Method and apparatus for assessing chronic cardiometabolic disease based on chest ct images

Abstract

This invention provides a method and apparatus for assessing chronic cardiovascular metabolic diseases based on chest CT images. The method includes: obtaining chest CT image data and clinical information of a subject; obtaining cardiac parameters of the subject based on the chest CT image data, including the volume of each chamber of the heart, left ventricular myocardial mass, epicardial fat volume, average epicardial fat density, and coronary artery calcium score; obtaining body composition data of the subject based on the chest CT image data, including bone data, muscle data, intramuscular fat data, visceral fat data, and subcutaneous fat data; and obtaining an assessment result of the subject based on the cardiac parameters, body composition data, and clinical information.

Description

Technical Field

[0001] This invention relates to disease assessment methods based on medical image data and artificial intelligence, and in particular to a method and apparatus for assessing chronic cardiovascular metabolic diseases based on chest CT images. Background Technology

[0002] Cardiometabolic diseases (CMDs) are not single diseases, but a collective term for a group of interconnected diseases centered on metabolic dysfunction and damage to the cardiovascular system. Therefore, interorgan crosstalk plays a role in the development of these diseases. Their core characteristic is that metabolic abnormalities and cardiovascular damage are mutually causal, forming a vicious cycle that significantly increases the risk of serious cardiovascular events such as myocardial infarction and stroke, as well as death. This group of diseases mainly includes metabolic disorders such as type 2 diabetes, prediabetes, obesity (especially abdominal obesity), and dyslipidemia, as well as cardiovascular diseases such as hypertension, coronary heart disease, stroke, and peripheral artery disease. Metabolic syndrome, as a collection of multiple metabolic abnormalities (such as the simultaneous presence of three or more of abdominal obesity, hypertension, hyperglycemia, and dyslipidemia), is an important health warning sign. This complex disease, encompassing multiple organ systems including the heart, blood vessels, brain, muscles, fat, bones, liver, kidneys, and pancreas, is quietly threatening global health and is one of the leading causes of death and disability worldwide. These diseases exhibit complex systemic interactions that collectively accelerate multi-organ damage and significantly increase the risk of all-cause mortality and cardiovascular mortality.

[0003] Chronic kidney disease (CKD) and cardiovascular metabolic diseases (CMDs) are bidirectionally linked and mutually causal comorbidities. They share a common pathogenesis, and the presence of one significantly accelerates the onset and progression of the other, creating a vicious cycle that ultimately greatly increases the risk of complications and death in patients. During the progression of CKD, changes in cardiac structure and function are key factors leading to poor prognosis. Early assessment of cardiac status and risk stratification in CKD patients plays a crucial role in preventing disease progression and complications.

[0004] Given the interrelationships among these chronic diseases, early and comprehensive assessment and risk prediction of patients are of great significance for implementing personalized interventions, helping patients manage their health, and preventing acute events.

[0005] Chest CT is a common examination in clinical radiology, and its application in routine physical examinations is becoming increasingly widespread. Its routine uses are primarily focused on screening and diagnosing lung diseases and cardiovascular diseases. However, as an imaging technique that can clearly display the thoracic cavity and surrounding tissue structures, the rich information contained in chest CT has not been fully explored and utilized. Summary of the Invention

[0006] This invention provides a method for assessing chronic cardiovascular metabolic diseases based on chest CT images, comprising: obtaining chest CT image data and clinical information of a subject; obtaining cardiac parameters of the subject based on the chest CT image data, the cardiac parameters including the volume of each chamber of the heart, left ventricular myocardial mass, epicardial fat volume, average epicardial fat density, and coronary artery calcium score; obtaining body composition data of the subject based on the chest CT image data, the body composition data including bone data, muscle data, intramuscular fat data, visceral fat data, and subcutaneous fat data; and obtaining the assessment result of the subject based on the cardiac parameters, body composition data, and clinical information.

[0007] In one embodiment, the method for obtaining the patient's cardiac parameters based on the chest CT image data includes: segmenting the chest CT image data using a TotalSegmentator image segmentation tool to obtain a left ventricular mask, a right ventricular mask, a left atrial mask, a right atrial mask, and a left ventricular myocardial mask; calculating the volume of each cardiac chamber and the left ventricular myocardial mass based on the segmentation results; segmenting the entire cardiac mask using a cardiac mask segmentation model trained on an nnU-Net model; obtaining adipose tissue with a CT value between -190 HU and -30 HU within the cardiac mask through threshold segmentation; calculating the epicardial fat volume and average epicardial fat density based on the segmented adipose tissue within the cardiac mask; segmenting the coronary artery mask using a coronary artery mask segmentation model; obtaining calcifications with a CT value greater than 130 HU through threshold segmentation; and calculating the coronary artery calcification integral based on the calcification area and weighting factors corresponding to different calcification densities.

[0008] In one embodiment, the method for obtaining the body composition data of the subject based on the chest CT image data includes: determining three standard analysis levels in the chest CT image data; segmenting the body composition data at each standard analysis level using a body composition segmentation model trained on an nnU-Net model to obtain skeletal masks, muscle masks, intermuscular fat masks, visceral fat masks, and subcutaneous fat masks; and calculating the area and average density of each type of mask region based on the segmentation results.

[0009] In one embodiment, the method for determining three standard analysis layers in the chest CT image data includes: segmenting the chest CT image data using the TotalSegmentator image segmentation tool to obtain a spinal mask and a rib mask; determining the lowest rib as the twelfth rib based on the rib mask, identifying the vertebral body connected to the twelfth rib in the spinal mask as the twelfth thoracic vertebra, traversing all layers of the twelfth thoracic vertebra along the axial direction, calculating the area of ​​the twelfth thoracic vertebra on each layer, and selecting the layer with the largest area of ​​the twelfth thoracic vertebra as the T12 analysis layer; determining the first vertebral body below the twelfth thoracic vertebra as the first lumbar vertebra; traversing all layers of the first lumbar vertebra along the axial direction, evaluating the integrity of the transverse process structure on each layer through image morphology, and selecting the layer with the most complete transverse process structure as the L1 analysis layer; and determining the layer containing the intervertebral disc between the lower edge of the twelfth thoracic vertebra and the upper edge of the first lumbar vertebra, where the anterior vertebral body and posterior spinous process are completely separated on the image, and defining this layer as the T13 analysis layer.

[0010] In one embodiment, the clinical information includes at least one of age, sex, systolic blood pressure, diastolic blood pressure, body mass index, history of diabetes, total cholesterol, high-density lipoprotein, low-density lipoprotein, triglycerides, and blood glucose.

[0011] In one embodiment, the assessment results of the subject include the predicted probability of the subject's all-cause mortality risk and / or cardiovascular disease mortality risk.

[0012] In one embodiment, the method for obtaining the assessment result of the subject based on the subject's cardiac parameters, body composition data, and clinical information includes: constructing a prediction model based on statistics or artificial intelligence, and using a large amount of real chronic cardiovascular metabolic disease patients' cardiac parameters, body composition data, clinical information, as well as time of death and cause of death accumulated in the hospital as training data for training, to obtain a trained cardiovascular metabolic disease patient mortality risk prediction model; inputting the subject's cardiac parameters, body composition data, and clinical information into the trained cardiovascular metabolic disease patient mortality risk prediction model to obtain the predicted probability of the subject's all-cause mortality risk and / or cardiovascular disease mortality risk.

[0013] In one embodiment, the evaluation results are presented in the form of a probability curve, where the horizontal axis represents future time and the vertical axis represents the predicted probability of survival or death.

[0014] In one embodiment, different predicted probabilities are classified into different risk levels according to a preset threshold.

[0015] On the other hand, the present invention also provides a chronic cardiovascular metabolic disease assessment device based on chest CT images, comprising: a data input module configured to obtain chest CT image data and clinical information of a subject; a cardiac parameter calculation module configured to obtain cardiac parameters of the subject based on the CT image data, the cardiac parameters including the volume of each heart chamber, left ventricular myocardial mass, epicardial fat volume, average epicardial fat density, and coronary artery calcium score; a body composition data calculation module configured to obtain body composition data of the subject based on the CT image data, the body composition data including bone data, muscle data, intramuscular fat data, visceral fat data, and subcutaneous fat data; and an assessment module configured to obtain the assessment result of the subject based on the cardiac parameters, body composition data, and clinical information of the subject. The body composition data calculation module is configured to perform the following steps: determine three standard analysis levels in the chest CT image data, namely the T12 analysis level, the L1 analysis level, and the T13 analysis level; use the body composition segmentation model to segment the bone mask, muscle mask, intermuscular fat mask, visceral fat mask, and subcutaneous fat mask on each standard analysis level; and calculate the area and average density of each type of mask region based on the segmentation results.

[0016] This invention, as a chronic cardiovascular metabolic disease assessment scheme based on chest CT images, effectively overcomes existing technological bottlenecks and demonstrates multiple core advantages in clinical application: First, it significantly improves assessment efficiency and consistency by relying on a deep learning model to achieve efficient and automatic extraction of cardiac parameters (volume of each cardiac chamber, left ventricular myocardial mass, etc.) and body composition data (muscle, fat, bone density, etc.), laying the foundation for large-scale screening; Second, it lowers the threshold for clinical application and the burden on patients by enabling further in-depth assessment of patients who have already undergone chest CT examinations based on routine chest CT images, maximizing the value of existing medical imaging resources; Third, it significantly improves risk assessment. Precision and targeting: The risk prediction model is trained on a large amount of data from patients with chronic cardiovascular and metabolic diseases. It can integrate multi-dimensional parameters (including cardiovascular, body composition and clinical information) and output the probability of all-cause mortality, the probability of cardiovascular mortality and low / medium / high risk stratification of patients within a specific time window in the future. This provides a reliable basis for clinical development of individualized intervention strategies and early prevention of acute events (such as myocardial infarction and stroke). Fourth, by extracting body composition data through standardized anatomical levels, it ensures the comparability of measurement data from different patients and different follow-up periods of the same patient. This helps doctors dynamically monitor disease progression, evaluate treatment effects and provide data support for optimizing long-term treatment plans. Attached Figure Description

[0017] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 Roadmap for automated analysis technology of cardiac parameters and body composition; Figure 2 Correlation between system-measured cardiac parameters and the gold standard; Figure 3 Correlation between calcification score and gold standard; Figure 4 Display of model segmentation effects at each level; Figure 5 Consistency in the measured area of ​​visceral fat and intermuscular fat; Figure 6 Area under the characteristic curve of subjects in a time-dependent mortality risk prediction model. Detailed Implementation

[0018] The present disclosure will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present disclosure, but do not limit the present disclosure in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present disclosure. These all fall within the protection scope of the present disclosure.

[0019] The core of this invention lies in providing a one-stop, automated medical image analysis and disease risk prediction device. This device automatically extracts comprehensive cardiac and body composition parameters (such as...) from conventional ungated chest CT images by integrating innovative deep learning models and anatomical localization algorithms. Figure 1 This technology, combined with clinical data, ultimately enables precise prediction of individualized mortality risk in patients with chronic cardiovascular and metabolic diseases. The technical solution will be described in detail below, module by module: 1. Cardiac Parameter Calculation Module: This module is responsible for accurately quantifying key cardiac parameters from raw CT images. Its technical implementation is as follows: 1.1 Cardiac Structure and Myocardial Segmentation Submodule: This submodule uses the pre-trained model TotalSegmentator based on the nnU-Net architecture. It is a three-dimensional fully convolutional neural network that can adaptively configure network parameters. Its input is the three-dimensional volume data (DICOM format) of the entire ungated, unenhanced chest CT scan.

[0020] Function and Principle: nnU-Net, through its encoder-decoder structure, progressively downsamples at the encoding end to extract multi-scale features of the image, and progressively upsamples at the decoding end, combining this with skip connections at the encoding end to achieve pixel-level accurate classification. This model is specifically trained to recognize various anatomical structures, including the ventricles of the left ventricle (LV), left atrium (LA), right ventricle (RV), right atrium (RA), and left ventricular myocardium (LV Myocardium).

[0021] The specific implementation steps include: 1) Data preprocessing: The input raw CT data is normalized and resampled according to the model requirements.

[0022] 2) Forward reasoning: The preprocessed volume data is input into the loaded TotalSegmentator model. The model outputs a segmentation map with the same size as the input, where each voxel is assigned a label representing the anatomical structure to which it belongs.

[0023] 3) Parameter Calculation: Based on the segmentation results, the system automatically calculates the volume (unit: mL) of each heart chamber. The specific calculation formula is "Volume = Total number of voxels marked as this heart chamber × Physical volume of a single voxel". Similarly, the mass (unit: g) of the left ventricular myocardium is calculated by "Mass = Myocardial volume × Myocardial tissue density (usually 1.05 g / cm³)".

[0024] 1.2 Epicardial Fat Quantification Submodule Since the TotalSegmentator tool does not provide complete cardiac region segmentation for epicardial fat (EAT) calculation, this invention specifically trained a custom 3D nnU-Net model. This model was fine-tuned using 50 manually annotated, finely detailed ungated chest CT scans with complete cardiac contours (including pericardial boundaries).

[0025] Function and Principle: The task of this custom model is to accurately segment the entire heart region, including the myocardium and epicardial fat. Its network structure and training process follow the standard nnU-Net framework, using a 3d_fullres (three-dimensional full resolution) pipeline and training with five-fold cross-validation to ensure the model's generalization ability and robustness.

[0026] The specific implementation steps include: 1) Heart region segmentation: A customized nnU-Net model is used to segment the CT volume data to obtain a binary 3D mask, where "1" represents the heart region (including the heart chambers, myocardium and epicardial space) and "0" represents the background.

[0027] 2) Adipose tissue identification: Within the segmented cardiac region, the system traverses all voxels and performs threshold filtering based on CT values ​​(Hounsfield Units, HU). Voxels with CT values ​​between -190 HU and -30 HU are identified as adipose tissue.

[0028] 3) EAT parameter calculation: EAT volume (unit: cm³ or mL): EAT volume = total number of voxels identified as fat × physical volume of a single voxel. EAT average density (unit: HU): Calculated as the arithmetic mean of the CT values ​​of all voxels identified as fat.

[0029] 4) Parameter Validation: The measurement results of the model's cardiac parameters were validated in data from 117 patients with gold standard coronary gated CTA. The measurements showed good consistency between the two methods. Relevant results are shown below. Figure 2 As shown.

[0030] 1.3 Coronary Artery Calcification Integral Calculation Submodule This submodule integrates a 3D ResU-Net calcification integral segmentation and computation model. This model has been trained on a large amount of paired data (non-gated chest CT and ECG-gated cardiac CT).

[0031] Function and principle: This network can automatically identify and segment the course areas of major coronary arteries such as the left main coronary artery, left anterior descending artery, left circumflex artery and right coronary artery, and detect high-density (calcified) plaques in them.

[0032] Specific implementation steps: 1) Calcified plaque detection and segmentation: Receive CT scan data and output the three-dimensional spatial location of the calcified plaque and the coronary artery to which it belongs.

[0033] 2) Calcification Score Calculation: Following the Agatston scoring method, a score is calculated for each detected calcified lesion. For each continuous calcified region (area ≥ 1 mm², CT value ≥ 130 HU), a weighting factor is assigned based on its maximum CT value (130-199 HU: weighting factor 1; 200-299 HU: weighting factor 2; 300-399 HU: weighting factor 3; ≥400 HU: weighting factor 4). Calcification score = Σ(calcified area × weighting factor). The final output is the score for each vessel and the total calcification score.

[0034] 3) Parameter accuracy verification: In 117 cases with gated calcification scores, manual measurement by doctors was used as the gold standard. The Bland-Altman correlation coefficient and Pearson correlation coefficient were calculated. Figure 3 The study found a high correlation between the calcification score automatically calculated based on chest CT and the gold standard, confirming the accuracy of the model calculation.

[0035] 2. Body composition analysis module based on multi-anatomical level localization: This module is the innovative core for achieving standardized measurement of whole-body body composition.

[0036] 2.1 Automatic Localization Submodule for Multiple Specific Anatomical Levels The input to this submodule is the segmentation mask of the thoracic vertebrae and ribs output by TotalSegmentator. Based on this, three standard planes are automatically located using a pre-set algorithm based on anatomical rules: the plane of the superior endplate of the twelfth thoracic vertebra (T12), the plane of the intervertebral disc between the twelfth thoracic vertebra and the first lumbar vertebra (T13), and the plane of the most complete transverse process of the first lumbar vertebra (L1).

[0037] Function and principle: By utilizing the stability of the skeletal structure and the spatial relative positional relationship, repeatable and accurate layer positioning is achieved by calculating geometric features (such as area and spatial coordinates), avoiding measurement deviations caused by different patient positioning or respiratory phases.

[0038] Specific implementation steps: 1) T12 Vertebral Body Localization: The system identifies the lowest pair of ribs (i.e., the 12th rib) in the rib segmentation mask. Then, in the thoracic vertebrae segmentation mask, it locates the vertebral body connected to the 12th rib; this is the T12 vertebral body. All slices of the T12 vertebral body are traversed along the axial direction (Z-axis), and the cross-sectional area of ​​the T12 vertebral body mask on each slice is calculated. The slice with the largest cross-sectional area is selected and defined as the T12 analysis slice.

[0039] 2) L1 Vertebral Body Localization: Based on the spatial continuity of the vertebral bodies, the system identifies the first vertebra below the T12 vertebra as the L1 vertebra. All slices of the L1 vertebra are traversed along the axial direction, and the integrity of the transverse process structure in each slice is evaluated using image morphology algorithms (e.g., calculating the connectivity or area of ​​pixels in the transverse process region). The slice with the most complete transverse process display is selected and defined as the L1 analysis slice.

[0040] 3) T13 Intervertebral Disc Localization: The intervertebral space between the lower edge of the T12 vertebral body and the upper edge of the L1 vertebral body is systematically located. Within this intervertebral space, a typical transverse section of the intervertebral disc is selected. At this level, the key anatomical landmarks are: the anterior bony structures of the vertebral body and the posterior bony structures of the spinous process are completely separated on the image, connected by the intervertebral disc tissue. This level is defined as the T13 analysis level.

[0041] 2.2 Multi-body component segmentation and quantification submodule A pre-trained 2D U-Net segmentation network is applied to three precisely localized 2D cross-sectional images: T12, T13, and L1. The network takes a single-slice CT image as input and outputs a probability map of each pixel in the image belonging to a different body component category.

[0042] Function and principle: U-Net network can effectively combine the contextual information of the image (through the encoder) and the precise localization information (through skip connections and decoder) in biomedical image segmentation to achieve fine segmentation of areas such as subcutaneous fat, muscle, and visceral fat.

[0043] Specific implementation steps: 1) Image extraction: Extract the two-dimensional CT images of the three localized layers, T12, T13 and L1, from the original CT body data.

[0044] 2) Component segmentation: The images of each layer are input into the two-dimensional U-Net model. The model outputs segmentation results, which usually include the following categories: subcutaneous fat, skeletal muscle, visceral fat, bone and background. The specific segmentation model training process is as follows: (1) Data collection and annotation: 100 cases of chest CT scan data are collected, and the subcutaneous fat, muscle, viscera, vertebral body and bone density areas are manually annotated at the L1, T12 and T13 levels, respectively, to be used as the ground truth of the multi-task segmentation model. (2) Data set partitioning: The dataset is randomly divided into training set (80%), validation set (10%) and test set (10%). (3) Data preprocessing: Before training and inference, all CT images are cropped using the 0.5 and 99.5 percentiles of the foreground voxels and normalized using the mean and standard deviation of the foreground. In order to deal with the heterogeneity of voxel spacing, all images are resampled to the same target spacing, which is defined as the median value of the voxel spacing in all training cases. When considering a two-dimensional (2D) segmentation model, this resampling process is performed on the two axes with the highest resolution. The image and mask are interpolated using cubic spline interpolation and nearest neighbor interpolation, respectively. (4) Model architecture: The backbone network for performing the thoracic and abdominal component segmentation is a 2D U-Net. The input image contains one channel. Our segmentation model consists of 6 encoder and 6 decoder convolutional blocks connected by a Bottle-neck block, downsampling the 640x640 voxel input image to 10x10 voxels and then upsampling it to a 640x640 voxel segmentation mask. Each encoder block consists of a convolutional block and a 2x2 max-pooling layer for downsampling. Each decoder block consists of a 2x2 transposed convolutional layer and a convolutional block for upsampling. The convolutional block consists of two consecutive 3x3 convolutional layers, each padded with 1x1 and followed by a LeakyReLU activation function with a negative slope of 0.01. The output layer is a 1x1 convolutional layer, and the number of filters in the segmentation model is 6. (5) Model results: The segmentation Dice for the vertebral body reached 0.89, the segmentation Dice for muscles reached 0.93, the segmentation Dice for visceral fat was 0.98, the segmentation Dice for subcutaneous fat was 0.94, the segmentation Dice for bone density was 0.89, and the correlation coefficient between the segmentation area of ​​fat between muscles and manually measured Pearson reached 0.83. (All are average values, and the display effect is as follows) Figure 4 (As shown) 3) Parameter Calculation: For each body component at each level, the system calculates the following parameters: Cross-sectional area (unit: cm²): Area = total number of pixels in the component × physical area of ​​a single pixel.

[0045] Average CT value / density (unit: HU): Calculates the average CT value of all pixels within this component region.

[0046] 4) Parameter Accuracy Validation: In a validation set of 200 cases, using manual measurements by physicians on chest CT scans as the gold standard, the model was validated to have high accuracy in calculating visceral fat and intramuscular fat parameters, with correlation coefficients of 0.98 and 0.83, respectively. (Results are as follows...) Figure 5 As shown.

[0047] 3. Risk prediction and stratification module for multimodal data fusion: This module is the final step in realizing the value of this invention.

[0048] This module integrates a machine learning prediction model, whose input consists of two main categories of data: one is all the quantitative imaging parameters automatically extracted by Module 1 and Module 2, and the other is the key clinical variables obtained from the Hospital Information System (HIS).

[0049] Mechanism of operation: This model learns the complex, nonlinear relationship between imaging parameters and clinical variables and patient prognosis (such as all-cause mortality and cardiovascular mortality), thereby establishing a comprehensive prediction function.

[0050] Specific implementation steps: 1) Feature Vector Construction: Construct a feature vector for each patient, including: Imaging features: cardiac chamber volume (LV, LA, RV, RA), left ventricular myocardial mass, EAT volume and density, total Agatston score, subcutaneous fat area, skeletal muscle area, visceral fat area and density at each level of T12 / T13 / L1, etc.

[0051] Clinical features include: age, sex, systolic blood pressure, diastolic blood pressure, body mass index (BMI), history of diabetes, total cholesterol, high-density lipoprotein (HDL), low-density lipoprotein (LDL), triglycerides, and blood glucose.

[0052] 2) Model Prediction: The constructed feature vectors are input into a pre-trained prediction model. This model can be a statistical model such as the Cox proportional hazards regression model, or an artificial intelligence model such as a random forest or a deep neural network. The model outputs one or more predicted values, such as a risk score, mortality or survival probability prediction, representing the relative risk of death at a specific future time (e.g., 1 year, 3 years, 5 years). Mortality probability: A value between 0 and 1, representing the individualized absolute probability of death.

[0053] like Figure 6 The figure shows the area under the subject characteristic curve of the time-dependent mortality risk prediction model, presented in the form of subject characteristic curves at different time points. The horizontal axis of the subject characteristic curve is the prediction specificity, and the vertical axis is the prediction sensitivity of survival or death.

[0054] 3) Risk Stratification and Output: Based on the calculated risk score or probability, the system classifies patients into different risk levels such as "low risk," "intermediate risk," and "high risk" according to preset thresholds (e.g., tertiles or quartiles). Finally, the system generates a structured risk assessment report, which is presented intuitively to clinicians in the form of text, charts, etc., to guide individualized treatment and management strategies.

[0055] The specific embodiments of this disclosure have been described above. It should be understood that this disclosure is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the substantive content of this disclosure. The above-described preferred features can be used in any combination without conflict.

[0056] The following is a unified explanation of the commonly used English abbreviations and terms in this art used in the specification and drawings: CT (Computed Tomography): a type of computed tomography scan. 3D ResU-Net: A three-dimensional residual U-shaped segmentation network; TotalSegmentator: A whole-body anatomical segmentation model; 3D nnU-Net: A three-dimensional adaptive U-shaped segmentation network; 2D U-Net: A two-dimensional U-shaped segmentation network; LV: Left ventricle; LA: left atrium; RV: Right ventricle; RA: Right atrium; LVMI: Left Ventricular Quality Index; Pearson r (Pearson Correlation, Pearson correlation coefficient): Pearson correlation coefficient; Mean diff (Mean difference): The average difference; LOA (Limits of Agreement): Consistency limits; Upper LOA: Consistency upper bound; Lower LOA: Lower bound for consistency; ICC (Intraclass Correlation Coefficient): Intraclass correlation coefficient; Bland-Altman: A method for analyzing consistency between Bland and Altman; Mean difference: the average difference; CI (Confidence Interval): Confidence interval; 95% CI: 95% confidence interval; 95% CI lower: Lower bound of the 95% confidence interval; 95% CI upper: The upper bound of the 95% confidence interval; L1: First lumbar vertebra; T12: Twelfth thoracic vertebra; ROC (Receiver Operating Characteristic): Receiver Operating Characteristic curve; AUC (Area Under Curve): Area under the curve; 1-Year AUC: Area under the curve for the 1-year period; 3-Year AUC: Area under the curve over 3 years; 5-Year AUC: Area under the curve over 5 years; Sensitivity: sensitivity; Specificity: uniqueness or specificity.

Claims

1. A method for evaluating chronic cardio metabolic disease based on a chest CT image, characterized by, include: Obtain chest CT image data and clinical information from the examinee; The patient's cardiac parameters were obtained based on the chest CT image data. These cardiac parameters included the volume of each heart chamber, the mass of the left ventricular myocardium, the volume of epicardial fat, the average density of epicardial fat, and the coronary artery calcification score. The body composition data of the subject is obtained based on the chest CT image data, which includes bone data, muscle data, intramuscular fat data, visceral fat data, and subcutaneous fat data. The assessment results of the examinee are obtained based on the examinee's cardiac parameters, body composition data, and clinical information; The method for obtaining the subject's body composition data based on the chest CT image data includes: determining three standard analysis levels in the chest CT image data, namely the T12 analysis level, the L1 analysis level, and the T13 analysis level; using a body composition segmentation model to segment the bone mask, muscle mask, intermuscular fat mask, visceral fat mask, and subcutaneous fat mask on each standard analysis level; and calculating the area and average density of each type of mask region based on the segmentation results.

2. The method of claim 1, wherein, The method for obtaining the subject's cardiac parameters based on the chest CT image data includes: The chest CT image data was segmented using the TotalSegmentator image segmentation tool to obtain the left ventricular mask, right ventricular mask, left atrial mask, right atrial mask, and left ventricular myocardial mask; the volume of each heart chamber and the mass of the left ventricular myocardium were calculated based on the segmentation results. The cardiac mask is segmented using a cardiac mask segmentation model trained on the nnU-Net model. Within the cardiac mask, adipose tissue with a CT value between -190HU and -30HU is segmented using a threshold. Based on the segmented adipose tissue within the cardiac mask, the volume and average density of epicardial fat are calculated. The coronary artery mask was obtained by segmentation based on the coronary artery mask segmentation model; calcifications with a CT value greater than 130 HU were obtained by threshold segmentation; and the coronary artery calcification score was calculated based on the area of ​​the calcifications and the weighting factors corresponding to different calcification densities.

3. The method of claim 1, wherein, The method for obtaining the subject's body composition data based on the chest CT image data includes: using a body composition segmentation model trained based on the nnU-Net model.

4. The method according to claim 3, characterized in that, The method for determining three standard analytical planes in the chest CT image data includes: The chest CT image data was segmented using the TotalSegmentator image segmentation tool to obtain the spinal mask and rib mask; The lowest rib is identified as the twelfth rib based on the rib mask. The vertebra connected to the twelfth rib is identified as the twelfth thoracic vertebra in the spine mask. All layers of the twelfth thoracic vertebra are traversed along the axis, and the area of ​​the twelfth thoracic vertebra on each layer is calculated. The layer with the largest area of ​​the twelfth thoracic vertebra is selected as the T12 analysis layer. The first vertebra below the twelfth thoracic vertebra was identified as the first lumbar vertebra. All layers of the first lumbar vertebra were traversed along the axial direction. The integrity of the transverse process structure of each layer was evaluated by image morphology. The layer with the most complete transverse process structure was selected as the L1 analysis layer. The T13 analysis plane is defined as the level at which the intervertebral disc between the lower edge of the twelfth thoracic vertebra and the upper edge of the first lumbar vertebra is located, where the anterior vertebral body and the posterior spinous process are completely separated on imaging.

5. The method of claim 1, wherein, The clinical information includes at least one of the following: age, sex, systolic blood pressure, diastolic blood pressure, body mass index, history of diabetes, total cholesterol, high-density lipoprotein, low-density lipoprotein, triglycerides, and blood glucose.

6. The method of claim 1, wherein, The assessment results of the subjects include the predicted probability of the subjects' all-cause mortality risk and / or cardiovascular disease mortality risk.

7. The method of claim 6, wherein, Methods for obtaining assessment results for the subject based on the subject's cardiac parameters, body composition data, and clinical information include: A prediction model based on statistics or artificial intelligence is constructed and trained using a large amount of real data on cardiac parameters, body composition, clinical information, time of death, and cause of death of patients with chronic cardiovascular and metabolic diseases accumulated in the hospital. The trained model is then used to predict the mortality risk of patients with cardiovascular and metabolic diseases. The subject's cardiac parameters, body composition data, and clinical information are input into the trained cardiovascular metabolic disease patient mortality risk prediction model to obtain the predicted probability of the subject's all-cause mortality risk and / or cardiovascular disease mortality risk.

8. The method of claim 7, wherein, The assessment results are presented in the form of a probability curve, where the horizontal axis represents future time and the vertical axis represents the predicted probability of survival or death.

9. The method of claim 7, wherein, Different predicted probabilities are classified into different risk levels according to preset thresholds.

10. A device for assessing chronic cardiovascular and metabolic diseases based on chest CT images, characterized in that, include: The data input module is configured to acquire chest CT image data and clinical information of the examinee. The cardiac parameter calculation module is configured to obtain the patient's cardiac parameters based on the CT image data. The cardiac parameters include the volume of each cardiac chamber, the mass of the left ventricular myocardium, the volume of epicardial fat, the average density of epicardial fat, and the coronary artery calcification integral. The body composition data calculation module is configured to obtain the subject's body composition data based on the CT image data. The body composition data includes bone data, muscle data, intramuscular fat data, visceral fat data, and subcutaneous fat data. The assessment module is configured to obtain the assessment results of the subject based on the subject's cardiac parameters, body composition data, and clinical information. The body composition data calculation module is configured to perform the following steps: determine three standard analysis levels in the chest CT image data, namely the T12 analysis level, the L1 analysis level, and the T13 analysis level; and use the body composition segmentation model to segment the bone mask, muscle mask, intermuscular fat mask, visceral fat mask, and subcutaneous fat mask on each standard analysis level. Based on the segmentation results, the area and average density of each type of mask region were calculated.

Images