Aorta motion feature extraction and disease risk prediction system based on medical image analysis

Abstract

This invention belongs to the field of medical image processing technology, specifically a system for aortic motion feature extraction and disease risk prediction based on medical image analysis. The system includes modules for acquiring multi-temporal image sequences and clinical data, dynamic segmentation of aortic anatomical structures, construction of a dynamic relative coordinate system, trajectory preprocessing and standardization, trajectory segmentation analysis based on physiological phases, multidimensional dynamic feature extraction, and risk stratification prediction model construction. This invention integrates multidimensional motion features with clinical indicators to construct a risk model based on machine learning or statistical regression, outputting a predicted value for the individual's cardiovascular disease risk. This invention can reduce global interference caused by respiratory and cardiac translation, improve the biological specificity and inter-individual comparability of features, achieve fully automated processing, and enhance the risk prediction ability for cardiovascular events such as heart failure, atrial fibrillation, and aortic disease, showing promising application prospects.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a system for extracting aortic motion features and predicting disease risk based on medical image analysis. Background Technology

[0002] Aortic motion characteristics are a crucial indicator for assessing cardiovascular function. With advancements in medical imaging technology, acquiring intracardiac cycle image sequences using four-dimensional computed tomography (CT) or magnetic resonance imaging (MRI) has become possible. However, current technologies still suffer from the following significant limitations in achieving automated aortic motion feature extraction and risk prediction:

[0003] On the one hand, global motion interference is difficult to eliminate. The local motion of the aorta caused by the heartbeat is often superimposed on the chest wall translation caused by respiration and the overall displacement of the heart within the thoracic cavity. Existing analysis methods are mostly based on a global fixed coordinate system, which cannot decouple local biomechanical features from global background motion, resulting in the extracted motion features lacking biological specificity.

[0004] On the other hand, the consistency and comparability of imaging data are poor. Due to differences in anatomical structure among individuals and variations in imaging equipment parameters and patient respiratory rates during acquisition, the directly extracted raw trajectory coordinates lack a unified measurement benchmark across different individuals, making them difficult to use directly for cross-sectional comparisons of large populations and for training deep learning models.

[0005] Secondly, the utilization rate of motion characteristics is insufficient. Current clinical diagnosis still relies heavily on static indicators such as the maximum diameter or area of ​​the aorta measured manually. There is a lack of effective quantitative means and predictive models for high-dimensional phenotypic characteristics such as the dynamic evolution of the aorta in different physiological phases (e.g., early systole and late diastole) and the geometric complexity of its trajectory.

[0006] Therefore, how to establish an aortic analysis system that can eliminate global motion interference, has high inter-individual comparability, and can automatically extract multidimensional dynamic characteristics is a technical problem that urgently needs to be solved in the field of cardiovascular imaging analysis. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, the present invention aims to provide a system for aortic motion feature extraction and disease risk prediction based on medical image analysis, which features strong anti-interference performance, high early prediction sensitivity, and significant inter-individual comparability.

[0008] The aortic motion feature extraction and disease risk prediction system based on medical image analysis provided by this invention includes:

[0009] (1) Multi-phase image sequence and clinical data acquisition module, used to acquire cardiovascular image sequence containing the complete cardiac cycle and corresponding clinical indicator data of the subject to be tested, wherein the cardiovascular image sequence includes but is not limited to cine-MRI, CT or ultrasound images;

[0010] (2) The aortic anatomical structure dynamic segmentation module uses a preset segmentation model (such as a deep neural network or an adaptive threshold algorithm) to perform structural segmentation on each frame of the image sequence, obtain the ascending aortic region and the descending aortic region, and calculate the feature point coordinate sequence of the two in each time frame;

[0011] (3) Dynamic relative coordinate system construction module: The geometric center of the descending aorta in each frame is taken as the origin of the coordinate system, and the direction from the descending aorta to the ascending aorta in the initial frame is taken as the reference axis to construct a dynamic reference coordinate system;

[0012] (4) Trajectory preprocessing and standardization module, which performs smoothing and noise reduction and start point normalization on the motion trajectory to ensure that the motion trajectories of all individuals are compared under a unified starting point and metric space;

[0013] (5) The trajectory segmentation analysis module based on physiological phase calculates the velocity derivative of the motion trajectory, identifies key dynamic turning points in the cardiac cycle (such as velocity troughs and acceleration inflection points), and maps the motion trajectory of the complete cardiac cycle to multiple physiological stages such as systole and diastole (including early filling and atrial auxiliary filling); at the same time, it divides the motion trajectory of the complete cardiac cycle into multiple motion stages corresponding to the phases of the physiological stages.

[0014] (6) Multidimensional dynamic feature extraction module, which extracts multidimensional features for the overall trajectory and each physiological stage, including displacement amplitude, trajectory geometric complexity, velocity, acceleration, stage ratio and dynamic change rate of intervascular distance;

[0015] (7) Risk stratification prediction model construction module, which integrates the above-mentioned multidimensional motion characteristics with clinical indicators to construct a risk model based on machine learning or statistical regression, and outputs the predicted value of individual cardiovascular disease risk.

[0016] Furthermore:

[0017] In the module for acquiring multi-temporal image sequences and clinical data, the image sequences include two-dimensional cross-sectional images, which can also be extended to three-dimensional temporal images.

[0018] In the dynamic segmentation module of aortic anatomy, the feature point coordinate sequence includes at least one of geometric center, centroid, and anatomical landmark.

[0019] In the dynamic relative coordinate system construction module, the construction of the dynamic reference coordinate system includes:

[0020] The direction of the line connecting the feature points of the ascending aorta and the descending aorta in the initial frame is used as the reference axis direction, and the consistency of this reference axis direction is maintained in subsequent frames. The dynamic reference coordinate system is constructed by taking the feature points of the descending aorta in each frame as the origin of the coordinate system and combining the reference axis direction, thereby decoupling the motion of the ascending aorta from the global translational motion.

[0021] In the trajectory segmentation analysis module based on physiological phase, the key dynamic turning point is determined by calculating the velocity derivative characteristics of the standardized motion trajectory. The characteristics include at least one of the following: local extreme points of the velocity magnitude curve, acceleration inflection points, or curvature change points.

[0022] The exercise phases include a first exercise sequence corresponding to ventricular systole and a second exercise sequence corresponding to ventricular diastole.

[0023] In the risk stratification prediction model construction module, the risk prediction model is a Cox proportional hazards model, a logistic regression model, a random forest model, or a deep neural network model.

[0024] The risk of developing cardiovascular diseases includes: heart failure, atrial fibrillation, cardiac conduction system diseases, and the risk of developing aortic aneurysm or aortic dissection.

[0025] The input variables of the risk prediction model include not only the multidimensional dynamic features extracted by the multidimensional dynamic feature extraction module, but also basic clinical indicators, which are selected from at least one of age, gender, blood pressure, body surface area, biochemical test indicators, smoking status, or history of diabetes.

[0026] The present invention also provides an electronic device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the operation of the system as described in any one of claims 1-9 when it invokes the computer program.

[0027] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the operation of the system as described in any one of claims 1-9.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] Extremely high anti-interference and robustness: By constructing a dynamic relative coordinate system, this invention can adaptively reduce global motion interference caused by respiratory motion, overall cardiac translation and rotation, thereby improving the robustness of image analysis in complex clinical environments.

[0030] Significantly improved inter-individual comparability: Traditional absolute trajectories are greatly affected by anatomical structures and image acquisition parameters. This invention, through standardized projection transformation, makes motion characteristics biologically specific, providing a standardized quantitative tool for risk assessment across centers and large populations;

[0031] This invention enables full-cycle dynamic micro-observation: unlike static analysis that only focuses on the maximum / minimum diameter, it can precisely identify trajectory inflection points and extract motion characteristics such as acceleration and trajectory complexity in each phase of systole and diastole. This helps to capture subtle pathological changes such as early aortic compliance decline, thereby significantly improving the early predictive sensitivity for diseases such as heart failure and atrial fibrillation.

[0032] Convenience and comprehensiveness of clinical applications: It realizes full-process automation from image input to risk output, and can not only predict single aortic diseases, but also comprehensively assess multiple cardiovascular events, providing richer data dimensions for clinical decision support;

[0033] The method of this invention has good versatility and scalability, can be applied to different types of cardiovascular imaging data, and can achieve automated processing, thus having high engineering application value. Attached Figure Description

[0034] Figure 1 This is a flowchart of the aortic motion feature extraction and disease risk prediction system of the present invention.

[0035] Figure 2 This is a schematic diagram of the aortic segmentation process.

[0036] Figure 3 Flowchart for building a risk prediction model. Detailed Implementation

[0037] The technical solution of the present invention will be further described below with reference to the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or modifications made by those skilled in the art without departing from the spirit and substance of the present invention should fall within the scope of protection of the present invention.

[0038] Example 1: A system for extracting aortic motion features and predicting cardiovascular disease risk based on medical image analysis, including modules for aortic segmentation, trajectory modeling, phased analysis and feature extraction, as well as a cardiovascular disease risk prediction model constructed based on the features.

[0039] In this embodiment, the cardiovascular imaging sequence of the subject includes MRI images of the thoracic aorta, see reference. Figure 2The images comprise multiple consecutive frames from a single cardiac cycle. The target image sequence may also be derived from modalities such as computed tomography angiography or ultrasound imaging, requiring data from different phases of the cardiac cycle and sufficient temporal resolution to capture vascular motion.

[0040] In one embodiment, a thoracic aortic cineMRI acquired at the level of the pulmonary trunk is used as the input image sequence, in which the ascending and descending aortas are clearly visible.

[0041] Aortic segmentation employs deep learning-based segmentation networks (such as U-Net or Transformer architectures), and in the absence of high-quality annotations, adaptive edge constraint algorithms can be used as an alternative.

[0042] Specifically, an image processing method based on edge detection and structural constraints can be adopted, including: performing Gaussian filtering on the image sequence to reduce noise interference, and performing grayscale intensity normalization on the filtered image to obtain a preprocessed image.

[0043] Furthermore, an adaptive thresholding segmentation method based on the Otsu method is adopted to calculate the global threshold of the image and extract the average gray value of the bright areas for subsequent brightness constraint construction.

[0044] Based on the preprocessed image, edge detection is performed using the Canny operator, wherein the low threshold and high threshold are adaptively determined according to the Otsu threshold and the average gray value of the bright area.

[0045] For edge images, the Hough circle transform method is used to detect candidate circle structures within a preset radius, resulting in a set of candidate circles. Each candidate circle includes its center coordinates, radius, and cumulative voting value. By adjusting the Hough cumulative spatial threshold, the number of candidate circles is ensured to meet preset requirements. The candidate circle set is then subjected to multi-dimensional screening, including: spatial location constraints: eliminating candidate circles located near the image boundary; brightness constraints: filtering candidate circles whose average gray value inside the circle is higher than a preset brightness threshold; edge noise constraints: calculating the edge noise level of the circumferential region and eliminating candidate circles with excessive noise; gray-level uniformity constraints: calculating the gray-level standard deviation inside the circle and eliminating candidate circles with uneven gray-level uniformity; and structural consistency constraints: calculating the Dice similarity coefficient between the candidate circle region and the segmented region based on the watershed segmentation method and eliminating candidate circles with similarity below a threshold. For candidate circles that pass the initial screening, their structural consistency score, edge noise score, and gray-level variance score are calculated, and the candidate circles are ranked according to each index.

[0046] Furthermore, a predetermined number of candidate circles ranking highly among each indicator are selected, and the intersection of the ranking results for different indicators is taken as a high-confidence candidate set. Spatial constraints are applied to the high-confidence candidate set, including: overlap constraint: candidate circles with significant overlap are eliminated; minimum spacing constraint: the distance between different candidate circles is ensured to be greater than a predetermined threshold; anatomical structure constraint: candidate circles are paired and screened according to the spatial distribution characteristics of the aorta and related vessels, and circle pairs that satisfy predetermined relative positional relationships and radius relationships are preferred. Finally, the target circle structure information that satisfies the above constraints is output, obtaining the regions of the ascending and descending aorta and their corresponding center coordinates.

[0047] Example 2: A reference coordinate system is constructed based on the center coordinates of the ascending and descending aortas in each time frame. The reference coordinate system is determined based on the relative spatial relationship between the ascending and descending aortas. After obtaining the center coordinates of the ascending and descending aortas in each frame of the image sequence according to the steps described in Example 1, a coordinate system transformation is first performed.

[0048] The reference system for the original coordinates obtained from image segmentation is the image coordinate system, with the origin located at the top left corner of the image, the positive x-axis pointing from left to right, and the positive y-axis pointing from top to bottom. To obtain the motion trajectory of the ascending aorta relative to the descending aorta, this embodiment constructs a novel coordinate system:

[0049] Let the first In the frame image, the geometric center of the ascending aorta is The geometric center of the descending aorta is .

[0050] The process of constructing a dynamic relative coordinate system is as follows:

[0051] Origin translation: Using the center of the descending aorta as the reference origin, calculate the relative position of the ascending aorta. ;

[0052] Coordinate rotation: Let the first frame ( vectors Using the reference vector as a reference, the rotation matrix of each subsequent frame relative to the reference vector is calculated, thereby eliminating the influence of the overall rotation of the thoracic cavity;

[0053] Trajectory representation: finally obtaining standardized trajectory coordinates This coordinate system purely reflects the dynamic coupling motion of the ascending aorta relative to the descending aorta, effectively filtering out the overall translation of respiration and the heart.

[0054] After obtaining the standardized motion trajectory, according to a specific implementation of an embodiment of the present invention, the step of identifying multiple key turning points based on the change characteristics of the motion trajectory and dividing the complete trajectory into multiple motion stages includes:

[0055] The motion velocity of the ascending aorta center in each frame was calculated using the difference method: , Determined by image temporal resolution;

[0056] Based on the curve of motion speed changing over time, the trough time points of the speed change curve within the cardiac cycle are identified. In one specific embodiment, three trough points can be found: the first trough point is designated as inflection point 1, the second trough point as inflection point 2, and the third trough point as inflection point 3.

[0057] Phase 1 corresponds to the trajectory of motion from the starting point to the turning point 1; Phase 2 corresponds to the trajectory of motion from the turning point 1 to the turning point 2; Phase 3 corresponds to the trajectory of motion from the turning point 2 to the turning point 3; and Phase 4 corresponds to the trajectory of motion from the turning point 3 to the ending point. Among these, Phase 1 and Phase 2 correspond to the movement of the aorta during systole, while Phase 3 and Phase 4 correspond to the movement of the aorta during diastole.

[0058] Example 3: Extracting multi-dimensional features based on the overall trajectory and the trajectories of each motion stage, including:

[0059] Displacement characteristics: Overall X / Y axis displacement amplitude, calculated by taking the maximum and minimum values ​​of the X / Y coordinates over time in the motion trajectory, and defining the difference between the two as the X / Y axis displacement amplitude; Overall displacement amplitude, calculated by taking the maximum distance between the coordinate points and the starting point in the motion trajectory;

[0060] Trajectory geometric features: Overall trajectory length, calculated by the finite difference method to determine the distance between each point in the trajectory, and then summed to obtain the length of the trajectory from the start point to the end point; Overall trajectory enclosed area, calculated by connecting the start point and the end point of the trajectory and determining the area of ​​the polygon enclosed by the trajectory.

[0061] Trajectory dynamics characteristics: Overall trajectory average velocity, calculated based on the finite difference method to determine the velocity of each point on the trajectory, and then the average value of all points is calculated;

[0062] Characteristics of each movement phase: Trajectory length of movement phases 1 / 2 / 3 / 4: After determining the movement phases in the trajectory according to the method described above, the method for calculating the overall trajectory length is applied to the trajectory of each movement phase to obtain the trajectory length of each movement phase; Displacement amplitude of movement phases 1 / 2 / 3 / 4: After determining the movement phases in the trajectory according to the method described above, the method for calculating the overall X / Y axis displacement amplitude is applied to the trajectory of each movement phase to obtain the X / Y axis displacement amplitude of each movement phase; Duration of movement phases 1 / 2 / 3 / 4: After determining the movement phases in the trajectory according to the method described above, the proportion of each movement phase to the overall movement time is calculated, and then multiplied by the cardiac cycle time to obtain the duration of each movement phase; Average / peak of movement phases 1 / 2 / 3 / 4. For velocity, after determining the motion stages in the trajectory according to the method described above, the method for calculating the overall average velocity of the trajectory is applied to the trajectory of each motion stage to obtain the average velocity and peak velocity of the trajectory in each motion stage; for the average acceleration of motion stages 1 / 2 / 3 / 4, the velocity, tangential acceleration, and normal acceleration of each coordinate point in the trajectory are calculated according to the difference method, and then the average tangential acceleration and normal acceleration within each motion stage are statistically analyzed; for ratio characteristics, after obtaining the trajectory length of each motion stage, the ratio of the trajectory length of motion stage 1 to motion stage 2, the ratio of the trajectory length of motion stage 3 to motion trajectory 4, and the ratio of the length of the systolic motion trajectory (motion stage 1 + motion stage 2) to the length of the diastolic motion trajectory (motion stage 3 + motion stage 4) are calculated;

[0063] Aortic spacing characteristics: For each frame, the distance between the center of the ascending aorta and the center of the descending aorta is calculated, and the average value, variation range, and relative variation range are calculated based on this distance.

[0064] Example 4: The construction of a cardiovascular disease risk prediction model is based on population cohort data or clinical imaging databases. It requires the inclusion of disease status and mortality information from sources such as basic clinical indicator data, imaging sequences, and electronic medical records. (See [reference needed]). Figure 3 The prediction of new-onset cardiovascular disease outcomes based on a Cox regression model includes the following steps:

[0065] The baseline was set at the time of the participant's image sequence acquisition, and the event was defined as the diagnosis of a new cardiovascular disease.

[0066] The inclusion criteria for the cohort were as follows: participants had complete baseline clinical data, covering psychometric, demographic, and cardiovascular function indicators; participants did not have cardiovascular disease or a history of cardiovascular disease at baseline; participants had a clear outcome status at the end of the follow-up; and the quality of the participants' imaging sequences was sufficient to obtain the aortic motion trajectory, and the acquired aortic motion characteristics were not within the abnormal range.

[0067] The features finally included in the Cox regression model were: age, sex, body surface area, smoking status, history of diabetes, blood pressure, plasma high-density lipoprotein level, left ventricular ejection fraction, left atrial ejection fraction, ascending aortic area, ascending aortic dilatation, and aortic motion characteristics.

[0068] Using a deca-fold cross-validation, the LASSO algorithm was employed to exclude highly correlated redundant features from the aortic motion features included in the Cox model. The final selected aortic motion features, metrological and demographic features were standardized and then incorporated into the Cox regression model for model training. Finally, based on the Cox model, the absolute risk of new cardiovascular disease at a specific time point was calculated, achieving stratification and prediction of cardiovascular disease risk. After incorporating the aortic motion features of this invention, the model's C-index improved by 0.2%-1.7% compared to traditional static index models in cardiovascular diseases such as heart failure, atrial fibrillation, cardiac nerve conduction disorders, and thoracic aortic disease, with a continuous net reclassification index reaching 6.5%-40.59%.

Claims

1. An aorta motion feature extraction and disease risk prediction system based on medical image analysis, characterized in that, include: (1) A multi-temporal image sequence and clinical data acquisition module is used to acquire cardiovascular image sequences containing the complete cardiac cycle and corresponding clinical indicator data of the subject to be tested. The cardiovascular image sequences include cine-MRI, CT or ultrasound images. (2) The dynamic segmentation module of aortic anatomy structure uses a preset segmentation model to perform structural segmentation on each frame of the image sequence to obtain the ascending aortic region and the descending aortic region, and calculates the feature point coordinate sequence of the two in each time frame. (3) Dynamic relative coordinate system construction module: The geometric center of the descending aorta in each frame is taken as the origin of the coordinate system, and the direction from the descending aorta to the ascending aorta in the initial frame is taken as the reference axis to construct a dynamic reference coordinate system; (4) Trajectory preprocessing and standardization module, which performs smoothing and noise reduction and start point normalization on the motion trajectory to ensure that the motion trajectories of all individuals are compared under a unified starting point and metric space; (5) The trajectory segmentation analysis module based on physiological phase calculates the velocity derivative of the motion trajectory, identifies key dynamic turning points in the cardiac cycle, and maps the motion trajectory of the complete cardiac cycle to multiple physiological stages of systole and diastole; at the same time, it divides the motion trajectory of the complete cardiac cycle into multiple motion stages corresponding to the phases of the physiological stages. (6) Multidimensional dynamic feature extraction module, which extracts multidimensional features for the overall trajectory and each physiological stage, including displacement amplitude, trajectory geometric complexity, velocity, acceleration, stage ratio and dynamic change rate of intervascular distance; (7) Risk stratification prediction model construction module, which integrates the above-mentioned multidimensional motion characteristics with clinical indicators to construct a risk model based on machine learning or statistical regression, and outputs the predicted value of individual cardiovascular disease risk.

2. The system of claim 1, wherein, In the module for acquiring multi-temporal image sequences and clinical data, the image sequences are two-dimensional cross-sectional images and three-dimensional temporal images.

3. The system of claim 1, wherein, In the dynamic segmentation module of aortic anatomy, the feature point coordinate sequence is at least one of geometric center, centroid, and anatomical landmark.

4. The system of claim 1, wherein, In the dynamic relative coordinate system construction module, the construction of the dynamic reference coordinate system includes: taking the direction of the line connecting the ascending aorta feature point and the descending aorta feature point in the initial frame as the reference axis direction, and maintaining the consistency of the reference axis direction in subsequent frames; taking the descending aorta feature point in each frame as the origin of the coordinate system, and combining the reference axis direction to complete the construction of the dynamic reference coordinate system, thereby achieving the decoupling of the ascending aorta motion and the global translational motion.

5. The system of claim 1, wherein, In the trajectory segmentation analysis module based on physiological phase, the key dynamic turning point is determined by calculating the velocity derivative characteristics of the standardized motion trajectory. The characteristics include at least one of the following: local extreme points of the velocity magnitude curve, acceleration inflection points, or curvature change points. The exercise phases include a first exercise sequence corresponding to ventricular systole and a second exercise sequence corresponding to ventricular diastole.

6. The system of claim 1, wherein, In the risk stratification prediction model construction module, the risk prediction model is a Cox proportional hazards model, a logistic regression model, a random forest model, or a deep neural network model.

7. The system according to claim 6, characterized in that, In the risk stratification prediction model construction module, the cardiovascular disease risk includes: the risk of new-onset heart failure, atrial fibrillation, cardiac conduction system diseases, aortic aneurysm, or aortic dissection. The input variables of the risk prediction model, in addition to the multidimensional dynamic features extracted by the multidimensional dynamic feature extraction module, also include basic clinical indicators, which are selected from at least one of age, gender, blood pressure, body surface area, biochemical test indicators, smoking status, or history of diabetes.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor, when invoked, performs the operation of the system as claimed in any one of claims 1-7.

Images