A coronary heart disease risk prediction method and system fusing multi-image omics features

Abstract

A coronary heart disease risk prediction method and system fusing multi-imaging features belong to the field of medical image analysis and disease risk prediction. The original chest plain CT data of a patient is obtained for preprocessing to obtain a pericardium, pericardial fat and coronary calcification structure mask, a direction field is constructed respectively, and the pericardium, pericardial fat and coronary calcification imaging features are extracted; the pericardium, pericardial fat and coronary calcification imaging features are weighted to obtain a structure level feature, and a coronary heart disease risk prediction result is output. The present application realizes collaborative modeling of multi-structure and multi-source imaging features, reflects the spatial heterogeneity characteristics of coronary heart disease related pathologies around the potential running direction of the coronary artery and the anatomical constraint relationship, and the directional imaging features realize end-to-end joint learning of the features through the structure level multi-layer attention mechanism, while carrying out lung disease screening, the coronary heart disease risk is evaluated synchronously, and technical support is provided for the comprehensive management of cardiopulmonary comorbidity.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis and disease risk prediction technology, and more specifically, to a method and system for predicting coronary heart disease risk by combining direction-aware radiomics with multi-level attention fusion. Background Technology

[0002] Coronary artery disease, also known as coronary atherosclerotic heart disease, is a heart condition caused by atherosclerosis of the coronary arteries, leading to narrowing or blockage of the blood vessels and subsequently causing myocardial ischemia, hypoxia, and even necrosis. Coronary angiography, the gold standard for diagnosing coronary artery disease, is an invasive procedure with risks such as contrast agent allergy, puncture site complications, and even death. Furthermore, it requires a high level of physician experience and advanced equipment, limiting its widespread adoption in primary care hospitals.

[0003] Existing methods for predicting coronary heart disease risk mainly fall into three categories. The first is based on clinical data, relying on indicators such as age, gender, and blood pressure. This provides limited information, makes it difficult to detect hidden pathological changes, and results in limited predictive accuracy, especially in the early stages of the disease. The second is based on a combination of clinical data and biomedical biomarkers. While this provides more comprehensive information, biomarkers require blood tests, increasing costs and invasiveness. Furthermore, some biomarkers lack large-scale clinical evidence, making them unsuitable for primary care settings. The third method integrates multimodal image and text features. While this method offers rich information, its complex examination process and high time and cost limit its widespread application.

[0004] Existing predictive models, such as the Diamond-Forrester model, rely solely on clinical data and do not incorporate radiomics features; some technologies rely on biomarkers or multimodal carotid artery imaging, which present problems such as invasiveness, high cost, or risk of inaccurate predictions.

[0005] Therefore, the existing technology has problems and needs further improvement and development. Summary of the Invention

[0006] (I) Purpose of the invention: In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a method and system for predicting the risk of coronary heart disease by integrating multiple radiomic features, so as to solve the problems of limited accuracy, strong invasiveness, high cost and difficulty in promotion of existing coronary heart disease risk prediction methods, and to achieve non-invasive, accurate and efficient prediction of coronary heart disease risk.

[0007] (II) Technical Solution: To address the aforementioned technical problems, this technical solution provides a method for predicting coronary heart disease risk by integrating multiple radiomics features, including:

[0008] Step 1: Obtain the patient's original chest CT scan data and perform preprocessing to obtain a structural mask of the pericardium, pericardial fat, and coronary artery calcification;

[0009] Step 2: Based on the structural masks of the pericardium, pericardial fat, and coronary artery calcification, construct orientation fields respectively, and extract radiomics features of the pericardium, pericardial fat, and coronary artery calcification.

[0010] Step 3: Weight the radiomic features of the pericardium, intrapericardial fat, and coronary artery calcification to obtain structural features;

[0011] Step 4: Output the coronary heart disease risk prediction results based on structural features.

[0012] Preferably, step 1 includes:

[0013] Step 101: Obtain the raw medical image data of the chest plain CT scan and the corresponding coronary heart disease label;

[0014] Step 102: Anonymize and standardize the raw medical image data of chest plain CT scan to obtain normalized raw medical image data of chest plain CT scan.

[0015] Step 103: Extract the three-dimensional segmentation mask of the pericardium, pericardial fat and coronary artery calcification using automatic or manual segmentation methods;

[0016] Step 104: Preprocess the three-dimensional segmentation mask of the pericardium, intrapericardial fat and coronary artery calcification to obtain the corresponding structural ROIs, which include the pericardium ROI, intrapericardial fat ROI and coronary artery calcification ROI.

[0017] Preferably, the raw medical image data of the chest plain CT scan is obtained when conducting lung disease screening without adding additional imaging examinations or radiation doses.

[0018] Preferably, the structural ROIs are resampled to a uniform resolution using trilinear interpolation. In the HU window, the pericardial ROI is truncated to the range of [-200, 300], the intracardiac fat ROI is truncated to the range of [-190, -30], and the coronary calcification ROI is truncated to the range of [130, 1000]. Finally, the three ROIs are normalized to the range of [0, 1] using Min-Max or Z-score.

[0019] Preferably, step 2 includes:

[0020] Step 201: Construct prior orientation fields for the pericardial ROI, intracardiac fat ROI, and coronary calcification ROI respectively;

[0021] Step 202: Under the constraint of the orientation field, extract first-order radiomics features and directional texture features, and perform parameter adaptive adjustment for different structures to obtain directional radiomics features.

[0022] Preferably, for the pericardial ROI, all voxel points Ω within the pericardial ROI are extracted: Calculate the centroid c of the pericardial ROI. For each voxel point x within the ROI i Calculate the direction vector in the x-direction. And calculate the direction vectors in the y and z directions to obtain the three-dimensional direction vector. ;x i The x-coordinate of voxel point i represents the x-axis coordinate; y i The z-axis coordinate represents the y-coordinate of voxel point i; i The z-coordinate of voxel point i is represented; i = 1, 2, 3, ..., N, where N is the total number of voxels within the pericardial ROI; M represents the set of all voxel points within the pericardial ROI.

[0023] For the pericardial fat region of interest (ROI), let S be... p It is the set of voxel points of the pericardial inner wall, and the set of all points in the pericardial region of interest (ROI) that have at least one neighbor that does not belong to the ROI. The goal is to extract the set of points on the surface of the pericardial inner wall from the pericardial ROI. s represents the spatial location (x, y, z) of a voxel; M p (s) represents the binary mask of the pericardial ROI; Let s' represent the 13-neighborhood of point s; for each pericardial fat somatic cell... M f As a binary mask for pericardial fat, the normal direction vector of the pericardial fat voxel x pointing towards the inner pericardial wall is calculated as follows: , For a pericardial fat body; Pericardial fat body The Euclidean shortest distance to the inner wall of the pericardium is determined by combining the first principal component direction of the local cubic neighborhood and introducing a direction consistency constraint. Obtain the set of directions of the pericardial fat ROI. ;

[0024] For coronary calcification ROIs, a 26-neighbor connectivity method is used to mask the coronary calcification ROI M. c Perform three-dimensional connected component labeling, aggregating spatially continuous calcified voxels into several independent calcified connected components, filtering out small-volume connected components, and representing the i-th connected component as a three-dimensional point set. Calculate the geometric center of a 3D point set. and the three-dimensional covariance matrix , ;x j N represents the coronary artery calcification voxel point in the i-th connected region; iThis represents the number of voxels in the i-th connected component; eigenvalue decomposition is performed on the 3D covariance matrix, and the direction of the first principal component is taken as the principal axis direction of the connected component. Forming a set of ROI directions for coronary artery calcification , N=1,2,...i.

[0025] Preferably, under the constraint of the orientation field, first-order imagemic features are extracted, including:

[0026] When extracting first-order imageomics features, modeling the directional constraint grayscale variation is performed in each direction. Upper definition of grayscale difference Where I represents the grayscale value function on the CT image; x represents the current voxel; δ represents the step size; d k Denotes the spatial vector originating from the current voxel; x+δdk represents the vector along the direction d. k The adjacent voxel positions, for all satisfying Mean value of voxel statistical grayscale change under conditions Standard deviation of grayscale variation Gray-scale variation skewness Gray-scale variation kurtosis and grayscale change energy .

[0027] Preferably, the directional texture features include directional gray-level co-occurrence matrix features and directional run-length matrix features;

[0028] The directional gray-level co-occurrence matrix features s represents the voxel coordinates within the ROI; g(s) represents the gray level of voxel s; (i,j) represents the gray level index; d k This is the eigenvector of the k-th direction;

[0029] Statistics along direction Gray combination of adjacent voxels Calculate contrast Correlation Homogeneity ,energy ,entropy , , Calculate j; for each direction d k get For k directions, ;

[0030] The directional gray-level run-length matrix statistically distributes the lengths of consecutive identical gray-level voxels along a given direction dk, extracting short run-length emphasis. Long travel itinerary Percentage of trip , Total number of tours The total number of voxels within the ROI; for each direction ,get For K directions, .

[0031] Preferably, after extracting first-order radiomics features and directional texture features, the grayscale discretization binwidth is set to a fixed HU range, specifically 20-30 HU for the pericardium, 10-20 HU for intrapericardial fat, and 40-50 HU for coronary artery calcification; the distance parameter for the ROI of coronary artery calcification is 1, the distance parameter for the ROI of intrapericardial fat is 2, and the distance parameter for the ROI of the pericardium is 5; 8-13 structurally relevant directions are extracted for each ROI to obtain the directional radiomics features of each ROI. r represents the pericardial ROI, intracardiac fat ROI, or coronary calcification ROI, K represents the number of directions, and F represents the number of radiomics features in each direction.

[0032] Preferably, the radiomics features of the pericardium, intrapericardial fat, and coronary artery calcification are weighted to obtain structural features, including: performing directional attention modeling on the directional radiomics features of the pericardium, intrapericardial fat, and coronary artery calcification structures through a structural fusion network, wherein the structural fusion network includes a directional feature embedding layer module, a directional attention aggregation module, a structure-specific mapping module, and a regularization and stabilization module;

[0033] The directional feature embedding layer module will embed the feature vector of each direction. , r belongs to the pericardial ROI, intracardiac fat ROI, and coronary calcification ROI; k represents the k-th direction vector; through linear transformation and activation function mapping to a unified latent space, execution is performed. ;

[0034] The directional attention aggregation module automatically identifies high-risk directions consistent with the potential coronary artery direction, and uses a preset attention weight formula. Calculate the importance weight for each direction, perform weighted aggregation on the embedded directional features, and output the result. ;

[0035] The structure-specific mapping module maps intermediate features of different structures to the same structural-level semantic space. , The weights of the three structural ROIs;

[0036] The regularization and stabilization module uses directional sparse regularization to encourage SFN to focus on key directions and output structural feature representations for each structure.

[0037] Preferably, step 4 outputs the coronary heart disease risk prediction result based on structural features, including: receiving the structural feature representation, and outputting the coronary heart disease risk prediction result through a classification head model, including CAD-RADS classification, whether it is coronary heart disease, and the probability of coronary heart disease risk;

[0038] The classification head uses a two-layer multilayer perceptron (MLP) combined with a softmax function. An end-to-end training approach was adopted, freezing the relevant parameters for target mask extraction in the chest CT data preprocessing unit, and using the cross-entropy loss function as the optimization objective. , where c is the category label and y is the actual label; To predict probabilities.

[0039] A coronary artery disease risk prediction system integrating multi-radiomics features includes a chest CT data preprocessing unit, which is used to acquire the patient's raw chest plain CT data for preprocessing to obtain structural masks of the pericardium, intrapericardial fat, and coronary artery calcification.

[0040] Based on the structural masks of the pericardium, intrapericardial fat, and coronary artery calcification, radiomic features were constructed to extract radiomic features of the pericardium, intrapericardial fat, and coronary artery calcification.

[0041] The radiomics feature fusion unit is used to weight the radiomics features of the pericardium, intrapericardial fat, and coronary artery calcification to obtain structural-level features;

[0042] The result prediction output unit is used to output the coronary heart disease risk prediction result based on structural features.

[0043] (III) Beneficial Effects: This invention provides a method and system for predicting coronary artery disease (CAD) risk by integrating multiple radiomics features. First, when coronary arteries cannot be directly visualized on plain CT scans, coronary artery-related anatomical structures are used to replace the invisible vessel body, achieving collaborative modeling of multi-structure and multi-source radiomics information, thus improving the robustness and accuracy of the model. Second, directional radiomics modeling is adopted, breaking through the limitations of traditional isotropic feature statistics. Multi-directional feature representations are constructed around the potential course of the coronary arteries and anatomical constraints, more fully reflecting the spatial heterogeneity of CAD-related pathological processes. Third, directional radiomics is upgraded from static feature engineering to a learnable networked modeling framework. End-to-end joint learning of features is achieved through a structural-level multi-layer attention mechanism, improving the model's discriminative ability and interpretability. Finally, relying on conventional chest CT data, CAD risk is simultaneously assessed while screening for lung diseases, effectively reducing medical costs and patient burden, increasing the early detection rate of CAD, and providing technical support for the comprehensive management of cardiopulmonary comorbidities. Attached Figure Description

[0044] Figure 1This is a flowchart illustrating the steps of a method for predicting the risk of coronary heart disease that integrates multiple radiomics features according to the present invention.

[0045] Figure 2 This is a flowchart of a method for predicting the risk of coronary heart disease that integrates multiple radiomics features according to the present invention;

[0046] Figure 3 This is a schematic diagram of the structure of a coronary heart disease risk prediction system that integrates multiple radiomics features according to the present invention. Detailed Implementation

[0047] The present invention will be further described in detail below with reference to preferred embodiments. More details are set forth in the following description in order to provide a full understanding of the present invention. However, the present invention can obviously be implemented in many other ways different from those described herein. Those skilled in the art can make similar extensions and derivations based on actual application situations without departing from the spirit of the present invention. Therefore, the scope of protection of the present invention should not be limited by the content of this specific embodiment.

[0048] The accompanying drawings are schematic diagrams of embodiments of the present invention. It should be noted that these drawings are for illustrative purposes only and are not drawn to scale, and should not be construed as limiting the actual scope of protection of the present invention.

[0049] A method for predicting coronary artery disease risk by integrating multiple radiomics features, such as Figure 1 , Figure 2 As shown, the specific steps include:

[0050] Step 1: Obtain the patient's original chest CT scan data and preprocess it to obtain a structural mask of the pericardium, pericardial fat, and coronary artery calcification.

[0051] Step 2: Based on the structural masks of the pericardium, pericardial fat, and coronary artery calcification, construct directional fields respectively, and extract radiomics features of the pericardium, pericardial fat, and coronary artery calcification.

[0052] Step 3: Weight the radiomic features of the pericardium, pericardial fat, and coronary artery calcification to obtain structural features.

[0053] Step 4: Output the coronary heart disease risk prediction results based on structural features.

[0054] A coronary artery disease risk prediction system integrating multiple radiomics features can be used to execute the aforementioned coronary artery disease risk prediction method integrating multiple radiomics features, such as... Figure 3As shown, the system includes a chest CT data preprocessing unit, a radiomics feature extraction unit, a radiomics feature fusion unit, and a result prediction output unit. The chest CT data preprocessing unit acquires the patient's raw chest plain CT data and preprocesses it to obtain structural masks of the pericardium, pericardial fat, and coronary artery calcifications. The radiomics feature extraction unit constructs orientation fields based on the structural masks of the pericardium, pericardial fat, and coronary artery calcifications, and extracts radiomics features for each of these components. The radiomics feature fusion unit weights the radiomics features of the pericardium, pericardial fat, and coronary artery calcifications to obtain structural-level features. The result prediction output unit outputs the coronary artery disease risk prediction result based on the structural-level features.

[0055] The chest CT data preprocessing unit in step 1 is specifically used for:

[0056] Step 101: Obtain the original medical image data of chest plain CT scan and the corresponding coronary heart disease label.

[0057] A chest CT scan is a non-contrast CT scan with a slice thickness typically between 1 and 5 millimeters, commonly used for screening and routine observation of lung lesions. While screening for lung diseases, it allows for the simultaneous assessment of a patient's coronary heart disease risk without additional imaging examinations or radiation doses. This invention utilizes raw medical imaging data from routinely acquired chest CT scans to automate the prediction of coronary heart disease risk without altering existing examination procedures.

[0058] The inclusion criteria for raw medical imaging data of chest plain CT scans described in this invention include a scan slice thickness ≤3 mm, a scan range covering the entire heart, and an interval of no more than 30 days between the chest plain CT scan and coronary angiography. Patients with a history of PCI (percutaneous coronary intervention) or CABG (coronary artery bypass grafting) prior to the chest plain CT scan are excluded.

[0059] The original medical image data of the chest plain CT scan was in DICOM format.

[0060] Based on the original medical imaging data of chest plain CT, and with the results of coronary angiography as the gold standard, three or more interventional cardiologists with the title of associate chief physician or above jointly determine whether it is coronary heart disease, the CAD-RADS (Coronary Artery Disease Reporting and Data System) classification, and the corresponding coronary heart disease label.

[0061] Coronary artery disease (CAD) is defined as at least one epicardial coronary artery with a diameter ≥2 mm and a stenosis of ≥50%. When there is no definite stenosis or plaque in an epicardial coronary artery with a diameter ≥2 mm, the CAD-RADS classification is 0, and the CAD label is 0. When there is mild stenosis in an epicardial coronary artery with a diameter ≥2 mm, and the stenosis is 1%–24%, with atherosclerotic plaques visible, the CAD-RADS classification is 1, and the CAD label is 1. When there is mild stenosis in an epicardial coronary artery with a diameter ≥2 mm, and the stenosis is 25%–49%, the CAD-RADS classification is 2, and the CAD label is 2. When there is moderate stenosis in an epicardial coronary artery with a diameter ≥2 mm, and the stenosis is 50%–69%, it is considered a borderline lesion, the CAD-RADS classification is 3, and the CAD label is 3. When there is severe stenosis in an epicardial coronary artery with a diameter ≥2 mm, and the stenosis is 70%–99%, the CAD-RADS classification is 4, and the CAD label is 4. When an epicardial coronary artery with a diameter ≥2 mm is 100% completely occluded, the CAD-RADS classification is 5, and the coronary artery disease label is 5.

[0062] Therefore, CAD-RADS classifications of 0, 1, and 2 are considered non-coronary artery disease; CAD-RADS classifications of 3, 4, and 5 are considered coronary artery disease.

[0063] Step 102: Anonymize and standardize the raw chest CT images to obtain normalized raw chest CT images.

[0064] The raw chest CT scan images were anonymized, removing patient-identification information from the DICOM header while preserving pixel spacing, slice thickness, and spatial location information to ensure the integrity of the image geometry. It should be noted that the acquisition and processing of the patient's raw medical image data were conducted with the patient's valid consent.

[0065] Due to differences in scanning parameters among different CT devices, the Hounsfield Unit (HU) values ​​of raw chest CT images may vary. To eliminate the influence of device differences, the HU values ​​of the raw chest CT images are truncated to [-1000, 1000], which can fully cover target structures such as fat, soft tissue, and calcification. Furthermore, a linear normalization method is used to map the HU values ​​to the [0, 1] interval.

[0066] The normalization methods include, but are not limited to, Min-Max or Z-score.

[0067] Step 103: Extract the three-dimensional segmentation mask of the pericardium, pericardial fat and coronary artery calcification using automatic or manual segmentation.

[0068] In cases where coronary arteries cannot be directly observed on non-contrast plain CT scans, their anatomical environment and pathological carrier structures are used as alternative representations. The pericardium, pericardial fat, and coronary artery calcification are listed as possible indicators for screening the risk of coronary heart disease and used for subsequent analysis.

[0069] Automatic segmentation refers to obtaining three-dimensional segmentation masks for the pericardium, intrapericardial fat, and coronary artery calcifications separately using a segmentation model. Manual segmentation refers to the process where an experienced senior thoracic radiologist adjusts the window width and level to independently depict the target contour layer by layer, forming a three-dimensional segmentation mask for the pericardium, intrapericardial fat, and coronary artery calcifications.

[0070] The automatic segmentation specifically includes:

[0071] The segmentation model refers to a 3D convolutional neural network based on an Encoder–Decoder architecture, comprising an Encoder part, a Decoder part, and an output layer. The Encoder part consists of 4-5 3D convolutional layers, each with a 3×3×3 kernel size and a stride of 2. Batch normalization is performed using BatchNorm layers, and ReLU is selected as the activation function. High-level semantic features of the image are extracted through layer-by-layer downsampling to capture the overall shape of the target structure. The Decoder part is symmetrical to the Encoder, achieving upsampling through transposed convolutional layers. Skip connections are introduced to fuse low-dimensional detail features from the corresponding Encoder layer with high-dimensional semantic features from the Decoder, ensuring the accuracy of the segmentation results. The output layer uses a 1×1×1 convolutional kernel to map features to two channels, outputting a voxel-level classification probability map.

[0072] The described 3D convolutional neural network extracts high-level semantic features through layer-by-layer downsampling and combines skip connections to recover spatial detail information during the decoding stage. The preferred 3D convolutional neural network is UNet or VNet, as these two models exhibit excellent spatial detail recovery capabilities in medical image segmentation tasks.

[0073] The three-dimensional convolutional neural network model was trained using a weighted combination of Dice loss and cross-entropy loss as the optimization objective, combined with the Adam optimizer, learning rate scheduling strategy, and data augmentation methods. Pericardial and coronary artery calcification masks were obtained by thresholding (0.5), and tissue with a HU value [0, -200] within the pericardial region was extracted as the intracardiac fat mask.

[0074] The weight ratio of the Dice loss to the cross-entropy loss is set to 0.7:0.3. The Dice loss is used to alleviate the class imbalance problem: Dice = 2 × |Y∩ | / (|Y|+| |), where Y is the real mask, A prediction mask is used. The cross-entropy loss is used to optimize the gradient update of the classification probability, improving the model's convergence speed. The Adam optimizer has an initial learning rate of 1e-4 and a weight decay coefficient of 1e-5. The learning rate scheduling adopts a cosine annealing scheduling strategy, decaying the learning rate to 0.9 times the current value every 10 epochs. When the validation set loss does not decrease for 5 consecutive epochs, the learning rate is reduced to 1 / 10 of its original value to avoid the model getting trapped in local optima. To improve the model's generalization ability, random data augmentation, including rotation, flipping, and intensity perturbation, is also performed on the input chest CT plain scan medical image data.

[0075] The complete raw medical image data from a plain chest CT scan is input into the three-dimensional convolutional neural network, which outputs voxel-level probability maps of pericardial and coronary artery calcification. Thresholding is applied; in this invention, the threshold is set to 0.5. Voxels with a probability ≥ 0.5 are identified as target structures, while voxels with a probability < 0.5 are identified as background, resulting in a binary three-dimensional segmentation mask for pericardial and coronary artery calcification.

[0076] The spatial range of the pericardium is determined based on the pericardial segmentation mask, and then voxels with HU values ​​between [0, -200] are extracted within this range to form a three-dimensional segmentation mask for pericardial fat.

[0077] The manual segmentation specifically includes: importing the raw medical image data from a chest CT scan into professional medical image annotation software, uniformly adjusting the window width and window level (in this invention, the window width is set to 350 HU and the window level to 50 HU); and using a layer-by-layer manual drawing method to complete the contour drawing of all layers, stacking the two-dimensional contours of all layers to generate a three-dimensional segmentation mask.

[0078] Simultaneous observation of the coronal, sagittal, and axial views of the original chest CT scan data ensures complete capture of the target structures. For the pericardium, the outline is drawn along the inner and outer walls, including the pericardial wall and surrounding connective tissue. For intrapericardial fat, all low-density areas within the pericardium are delineated, ensuring coverage of adipose tissue anterior, posterior, and lateral to the pericardium. For coronary artery calcification, all high-density calcifications are precisely delineated to avoid misinterpreting bone, metal artifacts, etc., as calcifications.

[0079] Step 104: Preprocess the three-dimensional segmentation mask of the pericardium, pericardial fat and coronary artery calcification to obtain a structured ROI.

[0080] For a three-dimensional segmentation mask of the pericardium, intrapericardial fat, and coronary artery calcification, traverse all voxels and record the minimum and maximum coordinates along the x, y, and z axes to obtain the coordinate point A(x). min ,y min ,z min ) and B(x max ,ymax ,z max The cuboid region with A and B as diagonal points is the initial ROI (Region of Interest). To avoid the ROI edge from disrupting texture continuity, the boundary of the initial ROI is extended by 5-10 mm in each direction, resulting in three target ROI regions for subsequent analysis.

[0081] The three structural ROIs were resampled to a uniform resolution using trilinear interpolation. To avoid global normalization diluting pathological comparisons, different ROIs were subjected to a differential HU window. Within the HU window, the pericardial ROI was truncated to the range of [-200, 300], the intracardiac fat ROI to the range of [-190, -30], and the coronary calcification ROI to the range of [130, 1000]. Finally, the three ROIs were uniformly normalized to the range of [0, 1] using Min-Max or Z-score.

[0082] The three ROI normalization methods are consistent: the pericardium, intrapericardial fat, and coronary artery calcification all use the Min-Max normalization method or the Z-score normalization method.

[0083] The radiomics feature extraction unit in step 2 specifically includes:

[0084] Based on the structured ROI output by the chest CT data preprocessing unit, a coronary artery prior orientation field is constructed, breaking the isotropic assumption of traditional radiomics, extracting multi-dimensional radiomics features with orientation specificity, and fully characterizing the spatial heterogeneity of coronary artery disease-related structures.

[0085] Step 201: Construct a prior orientation field for each structural ROI.

[0086] We explored the spatial correlation between three types of ROI structures and the potential course of coronary arteries, and constructed directional fields for pericardial ROI, intrapericardial fat ROI, and coronary calcification ROI, respectively, to provide directional constraints for feature extraction, enabling features to accurately reflect pathological changes along the coronary artery path.

[0087] For the pericardial region of interest (ROI), the centroid c of the pericardial ROI is first calculated. The centroid refers to the geometric center of the three-dimensional spatial region of the pericardium, which is obtained by averaging the physical spatial coordinates of all voxels in the three-dimensional segmentation mask of the pericardium. The centroid serves as a global reference point for the construction of the orientation field, providing a stable spatial benchmark for subsequent fat orientation modeling and coronary artery-related spatial relationship analysis.

[0088] Specifically, extract all voxel points Ω within the pericardial ROI: , where x i The x-coordinate of voxel point i represents the x-axis coordinate; y iThe z-axis coordinate represents the y-coordinate of voxel point i; i The z-coordinate of voxel point i is represented; i = 1, 2, 3, ..., N, where N is the total number of voxels within the pericardial ROI; M represents the set of all voxel points within the pericardial ROI.

[0089] Calculate the arithmetic mean of the physical coordinates of all voxel points within the pericardial ROI: c = .

[0090] For each voxel point x within the ROI i Calculate the direction vector in the x-direction. Similarly, the direction vectors in the y and z directions are calculated, ultimately yielding the three-dimensional direction vectors. Describing the radial tissue changes of the pericardial structure relative to the center of the heart indirectly reflects the anatomical and mechanical state of the heart.

[0091] Intracardiac fat plays a crucial role in the development and progression of coronary heart disease, regulating inflammation and metabolism. Its spatial distribution and extension often follow the inner pericardial wall, exhibiting a distinct directional structural characteristic. To characterize the anatomical constraints and local extension trends of intracardiac adipose tissue in space, this invention constructs a set of combined directions for the pericardial fat region of interest (ROI). This set is composed of the shortest distance direction from the fat to the inner pericardial wall (normal) and the first principal component direction of the PCA (polyacrylamide ablation) of the local fat morphology, thus forming a directional description with clear anatomical and morphological significance.

[0092] For the pericardial fat region of interest (ROI), let S... p It is the set of all points in the pericardial region of interest (ROI) that have at least one neighbor that does not belong to the ROI. The set of points on the inner wall surface of the pericardium is first extracted from the pericardial ROI according to the following formula. S p The set of voxel points representing the inner wall of the pericardium is the set of all points in the pericardial region of interest (ROI) that have at least one neighbor that is not in the ROI; s represents the spatial location (x, y, z) of a voxel; M p (s) represents the binary mask of the pericardial ROI; s represents the 13-neighborhood of point s; s' represents a neighboring point of s that does not belong to the pericardial ROI.

[0093] Secondly, for each pericardial fat body... M f A binary mask is used to represent the pericardial fat; calculate its Euclidean shortest distance to the inner pericardial wall. The normal direction vector of the pericardial fat saturates x pointing towards the inner wall of the pericardium is... This indicates that the pericardial fat at this location is oriented towards the main anatomical direction of the pericardium. Among these, For a pericardial fat body; Pericardial fat body The Euclidean shortest distance to the inner wall of the pericardium. Each fat cell within the pericardium has a unique shortest distance direction vector, with the direction form being... , represents the direction vector along the x-axis, y-axis, and z-axis.

[0094] Define a local cubic neighborhood centered on pericardial fat spore x. , x represents the current intracardiac fat cell (x0, y0, z0); y represents any intracardiac fat cell (x, y, z) in the local cubic neighborhood; This indicates that the local cubic neighborhood is restricted to a cube window, where r represents the radius of the local cubic neighborhood. This indicates that voxel y lies within the neighborhood of a cube centered at x with a side length of 2r+1; This represents a pericardial fat ROI mask, meaning it only retains pericardial fat voxels belonging to this structure. In this invention, r = 3~5 voxels, representing the fat mass in the local cubic neighborhood as a three-dimensional point set. , representing the local cubic neighborhood N r The i-th intracardiac fat body y in (x), x i The x-axis coordinate of the pericardial fat somatic pigment y represents the pericardial fat somatic pigment y. i The z-axis coordinates of the pericardial fat somatic pigment are y-axis and z-axis. i The z-axis coordinate of the pericardial fat somatic pigment (y) is represented. Calculate the covariance matrix. Where N is the total number of voxels within the local cubic fat mass. This represents the geometric center of the point set. Eigenvalue decomposition is performed on the covariance matrix. The covariance matrix is ​​a 3×3 vector, Σ(x)= This describes the joint distribution and extension relationship of the local point cloud in three spatial directions; v j λ represents the j-th eigenvector, which is a three-dimensional spatial direction vector; j Let be the j-th eigenvalue, representing the data in direction v. j The variance of the variance is taken as follows. The direction of the first principal component represents the direction of maximum extension of pericardial fat within the local cubic neighborhood: .

[0095] To ensure directional stability, a directional consistency constraint is introduced: Ultimately, for each adiposome x, a set of directional regions of interest (ROIs) for intracardiac fat is formed. The format is unified as

[0096] In non-contrast plain CT scans, coronary artery calcification typically appears as strip-shaped, chain-like, or arc-shaped high-density structures distributed along the course of the coronary arteries. Although the coronary vessels themselves cannot be directly segmented, the spatial morphology of the calcification still implies information about the main course of the underlying vessel. To explicitly characterize this directional structural feature, three-dimensional connected component analysis was used to divide the discrete coronary artery calcification region into multiple independent lesion units for the coronary artery calcification ROI. Three-dimensional PCA analysis was performed on each connected region, and finally, the first principal component feature vector was used as the principal axis direction of the coronary artery calcification connected region to characterize its course along the underlying coronary artery.

[0097] For coronary calcification ROIs, using 26-neighbor connectivity, voxel x and y are defined as connected if and only if... x represents the current voxel, and y represents the voxel adjacent to x. When x and y are infinitely close, x and y are connected. For the coronary artery calcification ROI mask M... c Perform three-dimensional connected component labeling Where CCA (Connected Components Analysis) represents three-dimensional connected component analysis; N represents the number of connected components in the coronary calcification ROI. This indicates that 26-neighbor connectivity is used in three-dimensional space to perform connectivity labeling on the coronary calcification binary mask Mc, which aggregates spatially continuous calcification voxels into several independent calcification connectivity regions.

[0098] To avoid the influence of noise, small connected components can be filtered. , It is typically set to 10-30 pixels. The i-th connected component is represented as a set of three-dimensional points: Calculate the geometric center of a 3D point set: ,in ;x j N represents the coronary artery calcification voxel point in the i-th connected region; i This represents the number of voxels in the i-th connected component. And the three-dimensional covariance matrix: Eigenvalue decomposition of the three-dimensional covariance matrix: v i,k λ represents the k-th eigenvector in the i-th connected component, which is a three-dimensional spatial direction vector; i,k Let k be the eigenvalue of the k-th direction vector in the i-th connected component; after eigenvalue decomposition, the first principal component eigenvector is taken as the principal axis direction of the connected component. This direction implies the potential course of the coronary artery, and a set of directions is constructed for coronary artery calcification ROIs: N is the number of connected components, N = 1, 2, ..., i. The principal direction can be obtained through weighted averaging. d iLet |C| be the directional feature vector of the i-th connected component of the coronary calcification ROI. i | represents the number of voxels in the i-th connected component, |C j | represents the number of voxels in the j-th connected component.

[0099] Step 202: Under the constraint of the orientation field, extract first-order radiomics features and directional texture features, and perform parameter adaptive adjustment for different structures to obtain directional radiomics features.

[0100] Traditional radiomics feature extraction typically assumes that image texture is spatially isotropic, meaning that texture statistics are direction-independent. However, in coronary artery disease-related structures (pericardium, intrapericardial fat, coronary artery calcification), image texture and grayscale distribution often exhibit significant direction dependence along the potential coronary artery course. Therefore, this invention introduces a structure-driven direction vector as a directional constraint for radiomics computation, enabling features to selectively respond to anatomical structure directions while maintaining the radiomics statistical framework. This invention constructs relevant directional priors based on pericardium, intrapericardial fat, and coronary artery calcification structures, and calculates first-order and higher-order radiomics features under this directional constraint. Through adaptive adjustments to spatial offset, grayscale discretization, and scale parameters in PyRadiomics, the radiomics features can explicitly reflect coronary artery course-related structural and texture changes, thereby improving the sensitivity and stability of non-contrast CT coronary artery disease screening.

[0101] Specifically, when extracting first-order imagemic features, the first step is to model the directional constraint grayscale variation in each direction. Upper definition of grayscale difference Where I represents the grayscale function on the CT image; x represents the current voxel; δ represents the step size, typically 1-2 voxels; d k Denotes the spatial vector originating from the current voxel; x+δdk represents the vector along the direction d. k The adjacent voxel positions.

[0102] The gray-level difference refers to the degree of difference in gray-level values ​​between adjacent or related voxels under given spatial orientation and distance constraints, and is used to characterize the roughness, drastic changes, and structural heterogeneity of image texture.

[0103] Secondly, within the ROI, statistics were performed on all voxels that met the conditions: Mean (mean of grayscale change), Standard Deviation (standard deviation of grayscale change), Skewness (skewness of grayscale change), Kurtosis (kurtosis of grayscale change), and Energy (energy of grayscale change). These statistical values ​​reflect the intensity and stability of grayscale changes along the coronary artery-related directions.

[0104] Specifically, all those that meet the requirements The set of voxels for the condition is: The corresponding set of grayscale changes is: Where N is the number of voxel points.

[0105] The formula for calculating Mean is: .

[0106] The formula for calculating Standard Deviation is as follows: .

[0107] The formula for calculating Skewness is as follows: .

[0108] The formula for calculating Kurtosis is: .

[0109] The formula for calculating Energy is: .

[0110] This enables the extraction of first-order radiomics features.

[0111] Directional texture features include directional GLCM (Gray-Level Co-occurrence Matrix) features and directional GLRLM (Run-Length Matrix) features.

[0112] The directional GLCM feature is as follows: , representing the number of times a voxel with gray value i appears along a given direction dk within the ROI, and the number of times a voxel with gray value j appears in its adjacent direction. Where s is the voxel coordinate within the ROI; g(s) is the gray level of voxel s; (i,j) is the gray level index; dk... k This is the eigenvector of the k-th direction.

[0113] Statistics only along the direction Gray combination of adjacent voxels Calculate Contrast: Correlation: Homogeneity: Energy: Entropy: ,in , The same applies to j.

[0114] To depict the texture thickness, uniformity, and other characteristics along a specific direction, for each direction d k get When there are k directions, .

[0115] In coronary artery disease-related structures, the grayscale of images often exhibits a continuous, extended, or discontinuous distribution along a specific spatial direction. For example, calcifications are distributed linearly or arc-shaped along the direction of the potential coronary artery; pericardial fat is distributed in strips along the direction of the blood vessel in the area near the coronary artery; and pericardial structures exhibit layered or shell-like textures in the normal direction.

[0116] The Directional Gray-Scale Run-Length Matrix (GLRLM) statistically analyzes the length distribution of consecutive identical gray-scale voxels in a given direction dk, and extracts Short Run Emphasis, Long Run Emphasis, and Run Percentage to characterize the continuity and fracture properties of the structure in different directions.

[0117] Specifically, statistics are performed only within the given structural ROIs of pericardial ROI, intracardiac fat ROI, and coronary artery calcification ROI, and the CT grayscale is discretized into a finite number of grayscale levels: G represents the grayscale level, which can be 32 or 64.

[0118] In direction Above, a run is defined as a continuous segment of voxels with identical gray values, and adjacent voxels are at positions of... Formal representation as ,satisfy If the previous voxel or the next voxel does not meet the condition, the journey ends.

[0119] Construct a directional GLRLM matrix, with direction d k The following GLRLM is defined as: Where i represents the gray level and r is the run length, i.e., the number of continuous voxels. Indicates in direction d k The number of runs with grayscale value i and length r.

[0120] In a fixed direction Above computational features: Short Run Emphasis: LongRun Emphasis: Run Percentage: ,in Total number of tours This represents the total number of voxels within the ROI. For each direction... ,get For K directions, .

[0121] After extracting first-order radiomics features and directional texture features, to ensure that the gray-level statistics of different structures are sensitive to pathological differences rather than noise, the pericardial binwith is set to a fixed HU range during the gray-level discretization stage. In this invention, a 20-30 HU range is preferred to adapt to the gray-level range of soft tissue. Pericardial fat itself has a narrow HU range; therefore, to improve resolution, the pericardial fat binwith is set to a smaller HU range, preferably 10-20 HU in this invention. Coronary artery calcification has a high HU range; therefore, a larger HU range can be set, preferably 40-50 HU in this invention to reduce redundancy.

[0122] Secondly, to capture the fine linear structure of calcification, the distance parameter of the ROI for coronary calcification was set to 1; to capture the slowly changing texture of fat, the distance parameter of the ROI for pericardial fat was set to 2; and to capture changes in pericardial morphology, the distance parameter of the ROI for pericardial fat was set to 5.

[0123] To avoid the curse of dimensionality caused by too many directions, 8–13 structure-related directions are extracted for each ROI.

[0124] Through the above steps, the directional radiomics features of each ROI are finally obtained. Where r is the pericardial ROI, intracardiac fat ROI, or coronary calcification ROI, K is the number of directions, and F is the number of radiomics features in each direction, thus achieving multi-directional and multi-dimensional feature representation.

[0125] Step 3 can be implemented through a radiomics feature fusion unit, and the specific implementation process includes:

[0126] In coronary artery disease (CAD) screening, the contributions of different anatomical structures to disease risk exhibit significant asymmetry and complementarity. Coronary artery calcification regions of interest (ROIs) are highly indicative and specific, pericardial fat is associated with inflammation and carries indirect risk, and pericardial morphology is constrained by overall anatomical and mechanical factors. Therefore, this invention utilizes a structural-level fusion network (SFN) to perform directional attention modeling on the directional radiomics features of the pericardium, pericardial fat, and coronary artery calcification structures. This enables end-to-end fusion of multi-ROI and multi-directional features, learns the contribution weights of different structures and directions to CAD risk, generates highly discriminative structural-level feature representations, and generates anatomically interpretable structural-level representations, providing highly discriminative feature representations for CAD risk assessment.

[0127] The structure-level fusion network includes a directional feature embedding layer module, a directional attention aggregation module, a structure-specific mapping module, and a regularization and stabilization module.

[0128] Since the image omics features from different directions differ in scale and statistical distribution, the directional feature embedding layer module embeds the feature vectors from each direction. Where r belongs to the pericardial ROI, intracardiac fat ROI, and coronary calcification ROI; k represents the k-th direction vector. A linear transformation and activation function are used to map to a unified latent space, and then... ,in W e b e For the learnable parameters of the feature embedding layer, It is a GELU or ReLU activation function.

[0129] Secondly, not all directions are highly correlated with the risk of coronary artery disease, especially in pericardial ROIs and intracardiac fat ROIs. The direction-level attention aggregation module automatically identifies high-risk directions consistent with potential coronary artery directions and calculates the importance weight of each direction using a preset attention weight formula. The attention weight calculation formula is as follows: The embedded directional features are weighted and aggregated to output... This allows the model to focus on directional features related to coronary artery pathological changes.

[0130] Because the pericardium, intracardiac fat, and coronary artery calcifications have different pathological significance, the structure-specific mapping module maps intermediate features of different structures to the same structural semantic space. , The weights of the three structural ROIs; It can be an activation function composed of normalization layers, dropout layers, and ReLU, which enhances the discriminative power of features and prevents overfitting.

[0131] Finally, the regularization and stabilization modules use sparse regularization to encourage SFN to focus on a few key directions. This suppresses interference from redundant directions and improves the generalization ability and stability of SFN.

[0132] The directional radiomics features of the pericardial ROI, intracardiac fat ROI, and coronary artery calcification ROI output from the chest CT data preprocessing unit are input into the structural fusion network. After directional feature embedding, directional attention aggregation, structure-specific mapping, and regularization, the structural feature representation of each structure is output, realizing deep fusion and optimization of multi-source features.

[0133] The result prediction output unit in step 4 specifically includes:

[0134] The result prediction output unit receives the structural feature representation output by the radiomics feature fusion unit, and outputs the final coronary heart disease risk prediction result through classification head model training and inference, providing a direct basis for clinical decision-making.

[0135] The classification head uses a two-layer multilayer perceptron (MLP) combined with a softmax function to achieve the mapping from structural features to prediction results. The specific calculation formula is as follows: ,in W1, b1, W2, and b2 are learnable model parameters; BN is batch normalization, and the Dropout layer is used to prevent overfitting.

[0136] An end-to-end training approach is adopted, freezing the relevant parameters for target mask extraction in the chest CT data preprocessing unit to avoid error propagation during preprocessing. The cross-entropy loss function is used as the optimization objective. Where c is the category label and y is the actual label. To predict probabilities, during training, the parameters of the structural fusion network and the classification head are iteratively updated using the training set, while the validation set monitors model performance to avoid overfitting.

[0137] The parameters of the front-end target structure mask extraction module are fixed and not updated. Backpropagation and parameter updates are only performed on the back-end structural feature fusion network and classification head. Parameter freezing methods include disabling gradient updates for all parameters in the target structure mask extraction network during the training initialization phase, or fixing the segmentation network to the eval state during the training phase. By freezing parameters, the target mask extraction module can be treated as a fixed, non-learnable preprocessing operator, making its output only the input to the downstream network.

[0138] The classification head receives the structural feature representation output by the radiomics feature fusion unit and outputs the coronary heart disease risk prediction results, including CAD-RADS classification, whether it is coronary heart disease, and the probability of coronary heart disease risk, providing a comprehensive and accurate risk assessment basis for clinical practice.

[0139] The CAD-RADS classification includes 6 levels of classification results, with outputs of 0, 1, 2, 3, 4, and 5 corresponding to CAD-RADS classification 0, 1, 2, 3, 4, and 5, respectively.

[0140] The determination of whether it is coronary heart disease is a binary classification result. According to the CAD-RADS classification, a score of 0, 1, or 2 indicates non-coronary heart disease; a score of 3, 4, or 5 indicates coronary heart disease.

[0141] The probability of coronary heart disease risk is directly output by a 2-layer MLP without going through an activation function, and ranges from 0 to 1. The closer it is to 0, the lower the risk, and the closer it is to 1, the higher the risk.

[0142] The following is a detailed description with reference to specific embodiments:

[0143] Example 1: For people undergoing health checkups and lung disease screening, chest CT scans are widely used for screening diseases such as lung nodules. This example achieves automated screening of high-risk individuals for coronary heart disease through the collaborative work of four units without adding additional imaging examinations.

[0144] The chest CT data preprocessing unit acquires plain chest CT data from the physical examination population, filters data that meet the standards, and anonymizes and standardizes them. It then uses automated segmentation to extract pericardial ROIs, intracardiac fat ROIs, and coronary artery calcification ROIs, completing the preprocessing. The radiomics feature extraction unit constructs a priori orientation field for the coronary arteries based on the preprocessed pericardial ROIs, intracardiac fat ROIs, and coronary artery calcification ROIs, extracting directional first-order features and texture features. The radiomics feature fusion unit inputs the features from the pericardial ROIs, intracardiac fat ROIs, and coronary artery calcification ROIs into a structural-level fusion network, generating structural-level features through attention fusion. The results prediction output unit outputs the individual's coronary artery disease risk probability in its classification head, used for high-risk population labeling or to recommend further clinical examinations.

[0145] Example 2: For patients suspected of having coronary heart disease, this provides a basis for decision-making regarding whether to perform CT angiography or coronary angiography, avoiding unnecessary invasive or high-cost examinations.

[0146] The chest CT data preprocessing unit acquires the patient's existing chest or cardiac plain CT data, performs data screening, standardization, and extraction of pericardial ROI, intracardiac fat ROI, and coronary artery calcification ROI. The radiomics feature extraction unit constructs a directional field and extracts directional radiomics features. The radiomics feature fusion unit achieves deep feature fusion. The results prediction output unit outputs coronary artery disease risk prediction results; low-risk patients are advised to follow up or conservative management, while intermediate- and high-risk patients are advised to undergo further CT angiography or coronary angiography.

[0147] Example 3: In view of the chronic progressive characteristics of coronary heart disease, the disease progression, efficacy and risk changes were evaluated by analyzing the CT data of the same patient at multiple stages.

[0148] The chest CT data preprocessing unit acquires plain CT data from the same patient at different time points and performs data preprocessing for each phase using the same workflow to ensure consistency in pericardial ROI, intracardiac fat ROI, and coronary artery calcification ROI across multiple phases. The radiomics feature extraction unit constructs a consistent orientation field based on an anatomical structure alignment strategy and extracts multi-phase directional radiomics features. The radiomics feature fusion unit outputs multi-phase structural features. The results prediction output unit combines the changing trends of multi-phase structural features to assess whether coronary artery disease has progressed, the effectiveness of treatment interventions, and the direction of risk changes.

[0149] This invention presents a method and system for predicting coronary artery disease (CAD) risk by integrating multiple radiomics features. First, when coronary arteries cannot be directly visualized on plain CT scans, coronary artery-related anatomical structures are used to replace the invisible vessel body, enabling collaborative modeling of multi-structure, multi-source radiomics information and improving the model's robustness and accuracy. Second, directional radiomics modeling is employed, overcoming the limitations of traditional isotropic feature statistics. Multi-directional feature representations are constructed around the potential course of the coronary arteries and anatomical constraints, more fully reflecting the spatial heterogeneity of CAD-related pathological processes. Third, directional radiomics is upgraded from static feature engineering to a learnable networked modeling framework. End-to-end joint learning of features is achieved through a structural-level multi-layer attention mechanism, enhancing the model's discriminative ability and interpretability. Finally, relying on conventional chest CT data, CAD risk is simultaneously assessed while screening for lung diseases, effectively reducing medical costs and patient burden, increasing the early detection rate of CAD, and providing technical support for the comprehensive management of cardiopulmonary comorbidities.

[0150] The above description illustrates preferred embodiments of the present invention and helps those skilled in the art to more fully understand the technical solution of the present invention. However, these embodiments are merely illustrative and should not be construed as limiting the specific implementation of the present invention to these embodiments. For those skilled in the art, several simple deductions and modifications can be made without departing from the inventive concept, and all such modifications should be considered within the protection scope of the present invention.

Claims

1. A method for predicting the risk of coronary heart disease by integrating multiple radiomics features, characterized in that, include: Step 1: Obtain the patient's original chest CT scan data and perform preprocessing to obtain a structural mask of the pericardium, pericardial fat, and coronary artery calcification; Step 2: Based on the structural masks of the pericardium, pericardial fat, and coronary artery calcification, construct orientation fields respectively, and extract radiomics features of the pericardium, pericardial fat, and coronary artery calcification. Step 3: Weight the radiomic features of the pericardium, intrapericardial fat, and coronary artery calcification to obtain structural features; Step 4: Output the coronary heart disease risk prediction results based on structural features.

2. The method for predicting coronary heart disease risk by fusing multiple radiomics features according to claim 1, characterized in that, Step 1 includes: Step 101: Obtain the raw medical image data of the chest plain CT scan and the corresponding coronary heart disease label; Step 102: Anonymize and standardize the raw medical image data of chest plain CT scan to obtain normalized raw medical image data of chest plain CT scan. Step 103: Extract the three-dimensional segmentation mask of the pericardium, pericardial fat and coronary artery calcification using automatic or manual segmentation methods; Step 104: Preprocess the three-dimensional segmentation mask of the pericardium, intrapericardial fat and coronary artery calcification to obtain the corresponding structural ROIs, which include the pericardium ROI, intrapericardial fat ROI and coronary artery calcification ROI.

3. The method for predicting coronary heart disease risk by fusing multiple radiomics features according to claim 2, characterized in that, The raw medical imaging data of the chest plain CT scan were obtained during lung disease screening without additional imaging examinations or radiation doses.

4. The method for predicting coronary heart disease risk by fusing multiple radiomics features according to claim 2, characterized in that, The structural ROIs were resampled to a uniform resolution using trilinear interpolation. In the HU window, the pericardial ROI was truncated to the range of [-200, 300], the intracardiac fat ROI was truncated to the range of [-190, -30], and the coronary calcification ROI was truncated to the range of [130, 1000]. Finally, the three ROIs were normalized to the range of [0, 1] using Min-Max or Z-score.

5. The method for predicting coronary heart disease risk by fusing multiple radiomics features according to claim 2, characterized in that, Step 2 includes: Step 201: Construct prior orientation fields for the pericardial ROI, intracardiac fat ROI, and coronary calcification ROI respectively; Step 202: Under the constraint of the orientation field, extract first-order radiomics features and directional texture features, and perform parameter adaptive adjustment for different structures to obtain directional radiomics features.

6. The method for predicting coronary heart disease risk by fusing multiple radiomics features according to claim 5, characterized in that, For the pericardial ROI, extract all voxel points Ω within the pericardial ROI: Calculate the centroid c of the pericardial ROI. For each voxel point x within the ROI i Calculate the direction vector in the x-direction. And calculate the direction vectors in the y and z directions to obtain the three-dimensional direction vector. ;x i The x-coordinate of voxel point i represents the x-axis coordinate; y i The z-axis coordinate represents the y-coordinate of voxel point i; i The z-coordinate of voxel point i is represented; i = 1, 2, 3, ..., N, where N is the total number of voxels within the pericardial ROI; M represents the set of all voxel points within the pericardial ROI. For the pericardial fat region of interest (ROI), let S be... p It is the set of voxel points of the pericardial inner wall, and the set of all points in the pericardial region of interest (ROI) that have at least one neighbor that does not belong to the ROI. The goal is to extract the set of points on the surface of the pericardial inner wall from the pericardial ROI. s represents the spatial location (x, y, z) of a voxel; M p (s) represents the binary mask of the pericardial ROI; Let s' represent the 13-neighborhood of point s; for each pericardial fat somatic cell... M f As a binary mask for pericardial fat, the normal direction vector of the pericardial fat voxel x pointing towards the inner pericardial wall is calculated as follows: , For a pericardial fat body; Pericardial fat body The Euclidean shortest distance to the inner wall of the pericardium is determined by combining the first principal component direction of the local cubic neighborhood and introducing a direction consistency constraint. Obtain the set of directions of the pericardial fat ROI. ; For coronary calcification ROIs, a 26-neighbor connectivity method is used to mask the coronary calcification ROI M. c Perform three-dimensional connected component labeling, aggregating spatially continuous calcified voxels into several independent calcified connected components, filtering out small-volume connected components, and representing the i-th connected component as a three-dimensional point set. Calculate the geometric center of a 3D point set. and the three-dimensional covariance matrix , ;x j N represents the coronary artery calcification voxel point in the i-th connected region; i This represents the number of voxels in the i-th connected component; eigenvalue decomposition is performed on the 3D covariance matrix, and the direction of the first principal component is taken as the principal axis direction of the connected component. Forming a set of ROI directions for coronary artery calcification , N=1,2,...i.

7. The method for predicting coronary heart disease risk by fusing multiple radiomics features according to claim 5, characterized in that, Under orientation field constraints, first-order imagemic features are extracted, including: When extracting first-order imageomics features, modeling the directional constraint grayscale variation is performed in each direction. Upper definition of grayscale difference Where I represents the grayscale value function on the CT image; x represents the current voxel; δ represents the step size; d k Denotes the spatial vector originating from the current voxel; x+δdk represents the vector along the direction d. k The adjacent voxel positions, for all satisfying Mean value of voxel statistical grayscale change under conditions Standard deviation of grayscale variation Gray-scale variation skewness Gray-scale variation kurtosis and grayscale change energy .

8. The method for predicting coronary heart disease risk by integrating multiple radiomics features according to claim 5, characterized in that, Directional texture features include directional gray-level co-occurrence matrix features and directional run-length matrix features; The directional gray-level co-occurrence matrix features s represents the voxel coordinates within the ROI; g(s) represents the gray level of voxel s. (i,j) is the grayscale index; d k This is the eigenvector of the k-th direction; Statistics along direction Gray combination of adjacent voxels Calculate contrast Correlation Homogeneity ,energy ,entropy , , Calculate j; for each direction d k get For k directions, ; The directional gray-level run-length matrix statistically distributes the lengths of consecutive identical gray-level voxels along a given direction dk, extracting short run-length emphasis. Long travel itinerary Percentage of trip , Total number of tours The total number of voxels within the ROI; for each direction ,get For K directions, .

9. The method for predicting coronary heart disease risk by fusing multiple radiomics features according to claim 5, characterized in that, After extracting first-order radiomics features and directional texture features, the grayscale discretization binwidth was set to a fixed HU interval, specifically 20-30 HU for pericardium, 10-20 HU for intrapericardial fat, and 40-50 HU for coronary artery calcification. The distance parameter for the ROI of coronary artery calcification was 1, the distance parameter for the ROI of intrapericardial fat was 2, and the distance parameter for the ROI of pericardium was 5. 8-13 structurally relevant orientations were extracted for each ROI to obtain the directional radiomics features for each ROI. r represents the pericardial ROI, intracardiac fat ROI, or coronary calcification ROI, K represents the number of directions, and F represents the number of radiomics features in each direction.

10. The method for predicting coronary heart disease risk by fusing multiple radiomics features according to claim 1, characterized in that, Weighted radiomics features of pericardium, intrapericardial fat, and coronary artery calcification are obtained to form structural features, including: directional attention modeling of directional radiomics features of pericardium, intrapericardial fat, and coronary artery calcification structures through a structural fusion network. The structural fusion network includes a directional feature embedding layer module, a directional attention aggregation module, a structure-specific mapping module, and a regularization and stabilization module. The directional feature embedding layer module will embed the feature vector of each direction. , r belongs to the pericardial ROI, intracardiac fat ROI, and coronary calcification ROI; k represents the k-th direction vector; through linear transformation and activation function mapping to a unified latent space, execution is performed. ; The directional attention aggregation module automatically identifies high-risk directions consistent with the potential coronary artery direction, and uses a preset attention weight formula. Calculate the importance weight for each direction, perform weighted aggregation on the embedded directional features, and output the result. ; The structure-specific mapping module maps intermediate features of different structures to the same structural-level semantic space. , The weights of the three structural ROIs; The regularization and stabilization module uses directional sparse regularization to encourage SFN to focus on key directions and output structural feature representations for each structure.

11. The method for predicting coronary heart disease risk by fusing multiple radiomics features according to claim 1, characterized in that, Step 4 outputs the coronary heart disease risk prediction result based on structural features, including: receiving the structural feature representation, and outputting the coronary heart disease risk prediction result through the classification head model, including CAD-RADS classification, whether it is coronary heart disease, and the probability of coronary heart disease risk; The classification head uses a two-layer multilayer perceptron (MLP) combined with a softmax function. An end-to-end training approach was adopted, freezing the relevant parameters for target mask extraction in the chest CT data preprocessing unit, and using the cross-entropy loss function as the optimization objective. , where c is the category label and y is the actual label; To predict probabilities.

12. A coronary heart disease risk prediction system integrating multi-radiomics features, characterized in that, The chest CT data preprocessing unit is used to acquire the patient's raw chest plain CT data and preprocess it to obtain structural masks of the pericardium, pericardial fat and coronary artery calcification. Based on the structural masks of the pericardium, intrapericardial fat, and coronary artery calcification, radiomic features were constructed to extract radiomic features of the pericardium, intrapericardial fat, and coronary artery calcification. The radiomics feature fusion unit is used to weight the radiomics features of the pericardium, intrapericardial fat, and coronary artery calcification to obtain structural-level features; The result prediction output unit is used to output the coronary heart disease risk prediction result based on structural features.

Images