Detection and risk prediction method for multiple pathological changes caused by myocardial ischemia and related equipment

Abstract

The invention discloses a detection and risk prediction method for multiple pathological changes caused by myocardial ischemia and related equipment, and the method comprises the steps: calling a preset multiple pathological change detection model through a data processing terminal to carry out multiple pathological change detection on CCTA time sequence images after myocardial ischemia, and calculating a myocardial rescue index according to a detection result; extracting a plurality of main adverse cardiovascular event factors according to a detection result, and generating a time sequence radiomics feature sequence in combination with a myocardial rescue index; and calling a preset cascade risk prediction framework to perform risk hierarchical prediction according to the time sequence clinical data and the time sequence radiomics feature sequence, and generating a structured risk report according to a prediction result. Therefore, a multi-pathological change detection model is adopted to detect on the basis of the CCTA time sequence image and time sequence clinical data, and meanwhile, risk hierarchical prediction is performed based on a cascade risk prediction framework, so that dynamic pathological change detection and risk prediction are more accurately performed on main adverse cardiovascular events.

Description

Technical Field

[0001] This invention relates to the field of pathological change detection technology, and in particular to a method and related equipment for detecting and predicting multiple pathological changes caused by myocardial ischemia. Background Technology

[0002] Myocardial infarction (MI) is myocardial necrosis caused by myocardial ischemia and an imbalance between oxygen supply and demand. Following acute myocardial infarction (AMI), myocardial tissue often undergoes various pathological changes, including edema, necrosis, hemorrhage, and microcirculatory obstruction (MVO). These changes are key factors determining the patient's prognosis. Currently, cardiac magnetic resonance imaging (CMR) with multiple sequence scans (such as T2WI, LGE, T2...) is crucial for prognosis. Coronary artery mapping is the gold standard for distinguishing the above pathological changes, but this technology has limitations such as long examination time, high cost, poor tolerance to acute patients, and low popularity. It also cannot complete coronary artery assessment in one stop, thus limiting its widespread application in the accurate diagnosis of acute myocardial infarction.

[0003] To compensate for the shortcomings of CMR, various alternative technologies have been developed in related fields, such as iodine-delayed enhancement technology based on coronary computed tomography angiography and AI-assisted diagnostic solutions.

[0004] However, the above methods cannot achieve simultaneous, rapid, and accurate identification of multiple pathological changes such as myocardial edema, necrosis, hemorrhage, and microcirculatory obstruction in the acute phase of AMI. They also lack the ability to integrate radiomics features and clinical time-series data to perform individualized dynamic risk stratification prediction of major adverse cardiovascular events (MACE), making it difficult to meet the integrated clinical needs for accurate diagnosis and prognostic assessment of AMI patients in the acute phase. Summary of the Invention

[0005] This invention provides a method and related equipment for detecting and predicting multiple pathological changes caused by myocardial ischemia. It solves the technical problems of traditional technologies, which cannot simultaneously, rapidly and accurately identify multiple pathological changes such as myocardial edema, necrosis, hemorrhage and microcirculation obstruction in the acute phase of acute myocardial infarction (AMI), and lack the ability to integrate radiomics features and clinical time-series data to perform individualized dynamic risk stratification prediction of major adverse cardiovascular events (MACE), making it difficult to meet the integrated clinical needs for accurate diagnosis and prognostic assessment of acute AMI patients.

[0006] The first aspect of this invention provides a method for detecting and predicting the risk of multiple pathological changes caused by myocardial ischemia, applied to a locally deployed data processing terminal, the method comprising: According to the detection cycle, CCTA time-series images and time-series clinical data were acquired after myocardial ischemia occurred in multiple detection targets; The preset multi-pathological change detection model is invoked to detect multiple pathological changes in each of the CCTA time-series images, and the myocardial salvage index is calculated according to the detection results. Based on the detection results, multiple major adverse cardiovascular event factors were extracted and combined with the myocardial salvage index to generate a time-series radiomics feature sequence. The system invokes a pre-defined cascaded risk prediction framework to perform risk stratification prediction based on the time-series clinical data and the time-series radiomics feature sequences, and generates a structured risk report based on the prediction results.

[0007] Optionally, the multi-pathological change detection model includes an encoder, a decoder, and a multi-task output head; the step of calling the preset multi-pathological change detection model to perform multi-pathological change detection on each of the CCTA time-series images, and calculating the myocardial salvage index according to the detection results includes: After standardizing each CCTA time series image, it is cropped into image blocks of the same size; The encoder is invoked to extract global pathological dependency features of different scales from each of the image blocks according to a multi-head attention mechanism; The decoder is invoked to perform multi-scale fusion of the global pathological dependent features and then the boundaries are sharpened to obtain the target pathological feature map. The multi-task output head is invoked to perform voxel binary classification on the target pathological feature map to determine the myocardial region and non-myocardial region. The multi-task output head is invoked to perform voxel six-class classification on the myocardial region to determine the probability of pathological changes corresponding to each type of pathological change in the myocardial region. The myocardial salvage index is calculated based on the probability of each pathological change and a preset calculation formula.

[0008] Optionally, the step of extracting multiple major adverse cardiovascular event factors according to the detection results and combining them with the myocardial salvage index to generate a time-series radiomics feature sequence includes: Multiple radiomics features are extracted from the pathological regions corresponding to each type of pathological change in the detection results to construct a high-dimensional feature matrix; After normalizing each of the radiomics features in the high-dimensional feature matrix, unsupervised dimensionality reduction is performed to obtain a low-dimensional feature matrix. The correlation between each low-dimensional feature in the low-dimensional feature matrix and major adverse cardiovascular events is calculated using a variational autoencoder. Based on the aforementioned relevance and feature interpretability, target low-dimensional features are selected from the low-dimensional feature matrix; The target low-dimensional features and the myocardial salvage index are spliced ​​together according to the detection cycle to generate a time-series radiomics feature sequence.

[0009] Optionally, the cascaded risk prediction framework includes a time-series processing layer, a risk prediction layer, and a report generation layer; the step of calling the preset cascaded risk prediction framework to perform risk stratification prediction based on the time-series clinical data and the time-series radiomics feature sequence, and generating a structured risk report according to the prediction results, includes: The time-series processing layer constructs individualized time-risk curves for each detection target based on time-series clinical data and time-series radiomics feature sequences within historical time periods. The risk prediction layer maps the individualized time-risk curve into a time-series feature vector of a preset dimension. The risk prediction layer uses the time-series feature vector, current time-series radiomics features, and current time-series clinical data to map to the semantic space and fuse them to generate a fusion feature corresponding to each detection target. The risk prediction layer performs risk classification according to all the fusion features to obtain the major adverse cardiovascular event level corresponding to each detection target; The report generation layer calls an attribution algorithm to classify the contribution of each time-series radiomics feature sequence, and combines the time-series clinical data and the major adverse cardiovascular event levels to generate a structured risk report corresponding to each detection target.

[0010] Optionally, the step of constructing individualized time-risk curves for each detection target based on time-series clinical data and time-series radiomics feature sequences within a historical time period through the time-series processing layer includes: The time-series processing layer aligns time-series clinical data and time-series radiomics feature sequences according to each time point within a historical time period, and constructs a multi-dimensional time-series matrix corresponding to each time point. The temporal processing layer invokes a temporal convolutional network to perform causal convolution and dilated convolution on the multidimensional time series matrix, and combines residual connections to determine the risk probability of adverse cardiovascular events at each time point; The time-series processing layer constructs individualized time-risk curves for each detection target by using the time point as the horizontal axis and the risk probability as the vertical axis.

[0011] Optionally, the step of mapping and fusing the temporal feature vector, current temporal radiomics features, and current temporal clinical data to the semantic space through the risk prediction layer to generate a fused feature corresponding to each detection target includes: The risk prediction layer calls an encoder based on contrast loss optimization to encode the temporal feature vector, the current temporal radiomics features, and the current temporal clinical data respectively, to obtain initial features of various modalities; The risk prediction layer maps various initial features to a semantic space of a preset dimension to obtain multiple cross-modal features corresponding to each detection target. The risk prediction layer performs weighted fusion of the cross-modal features according to the attention mechanism to generate a fused feature corresponding to each detection target.

[0012] Optionally, the data processing terminal is communicatively connected to a cloud server, and the method further includes: The cloud server periodically sends the initial global model to multiple data processing terminals through an encrypted channel; After privacy-desensitizing newly added time-series clinical data stored locally, local fine-tuning data is constructed by combining newly added CCTA images; After training the initial global model using the local fine-tuning data, the model update weights corresponding to all model parameters are extracted. Calculate the weight update amount between each of the model update weights and the original model weights of the initial global model, and encrypt and send it to the cloud server; Once the cloud server receives and decrypts all encrypted weight updates, it performs a weighted average of the weight updates to obtain the global update value. The cloud server uses the global update amount to update the model parameters of the initial global model to obtain a new initial global model, and then jumps to execute the step of the cloud server periodically sending the initial global model to multiple data processing terminals through an encrypted channel.

[0013] Optionally, the method further includes: If the target being detected meets the preset physical condition, then the plain CT image of the target being detected is obtained according to the detection cycle; A preset generative adversarial network is invoked to generate CCTA time-series images corresponding to each of the aforementioned plain CT images; Jump to execute the step of calling the preset multi-pathological change detection model to perform multi-pathological change detection on each of the CCTA time series images, and calculate the myocardial salvage index according to the detection results.

[0014] A second aspect of the present invention provides a device for detecting and predicting the risk of multiple pathological changes caused by myocardial ischemia, applied to a locally deployed data processing terminal, the device comprising: The data periodic acquisition module is used to acquire CCTA time-series images and time-series clinical data of multiple detection targets after myocardial ischemia occurs, according to the detection cycle; The multi-pathological change detection module is used to call a preset multi-pathological change detection model to perform multi-pathological change detection on each of the CCTA time-series images, and calculate the myocardial salvage index according to the detection results. The feature generation module is used to extract multiple major adverse cardiovascular event factors according to the detection results, and combine them with the myocardial salvage index to generate a time-series radiomics feature sequence. The risk stratification prediction module is used to call a preset cascaded risk prediction framework to perform risk stratification prediction based on the time-series clinical data and the time-series radiomics feature sequence, and generate a structured risk report according to the prediction results.

[0015] A third aspect of the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the method for detecting and predicting multiple pathological changes caused by myocardial ischemia as described in any one of the first aspects of the present invention.

[0016] As can be seen from the above technical solutions, the present invention has the following advantages: A locally deployed data processing terminal acquires CCTA time-series images and time-series clinical data of multiple targets after myocardial ischemia according to the detection cycle. A pre-defined multi-pathological change detection model is invoked to detect multiple pathological changes in each CCTA time-series image, and the myocardial salvage index is calculated based on the detection results. Multiple major adverse cardiovascular event factors are extracted based on the detection results and combined with the myocardial salvage index to generate a time-series radiomics feature sequence. A pre-defined cascaded risk prediction framework is invoked to perform risk stratification prediction based on time-series clinical data and time-series radiomics feature sequence, and a structured risk report is generated based on the prediction results. Therefore, by employing a multi-pathological change detection model based on CCTA time-series images and time-series clinical data for multi-pathological change detection, and simultaneously performing risk stratification prediction based on the cascaded risk prediction framework, more accurate dynamic pathological change detection and risk prediction of major adverse cardiovascular events can be achieved. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the steps of a method for detecting and predicting the risk of multiple pathological changes caused by myocardial ischemia, as provided in an embodiment of the present invention. Figure 2 This is a structural block diagram of a device for detecting and predicting the risk of multiple pathological changes caused by myocardial ischemia, provided in an embodiment of the present invention. Detailed Implementation

[0019] This invention provides a method and related equipment for detecting and predicting multiple pathological changes caused by myocardial ischemia. It addresses the technical problem that the system cannot simultaneously, rapidly, and accurately identify multiple pathological changes such as myocardial edema, necrosis, hemorrhage, and microcirculatory obstruction in the acute phase of acute myocardial infarction (AMI), and lacks the ability to integrate radiomics features and clinical time-series data to perform individualized dynamic risk stratification prediction of major adverse cardiovascular events (MACE), thus failing to meet the integrated clinical needs for accurate diagnosis and prognostic assessment of acute AMI patients.

[0020] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0021] Please see Figure 1 , Figure 1 This is a flowchart illustrating the steps of a method for detecting and predicting the risk of multiple pathological changes caused by myocardial ischemia, as provided in an embodiment of the present invention.

[0022] This invention provides a method for detecting and predicting the risk of multiple pathological changes caused by myocardial ischemia, applied to a locally deployed data processing terminal. The method includes: Step 101: Acquire CCTA time-series images and time-series clinical data of multiple detection targets after myocardial ischemia occurs, according to the detection cycle; CCTA time-series images refer to a series of images obtained by performing coronary CT angiography on the target at different time points (such as the acute phase and the recovery phase). They are used to observe the dynamic evolution of myocardial pathological changes, including but not limited to single-energy CCTA or dual-energy CCTA images, and can cover various reconstruction forms such as volume rendering and surface reconstruction, to reflect the temporal evolution of myocardial pathological changes.

[0023] Time-series clinical data refers to clinical information that changes dynamically over time and is synchronized with the CCTA time-series image acquisition cycle. It includes basic information (age, gender), biochemical indicators (myocardial enzyme profile, N-terminal pro-B-type natriuretic peptide (NT-proBNP)), cardiac function indicators (ejection fraction (EF)), comorbidities (history of diabetes, history of hypertension), etc.

[0024] In this embodiment of the invention, since time-series clinical data and CCTA time-series images may involve the privacy of the detection target, to avoid data leakage, the data processing terminal is a locally deployed terminal device that is not connected to the network. The data processing terminal automatically retrieves the CCTA image data and corresponding electronic medical record data of the target patient from the hospital information system or image archiving system according to a preset follow-up plan (such as day 1, day 7, day 30, day 90 after the occurrence of myocardial ischemia event). These clinical data include indicators such as myocardial enzyme spectrum, NT-proBNP concentration, and ventricular ejection fraction measured by transthoracic echocardiography at that time point, and are time-aligned and correlated in the local database to form a structured longitudinal dataset of the patient.

[0025] After acquiring CCTA image data, the format can be standardized, Z-score normalization can be used for time-series clinical data to eliminate dimensional differences, and Gaussian process interpolation can be used to complete missing data to ensure data integrity and consistency.

[0026] Step 102: Call the preset multi-pathological change detection model to perform multi-pathological change detection on each CCTA time series image, and calculate the myocardial salvage index according to the detection results. The multi-pathological change detection model refers to an end-to-end multi-task deep learning model built based on an encoder, decoder, and multi-task output head. The encoder-decoder structure is constructed based on a hybrid architecture of 3D Transformer and U-Net, enabling simultaneous left ventricular myocardial segmentation and multi-pathological differentiation, outputting six-classification results for the left ventricular myocardial region and intramyocardial edema, necrosis, hemorrhage, microcirculatory obstruction (MVO), radiosclerosis artifacts, and normal myocardium.

[0027] The myocardial salvage index (MSI) is a key indicator used to quantify the proportion of myocardial tissue that can be salvaged after myocardial ischemia. Its calculation formula is MSI = (AAR - IS) / AAR, where AAR (risk area) refers to the myocardial area at risk of necrosis (the sum of edema, necrosis, hemorrhage, and MVO areas), and IS (infarct area) refers to the myocardial area that has suffered irreversible damage (the sum of necrosis, hemorrhage, and MVO areas).

[0028] In this embodiment, the terminal is loaded with a pre-trained multi-pathological change detection model. After acquiring the CCTA time-series images and time-series clinical data corresponding to each detection cycle, the pre-processed CCTA time-series data is input into the multi-pathological change detection model. The encoder in the model captures the long-range contextual dependencies of myocardial pathological regions, such as distinguishing between edema and MVO with similar textures. Combined with the decoder, fine upsampling is performed to output a left ventricular myocardial binary mask (myocardial / non-myocardial) and a six-class probability map of myocardial tissue (1. edema, 2. necrosis, 3. hemorrhage, 4. MVO, 5. radiosclerosis artifact, 6. normal myocardium).

[0029] After further pixel spacing calibration based on CT equipment, the number of voxels in each pathological region is counted based on the classification results and converted into actual area to obtain AAR (sum of areas of edema, necrosis, hemorrhage and MVO) and IS (sum of areas of necrosis, hemorrhage and MVO). These are then substituted into the MSI calculation formula to automatically calculate the myocardial salvage index for each detection cycle.

[0030] Furthermore, for this multi-pathological change detection model, end-to-end training can be performed using multi-center labeled datasets during the pre-training phase. The labeled datasets include cardiac magnetic resonance (CMR) multi-sequence imaging (T2WI, LGE, T2...). The mapping result is considered the gold standard. Further training is performed using a joint loss function (α×Dice loss + β×Focal loss + γ×MSE loss), balancing the weights α, β, and γ across the three tasks to ensure both segmentation accuracy and pathological differentiation accuracy.

[0031] In one example of the present invention, the multi-pathological change detection model includes an encoder, a decoder, and a multi-task output head; step 102 includes the following sub-steps: After standardizing each CCTA time series image, it is cropped into image blocks of the same size; The encoder is invoked to extract global pathological dependency features at different scales from each image patch using a multi-head attention mechanism; The decoder is invoked to perform multi-scale fusion of various global pathological dependent features and then the boundaries are sharpened to obtain the target pathological feature map; The multi-task output head is invoked to perform voxel binary classification on the target pathological feature map to determine the myocardial region and non-myocardial region. The multi-task output head is invoked to perform voxel six-class classification on the myocardial region to determine the probability of pathological changes corresponding to each type of pathological change in the myocardial region. The myocardial salvage index is calculated based on the probability of each pathological change and a preset calculation formula.

[0032] In this embodiment, after obtaining CCTA time-series images, images generated by different devices are uniformly converted to the clinically common DICOM format. Gray-level normalization is performed based on the image gray-level histogram to eliminate gray-level deviations under different scanning conditions. A preset ray-sclerotherapy artifact correction algorithm is used to suppress strip-shaped or cup-shaped low-density artifacts around high-density materials (such as calcification and metal), restoring the true density information of the myocardium. Subsequently, based on the anatomical location of the left ventricle (referencing cardiac anatomical landmarks and previous coarse localization model results), image blocks containing the complete left ventricular myocardium are cropped from the standardized images to ensure that all image blocks have consistent coverage and size, avoiding the impact of image size differences on the accuracy of model feature extraction.

[0033] The encoder initiates a multi-head attention mechanism through the 3D Transformer unit. One attention head focuses on local areas within the image patch to capture the detailed features of small pathological lesions (such as microcirculation obstruction), while another attention head focuses on the global myocardial structure to learn the pathological associations between different segments of the left ventricular wall. Through multi-scale feature fusion, local detailed features and global structural features are integrated to generate global pathological dependency features containing information at different scales, laying the foundation for accurate identification of subsequent pathological regions.

[0034] The decoder, based on the U-Net architecture, is used to restore the global pathological dependency features extracted by the encoder into feature maps that match the size of the original image patches. After obtaining global pathological dependency features at different scales, the high-dimensional features are first mapped to a size matching the original image patches through upsampling. Then, the features at different levels are fused according to their spatial locations to ensure that the detailed features of small pathological lesions are preserved without losing the overall structural information of the myocardium. After fusion, a boundary sharpening algorithm is activated to enhance the edge contrast between the pathological area and the normal myocardial area and eliminate boundary blurring. Finally, a target pathological feature map is generated, which can intuitively reflect the location, extent, and boundary morphology of potential pathological areas within the myocardium.

[0035] The multi-task output head can simultaneously perform multiple classification tasks such as segmentation and pathological identification. After obtaining the target pathological feature map, the binary classification branch within the multi-task output head classifies each voxel into a myocardial region or a non-myocardial region based on the density, texture, and spatial location features of voxels in the target pathological feature map, combined with preset myocardial anatomical features (such as the morphology and location of the left ventricular wall). A binary classification result mask is output, where the portion marked as a myocardial region will be used as the scope for subsequent pathological identification, while non-myocardial regions are masked to avoid interference from irrelevant structures in the pathological analysis. The six-classification branch uses the myocardial region within the binary classification result as the output. Based on the density features, texture patterns, and association with surrounding tissues of the voxels within this region (referencing the pathological features labeled in the gold standard of cardiac magnetic resonance imaging), it calculates the classification probability for each voxel in six categories. Specifically, for edema, necrosis, hemorrhage, and microcirculatory obstruction, the focus is on the typical density manifestations and distribution patterns of each pathological change in CCTA images. For radiation sclerosis artifacts, the focus is on identifying characteristic density abnormalities around high-density materials. For normal myocardium, the focus is on identifying myocardial density and texture consistent with physiological conditions. Finally, the probability of six pathological change types corresponding to each voxel is output, with the type with the highest probability initially identified as the pathological type for that voxel. The number of voxels corresponding to each pathological type within the myocardial region is counted, and the danger area (the total number of voxels corresponding to edema, necrosis, hemorrhage, and microcirculatory obstruction, converted to the actual myocardial region) and infarct area (the total number of voxels corresponding to necrosis, hemorrhage, and microcirculatory obstruction, converted to the actual myocardial region) are calculated. Substituting the danger area and infarct area into a preset calculation formula, the myocardial salvage index is automatically calculated. This index reflects the proportion of reversibly damaged myocardium after myocardial ischemia. The calculation formula is as follows:

[0036] in, For voxel danger area, Voxel infarction area It is the myocardial salvage index.

[0037] Step 103: Extract multiple major adverse cardiovascular event factors based on the detection results, and combine them with the myocardial salvage index to generate a time-series radiomics feature sequence; Major adverse cardiovascular event factors refer to radiomics features extracted from pathological segmentation regions of CCTA time-series images that are strongly correlated with the risk of major adverse cardiovascular events (MACE). These include, but are not limited to, first-order features (gray-level mean, variance, skewness), texture features (entropy, contrast, correlation of gray-level co-occurrence matrix), and wavelet features (multi-scale frequency domain features).

[0038] Temporal radiomics feature sequences refer to feature sequences formed by integrating MACE factors and myocardial salvage indices of each detection cycle in chronological order. These sequences can reflect the temporal evolution of pathological changes and prognostic indicators.

[0039] In this embodiment, after obtaining the detection results output by the multi-pathological change detection model, various radiomics features are extracted from the pathological regions corresponding to each type of pathological change within the detection results (e.g., edema, necrosis, hemorrhage, MVO, etc.) to obtain an initial high-dimensional feature set, including but not limited to first-order grayscale features (such as the mean and variance of CT values ​​of the pathological region), texture features (such as the entropy and contrast calculated from the gray-level co-occurrence matrix), and wavelet features (such as the mean of low-frequency coefficients after wavelet transform). The maximum correlation minimum redundancy algorithm is used to screen features strongly correlated with MACE, and combined with minimum absolute contraction and the LASSO regression selection operator, dimensionality reduction and redundancy removal are further performed to obtain the main adverse cardiovascular event factors for each detection cycle. Furthermore, the MACE factors of each detection cycle are concatenated with the MSI calculated in the corresponding cycle in chronological order, such as 64×3+1×3=195 dimensions for 3 detection cycles, forming a temporal radiomics feature sequence. The time dimension of the feature sequence corresponds one-to-one with the detection cycle, ensuring that it reflects the dynamic changes of pathological changes and MSI, providing temporal information support for subsequent risk prediction.

[0040] In one example of the present invention, step 103 may include the following sub-steps: Multiple radiomics features were extracted from the pathological regions corresponding to each type of pathological change in the detection results to construct a high-dimensional feature matrix. After normalizing the image omics features within the high-dimensional feature matrix, unsupervised dimensionality reduction is performed to obtain a low-dimensional feature matrix. The correlation between each low-dimensional feature in the low-dimensional feature matrix and major adverse cardiovascular events is calculated using a variational autoencoder. Based on relevance and feature interpretability, target low-dimensional features are selected from the low-dimensional feature matrix; Based on the detection cycle, target low-dimensional features and myocardial salvage index are spliced ​​together to generate a time-series radiomics feature sequence.

[0041] The pathological change type refers to the pathological state related to myocardial ischemia identified by the multi-pathological change detection model, including myocardial edema, myocardial necrosis, myocardial hemorrhage, myocardial microcirculation obstruction (MVO), radiographic artifacts, and normal myocardium. Each type corresponds to an independent pathological region.

[0042] Variational autoencoders (VAEs) are generative models that combine autoencoders and variational Bayesian inference. In this embodiment, they are used to compress features into a low-dimensional interpretable space and constrain them based on information bottlenecks, in order to screen out sparse features that are independently associated with major adverse cardiovascular events and are physically interpretable without the need for additional labels.

[0043] In this embodiment of the invention, multiple radiomics features are extracted for each pathological region corresponding to each pathological change, such as myocardial edema, necrosis, hemorrhage, and microcirculation obstruction. These features are then sorted and integrated according to the detection cycle. For example, each row corresponds to samples from one detection cycle, and each column corresponds to one radiomics feature, constructing a high-dimensional feature matrix. This ensures that each element in the matrix accurately corresponds to the feature value of a specific pathological region, and that no artifact region features are mixed in. Each radiomics feature in the high-dimensional feature matrix is ​​normalized, for example, using the Z-score normalization method. Subsequently, based on feature redundancy analysis, an unsupervised dimensionality reduction algorithm (such as principal component analysis (PCA) or t-distributed random neighborhood embedding (t-SNE)) is selected. PCA is suitable for linearly correlated features, while t-SNE is suitable for nonlinearly distributed features. The feature dimension after dimensionality reduction is determined by calculating the feature variance contribution (e.g., PCA retains principal components with a cumulative variance contribution ≥ 85%), transforming the high-dimensional feature matrix into a low-dimensional feature matrix. This ensures that the low-dimensional features retain the pathological information association of the original features to the greatest extent possible. The VAE encoder maps low-dimensional features to the mean and variance of the latent space through a fully connected layer. Based on the probability distribution of the latent space, it calculates the correlation between the latent variable corresponding to each low-dimensional feature and the MACE occurrence label (e.g., whether MACE occurred during the follow-up period). The overlap between the latent variable distribution and the MACE event distribution is quantified through mutual information, or through VAE decoder reconstruction error analysis. Specifically, if a low-dimensional feature is strongly correlated with MACE, removing that feature significantly increases the reconstruction error of MACE-related samples. Finally, the correlation quantification result between each low-dimensional feature and MACE is output. Highly correlated features are selected according to a correlation threshold and their interpretability is verified. For example, by mapping low-dimensional features back to the pathological region features of the original CCTA image, it is determined whether the feature can be associated with a clear pathological significance.

[0044] Specifically, if a low-dimensional feature corresponds to the first-order gray mean of the myocardial necrosis area, it has the interpretability of "abnormal density of necrosis area and MACE"; if a feature cannot correspond to specific pathological changes (such as abstract features without clear imaging or clinical significance), it is removed; finally, features that simultaneously satisfy high correlation and strong interpretability are retained as target low-dimensional features.

[0045] Finally, following the order of detection cycles (acute, subacute, chronic), the target low-dimensional features of each detection cycle are concatenated with the myocardial salvage index (MSI) calculated for the corresponding cycle. For example, if the target low-dimensional feature of a certain cycle is a D-dimensional vector and the MSI is a single quantized value, the concatenation forms a D+1-dimensional single-cycle feature vector. The D+1-dimensional feature vectors of all detection cycles are arranged in chronological order to form a time-series radiomics feature sequence. The time dimension of the sequence corresponds one-to-one with the detection cycle of the CCTA time-series images, ensuring that the sequence can fully reflect the feature evolution of the detection target at different pathological stages, providing time-series multi-dimensional feature input for the subsequent two-level cascaded risk prediction framework.

[0046] Step 104: Invoke the preset cascaded risk prediction framework to perform risk stratification prediction based on time-series clinical data and time-series radiomics feature sequences, and generate a structured risk report according to the prediction results.

[0047] The cascaded risk prediction framework refers to a machine learning architecture that combines temporal dynamic feature extraction with multimodal feature fusion to process longitudinal data and output personalized risk predictions. It includes a temporal processing layer, a risk prediction layer, and a report generation layer, performing risk stratification prediction by fusing time-series clinical data and time-series radiomics feature sequences.

[0048] A structured risk report is a standardized report that includes risk stratification results, quantitative analysis of key influencing factors, and quantitative pathological indicators, and is generated using the SHAP (SHapley Additive exPlanations) attribution algorithm.

[0049] In this embodiment of the invention, the terminal is also deployed with a pre-trained cascaded risk prediction framework. After obtaining time-series clinical data and time-series radiomics feature sequences, its time-series processing layer calls a time-series convolutional network to generate a personalized baseline risk curve that predicts the probability of MACE occurring at different future time points. Further, using this risk curve, current time-series radiomics features, and current time-series clinical data as input, the risk prediction layer uses a cross-modal alignment network based on contrastive learning to map these heterogeneous features to a unified semantic space and perform weighted fusion to generate fused features. A differentiable ranking learning layer then performs risk classification according to these fused features to determine the major adverse cardiovascular event (MAE) level corresponding to the detection target. Finally, the SHAP attribution algorithm is called to quantify the contribution of each input feature (such as MVO volume, MSI, NT-proBNP) to the risk prediction result (e.g., MVO volume contributes 30%), generating a structured risk report containing basic information about the detection target, quantitative pathological indicators for each cycle (such as edema volume, MSI), risk stratification level, key influencing factors, and their contribution percentages.

[0050] In one example of the present invention, the cascaded risk prediction framework includes a time-series processing layer, a risk prediction layer, and a report generation layer; step 104 may include the following sub-steps: S11. Based on the time-series clinical data and time-series radiomics feature sequences within historical time periods, the time-series processing layer constructs individualized time-risk curves for each detection target. In this embodiment, the historical time period can include at least one time point within each follow-up cycle of the key stage of pathological changes caused by myocardial ischemia. After obtaining the time-series clinical data and time-series radiomics feature sequences, other times besides the current time are used as the historical time period. The time-series processing layer retrieves the time-series clinical data and time-series radiomics feature sequences within the historical time period. Taking each detection cycle as the time point, the clinical data at the same time point, such as the myocardial enzyme spectrum detection value and EF value of that cycle, and the time-series radiomics features are classified according to feature type to ensure that the clinical and imaging features at each time point are completely aligned. Subsequently, the data is organized according to the "time point-feature" dimension, with rows corresponding to each historical time point and columns corresponding to each clinical / imaging feature, to construct a multi-dimensional time series matrix. Missing time point data in the matrix are filled in using Gaussian interpolation to ensure the integrity of the matrix.

[0051] Furthermore, S11 may include the following sub-steps: By aligning time-series clinical data and time-series radiomics feature sequences according to each time point within a historical time period through a time-series processing layer, a multi-dimensional time-series matrix corresponding to each time point is constructed. The temporal processing layer calls the temporal convolutional network to perform causal convolution and dilated convolution on the multidimensional time series matrix, and combines residual connections to determine the risk probability of adverse cardiovascular events at each time point; By using the time-series processing layer with time points as the horizontal axis and risk probability as the vertical axis, individualized time-risk curves are constructed for each detection target.

[0052] In this embodiment of the invention, the temporal processing layer strictly aligns the temporal clinical data (including myocardial enzyme profiles, N-terminal pro-B-type natriuretic peptide, ejection fraction, etc.) and temporal radiomics feature sequences (including radiomics features of each pathological change area and myocardial salvage index) at each time point within a historical time period. After imputing missing data using Gaussian interpolation, a multidimensional time series matrix corresponding to each time point is constructed with time points as rows and clinical / imaging features as columns. Then, a pre-trained temporal convolutional network (TCN) is invoked. The multidimensional time series matrix is ​​first processed by causal convolution combined with left padding to ensure that only current and historical features are used and to avoid future information leakage. Then, dilated convolution with increasing dilation rate is used to expand the receptive field to capture the long-term dependence of features at different time points. At the same time, residual connections are used to superimpose the input and output of the convolutional layer to alleviate gradient vanishing. Finally, the risk probability of major adverse cardiovascular events (MACE) occurring at each time point is output. Finally, using historical time points as the horizontal axis and corresponding MACE risk probabilities as the vertical axis, each coordinate point is connected by linear interpolation to construct an individualized time-risk curve for each detection target, intuitively presenting the dynamic trend of risk changes over time.

[0053] S12. The individualized time-risk curve is mapped to a time-series feature vector of a preset dimension through the risk prediction layer; In this embodiment of the invention, key features of the individualized time-risk curve (such as the peak risk time point, the slope of risk change, and the average risk of each period) are extracted, and these key features are mapped into vectors of a preset dimension through a fully connected network. L2 normalization is used in the mapping process to ensure that the vector magnitude is uniform and to avoid the fusion effect being affected by the difference in dimension. Finally, the time-series feature vector is output, which retains the core trend information of the time-risk curve.

[0054] S13. Through the risk prediction layer, the temporal feature vector, the current temporal radiomics features and the current temporal clinical data are mapped to the semantic space and fused to generate the fusion feature corresponding to each detection target. Furthermore, S13 may include the following sub-steps: By calling an encoder based on contrast loss optimization through the risk prediction layer, the temporal feature vector, the current temporal radiomics features, and the current temporal clinical data are encoded respectively to obtain initial features of various modalities; The risk prediction layer maps various initial features to a semantic space of a preset dimension to obtain multiple cross-modal features corresponding to each detection target. The risk prediction layer uses an attention mechanism to weight and fuse cross-modal features to generate fused features for each detection target.

[0055] Current time-series radiomics features refer to the radiomics features obtained in the current examination or the most recent detection cycle.

[0056] In this embodiment, the risk prediction layer invokes three pre-defined dedicated encoders: a temporal feature encoder (LSTM architecture) processes temporal feature vectors to capture the dynamic correlation of risk trends; an image feature encoder (residual MLP architecture) processes current temporal radiomics features to extract core features of current pathological changes (such as edema and necrosis); and a clinical data encoder (gated MLP architecture) processes current temporal clinical data to filter key indicators (such as current NT-proBNP and EF values). All three encoders are pre-trained and optimized using contrastive loss to ensure that the encoded initial features are comparable across modalities, ultimately outputting initial features for three different modalities: temporal, image, and clinical. After obtaining initial features for multiple modalities, a linear transformation maps each initial feature to a semantic space of a pre-defined dimension, converting each initial feature into a vector of a pre-defined dimension. During the mapping process, due to the constraints of contrastive loss, the three modal features of the same detection target are close to each other in the semantic space, while the three modal features of different detection targets are separated, thus obtaining the three cross-modal features (temporal, image, and clinical) corresponding to each detection target.

[0057] After obtaining the three cross-modal features, the three cross-modal features are concatenated to obtain a feature matrix. The importance score of each feature is calculated by a single-layer MLP. For example, the image feature score of acute detection targets is high, and the temporal feature score of chronic detection targets is high. Then, the attention weight is obtained by Softmax normalization, and the weights are summed to 1. Finally, the three cross-modal features are weighted and summed according to the weights to generate fusion features, ensuring that the fusion features preferentially reflect the modal information that contributes more to risk prediction.

[0058] S14. Risk is classified according to all fusion features through the risk prediction layer to obtain the major adverse cardiovascular event level corresponding to each detection target. Major adverse cardiovascular events (MACE) are risk levels classified based on fusion characteristics, typically categorized as high risk, medium risk, and low risk.

[0059] In this embodiment, the risk prediction layer inputs the fusion features of all detected targets into a classification model such as XGBoost or a differentiable ranking learning layer. The classification model performs risk classification on the fusion features based on the MACE follow-up labels during the training phase, such as "high risk = MACE occurred within 1 year" and "low risk = no MACE occurred within 3 years". Cross-validation is used to optimize the model parameters during the classification process, and finally the MACE level (high / medium / low risk) of each detected target is output.

[0060] S15. The attribution algorithm is invoked through the report generation layer to classify the contribution of each time-series radiomics feature sequence, and combined with time-series clinical data and the level of major adverse cardiovascular events, a structured risk report corresponding to each detection target is generated.

[0061] In this embodiment of the invention, the report generation layer calls the SHAP attribution algorithm, inputs the time-series radiomics feature sequence and MACE level, and the algorithm quantifies the contribution of each radiomics feature (such as edema area texture features, MSI) to the MACE level. The features are classified into high contribution (e.g., contribution ratio ≥20%), medium contribution (10%-20%), and low contribution (<10%) according to the contribution level. Then, the classification results, time-series clinical data (such as myocardial enzyme spectrum and EF value in each cycle) and MACE level are integrated to generate a structured risk report according to the structure of "basic information → summary of time-series clinical data → MACE level → contribution of key imaging features → clinical recommendations".

[0062] In one example of the present invention, the data processing terminal is communicatively connected to a cloud server, and the method further includes the following steps: The cloud server periodically sends the initial global model to multiple data processing terminals through an encrypted channel; After privacy-desensitizing newly added time-series clinical data stored locally, local fine-tuning data is constructed by combining newly added CCTA images; After training the initial global model using local fine-tuning data, the model update weights corresponding to all model parameters are extracted. Calculate the weight update amount between the updated weights of each model and the original weights of the initial global model, and encrypt and send it to the cloud server; Once the cloud server receives and decrypts all encrypted weight updates, it performs a weighted average of the weight updates to obtain the global update value. The cloud server updates the model parameters of the initial global model using a global update variable to obtain a new initial global model, and then jumps to execute the step of the cloud server periodically sending the initial global model to multiple data processing terminals through an encrypted channel.

[0063] The initial global model refers to the basic model (such as a multi-pathological change detection model or a two-layer cascaded risk prediction framework) that is preset on the cloud server and used for multi-center collaborative training. Its parameters are optimized through multi-center data pre-training.

[0064] Local fine-tuning data refers to training data formed by integrating newly added time-series clinical data (after privacy desensitization) with newly added CCTA images (single-energy / dual-energy CCTA) on the local data processing terminal, which is used to fine-tune the model to adapt to the local data distribution.

[0065] In this embodiment of the invention, the cloud server sends the initial global model to the data processing terminals deployed in each hospital through an encrypted channel according to a preset cycle, such as once a month. Each data processing terminal performs privacy desensitization on the newly added time-series clinical data stored locally, such as myocardial enzyme spectrum and EF value of myocardial ischemia patients in the past month, such as adding differential privacy noise to mask individual characteristics, and integrates it with newly added CCTA images collected at the same time in a unified DICOM format to construct local fine-tuning data.

[0066] The terminal's model training module is invoked to perform local training on the initial global model using locally tuned data. Specifically, the model backbone is frozen, and only the output layer parameters are fine-tuned. After training, the model update weights corresponding to all model parameters are extracted. The difference between the updated model weights and the original weights of the initial global model is calculated to obtain the weight update amount, which is then sent to the cloud server via an encrypted channel.

[0067] The cloud server receives and decrypts the encrypted weight updates from all data processing terminals. It then performs a weighted average of these updates based on the proportion of data volume from each terminal, with larger data volumes resulting in higher weights. This generates a global update, which is used to adjust the parameters of the initial global model, resulting in a new initial global model. This process repeats until the next periodic transmission of the initial global model. This achieves collaborative evolution of the multi-center model without transmitting raw data, effectively protecting data privacy.

[0068] In addition, before the weighted average of the weighted update amounts, the cloud will conduct a quality assessment of the model updates uploaded by each node, and demote or remove low-quality updates that may be caused by poor data quality or malicious attacks, so as to ensure the quality of global model evolution.

[0069] In one example of the present invention, the method further includes the following steps: If the target meets the preset physical condition, the plain CT image of the target will be obtained according to the detection cycle. The preset generative adversarial network is invoked to generate CCTA time-series images corresponding to each plain CT image; Jump to execute the steps of calling the preset multi-pathological change detection model to perform multi-pathological change detection on each CCTA time series image, and calculate the myocardial salvage index according to the detection results.

[0070] The pre-set physical conditions refer to the contraindications of the test target that prevent the use of iodine contrast agents, such as renal insufficiency (estimated glomerular filtration rate eGFR < 30 mL / min) or iodine contrast agent allergy.

[0071] Plain CT images refer to computed tomography images that do not require the injection of iodine contrast agents and can display basic structures such as the myocardium and lungs.

[0072] Generative Adversarial Networks (GANs) are deep learning-based generative models that include generators and discriminators. They can learn the mapping relationship between plain CT and CCTA images and generate CCTA time-series images with pathological details.

[0073] In this embodiment, the data processing terminal can check whether there is a history of iodine contrast agent allergy through electronic medical records and confirm whether the eGFR is below the threshold through laboratory tests to determine whether the target meets the preset physical condition. If it does, the terminal collects plain CT images of the target according to the detection cycle to ensure that the images cover the complete left ventricular myocardial region. Then, the terminal calls the preset generative adversarial network (GAN) and inputs the plain CT image into the generator. The generator generates CCTA time-series images corresponding to the time points of the plain CT scan based on the pre-trained plain CT-CCTA feature mapping relationship. The discriminator simultaneously optimizes the pathological details of the generated images, such as simulating the contrast-enhanced blood vessels and myocardial regions. After generating the CCTA time-series images, the terminal jumps to step 102 to avoid diagnostic omissions caused by the inability to collect real CCTA images due to physical condition.

[0074] For the training of the pre-defined generative adversarial network, the training data comes from paired data of multi-center labeled plain CT scans and real CCTA scans.

[0075] In another example of the invention, a pure CNN network (such as 3D ResNet) can be used instead of the Transformer+U-Net hybrid architecture; or a recurrent neural network (RNN) can be used to process time-series data; or different feature attribution methods (such as LIME) can be used.

[0076] In this embodiment of the invention, a locally deployed data processing terminal acquires CCTA time-series images and time-series clinical data of multiple targets after myocardial ischemia according to a detection cycle; a preset multi-pathological change detection model is invoked to detect multiple pathological changes in each CCTA time-series image, and the myocardial salvage index is calculated based on the detection results; multiple major adverse cardiovascular event factors are extracted according to the detection results, and combined with the myocardial salvage index to generate a time-series radiomics feature sequence; a preset cascaded risk prediction framework is invoked to perform risk stratification prediction based on time-series clinical data and time-series radiomics feature sequence, and a structured risk report is generated according to the prediction results. Thus, by employing a multi-pathological change detection model based on CCTA time-series images and time-series clinical data to detect multiple pathological changes, and simultaneously performing risk stratification prediction based on the cascaded risk prediction framework, more accurate dynamic pathological change detection and risk prediction of major adverse cardiovascular events are achieved.

[0077] Please see Figure 2 , Figure 2 This is a structural block diagram of a device for detecting and predicting the risk of multiple pathological changes caused by myocardial ischemia, provided in an embodiment of the present invention.

[0078] This invention provides a device for detecting and predicting the risk of multiple pathological changes caused by myocardial ischemia, applied to a locally deployed data processing terminal. The device includes: The data periodic acquisition module 201 is used to acquire CCTA time-series images and time-series clinical data of multiple detection targets after myocardial ischemia occurs according to the detection cycle; The multi-pathological change detection module 202 is used to call the preset multi-pathological change detection model to perform multi-pathological change detection on each CCTA time series image, and calculate the myocardial salvage index according to the detection results. The feature generation module 203 is used to extract multiple major adverse cardiovascular event radiomics feature factors according to the detection results, and combine them with the myocardial salvage index to generate a time-series radiomics feature sequence. The risk stratification prediction module 204 is used to call a preset cascaded risk prediction framework to perform risk stratification prediction based on time-series clinical data and time-series radiomics feature sequences, and generate a structured risk report according to the prediction results.

[0079] Optionally, the multi-pathological change detection model includes an encoder, a decoder, and a multi-task output head; the multi-pathological change detection module 202 is specifically used for: After standardizing each CCTA time series image, it is cropped into image blocks of the same size; The encoder is invoked to extract global pathological dependency features at different scales from each image patch using a multi-head attention mechanism; The decoder is invoked to perform multi-scale fusion of various global pathological dependent features and then the boundaries are sharpened to obtain the target pathological feature map; The multi-task output head is invoked to perform voxel binary classification on the target pathological feature map to determine the myocardial region and non-myocardial region. The multi-task output head is invoked to perform voxel six-class classification on the myocardial region to determine the probability of pathological changes corresponding to each type of pathological change in the myocardial region. The myocardial salvage index is calculated based on the probability of each pathological change and a preset calculation formula.

[0080] Optionally, the feature generation module 203 is specifically used for: Multiple radiomics features were extracted from the pathological regions corresponding to each type of pathological change in the detection results to construct a high-dimensional feature matrix. After normalizing the image omics features within the high-dimensional feature matrix, unsupervised dimensionality reduction is performed to obtain a low-dimensional feature matrix. The correlation between each low-dimensional feature in the low-dimensional feature matrix and major adverse cardiovascular events is calculated using a variational autoencoder. Based on relevance and feature interpretability, target low-dimensional features are selected from the low-dimensional feature matrix; Based on the detection cycle, target low-dimensional features and myocardial salvage index are spliced ​​together to generate a time-series radiomics feature sequence.

[0081] Optionally, the cascaded risk prediction framework includes a time-series processing layer, a risk prediction layer, and a report generation layer; the risk stratification prediction module 204 includes: The risk curve generation submodule is used to construct individualized time-risk curves for each detection target based on time-series clinical data and time-series radiomics feature sequences within a historical time period through the time-series processing layer. The curve mapping submodule is used to map individualized time-risk curves into time-series feature vectors of a preset dimension through the risk prediction layer; The semantic fusion submodule is used to map and fuse temporal feature vectors, current temporal radiomics features and current temporal clinical data into the semantic space through the risk prediction layer, and generate fusion features corresponding to each detection target. The risk classification submodule is used to classify risks according to all fused features through the risk prediction layer to obtain the major adverse cardiovascular event level corresponding to each detection target. The report generation submodule is used to call the attribution algorithm through the report generation layer to classify the contribution of each time series radiomics feature sequence, and combine time series clinical data and the level of major adverse cardiovascular events to generate a structured risk report corresponding to each detection target.

[0082] Optionally, the risk curve generation submodule is specifically used for: By aligning time-series clinical data and time-series radiomics feature sequences according to each time point within a historical time period through a time-series processing layer, a multi-dimensional time-series matrix corresponding to each time point is constructed. The temporal processing layer calls the temporal convolutional network to perform causal convolution and dilated convolution on the multidimensional time series matrix, and combines residual connections to determine the risk probability of adverse cardiovascular events at each time point; By using the time-series processing layer with time points as the horizontal axis and risk probability as the vertical axis, individualized time-risk curves are constructed for each detection target.

[0083] Optionally, the semantic fusion submodule is specifically used for: By calling an encoder based on contrast loss optimization through the risk prediction layer, the temporal feature vector, the current temporal radiomics features, and the current temporal clinical data are encoded respectively to obtain initial features of various modalities; The risk prediction layer maps various initial features to a semantic space of a preset dimension to obtain multiple cross-modal features corresponding to each detection target. The risk prediction layer uses an attention mechanism to weight and fuse cross-modal features to generate fused features for each detection target.

[0084] Optionally, the data processing terminal communicates with a cloud server, and the device also includes a federated training module, specifically used for: The cloud server periodically sends the initial global model to multiple data processing terminals through an encrypted channel; After privacy-desensitizing newly added time-series clinical data stored locally, local fine-tuning data is constructed by combining newly added CCTA images; After training the initial global model using local fine-tuning data, the model update weights corresponding to all model parameters are extracted. Calculate the weight update amount between the updated weights of each model and the original weights of the initial global model, and encrypt and send it to the cloud server; Once the cloud server receives and decrypts all encrypted weight updates, it performs a weighted average of the weight updates to obtain the global update value. The cloud server updates the model parameters of the initial global model using a global update variable to obtain a new initial global model, and then jumps to execute the step of the cloud server periodically sending the initial global model to multiple data processing terminals through an encrypted channel.

[0085] Optionally, the device also includes an image updating module, specifically used for: If the target meets the preset physical condition, the plain CT image of the target will be obtained according to the detection cycle. The preset generative adversarial network is invoked to generate CCTA time-series images corresponding to each plain CT image; Jump to execute the steps of calling the preset multi-pathological change detection model to perform multi-pathological change detection on each CCTA time series image, and calculate the myocardial salvage index according to the detection results.

[0086] This invention also provides an electronic device, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the method for detecting and predicting multiple pathological changes caused by myocardial ischemia as described in any embodiment of this invention.

[0087] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0088] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0089] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0090] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0091] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting and predicting the risk of multiple pathological changes caused by myocardial ischemia, characterized in that, The method, applied to a locally deployed data processing terminal, includes: According to the detection cycle, CCTA time-series images and time-series clinical data were acquired after myocardial ischemia occurred in multiple detection targets; The preset multi-pathological change detection model is invoked to detect multiple pathological changes in each of the CCTA time-series images, and the myocardial salvage index is calculated according to the detection results. Based on the detection results, multiple major adverse cardiovascular event factors were extracted and combined with the myocardial salvage index to generate a time-series radiomics feature sequence. The system invokes a pre-defined cascaded risk prediction framework to perform risk stratification prediction based on the time-series clinical data and the time-series radiomics feature sequences, and generates a structured risk report based on the prediction results.

2. The method according to claim 1, characterized in that, The multi-pathological change detection model includes an encoder, a decoder, and a multi-task output head; the step of calling the preset multi-pathological change detection model to perform multi-pathological change detection on each of the CCTA time-series images, and calculating the myocardial salvage index according to the detection results includes: After standardizing each CCTA time series image, it is cropped into image blocks of the same size; The encoder is invoked to extract global pathological dependency features of different scales from each of the image blocks according to a multi-head attention mechanism; The decoder is invoked to perform multi-scale fusion of the global pathological dependent features and then the boundaries are sharpened to obtain the target pathological feature map. The multi-task output head is invoked to perform voxel binary classification on the target pathological feature map to determine the myocardial region and non-myocardial region. The multi-task output head is invoked to perform voxel six-class classification on the myocardial region to determine the probability of pathological changes corresponding to each type of pathological change in the myocardial region. The myocardial salvage index is calculated based on the probability of each pathological change and a preset calculation formula.

3. The method according to claim 2, characterized in that, The step of extracting multiple major adverse cardiovascular event factors according to the detection results and combining them with the myocardial salvage index to generate a time-series radiomics feature sequence includes: Multiple radiomics features are extracted from the pathological regions corresponding to each type of pathological change in the detection results to construct a high-dimensional feature matrix; After normalizing each of the radiomics features in the high-dimensional feature matrix, unsupervised dimensionality reduction is performed to obtain a low-dimensional feature matrix. The correlation between each low-dimensional feature in the low-dimensional feature matrix and major adverse cardiovascular events is calculated using a variational autoencoder. Based on the aforementioned relevance and feature interpretability, target low-dimensional features are selected from the low-dimensional feature matrix; The target low-dimensional features and the myocardial salvage index are spliced ​​together according to the detection cycle to generate a time-series radiomics feature sequence.

4. The method according to claim 1, characterized in that, The cascaded risk prediction framework includes a time-series processing layer, a risk prediction layer, and a report generation layer; the step of invoking the preset cascaded risk prediction framework to perform risk stratification prediction based on the time-series clinical data and the time-series radiomics feature sequences, and generating a structured risk report according to the prediction results, includes: The time-series processing layer constructs individualized time-risk curves for each detection target based on time-series clinical data and time-series radiomics feature sequences within historical time periods. The risk prediction layer maps the individualized time-risk curve into a time-series feature vector of a preset dimension. The risk prediction layer uses the time-series feature vector, current time-series radiomics features, and current time-series clinical data to map to the semantic space and fuse them to generate a fusion feature corresponding to each detection target. The risk prediction layer performs risk classification according to all the fusion features to obtain the major adverse cardiovascular event level corresponding to each detection target; The report generation layer calls an attribution algorithm to classify the contribution of each time-series radiomics feature sequence, and combines the time-series clinical data and the major adverse cardiovascular event levels to generate a structured risk report corresponding to each detection target.

5. The method according to claim 4, characterized in that, The step of constructing individualized time-risk curves for each detection target based on time-series clinical data and time-series radiomics feature sequences within a historical time period through the time-series processing layer includes: The time-series processing layer aligns time-series clinical data and time-series radiomics feature sequences according to each time point within a historical time period, and constructs a multi-dimensional time-series matrix corresponding to each time point. The temporal processing layer invokes a temporal convolutional network to perform causal convolution and dilated convolution on the multidimensional time series matrix, and combines residual connections to determine the risk probability of adverse cardiovascular events at each time point; The time-series processing layer constructs individualized time-risk curves for each detection target by using the time point as the horizontal axis and the risk probability as the vertical axis.

6. The method according to claim 4, characterized in that, The step of mapping and fusing the temporal feature vector, current temporal radiomics features, and current temporal clinical data to the semantic space through the risk prediction layer to generate fused features corresponding to each detection target includes: The risk prediction layer calls an encoder based on contrast loss optimization to encode the temporal feature vector, the current temporal radiomics features, and the current temporal clinical data respectively, to obtain initial features of various modalities; The risk prediction layer maps various initial features to a semantic space of a preset dimension to obtain multiple cross-modal features corresponding to each detection target. The risk prediction layer performs weighted fusion of the cross-modal features according to the attention mechanism to generate a fused feature corresponding to each detection target.

7. The method according to claim 1, characterized in that, The data processing terminal is communicatively connected to a cloud server, and the method further includes: The cloud server periodically sends the initial global model to multiple data processing terminals through an encrypted channel; After privacy-desensitizing newly added time-series clinical data stored locally, local fine-tuning data is constructed by combining newly added CCTA images; After training the initial global model using the local fine-tuning data, the model update weights corresponding to all model parameters are extracted. Calculate the weight update amount between each of the model update weights and the original model weights of the initial global model, and encrypt and send it to the cloud server; Once the cloud server receives and decrypts all encrypted weight updates, it performs a weighted average of the weight updates to obtain the global update value. The cloud server uses the global update amount to update the model parameters of the initial global model to obtain a new initial global model, and then jumps to execute the step of the cloud server periodically sending the initial global model to multiple data processing terminals through an encrypted channel.

8. The method according to claim 1, characterized in that, The method further includes: If the target being detected meets the preset physical condition, then the plain CT image of the target being detected is obtained according to the detection cycle; A preset generative adversarial network is invoked to generate CCTA time-series images corresponding to each of the aforementioned plain CT images; Jump to execute the step of calling the preset multi-pathological change detection model to perform multi-pathological change detection on each of the CCTA time series images, and calculate the myocardial salvage index according to the detection results.

9. A device for detecting and predicting the risk of multiple pathological changes caused by myocardial ischemia, characterized in that, The device is used in a locally deployed data processing terminal and includes: The data periodic acquisition module is used to acquire CCTA time-series images and time-series clinical data of multiple detection targets after myocardial ischemia occurs, according to the detection cycle; The multi-pathological change detection module is used to call a preset multi-pathological change detection model to perform multi-pathological change detection on each of the CCTA time-series images, and calculate the myocardial salvage index according to the detection results. The feature generation module is used to extract multiple major adverse cardiovascular event factors according to the detection results, and combine them with the myocardial salvage index to generate a time-series radiomics feature sequence. The risk stratification prediction module is used to call a preset cascaded risk prediction framework to perform risk stratification prediction based on the time-series clinical data and the time-series radiomics feature sequence, and generate a structured risk report according to the prediction results.

10. An electronic device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the method for detecting and predicting the risk of multiple pathological changes caused by myocardial ischemia as described in any one of claims 1-9.

Images