Four-dimensional echocardiography dynamic segmentation system for congenital heart disease

Abstract

The present application relates to the field of medical image processing, particularly to a four-dimensional echocardiogram dynamic segmentation system for congenital heart disease, which acquires four-dimensional echocardiogram data of a fetal heart through an image acquisition module, corrects fetal movement artifacts by using optical flow field time series analysis technology in an optical flow field correction module, improves image quality, aligns heart structures and extracts interventricular septal defect regions through a multi-scale similarity transformation strategy in a heart structure time series alignment module, calculates the edge line diameter and shunt direction of the defect site in a defect analysis module, generates defect parameter information, and generates a congenital heart disease screening report in a result output module, which effectively eliminates image artifacts caused by fetal movement, lays a foundation for subsequent analysis, and accurately analyzes by accurately aligning heart structures to keep the defect site in a stable position in the time sequence.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image processing, in particular to a four-dimensional echocardiogram dynamic segmentation system for congenital heart disease, which can be applied to the automatic screening and analysis of fetal congenital heart disease in prenatal diagnosis. BACKGROUND

[0002] Congenital heart disease (congenital heart disease) is the most common congenital malformation in newborns, with an incidence of about 8‰, of which ventricular septal defect is one of the most common types of congenital heart disease. Prenatal ultrasound diagnosis is an important means of screening for congenital heart disease, and four-dimensional echocardiogram technology allows doctors to observe the spatial structure and temporal changes of the fetal heart, providing an important basis for the diagnosis of congenital heart disease.

[0003] However, the existing fetal heart ultrasound image analysis method has many challenges: first, the image artifacts caused by fetal movement seriously affect the quality of diagnosis; second, the dynamic changes of the heart structure during systole and diastole make image alignment difficult; third, the edge detection and measurement of the defect site have long relied on manual operation by doctors, which is subjective and inefficient. Especially in the early and middle stages of pregnancy, the small size and frequent movement of the fetal heart make accurate detection more difficult.

[0004] In the prior art, although some automatic segmentation algorithms have been applied to echocardiogram analysis, most of the methods are designed for adult hearts and cannot be directly applied to fetal heart analysis. At the same time, most of the existing methods fail to adequately address the artifact problem caused by fetal movement, and lack effective alignment strategies for the temporal changes of the heart structure, resulting in unstable diagnostic results. In addition, the existing technology has limited accuracy in automatic measurement of the defect site, especially for complex ventricular septal defects, making it difficult to provide accurate diameter and shunt direction information.

[0005] Therefore, it is urgent to develop an intelligent system that can effectively handle fetal movement artifacts, accurately align the heart structure, and accurately segment and measure the defect site, in order to improve the accuracy and efficiency of prenatal screening for congenital heart disease. SUMMARY

[0006] The purpose of the present application is to provide a four-dimensional echocardiogram dynamic segmentation system for congenital heart disease, which solves the technical problems of fetal movement artifact correction, heart structure temporal alignment, and accurate measurement of the defect site through innovative image processing and artificial intelligence technology, and realizes early and accurate diagnosis of congenital heart disease.

[0007] The present application proposes a four-dimensional echocardiogram dynamic segmentation system for congenital heart disease, comprising:

[0008] An image acquisition module for acquiring four-dimensional echocardiogram data of the fetal heart;

[0009] The optical flow field correction module is connected with the image acquisition module, and is configured to receive the four-dimensional echocardiogram data, correct fetal movement artifacts based on an optical flow field time series analysis technology, and generate corrected four-dimensional echocardiogram data.

[0010] The cardiac structure time series alignment module is connected with the optical flow field correction module, and is configured to receive the corrected four-dimensional echocardiogram data, align the cardiac structure through a multi-scale similarity transformation strategy, extract an interventricular septal defect region, and generate aligned cardiac structure data.

[0011] The defect analysis module is connected with the cardiac structure time series alignment module, and is configured to receive the aligned cardiac structure data, calculate a defect site edge line diameter and a shunt direction, and generate defect parameter information.

[0012] The result output module is connected with the defect analysis module, and is configured to receive the defect parameter information and generate a congenital heart disease screening report.

[0013] Preferably, the cardiac structure time series alignment module comprises:

[0014] A multi-scale pyramid construction unit is configured to construct a multi-level pyramid structure for the received four-dimensional echocardiogram data, each level containing cardiac images of different resolutions.

[0015] A feature extraction unit is connected with the multi-scale pyramid construction unit, and is configured to extract cardiac structure features in different scale spaces.

[0016] A feature fusion unit is connected with the feature extraction unit, and is configured to weight and integrate features of different levels to generate multi-scale feature descriptions.

[0017] A model matching unit is connected with the feature fusion unit, and is configured to retrieve a most similar reference model from a pre-established cardiac structure model library based on the multi-scale feature descriptions.

[0018] A non-rigid transformation unit is connected with the model matching unit, and is configured to perform non-rigid transformation based on the reference model to align the cardiac structure.

[0019] Preferably, the non-rigid transformation unit comprises:

[0020] A global transformation processor is configured to perform overall coarse alignment of the cardiac structure.

[0021] A local transformation processor is connected with the global transformation processor, and is configured to perform local fine transformation based on the global alignment.

[0022] An iterative optimization processor is connected with the local transformation processor, and is configured to adjust transformation parameters through multiple iterations until a preset alignment quality threshold is reached.

[0023] As preferred, the optical flow field correction module comprises:

[0024] an optical flow field estimation unit configured to estimate the motion field of the fetal heart by analyzing the image changes between adjacent ultrasound frames;

[0025] a motion separation unit connected to the optical flow field estimation unit, configured to separate the fetal movement from the heart movement itself;

[0026] an artifact elimination unit connected to the motion separation unit, configured to eliminate the image artifacts caused by the fetal movement based on the separated motion information.

[0027] As preferred, the defect analysis module comprises:

[0028] an edge detection unit configured to identify the interventricular septal defect edge in the aligned cardiac structure data;

[0029] a size measurement unit connected to the edge detection unit, configured to calculate the diameter of the interventricular septal defect;

[0030] a shunt direction analysis unit connected to the edge detection unit, configured to determine the shunt direction of the defect based on the morphological features and blood flow information of the interventricular septal defect;

[0031] a severity evaluation unit connected to the size measurement unit and the shunt direction analysis unit, configured to comprehensively analyze the defect parameters to evaluate the severity of the congenital heart disease.

[0032] As preferred, the model matching unit comprises:

[0033] a feature vector generator configured to convert the current ultrasound sequence into a feature vector;

[0034] a similarity calculator connected to the feature vector generator, configured to calculate the similarity between the current feature vector and the reference models stored in the model library;

[0035] an optimal matching selector connected to the similarity calculator, configured to select the reference model with the highest similarity as the matching result.

[0036] As preferred, it further comprises:

[0037] a spatio-temporal fusion network module arranged between the cardiac structure time series alignment module and the defect analysis module, configured to integrate the features in the time dimension and the spatial dimension of the aligned cardiac structure data, and enhance the feature expression of the defect region.

[0038] As preferred, the spatio-temporal fusion network module comprises:

[0039] a temporal channel for processing time series features of cardiac dynamics;

[0040] a spatial channel for processing spatial geometry features of cardiac structure;

[0041] a fusion layer connected with the temporal channel and the spatial channel for integrating time and space features to generate enhanced expression of defect region.

[0042] As preferred, the result output module comprises:

[0043] a defect visualization unit for converting the defect parameter information into intuitive graphical representation;

[0044] a report generation unit connected with the defect visualization unit for generating a congenital heart disease screening report containing diagnostic results and suggestions based on the defect parameter information and the graphical representation;

[0045] a data storage unit connected with the report generation unit for saving the congenital heart disease screening report and related analysis data.

[0046] As preferred, it further comprises:

[0047] an adaptive model updating module connected with the cardiac structure temporal alignment module for updating reference models in the cardiac structure model library based on new analysis cases to improve the adaptability of the system to different fetal physiological characteristics.

[0048] The beneficial effects of the present application include:

[0049] 1. By optical flow field temporal correction technology, image artifacts caused by fetal movement are effectively eliminated, image quality is improved, and a foundation is laid for subsequent analysis;

[0050] 2. The use of multi-scale similarity transformation strategy realizes accurate alignment of cardiac structure, so that the defect position remains stable in the time series, facilitating accurate analysis;

[0051] 3. Combined with spatio-temporal fusion network technology, the feature expression of the defect region is enhanced, and the segmentation accuracy is improved;

[0052] 4. The edge line diameter and shunt direction of the defect position are automatically calculated, the detection efficiency is improved by 15 times compared with manual labeling, and the congenital heart disease screening as early as 16 weeks of pregnancy is supported;

[0053] 5. The system performance is continuously optimized through the adaptive model library, and the adaptability to different fetal physiological characteristics is improved.

[0054] In summary, the present application significantly improves the accuracy, efficiency and early screening ability of prenatal diagnosis of congenital heart disease, and has important clinical application value. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 The whole structure block diagram of the congenital heart disease four-dimensional echocardiogram dynamic segmentation system of the application is shown in the figure;

[0056] Figure 2 The structure block diagram of the heart structure time sequence alignment module of the application is shown in the figure;

[0057] Figure 3 The structure block diagram of the optical flow field correction module of the application is shown in the figure;

[0058] Figure 4 The structure block diagram of the defect analysis module of the application is shown in the figure;

[0059] Figure 5 The structure block diagram of the space-time fusion network module of the application is shown in the figure;

[0060] Figure 6 The multi-scale pyramid structure schematic diagram of the application is shown in the figure;

[0061] Figure 7 The non-rigid transformation optimization process schematic diagram of the application is shown in the figure;

[0062] Figure 8 The ventricular septal defect measurement result example diagram of the application is shown in the figure. DETAILED DESCRIPTION

[0063] Please refer to the accompanying Figures 1-8 The specific embodiments of the application will be further described in detail below in combination with the accompanying drawings.

[0064] Referring to Figure 1 The congenital heart disease four-dimensional echocardiogram dynamic segmentation system provided by the application comprises an image acquisition module 1, an optical flow field correction module 2, a heart structure time sequence alignment module 3, a defect analysis module 4, a result output module 5, an adaptive model updating module 6 and a space-time fusion network module 7.

[0065] The image acquisition module 1 is used for acquiring four-dimensional echocardiogram data of a fetal heart. In an embodiment of the application, the four-dimensional echocardiogram data acquired by the module generally has an acquisition frequency of 25 frames per second, a spatial resolution of 224x224x96 voxels, and can clearly capture the three-dimensional structure of the fetal heart and the four-dimensional information changing with time.

[0066] The optical flow field correction module 2 is connected with the image acquisition module 1 and is used for receiving the above-mentioned four-dimensional echocardiogram data, correcting fetal movement artifacts based on optical flow field time sequence analysis technology, and generating corrected four-dimensional echocardiogram data. The module solves the problem of interference of fetal movement on the quality of the ultrasound image, and lays a foundation for subsequent accurate analysis.

[0067] The cardiac structure time series alignment module 3 is connected with the optical flow field correction module 2, and is configured to receive the corrected four-dimensional echocardiogram data, align the cardiac structure through a multi-scale similarity transformation strategy, extract an interventricular septal defect region, and generate aligned cardiac structure data. The module is a core innovative part of the system, and solves the problem of structure alignment difficulty caused by dynamic changes of the heart.

[0068] The defect analysis module 4 is connected with the cardiac structure time series alignment module 3, and is configured to receive the aligned cardiac structure data, calculate a defect site edge line diameter and a shunt direction, and generate defect parameter information. The module realizes automatic and accurate measurement of the interventricular septal defect.

[0069] The result output module 5 is connected with the defect analysis module 4, and is configured to receive the defect parameter information and generate a congenital heart disease screening report. The module converts the analysis result into diagnostic information that can be understood by a clinician.

[0070] Preferably, the system further comprises a space-time fusion network module 7 arranged between the cardiac structure time series alignment module 3 and the defect analysis module 4, and configured to integrate features in the time dimension and the spatial dimension for the aligned cardiac structure data, and enhance feature expression of the defect region.

[0071] In addition, the system further comprises an adaptive model updating module 6 connected with the cardiac structure time series alignment module 3, and configured to update a reference model in a cardiac structure model library based on a new analysis case, and improve adaptability of the system to different fetal physiological characteristics.

[0072] The specific implementation of each module is described in detail below.

[0073] Referring to Figure 3 The optical flow field correction module 2 comprises an optical flow field estimation unit 21, a motion separation unit 22, and an artifact elimination unit 23.

[0074] The optical flow field estimation unit 21 is configured to estimate a motion field of the fetal heart by analyzing image changes between adjacent ultrasound frames. Specifically, the unit calculates a motion vector of each pixel point in the ultrasound image sequence based on the optical flow principle.

[0075] In the implementation process, the optical flow field estimation adopts an improved Lucas-Kanade algorithm, and the calculation formula is as follows:

[0076] ,

[0077] wherein: is an image space gradient, which is a three-dimensional vector representing a brightness change rate of the image in the x, y, and z directions; is a velocity vector field to be solved, which is a three-dimensional vector , represents the displacement rate of the pixel in the x, y, z direction; is the image time derivative, representing the rate of change of image brightness with time.

[0078] In fetal heart ultrasound image processing, the above optical flow equation is based on the brightness constant assumption, that is, assuming that the same tissue keeps the brightness unchanged in a short time, only the position changes. This assumption is applicable to most heart tissues in the ultrasound sequence, but special processing is required in areas with large brightness changes such as ventricular septal defect areas.

[0079] In a specific implementation, for each voxel point , the velocity vector is obtained by solving the following equation set :

[0080] ,

[0081] wherein: , , respectively, the partial derivative of the image in the x, y, z three spatial directions, representing the brightness change rate of the image in each direction; is the image time derivative, representing the rate of change of image brightness with time. , , respectively, the x, y, z direction velocity components to be solved, with the unit of pixel / frame.

[0082] To improve the stability of the calculation, the present application adopts a 5x5x5 Gaussian kernel for spatial filtering, and the standard deviation σ is set to 1.5. In actual application, the selection of the Gaussian kernel size and the standard deviation is based on the size and motion characteristics of the fetal heart. For a typical mid-pregnancy (20-28 weeks) fetal heart, a kernel size of 5x5x5 can balance the needs of local detail preservation and noise suppression, and σ=1.5 ensures appropriate smoothing effect.

[0083] The motion separation unit 22 is connected with the optical flow field estimation unit 21, and is used for separating fetal movement and heart motion. This unit distinguishes the motion from different sources based on frequency characteristics, wherein fetal movement usually presents as low-frequency overall displacement, and heart motion presents as periodic local deformation.

[0084] In a specific implementation, a frequency domain filtering method is adopted, and a band-pass filter is designed to extract the heart periodic motion:

[0085] ,

[0086] wherein: is the signal frequency, with the unit of hertz (Hz); Low cutoff frequency, set to 0.5 Hz, for filtering out breathing and other low frequency motion; High pass region start frequency, set to 1.0 Hz; High pass region cutoff frequency, set to 2.5 Hz; Half of the signal sampling frequency, i.e. the Nyquist frequency, for a sampling rate of 25 frames / s, Hz; Frequency response function, value range [0, 1], representing the retention ratio of the corresponding frequency component.

[0087] These parameter settings are based on the actual situation that the fetal heart rate is usually 110-160 times per minute (about 1.8-2.7 Hz). In the application of different gestational weeks, these parameters can be dynamically adjusted according to the change of fetal heart rate to obtain the best separation effect. For example, when the fetal heart rate is high in early pregnancy (16-20 weeks), the and can be increased to 1.2 Hz and 2.8 Hz.

[0088] The artifact elimination unit 23 is connected with the motion separation unit 22, and is used for eliminating image artifacts caused by fetal movement based on the separated motion information. The unit applies inverse transformation to compensate the inter-frame fetal movement back to realize image stabilization.

[0089] The core of the compensation algorithm is:

[0090] ,

[0091] Wherein: is the corrected image; is the original image; is the estimated fetal movement speed vector, unit: pixel / frame; is the inter-frame time interval, unit: second, for a sampling rate of 25 frames / s, second; is the voxel coordinate in the image; is the time point, unit: second.

[0092] In actual fetal heart ultrasound image processing, the displacement calculated by the fetal movement speed vector is usually a non-integer pixel, so pixel resampling is needed. The present application adopts bicubic interpolation algorithm for resampling to ensure image quality. The bicubic interpolation algorithm can effectively suppress artifact generation while maintaining image details, and is particularly suitable for fetal heart ultrasound and other medical images with rich details.

[0093] For the detection of ventricular septal defect, artifact elimination is particularly important. Clinical verification shows that after processing by the optical flow field correction module of the present application, the image artifacts caused by fetal movement can be reduced by more than 85%, especially in the middle and late pregnancy (24-34 weeks) when fetal movement is frequent. For example, in the test of 30 fetuses in the middle and late pregnancy, the signal-to-noise ratio of the ventricular septum region before correction is 15.3 dB on average, and after correction, it is increased to 23.7 dB, providing a high-quality image basis for subsequent defect detection.

[0094] Referring to Figure 2 , the heart structure time sequence alignment module 3 is the core innovative part of the system, including a multi-scale pyramid construction unit 31, a feature extraction unit 32, a feature fusion unit 33, a model matching unit 34, and a non-rigid transformation unit 35.

[0095] The multi-scale pyramid construction unit 31 is used to construct a multi-level pyramid structure for the received four-dimensional echocardiogram data, each level containing heart images of different resolutions. Referring to Figure 6 , this unit constructs an image pyramid usually containing 5 layers through recursive downsampling, realizing multi-scale expression of the heart structure from macro to micro.

[0096] In an embodiment of the present application, the pyramid construction adopts the following recursive formula:

[0097] ,

[0098] wherein: represents the pyramid image of the i-th layer, the value range of i is [1, 5], the original input image; is a downsampling function, representing the operation of reducing the resolution of the input image; is a downsampling coefficient, usually set to 0.5, representing that the linear resolution of each layer is half of the previous layer.

[0099] In fetal heart ultrasound image processing, the original image is usually three-dimensional data of 224x224x96 voxels. After 5-layer pyramid construction, the sizes of the images of each layer are: , , , and . This multi-scale representation can capture both the overall morphology of the heart (low-resolution layer) and the detailed features (high-resolution layer), and is particularly suitable for the detection of ventricular septal defects of different sizes.

[0100] ​The feature extraction unit 32 is connected with the multi-scale pyramid construction unit 31, and is used for extracting features of a cardiac structure in different scale spaces. The unit extracts key features such as a ventricular contour, an interventricular septum and a myocardial boundary at each pyramid level.

[0101] In a preferred embodiment of the application, the feature extraction adopts an improved SIFT (Scale Invariant Feature Transform) algorithm which is optimized for the characteristics of three-dimensional ultrasound images. For each voxel point , a local feature descriptor is calculated:

[0102] ,

[0103] Wherein: is a feature descriptor at the voxel point , and is a dimensional vector; represents a 3D gradient histogram centered at the point, the value range of which is [1, n], and is usually 128; each represents the cumulative intensity of the gradient in a specific direction, and is a dimensionless normalized value.

[0104] In fetal cardiac ultrasound image processing, the feature extraction process usually divides the space into sub-blocks, and an 8-direction gradient histogram is calculated for each sub-block to form a 128-dimensional feature vector. In order to adapt to the noise characteristics of the ultrasound image, an anisotropic diffusion filter is used for pretreatment before gradient calculation, and the parameter setting is a diffusion coefficient K=25 and an iteration number of 10.

[0105] These parameter settings are determined based on the characteristics of fetal cardiac ultrasound images. The diffusion coefficient K=25 is suitable for retaining the edges of key structures such as the interventricular septum while effectively suppressing speckle noise; and the 10 iterations strike a balance between image smoothing and calculation efficiency. In actual application, for ultrasound images of different qualities, the parameters can be fine-tuned within the range of K=20-30 to obtain the best effect.

[0106] The feature fusion unit 33 is connected with the feature extraction unit 32, and is used for weighting and integrating features of different levels to generate multi-scale feature descriptions. The unit realizes effective fusion of multi-scale information and improves the robustness of feature expression.

[0107] The feature fusion adopts an adaptive weight scheme:

[0108] ,

[0109] Wherein: is the fused feature, and is a vector with the same dimension as . the feature extracted for the i-th layer, , where n is the number of pyramid layers minus 1, usually 4. is the corresponding weight factor, indicating the importance of the i-th layer feature, satisfying denotes the vector weighted sum operation.

[0110] In fetal heart ultrasound image processing, the weight is dynamically determined by a feature quality evaluation function:

[0111] ,

[0112] where: is the quality score of the i-th layer feature, a non-negative real number; denotes the sum of all layer feature quality scores. The feature quality score is calculated based on gradient magnitude, contrast, and entropy, etc.

[0113] In ventricular septal defect detection, a typical weight distribution is: (original layer), (first layer), = 0.20 (second layer), (third layer), (fourth layer). This distribution emphasizes the edge details of the original high-resolution layer, while taking into account the overall structure information of the low-resolution layer, and is particularly suitable for ventricular septal defects of different sizes.

[0114] The model matching unit 34 is connected to the feature fusion unit 33 and is used to retrieve the most similar reference model from a pre-established heart structure model library based on multi-scale feature description. This unit realizes intelligent matching of the current ultrasound sequence with standard cases in the model library.

[0115] The model matching unit 34 includes a feature vector generator 341, a similarity calculator 342, and an optimal matching selector 343.

[0116] The feature vector generator 341 is used to convert the current ultrasound sequence into a feature vector. In fetal heart ultrasound image processing, the feature vector usually contains heart anatomical structure features (such as ventricular septal position, thickness, etc.), texture features, and dynamic features, etc. to form a high-dimensional vector that comprehensively represents the current case.

[0117] The similarity calculator 342 is connected to the feature vector generator 341 and is used to calculate the similarity between the current feature vector and the reference model stored in the model library. In the present invention, the similarity calculation uses the cosine distance:

[0118] ,

[0119] wherein: is the similarity value, ranging from [-1, 1], the larger the value, the more similar; is the feature vector of the current ultrasound sequence; is the feature vector of the m-th model in the model library; denotes the dot product of two vectors; and denote the Euclidean norm (i.e., vector length) of two vectors, respectively.

[0120] In the fetal heart ultrasound image, when the similarity value is greater than 0.9, it is generally considered that the current sequence has high similarity with the reference model, and is suitable for the initial parameter setting of subsequent non-rigid transformation.

[0121] The optimal matching selector 343 is connected with the similarity calculator 342, for selecting the reference model with the highest similarity as the matching result. In the actual system, in order to improve the reliability of the matching, the clinical factors such as gestational age and fetal position are usually considered for auxiliary screening.

[0122] In an embodiment of the present application, the model library contains 1000 groups of standard images, which are classified and indexed according to gestational age (16-40 weeks), maternal characteristics (such as BMI index), fetal position (head position, breech position, transverse position) and heart structure type (normal, ventricular septal defect, other congenital heart disease). This structured organization significantly improves the retrieval efficiency, and usually can complete the optimal matching search within 0.5 seconds.

[0123] The non-rigid transformation unit 35 is connected with the model matching unit 34, for performing non-rigid transformation based on the reference model, and aligning the heart structure. This is the core step of the entire alignment process, which realizes the accurate modeling of the complex deformation of the heart structure.

[0124] The non-rigid transformation unit 35 includes a global transformation processor 351, a local transformation processor 352 and an iterative optimization processor 353.

[0125] The global transformation processor 351 is used to perform the overall rough alignment of the heart structure. This processor implements rigid transformation, including translation, rotation and scaling, and the basic transformation matrix is:

[0126] ,

[0127] wherein: is the global transformation matrix, with a size of 3x3 (two-dimensional case) or 4x4 (three-dimensional case); is the scaling factor, representing the proportion of overall size adjustment, usually in the range of 0.8-1.2; is the rotation angle in radian, representing the rotation amount around the center point; is the translation vector in pixel, representing the translation amount in x and y direction. For 3D data, the transformation matrix is extended to 4x4, adding the transformation parameter in z direction.

[0128] In fetal heart ultrasound image processing, the global transformation parameters are usually initialized based on the matched reference model. For example, for a 24-week pregnant fetus with head position, the typical transformation parameters can be: = 1.05 (slightly enlarged), = 0.1 radian (about 5.7 degrees of rotation), = 5 pixels, = 3 pixels. These parameters vary with gestational age and fetal position.

[0129] The local transformation processor 352 is connected with the global transformation processor 351, and is used to perform local fine transformation on the basis of global alignment. The processor uses a free form deformation (FFD) model to realize local deformation based on B-spline function.

[0130] The specific formula is:

[0131] ,

[0132] Wherein: is the local transformation function, representing the new position of point after transformation; , , is the cubic B-spline basis function, , , The value range of is [0, 3]; is the control point coordinate in the control point grid, wherein , , is the control point grid spacing; is the normalized coordinate, and the calculation method is , , .

[0133] The cubic B-spline basis function is defined as:

[0134] ,

[0135] ,

[0136] ,

[0137] ,

[0138] wherein: is the normalized coordinate, whose value range is [0, 1]; are four basis functions, used for interpolation calculation. In fetal heart ultrasound image processing, the control point grid spacing is usually set to 8-16 voxels. For example, for a 224 image, the control point grid spacing can be set to , generating about 14x14x12 control points. In the high gradient area of the interventricular septum, the control point can be dynamically encrypted to 8x8x4 spacing to obtain more detailed deformation control.

[0139] The iterative optimization processor 353 is connected with the local transformation processor 352, used for adjusting the transformation parameters through multiple iterations until the preset alignment quality threshold is reached. The processor realizes the automatic optimization process of the transformation parameters.

[0140] The optimization objective function is:

[0141] ,

[0142] wherein: is the total objective function, which is a non-negative real number, and the smaller the value, the higher the alignment quality; is the image similarity measure, usually using mutual information or normalized cross-correlation; is the deformation field smoothness constraint, used to prevent over-deformation; and are weighting factors, used to balance the importance of similarity and smoothness, usually set to , whose value range is 0.01-0.1.

[0143] In fetal heart ultrasound image processing, the calculation formula of mutual information (MI) is:

[0144] ,

[0145] wherein: is the mutual information value, in bits, and the larger the value, the higher the matching degree of the two images; and are two images to be aligned; and are the entropies of images and respectively, is the joint entropy of and .

[0146] The calculation formula of the deformation field smoothness constraint is:

[0147] ,

[0148] wherein: is a smoothness constraint term, the smaller the value is, the smoother the deformation field is; is an image domain; is a Laplacian operator, representing a second-order derivative; is a transformation function; represents a Euclidean norm.

[0149] The iterative optimization adopts a gradient descent method, the learning rate is set to 0.01, the maximum number of iterations is 100, or the iteration is terminated in advance when the objective function changes less than 0.001 for two consecutive iterations. In practical applications, 15-30 iterations can achieve convergence.

[0150] After the heart structure time sequence alignment module 3 processing, the heart structures on different time frames are accurately aligned, especially the interventricular septum defect area maintains a stable position in the entire sequence, laying a foundation for subsequent accurate analysis. Clinical verification shows that this module improves the alignment accuracy of the heart structure from the original pixel level (about 1-2 mm) to the sub-pixel level (about 0.3-0.5 mm), significantly improving the accuracy of subsequent defect analysis.

[0151] For example, in the test of 40 cases of different gestational age (18-32 weeks) interventricular septum defect cases, the overlap coefficient (Dice coefficient) of the interventricular septum area before and after alignment is improved from an average of 0.68 to 0.92, especially for cases with obvious fetal movement, the improvement effect is more significant.

[0152] Referring to Figure 4 , the defect analysis module 4 includes an edge detection unit 41, a size measurement unit 42, a shunt direction analysis unit 43, and a severity evaluation unit 44.

[0153] The edge detection unit 41 is used to identify the interventricular septum defect edge in the aligned heart structure data. This unit accurately segments the interventricular septum defect area to generate a defect edge contour.

[0154] In the preferred embodiment of the present application, the edge detection adopts an improved level set method, and its evolution equation is:

[0155] ,

[0156] wherein: is a level set function, representing an implicit surface, is the edge position; is an evolution time, representing the number of iteration steps; is the partial derivative of the level set function with respect to time, representing the surface evolution rate; is the Dirac function, used to restrict evolution only near the edges; is the image gradient based edge stopping function; is the curvature term, used to keep the surface smooth; is the gradient of the level set function; is the gradient of the edge stopping function; and is the control parameter, usually set to .

[0157] Dirac function is defined as:

[0158] ,

[0159] where: ϵ is the bandwidth parameter, controlling the range of action, usually set to 1.5 pixels.

[0160] Edge stopping function is defined as:

[0161] ,

[0162] where: is the Gaussian kernel with standard deviation σ, used to smooth the image; * denotes the convolution operation; I is the input image; ∇ is the gradient operator; denotes the gradient magnitude of the smoothed image; g is the edge stopping function, which tends to 0 at the edges (large gradient) and to 1 in the smooth regions (small gradient).

[0163] In fetal heart ultrasound image processing, σ is usually set to 1.0, which is based on the characteristics of ultrasound images and the typical size of ventricular septal defects. The initial contour of the level set method is usually set based on the candidate region in the previous step of heart structure alignment. This "warm start" strategy significantly improves the efficiency and accuracy of segmentation.

[0164] The size measurement unit 42 is connected to the edge detection unit 41 for calculating the diameter of the ventricular septal defect. Based on the defect edge contour, the unit calculates the maximum diameter and effective diameter of the defect.

[0165] In practical applications, since the ventricular septal defect usually presents an irregular shape, the system calculates the diameter in multiple directions and takes the maximum value as the maximum diameter:

[0166] ,

[0167] where: is the maximum diameter, in millimeters; is the set of defect contour points; and are any two points on the contour. represents the Euclidean distance between two points; represents the max operation. For three-dimensional data, point point and are three-dimensional coordinates .

[0168] To more comprehensively describe the size of the defect, the system also calculates the equivalent circular diameter as the effective diameter:

[0169] ,

[0170] where: is the effective diameter in millimeters; is the area of the defect region in square millimeters; is the constant pi, approximately equal to 3.14159. For three-dimensional data, the equivalent spherical diameter of the defect is calculated: , where is the volume of the defect in cubic millimeters.

[0171] In clinical practice, the size of the interventricular septal defect is an important indicator for assessing the severity. Typically, is considered a small defect, is a moderate defect, is a large defect. These thresholds are determined based on a large amount of clinical data statistics and expert consensus, and are suitable for fetal heart assessment of fetuses of different gestational ages.

[0172] The shunt direction analysis unit 43 is connected to the edge detection unit 41 and is used to determine the shunt direction of the defect based on the morphological features and blood flow information of the interventricular septal defect. This unit identifies the direction of blood flow from the left ventricle to the right ventricle (left-to-right shunt) or from the right ventricle to the left ventricle (right-to-left shunt).

[0173] In one embodiment of the present application, the shunt direction analysis combines morphological features and Doppler data. Based on morphological analysis, the system calculates the estimated pressure gradient on both sides of the defect area:

[0174] ,

[0175] where: is the estimated pressure gradient in millimeters of mercury (mmHg); is the proportionality coefficient, with an empirical value of 0.05, which is calibrated based on fluid dynamics models and clinical data; is the effective diameter in millimeters; and are the estimated radii of the left and right ventricles, respectively, in millimeters.

[0176] The sign of the pressure gradient indicates the shunt direction: when it is judged as left-to-right shunt (normal physiological state); when it is judged as right-to-left shunt (pathological state, usually seen in severe congenital heart disease); when it is judged as bidirectional shunt.

[0177] In practical applications, the system will further verify in combination with Doppler color flow data. For the standard four-chamber heart section, left-to-right shunt is usually displayed as a red or blue jet from the left ventricle to the right ventricle (depending on the Doppler angle setting) in the color Doppler mode.

[0178] The severity evaluation unit 44 is connected with the size measurement unit 42 and the shunt direction analysis unit 43, and is used for comprehensive analysis of the defect parameters to evaluate the severity of congenital heart disease. Based on factors such as defect diameter, shunt direction and location, the unit grades the severity of congenital heart disease.

[0179] In clinical applications, the system uses the following grading standards:

[0180] Mild: defect diameter < 3mm, left-to-right shunt, located in the membranous part

[0181] Moderate: defect diameter 3-6mm, left-to-right shunt, or small defect located in the muscular part

[0182] Severe: defect diameter > 6mm, or any right-to-left shunt, or defect located in the outlet part

[0183] This grading standard is based on international cardiology guidelines and a large amount of clinical data statistics, and is suitable for the evaluation of congenital heart disease of fetuses in different gestational age stages. In practical applications, the system will also consider factors such as gestational age for adjustment. For example, for early gestational period (18-24 weeks) fetuses, the defect diameter threshold will be adjusted proportionally according to the size of the heart.

[0184] At the same time, the system also considers the impact of defect location on prognosis. In interventricular septal defect, it can be divided into membranous defect, muscular defect and outlet defect according to the anatomical location. Among them, the membranous defect usually has a good prognosis, while the muscular defect depends on the specific location and size, and the outlet defect often coexists with other heart abnormalities and has a poor prognosis. The system automatically judges the defect type by analyzing the positional relationship of the defect relative to the aortic valve and tricuspid valve.

[0185] After processing by the defect analysis module 4, the system can automatically provide key diagnostic parameters of interventricular septal defect, including diameter, shunt direction and severity evaluation, providing strong support for clinical decision-making. In the clinical verification of 50 cases of interventricular septal defect, the consistency (Kappa coefficient) of the grading results of the system and three experienced ultrasound doctors reached 0.87, showing good clinical practicability.

[0186] The spatio-temporal fusion network module 7 of the present application is arranged between the cardiac structure time alignment module 3 and the defect analysis module 4, and is used for integrating the features in the time dimension and the space dimension of the aligned cardiac structure data, so as to enhance the feature expression of the defect region.

[0187] As shown in FIG. 7, the spatio-temporal fusion network module 7 includes a time sequence channel 71, a space channel 72, and a fusion layer 73. Figure 5

[0188] The time sequence channel 71 is used for processing the time sequence features of the dynamic changes of the heart. The channel adopts a long short-term memory (LSTM) structure, and can capture the time sequence dependency of the heart movement.

[0189] In the processing of fetal heart ultrasound images, the number of hidden units of the LSTM unit is usually set to 128, which is determined based on the input data dimension and the balance of computing resources. The sequence length is set to the number of cardiac cycle frames, usually 20-30 frames (corresponding to one cardiac cycle under a sampling rate of 25 frames / second).

[0190] The LSTM network can effectively capture the dynamic change features of the interventricular septum in the systolic and diastolic processes, which is crucial for distinguishing the real interventricular septal defect from the false appearance similar to the defect (such as the normal anatomical structure under certain angles). For example, the real interventricular septal defect exists throughout the cardiac cycle, while some false appearances may only appear at a specific cardiac cycle stage.

[0191] The space channel 72 is used for processing the spatial geometric features of the cardiac structure. The channel adopts a multi-layer coding and decoding structure, and is implemented based on an improved U-Net architecture to extract spatial features.

[0192] The encoding part of the U-Net includes 4 consecutive down-sampling blocks, each block including:

[0193] Two 3×3×3 three-dimensional convolution layers, each layer followed by batch normalization and ReLU activation function

[0194] A 2×2×2 max-pooling layer for reducing the spatial resolution

[0195] The specific parameter settings of the encoding part of the U-Net are as follows:

[0196] Input layer: 224×224×96×1 (size×channel number);

[0197] First down-sampling block: 112×112×48×64;

[0198] Second down-sampling block: 56×56×24×128;

[0199] ​Third down-sampling block: 28x28x12x256;

[0200] Fourth down-sampling block: 14x14x6x512;

[0201] The decoding part corresponds to 4 up-sampling blocks, each block contains:

[0202] A 2x2x2 transpose convolution up-sampling layer to increase the spatial resolution;

[0203] Feature connection (skip connection) with the corresponding layer of the encoding part to fuse low-level and high-level features

[0204] Two 3x3x3 three-dimensional convolution layers, each followed by batch normalization and ReLU activation function;

[0205] In addition, a residual connection is added after each encoding block to improve feature transmission efficiency and alleviate the problem of gradient disappearance. The implementation formula of the residual connection is:

[0206] ,

[0207] Where: is the output feature of the block; is the output of the regular processing path (convolution, activation, etc.); is the input feature of the block.

[0208] In fetal heart ultrasound image processing, each convolution layer of the U-Net architecture uses a small convolution kernel of 3x3x3, which can reduce the number of parameters while maintaining the receptive field, suitable for processing ultrasound images with complex textures. At the same time, the design of the skip connection ensures the transmission of high-resolution features, which is particularly important for detecting small defects.

[0209] The fusion layer 73 is connected with the time channel 71 and the space channel 72, used to integrate time and space features, and generate enhanced defect area expression. This layer realizes the effective fusion of time and space information, and improves the recognition degree of defect area features.

[0210] The fusion layer uses attention mechanism to realize adaptive fusion of time and space features:

[0211] ,

[0212] Where: is the fused feature, with the same dimension as the input feature; is the time sequence feature, coming from the output of the time channel; is the spatial feature, coming from the output of the space channel; and are learning weights, which are trainable parameter matrices; Bias parameters, which are also trainable parameter vectors; Sigmoid function, which maps the input to the interval [0, 1]; Represents an element-wise multiplication operation.

[0213] This attention mechanism design allows the network to adaptively adjust the weights of the temporal and spatial features according to the characteristics of different regions. For example, in the region of ventricular septal defect, the temporal feature usually obtains a higher weight due to its significant characteristic change in the cardiac cycle; while in the region of myocardium, which is relatively stable, the spatial feature obtains a higher weight.

[0214] The spatio-temporal fusion network module 7 uses data augmentation techniques in the training phase, including random rotation (±15°), scaling (0.9-1.1 times) and elastic deformation, to improve the model's generalization ability. The optimization uses the Adam algorithm, with an initial learning rate of 0.0001, decaying to 0.5 of the original every 50 epochs, a batch size of 4, and training for 200 epochs or stopping early when the validation loss does not decrease for 20 consecutive epochs.

[0215] The learning rate is set based on extensive experimental testing, with an initial value of 0.0001 to ensure stable convergence while providing sufficient learning speed. Decaying the learning rate to 0.5 of the original every 50 epochs is a common learning rate scheduling strategy that helps the model converge to a local optimal solution more finely in the later stages of training. The batch size is set to 4, taking into account the complexity of the network structure and the limitations of GPU memory, which can effectively run on most medical-grade GPU workstations.

[0216] After processing by the spatio-temporal fusion network module 7, the feature representation of the defect region is significantly enhanced, especially for small size defects and fuzzy boundaries, the detection ability is obviously improved. Clinical verification shows that this module improves the Dice coefficient of defect region segmentation from 0.77 to 0.89, especially for small defect detection in early fetal heart of 16-22 weeks of gestation.

[0217] For example, in the test of 35 cases of ventricular septal defect in early pregnancy (16-22 weeks), without using the spatio-temporal fusion network, the system's detection rate for small defects with a diameter <3mm was 76.5%; after adding the spatio-temporal fusion network, the detection rate increased to 92.3%. This has important clinical value for early prenatal diagnosis, and can provide more time window for clinical intervention and family decision-making.

[0218] The result output module 5 includes a defect visualization unit 51, a report generation unit 52 and a data storage unit 53.

[0219] The defect visualization unit 51 is used to transform the defect parameter information into intuitive graphical representation. This unit generates visualization images including defect region segmentation, diameter annotation and shunt direction marking, which facilitates intuitive understanding by doctors.

[0220] In one embodiment of the present application, the visualization displays the defect region using pseudo-color coding, using different colors to represent different analysis results: red for defect region, green for normal interventricular septum, and blue arrow for shunt direction. Meanwhile, the system provides interactive display interfaces of multiple views (sagittal, coronal and transverse), and doctors can rotate and zoom to view the results from different angles.

[0221] Specifically, the system generates the following standard visualization views:

[0222] Four-chamber heart section: shows the position and size of the defect in the four chambers of the heart and the interventricular septum;

[0223] Left ventricular outflow tract section: shows the relationship between the defect and the aorta, especially suitable for observing outlet defects;

[0224] Short-axis section: shows the transverse morphology of the defect, suitable for observing muscular defects;

[0225] Three-dimensional reconstruction view: shows the three-dimensional morphology and relative position relationship of the defect;

[0226] Time sequence dynamic display: shows the dynamic changes of the defect in the cardiac cycle;

[0227] The report generation unit 52 is connected with the defect visualization unit 51, and is used to generate a congenital heart disease screening report containing diagnosis results and suggestions based on defect parameter information and graphical representation. This unit converts quantitative analysis results into standardized diagnostic reports.

[0228] The standard report includes the following contents:

[0229] Basic information: examination date, gestational age, basic information of the fetus (such as gestational age, fetal position, estimated weight, etc.);

[0230] Defect parameters: location (membrane, muscle or outlet), maximum diameter, effective diameter, shunt direction;

[0231] Severity assessment: mild, moderate or severe, with assessment basis;

[0232] Diagnosis suggestions: follow-up interval, precautions, prenatal counseling suggestions, etc.;

[0233] Visualization images: defect display images of key sections, usually including static images and dynamic sequence links;

[0234] The report adopts a structured format, allowing doctors to quickly obtain key information while retaining detailed quantitative data for professional analysis. The system automatically adjusts the format based on different hospital report standards to meet the needs of different clinical environments.

[0235] The data storage unit 53 is connected to the report generation unit 52 and is used to save the congenital heart disease screening report and related analysis data. This unit realizes the functions of long-term data storage and retrospective analysis.

[0236] The system supports multiple data storage methods, including local databases and cloud storage, and provides permission control and encryption protection measures. Patient data is stored in DICOM format and custom format, with raw ultrasound data and analysis results saved separately to optimize storage space and access efficiency.

[0237] At the same time, the system provides historical data comparison and analysis functions, which can track the changes in the same fetus's examination results at different gestational weeks, and assist in assessing the progress of the disease. For example, for a diagnosed ventricular septal defect, the system can automatically compare the size changes of the defect at different times, providing objective evidence for disease prognosis assessment.

[0238] The adaptive model updating module 6 is connected to the cardiac structure time alignment module 3, which is used to update the reference models in the cardiac structure model library based on new analysis cases, improving the system's adaptability to different fetal physiological characteristics.

[0239] This module realizes the self-learning and continuous optimization capability of the system, mainly including three functions:

[0240] First, case evaluation function: quality evaluation of each new analysis case, including image quality, segmentation accuracy and diagnosis consistency, only high-quality cases can be used for model updating. Quality score criteria include signal-to-noise ratio (> 20 dB), contrast ratio (> 0.4), segmentation consistency (Dice> 0.85), etc.

[0241] These score thresholds are determined based on a large amount of clinical data analysis. For example, signal-to-noise ratio > 20 dB is based on the technical specifications of ultrasound equipment and clinical image quality standards, images below this value often have difficulty in clearly displaying the details of the ventricular septum; contrast ratio > 0.4 is based on the normal brightness difference between the ventricular septum and surrounding tissues, images below this value often have unclear boundaries of the ventricular septum; Dice> 0.85 is based on professional standards in the field of medical image segmentation, indicating high consistency between segmentation results and expert annotations.

[0242] Second, model incremental updating function: integrate high-quality new cases into the existing model library to update the feature representation of similar cases. The update uses a sliding window weighted average method:

[0243] ,

[0244] where: is the updated model, is the feature vector of the new case, and and are feature vectors of the same dimension; is the original model, which is the feature vector stored in the model library; is the new case feature, which is the feature vector extracted from the current case; is the update weight, with a value range of [0, 1], usually set to 0.1-0.3, and dynamically adjusted according to the quality score of the new case. The higher the quality score, the greater the value, indicating that the new case has a greater impact on the model.

[0245] The dynamic adjustment formula for the update weight is:

[0246] ,

[0247] where: is the basic update weight, usually set to 0.3; is the quality score of the current case; is the maximum possible value of the quality score, usually 1.0. This dynamic adjustment mechanism ensures that high-quality cases can significantly affect the model, while low-quality cases have relatively less impact.

[0248] Finally, the model library optimization function: regularly assesses the usage frequency and performance of each model in the model library, removes low-frequency usage or poor performance models, and maintains the efficiency of the model library. The system will maintain a usage counter and performance score, and when a model has not been used for 30 consecutive days or the performance score is below the threshold of 0.6, it will be considered for archiving or removal.

[0249] The performance score calculation formula is:

[0250] ,

[0251] where: is the performance score, with a value range of [0, 1]; is the accuracy (Accuracy); is the precision (Precision); is the recall (Recall). The weights of each indicator (0.5, 0.3, 0.2) are determined based on the different tolerances of false positives and false negatives in clinical diagnosis. In congenital heart disease screening, false negatives (missed diagnosis) are usually more serious than false positives (misdiagnosis), so the weight of recall is relatively low.

[0252] Through the adaptive model updating module 6, the system can continuously learn new cases, adapt to different population characteristics and device conditions, and continuously improve the diagnostic performance. Clinical application shows that after 3 months of adaptive learning, the diagnostic accuracy of the system for the Chinese population has increased from 89.5% at the beginning to 94.2%, significantly improving the value of the system in actual application.

[0253] This continuous learning ability is particularly important for the adaptability of the system to different regions and populations. For example, there are subtle differences in the morphology of fetal hearts of different races. Through adaptive learning, the system can gradually adjust the model parameters to adapt to the characteristics of the local population. Similarly, for different brands and models of ultrasound devices, the system can also gradually adapt to their image characteristics through learning, ensuring good diagnostic performance on different devices.

[0254] The overall workflow of the congenital heart disease four-dimensional echocardiogram dynamic segmentation system provided by the application is as follows:

[0255] 1. The image acquisition module 1 acquires four-dimensional echocardiogram data of the fetal heart;

[0256] 2. The optical flow field correction module 2 receives the four-dimensional echocardiogram data, corrects the fetal movement artifacts based on the optical flow field time series analysis technology, and generates corrected four-dimensional echocardiogram data;

[0257] 3. The heart structure time alignment module 3 receives the corrected four-dimensional echocardiogram data, aligns the heart structure through a multi-scale similarity transformation strategy, extracts the interventricular septal defect area, and generates aligned heart structure data;

[0258] 4. The spatio-temporal fusion network module 7 integrates the features of the aligned heart structure data in the time dimension and the spatial dimension, enhancing the feature expression of the defect area;

[0259] 5. The defect analysis module 4 receives the enhanced heart structure data, calculates the defect site edge line diameter and shunt direction, and generates defect parameter information;

[0260] 6. The result output module 5 receives the defect parameter information and generates a congenital heart disease screening report;

[0261] 7. The adaptive model updating module 6 updates the reference model in the heart structure model library based on new analysis cases, improving the adaptability of the system to different fetal physiological characteristics.

[0262] In actual application, the entire processing flow usually takes 30-45 seconds to complete (based on standard medical workstation configuration: Intel Xeon E5-2690 CPU, NVIDIA RTX 3090 GPU, 64GB RAM), while traditional manual analysis of the same data usually takes 10-15 minutes, with significant efficiency improvement.

[0263] The system of the present application has been clinically verified in multiple top-tier hospitals, and a total of 350 fetal echocardiogram data have been collected, including 120 normal fetuses and 230 fetuses with different types of congenital heart disease (including 150 cases of ventricular septal defect). The verification results show that:

[0264] 1. Accuracy: The detection accuracy of the system for ventricular septal defect is 94.2%, the sensitivity is 92.5%, the specificity is 95.8%, and the average diagnostic consistency (Kappa coefficient) with three experienced ultrasound doctors is 0.89;

[0265] 2. Efficiency: The average diagnosis time of the system is 45 seconds per case, which is 15 times faster than manual diagnosis (average 11 minutes per case);

[0266] 3. Early screening ability: The system can detect ventricular septal defect from 16 weeks of pregnancy, which is 4-8 weeks earlier than conventional manual screening (usually 20-24 weeks), providing a time window for early intervention;

[0267] 4. Measurement accuracy: The average measurement error of the system for defect diameter is 0.3mm, which is significantly better than manual measurement (average error 0.8mm).

[0268] At present, the system has been put into clinical trial in some hospitals, and the feedback shows that the system has good practicability and stability, and can effectively assist doctors in congenital heart disease screening, especially in primary medical institutions and scenes with high screening pressure.

[0269] The above only describes the preferred embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A system for dynamic segmentation of four-dimensional echocardiograms of congenital heart disease, characterized in that, The method comprises the following steps: An image acquisition module is configured to acquire four-dimensional echocardiogram data of a fetal heart; A light flow field correction module is connected to the image acquisition module and configured to receive the four-dimensional echocardiogram data, correct fetal movement artifacts based on a light flow field time series analysis technique, and generate corrected four-dimensional echocardiogram data; A heart structure time series alignment module is connected to the light flow field correction module and configured to receive the corrected four-dimensional echocardiogram data, align heart structures through a multi-scale similarity transformation strategy, extract a ventricular septal defect region, and generate aligned heart structure data; wherein the heart structure time series alignment module comprises: A multi-scale pyramid construction unit is configured to construct a multi-level pyramid structure for the received four-dimensional echocardiogram data, each level containing heart images of different resolutions; The feature extraction unit is connected with the multi-scale pyramid construction unit, and is configured to extract a cardiac structure feature in different scale spaces. For each voxel point, a local feature descriptor is calculated by using an improved SIFT algorithm. ， wherein: is a feature descriptor at the voxel point is a dimensional vector; represents a 3D gradient histogram centered at the point, has a value range of [1, m]; each represents the cumulative intensity of the gradient in a specific direction, with a unit of a dimensionless normalized value; A feature fusion unit is connected to the feature extraction unit and configured to weight and integrate features of different levels to generate multi-scale feature descriptions, and adopt an adaptive weight scheme; ， wherein: is the fused feature, a vector with the same dimension as is the feature extracted from the layer, whose value range is [0, n], is the number of pyramid layers minus 1; is the corresponding weight factor, indicating the importance of the feature of the layer, satisfying ; represents the vector weighted summation operation; Weight Determined dynamically by a feature quality assessment function: ， wherein: is the first the quality score of the layer feature is a non-negative real number; denotes the sum of all layer feature quality scores; A model matching unit is connected to the feature fusion unit and configured to search for a most similar reference model from a pre-established heart structure model library based on the multi-scale feature descriptions, and adopt a cosine distance for similarity calculation; A non-rigid transformation unit is connected to the model matching unit and configured to perform non-rigid transformation based on the reference model to align heart structures; A defect analysis module is connected to the heart structure time series alignment module and configured to receive the aligned heart structure data, calculate a defect site edge line diameter and a shunt direction, and generate defect parameter information; A result output module is connected to the defect analysis module and configured to receive the defect parameter information and generate a congenital heart disease screening report. Further comprising: A space-time fusion network module is arranged between the heart structure time series alignment module and the defect analysis module and configured to integrate features in the time dimension and the space dimension for the aligned heart structure data to enhance feature expression of a defect region; The space-time fusion network module comprises: A time series channel is configured to process time series features of dynamic changes of the heart; A space channel is configured to process spatial geometric features of the heart structure; A fusion layer is connected to the time series channel and the space channel and configured to integrate time and space features to generate enhanced defect region expression.

2. The CHD four-dimensional echocardiogram dynamic segmentation system of claim 1, wherein, The heart structure time series alignment module comprises: A multi-scale pyramid construction unit is configured to construct a multi-level pyramid structure for the received four-dimensional echocardiogram data, each level containing heart images of different resolutions; A feature extraction unit is connected to the multi-scale pyramid construction unit and configured to extract heart structure features in different scale spaces; A feature fusion unit is connected to the feature extraction unit and configured to weight and integrate features of different levels to generate multi-scale feature descriptions; A model matching unit is connected to the feature fusion unit and configured to search for a most similar reference model from a pre-established heart structure model library based on the multi-scale feature descriptions; A non-rigid transformation unit is connected to the model matching unit and configured to perform non-rigid transformation based on the reference model to align heart structures.

3. The CHD-4D echocardiogram dynamic segmentation system of claim 2, wherein, The non-rigid transformation unit comprises: a global transformation processor configured to perform a global coarse alignment of the cardiac structure; a local transformation processor connected to the global transformation processor and configured to perform a local fine transformation based on the global alignment; an iterative optimization processor connected to the local transformation processor and configured to adjust transformation parameters through multiple iterations until a preset alignment quality threshold is reached.

4. The CHD-4D echocardiogram dynamic segmentation system of claim 1, wherein, The optical flow field correction module includes: an optical flow field estimation unit configured to estimate the motion field of the fetal heart by analyzing image changes between adjacent ultrasound frames; a motion separation unit connected to the optical flow field estimation unit and configured to separate fetal movement from the heart's own movement; an artifact removal unit connected to the motion separation unit and configured to remove image artifacts caused by fetal movement based on the separated motion information.

5. The CHD-4D echocardiogram dynamic segmentation system of claim 1, wherein, The defect analysis module includes: an edge detection unit configured to identify the interventricular septal defect edge in the aligned cardiac structure data; a size measurement unit connected to the edge detection unit and configured to calculate the diameter of the interventricular septal defect; a shunt direction analysis unit connected to the edge detection unit and configured to determine the shunt direction of the defect based on the morphological features and blood flow information of the interventricular septal defect; a severity assessment unit connected to the size measurement unit and the shunt direction analysis unit and configured to comprehensively analyze defect parameters to assess the severity of congenital heart disease.

6. The four-dimensional echocardiographic dynamic segmentation system of congenital heart disease of claim 2, wherein, The model matching unit includes: a feature vector generator configured to convert the current ultrasound sequence into a feature vector; a similarity calculator connected to the feature vector generator and configured to calculate the similarity between the current feature vector and reference models stored in the model library; an optimal matching selector connected to the similarity calculator and configured to select the reference model with the highest similarity as the matching result.

7. The CHD-4D echocardiogram dynamic segmentation system of claim 1, wherein, The result output module includes: a defect visualization unit configured to convert the defect parameter information into an intuitive graphical representation; a report generation unit connected to the defect visualization unit and configured to generate a congenital heart disease screening report containing diagnostic results and recommendations based on the defect parameter information and the graphical representation; a data storage unit connected to the report generation unit and configured to save the congenital heart disease screening report and related analysis data.

8. The CHD-4D echocardiogram dynamic segmentation system of claim 1, wherein, Further comprising: an adaptive model update module connected to the cardiac structure temporal alignment module and configured to update the reference models in the cardiac structure model library based on new analysis cases, improving the system's adaptability to different fetal physiological characteristics.

Images