DCM early noninvasive analysis method based on multi-radiomics and serum markers

Abstract

The invention relates to the technical field of medical diagnosis, and discloses a DCM early noninvasive analysis method based on multi-radiomics and serum markers. Collecting image data through a multi-modal medical imaging device, and collecting serum marker data through a blood detection device; respectively generating a radiomics feature set and a serum marker time sequence feature set by using a multi-scale feature extraction algorithm and a time sequence analysis model; fusing the features by adopting a dynamic weighted fusion strategy to generate a joint feature matrix; inputting the model into a pre-trained multi-task deep learning model to predict a DCM risk probability; and finally, based on a genetic algorithm, optimizing the diagnosis decision tree and outputting an early DCM diagnosis result. The method is noninvasive and accurate, and can effectively improve the early diagnosis accuracy of DCM.

Description

Technical Field

[0001] The present invention relates to the field of medical diagnostic technology, and in particular to an early non-invasive analysis method for DCM based on multi-imaging omics and serum markers. Background Art

[0002] Dilated cardiomyopathy (DCM) is a serious cardiovascular disease characterized by ventricular enlargement and decreased myocardial contractile function, which can lead to heart failure, arrhythmias, and even sudden death. Early diagnosis of DCM is particularly critical in children, as early intervention can significantly improve prognosis, quality of life, and survival. However, early diagnosis of DCM currently faces many challenges.

[0003] Traditional diagnostic methods rely primarily on invasive procedures, such as endomyocardial biopsy. While endomyocardial biopsy can provide a histological basis for diagnosis, it is invasive and can lead to complications such as infection, bleeding, and cardiac perforation, causing pain and risks to patients. Furthermore, this method is difficult to widely use for large-scale screening and repeated examinations in clinical practice.

[0004] When it comes to non-invasive diagnosis, single medical imaging technologies have limitations. While cardiac MRI can provide high-resolution information on cardiac structure and function, it lacks a comprehensive understanding of myocardial tissue characteristics. Echocardiography is simple to use and can dynamically observe cardiac activity in real time, but image quality is easily affected by factors such as patient size and lung function, resulting in limited diagnostic accuracy. Cardiac CT can clearly display cardiac and vascular structures, but it is inadequate for assessing myocardial function and tissue characteristics.

[0005] Serum marker testing is an important adjunct to the diagnosis of DCM, but most serum markers lack high specificity and sensitivity, and a single serum marker cannot accurately reflect the progression and severity of the disease. Furthermore, current clinical analysis of serum markers is often based on static test results, ignoring their dynamic changes over time, resulting in an inability to capture subtle early changes in the disease.

[0006] In terms of data processing and analysis, existing diagnostic models often fail to fully integrate the advantages of multi-source, heterogeneous data. Multimodal medical imaging data and serum marker data have different data structures and characteristics. Traditional methods struggle to effectively integrate these data and mine the potential diagnostic information, making it difficult to achieve ideal diagnostic accuracy and reliability.

[0007] Due to the lack of efficient, non-invasive early diagnosis methods, many DCM patients are not diagnosed until the disease has progressed to the middle or late stages, missing the optimal treatment opportunity. Therefore, developing an early, non-invasive analysis method based on multi-radiomic and serum markers has important clinical and social value for improving the accuracy of early diagnosis of DCM and improving patient prognosis. Summary of the Invention

[0008] The purpose of the present invention is to provide a non-invasive analysis method for early DCM based on multi-imaging omics and serum markers to solve the problems raised in the above background technology.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] A non-invasive early-stage DCM analysis method based on multi-radiomic and serum markers, comprising:

[0011] collecting imaging data of the pediatric patient using a multimodal medical imaging device, wherein the multimodal medical imaging includes cardiac MRI, echocardiography, and cardiac CT;

[0012] Collect serum marker data through blood testing equipment;

[0013] Performing radiomics feature extraction on the image data based on a multi-scale feature extraction algorithm to generate a radiomics feature set;

[0014] Performing dynamic change modeling on the serum marker data based on a time series analysis model to generate a serum marker time series feature set;

[0015] A dynamic weighted fusion strategy is used to perform cross-modal feature fusion on the radiomics feature set and the serum marker temporal feature set to generate a joint feature matrix;

[0016] Inputting the joint feature matrix into a pre-trained multi-task deep learning model, the multi-task deep learning model includes a classification module and a reconstruction module, wherein the classification module is used to perform a DCM classification task to predict the disease risk probability, and the reconstruction module is used to perform a feature reconstruction task to reconstruct the joint feature matrix, interactively learning the node embedding features of the classification module and the reconstruction module through a graph neural network architecture, and synchronously optimizing the two types of tasks based on a joint loss function to output the DCM risk probability;

[0017] A personalized diagnosis decision tree is constructed according to the DCM risk probability, and the branch threshold of the decision tree is optimized based on a genetic algorithm to generate an optimal diagnostic path; and an early DCM diagnosis result is output based on the optimal diagnostic path.

[0018] Optionally, performing radiomics feature extraction on the image data based on a multi-scale feature extraction algorithm to generate a radiomics feature set includes:

[0019] Perform three-dimensional wavelet transform on cardiac MRI data to extract texture features, morphological features and intensity distribution features in multiple frequency bands;

[0020] Dynamic time warping is performed on echocardiographic data to generate cardiac cycle aligned temporal features, and local texture features are extracted using the gray-level co-occurrence matrix.

[0021] performing vascular segmentation on cardiac CT data, extracting topological and geometric features based on the segmented vascular structure; and normalizing the texture features, morphological features, intensity distribution features, temporal features, local texture features, topological features, and geometric features to eliminate dimensional differences between modalities.

[0022] Principal component analysis was used to reduce the dimension of the standardized features, remove redundant features, and generate a radiomics feature set.

[0023] Optionally, performing dynamic change modeling on the serum marker data based on a time series analysis model to generate a serum marker time series feature set includes:

[0024] Collecting serum marker concentration data at multiple time points to construct a time series matrix; performing trend decomposition on the time series matrix using an autoregressive integrated moving average model to obtain trend terms, seasonal terms, and residual terms;

[0025] The time series data of multiple patients were aligned based on the dynamic time warping algorithm to eliminate the sampling time deviation between individuals. A time series prediction model based on the long short-term memory network was constructed, and the aligned time series data was input to predict the changes in marker concentrations at future time points.

[0026] The hidden state vector of the time series prediction model is extracted as a dynamic change feature; and a serum marker time series feature set is generated by combining the trend term, the seasonal term, the residual term and the dynamic change feature.

[0027] Optionally, the cross-modal feature fusion of the radiomics feature set and the serum marker time series feature set using a dynamic weighted fusion strategy to generate a joint feature matrix includes:

[0028] An inter-modality similarity metric function was constructed, and the correlation between radiomics features and serum marker features was calculated based on cosine similarity. A dynamic weight was assigned to each radiomics feature based on the correlation, with the weight value being positively correlated with the correlation degree. The serum marker temporal features were standardized and weighted, with the weight value determined based on the inverse of the feature variance.

[0029] The weighted imaging genomics features and the standardized weighted serum marker temporal features were horizontally spliced to generate a preliminary fusion matrix. Kernel principal component analysis was used to perform nonlinear dimensionality reduction on the preliminary fusion matrix to eliminate cross-modal redundant information and generate a joint feature matrix.

[0030] Optionally, the steps of constructing the multi-task deep learning model include:

[0031] Construct a patient feature graph, where nodes represent joint feature vectors of individual patients and edges represent clinical similarities between patients, calculated based on age, gender, and underlying disease matching.

[0032] A graph attention mechanism is used to aggregate node features and generate node embedding vectors. A classification task branch is designed to input the node embedding vectors into the fully connected layer and output the DCM risk probability.

[0033] Design a feature reconstruction task branch to reconstruct the original joint feature matrix from the node embedding vector through the decoder network;

[0034] A joint loss function is constructed, wherein the joint loss function includes a weighted sum of classification cross entropy loss and reconstruction mean square error loss; and an alternating optimization strategy is used to update network parameters of the classification task and the reconstruction task until the model converges.

[0035] Optionally, the optimizing the branch threshold of the decision tree based on the genetic algorithm to generate the optimal diagnostic path includes:

[0036] Each branch node of the diagnostic decision tree is encoded as a gene position, and the branch threshold is encoded as a gene value. The population is initialized and multiple individuals with different gene values are randomly generated. A fitness function is constructed, which is calculated based on the confusion matrix indicators of the diagnostic path and includes a weighted combination of sensitivity, specificity, and F1 score. A tournament selection strategy is used to screen high-fitness individuals, and a progeny population is generated through single-point crossover and mutation operations. The population is iteratively updated until the maximum evolutionary generation is reached, and the individual with the highest fitness is selected and decoded as the optimized branch threshold to generate the optimal diagnostic path.

[0037] Optionally, performing three-dimensional wavelet transform on cardiac MRI data to extract texture features, morphological features, and intensity distribution features in multiple frequency bands includes:

[0038] Cardiac MRI data were decomposed into low-frequency approximate components and high-frequency detail components using a three-dimensional discrete wavelet transform. Intensity distribution features based on grayscale histograms, including mean, variance, and skewness, were extracted from the low-frequency approximate components. Texture features based on Gabor filters, including energy, entropy, and contrast, were extracted from each high-frequency detail component. The three-dimensional volume, surface area, and sphericity index of the cardiac chambers were extracted based on morphological operations. The intensity distribution features, texture features, and morphological features were hierarchically integrated by frequency band to generate multi-scale imaging omics features.

[0039] Optionally, constructing a time series prediction model based on a long short-term memory network, inputting the aligned time series data, and predicting changes in marker concentrations at future time points includes:

[0040] The time series data is divided into sliding windows, each containing a sequence of marker concentrations at consecutive time points; a bidirectional long short-term memory network is constructed, which inputs the window sequence and outputs the concentration prediction value at the next time point; a self-attention mechanism is introduced in the hidden layer to dynamically adjust the weights of different time steps; the mean square error is used as the loss function, and the network parameters are updated through the time backpropagation algorithm; the early stopping method is used to prevent overfitting, and training is terminated when the validation set loss does not decrease for three consecutive times.

[0041] Optionally, aggregating node features using a graph attention mechanism to generate a node embedding vector includes:

[0042] The attention coefficient between nodes is calculated, and the attention coefficient is based on the weighted sum of the cosine similarity and clinical similarity of the node features. The attention coefficient is normalized by the Softmax function to generate normalized attention weights. The features of neighboring nodes are weighted summed according to the normalized attention weights to update the embedding vector of the central node. Multiple layers of graph attention layers are stacked, and the output of each layer is processed by Layer Normalization and ReLU activation function. Finally, the node embedding vector of the last layer is extracted as the input of the classification and reconstruction tasks.

[0043] Optionally, the adopting of a tournament selection strategy to screen individuals with high fitness and generating a progeny population through single-point crossover and mutation operations includes:

[0044] K individuals are randomly selected from the current population to form a tournament group, and the individual with the highest fitness is selected to enter the mating pool; the tournament selection is repeated until the mating pool is filled; the individuals in the mating pool are paired, and crossover points are randomly selected for gene exchange to generate offspring individuals; each gene position of the offspring individual is randomly mutated with a preset probability, and the mutation range is the feasible interval of the branching threshold; the individuals with the top 10% fitness in the parent population are retained and directly enter the next generation.

[0045] Compared with the prior art, the present invention has at least the following beneficial effects:

[0046] The non-invasive early diagnosis method for DCM based on multi-imaging omics and serum markers proposed in this invention improves the accuracy and efficiency of early diagnosis of DCM in many aspects.

[0047] In terms of data acquisition and integration, this method uses multimodal medical imaging equipment to acquire cardiac MRI, echocardiography, and cardiac CT imaging data, while simultaneously combining blood testing to obtain serum marker data, achieving comprehensive collection of multi-source heterogeneous data. This multimodal data fusion strategy overcomes the limitations of a single detection method and comprehensively utilizes the advantages of different data in reflecting cardiac structure, function, and disease characteristics. The high-resolution structural information of cardiac MRI, the dynamic functional information of echocardiography, the vascular characteristics of cardiac CT, and the biological information of serum markers complement each other, providing a rich basis for the accurate diagnosis of DCM.

[0048] In the imaging genomics feature extraction process, specialized multi-scale feature extraction algorithms are used for different imaging data. Three-dimensional wavelet transforms are performed on cardiac MRI to comprehensively extract texture, morphology, and intensity distribution features in multiple frequency bands, reflecting the characteristics of myocardial tissue from different dimensions. Echocardiography, through dynamic time warping and grayscale co-occurrence matrix analysis, obtains cardiac cycle-aligned temporal features and local texture features, effectively capturing subtle changes during cardiac motion. Cardiac CT extracts topological and geometric features through vascular segmentation, highlighting the association between vascular structure and disease. Through standardization and principal component analysis dimensionality reduction, inter-modality dimensional differences are eliminated, redundant features are removed, and a high-quality imaging genomics feature set is generated, improving data availability and analysis efficiency.

[0049] During the generation of serum marker time series features, serum marker concentration data from multiple time points are collected to construct a time series matrix. Trend decomposition is performed using an autoregressive integrated moving average model. A dynamic time warping algorithm is then used to align sampling time deviations between individuals. A time series prediction model based on a long short-term memory network is then used to predict future marker concentration changes, fully tapping into the dynamic changes in serum markers. This dynamic analysis of serum markers can more keenly capture disease trends and better reflect the true state of the disease than traditional static testing.

[0050] Cross-modal feature fusion employs a dynamic weighted fusion strategy, assigning dynamic weights based on the correlation between radiomics and serum marker features. This strategy also considers the variance of serum marker features to determine their weights, effectively integrating the strengths of both feature sets. The resulting joint feature matrix, generated after nonlinear dimensionality reduction using kernel principal component analysis, retains key diagnostic information while reducing cross-modal redundancy, providing more representative input for subsequent diagnostic models.

[0051] The multi-task deep learning model constructs a patient feature graph, aggregates node features using a graph attention mechanism, designs classification and reconstruction task branches, and simultaneously optimizes them using a joint loss function. The classification module predicts the risk of DCM, while the reconstruction module assists in model optimization by reconstructing the joint feature matrix. The two work together to improve model accuracy and stability. The graph neural network architecture enables interactive learning of node embedding features, fully exploiting clinical similarities between patients and further improving model performance.

[0052] A genetic algorithm optimizes the branch thresholds of the diagnostic decision tree, constructs a fitness function using confusion matrix metrics, and generates the optimal diagnostic path through multiple rounds of selection, crossover, and mutation. This approach can tailor diagnostic strategies to the specific characteristics of each patient, improving diagnostic accuracy and reliability while reducing misdiagnosis and missed diagnoses. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a diagram showing the working principle of the non-invasive early-stage DCM analysis method based on multi-imaging omics and serum markers according to the present invention;

[0054] Figure 2 Flowchart of cross-modal feature fusion;

[0055] Figure 3 Flowchart for building a multi-task deep learning model;

[0056] Figure 4 Flowchart for optimizing diagnostic decision tree based on genetic algorithm. DETAILED DESCRIPTION

[0057] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the 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.

[0058] See also Figure 1-4 The present invention provides a non-invasive analysis method for early DCM based on multi-imaging omics and serum markers, which specifically includes the following steps:

[0059] Data Collection: Children's imaging data is collected using multimodal medical imaging equipment, including cardiac MRI, echocardiography, and cardiac CT. These images present information about the heart's structure and function from different angles and in different ways. Blood testing equipment is also used to collect serum marker data, which reflects the patient's biochemical state and is crucial for disease diagnosis and monitoring.

[0060] Feature extraction: Radiomics features are extracted from the collected image data using a multi-scale feature extraction algorithm to generate a radiomics feature set. Radiomics features can more comprehensively and meticulously describe the information in the image, providing a rich data foundation for subsequent analysis.

[0061] Dynamic modeling: Using a time series analysis model to model the dynamic changes in serum marker data, we generate a time series feature set for serum markers. This step considers the changing trends of serum markers over time, helping to more accurately understand the progression of the disease.

[0062] Feature fusion: A dynamic weighted fusion strategy is used to perform cross-modal feature fusion of the radiomics feature set and the serum marker time series feature set to generate a joint feature matrix. This approach integrates the advantages of different modalities and improves the accuracy and reliability of the analysis.

[0063] Model prediction: The joint feature matrix is fed into a pre-trained multi-task deep learning model. This model consists of a classification module and a reconstruction module. The classification module performs the DCM classification task and predicts disease risk probability, while the reconstruction module performs the feature reconstruction task and reconstructs the joint feature matrix. A graph neural network architecture is used to interactively learn the node embedding features of the classification and reconstruction modules. Both tasks are simultaneously optimized based on a joint loss function, ultimately outputting the DCM risk probability.

[0064] Diagnostic decision-making: Based on the obtained DCM risk probability, a personalized diagnostic decision tree is constructed. Using a genetic algorithm, the branch thresholds of the decision tree are optimized to generate the optimal diagnostic path. Finally, based on the optimal diagnostic path, early DCM diagnosis results are output, providing strong support for clinical diagnosis.

[0065] The implementation of the present invention will be further described below with reference to Examples 1 to 5.

[0066] Example 1:

[0067] In this embodiment, the specific process of extracting radiomic features from image data based on a multi-scale feature extraction algorithm and generating a radiomic feature set is described in detail, especially the processing of cardiac MRI data.

[0068] First, a 3D wavelet transform is performed on the cardiac MRI data. Using a 3D discrete wavelet transform, the cardiac MRI data is decomposed into low-frequency approximate components and high-frequency detail components. The low-frequency approximate components retain the main contours and general features of the image, while the high-frequency detail components contain fine information such as edges and textures.

[0069] Extract intensity distribution features based on the grayscale histogram from the low-frequency approximate components. The grayscale histogram reflects the frequency of occurrence of different grayscale values in the image. Calculate the mean, which represents the average grayscale level of the image and reflects the overall brightness. The variance measures the dispersion of grayscale values relative to the mean; a larger variance indicates a more dispersed distribution of grayscale values. The skewness describes the asymmetry of the grayscale distribution and helps identify areas of abnormal grayscale distribution in the image.

[0070] Texture features are extracted from each high-frequency detail component using a Gabor filter. The Gabor filter is a commonly used filter in image processing that effectively extracts texture information. The extracted energy feature indicates the activity of the image texture; higher energy values indicate more pronounced texture. Entropy reflects the complexity of the texture; higher entropy values indicate more complex textures. Contrast reflects the degree of difference between different grayscale regions in the image; higher contrast values indicate clearer image details.

[0071] Morphological operations are used to extract the 3D volume, surface area, and sphericity index of the cardiac chambers. Morphological operations accurately extract the geometric features of the cardiac chambers by performing operations such as erosion and dilation on the image. The 3D volume reflects the size of the cardiac chamber, the surface area reflects the surface area of the cardiac chamber, and the sphericity index measures the degree to which the shape of the cardiac chamber resembles a sphere. These features are important for assessing cardiac structure and function.

[0072] The intensity distribution features, texture features, and morphological features extracted above are hierarchically integrated by frequency band to generate multi-scale radiomics features. The resulting features can comprehensively describe the information of cardiac MRI images from different scales and angles.

[0073] Dynamic time warping is performed on echocardiographic data. Because cardiac cycles may vary between patients, dynamic time warping can align echocardiographic data from different patients in time, generating temporal features that align cardiac cycles. Local texture features are then extracted using a gray-level co-occurrence matrix. The gray-level co-occurrence matrix describes the spatial distribution relationship between different gray-level pairs in an image. Texture features such as contrast, correlation, energy, and entropy can be extracted from this matrix. These features can reflect the texture information of local regions in the echocardiogram.

[0074] Perform vessel segmentation on cardiac CT data. Using a specialized vessel segmentation algorithm, the vascular structures in cardiac CT images are separated from other tissues. Topological and geometric features are extracted based on the segmented vascular structures. Topological features primarily describe the connections and network structure of the vessels, such as the number of branches and connectivity. Geometric features include vessel length, diameter, and curvature, which are valuable for understanding the morphology and structure of cardiovascular vessels.

[0075] Finally, the extracted texture features, morphological features, intensity distribution features, temporal features, local texture features, topological features, and geometric features are standardized. The purpose of standardization is to eliminate dimensional differences between modal data and make different features comparable. A common approach is to map the value of each feature to a specific interval, such as [0, 1] or [-1, 1]. Principal component analysis is then used to reduce the dimensionality of the standardized features. Principal component analysis uses linear transformations to convert the original features into a set of uncorrelated principal components. These principal components are sorted by variance, retaining those with larger variances and removing redundant features, thereby generating a streamlined and effective radiomics feature set.

[0076] Example 2:

[0077] This embodiment describes in detail the process of modeling the dynamic changes of serum marker data based on a time series analysis model to generate a serum marker time series feature set, with an emphasis on the application of long short-term memory networks.

[0078] Collect serum marker concentration data at multiple time points and construct a time series matrix. These time points should be representative and reflect changes in serum marker concentrations over time. For example, data can be collected at different time points during a patient's regular physical examination or during treatment.

[0079] An autoregressive integrated moving average model is used to perform trend decomposition on the time series matrix. This model decomposes the time series into a trend term, a seasonal term, and a residual term. The trend term reflects the long-term variation in serum marker concentrations over time. For example, during the course of a disease, some serum markers may exhibit a gradual upward or downward trend. The seasonal term reflects the cyclical variation in the data. If serum marker concentrations are affected by seasonal factors, this will be reflected in the seasonal term. The residual term includes other random fluctuations that cannot be explained by the trend and seasonal terms.

[0080] Align time series data from multiple patients using the Dynamic Time Warping algorithm. Since different patients may have different sampling times, this can affect subsequent analysis results. The Dynamic Time Warping algorithm can align time series data from different patients by finding the optimal matching path between time series, eliminating sampling time deviations between individuals.

[0081] Build a time series prediction model based on a long short-term memory (LSTM) network. Divide the time series data into sliding windows, with each window containing a sequence of marker concentrations at consecutive time points. For example, each window can contain serum marker concentration data for five consecutive time points. Build a bidirectional LSTM network, which can simultaneously learn both forward and backward information about the time series, better capturing long-term dependencies in the data. Input the window sequence into the bidirectional LSTM network, which then outputs the predicted concentration value for the next time point.

[0082] A self-attention mechanism is introduced in the hidden layer. This mechanism dynamically adjusts the weights of different time steps, allowing the network to focus more on time steps that are important for prediction. For example, at key time points when a disease is changing, the network automatically assigns higher weights to these time steps, thereby improving prediction accuracy.

[0083] The mean square error is used as the loss function. The network parameters are updated through the time backpropagation algorithm, which is a backpropagation algorithm trained on time series data. It can effectively propagate the error from the output layer to each layer of the network, thereby adjusting the weights and bias of the network.

[0084] Use early stopping to prevent overfitting. During training, the dataset is divided into a training set and a validation set. If the validation set loss fails to decrease for three consecutive times, the model is considered overfitting and training is terminated. This prevents the model from overfitting the training set and improves its generalization ability.

[0085] Finally, the hidden state vector of the time series prediction model is extracted as a dynamic change feature. Combined with the previously obtained trend term, seasonal term, residual term, and dynamic change feature, a serum marker time series feature set is generated. These features comprehensively reflect the dynamic changes in serum marker concentrations, providing an important basis for subsequent disease diagnosis.

[0086] Example 3:

[0087] This example introduces in detail the specific steps of using a dynamic weighted fusion strategy to perform cross-modal feature fusion on the radiomics feature set and the serum marker time series feature set to generate a joint feature matrix.

[0088] A similarity metric function between modalities was constructed. The correlation between radiomics features and serum marker features was calculated based on cosine similarity. Cosine similarity measures the similarity between two vectors by calculating the cosine value of the angle between them. The calculation formula is:

[0089]

[0090] in and Represent the radiomics feature vector and serum marker feature vector, respectively. The higher the correlation, the stronger the similarity between the two features.

[0091] A dynamic weight is assigned to each radiomics feature based on the calculated correlation. The weight is positively correlated with the correlation; the higher the correlation, the greater the weight assigned to the radiomics feature. This helps highlight radiomics features that are closely associated with serum markers and improves fusion results.

[0092] Normalized weighting is applied to serum marker time series features. The weight is determined based on the inverse of the feature variance. Variance reflects the degree of dispersion of feature data. A smaller variance indicates a more stable feature and a greater potential contribution to the fusion results. Therefore, using the inverse of the variance as the weight assigns greater weight to stable serum marker time series features.

[0093] The weighted radiomics features were horizontally concatenated with the standardized weighted serum marker temporal features to generate a preliminary fusion matrix. This concatenation process merged the two feature sets along the columns, combining the radiomics features and serum marker temporal features for each sample.

[0094] Kernel principal component analysis (KPCA) was used to perform nonlinear dimensionality reduction on the initial fusion matrix. KPCA uses a kernel function to map the raw data into a high-dimensional space, enabling it to handle nonlinear data. In this high-dimensional space, it can identify a new set of uncorrelated principal components that better preserve the data's characteristic information. KPCA eliminates cross-modal redundant information and generates a joint feature matrix. This joint feature matrix retains the important information of both radiomics and serum biomarker temporal features while removing redundancy, providing higher-quality data for subsequent model training.

[0095] Embodiment 4:

[0096] This embodiment describes in detail the steps of building a multi-task deep learning model, especially the application of the graph attention mechanism therein.

[0097] Construct a patient feature graph. Nodes represent joint feature vectors for individual patients, obtained through the previous feature fusion step. Edges represent clinical similarity between patients, calculated based on the matching degree of age, gender, and underlying disease. For example, a quantitative approach can be used to convert age differences, gender identity, and underlying disease similarity into a single numerical value. A larger numerical value indicates greater clinical similarity between patients.

[0098] The graph attention mechanism is used to aggregate node features and generate node embedding vectors. First, the attention coefficient between nodes is calculated. The attention coefficient is based on the weighted sum of the cosine similarity of node features and clinical similarity. Assume that the feature vectors of node i and node j are and Clinical similarity is S ij , then the attention coefficient e ij The calculation formula is:

[0099]

[0100] where α is a weight parameter used to balance the importance of cosine similarity and clinical similarity.

[0101] The attention coefficient is normalized by the Softmax function to generate the normalized attention weight. The calculation formula of the Softmax function is:

[0102]

[0103] where N i is the set of neighbor nodes of node i, β ij It is the normalized attention weight, which indicates the importance of node j to node i.

[0104] According to the normalized attention weight, the features of the neighboring nodes are weighted summed and the embedding vector of the central node is updated. That is:

[0105]

[0106] in is the updated embedding vector of node i.

[0107] Multiple layers of graph attention layers are stacked, and the output of each layer is processed using Layer Normalization and the ReLU activation function. Layer Normalization is a normalization method that normalizes the features of each node, accelerating model training convergence. The ReLU activation function is used to introduce nonlinearity and enhance the model's expressiveness. Processing through multiple layers of graph attention layers allows for better extraction of node feature information.

[0108] Finally, the node embedding vector of the last layer is extracted as the input for classification and reconstruction tasks.

[0109] Design a classification task branch and input the node embedding vector into a fully connected layer. A fully connected layer is a common neural network layer in which each neuron is connected to all neurons in the previous layer. The fully connected layer calculates and outputs the DCM risk probability.

[0110] The feature reconstruction task branch is designed to reconstruct the original joint feature matrix from the node embedding vectors through the decoder network. The decoder network can adopt a multi-layer neural network structure to map the low-dimensional node embedding vectors back to the high-dimensional joint feature matrix space.

[0111] Construct a joint loss function, which includes the weighted sum of classification cross entropy loss and reconstruction mean square error loss. Classification cross entropy loss is used to measure the difference between the predicted result of the classification task and the true label. Its calculation formula is:

[0112]

[0113] Where C is the number of categories, y i is the i-th component of the true label, p i Is the i-th component of the predicted probability. The reconstruction mean square error loss is used to measure the difference between the reconstructed joint feature matrix and the original joint feature matrix, and the calculation formula is:

[0114]

[0115] Where n is the number of features, x i is the i-th element of the original joint feature matrix, is the i-th element of the reconstructed joint feature matrix. The joint loss function L = λCE + (1-λ)MSE, where λ is a weight parameter used to balance the importance of classification loss and reconstruction loss.

[0116] An alternating optimization strategy is used to update the network parameters for the classification and reconstruction tasks. During training, the network parameters for the reconstruction task are first fixed and the parameters for the classification task are updated. Then, the parameters for the classification task are fixed and the parameters for the reconstruction task are updated. This alternating process continues until the model converges. This approach allows for simultaneous optimization of both the classification and reconstruction tasks, improving model performance.

[0117] Example 5:

[0118] When diagnosing early-stage DCM, the optimal diagnostic path is generated by optimizing the decision tree branch threshold based on a genetic algorithm. The specific steps include:

[0119] A personalized diagnostic decision tree is constructed based on the DCM risk probability, and the branch threshold of the decision tree is optimized using a genetic algorithm to generate an optimal diagnostic path. An early DCM diagnosis result is output based on the optimal diagnostic path. The process of constructing a personalized diagnostic decision tree based on the DCM risk probability is as follows:

[0120] 1) Input data definition

[0121] Input features: DCM risk probability (0-1), patient clinical indicators (such as left ventricular ejection fraction, electrocardiogram characteristics), and multimodal examination data (ultrasound, MRI).

[0122] Decision variables: Decision tree nodes correspond to diagnostic branches (e.g., “high risk requires interventional treatment”, “medium risk requires regular follow-up”).

[0123] 2) Decision tree initialization design

[0124] Root node: The first fork is performed based on the DCM risk probability, and the initial threshold is set to 0.5 (high risk / low risk).

[0125] Intermediate nodes: Combine clinical indicators (such as LVEF < 40%) with the feature reconstruction error (outlier detection) output by the reconstruction module to generate secondary branches.

[0126] Leaf node: outputs the diagnosis conclusion (confirmed DCM, suspected case, excluded risk) and confidence level.

[0127] 3) Feature interaction enhancement

[0128] The node embedding features of the classification module and reconstruction module are extracted through the graph neural network (GNN), and the joint feature matrix is constructed as the decision tree input.

[0129] Use the attention mechanism to dynamically adjust the weights of different features in the decision tree (e.g., high-risk patients pay more attention to MRI features).

[0130] The branch threshold of the decision tree is optimized based on the genetic algorithm to generate the optimal diagnostic path as follows:

[0131] The diagnostic decision tree is encoded. Each branch node in the decision tree is considered a gene position, and the branch threshold is encoded as a gene value. This encoding method is like giving the decision tree a unique "genetic code," enabling genetic algorithms to operate and optimize it. Decision trees play a vital role in classification decision-making throughout the diagnostic process. They make diagnostic judgments by layering features extracted from multi-radiological and serum marker data. This encoding of key parameters is the foundation for subsequent optimization using genetic algorithms.

[0132] Multiple individuals with different genetic values are randomly generated to form the initial population. Each individual represents a possible combination of decision tree branch thresholds, much like randomly selecting initial "exploration directions" from among numerous possible diagnostic paths. For example, 50 individuals are randomly generated, each containing the corresponding threshold genetic values for each branch node of the decision tree. These initial values are completely randomly determined to ensure population diversity and provide sufficient "material" for subsequent optimization processes.

[0133] Construct a fitness function to evaluate the diagnostic performance of each individual. This function is calculated based on the confusion matrix of the diagnostic path. The confusion matrix is a commonly used tool in classification tasks that clearly demonstrates the difference between the actual classification results and the true situation. The main metrics considered here are sensitivity, specificity, and F1 score.

[0134] Sensitivity reflects the proportion of people who are actually ill and correctly diagnosed as having the disease. Simply put, the higher the proportion of people who are accurately detected among all those who are actually ill, the better the sensitivity, which is crucial for timely disease detection. Specificity reflects the proportion of people who are actually not ill and correctly diagnosed as not having the disease. It measures the accuracy of the diagnostic method in excluding healthy people. The F1 score is an indicator that comprehensively considers sensitivity and specificity. When both sensitivity and specificity are high, the F1 score is also high, which more comprehensively reflects the accuracy of the diagnosis.

[0135] These indicators are weighted together to form a fitness function, such as the fitness function F = 0.4 × Sensitivity + 0.3 × Specificity + 0.3 × F1. The weight coefficients (0.4, 0.3, 0.3) are set based on actual needs and experience. Different weights will have different effects on individual fitness assessments. By properly adjusting the weights, we can emphasize the importance of certain indicators in diagnosis.

[0136] A tournament selection strategy is used to select individuals. K individuals are randomly selected from the current population to form tournament groups. Here, k can be set to 3 based on actual conditions. Within each tournament group, the individual with the highest fitness is selected to enter the mating pool. This process is like a series of small "contests," each time selecting relatively outstanding individuals to enter the "candidate breeding pool." Tournament selection is repeated until the mating pool is full.

[0137] Once the mating pool is prepared, individuals in the pool are paired, and genes are exchanged at randomly selected crossover points to produce offspring. This crossover operation simulates genetic recombination in biological inheritance. By exchanging partial genes between two individuals, it is possible to create new individuals with greater advantages. For example, consider individuals A and B. After exchanging their gene values at a randomly selected crossover point, new offspring are produced. These offspring inherit some of the characteristics of their parents, but may also produce new combinations, providing new possibilities for finding more optimal diagnostic pathways.

[0138] The offspring individuals are also subjected to mutation. Each gene position is randomly mutated with a preset probability, and the mutation range is limited to the feasible interval of the branching threshold. The preset probability is usually a small value, such as 0.05, which means that each gene position has a 5% probability of mutation. The purpose of mutation is to prevent the algorithm from prematurely falling into a local optimal solution. By introducing new gene values, the algorithm can explore a wider solution space and potentially discover better branching threshold combinations.

[0139] During each generation of evolution, the top 10% of individuals in the parent population by fitness are retained and passed directly to the next generation. This ensures that high-performing individuals are not lost due to genetic manipulation, and their excellent "genes" can continue to be passed on in the population, accelerating the algorithm's convergence to the optimal solution.

[0140] The population is continuously updated iteratively, repeating selection, crossover, and mutation operations until a pre-set maximum number of evolutionary generations is reached. When this maximum number of evolutionary generations is reached, the individual with the highest fitness is selected from the population. This optimal individual is decoded to obtain optimized branch thresholds, which form the optimal diagnostic path. Based on this optimal diagnostic path, more accurate and scientific early-stage DCM diagnoses can be generated, providing powerful diagnostic support to clinicians, helping them more accurately determine whether a patient has early-stage DCM and promptly initiate appropriate treatment measures.

[0141] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0142] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A non-invasive early-stage DCM analysis method based on multi-imaging omics and serum markers, characterized by: include: collecting imaging data of the pediatric patient using a multimodal medical imaging device, wherein the multimodal medical imaging includes cardiac MRI, echocardiography, and cardiac CT; Collect serum marker data through blood testing equipment; Performing radiomics feature extraction on the image data based on a multi-scale feature extraction algorithm to generate a radiomics feature set; Performing dynamic change modeling on the serum marker data based on a time series analysis model to generate a serum marker time series feature set; A dynamic weighted fusion strategy is used to perform cross-modal feature fusion on the radiomics feature set and the serum marker temporal feature set to generate a joint feature matrix; Inputting the joint feature matrix into a pre-trained multi-task deep learning model, the multi-task deep learning model includes a classification module and a reconstruction module, wherein the classification module is used to perform a DCM classification task to predict the disease risk probability, and the reconstruction module is used to perform a feature reconstruction task to reconstruct the joint feature matrix, interactively learning the node embedding features of the classification module and the reconstruction module through a graph neural network architecture, and synchronously optimizing the two types of tasks based on a joint loss function to output the DCM risk probability; A personalized diagnosis decision tree is constructed according to the DCM risk probability, and the branch threshold of the decision tree is optimized based on a genetic algorithm to generate an optimal diagnostic path; and an early DCM diagnosis result is output based on the optimal diagnostic path.

2. The non-invasive analysis method for early DCM based on multi-imaging omics and serum markers according to claim 1, characterized in that: The extracting radiomics features from the image data based on a multi-scale feature extraction algorithm to generate a radiomics feature set includes: Perform three-dimensional wavelet transform on cardiac MRI data to extract texture features, morphological features and intensity distribution features in multiple frequency bands; Dynamic time warping is performed on echocardiographic data to generate cardiac cycle aligned temporal features, and local texture features are extracted using the gray-level co-occurrence matrix. performing vascular segmentation on cardiac CT data, extracting topological and geometric features based on the segmented vascular structure; and normalizing the texture features, morphological features, intensity distribution features, temporal features, local texture features, topological features, and geometric features to eliminate dimensional differences between modalities. Principal component analysis was used to reduce the dimension of the standardized features, remove redundant features, and generate a radiomics feature set.

3. The non-invasive analysis method for early DCM based on multi-radiology and serum markers according to claim 1, characterized in that: The step of performing dynamic change modeling on the serum marker data based on the time series analysis model to generate a serum marker time series feature set includes: Collecting serum marker concentration data at multiple time points to construct a time series matrix; performing trend decomposition on the time series matrix using an autoregressive integrated moving average model to obtain trend terms, seasonal terms, and residual terms; The time series data of multiple patients were aligned based on the dynamic time warping algorithm to eliminate the sampling time deviation between individuals. A time series prediction model based on the long short-term memory network was constructed, and the aligned time series data was input to predict the changes in marker concentrations at future time points. The hidden state vector of the time series prediction model is extracted as a dynamic change feature; and a serum marker time series feature set is generated by combining the trend term, the seasonal term, the residual term and the dynamic change feature.

4. The non-invasive analysis method for early DCM based on multi-radiology and serum markers according to claim 1, characterized in that: The cross-modal feature fusion of the imaging omics feature set and the serum marker time series feature set using a dynamic weighted fusion strategy to generate a joint feature matrix includes: An inter-modality similarity metric function was constructed, and the correlation between radiomics features and serum marker features was calculated based on cosine similarity. A dynamic weight was assigned to each radiomics feature based on the correlation, with the weight value being positively correlated with the correlation degree. The serum marker temporal features were standardized and weighted, with the weight value determined based on the inverse of the feature variance. The weighted imaging genomics features and the standardized weighted serum marker temporal features were horizontally spliced to generate a preliminary fusion matrix. Kernel principal component analysis was used to perform nonlinear dimensionality reduction on the preliminary fusion matrix to eliminate cross-modal redundant information and generate a joint feature matrix.

5. The non-invasive analysis method for early DCM based on multi-radiology and serum markers according to claim 1, characterized in that: The steps of constructing the multi-task deep learning model include: Construct a patient feature graph, where nodes represent joint feature vectors of individual patients and edges represent clinical similarities between patients, calculated based on age, gender, and underlying disease matching. A graph attention mechanism is used to aggregate node features and generate node embedding vectors. A classification task branch is designed to input the node embedding vectors into the fully connected layer and output the DCM risk probability. Design a feature reconstruction task branch to reconstruct the original joint feature matrix from the node embedding vector through the decoder network; A joint loss function is constructed, wherein the joint loss function includes a weighted sum of classification cross entropy loss and reconstruction mean square error loss; and an alternating optimization strategy is used to update network parameters of the classification task and the reconstruction task until the model converges.

6. The non-invasive analysis method for early DCM based on multi-radiology and serum markers according to claim 1, characterized in that: The step of optimizing the branch threshold of the decision tree based on the genetic algorithm to generate the optimal diagnostic path includes: Each branch node of the diagnostic decision tree is encoded as a gene position, and the branch threshold is encoded as a gene value. The population is initialized and multiple individuals with different gene values are randomly generated. A fitness function is constructed, which is calculated based on the confusion matrix indicators of the diagnostic path and includes a weighted combination of sensitivity, specificity, and F1 score. A tournament selection strategy is used to screen high-fitness individuals, and a progeny population is generated through single-point crossover and mutation operations. The population is iteratively updated until the maximum evolutionary generation is reached, and the individual with the highest fitness is selected and decoded as the optimized branch threshold to generate the optimal diagnostic path.

7. The non-invasive analysis method for early DCM based on multi-imaging omics and serum markers according to claim 2, characterized in that: The three-dimensional wavelet transform is performed on the cardiac MRI data to extract the texture features, morphological features and intensity distribution features in multiple frequency bands, including: Cardiac MRI data were decomposed into low-frequency approximate components and high-frequency detail components using a three-dimensional discrete wavelet transform. Intensity distribution features based on grayscale histograms, including mean, variance, and skewness, were extracted from the low-frequency approximate components. Texture features based on Gabor filters, including energy, entropy, and contrast, were extracted from each high-frequency detail component. The three-dimensional volume, surface area, and sphericity index of the cardiac chambers were extracted based on morphological operations. The intensity distribution features, texture features, and morphological features were hierarchically integrated by frequency band to generate multi-scale imaging omics features.

8. The non-invasive analysis method for early DCM based on multi-radiology and serum markers according to claim 3, characterized in that: The method of constructing a time series prediction model based on a long short-term memory network, inputting the aligned time series data, and predicting changes in marker concentrations at future time points includes: The time series data is divided into sliding windows, each containing a sequence of marker concentrations at consecutive time points. A bidirectional long-short-term memory network is constructed, which inputs the window sequence and outputs the predicted concentration value at the next time point. A self-attention mechanism is introduced in the hidden layer to dynamically adjust the weights of different time steps. The mean square error is used as the loss function, and the network parameters are updated through the time backpropagation algorithm; the early stopping method is used to prevent overfitting, and the training is terminated when the validation set loss does not decrease for three consecutive times.

9. The non-invasive analysis method for early DCM based on multi-radiology and serum markers according to claim 5, characterized in that: The graph attention mechanism is used to aggregate node features and generate node embedding vectors, including: Calculating the attention coefficient between nodes, where the attention coefficient is based on the weighted sum of the cosine similarity and clinical similarity of the node features; normalizing the attention coefficient using the Softmax function to generate normalized attention weights; performing weighted summation of the features of neighboring nodes based on the normalized attention weights to update the embedding vector of the central node; Multiple layers of graph attention layers are stacked, and the output of each layer is processed by Layer Normalization and ReLU activation function; finally, the node embedding vector of the last layer is extracted as the input of the classification and reconstruction tasks.

10. The non-invasive analysis method for early DCM based on multi-radiology and serum markers according to claim 6, characterized in that: The tournament selection strategy is used to screen individuals with high fitness, and the offspring population is generated through single-point crossover and mutation operations, including: K individuals are randomly selected from the current population to form a tournament group, and the individual with the highest fitness is selected to enter the mating pool; the tournament selection is repeated until the mating pool is filled; the individuals in the mating pool are paired, and crossover points are randomly selected for gene exchange to generate offspring individuals; each gene position of the offspring individual is randomly mutated with a preset probability, and the mutation range is the feasible interval of the branching threshold; the individuals with the top 10% fitness in the parent population are retained and directly enter the next generation.

Images