Carotid plaque analysis method based on morphological evolution and negative sample perception mechanism

Abstract

The application discloses a carotid plaque analysis method based on morphological evolution and negative sample perception mechanism and belongs to the technical field of medical ultrasound image processing and artificial intelligence auxiliary diagnosis. The application discloses a carotid plaque analysis method based on morphological evolution and negative sample perception mechanism and belongs to the technical field of medical ultrasound image processing and artificial intelligence auxiliary diagnosis. The application discloses a carotid plaque analysis method based on morphological evolution and negative sample perception mechanism and belongs to the technical field of medical ultrasound image processing and artificial intelligence auxiliary diagnosis. The application discloses a carotid plaque analysis method based on morphological evolution and negative sample perception mechanism and belongs to the technical field of medical ultrasound image processing and artificial intelligence auxiliary diagnosis. The application discloses a carotid plaque analysis method based on morphological evolution and negative sample perception mechanism and belongs to the technical field of medical ultrasound image processing and artificial intelligence auxiliary diagnosis.

Description

Technical Field

[0001] This invention relates to the fields of medical ultrasound image processing and artificial intelligence-assisted diagnosis, and in particular to a method for carotid plaque analysis based on morphological evolution and negative sample perception mechanisms. Background Technology

[0002] The carotid artery, a vital blood supply artery in the human body, directly impacts the safety of blood supply to the brain. The formation of atherosclerotic plaques in the carotid arteries is a significant contributing factor to serious cardiovascular and cerebrovascular diseases such as stroke. Therefore, accurate detection and morphological assessment of these plaques are crucial for early disease screening, risk stratification, and clinical intervention.

[0003] Currently, ultrasound imaging has become the primary method for carotid artery plaque screening due to its advantages such as being non-invasive, real-time, and low-cost. Existing carotid artery plaque detection and assessment mainly fall into two categories of known technologies: manual image interpretation and traditional medical image analysis.

[0004] (1) Manual image interpretation scheme: After obtaining static ultrasound images of the carotid artery, medical staff observe the grayscale differences of the images frame by frame and rely on subjective experience to distinguish normal blood vessel walls, soft tissues and plaque areas; calcified plaques and non-calcified plaques are roughly divided according to the grayscale level; the thickness of the intima-media of the carotid artery and the size of the plaque are manually measured using a manual ruler, and finally the plaque and stroke risk level are qualitatively judged.

[0005] (2) Traditional image and shallow AI analysis scheme: The ultrasound image is segmented simply using basic algorithms such as fixed threshold and edge detection; single gray level and shallow texture features of the image are extracted, and the basic convolutional neural network is used to perform simple binary classification of the patches; the model has not been lightweighted, has a large number of parameters, and only supports offline batch image analysis after screening to make a preliminary judgment on the presence or absence of basic patches.

[0006] However, existing technologies still have significant shortcomings in practical applications, mainly in the following aspects: ① Existing patch recognition relies on subjective interpretation by human eyes, and traditional algorithms only use a single gray-level shallow feature, which cannot integrate multi-dimensional features such as morphology, texture, and echo. The gray-level features of different types of patches highly overlap, resulting in low accuracy of patch fine classification, high false negative rate of sidewall patches, and easy false detection or false negative of small patches.

[0007] ②Since the thickness of the carotid intima-media, the area and size of the plaque are all measured manually, there is a lack of a unified automated quantitative calculation mechanism, which leads to large subjective errors in measurement and poor standardization and repeatability of test results.

[0008] ③ Existing technologies lack vascular skeleton extraction and ICP (Iterative Closest Point) image registration mechanisms. Carotid artery images at different screening time points cannot achieve precise spatial alignment, and cannot dynamically track plaque size, morphology, carotid intima-media thickness, and luminal lesions for longitudinal comparison. They also cannot dynamically track the evolution trend of plaques and vascular lesions, making it difficult to support long-term follow-up assessment.

[0009] ④ Existing technologies do not incorporate clinical guidelines to build standardized assessment models, and can only achieve qualitative description of plaques. They cannot automatically complete multi-level risk grading of stroke and are difficult to output interpretable standardized screening reports. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this invention solves problems such as the scarcity of unstable patch samples, high false detection rate of background artifacts, lack of anatomical logical constraints on recognition results, and difficulty in conducting long-term dynamic follow-up by using morphological cascade evolution, negative sample perception segmentation, patch quantification analysis, and multi-temporal image registration.

[0011] To achieve the above objectives, this invention provides a method for carotid plaque analysis based on morphological evolution and negative sample perception mechanisms, comprising the following steps: (1) Obtain two-dimensional grayscale ultrasound images of the carotid artery, complete the association of case information, construct an ultrasound image sample set, and perform preprocessing; (2) The preprocessed ultrasound image sample set is simulated for pathological evolution through a morphological cascade evolution simulation mechanism to generate an evolution-enhanced sample set; (3) Construct a negative sample perception segmentation network and input ultrasound images and evolution-enhanced samples into the negative sample perception segmentation network for training; The negative sample-aware segmentation network uses U-Net as its basic segmentation architecture and introduces a negative sample-aware loss function with anatomical topological constraints during the training phase. (4) Input the two-dimensional grayscale ultrasound image of the carotid artery to be identified into the negative sample perception segmentation network for plaque segmentation, measure the plaque area parameters, carotid intima-media thickness parameters, and lumen stenosis rate parameters, and extract the echo characteristics and morphological characteristics of the plaque, and output the plaque stability assessment and stroke risk related results.

[0012] Furthermore, the preprocessing includes: denoising, grayscale normalization, and size standardization of the original ultrasound image to eliminate grayscale differences and speckle noise caused by different equipment, gain, and scanning environment.

[0013] Furthermore, the morphological cascade evolution simulation mechanism includes: (2.1) Artificial plaque region masking is performed on the preprocessed ultrasound image samples, and the corresponding blood vessel wall boundary information is obtained; (2.2) The continuous deformation vector generated by the elastic deformation operator is used to perform pixel-level spatial mapping of the patch region to simulate the local displacement and irregular deformation of the patch; (2.3) The edge erosion operator is used to perform directional morphological erosion operation on the outer edge of the plaque area to simulate the pathological process of fibrous cap destruction, local defects and ulcer formation; (2.4) The elastic deformation results and edge corrosion results are cascaded and fused to generate simulated unstable patch samples.

[0014] Furthermore, the implementation process of the elastic deformation operator is as follows: a random displacement field is generated and a continuous deformation vector is formed by Gaussian smoothing. The stable patch area is spatially mapped at the pixel level to realize the local offset of the patch contour, the irregular change of the surface, and the simulation of the tissue compression morphology, so that the patch gradually changes from a regular boundary to an irregular morphology. The implementation process of the edge erosion operator is as follows: local morphological erosion operation is performed on the boundary of the patch area using preset structuring elements to generate features such as patch surface roughening, local depression, boundary breakage, and discontinuous contour, so that the patch presents irregular morphological features close to the real high-risk lesion.

[0015] Furthermore, the morphological cascade evolution simulation mechanism can also use generative adversarial networks (GANs) or diffusion models to generate pathological style transfer of plaque surface roughness in preprocessed ultrasound image sample sets under supervised or semi-supervised conditions, or use computational fluid dynamics (CFD) blood flow shear force simulation deformation to replace pure geometric morphological deformation, thereby generating and amplifying lesion evolution samples.

[0016] Furthermore, step (3) specifically involves: Preprocessed ultrasound images were used to locate the vascular wall structure and establish the vascular wall base topology information. A ROI-restricted mask is established using the spatial boundary of the blood vessel wall as a constraint condition; U-Net is used for patch segmentation. When the network misidentifies background artifacts as patches, the loss penalty weight of the corresponding region is increased. For real patch regions that conform to the topological relationship of the blood vessel wall, the probability of mis-penalty is reduced, thereby achieving reverse removal of background artifacts.

[0017] Furthermore, step (3) can also be replaced by a two-stage cascaded network; In the first stage of the network, Attention U-Net is used to segment the blood vessel wall structure and generate a region of interest mask for the preprocessed ultrasound image. The second-stage network uses the ultrasound image constrained by the region of interest mask as input and employs U-Net to perform fine segmentation of the plaque region. By limiting the segmentation range, false detections caused by background artifacts outside the blood vessel wall are reduced, realizing a cascaded segmentation process of first locating the blood vessel wall structure and then performing plaque recognition; or a Transformer-based segmentation network is used, which enhances the model's ability to focus on the features of the blood vessel wall region and plaque region by introducing an attention mechanism.

[0018] Furthermore, step (4) specifically involves: The patch area parameters are obtained by reading the patch segmentation mask and counting the number of pixels in the mask area; after completing the image scale calibration, the pixel area is converted into the actual anatomical area to obtain the patch area parameters, including the total patch area, the maximum cross-sectional area, and the cumulative area of ​​multiple patches. Carotid intima-media thickness parameters were obtained by extracting the proximal and distal boundaries of the vessel wall, identifying the intima-lumen interface and the media-adventitia interface, calculating the vertical distance between them, obtaining the IMT value, and further obtaining the average IMT, maximum IMT, and IMT distribution at multiple locations. The stenosis rate parameter is calculated by measuring the minimum lumen diameter at the lesion site and using the diameter of the adjacent normal blood vessel as a reference value; it includes the minimum lumen diameter, the normal reference diameter, and the stenosis rate percentage. Echo characteristics include: low echo, high echo, high echo, and the degree of echo heterogeneity. Morphological features include: surface roughness, boundary irregularity index, ulcer-like changes, and shape complexity; Stability assessment, based on plaque area parameters, carotid intima-media thickness parameters, lumen stenosis rate parameters, echogenicity and morphological characteristics, outputs stable plaques, unstable plaques, or high-risk vulnerable plaques; Stroke risk is assessed based on plaque area parameters, carotid intima-media thickness parameters, lumen stenosis rate parameters, echogenicity, and morphological characteristics, resulting in a stroke-related risk level evaluation.

[0019] Furthermore, it also includes step (5): based on the multi-temporal ultrasound image registration mechanism, spatial alignment and evolution analysis are performed on historical ultrasound images, current ultrasound images, and corresponding plaque segmentation results from different screening cycles, and finally plaque stability assessment and stroke risk-related results are output, specifically: Based on historical ultrasound images, current ultrasound images, and corresponding plaque segmentation results from different screening cycles, skeletonization processing is performed on the vascular region, and the vascular skeleton point set, bifurcation position, and geometric topology are obtained through a centerline extraction algorithm. The iterative nearest point (ICP) algorithm is used to establish the nearest neighbor correspondence between the skeleton points of historical ultrasound images and current ultrasound images. The spatial transformation matrix is ​​iteratively optimized to achieve rigid or affine alignment of the two sets of images. After alignment, longitudinal comparative analysis was performed on changes in plaque area parameters, trends in carotid intima-media thickness parameters, changes in lumen stenosis rate parameters, and morphological evolution, ultimately generating a lesion evolution analysis report. The multi-temporal ultrasound image registration mechanism can also use the deep learning-based end-to-end registration network Voxelmorph to achieve flexible spatiotemporal registration, or use the carotid bifurcation as a feature anchor point combined with SIFT / ORB feature operators to complete image registration, thereby achieving spatial alignment of images from different screening cycles.

[0020] This invention also provides a carotid plaque analysis system based on morphological evolution and negative sample perception mechanisms, comprising: The ultrasound image input and preprocessing module is used to acquire a set of two-dimensional grayscale ultrasound images of the carotid artery and perform preprocessing. The morphological cascade evolution enhancement module is used to simulate the pathological evolution process of the preprocessed ultrasound image sample set through the morphological cascade evolution simulation mechanism, thereby generating an evolution enhancement sample set; The negative sample perception patch segmentation module is used to construct a negative sample perception segmentation network. Ultrasound images and evolution-enhanced samples are jointly input into the negative sample perception segmentation network for training. The negative sample-aware segmentation network uses U-Net as its basic segmentation architecture and introduces a negative sample-aware loss function with anatomical topological constraints during the training phase. The plaque parameter quantification and stability analysis module is used to input the two-dimensional grayscale ultrasound image of the carotid artery to be identified into the negative sample perception segmentation network for plaque segmentation, measure plaque area parameters, carotid intima-media thickness parameters, and lumen stenosis rate parameters, and extract the echo and morphological features of the plaque at the same time, and perform plaque stability assessment and stroke risk level assessment. The multi-temporal image registration module is used to perform spatial alignment and evolution analysis on historical ultrasound images, current ultrasound images and corresponding plaque segmentation results from different screening cycles based on the multi-temporal ultrasound image registration mechanism. The lesion evolution analysis and risk output module is used to output the final plaque stability assessment and stroke risk-related results.

[0021] The beneficial effects of this invention are: 1. Overcome the training bottlenecks caused by scarce patch samples and imbalanced datasets.

[0022] Traditional deep learning segmentation networks heavily rely on massive amounts of real pathological data. However, image samples of unstable and ulcerated high-risk plaques are extremely difficult to obtain due to their rapid onset and the need for immediate surgery upon detection. This invention constructs an algorithm based on "elastic deformation + edge erosion" to mathematically simulate the dynamic pathological process of plaque evolution from a "stable state" to an "ulcerated state." This allows the model to achieve high generalization ability in recognizing high-risk unstable plaques without relying on a large number of rare clinical samples.

[0023] 2. By endowing the network anatomical structure with a priori and logical proof capabilities, the misdiagnosis rate of clinical artifacts is significantly reduced.

[0024] Conventional segmentation algorithms, lacking anatomical logical constraints, are prone to misclassifying common background artifacts in ultrasound images, such as acoustic shadowing and reverberation, as plaques. This invention introduces a negative sample perception mechanism with spatial topological constraints into the loss function, forcing the model to follow a dual feedback logic during optimization: first, the anatomical structure of the vessel wall base must be extracted and confirmed; then, using this boundary as a precondition, background grayscale artifacts in non-vessel wall areas are forcibly eliminated in reverse, significantly reducing the false negative and false positive rates of small plaques and lateral wall plaques.

[0025] 3. Achieve dynamic alignment and longitudinal evolution tracking of ultrasound images across multiple temporal spaces and morphologies.

[0026] This invention overcomes the limitations of traditional methods that can only perform coarse binary classification based on single-shot, offline, and static plaque presence or absence. It introduces an image registration mechanism combining vascular skeleton extraction and ICP (Iterative Closest Point), which automatically overcomes the non-rigid differences in image space caused by probe angle and patient position at different screening time points, enabling long-term, fully automated dynamic tracking of the evolution trend of carotid intima-media thickness (IMT) and plaque geometry.

[0027] 4. End-to-end automated quantification throughout the entire process eliminates heterogeneity and subjective errors caused by human measurement.

[0028] By integrating multi-dimensional features such as morphology, texture, and echo, the system automatically measures core parameters related to plaques, fundamentally avoiding subjective experience and measurement errors caused by traditional manual image reading or manual rulers. This significantly improves the data standardization and clinical interpretability of large-scale early screening for stroke at the community level. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the carotid plaque analysis method based on morphological evolution and negative sample perception mechanism according to an embodiment of the present invention.

[0030] Figure 2 This is a schematic diagram of the morphological cascade evolution process in an embodiment of the present invention.

[0031] Figure 3 This is a schematic diagram of the patch segmentation process based on the negative sample perceptual segmentation network in an embodiment of the present invention.

[0032] Figure 4 This is a schematic diagram of the plaque parameter quantification analysis process according to an embodiment of the present invention.

[0033] Figure 5 This is a schematic diagram of the multi-temporal ultrasound image registration mechanism according to an embodiment of the present invention. Detailed Implementation

[0034] The present invention will be further explained and described below with reference to the accompanying drawings and embodiments.

[0035] like Figure 1 As shown, this invention provides a method for carotid plaque analysis based on morphological evolution and negative sample perception mechanisms, comprising the following steps: S101. Obtain two-dimensional grayscale ultrasound images of the carotid artery, complete the association of case information, construct an ultrasound image sample set, and perform preprocessing.

[0036] Two-dimensional grayscale ultrasound images were acquired using a carotid ultrasound device, and case information was associated, including the corresponding case number and examination time. An ultrasound image sample set was constructed, and the original images of the sample set were standardized and preprocessed.

[0037] Preprocessing includes denoising, grayscale normalization, and size standardization of the original ultrasound images to eliminate grayscale differences and speckle noise caused by different equipment, gain, and scanning environment.

[0038] S102. The preprocessed ultrasound image sample set is simulated for pathological evolution through a morphological cascade evolution simulation mechanism to generate an evolution-enhanced sample set.

[0039] Because high-risk unstable plaques, especially lesions that have formed ulcer-like changes, have a low incidence, are difficult to obtain, and have data privacy restrictions, existing deep learning models struggle to obtain sufficient training samples, which can lead to insufficient ability to identify complex lesions.

[0040] To address the aforementioned problems, this invention expands the sample set by constructing unstable plaque samples based on morphological cascade evolution, generating an evolution-enhanced sample set. This method uses stable plaque ultrasound images as input and simulates the dynamic process of plaque evolution from a stable to an unstable state at the geometric morphology and pathological feature levels through a cascaded combination of elastic deformation operators and edge erosion operators.

[0041] like Figure 2 As shown, the specific process includes: (1) Artificial plaque region masking is performed on the preprocessed ultrasound image samples, and the corresponding blood vessel wall boundary information is obtained.

[0042] (2) The continuous deformation vector generated by the elastic deformation operator is used to perform pixel-level spatial mapping of the patch area to simulate the local displacement and irregular deformation of the patch.

[0043] To simulate the local displacement and irregular deformation of plaques under long-term blood flow impact and tissue pressure, this embodiment of the invention employs an elastic deformation operator for spatial perturbation. The implementation process of the elastic deformation operator involves generating a random displacement field and forming a continuous deformation vector through Gaussian smoothing. This vector is then used for pixel-level spatial mapping of the stable plaque region, simulating local offset of the plaque contour, irregular surface changes, and tissue pressure morphology, causing the plaque to gradually transform from a regular boundary to an irregular shape.

[0044] (3) Perform directional morphological corrosion operation on the outer edge of the plaque by using the edge corrosion operator to simulate the pathological process of fiber cap destruction, local defects and ulcer formation.

[0045] The implementation process of the edge erosion operator is as follows: local morphological erosion operation is performed on the patch boundary using preset structuring elements to generate features such as patch surface roughening, local depression, boundary breakage, and discontinuous contour, so that the patch presents irregular morphological features close to the real high-risk lesion.

[0046] (4) The elastic deformation results and edge corrosion results are cascaded and fused to generate simulated unstable patch samples.

[0047] Finally, the elastic deformation results and edge corrosion results are cascaded and fused to generate simulated unstable patch samples. These include: a. evolutionary ultrasound images; b. evolutionary mask labels; c. an unstable patch training sample set.

[0048] In addition to using a combination of morphological operators, such as "elastic deformation + edge erosion", morphological simulation evolution algorithms can also be replaced by using generative adversarial networks (GANs) or diffusion models to generate pathological style transfer of plaque surface roughness under supervised or semi-supervised conditions, or by using computational fluid dynamics (CFD) blood flow shear force to simulate deformation instead of pure geometric morphological deformation, thereby simulating the evolution of plaques from stability to ulceration.

[0049] S103. Construct a negative sample perception segmentation network by inputting ultrasound images and evolution-enhanced samples into the negative sample perception segmentation network for training.

[0050] Traditional deep learning segmentation methods typically learn patch regions directly based on grayscale features, lacking the necessary logical constraints of anatomical structure. Therefore, they are prone to misidentifying acoustic shadows, reverberation, and speckle artifacts in ultrasound images as lesion areas.

[0051] To address the aforementioned issues, this invention introduces a U-Net plaque segmentation method based on a negative sample perception mechanism. U-Net is used as the basic segmentation network, and a negative sample perception mechanism is incorporated into its loss function design to achieve priority identification of the blood vessel wall and reverse removal of background artifacts.

[0052] like Figure 3 As shown, the specific process is as follows: Preprocessed ultrasound images are used to locate the vascular wall structure. Because plaque lesions have strict anatomical attachment relationships, they must be located in the vicinity of the vascular wall; therefore, the system first establishes the basal topological information of the vascular wall.

[0053] A Region of Interest (ROI) constraint mask is established using the spatial boundary of the vessel wall as a constraint. Only candidate regions located within the neighborhood of the vessel wall are allowed to enter subsequent plaque segmentation calculations.

[0054] Building upon this foundation, the U-Net encoder extracts multi-scale texture and echo features, while the decoder restores spatial resolution layer by layer, achieving fine segmentation of plaque contours. A negative sample perception weighting term is introduced into the training loss function. When the network misidentifies background artifacts as plaques, the corresponding region's loss penalty weight is increased; conversely, for genuine plaque regions conforming to the topological relationship of the blood vessel wall, the probability of mis-penalty is reduced, achieving reverse removal of background artifacts.

[0055] The backpropagation phase of the network establishes a dual feedback mechanism: first, the vascular wall structure is identified, and then non-vascular wall background areas are excluded. This mechanism significantly enhances the model's ability to identify lateral wall plaques, micro-plaques, and complex high-risk lesions.

[0056] The above operation is a single-stage segmentation method, which can also be replaced by a two-stage cascaded network. The first stage network uses Attention U-Net to segment the blood vessel wall structure and generate a region of interest mask for the preprocessed ultrasound image. The second stage network uses the ultrasound image constrained by the region of interest mask as input and uses U-Net to perform fine segmentation of the plaque region. By limiting the segmentation range, false detections caused by background artifacts outside the blood vessel wall are reduced, realizing a cascaded segmentation process of first locating the blood vessel wall structure and then identifying plaques. Alternatively, a Transformer-based segmentation network can be used, which enhances the model's ability to focus on the features of the blood vessel wall region and plaque region by introducing an attention mechanism.

[0057] S104. Input the two-dimensional grayscale ultrasound image of the carotid artery to be identified into the negative sample perception segmentation network for plaque segmentation, measure the plaque area parameters, carotid intima-media thickness parameters, and lumen stenosis rate parameters, and extract the echogenicity and morphological features of the plaque, and output the plaque stability assessment and stroke risk-related results.

[0058] After completing plaque region segmentation based on the negative sample-aware U-Net network, further plaque parameter quantification and stability analysis are performed to achieve objective measurement and risk assessment of carotid artery lesions.

[0059] like Figure 4 As shown, this step uses the plaque segmentation result mask as input and achieves a comprehensive quantitative assessment of plaque stability through area measurement, carotid intima-media thickness measurement, stenosis rate calculation, echo feature analysis, and morphological feature extraction.

[0060] The plaque area parameter is obtained by reading the plaque segmentation mask and counting the number of pixels in the mask area; after image scale calibration, the pixel area is converted into the actual anatomical area to obtain the plaque area parameter. The calculation results can be expressed as: total plaque area, maximum cross-sectional area, and cumulative area of ​​multiple plaques. This parameter can reflect the degree of lesion burden and the trend of lesion development.

[0061] The carotid intima-media thickness (IMT) parameter extracts the proximal and distal boundaries of the vessel wall, identifies the intima-lumen interface and the media-adventitia interface, and calculates the vertical distance between them to obtain the IMT value. Output includes: mean IMT, maximum IMT, and IMT distribution at multiple locations. IMT is an important indicator of early atherosclerotic lesions and can be used to assess the degree of vessel wall thickening and the risk of stroke.

[0062] The stenosis rate parameter is calculated by measuring the minimum lumen diameter at the lesion site and using the diameter of an adjacent normal vessel as a reference. Output results include: minimum lumen diameter, normal reference diameter, and stenosis rate percentage. This indicator reflects the degree of blood flow restriction and provides a basis for clinical intervention decisions.

[0063] Echo characteristics include: low echo, high echo, high echo, and the degree of echo heterogeneity.

[0064] Morphological features include: surface roughness, boundary irregularity index, ulcer-like changes, and shape complexity.

[0065] Since unstable plaques typically exhibit heterogeneous echogenicity, rough edges, and localized ulceration, these characteristics can effectively enhance the ability to determine stability.

[0066] Stability assessment integrates plaque area parameters, carotid intima-media thickness parameters, luminal stenosis rate parameters, echogenicity, and morphological characteristics to establish a stability classification model. Outputs include: stable plaques, unstable plaques, or high-risk vulnerable plaques. Simultaneously, stroke risk assessment results are generated. Stroke risk, based on plaque area parameters, carotid intima-media thickness parameters, luminal stenosis rate parameters, echogenicity, and morphological characteristics, outputs a stroke-related risk level evaluation, providing auxiliary decision-making basis for clinical screening and follow-up.

[0067] This step automates the conversion from single segmentation results to multidimensional pathological parameter analysis.

[0068] S105. Based on the multi-temporal ultrasound image registration mechanism, spatial alignment and evolution analysis are performed on historical ultrasound images, current ultrasound images, and corresponding plaque segmentation results from different screening cycles, and finally plaque stability assessment and stroke risk-related results are output.

[0069] Traditional ultrasound screening can usually only provide information on a single, static lesion, which is difficult to accurately reflect the spatial evolution trend of plaques during long-term follow-up.

[0070] Because ultrasound images at different examination time points are affected by probe angle, patient position, and respiratory movements, there are often significant spatial offsets and distortions between them. Therefore, directly comparing images from different times can easily lead to large errors. Thus, this invention further proposes a multi-temporal image registration mechanism based on vascular skeleton extraction and the ICP algorithm. Using carotid ultrasound images from different screening time points as input, and through vascular skeleton centerline extraction and spatial alignment, long-term dynamic tracking of lesions is achieved.

[0071] like Figure 5 As shown, the specific process is as follows: (1) Since the carotid artery centerline has high anatomical stability, the vascular region is first subjected to skeletonization processing based on historical ultrasound images, current ultrasound images, and corresponding plaque segmentation results from different screening cycles. The vascular skeleton point set, bifurcation positions, and geometric topology are obtained through a centerline extraction algorithm. This step can significantly reduce the impact of local image noise and boundary differences on the registration results.

[0072] (2) After obtaining the skeleton point set, the Iterative Closest Point (ICP) algorithm is used to establish the nearest neighbor correspondence between the skeleton points of historical ultrasound images and current ultrasound images. The spatial transformation matrix is ​​iteratively optimized to achieve rigid or affine alignment of the two sets of images. During the registration process, the distance error between corresponding points is continuously minimized until the convergence condition is met. The output includes: a. Spatial transformation matrix; b. Registration error; c. Aligned vascular image. (3) After alignment, longitudinal comparative analysis of plaque area parameter changes, carotid intima-media thickness parameter changes, lumen stenosis rate parameter changes and morphological feature evolution is performed to finally form a lesion evolution analysis report; This step breaks through the traditional single static screening model and enables long-term dynamic monitoring of carotid artery lesions.

[0073] In multi-temporal image alignment, in addition to the ICP algorithm based on the vascular skeleton, alternatives include: achieving spatiotemporal flexible registration based on a deep learning end-to-end registration network (such as Voxelmorph); or using the carotid bifurcation as a feature anchor point and combining it with traditional SIFT / ORB feature operators to perform image registration, thereby achieving spatial alignment of images from different screening cycles.

[0074] This invention also provides a carotid plaque analysis system based on morphological evolution and negative sample perception mechanisms, including: The ultrasound image input and preprocessing module is used to acquire a set of two-dimensional grayscale ultrasound images of the carotid artery and perform preprocessing.

[0075] The morphological cascade evolution enhancement module is used to simulate the pathological evolution process of the preprocessed ultrasound image sample set through a morphological cascade evolution simulation mechanism, thereby generating an evolution enhancement sample set.

[0076] The negative sample perception patch segmentation module is used to construct a negative sample perception segmentation network. Ultrasound images and evolution-enhanced samples are input into the negative sample perception segmentation network for training.

[0077] The Negative Sample Aware Segmentation Network uses U-Net as its basic segmentation architecture and introduces a negative sample awareness loss function with anatomical topological constraints during the training phase.

[0078] The plaque parameter quantification and stability analysis module is used to input the two-dimensional grayscale ultrasound image of the carotid artery to be identified into the negative sample perception segmentation network for plaque segmentation, measure plaque area parameters, carotid intima-media thickness parameters, and lumen stenosis rate parameters, and extract the echo and morphological features of the plaque, while also performing plaque stability assessment and stroke risk level assessment.

[0079] The multi-temporal image registration module is used to perform spatial alignment and evolution analysis on historical ultrasound images, current ultrasound images, and corresponding plaque segmentation results from different screening cycles based on the multi-temporal ultrasound image registration mechanism.

[0080] The lesion evolution analysis and risk output module is used to output the final plaque stability assessment and stroke risk-related results.

[0081] Finally, it should be noted that the above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for carotid plaque analysis based on morphological evolution and negative sample perception mechanism, characterized in that, Includes the following steps: (1) Obtain two-dimensional grayscale ultrasound images of the carotid artery, complete the association of case information, construct an ultrasound image sample set, and perform preprocessing; (2) The preprocessed ultrasound image sample set is simulated for pathological evolution through a morphological cascade evolution simulation mechanism to generate an evolution-enhanced sample set; (3) Construct a negative sample perception segmentation network and input ultrasound images and evolution-enhanced samples into the negative sample perception segmentation network for training; The negative sample-aware segmentation network uses U-Net as its basic segmentation architecture and introduces a negative sample-aware loss function with anatomical topological constraints during the training phase. (4) Input the two-dimensional grayscale ultrasound image of the carotid artery to be identified into the negative sample perception segmentation network for plaque segmentation, measure the plaque area parameters, carotid intima-media thickness parameters, and lumen stenosis rate parameters, and extract the echo characteristics and morphological characteristics of the plaque, and output the plaque stability assessment and stroke risk related results.

2. The carotid plaque analysis method based on morphological evolution and negative sample perception mechanism according to claim 1, characterized in that, The preprocessing includes: denoising, grayscale normalization, and size standardization of the original ultrasound images to eliminate grayscale differences and speckle noise caused by different equipment, gain, and scanning environment.

3. The carotid plaque analysis method based on morphological evolution and negative sample perception mechanism according to claim 1, characterized in that, The morphological cascade evolution simulation mechanism includes: (2.1) Artificial plaque region masking is performed on the preprocessed ultrasound image samples, and the corresponding blood vessel wall boundary information is obtained; (2.2) The continuous deformation vector generated by the elastic deformation operator is used to perform pixel-level spatial mapping of the patch region to simulate the local displacement and irregular deformation of the patch; (2.3) The edge erosion operator is used to perform directional morphological erosion operation on the outer edge of the plaque area to simulate the pathological process of fibrous cap destruction, local defects and ulcer formation; (2.4) The elastic deformation results and edge corrosion results are cascaded and fused to generate simulated unstable patch samples.

4. The carotid plaque analysis method based on morphological evolution and negative sample perception mechanism according to claim 3, characterized in that, The implementation process of the elastic deformation operator is as follows: a random displacement field is generated and a continuous deformation vector is formed by Gaussian smoothing. The stable patch area is spatially mapped at the pixel level to realize the local offset of the patch contour, surface irregularity change, and tissue compression morphology simulation, so that the patch gradually transforms from a regular boundary to an irregular morphology. The implementation process of the edge erosion operator is as follows: local morphological erosion operation is performed on the boundary of the patch area using preset structuring elements to generate features such as patch surface roughening, local depression, boundary breakage, and discontinuous contour, so that the patch presents irregular morphological features close to the real high-risk lesion.

5. The carotid plaque analysis method based on morphological evolution and negative sample perception mechanism according to claim 1, characterized in that, The morphological cascade evolution simulation mechanism can also use generative adversarial networks (GANs) or diffusion models to generate pathological style transfer of plaque surface roughness in preprocessed ultrasound image sample sets under supervised or semi-supervised conditions, or use computational fluid dynamics (CFD) blood flow shear force simulation deformation to replace pure geometric morphological deformation, thereby generating and amplifying lesion evolution samples.

6. The carotid plaque analysis method based on morphological evolution and negative sample perception mechanism according to claim 1, characterized in that, The specific steps (3) are as follows: Preprocessed ultrasound images were used to locate the vascular wall structure and establish the vascular wall base topology information. A ROI-restricted mask is established using the spatial boundary of the blood vessel wall as a constraint condition; U-Net is used for patch segmentation. When the network misidentifies background artifacts as patches, the loss penalty weight of the corresponding region is increased. For real patch regions that conform to the topological relationship of the blood vessel wall, the probability of mis-penalty is reduced, thereby achieving reverse removal of background artifacts.

7. The carotid plaque analysis method based on morphological evolution and negative sample perception mechanism according to claim 1, characterized in that: Step (3) can also be replaced by a two-stage cascaded network; In the first stage of the network, Attention U-Net is used to segment the blood vessel wall structure and generate a region of interest mask for the preprocessed ultrasound image. The second-stage network uses the ultrasound image constrained by the region of interest mask as input and employs U-Net to perform fine segmentation of the plaque region. By limiting the segmentation range, false detections caused by background artifacts outside the blood vessel wall are reduced, realizing a cascaded segmentation process of first locating the blood vessel wall structure and then performing plaque recognition; or a Transformer-based segmentation network is used, which enhances the model's ability to focus on the features of the blood vessel wall region and plaque region by introducing an attention mechanism.

8. The carotid plaque analysis method based on morphological evolution and negative sample perception mechanism according to claim 1, characterized in that, Step (4) specifically involves: The patch area parameter is obtained by reading the patch segmentation mask and counting the number of pixels in the mask area; after completing the image scale calibration, the pixel area is converted into the actual anatomical area to obtain the patch area parameter. This includes the total area of ​​the plaques, the maximum cross-sectional area, and the cumulative area of ​​multiple plaques. Carotid intima-media thickness parameters were obtained by extracting the proximal and distal boundaries of the vessel wall, identifying the intima-lumen interface and the media-adventitia interface, calculating the vertical distance between them, obtaining the IMT value, and further obtaining the average IMT, maximum IMT, and IMT distribution at multiple locations. The stenosis rate parameter is calculated by measuring the minimum lumen diameter at the lesion site and using the diameter of the adjacent normal blood vessel as a reference value. This includes the minimum lumen diameter, the normal reference diameter, and the percentage of stenosis. Echo characteristics include: low echo, high echo, high echo, and the degree of echo heterogeneity. Morphological features include: surface roughness, boundary irregularity index, ulcer-like changes, and shape complexity; Stability assessment, based on plaque area parameters, carotid intima-media thickness parameters, lumen stenosis rate parameters, echogenicity and morphological characteristics, outputs stable plaques, unstable plaques, or high-risk vulnerable plaques; Stroke risk is assessed based on plaque area parameters, carotid intima-media thickness parameters, lumen stenosis rate parameters, echogenicity, and morphological characteristics, resulting in a stroke-related risk level evaluation.

9. The carotid plaque analysis method based on morphological evolution and negative sample perception mechanism according to claim 1, characterized in that, It also includes step (5): Based on the multi-temporal ultrasound image registration mechanism, spatial alignment and evolution analysis are performed on historical ultrasound images, current ultrasound images, and corresponding plaque segmentation results from different screening cycles, and finally, plaque stability assessment and stroke risk-related results are output, specifically: Based on historical ultrasound images, current ultrasound images, and corresponding plaque segmentation results from different screening cycles, skeletonization processing is performed on the vascular region, and the vascular skeleton point set, bifurcation position, and geometric topology are obtained through a centerline extraction algorithm. The iterative nearest point (ICP) algorithm is used to establish the nearest neighbor correspondence between the skeleton points of historical ultrasound images and current ultrasound images. The spatial transformation matrix is ​​iteratively optimized to achieve rigid or affine alignment of the two sets of images. After alignment, longitudinal comparative analysis was performed on changes in plaque area parameters, trends in carotid intima-media thickness parameters, changes in lumen stenosis rate parameters, and morphological evolution, ultimately generating a lesion evolution analysis report. The multi-temporal ultrasound image registration mechanism can also use the deep learning-based end-to-end registration network Voxelmorph to achieve flexible spatiotemporal registration, or use the carotid bifurcation as a feature anchor point combined with SIFT / ORB feature operators to complete image registration, thereby achieving spatial alignment of images from different screening cycles.

10. A carotid plaque analysis system based on morphological evolution and negative sample perception mechanism, characterized in that, include: The ultrasound image input and preprocessing module is used to acquire a set of two-dimensional grayscale ultrasound images of the carotid artery and perform preprocessing. The morphological cascade evolution enhancement module is used to simulate the pathological evolution process of the preprocessed ultrasound image sample set through the morphological cascade evolution simulation mechanism, thereby generating an evolution enhancement sample set; The negative sample perception patch segmentation module is used to construct a negative sample perception segmentation network. Ultrasound images and evolution-enhanced samples are jointly input into the negative sample perception segmentation network for training. The negative sample-aware segmentation network uses U-Net as its basic segmentation architecture and introduces a negative sample-aware loss function with anatomical topological constraints during the training phase. The plaque parameter quantification and stability analysis module is used to input the two-dimensional grayscale ultrasound image of the carotid artery to be identified into the negative sample perception segmentation network for plaque segmentation, measure plaque area parameters, carotid intima-media thickness parameters, and lumen stenosis rate parameters, and extract the echo and morphological features of the plaque at the same time, and perform plaque stability assessment and stroke risk level assessment. The multi-temporal image registration module is used to perform spatial alignment and evolution analysis on historical ultrasound images, current ultrasound images and corresponding plaque segmentation results from different screening cycles based on the multi-temporal ultrasound image registration mechanism. The lesion evolution analysis and risk output module is used to output the final plaque stability assessment and stroke risk-related results.

Images