Early Non-invasive Analysis Methods for DCM Based on Multi-radiomics and Serum Biomarkers

Abstract

This invention relates to the field of medical diagnostic technology and discloses a non-invasive early DCM analysis method based on multi-radiomics and serum biomarkers. The method involves acquiring image data using multimodal medical imaging equipment and serum biomarker data using blood testing equipment; generating radiomics feature sets and serum biomarker time-series feature sets using multi-scale feature extraction algorithms and temporal analysis models, respectively; fusing these features using a dynamic weighted fusion strategy to generate a joint feature matrix; inputting this matrix into a pre-trained multi-task deep learning model to predict the probability of DCM risk; and finally, optimizing the diagnostic decision tree based on a genetic algorithm to output the early DCM diagnostic result. This method is non-invasive, accurate, and can effectively improve the accuracy of early DCM diagnosis.

Description

Technical Field

[0001] This invention relates to the field of medical diagnostic technology, specifically to a non-invasive early analysis method for DCM based on multi-radiomics and serum biomarkers. Background Technology

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

[0003] Traditional diagnostic methods primarily rely on invasive techniques, such as endomyocardial biopsy. While endomyocardial biopsy can provide histological diagnostic evidence, it is invasive and may cause complications such as infection, bleeding, and cardiac perforation, bringing pain and risks to patients. Furthermore, this method is difficult to widely apply in clinical practice for large-scale screening and repeat testing.

[0004] In non-invasive diagnostics, single medical imaging techniques have limitations. While cardiac MRI can provide high-resolution information on cardiac structure and function, it does not comprehensively reflect the characteristics of myocardial tissue. Echocardiography is easy to operate and can dynamically observe cardiac activity in real time, but image quality is easily affected by factors such as patient size and lung capacity, resulting in limited diagnostic accuracy. Cardiac CT can clearly display cardiac vascular structures, but it is insufficient in assessing myocardial function and tissue characteristics.

[0005] Serum biomarker testing is an important auxiliary means of diagnosing DCM, but most serum biomarkers lack high specificity and sensitivity, and a single serum biomarker is insufficient to accurately reflect the disease progression and severity. Furthermore, current clinical analysis of serum biomarkers is mostly based on static test results, ignoring their dynamic changes over time, thus failing to capture early, subtle changes in the disease in a timely manner.

[0006] In terms of data processing and analysis techniques, existing diagnostic models often fail to fully integrate the advantages of multi-source heterogeneous data. Multimodal medical imaging data and serum biomarker data have different data structures and characteristics, making it difficult for traditional methods to effectively fuse these data and extract potential diagnostic information, thus hindering the achievement of ideal levels of diagnostic accuracy and reliability.

[0007] Due to the lack of efficient early non-invasive diagnostic methods, many patients with diabetic comorbidities (DCM) are only diagnosed when the disease has progressed to the middle or late stages, missing the optimal treatment window. Therefore, developing an early non-invasive analysis method based on multi-radiomics and serum biomarkers is of significant 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 this invention is to provide a non-invasive early analysis method for DCM based on multi-radiomics and serum biomarkers, in order to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A non-invasive early analysis method for DCM based on multi-radiomics and serum biomarkers, the method comprising:

[0011] Imaging data of pediatric patients are acquired using multimodal medical imaging equipment, including cardiac MRI, echocardiography, and cardiac CT.

[0012] Serum biomarker data are collected using blood testing equipment;

[0013] Based on a multi-scale feature extraction algorithm, radiomics features are extracted from the image data to generate a radiomics feature set.

[0014] Based on the time series analysis model, the serum biomarker data is dynamically modeled to generate a serum biomarker 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 biomarker time-series feature set to generate a joint feature matrix;

[0016] The joint feature matrix is ​​input into a pre-trained multi-task deep learning model, which includes a classification module and a reconstruction module. 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. The node embedding features of the classification module and the reconstruction module are interactively learned through a graph neural network architecture, and the two tasks are simultaneously optimized based on a joint loss function to output the DCM risk probability.

[0017] A personalized diagnostic decision tree is constructed based on the DCM risk probability. The branch threshold of the decision tree is optimized based on a genetic algorithm to generate the optimal diagnostic path. The early DCM diagnostic results are output based on the optimal diagnostic path.

[0018] Optionally, the step of extracting radiomics features from the image data based on a multi-scale feature extraction algorithm to generate a radiomics feature set includes:

[0019] Three-dimensional wavelet transform was performed 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 temporal features aligned with cardiac cycles, and local texture features are extracted using a gray-level co-occurrence matrix.

[0021] Cardiac CT data is segmented into blood vessels, and topological and geometric features are extracted based on the segmented blood vessel structures. The texture features, morphological features, intensity distribution features, temporal features, local texture features, topological features, and geometric features are then standardized to eliminate dimensional differences between modalities.

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

[0023] Optionally, the step of dynamically modeling the changes in the serum biomarker data based on the time-series analysis model to generate a serum biomarker time-series feature set includes:

[0024] Serum biomarker concentration data were collected at multiple time points to construct a time series matrix. An autoregressive integral moving average model was used to decompose the time series matrix into trend terms, seasonal terms, and residual terms.

[0025] The time series data of multiple patients are aligned using a dynamic time warping algorithm to eliminate sampling time bias between individuals; a time series prediction model based on a long short-term memory network is constructed, and the aligned time series data is input to predict the changes in biomarker concentrations at future time points.

[0026] The hidden state vector of the time-series prediction model is extracted as a dynamic change feature; the trend term, seasonal term, residual term and dynamic change feature are combined to generate a time-series feature set of serum biomarkers.

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

[0028] A modal similarity measurement function is constructed, and the correlation between radiomics features and serum biomarker features is calculated based on cosine similarity. Dynamic weights are assigned to each radiomics feature according to the correlation, and the weight values ​​are positively correlated with the correlation. The time-series features of serum biomarkers are standardized and weighted, and the weight values ​​are determined based on the inverse of the feature variance.

[0029] The weighted radiomics features and the standardized weighted serum biomarker time-series features are horizontally concatenated to generate a preliminary fusion matrix. Kernel principal component analysis is then 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 construction steps of the multi-task deep learning model include:

[0031] Construct a patient feature map, where nodes represent the joint feature vector of a single patient and edges represent the clinical similarity between patients, which is calculated based on the matching degree of age, gender, and underlying diseases;

[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 a fully connected layer and output the DCM risk probability.

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

[0034] A joint loss function is constructed, which is a weighted sum of classification cross-entropy loss and reconstruction mean squared error loss; the network parameters for classification and reconstruction tasks are updated using an alternating optimization strategy until the model converges.

[0035] Optionally, the optimization of 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 location, 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 index of the diagnostic path, including a weighted combination of sensitivity, specificity, and F1 score. A tournament selection strategy is used to screen individuals with high fitness, and offspring populations are generated through single-point crossover and mutation operations. The population is iteratively updated until the maximum number of generations 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, the step of performing three-dimensional wavelet transform on cardiac MRI data to extract texture features, morphological features, and intensity distribution features across multiple frequency bands includes:

[0038] The cardiac MRI data was decomposed into low-frequency approximation components and high-frequency detail components using three-dimensional discrete wavelet transform. Intensity distribution features based on gray-level histograms, including mean, variance, and skewness, were extracted from the low-frequency approximation 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 then integrated in a hierarchical manner according to frequency bands to generate multi-scale radiomics features.

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

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

[0041] Optionally, the step of aggregating node features using a graph attention mechanism to generate node embedding vectors includes:

[0042] The attention coefficients between nodes are calculated, which are weighted sums of cosine similarity and clinical similarity of node features. The attention coefficients are normalized using the Softmax function to generate normalized attention weights. The features of neighboring nodes are weighted and summed according to the normalized attention weights to update the embedding vector of the center node. Multiple graph attention layers are stacked, and the output of each layer is processed by Layer Normalization and ReLU activation functions. Finally, the node embedding vectors of the last layer are extracted as inputs for classification and reconstruction tasks.

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

[0044] Randomly select k individuals from the current population to form a tournament group, and select the individual with the highest fitness to enter the mating pool; repeat the tournament selection until the mating pool is full; pair up the individuals in the mating pool, randomly select the crossover point to exchange genes, and generate offspring individuals; perform random mutation on each gene position of the offspring individuals with a preset probability, and the mutation range is the feasible interval of the branch threshold; retain the top 10% of the individuals with the highest fitness in the parent population to 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 analysis method for DCM based on multi-radiomics and serum biomarkers proposed in this invention improves the accuracy and efficiency of early diagnosis of DCM in many ways.

[0047] In terms of data acquisition and integration, this method acquires cardiac MRI, echocardiography, and cardiac CT imaging data using multimodal medical imaging equipment, while simultaneously combining blood tests to obtain serum biomarker data, achieving comprehensive collection of multi-source heterogeneous data. This multimodal data fusion strategy overcomes the limitations of single detection methods, comprehensively utilizing 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 features of cardiac CT, and the biological information of serum biomarkers complement each other, providing rich evidence for the accurate diagnosis of DCM.

[0048] In the radiomics feature extraction stage, specialized multi-scale feature extraction algorithms are employed for different image data. Three-dimensional wavelet transform is performed on cardiac MRI to comprehensively extract texture, morphology, and intensity distribution features across multiple frequency bands, reflecting myocardial tissue characteristics from different dimensions. Echocardiography undergoes dynamic time warping and gray-level co-occurrence matrix analysis to obtain temporal features aligned with the cardiac cycle and local texture features, effectively capturing subtle changes during cardiac motion. Cardiac CT extracts topological and geometric features through vascular segmentation, highlighting the correlation between vascular structure and disease. After standardization and principal component analysis for dimensionality reduction, intermodal dimensional differences are eliminated, redundant features are removed, and a high-quality radiomics feature set is generated, improving data usability and analytical efficiency.

[0049] In the process of generating time-series features of serum biomarkers, serum biomarker concentration data from multiple time points are collected to construct a time-series matrix. An autoregressive integral moving average model is used for trend decomposition, and a dynamic time warping algorithm is combined to align sampling time deviations between individuals. Finally, a time-series prediction model based on a long short-term memory network is used to predict future biomarker concentration changes, fully exploring the dynamic changes in serum biomarkers. This dynamic analysis of serum biomarkers can more sensitively capture disease development trends and reflects the true state of the disease better than traditional static detection.

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

[0051] The multi-task deep learning model constructs a patient feature map, aggregates node features using a graph attention mechanism, designs classification and reconstruction task branches, and simultaneously optimizes using a joint loss function. The classification module predicts the probability of DCM risk, while the reconstruction module assists in optimizing the model by reconstructing the joint feature matrix. The collaboration between these two modules improves the model's accuracy and stability. The graph neural network architecture enables interactive learning of node embedded features, fully leveraging clinical similarity information among patients and further enhancing model performance.

[0052] This approach optimizes the branch thresholds of a diagnostic decision tree using a genetic algorithm, constructs a fitness function based on a confusion matrix, and generates the optimal diagnostic path through multiple rounds of selection, crossover, and mutation operations. This method enables the development of personalized diagnostic strategies based on the specific characteristics of different patients, improving the accuracy and reliability of diagnosis and reducing the occurrence of misdiagnosis and missed diagnosis. Attached Figure Description

[0053] Figure 1 This is a schematic diagram illustrating the working principle of the early non-invasive analysis method for DCM based on multiple radiomics and serum biomarkers described in this invention.

[0054] Figure 2 A flowchart for cross-modal feature fusion;

[0055] Figure 3 A flowchart for constructing a multi-task deep learning model;

[0056] Figure 4 This is a flowchart of optimizing a diagnostic decision tree based on a genetic algorithm. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0058] Please see Figure 1-4 This invention provides a non-invasive early analysis method for DCM based on multi-radiomics and serum biomarkers, the specific steps of which include:

[0059] Data Acquisition: Imaging data of pediatric patients was acquired using multimodal medical imaging equipment, including cardiac MRI, echocardiography, and cardiac CT. These images present structural and functional information of the heart from different angles and in different ways. Simultaneously, serum biomarker data were collected using blood testing equipment. This data reflects the patient's biochemical state and is of great significance for disease diagnosis and monitoring.

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

[0061] Dynamic modeling: A time-series analysis model is used to model the dynamic changes of serum biomarker data, generating a time-series feature set of serum biomarkers. This step considers the changing trends of serum biomarkers over time, which helps to more accurately grasp the progression of the disease.

[0062] Feature fusion: A dynamic weighted fusion strategy is employed to fuse radiomics feature sets and serum biomarker time-series feature sets across modalities, generating a joint feature matrix. This approach integrates the advantages of different modalities, improving the accuracy and reliability of the analysis.

[0063] Model Prediction: The joint feature matrix is ​​input into a pre-trained multi-task deep learning model. This model includes a classification module and a reconstruction module. The classification module performs a DCM classification task to predict disease risk probabilities; the reconstruction module performs a feature reconstruction task to reconstruct the joint feature matrix. A graph neural network architecture is used to interactively learn the node embedding features of the classification and reconstruction modules, and the two tasks are simultaneously optimized based on a joint loss function, ultimately outputting the DCM risk probability.

[0064] Diagnostic Decision: A personalized diagnostic decision tree is constructed based on the obtained DCM risk probability. The branch thresholds of the decision tree are optimized using a genetic algorithm to generate the optimal diagnostic path. Finally, early DCM diagnostic results are output based on the optimal diagnostic path, 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] This embodiment elaborates in detail the specific process of extracting radiomics features from image data based on a multi-scale feature extraction algorithm to generate a radiomics feature set, especially the processing of cardiac MRI data.

[0068] First, a three-dimensional wavelet transform was performed on the cardiac MRI data. Using a three-dimensional discrete wavelet transform, the cardiac MRI data was decomposed into low-frequency approximation components and high-frequency detail components. The low-frequency approximation components preserved the main contours and general features of the image, while the high-frequency detail components contained fine information such as edges and textures.

[0069] Intensity distribution features based on gray-level histograms are extracted from low-frequency approximation components. A gray-level histogram reflects the frequency of different gray-level values ​​in an image. The mean is calculated, representing the average level of gray-level values ​​in the image and reflecting the overall brightness. Variance measures the dispersion of gray-level values ​​relative to the mean; a larger variance indicates a more dispersed distribution of gray-level values. Skewness describes the asymmetry of the gray-level distribution, helping to identify potentially anomalous gray-level distribution areas in the image.

[0070] Texture features based on Gabor filters are extracted from various high-frequency detail components. Gabor filters are commonly used in image processing and can effectively extract texture information from images. The extracted energy features represent the activity level of the image texture; the higher the energy value, the more obvious the texture. Entropy reflects the complexity of the texture; the higher the entropy value, the more complex the texture. Contrast reflects the degree of difference between different gray-level regions in the image; the higher the contrast, the clearer the image details.

[0071] Morphological operations were used to extract the three-dimensional volume, surface area, and sphericity index of heart chambers. These operations, such as erosion and dilation, accurately extract the geometric features of the heart chambers. The three-dimensional volume reflects the size of the heart chambers, the surface area represents their surface extent, and the sphericity index measures how closely the shape of the heart chambers resembles a sphere. These features are crucial for assessing the structure and function of the heart.

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

[0073] For echocardiographic data, dynamic time warping is performed. Since cardiac cycles may differ among patients, dynamic time warping aligns the echocardiographic data from different patients temporally, generating time-series features aligned with cardiac cycles. Then, local texture features are extracted using a gray-level co-occurrence matrix (GLCM). The GLCM describes the spatial distribution relationship between different gray-level pairs in the image, from which texture features such as contrast, correlation, energy, and entropy can be extracted. These features reflect the texture information of local regions in the echocardiogram.

[0074] Vascular segmentation is performed on cardiac CT data. A specialized vascular segmentation algorithm is used to separate the vascular structures from other tissues in cardiac CT images. Topological and geometric features are then extracted from the segmented vascular structures. Topological features primarily describe the connectivity and network structure of the vessels, such as the number of branches and connectivity; geometric features include the length, diameter, and curvature of the vessels. These features are valuable for understanding the morphology and structure of cardiac 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 different modalities, making different features comparable. A common method is to map the value of each feature to a specific interval, such as [0,1] or [-1,1]. Then, principal component analysis (PCA) is used to reduce the dimensionality of the standardized features. PCA transforms the original features into a set of uncorrelated principal components through linear transformation. These principal components are sorted by variance, retaining those with higher variances and removing redundant features, thus generating a concise and effective set of imagemics features.

[0076] Example 2:

[0077] This embodiment details the process of dynamically modeling serum biomarker data based on a time-series analysis model to generate a time-series feature set of serum biomarkers, with a focus on the application of long short-term memory networks.

[0078] Serum biomarker concentration data were collected at multiple time points to construct a time series matrix. These time points should be representative and reflect the changes in serum biomarker concentrations over a period of time. For example, data could be collected at different time points during a patient's regular checkups or treatment.

[0079] An autoregressive integral moving average model was 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 trend of serum biomarker concentrations over time; for example, during disease development, certain serum biomarkers may show a gradual increase or decrease. The seasonal term reflects the periodic variations in the data; if serum biomarker concentrations are affected by seasonal factors, this will be reflected in the seasonal term. The residual term contains other random fluctuations that cannot be explained by the trend and seasonal terms.

[0080] This study uses a dynamic time warping algorithm to align time series data from multiple patients. Since sampling times may differ between patients, this can affect subsequent analysis results. The dynamic time warping algorithm can find the optimal matching path between time series, aligning the time series data of different patients temporally and eliminating inter-individual sampling time biases.

[0081] A time-series prediction model based on a Long Short-Term Memory (LSTM) network is constructed. The time-series data is divided into sliding windows, each containing a biomarker concentration sequence at consecutive time points. For example, each window can contain serum biomarker concentration data from five consecutive time points. A bidirectional LSM network is constructed, capable of simultaneously learning both forward and backward information from the time series, better capturing long-term dependencies in the data. The window sequence is input into the bidirectional LSM network, which outputs the predicted concentration value for the next time point.

[0082] A self-attention mechanism is introduced in the hidden layers. This mechanism dynamically adjusts the weights of different time steps, allowing the network to focus more on time steps that are important to the prediction outcome. For example, at critical points in the disease's progression, the network automatically assigns higher weights to these time steps, thereby improving prediction accuracy.

[0083] Mean squared error is used as the loss function. Network parameters are updated using the time-backpropagation algorithm, a backpropagation algorithm trained on time-series data that effectively propagates errors from the output layer to all layers of the network, thereby adjusting the network's weights and biases.

[0084] Early stopping is used to prevent overfitting. During training, the dataset is divided into a training set and a validation set. If the loss on the validation set fails to decrease for three consecutive times, the model is considered overfitting, and training is terminated. This avoids overfitting the model to the training set and improves the model's generalization ability.

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

[0086] Example 3:

[0087] This embodiment details the specific steps of using a dynamic weighted fusion strategy to perform cross-modal feature fusion of radiomics feature sets and serum biomarker time-series feature sets to generate a joint feature matrix.

[0088] Construct an intermodal similarity measurement function. Calculate the correlation between radiomics features and serum biomarker features based on cosine similarity. Cosine similarity measures the degree of similarity between two vectors by calculating the cosine of the angle between them; the formula is:

[0089]

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

[0091] Dynamic weights are assigned to each radiomics feature based on the calculated correlation. The weight values ​​are positively correlated with the correlation; that is, radiomics features with higher correlation receive greater weights. This approach highlights radiomics features closely associated with serum biomarkers, improving the fusion effect.

[0092] The time-series features of serum biomarkers are standardized and weighted. The weight values ​​are determined based on the inverse of the feature variance. Variance reflects the dispersion of the feature data; the smaller the variance, the higher the stability of the feature, and the greater its potential contribution to the fusion result. Therefore, using the inverse of variance as the weight allows for assigning greater weight to stable time-series features of serum biomarkers.

[0093] The weighted radiomics features and the standardized weighted serum biomarker time-series features are horizontally concatenated to generate a preliminary fusion matrix. Horizontal concatenation involves merging the two feature sets along the column direction, combining the radiomics features and serum biomarker time-series features of each sample.

[0094] Kernel principal component analysis (KPCA) was used to perform nonlinear dimensionality reduction on the preliminary fusion matrix. KPCA, by introducing a kernel function, maps the original data to a high-dimensional space, thus enabling the processing of nonlinear data. In this high-dimensional space, it can find a new set of uncorrelated principal components that better preserve the characteristic information of the data. KPCA can eliminate cross-modal redundancy and generate a joint feature matrix. This joint feature matrix retains important information from both radiomics features and serum biomarker time-series features while removing redundant parts, providing higher-quality data for subsequent model training.

[0095] Example 4:

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

[0097] A patient feature map is constructed, where nodes represent the joint feature vectors of individual patients, obtained through the preceding feature fusion step. Edges represent the clinical similarity between patients, calculated based on the matching degree of age, gender, and underlying diseases. For example, a quantitative method can be used to convert age difference, gender similarity, and the degree of similarity of underlying diseases into a numerical value; the higher the value, the higher the clinical similarity between patients.

[0098] A graph attention mechanism is used to aggregate node features and generate node embedding vectors. First, the attention coefficients between nodes are calculated, which are weighted sums of the cosine similarity and clinical similarity of the node features. Assume 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] Here, α is a weighting parameter used to balance the importance of cosine similarity and clinical similarity.

[0101] The attention coefficients are normalized using the Softmax function to generate normalized attention weights. The formula for calculating the Softmax function is as follows:

[0102]

[0103] Where N i It is the set of neighboring nodes of node i, β ij It is the normalized attention weight, which represents the importance of node j to node i.

[0104] The embedding vector of the center node is updated by weighted summation of the features of neighboring nodes based on normalized attention weights. That is:

[0105]

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

[0107] Multiple graph attention layers are stacked, with each layer's output processed by Layer Normalization and the ReLU activation function. Layer Normalization is a normalization method that normalizes the features of each node, accelerating model training and convergence. The ReLU activation function introduces non-linearity, enhancing the model's expressive power. This multi-layer graph attention process allows for better extraction of node feature information.

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

[0109] The design incorporates a classification task branch, embedding nodes into vectors and inputting them into a fully connected layer. A fully connected layer is a common type of neural network layer where each neuron is connected to all neurons in the layer above it. The fully connected layer then calculates the DCM risk probability, outputting the result.

[0110] The design branch focuses on feature reconstruction, reconstructing the original joint feature matrix from node embedding vectors using a decoder network. The decoder network can employ a multi-layer neural network structure to map low-dimensional node embedding vectors back to a high-dimensional joint feature matrix space.

[0111] A joint loss function is constructed, which is a weighted sum of the classification cross-entropy loss and the reconstruction mean squared error loss. The classification cross-entropy loss measures the difference between the predicted result and the true label in a classification task, and its calculation formula is as follows:

[0112]

[0113] Where C is the number of categories, y i p is the i-th component of the real label. i This is the i-th component of the predicted probability. The reconstruction mean squared error loss measures the difference between the reconstructed joint feature matrix and the original joint feature matrix, and is calculated using the following formula:

[0114]

[0115] Where n is the number of features, x i It is the i-th element of the original joint feature matrix. It is the i-th element of the reconstructed joint feature matrix. The joint loss function is 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 employed to update the network parameters for both the classification and reconstruction tasks. During training, the network parameters for the reconstruction task are first fixed while the parameters for the classification task are updated; then, the parameters for the classification task are fixed while the parameters for the reconstruction task are updated, and this process is repeated alternately until the model converges. This approach allows for simultaneous optimization of both classification and reconstruction tasks, improving model performance.

[0117] Example 5:

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

[0119] A personalized diagnostic decision tree is constructed based on the DCM risk probability. The branch thresholds of the decision tree are optimized using a genetic algorithm to generate the optimal diagnostic path. Early DCM diagnostic results are then output based on the optimal diagnostic path. The process of constructing the 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" or "medium risk requires regular follow-up").

[0123] 2) Decision tree initialization design

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

[0125] Intermediate Nodes: Secondary branches are generated by combining clinical indicators (such as LVEF < 40%) with the feature reconstruction error (outlier detection) output by the reconstruction module.

[0126] Leaf nodes: Output diagnostic conclusions (confirmed DCM, suspected cases, risk excluded) and confidence levels.

[0127] 3) Feature Interaction Enhancement

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

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

[0130] The optimal diagnostic path is generated by optimizing the branch threshold of the decision tree using a genetic algorithm, as follows:

[0131] The diagnostic decision tree is encoded. Each branch node of the decision tree is treated as a gene locus, 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 manipulate and optimize it. The decision tree plays a crucial role in classification and decision-making throughout the diagnostic process. It makes diagnostic judgments by layering features extracted from multi-radiomics and serum biomarker data. Encoding its key parameters in this way forms the basis for subsequent optimization using genetic algorithms.

[0132] Multiple individuals with different gene values ​​are randomly generated, forming the initial population. Each individual represents a possible combination of decision tree branch thresholds, much like randomly selecting some initial "exploration directions" from many possible diagnostic paths. For example, 50 individuals are randomly generated, each containing the threshold gene values ​​corresponding to each branch node of the decision tree. These initial values ​​are determined completely randomly to ensure population diversity and provide sufficiently rich "materials" for the subsequent optimization process.

[0133] A fitness function is constructed to evaluate the diagnostic performance of each individual. It is calculated based on the confusion matrix of the diagnostic path. The confusion matrix is ​​a commonly used tool in classification tasks, clearly showing the difference between the actual classification results and the true situation. Here, we mainly consider the sensitivity, specificity, and F1 score.

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

[0135] These indicators are weighted and combined 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 the fitness assessment of an individual. By reasonably adjusting the weights, we can pay more attention to the importance of certain indicators in diagnosis.

[0136] A tournament selection strategy is used when selecting individuals. K individuals are randomly selected from the current population to form a tournament group; k can be set according to the actual situation, for example, 3. 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 "competitions," each time selecting relatively superior individuals to enter the "candidate breeding pool." The tournament selection is repeated continuously until the mating pool is full.

[0137] Once the mating pool is ready, individuals in the pool are paired up, and a random crossover point is selected for gene exchange, thus generating offspring. This crossover operation simulates gene recombination in biological heredity; by exchanging partial genes between two individuals, it is possible to create new individuals with greater advantages. For example, individuals A and B, after randomly selecting a crossover point, exchange their gene values, producing new offspring. These offspring inherit some characteristics from their parents and may also generate new combinations, providing new possibilities for finding better diagnostic pathways.

[0138] For offspring individuals, a mutation operation is also performed. Each gene locus is randomly mutated with a preset probability, and the mutation range is limited to the feasible interval of the branch threshold. The preset probability is usually a small value, such as 0.05, meaning that each gene locus has a 5% probability of mutating. The purpose of mutation is to prevent the algorithm from getting trapped in local optima too early. By introducing new gene values, the algorithm can explore a wider solution space, potentially discovering better combinations of branch thresholds.

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

[0140] The population is continuously iterated and updated, with selection, crossover, and mutation operations repeated until a pre-set maximum number of generations is reached. When the maximum number of 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 constitute the optimal diagnostic path. Based on this optimal diagnostic path, more accurate and scientific early DCM diagnostic results can be output, providing strong diagnostic support for clinicians and helping them more accurately determine whether a patient has early DCM, thus enabling timely and appropriate treatment.

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

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

Claims

1. A method for early non-invasive analysis of DCM based on multi- image omics and serum markers, characterized in that, The method comprises the following steps: acquiring image data of a child patient by a multi-modal medical imaging device, wherein the multi-modal medical imaging device comprises a cardiac MRI, an echocardiogram and a cardiac CT; acquiring serum marker data by a blood detection device; extracting imageomics features from the image data based on a multi-scale feature extraction algorithm to generate an imageomics feature set; modeling dynamic changes of the serum marker data based on a time series analysis model to generate a serum marker time series feature set, including: acquiring serum marker concentration data at multiple time points, constructing a time series matrix; decomposing the time series matrix into a trend item, a seasonal item and a residual item by using an autoregressive integrated moving average model; aligning time series data of multiple patients based on a dynamic time warping algorithm to eliminate individual sampling time bias; constructing a time series prediction model based on a long short-term memory network, inputting the aligned time series data to predict marker concentration changes at future time points; extracting hidden state vectors of the time series prediction model as dynamic change features; combining the trend item, the seasonal item, the residual item and the dynamic change features to generate the serum marker time series feature set; performing cross-modal feature fusion on the imageomics feature set and the serum marker time series feature set by using a dynamic weighted fusion strategy to generate a joint feature matrix; inputting the joint feature matrix into a pre-trained multi-task deep learning model, wherein the multi-task deep learning model comprises a classification module and a reconstruction module, the classification module is used to perform a DCM classification task to predict a disease risk probability, and the reconstruction module is used to perform a feature reconstruction task to reconstruct the joint feature matrix, the node embedding vectors of the classification module and the reconstruction module are interactively learned through a graph neural network architecture, and the two types of tasks are simultaneously optimized based on a joint loss function, and a DCM risk probability is output; wherein the construction steps of the multi-task deep learning model comprise: constructing a patient feature graph, wherein a node represents a joint feature vector of a single patient, and an edge represents the clinical similarity between patients, and the clinical similarity is calculated based on age, gender and underlying disease matching degree; aggregating node features by using a graph attention mechanism to generate node embedding vectors; designing a classification task branch, inputting the node embedding vectors into a fully connected layer, and outputting a DCM risk probability; constructing a personalized diagnosis decision tree according to the DCM risk probability, wherein the initialization design of the decision tree comprises: a root node: based on the DCM risk probability, the first branch is divided into high risk and low risk; an intermediate node: combining clinical indicators and feature reconstruction errors output by the reconstruction module to generate a secondary branch; a leaf node: outputting a diagnosis conclusion and a confidence degree, wherein the diagnosis conclusion comprises a confirmed DCM, a suspected case and an excluded risk; optimizing the branch threshold of the decision tree based on a genetic algorithm to generate an optimal diagnosis path; and outputting an early DCM diagnosis result based on the optimal diagnosis path.

2. The method for early non-invasive analysis of DCM based on multi-imagomics and serum markers according to claim 1, characterized in that, The method for extracting imageomics features from the image data based on a multi-scale feature extraction algorithm to generate an imageomics feature set comprises: performing three-dimensional wavelet transform on the cardiac MRI data to extract texture features, morphological features and intensity distribution features under multiple frequency bands; the intensity distribution features are intensity distribution features based on a gray histogram, including mean, variance and skewness; performing dynamic time warping on the echocardiogram data to generate time sequence features aligned with a cardiac cycle, and extracting local texture features through a gray level co-occurrence matrix; performing blood vessel segmentation on the cardiac CT data, and extracting topological features and geometric features based on the segmented blood vessel structure; performing standardization processing on the texture features, morphological features, intensity distribution features, time sequence features, local texture features, topological features and geometric features to eliminate inter-modality dimensional differences; performing dimensionality reduction on the standardized features by using principal component analysis to remove redundant features, and generating an image feature set.

3. The method for early non-invasive analysis of DCM based on multi-imaging and serum markers according to claim 1, characterized in that, the cross-modality feature fusion of the image feature set and the serum marker time sequence feature set by using a dynamic weighted fusion strategy to generate a joint feature matrix includes: constructing an inter-modality similarity measurement function to calculate the correlation degree of the image features and the serum marker features based on cosine similarity; assigning a dynamic weight to each image feature according to the correlation degree, and the weight value is positively correlated with the correlation degree; performing standardization weighting on the serum marker time sequence features, and the weight value is determined based on the reciprocal of the feature variance; performing horizontal splicing on the weighted image features and the standardized weighted serum marker time sequence features to generate a preliminary fusion matrix; performing nonlinear dimensionality reduction on the preliminary fusion matrix by using kernel principal component analysis to eliminate cross-modality redundant information, and generating a joint feature matrix.

4. The method for early non-invasive analysis of DCM based on multi- imaging omics and serum markers according to claim 1, characterized in that, after outputting the DCM risk probability, further comprising: designing a feature reconstruction task branch to reconstruct the original joint feature matrix from the node embedding vector through a decoder network; constructing a joint loss function, which includes a weighted sum of classification cross-entropy loss and reconstruction mean square error loss; updating the network parameters of the classification task and the reconstruction task by using an alternating optimization strategy until the model converges.

5. The method for early non-invasive analysis of DCM based on multi-imagomics and serum markers according to claim 1, characterized in that, the optimization of the branch threshold value of the decision tree based on the genetic algorithm to generate an optimal diagnostic path includes: encoding each branch node of the diagnostic decision tree as a gene bit and encoding the branch threshold value as a gene value; initializing a population to randomly generate multiple individuals containing different gene values; constructing a fitness function based on the confusion matrix indicators of the diagnostic path, including a weighted combination of sensitivity, specificity and F1 score; selecting high fitness individuals by using a tournament selection strategy, generating a child population through single-point crossover and mutation operations; iteratively updating the population until the maximum evolution generation is reached, selecting the individual with the highest fitness to decode as the optimized branch threshold value to generate the optimal diagnostic path.

6. The method for early non-invasive analysis of DCM based on multi-imaging and serum markers according to claim 2, characterized in that, the performing of three-dimensional wavelet transform on the cardiac MRI data to extract texture features, morphological features and intensity distribution features under multiple frequency bands includes: The three-dimensional discrete wavelet transform is used to decompose the cardiac MRI data into a low-frequency approximation component and a high-frequency detail component; the intensity distribution features based on the gray histogram are extracted from the low-frequency approximation component, including the mean, variance and skewness; the texture features based on the Gabor filter are extracted from each high-frequency detail component, including the energy, entropy and contrast; the three-dimensional volume, surface area and sphericity index of the cardiac chamber are extracted based on the morphological operation; and the intensity distribution features, texture features and morphological features are integrated in layers according to the frequency band to generate the multi-scale imageomics features.

7. The method for early non-invasive analysis of DCM based on multi-imaging and serum markers according to claim 1, characterized in that, The time series prediction model based on the long short-term memory network is constructed, the aligned time series data is input, and the marker concentration change at a future time point is predicted, including: The time series data is divided into sliding windows, each window containing a sequence of marker concentrations at consecutive time points; a bidirectional long short-term memory network is constructed, the window sequence is input, and the concentration prediction value at the next time point is output; a self-attention mechanism is introduced in the hidden layer to dynamically adjust the weight of different time steps; The mean square error is used as the loss function, and the network parameters are updated by 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.

8. The method for early non-invasive analysis of DCM based on multi-imaging and serum markers according to claim 4, characterized in that, The graph attention mechanism is used to aggregate node features to generate node embedding vectors, including: The attention coefficients between nodes are calculated, which are based on the weighted sum of the cosine similarity of node features and clinical similarity; the attention coefficients are normalized by the Softmax function to generate normalized attention weights; the embedding vectors of the center node are updated by weighted sum of the features of the neighbor nodes according to the normalized attention weights; Multiple graph attention layers are stacked, and each layer outputs are processed by Layer Normalization and ReLU activation function; finally, the node embedding vectors of the last layer are extracted as the input of the classification and reconstruction tasks.

9. The method for early non-invasive analysis of DCM based on multi-imaging and serum markers according to claim 5, characterized in that, The tournament selection strategy is used to screen high fitness individuals, and the offspring population is generated by single-point crossover and mutation operation, including: k individuals are randomly selected from the current population to form a tournament group, and the individual with the highest fitness is selected into the mating pool; the tournament selection is repeated until the mating pool is filled; the individuals in the mating pool are paired, a crossover point is randomly selected for gene exchange to generate offspring individuals; each gene site of the offspring individuals is randomly mutated with a preset probability, and the mutation range is the feasible interval of the branch threshold; the top 10% of individuals in the parent population are directly entered into the next generation.

Images