] in, For binary cross-entropy loss, For mean square error loss, To balance the hyperparameters for classification and regression tasks; The calculation method is as follows:

[0017] in, For the first True classification labels for various motor symptoms For the first Predictive classification labels for various motor symptoms; The calculation method is as follows:

[0018] in, For the first The true score of the motor symptoms, For the first Predictive scores for various motor symptoms.

[0019] Optionally, the data preprocessing module specifically performs the following steps: N4 bias correction, MNI template registration, and voxel normalization on the MRI data sequentially; the target range for voxel normalization is [0,1]; standardized uptake values ​​(SUVs) are sequentially normalized on the SPECT or PET-CT data, and the normalized SPECT or PET-CT data are spatially registered with the preprocessed MRI data; the MNI template is the MNI152 template; and the voxel normalization uses maximum-minimum normalization.

[0020] Secondly, the present invention provides a method for operating a Parkinson's disease motor symptom auxiliary diagnostic system, comprising: Step 1: Acquire multimodal imaging data of Parkinson's disease patients, including SPECT or PET-CT data and MRI data, including T1-weighted images, T2-weighted images, diffusion tensor imaging (DTI), and functional magnetic resonance imaging (fMRI). Step 2: Preprocess the multimodal image data. Perform N4 bias correction, MNI template registration, and voxel normalization within the [0,1] range on the MRI data in sequence. Perform SUV normalization on the SPECT or PET-CT data in sequence, and spatially register the normalized SPECT or PET-CT data with the preprocessed MRI data. Step 3: Process and predict the preprocessed multimodal image data, extract the structural and functional metabolic features of the preprocessed image data; screen the extracted features and retain those with local causal relationship with the four motor symptoms; fuse the screened features and perform multi-task learning to output classification results representing symptom coexistence and scoring results representing symptom severity.

[0021] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described operation method.

[0022] [Beneficial Effects] (1) The multimodal feature extraction module of this invention adopts a dual-branch feature extraction architecture combining 3D CNN and Vision Transformer, realizing multi-dimensional and in-depth feature mining, and making full use of the global information of functional images. DenseNet-3D and ResNet-3D combined with SE module can efficiently extract local detailed features of MRI structural images, such as substantia nigra atrophy and white matter fiber tract integrity. The adaptive channel weighting of SE module significantly enhances the feature expression of key brain regions that are crucial for the diagnosis of Parkinson's disease and improves the model's sensitivity to subtle structural changes. The Vision Transformer branch, through its powerful global modeling capabilities, effectively captures whole-brain and long-range dependent patterns in SPECT / PET-CT metabolic images and fMRI functional images, such as abnormal dopamine neurotransmitter distribution and changes in whole-brain functional connectivity, making up for the shortcomings of traditional CNN in processing the global characteristics of functional images.

[0023] (2) The causal feature screening module of this invention innovatively introduces the IAMB algorithm, enhancing the interpretability and reliability of the model, and improving its generalization ability and robustness. This module can screen out biomarkers with local causal relationships with four motor symptoms from massive features, such as the association between substantia nigra volume and tremor, and the association between basal ganglia metabolism and bradykinesia, rather than just related features. This shifts the model's decision-making basis from a "black box" to a "white box," providing clinicians with credible decision-making basis and pointing out key imaging evidence leading to symptoms. At the same time, by eliminating a large number of redundant features and noise unrelated to symptoms, the risk of model overfitting is significantly reduced. This allows this invention to maintain excellent performance even in learning scenarios like Parkinson's disease where labeled data is scarce and samples are small.

[0024] (3) The multimodal fusion and multi-task learning module of this invention adopts joint co-learning fusion supplemented by the Excitation Network to perform multi-task learning in parallel, realizing information complementarity and adaptive fusion, which can meet the dual clinical needs in one stop. The Excitation Network can adaptively weigh the importance of MRI structure and SPECT / PET-CT function and fMRI metabolic modality features through the channel attention mechanism, giving higher weight to the modality features most relevant to the current diagnostic task, avoiding information redundancy or loss caused by simple splicing or fusion. The multi-task output layer performs "symptom comorbidity classification" and "symptom score regression" at the same time, which enables the system to output two key clinical conclusions at one time: "which symptoms the patient has at the same time" and "how severe each symptom is", directly aligning with the assessment needs of the UPDRS scale, providing unprecedented comprehensive quantitative assessment results, and greatly improving the efficiency and practicality of clinical applications.

[0025] In summary, this invention achieves comprehensive mining of deep features of structural and functional images through a multimodal feature extraction module; enhances the interpretability, robustness, and generalization ability of the model through a causal feature selection module; and achieves adaptive and effective fusion of information and simultaneous output of classification and quantitative scoring results through a multimodal fusion and multi-task learning module. The entire system works collaboratively to successfully construct an accurate, reliable, interpretable, and fully clinically-compliant auxiliary diagnostic platform for motor symptoms of Parkinson's disease. It effectively addresses many pain points in existing technologies, such as insufficient multimodal utilization, "black box" models, strong subjective assessment, and inability to quantify scoring. While improving diagnostic accuracy, it possesses high clinical practical value and promising prospects for widespread application. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 The basic operation flowchart of the auxiliary diagnostic system for motor symptoms of Parkinson's disease provided by the present invention is shown.

[0028] Figure 2 This is a schematic diagram of the structural composition of the auxiliary diagnostic system for motor symptoms of Parkinson's disease provided by the present invention.

[0029] Figure 3 This is a schematic diagram of the structure of the multimodal image preprocessing module provided by the present invention.

[0030] Figure 4 This is a schematic diagram of the structure of the 3D CNN branch provided by the present invention.

[0031] Figure 5 This is a schematic diagram of the Vision Transformer branch provided by the present invention.

[0032] Figure 6 A schematic diagram of the causal feature screening module provided by the present invention.

[0033] Figure 7 This is a schematic diagram of the structure of the multimodal fusion and multi-task learning module provided by the present invention.

[0034] Figure 8 The original MRI image of the brain of a Parkinson's disease patient provided by this invention.

[0035] Figure 9 The original SPECT / PET-CT image of the brain of a Parkinson's disease patient provided by this invention. Detailed Implementation

[0036] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Example 1: This embodiment provides an artificial intelligence-based auxiliary diagnostic system for motor symptoms of Parkinson's disease, the basic operation process of which is as follows: Figure 1 As shown, the system structure is as follows: Figure 2 As shown, the system includes: The data acquisition module is configured to retrieve raw multimodal imaging data of Parkinson's disease patients from the PPMI (Program for Progression Biomarkers in Parkinson's Disease) database, including magnetic resonance imaging (MRI) data and nuclear medicine imaging data. The MRI data includes T1-weighted images, T2-weighted images, diffusion tensor imaging (DTI), and functional magnetic resonance imaging (fMRI) data, stored in NIfTI format. The nuclear medicine imaging data includes SPECT or PET-CT data (such as dopamine transporter imaging (DAT-SPECT) or glucose metabolism imaging (FDG-PET), stored in DICOM format, which includes standardized uptake value (SUV) information.

[0038] The data preprocessing module is configured to perform standardization and spatial registration on the raw multimodal image data. See [link / details]. Figure 3 Regarding MRI data: First, the N4ITK algorithm is used for N4 bias correction to eliminate intensity deviations caused by magnetic field inhomogeneity. This process involves iteratively estimating the bias field. For each voxel gray value Perform correction: The corrected data were then used for MNI template registration using the SyN (Symmetric Normalization) algorithm from the ANTs toolkit, non-linearly registering individual brain structures to the MNI152 standard space (resolution ). Specifically, this is achieved by optimizing translation, rotation, scaling, and nonlinear deformation parameters to match the coordinates of each brain region with a standard template. The registered MRI data is then voxel-normalized, scaling the grayscale values ​​to the range [0, 1], using the following formula:

[0039] in, and These are the minimum and maximum voxel values ​​of the MRI sequence, respectively, to ensure the numerical stability of the input data during subsequent 3D CNN training.

[0040] For SPECT / PET-CT data: First, the SUV uptake value is normalized using the standardized uptake value. The SUV calculation formula is:

[0041] This formula converts uptake values ​​into standardized values, highlighting areas of abnormal distribution of neurotransmitters, such as dopamine. Spatial registration is then performed, using the MRI-registered MNI space as a reference. Affine transformation combined with nonlinear registration ensures that the coordinates of each voxel in the SPECT / PET-CT scan are perfectly aligned with the MRI, facilitating subsequent multimodal feature fusion.

[0042] After preprocessing, the MRI data is kept as a 3D voxel matrix with a size of 176×240×256, which is used for subsequent 3DCNN to extract structural features; the SPECT / PET-CT data are used to extract the region of interest (ROI), such as the mean SUV of the basal ganglia and substantia nigra to form feature vectors, which are used for subsequent Vision Transformer branches to capture whole-brain metabolic patterns or as the raw input of the causal feature screening module.

[0043] The entire preprocessing process eliminates data bias, unifies spatial coordinates and numerical ranges, and provides standardized, highly comparable, high-quality data for the subsequent multimodal feature extraction module, ensuring that the complementary information between structural MRI data and functional SPECT / PET-CT data is effectively preserved and utilized.

[0044] The Parkinson's disease motor symptom prediction module is configured to output the comorbidity classification results of four motor symptoms in Parkinson's disease patients and the severity score of each motor symptom based on preprocessed multimodal image data. The four motor symptoms include resting tremor, bradykinesia, rigidity, and postural balance disorder. The Parkinson's disease motor symptom prediction module includes a multimodal feature extraction module, a causal feature screening module, and a multimodal fusion and multi-task learning module.

[0045] The multimodal feature extraction module adopts a dual-branch architecture, including a 3D CNN branch and a Vision Transformer branch, and is configured to extract features from preprocessed multimodal image data.

[0046] See Figure 4 The 3D CNN branch includes the DenseNet-3D network and the ResNet-3D network, which incorporates a squeeze-and-excitation (SE) module. The DenseNet-3D network extracts structural features from MRI data by stacking dense blocks; its computation method is as follows: Let the input feature map be After the first The output feature map is obtained after convolution operations on dense blocks. ,Right now:

[0047] in, Indicates the first Convolution operation on dense blocks Indicates the previous The feature maps of the layers are stitched together.

[0048] A ResNet-3D network with an SE module captures deep structural features from MRI data through residual connections. The SE module further enhances the feature weights of key brain regions, which are identified based on prior medical knowledge of Parkinson's disease, including the substantia nigra (a core brain region for dopaminergic neuronal degeneration) and the basal ganglia (a key brain region for motor circuit regulation). These brain regions are pre-determined as core pathologically relevant areas of Parkinson's disease based on medical research. Building upon this, the SE module in the model further adaptively learns refined subregions and their feature weights within these key brain regions through channel attention mechanisms, achieving an organic combination of prior knowledge and data-driven approaches.

[0049] The residual block is calculated as follows:

[0050] in, Represents residual mapping, As weight, For the input of the residual block, This is the output of the residual block; The SE module specifically performs global average pooling on the feature map. Convert to channel descriptor ,Right now:

[0051] in, , , These are the height, width, and depth of the feature map, respectively. Representation of feature map In coordinates The pixel value at that location is then used to generate channel attention weights through two fully connected layers and an activation function. ,Will Compared with the original feature map Element-wise multiplication yields a feature map with enhanced weights for key brain regions. The first fully connected layer uses the ReLU activation function to introduce a non-linear transformation to enhance feature representation. The second fully connected layer uses the Sigmoid activation function to map channel attention weights to the 0-1 range to reasonably allocate the attention weights of different channels.

[0052] See Figure 5 The Vision Transformer branch is configured to segment 3D slices of nuclear medicine imaging SPECT or PET-CT data and functional magnetic resonance imaging (fMRI) data into non-overlapping 3D slice blocks. Each 3D slice block is linearly projected and its location encoded. Then, a multi-head attention mechanism is used to capture whole-brain metabolic patterns from SPECT or PET-CT data and whole-brain functional connectivity patterns from fMRI data. The computation process is as follows: Assuming the image is segmented into Non-overlapping 3D slices Each slice block is obtained after linear projection. ,in It's a projection matrix; then position encoding is added. get In multi-head attention mechanisms, queries ,key Sum Through respectively , , The attention score was calculated. for

[0053] in, For key The dimension; the final output ,in This is the output of the Vision Transformer branch. The multimodal feature extraction module focuses on mining features from different dimensions in multimodal image data, providing a rich information foundation for subsequent analysis.

[0054] The causal feature selection module employs the Incremental Association Markov Boundary (IAMB) algorithm, which is configured to filter extracted features, retain those with local causal relationships to the four motor symptoms, remove irrelevant noise, and enhance the model's sensitivity to biomarkers related to motor symptoms.

[0055] In practice, the multimodal feature extraction module extracted a total of 128 to 512 features from MRI, SPECT / PET-CT, and fMRI data. After processing by the causal feature selection module using the IAMB algorithm, features strongly associated with the four motor symptoms were retained, reducing the number of features to 32 to 64, significantly reducing feature dimensionality and improving model efficiency.

[0056] In this embodiment, local causality refers to a potential causal relationship between imaging features of specific brain regions and motor symptoms, as confirmed by medical research. It is not absolute causality in a philosophical sense, but rather a strong predictive association verified at the machine learning level using the IAMB algorithm. This association possesses strong correlation, strong predictability, and can exclude spurious associations. Prior medical knowledge includes core pathological brain regions associated with Parkinson's disease, such as the substantia nigra (the core brain region for dopaminergic neuron degeneration) and the basal ganglia (a key brain region for motor circuit regulation).

[0057] See Figure 6 The IAMB algorithm is as follows: Initialize Markov boundary set Let the set of features be empty. Target variable It is any one of the four motor symptoms; For each feature ,calculate and A subset of the currently determined Markov boundaries Conditional mutual information:

[0058] in, for , and The joint probability distribution, for under conditions and The conditional joint probability distribution, and They are respectively under conditions and The conditional probability distribution of ; where, Representing a subset of Markov boundaries The specific value instance refers to the specific content of the currently determined candidate feature set related to the target symptom. This is achieved by calculating within a given subset. Values Under the condition, characteristics With target symptoms The strength of the correlation between features is used to filter out features with strong causal relationships. For example, when calculating the conditional mutual information between substantia nigra volume features and resting tremor symptoms, a subset... It may include established relevant features such as basal ganglia metabolic values ​​and white matter fiber integrity. This refers to the specific numerical combination of these features in the current calculation.

[0059] like Greater than the preset threshold Then Add Markov boundary set This continues until all features are traversed, yielding the result that matches the target variable. A set of features that have local causal relationships.

[0060] Preset threshold pass The cross-validation method was used to determine the optimal model based on its predictive performance (such as classification accuracy, regression correlation, etc.) of Parkinson's motor symptoms after feature selection. The value is then used consistently in subsequent models. It is not dynamically and adaptively calculated at runtime, but remains unchanged after being determined during the training phase. For example, through 5-fold cross-validation testing. Multiple candidate values ​​ranging from 0.01 to 0.1 were considered, and the value that resulted in the highest prediction accuracy for resting tremor symptoms was ultimately selected. As a fixed threshold.

[0061] See Figure 7 The multimodal fusion and multi-task learning module is configured to integrate the filtered features and output comorbidity classification results and severity scores through multi-task learning. The multimodal fusion and multi-task learning module includes a joint co-learning fusion submodule and a multi-task output layer submodule.

[0062] The multimodal fusion and multi-task learning module first processes features through a joint co-learning fusion submodule: It first concatenates the feature vectors output from the 3D CNN branch and the Vision Transformer branch. Let the feature vector output from the 3D CNN branch be:

[0063] The feature vector output by the Vision Transformer branch is:

[0064] The concatenated feature vector is:

[0065] Next, the concatenated feature vectors are processed through a fully connected layer. After processing, the computation of the fully connected layer can be expressed as:

[0066] in, It is the weight matrix of the fully connected layer. The bias vector of the fully connected layer. This step enables further transformation and combination of features to generate shared representations. .

[0067] To highlight modal information related to motor symptoms, an Excitation Network (EN) was introduced to address shared representations. Channel attention weighting is applied, and the EN network specifically includes two fully connected layers: the number of neurons in the first fully connected layer is set to share representations. The dimension is reduced to 1 / 4, and a non-linear transformation is introduced using the ReLU activation function to enhance feature representation; the number of neurons in the second fully connected layer is... While maintaining consistent dimensions, the Sigmoid activation function is used to map the output to the 0-1 range, generating channel attention weights, specifically: First, the shared representation Perform global average pooling to obtain the channel descriptors, as shown in the following formula:

[0068] in, for The characteristic length, Index the channel; then, The input is fed into two fully connected layers, and the output of the first fully connected layer is:

[0069] The output of the second fully connected layer is:

[0070] in, , It is the weight matrix of the fully connected layer. , It is the bias vector.

[0071] Finally, and Element-wise multiplication yields the weighted eigenvector. This increases the weight of key features and highlights modal information related to motor symptoms.

[0072] Subsequently, the multi-task output layer submodule receives the weighted features and performs symptom comorbidity classification and symptom severity score regression tasks in parallel.

[0073] The symptom comorbidity classification task is a multi-label classification task. The weighted feature vectors... The input is fed into a fully connected layer, whose output dimension is 4, corresponding to the four motor symptoms. Then, the output is mapped to the [0, 1] interval using a sigmoid activation function, resulting in a four-dimensional binary vector. ,in Indicates the first Do these symptoms exist? Indicates the first The motor symptoms are present. Indicates the first The type of motor symptom is absent). For the first Output of symptoms:

[0074] in, The fully connected layer corresponding to the classification task is the first... The weight vector of each symptom Bias terms of fully connected layers for classification tasks The activation function is Sigmoid. The loss function for this task is binary cross-entropy loss, calculated as follows:

[0075] in, It is the first True label of symptoms ； For the first Predictive classification labels for various motor symptoms.

[0076] The implementation of the symptom severity score regression task is as follows: The weighted feature vector is then used... The input is fed into a fully connected layer, and the output dimension is 4, corresponding to the specific scores of the four symptoms. Here, a linear activation function is used, meaning the output is directly:

[0077] in, It is the weight matrix of the fully connected layer for the regression task. This is the bias vector of the fully connected layer in the regression task. The loss function for this task uses the mean squared error (MSE) loss, calculated as follows:

[0078] in, It is the first The true score of the symptoms It is the model's prediction of the first The scores for each symptom.

[0079] The total loss function for the entire multi-task learning process is

[0080] in, It is a hyperparameter used to balance the weights of classification and regression tasks, and its optimal value can usually be determined by methods such as cross-validation.

[0081] The multimodal fusion and multi-task learning module effectively integrates the features of multimodal image data through the above methods, and simultaneously completes two tasks: symptom comorbidity classification and symptom severity score regression. This provides comprehensive and accurate information for the auxiliary diagnosis of Parkinson's disease, helping doctors to formulate more scientific and reasonable treatment plans.

[0082] This embodiment is based on multimodal imaging data (MRI data) of 200 Parkinson's disease patients (120 in the prodromal stage and 80 in the clinical stage) from the PPMI database; see example. Figure 8 With SPECT / PET-CT data, see example. Figure 9 The technical effects of the system proposed in this invention are verified and explained. Hierarchical 10-fold cross-validation is used to compare the system with a single-modal model, a traditional multimodal model, and the causal feature selection system (MM-CFM) based on multimodal fusion and multi-task learning proposed in this invention.

[0083] The comparison results of the symptom comorbidity classification task are as follows: Table 1: Performance Comparison Results of Symptom Comorbidity Classification Task

[0084] The results are shown in Table 1. The average accuracy of the single-modal MRI model (3D DenseNet) was 89.2%, but due to the difficulty of traditional CNNs in capturing abnormalities in whole-brain functional networks, the model had a high rate of missed diagnoses of postural balance disorders (recall rate 78.5%). The accuracy of the multimodal model with SPECT / PET-CT (without causal screening) improved to 92.3%, but the AUC was only 0.912, indicating insufficient discriminative power for symptom coexistence. The system of this invention (MM-CFM) achieved an accuracy of 95.6% and an AUC of 0.968 by screening causal features (such as the association between substantia nigra volume and tremor, and the association between basal ganglia SUV and myotonia) through IAMB screening.

[0085] The comparison results of the symptom severity score regression task are as follows: Table 2: Performance Comparison Results of Symptom Severity Score Regression Task

[0086] The results are shown in Table 2. In the regression results of the UPDRS sub-item scores, the MAE of the single-modal SPECT / PET-CT model (ViT) was 1.23. Because functional imaging is not sensitive to early structural damage, the scoring error for bradykinesia was relatively large (RMSE=1.58). The traditional multimodal model reduced the MAE to 1.01 through feature stitching, but the RMSE still reached 1.32, indicating a risk of overfitting. The system of this invention, through the correlation between attention-weighted key features of the EN channel, such as the FA value of the DTI, and the postural balance disorder score, reduced the MAE to 0.82 and the RMSE to 1.15, with a Pearson correlation coefficient of 0.921, showing a high degree of consistency with clinical scores.

[0087] Experimental results show that the system of this invention significantly outperforms single-modal and traditional multimodal methods in both symptom comorbidity classification and symptom severity score regression tasks. This is primarily due to the effective noise filtering of the IAMB algorithm, the enhancement of key modalities through EN channel attention, and the synergistic optimization of multi-task learning to meet the dual clinical needs. This system provides an interpretable and robust solution for the accurate diagnosis of motor symptoms in Parkinson's disease, with significant advantages, especially in the identification of comorbidities in the early prodromal stage.

[0088] Example 2: This embodiment provides an operating method for an auxiliary diagnostic system for motor symptoms of Parkinson's disease, including: Step 1: Obtain multimodal imaging data of Parkinson's disease patients from the PPMI database. The multimodal imaging data includes SPECT or PET-CT data and MRI data. The MRI data includes T1-weighted images, T2-weighted images, diffusion tensor imaging (DTI), and functional magnetic resonance imaging (fMRI). Step 2: Preprocess the multimodal image data. Perform N4 bias correction, MNI template registration, and voxel normalization within the [0,1] range on the MRI data in sequence. Perform SUV normalization on the SPECT or PET-CT data in sequence, and spatially register the normalized SPECT or PET-CT data with the preprocessed MRI data. Step 3: Process and predict the preprocessed multimodal image data, extract the structural and functional metabolic features of the preprocessed image data; screen the extracted features and retain those with local causal relationship with the four motor symptoms; fuse the screened features and perform multi-task learning to output classification results representing symptom coexistence and scoring results representing symptom severity.

[0089] Example 3: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described operation method.

[0090] From a clinical application perspective, the multimodal fusion and multi-task learning module of this invention achieves dual outputs of "comorbidity classification" and "quantitative scoring" for motor symptoms: by outputting binary vectors indicating the presence or absence of four symptoms through the Sigmoid activation function, and combining linear regression to output specific scores for corresponding UPDRS sub-items, it not only meets the clinical need for rapid judgment of symptom coexistence, but also provides objective and quantitative indicators for disease assessment, effectively reducing the human bias of traditional subjective scoring.

[0091] From the perspective of technological innovation and scalability, the system of this invention breaks through the limitations of traditional single-modal analysis. Its joint co-learning fusion strategy (such as feature splicing and EN attention weighting) and the design of a multi-task loss function that balances classification and regression tasks provide a universal framework for the auxiliary diagnosis of Parkinson's disease and other neurodegenerative diseases. With the accumulation of long-term follow-up data from databases such as PPMI and the integration of multi-center imaging data, the model can continuously improve its performance by updating training data and optimizing hyperparameters. It has broad application prospects in early diagnosis, efficacy evaluation, and the formulation of individualized treatment plans, and powerfully promotes the development of Parkinson's disease diagnosis and treatment towards precision and objectivity.

[0092] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. An artificial intelligence-based auxiliary diagnostic system for motor symptoms of Parkinson's disease, characterized in that, The system includes: The data acquisition module is configured to acquire multimodal image data of Parkinson's disease patients, including magnetic resonance imaging (MRI) data and nuclear medicine imaging (SPECT) or PET-CT data; the MRI data includes T1-weighted images, T2-weighted images, diffusion tensor imaging (DTI), and functional magnetic resonance imaging (fMRI). The data preprocessing module is configured to perform standardization and spatial registration processing on the multimodal image data; The Parkinson's disease motor symptom prediction module is configured to output the comorbidity classification results of four motor symptoms of Parkinson's disease patients and the severity score of each motor symptom based on preprocessed multimodal image data. The four motor symptoms include resting tremor, bradykinesia, rigidity and postural balance disorder; wherein the severity score is matched with the score of the corresponding item in the Unified Parkinson's Disease Rating Scale. The Parkinson's disease motor symptom prediction module includes a multimodal feature extraction module, a causal feature screening module, and a multimodal fusion and multi-task learning module; The multimodal feature extraction module is configured to extract features from preprocessed multimodal image data; the causal feature filtering module is configured to filter the extracted features and retain features that have a strong predictive association with the four motor symptoms; the multimodal fusion and multi-task learning module is configured to integrate the filtered features and output comorbidity classification results and severity scores through multi-task learning.

2. The system according to claim 1, characterized in that, The multimodal feature extraction module includes a 3D CNN branch and a Vision Transformer branch. The 3D CNN branch includes a DenseNet-3D network and a ResNet-3D network combined with an SE module. The DenseNet-3D network extracts structural features from MRI data by stacking dense blocks, while the ResNet-3D network combined with an SE module captures deep structural features from MRI data through residual connections and enhances the feature weights of key brain regions through the SE module. The key brain regions are determined based on medical prior knowledge and include the substantia nigra and basal ganglia. The model further adaptively learns refined brain region subregions and feature weights based on the key brain regions through the SE module. The Vision Transformer branch is configured to segment 3D slices of SPECT or PET-CT data and fMRI data into non-overlapping 3D slice blocks, perform linear projection on each 3D slice block and add position encoding, and then capture the whole-brain metabolic patterns of SPECT or PET-CT data and the whole-brain functional connectivity patterns of fMRI data through a multi-head attention mechanism. The calculation method for feature extraction by the DenseNet-3D network is as follows: Let the input feature map be... After the first The output feature map is obtained after convolution operations on dense blocks. ,Right now: ,in, Indicates the first Convolution operation on dense blocks Indicates the previous The feature maps of the layers are stitched together; In the ResNet-3D network incorporating the SE module, the residual blocks are calculated as follows: ,in Represents residual mapping, As weight, For the input of the residual block, The output is the residual block; specifically, the SE module performs global average pooling to process the feature map. Convert to channel descriptor ,Right now: ,in , , These are the height, width, and depth of the feature map, respectively. Representation of feature map In coordinates The pixel value at that location is then used to generate channel attention weights through two fully connected layers and an activation function. The first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the Sigmoid activation function; Compared with the original feature map Element-wise multiplication yields a feature map with enhanced weights for key brain regions. The calculation process of the Vision Transformer branch is as follows: assuming the image is segmented into... Non-overlapping 3D slices Each slice block is obtained after linear projection. ,in It's a projection matrix; then position encoding is added. get In multi-head attention mechanisms, queries ,key Sum Through respectively , , The attention score was calculated. for: ,in, For key The dimension; the final output ,in This is the output of the Vision Transformer branch.

3. The system according to claim 1, characterized in that, The causal feature selection module uses the IAMB algorithm for feature selection, specifically: initializing the Markov boundary set. Let the set be empty, and let the characteristic set be... Target variable It is any one of the four motor symptoms; for each feature ,calculate and A subset of the currently determined Markov boundaries Conditional mutual information: ,in, for , and The joint probability distribution, for under conditions and The conditional joint probability distribution, and They are respectively under conditions and The conditional probability distribution; if Greater than the preset threshold Then Add Markov boundary set This continues until all features are traversed, yielding the result that matches the target variable. A set of features that have local causal relationships.

4. The system according to claim 3, characterized in that, The preset threshold pass The cross-validation method was determined, among which... For integers greater than or equal to 2, by evaluating different The optimal threshold is selected based on the predictive performance of the model after feature filtering.

5. The system according to claim 1, characterized in that, The multimodal fusion and multi-task learning module includes a joint co-learning fusion submodule and a multi-task output layer submodule; The calculation process of the joint co-learning fusion submodule is as follows: the 3D CNN branch feature vectors output by the multimodal feature extraction module are processed... With Vision Transformer branch feature vectors The concatenation is performed to obtain the concatenated feature vector. ; through fully connected layers Perform a transformation to obtain a shared representation. ,in This is the weight matrix of the fully connected layer. The bias vector of the fully connected layer is used; an Excitation Network is introduced to... Channel attention weighting is performed to obtain weighted feature vectors. The Excitation Network comprises two fully connected layers: the first fully connected layer has neurons that share representations. One-quarter of the dimension, using the ReLU activation function; the number of neurons in the second fully connected layer is... Since the dimensions are the same, the Sigmoid activation function is used. The multi-task output layer submodule is configured to receive weighted features and perform symptom co-occurrence classification and symptom severity score regression tasks in parallel. The symptom co-occurrence classification task uses a fully connected layer with a Sigmoid activation function to output a four-dimensional binary vector to determine the presence of four symptoms. The symptom severity score regression task uses a fully connected layer with a linear activation function to output a four-dimensional continuous vector, corresponding to the severity scores of four types of symptoms.

6. The system according to claim 5, characterized in that, The calculation method of the Excitation Network is as follows: Shared representation Perform global average pooling to obtain the channel descriptors, as shown in the following formula: in for The characteristic length, For channel indexing; Will Input two fully connected layers, the output of the first fully connected layer is The output of the second fully connected layer is ,in , This is the weight matrix of the fully connected layer. , This is the bias vector of the fully connected layer; Will and Element-by-element multiplication yields .

7. The system according to claim 5, characterized in that, The symptom comorbidity classification task is implemented as follows: The input is a fully connected layer, which is activated by the Sigmoid function and outputs a four-dimensional binary vector. ,Right now: ,in, The fully connected layer for the classification task corresponds to the first... The weight vector of the various movement symptoms, For the corresponding bias term of the fully connected layer in the classification task, It is the Sigmoid activation function. Indicates the first The motor symptoms are present. Indicates the first The motor symptoms are absent; The symptom severity score regression task is implemented as follows: The input is a fully connected layer, which is then processed by a linear activation function to output a four-dimensional continuous vector. ,Right now: ,in, This is the weight matrix of the fully connected layer for the regression task. This represents the bias vector corresponding to the fully connected layer in the regression task. For the first The severity score of the motor symptoms; The total loss function for the multi-task learning is: ,in, For binary cross-entropy loss, For mean square error loss, To balance the hyperparameters of classification and regression tasks; The calculation method is as follows: in, For the first True classification labels for various motor symptoms For the first Predictive classification labels for various motor symptoms; The calculation method is as follows: in, For the first The true score of the motor symptoms, For the first Predictive scores for various motor symptoms.

8. The system according to claim 1, characterized in that, The specific process of the data preprocessing module is as follows: the magnetic resonance imaging (MRI) data are sequentially subjected to N4 bias correction, MNI template registration, and voxel normalization; the target range of the voxel normalization is [0,1]; the nuclear medicine imaging (SPECT) or PET-CT data are sequentially subjected to standardized uptake value (SUV) normalization, and the normalized SPECT or PET-CT data are spatially registered with the preprocessed MRI data; the MNI template is the MNI152 template; the voxel normalization uses maximum-minimum normalization.

9. A method for operating a Parkinson's disease motor symptom auxiliary diagnostic system, characterized in that, The methods include: Step 1: Acquire multimodal imaging data of Parkinson's disease patients, including SPECT or PET-CT data and MRI data, including T1-weighted images, T2-weighted images, diffusion tensor imaging (DTI), and functional magnetic resonance imaging (fMRI). Step 2: Preprocess the multimodal image data. Perform N4 bias correction, MNI template registration, and voxel normalization within the [0,1] range on the MRI data in sequence. Perform SUV normalization on the SPECT or PET-CT data in sequence, and spatially register the normalized SPECT or PET-CT data with the preprocessed MRI data. Step 3: Process and predict the preprocessed multimodal image data, extract the structural and functional metabolic features of the preprocessed image data; screen the extracted features and retain those that have a local causal relationship with the four motor symptoms; The filtered features are fused and multi-task learning is performed to output classification results representing symptom coexistence and scores representing symptom severity.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method as described in claim 9.

Images