A method and device for identifying vascular structural features based on intravascular ultrasound images

Abstract

This invention relates to the field of image recognition, and more particularly to a method and apparatus for recognizing vascular structural features based on intravascular ultrasound images. The method includes the following steps: introducing an ultrasound catheter into a target blood vessel and performing a uniform retraction scan along the vessel axis to obtain a standard ultrasound image sequence; performing pixel-level segmentation and frame-by-frame contour feature calculation based on the standard ultrasound image sequence to generate contour structural feature parameters for different frames; performing vascular axial direction position analysis on the standard ultrasound image sequence to generate three-dimensional pose information of the vascular contour for each frame; identifying key contour feature points and modeling discrete point clouds based on the three-dimensional pose information to construct a three-dimensional mesh model; and performing surface smoothing filtering and parameter constraint adjustment on the three-dimensional mesh model based on the contour structural feature parameters to construct a global three-dimensional model. This invention improves the accuracy of vascular structure recognition and visualization analysis capabilities, providing a direct visual representation of vascular structures.

Description

Technical Field

[0001] This invention relates to the field of image recognition, and in particular to a method and apparatus for recognizing vascular structural features based on intravascular ultrasound images. Background Technology

[0002] In the long-term application of intravascular ultrasound imaging, due to the complex imaging environment, blood flow disturbances, and the characteristics of the equipment itself, the acquired images often suffer from problems such as speckle noise, ring artifacts, and guidewire obstruction. These interference factors significantly affect the identification of vascular structure boundaries. The tortuous, multi-branched, and irregular morphology of blood vessels cause significant variations between different image frames regarding the vascular lumen and adventitia, increasing the difficulty of structural feature extraction. In this situation, relying solely on a single frame image for analysis can easily lead to discontinuous or inaccurate structural identification. Existing vascular structure identification methods are mostly based on two-dimensional image processing techniques, extracting vascular contours through edge detection, threshold segmentation, or simple machine learning methods. While these methods can achieve preliminary identification of the vascular lumen to some extent, they lack modeling of the spatial relationships between multiple images. Analysis is limited to a two-dimensional plane and cannot reflect the true morphology of blood vessels in three-dimensional space, resulting in limitations in assessing the degree of vascular stenosis and lesion distribution. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a method and apparatus for identifying vascular structural features based on intravascular ultrasound images, thereby resolving at least one of the aforementioned technical problems.

[0004] To achieve the above objectives, the present invention provides a method for identifying vascular structural features based on intravascular ultrasound images, comprising the following steps: Step S1: An ultrasound catheter is introduced into the target blood vessel and a uniform retraction scan is performed along the vascular axis to obtain a standard ultrasound image sequence; Step S2: Perform pixel-level segmentation and frame-by-frame contour feature calculation based on standard ultrasound image sequences to generate contour structure feature parameters for different frames of images; Step S3: Analyze the axial position of blood vessels in the standard ultrasound image sequence to generate three-dimensional pose information of the blood vessel contour in each frame; Step S4: Based on the 3D pose information, identify key contour feature points and model discrete point clouds to construct a 3D mesh model; Step S5: Based on the contour structure feature parameters, perform surface smoothing filtering and parameter constraint adjustment on the 3D mesh model to construct a global 3D model; Step S6: Display the global 3D model with layered coloring to construct a visualized vascular structure model.

[0005] In this invention, step S1 specifically involves the following steps: An ultrasound catheter is introduced into the target blood vessel and performs a uniform retraction scan along the vessel axis to continuously acquire ultrasound image sequences of the vessel cross-section and acquisition timestamps. The ultrasound image sequence is time-synchronized and corrected according to the acquisition timestamp to generate a first image sequence; Identify speckle noise and ring artifacts in the first image sequence; perform dynamic filtering on the speckle noise and ring artifacts to generate a second image sequence; Gray-level distribution recognition is performed on the second image sequence to obtain gray-level distribution features; Based on the grayscale distribution characteristics, the guidewire-occluded area is determined and removed to generate a third image sequence; The original polar coordinates of the third image sequence are transformed into Cartesian coordinates to output a standard ultrasound image sequence.

[0006] In this invention, step S2 specifically involves the following steps: Pixel-level segmentation is performed based on standard ultrasound image sequences, and the pixel-level segmentation results are extracted. Visual contour recognition is performed based on pixel-level segmentation results to mark the vascular lumen boundary curve and the adventitia boundary curve. The vascular lumen boundary curve and the adventitia boundary curve are smoothed, and the contour features are calculated frame by frame to generate contour structure feature parameters of different frames.

[0007] In this invention, step S3 specifically involves the following steps: Calculate the retraction speed information of the uniform retraction scan; Based on the retraction speed information, the axial position of blood vessels in the standard ultrasound image sequence is analyzed to obtain the initial spatial position relationship. Image registration processing is performed on the standard ultrasound image sequence based on the initial spatial positional relationship; After image registration processing, the center point of the blood vessel lumen of the standard ultrasound image sequence is extracted; Three-dimensional analysis of blood vessel contours is performed on multiple frames of images based on the center point of the blood vessel lumen to generate three-dimensional pose information of the blood vessel contours in each frame.

[0008] In this invention, step S4 specifically involves the following steps: Key contour feature points are identified and marked based on the vascular lumen boundary curve and the adventitia boundary curve. Based on three-dimensional pose information, multiple contour feature points are uniformly mapped to generate three-dimensional point cloud data of blood vessel structure. Based on the three-dimensional point cloud data, adjacent contour sections are connected in sequence according to the axial direction of the blood vessels to construct a continuous tubular structure. Discrete point cloud modeling is performed on continuous tubular structures to construct a three-dimensional mesh model.

[0009] In this invention, step S5 specifically involves the following steps: The surface of the 3D mesh model is smoothed by filtering to obtain the first optimized 3D model. The filtered mesh model is subjected to structural anomaly identification, and anomaly points are marked; the anomaly points include holes, overlaps, and topological errors. Based on the anomalies, automatic repair and mesh reconstruction are performed to construct a second optimized 3D model. Based on the contour structure feature parameters, the local deviation of the three-dimensional parameters of the second optimized three-dimensional model is identified, and the location of the structural deviation is marked. Consistency parameter constraints are adjusted to address structural deviations, and a global 3D model is constructed.

[0010] In this invention, step S6 specifically involves the following steps: Based on pixel-level segmentation results, the global 3D model is subjected to region identification and layered coloring display to generate a layered coloring model. Based on the contour structure feature parameters, the structural parameters of the layered coloring model are located, and attribute visualization mapping is performed to construct a visualized vascular structure model.

[0011] This specification provides a vascular structure feature recognition device based on intravascular ultrasound images, used to perform the vascular structure feature recognition method based on intravascular ultrasound images as described above, including: The scanning imaging module is used to introduce an ultrasound catheter into the target blood vessel and perform a uniform retraction scan along the vascular axis to obtain a standard ultrasound image sequence. The feature calculation module is used to perform pixel-level segmentation and frame-by-frame contour feature calculation based on standard ultrasound image sequences, and generate contour structure feature parameters for different frames of images. The position analysis module is used to perform axial position analysis of blood vessels in standard ultrasound image sequences and generate three-dimensional pose information of blood vessel contours for each frame. The modeling module is used to identify key contour feature points and model discrete point clouds based on 3D pose information, and to construct a 3D mesh model. The constraint adjustment module is used to perform surface smoothing filtering and parameter constraint adjustment on the three-dimensional mesh model based on the contour structure feature parameters, and to construct a global three-dimensional model. The visualization module is used to display the global 3D model in layers with coloring, and to build a visualized vascular structure model.

[0012] The beneficial effects of this invention are as follows: By establishing a stable correspondence between catheter displacement and acquisition time through uniform retraction scanning, each frame of ultrasound image possesses a clear axial spatial position. This avoids scale inconsistencies caused by fluctuations in retraction speed, improving the spatial continuity and consistency of the image sequence. Pixel-level segmentation transforms ultrasound grayscale information into structured regions, making the boundaries of the lumen, adventitia, and plaques clearly discernible. Geometric feature parameters are obtained through contour extraction, enabling quantitative expression of image information. An axial spatial mapping relationship is established between retraction speed and time information, giving the two-dimensional image sequence three-dimensional spatial positioning capabilities. Center point extraction describes changes in vascular orientation. This effectively solves the problem of lacking spatial coordinates in images, while reducing the impact of catheter offset and vascular curvature, making the expression of vascular spatial structure more continuous and realistic. By mapping two-dimensional contours to three-dimensional space to form a point cloud and constructing a mesh model, the transformation of vascular structure from discrete cross-sections to continuous three-dimensional morphology is achieved. This process preserves the local geometrical change characteristics of the blood vessel, giving the overall structure complete spatial expression capabilities, thus forming a computable and analyzable three-dimensional vascular geometric model. By introducing contour structure feature parameters to constrain and adjust the 3D model, the model's geometry is made consistent with actual measurement results, effectively correcting local noise and deformation errors. This improves the model's quantitative accuracy, ensuring consistent spatial representation of parameters such as vessel area, diameter, and degree of stenosis, thus enhancing the model's reliability. Layered coloring distinguishes the structure of the 3D model, and combined with parameter mapping, provides a visual representation of lesion areas, making different tissue structures clearly distinguishable in space. This enhances the model's readability, allowing for a more intuitive display of vascular structures and lesion distribution, thereby improving analytical efficiency and application value. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the steps of a method for recognizing vascular structural features based on intravascular ultrasound images according to the present invention. Figure 2 This is a detailed flowchart illustrating the implementation steps of step S1. Figure 3 This is a detailed flowchart illustrating the implementation steps of step S2; Figure 4 This is a schematic diagram of a standard ultrasound image sequence; Figure 5 This is a schematic diagram of a 3D mesh model at one angle. Figure 6 This is a schematic diagram of a 3D mesh model from another angle. Figure 7 This is a schematic diagram of the visualized vascular structure model obtained after final processing. Detailed Implementation

[0014] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0015] This application provides a method and apparatus for recognizing vascular structural features based on intravascular ultrasound images. The execution entities of the method and apparatus for recognizing vascular structural features based on intravascular ultrasound images include, but are not limited to, mechanical equipment, data processing platforms, cloud server nodes, network upload devices, etc., which can be considered as general computing nodes in this application. The data processing platform includes, but is not limited to, at least one of an audio image management system, an information management system, and a cloud data management system.

[0016] Please see Figures 1 to 7 This invention provides a method for identifying vascular structural features based on intravascular ultrasound images, comprising the following steps: Step S1: An ultrasound catheter is introduced into the target blood vessel and a uniform retraction scan is performed along the vascular axis to obtain a standard ultrasound image sequence; Step S2: Perform pixel-level segmentation and frame-by-frame contour feature calculation based on standard ultrasound image sequences to generate contour structure feature parameters for different frames of images; Step S3: Analyze the axial position of blood vessels in the standard ultrasound image sequence to generate three-dimensional pose information of the blood vessel contour in each frame; Step S4: Based on the 3D pose information, identify key contour feature points and model discrete point clouds to construct a 3D mesh model; Step S5: Based on the contour structure feature parameters, perform surface smoothing filtering and parameter constraint adjustment on the 3D mesh model to construct a global 3D model; Step S6: Display the global 3D model with layered coloring to construct a visualized vascular structure model.

[0017] In one specific embodiment, a 30 mm long coronary artery was used as the analysis object. Intravascular ultrasound (IVUS) was employed for continuous retraction scanning imaging. The mechanical retraction speed was output by the catheter drive system, and this speed was measured in real time by an encoder built into the device. The encoder resolution was 0.01 mm, and the instantaneous velocity sequence was calculated by combining the output displacement difference with a timestamp. Let the retraction speed be... Its calculation source is the ratio of the difference between adjacent sampling positions to the time difference, i.e.

[0018] in Originating from the displacement change of the duct encoder, Derived from the system clock (40 fps frame cycle), the average stable speed is used in this embodiment; the frame rate is set by the ultrasound host.

[0019] Its corresponding time resolution is provided by a hardware crystal oscillator, therefore the inter-frame time interval is...

[0020] Axial spatial resolution is determined by both velocity and time, and is as follows: ; This parameter represents the actual spatial interval between two adjacent frames along the blood vessel axis, derived from the fusion calculation of the device calibration system and real-time speed.

[0021] Based on the above parameters, a spatial-temporal mapping relationship is established for each frame of the image. The timestamp is directly output by the system clock and is denoted as […]. Its definition originates from the recording of the trigger signal time, therefore:

[0022] The corresponding spatial position is obtained by integrating the retreat velocity, and is expressed as:

[0023] in This represents the actual physical location of the i-th frame image along the blood vessel axis, and is used as the spatial constraint input for subsequent 3D reconstruction.

[0024] In the image denoising stage, the original ultrasound image is generated from the transducer echo signal, and its physical model is the multiplicative speckle noise model. This noise originates from the ultrasound coherence imaging mechanism and is expressed as:

[0025] in This is a real tissue reflection signal, originating from the difference in acoustic impedance between vascular tissues. The speckle noise term is generated by coherent superposition interference. To recover the true signal, Lee filtering is used, and its calculation is based on local statistics (obtained from a 7×7 image pixel window), expressed as:

[0026] in It is a local mean (calculated directly from the pixels of the window). For local variance, This represents the system noise variance (obtained from equipment calibration experiments). For example, it can be statistically obtained within a specific local window. , , , The output will be:

[0027] This value represents the true echo intensity of the tissue after noise reduction.

[0028] During guidewire removal, the global grayscale statistics are derived from the pixel histogram of the entire image and are defined as follows:

[0029] The statistical results are as follows:

[0030] The guide wire region has a high-brightness, low-texture structure, and its judgment threshold is determined by statistical distribution as follows:

[0031] Simultaneously, considering the local variance, it is calculated using a 3×3 window, with a threshold of 15, when the following conditions are met:

[0032] If it is found to be an area blocked by the guidewire, it will be removed.

[0033] Pixel-level segmentation results are derived from a trained semantic segmentation network (input is a 512×512 standard image, output is a three-class probability map). The lumen region is learned from low-intensity blood echo regions, the adventitia region from mid-to-high echo boundary structures, and the plaque region from abnormally high or low echo regions. The lumen area is calculated based on the segmentation results, derived from the conversion between pixel count and spatial calibration coefficients.

[0034] in The physical scale corresponding to a pixel is obtained from probe calibration. In this embodiment, it is taken as 0.01 mm / pixel. For example, the calculation is as follows:

[0035] The equivalent diameter is obtained by inversely calculating the area of ​​the circle model:

[0036] The degree of stenosis is derived from the statistical mean of a normal blood vessel database, obtained by averaging healthy samples.

[0037] in ,get

[0038] The three-dimensional center point is calculated from the centroid of the segmented contour, and is defined as follows:

[0039] And combined with axial position This constitutes a three-dimensional coordinate system. The tangent vector originates from the difference between the center points of adjacent frames.

[0040] This vector is used to describe the direction of blood vessel movement.

[0041] A 3D point cloud is obtained from 2D contour points through spatial transformation. Its coordinate mapping comes from the combination of a local coordinate system (N, B, T) and the center point.

[0042] With a 5° interval, the angle sampling is derived from the contour discretization, therefore each section has 72 points, for a total of [number of points].

[0043] Model optimization employs neighborhood Laplace smoothing, with the neighborhood mean calculated from topological connectivity: ; in Derived from the average value of first-order adjacent vertices, The curvature constraint is derived from discrete geometric differential calculations:

[0044] The restrictions are:

[0045] The error was obtained by comparing the model with IVUS measurements:

[0046] For example:

[0047] The visualization mapping function is derived from the normalized results of the structural parameters, and its color is defined as: assigning color according to the segmented regions: Lumen: RGB(255,0,0); Outer membrane: RGB(0,255,0); Plaque: RGB(0,0,255); Transparency is derived from area normalization:

[0048] Finally, a three-dimensional global vascular structure model with a layered color model and quantitative annotation of stenosis is output, realizing a complete mapping chain from the original IVUS signal to an interpretable three-dimensional anatomical structure.

[0049] Furthermore, in another specific embodiment, the superficial femoral artery, approximately 50 mm in length, is used as the analysis object. This vessel is a peripheral artery, characterized by its tortuous course, high plaque load, and complex echogenic structure. Therefore, higher requirements are placed on noise suppression and structure preservation during image processing. Intravascular ultrasound imaging is performed by a catheter retraction system. The catheter position is obtained in real time by an encoder with a resolution of 0.005 mm, used to record axial displacement changes. The retraction speed is set by the device control system and a stable value is obtained through multi-frame averaging. In this embodiment, the retraction speed is...

[0050] The imaging frame rate is set by the IVUS host hardware.

[0051] The time sampling interval can be obtained from this. ; The axial spatial spacing is determined by both velocity and time.

[0052] This parameter is used to establish the spatial index relationship of the image sequence along the blood vessel axis.

[0053] In the image preprocessing stage, the original ultrasound images are first subjected to time synchronization correction. The timestamp is directly recorded by the system's crystal oscillator clock, and it is expressed as follows:

[0054] The spatial position is obtained by encoder integration.

[0055] This process achieves a one-to-one mapping between image sequences and the actual axial positions of blood vessels, thereby eliminating spatial errors caused by minor fluctuations in retraction speed.

[0056] In the noise reduction process, the ultrasound images are described using a multiplicative speckle model.

[0057] in This represents the actual tissue echo signal, which originates from the difference in acoustic impedance of vascular tissue. This represents speckle noise, which originates from the interference superposition effect in the ultrasonic coherent imaging mechanism. To address this noise characteristic, an adaptive filtering method based on local statistics is employed. Its core calculation form is as follows:

[0058] in and The results were obtained from statistics of local windows (9×9 pixels). The system noise variance obtained from equipment calibration. The purpose of this process is to suppress random speckle while preserving the vascular boundary structure.

[0059] In handling strong reflection interference (guide wire and calcified areas), threshold determination is performed using global grayscale statistics. The global mean and variance are calculated from the histogram of the entire image.

[0060] Thresholds are set based on statistical characteristics.

[0061] When the pixel grayscale is higher than the threshold and the local texture variance is low, the region is identified as a non-vascular structure and removed, thereby reducing subsequent segmentation errors.

[0062] In the structure extraction stage, a semantic segmentation model is used to perform pixel-level classification on the standardized images, outputting three categories of labels: lumen region, external elastic membrane region, and plaque region. The pixel spatial scale was obtained from the probe calibration experiment.

[0063] The lumen area is calculated based on the segmentation results, and is defined as the product of the number of pixels and the square of the spatial scale:

[0064] For example, statistics show that... ,but

[0065] The equivalent diameter is calculated using a circular equivalent model.

[0066] The degree of stenosis was obtained by comparing with a health reference database.

[0067] in ,get

[0068] Used to characterize the degree of moderate to severe stenosis.

[0069] In the 3D reconstruction stage, the spatial position of each frame of the image is calculated from the centroid of the contour.

[0070] And combined with axial position These points form a three-dimensional coordinate system. The tangent vector is obtained by differencing adjacent center points.

[0071] This vector is used to describe changes in blood vessel orientation and serves as a directional constraint for 3D reconstruction.

[0072] In the point cloud construction process, a local coordinate system (normal N, binormal B, tangential T) is used to map the two-dimensional contour points to three-dimensional space, which is expressed as follows:

[0073] Each cross-section is discretized using 90 points (angle step size 4°), thereby improving the accuracy of irregular structure representation. The total point cloud size is approximately...

[0074] During the model optimization phase, Laplacian smoothing is performed on the point cloud.

[0075] in Derived from the average value of the first-order neighborhood vertices, Used to control smoothing intensity. Curvature constraint is defined as...

[0076] This is used to ensure the geometric continuity of blood vessels. Model error is obtained by comparing with IVUS measurement results.

[0077] For example, calculations yield...

[0078] This indicates that the deviation between the model and the actual structure is within an acceptable range.

[0079] During the visualization phase, color mapping is performed based on the degree of narrowness.

[0080] For example

[0081] Transparency is determined by area ratio

[0082] This enables a unified visual representation of vascular structure, lesion severity, and spatial location.

[0083] In this embodiment, see Figure 2 The diagram below illustrates the detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include: An ultrasound catheter is introduced into the target blood vessel and performs a uniform retraction scan along the vessel axis to continuously acquire ultrasound image sequences of the vessel cross-section and acquisition timestamps. The ultrasound image sequence is time-synchronized and corrected according to the acquisition timestamp to generate a first image sequence; Identify speckle noise and ring artifacts in the first image sequence; perform dynamic filtering on the speckle noise and ring artifacts to generate a second image sequence; Gray-level distribution recognition is performed on the second image sequence to obtain gray-level distribution features; Based on the grayscale distribution characteristics, the guidewire-occluded area is determined and removed to generate a third image sequence; The original polar coordinates of the third image sequence are transformed into Cartesian coordinates to output a standard ultrasound image sequence.

[0084] In this embodiment, an ultrasound catheter is introduced into the target blood vessel and performs a uniform retraction scan along the vessel axis to continuously acquire ultrasound image sequences of the vessel cross-section and corresponding acquisition timestamps. The retraction scan is controlled by the catheter drive system, and its retraction speed is denoted as v. This speed is directly set by the device's mechanical control module and obtained in real time. In this embodiment, v is preferably set to 0.7 mm / s, with an allowable fluctuation range controlled within ±0.05 mm / s to ensure axial spatial sampling consistency. The image acquisition frame rate is denoted as f and is controlled by the internal clock of the ultrasound imaging host. In this embodiment, it is set to 40 fps. Each frame is acquired by the probe transducer, and a unique timestamp ti is automatically generated by the system. The timestamp originates from the device's high-precision crystal oscillator clock module, with a time resolution of 1 ms. This acquisition process yields the original ultrasound image sequence and its corresponding time series data.

[0085] All image frames are sorted according to timestamp ti, and a fixed time interval Δt = 1 / f is used as the standard time base, where Δt is 25 ms in this embodiment. Due to slight fluctuations in mechanical retraction speed and non-uniform sampling between frames during actual acquisition, a time resampling strategy is used for correction. Missing frames are reconstructed between adjacent frames using linear interpolation, ensuring the sequence maintains an evenly spaced distribution on the time axis. The interpolation is based on constraints including grayscale similarity between adjacent frames and spatial structural continuity, thus avoiding structural jumps. After this processing, a first image sequence is obtained, which satisfies the requirements of temporal consistency and sequence continuity.

[0086] The speckle noise and ring artifacts in the first image sequence are identified and dynamically filtered to generate a second image sequence. Speckle noise is defined as multiplicative random noise caused by coherent ultrasonic echo interference, statistically characterized by random fluctuations in local grayscale. Ring artifacts are defined as periodic or quasi-periodic brightness enhancement bands centered on the image center, originating from probe rotation errors or system structure reflections. A local window analysis method is used in the spatial domain, with a window size of 7×7 pixels. The local mean and local variance are calculated to assess noise intensity. The filtering intensity is adaptively adjusted based on the local variance; a larger variance results in a higher filtering weight, achieving noise suppression under edge protection conditions. For ring artifacts, the image is transformed to a polar coordinate system, and the grayscale energy distribution curve is calculated in the radial direction. If a stable abnormal peak appears within a fixed radius range, it is identified as a ring artifact region, and radial low-pass filtering and attenuation functions are used for suppression. The final output is the second image sequence, which effectively reduces noise interference while preserving structural boundaries.

[0087] For each frame of the image, a grayscale histogram is calculated, with the grayscale range defined as 0 to 255, where 0 represents an echo-free region and 255 represents a highly reflective region. Further, within a local window (set to 15×15 pixels), the local grayscale mean μ and standard deviation σ are calculated to characterize the local texture complexity and structural variation. The global grayscale distribution function H(g) is calculated to describe the intensity distribution characteristics of the entire image. This grayscale distribution characteristic is directly calculated from the image acquired by the device, without relying on external prior data; its parameters are derived from the pixel statistics of the original image.

[0088] The guide wire occlusion region is defined as a region that simultaneously satisfies both high brightness and low texture complexity. The high brightness condition is that the pixel grayscale value is greater than the global mean μ_global plus twice the standard deviation 2σ_global, which is calculated from the overall grayscale statistics of the current frame. The low texture condition is that the local variance is less than a preset threshold σ_th, which is set to 15 in this embodiment. Regions meeting both conditions are identified as guide wire occlusion regions and marked using a binary mask. Processing methods for this region include zeroing out pixels or interpolation filling based on neighboring pixels to avoid structural breaks. After processing, a third image sequence is generated, which enhances structural continuity and effectively removes guide wire interference.

[0089] Each pixel is represented in polar coordinates by a radial distance *r* and an angle *θ*, where *r* is determined by the radial sampling depth of the ultrasound probe (in this embodiment, the radial resolution is set to 0.01 mm), and *θ* is determined by the probe's rotation sampling angle (in this embodiment, the angular resolution is 0.5 degrees). Polar coordinates are converted to Cartesian coordinates using geometric mapping, and bilinear interpolation is used to reconstruct grayscale values ​​for non-integer coordinates to ensure image continuity and smooth boundaries. Simultaneously, a combination of mirror expansion and zero-padding is used in image boundary regions to avoid missing edge information. The final output is a standard ultrasound image sequence, which is spatially unified in Cartesian coordinates and possesses scale consistency and structural analyzability.

[0090] In this embodiment, see Figure 3 The diagram below illustrates the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include: Pixel-level segmentation is performed based on standard ultrasound image sequences, and the pixel-level segmentation results are extracted. Visual contour recognition is performed based on pixel-level segmentation results to mark the vascular lumen boundary curve and the adventitia boundary curve. The vascular lumen boundary curve and the adventitia boundary curve are smoothed, and the contour features are calculated frame by frame to generate contour structure feature parameters of different frames.

[0091] The pixel-level segmentation results include the lumen region, the external elastic membrane region, and the patch region; the contour structure feature parameters include the lumen area, the equivalent diameter, and the degree of local narrowing.

[0092] Standard ultrasound image sequences were used as input data, having undergone polar to Cartesian coordinate transformation and possessing uniform spatial resolution. Pixel-level segmentation was achieved using a deep learning-based semantic segmentation model, preferably an improved U-Net network. The encoder extracted multi-scale features, while the decoder restored spatial resolution and generated pixel-level classification results. Input images were uniformly resized to 512×512 pixels, and grayscale values ​​were normalized to the 0–1 range to eliminate brightness differences between devices. The model output consisted of three-channel segmentation results, corresponding to the lumen region, external elastic membrane region, and plaque region, with each category represented as a probability map with probability values ​​ranging from 0 to 1. The loss function used during training was a weighted combination of Dice Loss and cross-entropy loss, with Dice Loss weight set to 0.7 and cross-entropy weight set to 0.3. The training data comes from an expert-annotated IVUS image dataset, with annotations based on vascular anatomy standards. The lumen region is defined as a blood-filled area, the external elastic lamina as the boundary between the media and adventitia, and plaque regions as areas of abnormal echo enhancement or hypoechoic structures. The final output is a pixel-level segmentation matrix M(x,y), where each pixel corresponds to one of three class labels.

[0093] The improved U-Net network described above uses an improved network structure based on the publicly available semantic segmentation model U-Net to process standard ultrasound image sequences. The input data is an IVUS image sequence that has undergone polar-to-Cartesian coordinate transformation and has a uniform spatial resolution. The input images are uniformly adjusted to 512×512 pixels, and the grayscale values ​​are normalized to 0–1 to reduce brightness differences between different devices. The network adopts a symmetrical encoder-decoder structure. The encoder extracts multi-scale features through multi-layer convolution and downsampling operations. A residual connection mechanism is preferably introduced during high-level feature extraction to enhance the expression of deep features; this mechanism can refer to the structure of the publicly available ResNet model. The decoder restores spatial resolution through layer-by-layer upsampling and feature fusion, and combines skip connections to achieve the fusion of low-level detail information and high-level semantic information. To improve the recognition accuracy of the vascular lumen boundary and the external elastic membrane boundary, an attention mechanism is introduced during the feature fusion process between the encoder and decoder. The corresponding structure can refer to Attention U-Net to enhance the response to the target region and suppress background noise interference. The network output is mapped to a three-channel probability map using a 1×1 convolution, corresponding to the lumen region, external elastic membrane region, and plaque region, respectively. The probability values ​​for each channel range from 0 to 1 and are normalized using the Softmax function. During training, a weighted combination of Dice Loss and cross-entropy loss is used as the optimization objective, with Dice Loss weight set to 0.7 and cross-entropy weight set to 0.3 to balance class imbalance and pixel-level classification accuracy. The training data comes from an IVUS image dataset annotated by experts according to vascular anatomy standards. The lumen region is defined as the blood-filled area, the external elastic membrane region as the boundary structure between the media and adventitia, and the plaque region as an area of ​​abnormal echo enhancement or hypoechoic structure. The final output is a pixel-level segmentation matrix M(x,y), where each pixel is assigned one of three labels, providing a foundation for subsequent extraction of vascular lumen boundary curves and adventitia boundary curves, as well as contour structure feature calculation.

[0094] The segmentation results are binarized to separate the lumen region from the background region, generating a binary mask for the lumen and an independent binary mask for the external elastic lamina region. A boundary extraction algorithm is used to detect the contours of each mask category, preferably employing an improved Marching Squares algorithm or an OpenCV contour extraction method. Contour points are defined as the set of pixels in the binary image whose grayscale values ​​change between 0 and 1, and each contour point is represented by two-dimensional coordinates (x, y). For the lumen boundary curve, it is defined as the boundary between the lumen region and the vessel wall region; for the adventitia boundary curve, it is defined as the structural boundary between the external elastic lamina and the peripheral tissue. To improve contour stability, the initial contour points are filtered for connectivity, retaining only the largest connected regions as valid contours, thereby eliminating isolated small regions caused by noise. Finally, two sets of structural curves are output, representing the lumen boundary curve and the adventitia boundary curve, respectively, and stored in an ordered manner according to the frame sequence.

[0095] The extracted contour point sequence for each frame is smoothed using a parametric curve fitting method, preferably cubic B-spline curves or Savitzky-Golay filtering. The sliding window length is set to 11 sampling points to ensure local structural smoothness while avoiding overfitting. After smoothing, the contour points are resampled to ensure a uniform spatial distribution of point spacing. The sampling interval is set to 0.1 mm to improve the stability of geometric calculations.

[0096] In terms of feature calculation, three core structural parameters are calculated for each frame of the image. The first is the lumen area, defined as the area of ​​the closed region enclosed by the lumen boundary curve. Its calculation is based on the polygon area integration method, obtained directly from the contour point coordinates, and the unit is square millimeters (mm). 2 The second is the equivalent diameter, defined as the diameter of a circle equivalent to the current lumen area. It is calculated by mapping the lumen area to the diameter of a circle with an equal area, and is used to characterize the degree of vascular dilation or contraction. The third is the degree of local stenosis, defined as the relative difference between the lumen area of ​​the current frame and the standard area of ​​a reference healthy blood vessel. The reference area is derived from the average lumen area of ​​a disease-free region distal to the same vessel or a standard value obtained from a database. This indicator is used to quantify the degree of vascular stenosis, with a value ranging from 0 to 1; a higher value indicates a more severe degree of stenosis.

[0097] In this embodiment, step S3 includes the following steps: Calculate the retraction speed information of the uniform retraction scan; Based on the retraction speed information, the axial position of blood vessels in the standard ultrasound image sequence is analyzed to obtain the initial spatial position relationship. Image registration processing is performed on the standard ultrasound image sequence based on the initial spatial positional relationship; After image registration processing, the center point of the blood vessel lumen of the standard ultrasound image sequence is extracted; Three-dimensional analysis of blood vessel contours is performed on multiple frames of images based on the center point of the blood vessel lumen to generate three-dimensional pose information of the blood vessel contours in each frame.

[0098] In this embodiment, the retraction speed information is acquired in real time by an ultrasonic catheter driving system. This driving system includes a stepper motor control module and a displacement feedback module, wherein the displacement feedback module is implemented using a grating ruler or encoder, with a preferred resolution of 0.01 mm. The retraction speed v is defined as the amount of displacement change of the catheter in the axial direction of the blood vessel per unit time, and its calculation method is based on the spatial displacement difference corresponding to consecutive timestamps. The sampling time interval Δt is determined by the imaging frame rate, preferably set to 25 ms corresponding to 40 fps. By performing differential calculation on the catheter displacement between consecutive frames, the instantaneous velocity sequence v(t) is obtained, and further smoothed by a moving average filter. The filter window length is set to 5 frames to eliminate velocity fluctuations caused by mechanical jitter. Finally, a stable retraction speed information v is output, the value range of which is usually controlled between 0.5–1.0 mm / s. In this embodiment, 0.7 mm / s is preferred as the nominal speed, and its small fluctuation curve over time is recorded as the basis for subsequent spatial modeling.

[0099] Each frame of a standard ultrasound image sequence is bound to its corresponding timestamp t_i, and the axial displacement is accumulated using the pullback velocity v_i to obtain the initial spatial coordinates z_i of each frame in the vascular axial direction. Specifically, the first frame is used as the reference zero point, and the position of subsequent frames is represented by the accumulated displacement; that is, z_i is the integral result of the velocity and time interval of the previous i frames. To improve robustness, linear interpolation compensation is applied to the velocity sequence to handle sampling gaps or velocity micro-judders. The spatial unit is uniformly set to millimeters (mm), and scale correction is performed using equipment calibration parameters. The calibration source is the catheter factory calibration data, with an error controlled within ±2%. Finally, a one-to-one correspondence is established between the image sequence and the axial spatial position, thus constructing the initial spatial position relationship model.

[0100] Image registration is performed on standard ultrasound image sequences based on initial spatial relationships. Due to the influence of vascular curvature, catheter offset, and cardiac cycle, non-rigid spatial misalignment may exist between adjacent frames, necessitating image registration correction. The preferred registration method is a non-rigid registration algorithm based on mutual information (MI), combined with a B-spline deformation model for local deformation correction. Using z_i provided by the initial spatial relationships as the initial alignment constraint, adjacent frames are iteratively optimized to maximize structural similarity. The similarity metric is determined statistically from the image grayscale distribution, and mutual information is calculated based on joint probability distribution estimation. The number of registration iterations is set to 10–20, stopping when the mutual information gain between adjacent frames is less than 0.01. The final output is a spatially aligned image sequence that satisfies structural continuity and spatial consistency in the axial direction.

[0101] After image registration, the center points of the blood vessel lumens in the standard ultrasound image sequence are extracted. The lumen region in each frame is binary segmented, and its geometric center point is calculated. The center point is defined as the geometric mean of the coordinates of all pixels in the lumen region, i.e., the average position of the x and y coordinates of all pixels. To improve the stability of the center point, morphological opening and closing operations are performed on the segmentation results, and the structuring element size is set to 3×3 pixels to eliminate boundary noise and the influence of small holes. In the presence of local irregular shapes, the minimum circumcircle center is introduced as an auxiliary correction mechanism to avoid eccentricity errors. Finally, the center point coordinates c_i(x_i, y_i) for each frame are obtained, and a center point trajectory sequence is formed in chronological order. This trajectory is used for subsequent 3D spatial modeling.

[0102] Using the center point sequence as a spatial reference, and combining it with the aforementioned axial position z_i, the two-dimensional contour is mapped to a three-dimensional spatial coordinate system, constructing a three-dimensional point set of (x, y, z). Here, x and y originate from the contour point coordinates of the current frame, and z originates from the axial spatial position of the corresponding frame. To describe the spatial pose of the blood vessel, a local coordinate system composed of a local tangent vector, a normal vector, and a binormal vector is introduced. This coordinate system is obtained by differential calculation of continuous center points, where the tangent vector is obtained by normalizing the difference between adjacent center points. Based on this local coordinate system, each frame's contour is spatially rotated and aligned to determine its true pose in three-dimensional space. Finally, the three-dimensional pose information of each frame's blood vessel contour is output, including spatial position coordinates, direction vectors, and local geometric pose parameters.

[0103] In this embodiment, step S4 includes the following steps: Key contour feature points are identified and marked based on the vascular lumen boundary curve and the adventitia boundary curve. Based on three-dimensional pose information, multiple contour feature points are uniformly mapped to generate three-dimensional point cloud data of blood vessel structure. Based on the three-dimensional point cloud data, adjacent contour sections are connected in sequence according to the axial direction of the blood vessels to construct a continuous tubular structure. Discrete point cloud modeling is performed on continuous tubular structures to construct a three-dimensional mesh model.

[0104] In this embodiment, the vascular lumen boundary curve and adventitia boundary curve of each frame are uniformly discretized, representing the continuous curve as an ordered discrete point sequence. The sampling interval between points is preferably set to 0.1 mm to ensure consistent spatial resolution. Key feature points are defined as representative points that can characterize the geometric changes of the curve, including curvature extrema, inflection points, and boundary abrupt change points. Curvature calculation is based on the three-point neighborhood method, that is, a local arc is constructed by three adjacent discrete points and the curvature value is calculated. The curvature threshold is determined by statistically analyzing the curvature distribution of the entire contour, and the mean curvature plus twice the standard deviation is taken as the judgment threshold. Feature points are extracted independently for the adventitia and vascular lumen curves, and a distance constraint condition is introduced, that is, the minimum distance between adjacent feature points is set to 0.5 mm to avoid redundancy caused by excessively dense feature points. All identified key points record their two-dimensional coordinates (x, y) and their curve category, and include curvature values ​​as attribute information, thereby forming a multi-level contour feature point set. This set is directly calculated from the images acquired by the device, without relying on external prior models.

[0105] The 3D pose information is obtained from the preceding steps, including the spatial position corresponding to each frame of the image. and local coordinate system direction parameters (tangent vector) Normal vector binormal vector For each contour feature point (x, y), it is normalized to a local coordinate system with the center point of the lumen as the origin, and then mapped to a three-dimensional spatial coordinate system through a coordinate transformation matrix. This transformation process consists of a rotation matrix and a translation vector, where the rotation matrix is ​​constructed from the three-directional vectors of the local coordinate system, and the translation vector is determined by the position of the center point in three-dimensional space. The point obtained after spatial mapping is defined as the basic unit of the three-dimensional point cloud P(x, y, z), where z is provided by the axial position of the corresponding frame. An independent subset of point clouds is generated for each frame, and they are summarized according to the time series to form a complete three-dimensional point cloud dataset of the vascular structure. The point cloud density is determined by the contour sampling interval and the frame interval, with the axial frame interval set to 0.7 mm (corresponding to the combination of pullback speed and frame rate) to ensure the continuity and spatial consistency of the three-dimensional structure in the axial direction.

[0106] All point clouds are sorted according to the axial coordinate z, and adjacent frame point clouds are defined as a set of continuous sections. Then, a topological correspondence is established between adjacent sections, i.e., based on angle parameters or the nearest neighbor distance principle, each point in the previous section is connected to the nearest point in the next section. The angle parameter is defined as the polar angle θ with the centerline as the reference, ranging from 0 to 360 degrees, and is matched at equal angular intervals. In this embodiment, the angle sampling step size is set to 5 degrees to ensure structural uniformity. A triangulation or quadrilateral mesh connection framework is constructed through this connection relationship, thus forming a preliminary tubular geometry. A smoothing constraint is introduced during the construction process, i.e., the change in the center distance between adjacent sections must not exceed 0.2 mm, to avoid abrupt changes in local structure. Finally, a tubular structure model continuously extending along the blood vessel axis is formed, which maintains topological connectivity and geometric continuity.

[0107] The tubular structure surface was resampled to convert the irregular point cloud into a uniformly distributed set of discrete points. The sampling interval was set to 0.2 mm in the surface direction and 0.7 mm in the axial direction to match the original frame interval. Surface fitting was performed using either Poisson reconstruction or Ball-Pivoting algorithms. The Poisson reconstruction method generates continuous surfaces by solving implicit functions and is suitable for noisy medical point cloud data. Normal consistency constraints were set during the reconstruction process, and the normal vectors were derived from the normal components in the aforementioned local coordinate system. To ensure consistency in surface orientation, the generated initial mesh model undergoes topology optimization, including hole filling, non-manifold edge processing, and self-intersection detection. Hole filling is performed using a local reconstruction method based on neighborhood interpolation. The final output is a standard 3D mesh model, which consists of a set of vertices, edges, and faces, and represents the intraluminal and adventitia structures of the blood vessel in the form of a triangular mesh, achieving a complete 3D geometric representation of the blood vessel.

[0108] The aforementioned three-dimensional mesh model is essentially a medical geometric reconstruction model. It is a parametric three-dimensional tubular geometric model based on intravascular ultrasound (IVUS) image sequence reconstruction. Its core is not a learning model, but is composed of the following three types of objects: spatial point set, topological connection relationship, and continuous surface mesh.

[0109] In this embodiment, the specific steps of step S5 are as follows: The surface of the 3D mesh model is smoothed by filtering to obtain the first optimized 3D model. The filtered mesh model is subjected to structural anomaly identification, and anomaly points are marked; the anomaly points include holes, overlaps, and topological errors. Based on the anomalies, automatic repair and mesh reconstruction are performed to construct a second optimized 3D model. Based on the contour structure feature parameters, the local deviation of the three-dimensional parameters of the second optimized three-dimensional model is identified, and the location of the structural deviation is marked. Consistency parameter constraints are adjusted to address structural deviations, and a global 3D model is constructed.

[0110] In this embodiment, the input 3D mesh model consists of a vertex set V, an edge set E, and a face set F, where the vertices are derived from the previous vascular point cloud reconstruction results. Due to unavoidable discrete sampling errors and local noise during 3D reconstruction, the model surface typically exhibits high-frequency geometric perturbations, requiring smoothing to improve geometric continuity. A combination of Laplacian smoothing and Taubin non-shrinking smoothing is used, where Laplacian smoothing eliminates sharp local noise, while Taubin filtering avoids overall model shrinkage. The smoothing iteration count is set to 10–20, with a neighborhood weight coefficient of 0.5 for each iteration to control the smoothing intensity. The neighborhood is defined as the set of first-order topologically adjacent vertices, i.e., vertices directly connected to the current vertex via edges. The vertex update process is based on neighborhood mean offset calculation, thereby gradually reducing local surface curvature fluctuations. Curvature constraints are introduced during the filtering process, limiting local curvature to below 0.02 mm. -1 The smoothing intensity is reduced to preserve the details of the vascular structure. The final output is a first optimized 3D model, which significantly reduces surface high-frequency noise while maintaining the overall structural morphology.

[0111] The filtered mesh model is subjected to structural anomaly identification, and anomaly points are marked. Anomalies include holes, overlaps, and topological errors. Topological consistency is checked on the first optimized 3D model by analyzing the mesh boundary loop structure to determine the presence of hole regions. A hole is defined as a non-closed boundary loop with more than 3 edges that cannot form a valid facet. Overlapping regions are detected by calculating the angle between the normal vectors of triangular facets; an overlap anomaly is identified when the angle between the normal vectors of adjacent facets is less than 5 degrees and the spatial overlap ratio exceeds 30%. Topological errors are determined by checking the Euler characteristic number; a topological anomaly is considered to exist when the model does not satisfy V-E+F=2 (simply connected condition). The number of vertices V, edges E, and faces F are directly obtained from the mesh structure without external assumptions. Spatial index structures (such as KD-trees) are used to accelerate neighborhood queries during anomaly detection, reducing the detection complexity from O(n^2) to O(n^2). 2 The value is reduced to O(n log n). The final output outlier set A = {( , typei)}, where ∈{holes, overlaps, topological errors}, and their spatial locations are marked and stored.

[0112] For the void region, a local surface reconstruction method based on boundary loops is adopted. Filling patches are generated by parametric interpolation of the void boundary points, with cubic spline curves used for interpolation to ensure boundary smoothness. For overlapping regions, a normal consistency adjustment method is used for correction, i.e., the normal direction of the overlapping patches is recalculated, and patch splitting and reconstruction are performed on areas with conflicting normal directions. For topological errors, a global repair method based on Poisson reconstruction is adopted, which refits the local geometry through implicit functions to restore topological consistency. During the repair process, a maximum repair radius of 2mm is set to limit the impact of local repair and avoid global structural shift. After mesh reconstruction, the repaired region and the original region are merged at the boundary using a weighted average method to ensure a smooth transition. The final output is a second optimized 3D model, which shows significant improvements in both topological integrity and geometric continuity.

[0113] The contour structural feature parameters include lumen area, equivalent diameter, and local stenosis degree. These parameters are derived from the previous two-dimensional contour calculation results and extended to three-dimensional space through axial mapping. In the second optimized three-dimensional model, the model is segmented along the axial direction, with each segment corresponding to one frame of image structure. The local cross-section of the three-dimensional model is matched and compared with the corresponding contour structural feature parameters. The lumen area deviation is defined as the relative error between the three-dimensional cross-sectional area and the two-dimensional measured area, with an error threshold set at 10%. The equivalent diameter deviation is calculated using a circular equivalent model, and an abnormal area is marked when the deviation exceeds 0.15 mm. The local stenosis degree deviation is determined by comparative analysis with a reference healthy blood vessel database, which is derived from the statistical mean of normal blood vessel samples. Through the above multi-parameter joint judgment mechanism, a set of structural deviation locations is generated and spatially labeled in the three-dimensional model.

[0114] A global optimization constraint model is introduced for the marked structural deviation regions, with geometric consistency and feature consistency as objective functions. Geometric consistency constraints include local curvature continuity constraints and axial smoothness constraints, where the rate of curvature change is limited to 0.03 mm. -1 Within a certain range; the feature consistency constraint requires that the adjusted lumen area deviates from the original feature parameters by no more than 5%. The optimization method adopts a global adjustment strategy based on weighted least squares, iteratively updating the vertex positions of the deviation region while keeping the non-deviation region unchanged or finely adjusted. A neighborhood propagation mechanism is introduced during the optimization process to gradually attenuate the adjustment effect and ensure the stability of the global structure. The number of iterations is set to 15–30 times, and the calculation stops when the overall error converges to below the threshold of 0.01. Finally, a global 3D model is constructed, which meets the unified constraint requirements in terms of geometric structure, topological structure, and feature parameter consistency, realizing a high-fidelity 3D representation of the vascular structure.

[0115] In this embodiment, the specific steps of step S6 are as follows: Based on pixel-level segmentation results, the global 3D model is subjected to region identification and layered coloring display to generate a layered coloring model. Based on the contour structure feature parameters, the structural parameters of the layered coloring model are located, and attribute visualization mapping is performed to construct a visualized vascular structure model.

[0116] In this embodiment, the input pixel-level segmentation results include three types of label information: lumen region, external elastic membrane region, and plaque region. These labels originate from the output of the preceding deep learning semantic segmentation model, with each pixel corresponding to a unique category identifier. Simultaneously, the global 3D model consists of a set of 3D mesh vertices and a set of faces, and topology repair and structural consistency optimization have been completed. To achieve the mapping relationship between the 2D segmentation results and the 3D model, a correspondence between 2D pixel coordinates and 3D spatial coordinates is established. This correspondence is jointly determined by the 3D pose information and axial position parameters from the preceding steps, where the axial position is derived from the integral calculation result of the pullback velocity. During the mapping process, the nearest neighbor projection method is used to project the 3D model surface points onto the corresponding 2D segmentation image plane, and the 3D mesh faces are classified and labeled according to the pixel category of the projected point. The classification rule is defined as follows: when the centroid projection position of a face falls in the lumen region, it is marked as an internal lumen region; when it falls in the external elastic membrane region, it is marked as a vessel wall region; and when it falls in the plaque region, it is marked as a lesion region. After region identification, different regions are assigned preset color mapping functions. Lumen regions are represented by a highly transparent blue with a transparency set to 0.3; adventitia regions are represented by green with a transparency set to 0.5; and plaque regions are highlighted in red with a transparency set to 0.8 to enhance lesion visibility. The color mapping is based on the RGBA color model, where the R, G, and B values ​​are determined by a preset standard color table, thereby generating a layered coloring model that spatially represents the layered structure of tissues.

[0117] Based on the contour structural feature parameters, the layered coloring model is used to locate structural parameters and perform attribute visualization mapping to construct a visualized vascular structure model. The contour structural feature parameters include lumen area, equivalent diameter, and local stenosis degree. These parameters originate from the preceding two-dimensional contour calculation module and are extended to the global model space coordinate system through three-dimensional axial mapping. First, the layered coloring model is axially segmented, with each segment corresponding to one frame of ultrasound image and its corresponding three-dimensional cross-sectional position. The segment spacing is determined by the retraction speed and imaging frame rate, and the axial resolution is set to 0.7 mm. The structural feature parameters of each segment are bound to the corresponding three-dimensional cross-section to achieve spatial positioning of structural parameters. The lumen area is located within the projection range of the three-dimensional cross-section of the lumen region, the equivalent diameter is calculated through the center point of the cross-section geometric fitting, and the local stenosis degree is mapped to the color gradient intensity parameter.

[0118] In the attribute visualization mapping process, structural parameters are converted into visualization variables. The lumen area is mapped to local transparency changes on the model surface; smaller areas have lower transparency to enhance visual prominence. The equivalent diameter is mapped to local cross-sectional dimension annotations and displayed as dynamically superimposed labels. Local stenosis is mapped to color heatmap values, with a value range of 0–1 mapped to a red intensity gradient through linear normalization, from light red to dark red indicating progressively worsening stenosis. To ensure visualization continuity, cubic interpolation smoothing is applied to adjacent cross-sectional parameters, ensuring a continuous spatial transition of parameter changes. The interpolation step size is set to 0.2 mm. Finally, a visualized vascular structure model is constructed. This model simultaneously expresses anatomical structures, tissue layers, and quantitative parameter distributions in three-dimensional space, achieving a unified expression of structural and functional information, and can be used for vascular lesion analysis and auxiliary diagnostic display.

[0119] In this embodiment, a vascular structure feature recognition device based on intravascular ultrasound images is provided, used to perform the vascular structure feature recognition method based on intravascular ultrasound images as described above, including: The scanning imaging module is used to introduce an ultrasound catheter into the target blood vessel and perform a uniform retraction scan along the vascular axis to obtain a standard ultrasound image sequence. The feature calculation module is used to perform pixel-level segmentation and frame-by-frame contour feature calculation based on standard ultrasound image sequences, and generate contour structure feature parameters for different frames of images. The position analysis module is used to perform axial position analysis of blood vessels in standard ultrasound image sequences and generate three-dimensional pose information of blood vessel contours for each frame. The modeling module is used to identify key contour feature points and model discrete point clouds based on 3D pose information, and to construct a 3D mesh model. The constraint adjustment module is used to perform surface smoothing filtering and parameter constraint adjustment on the three-dimensional mesh model based on the contour structure feature parameters, and to construct a global three-dimensional model. The visualization module is used to display the global 3D model in layers with coloring, and to build a visualized vascular structure model.

[0120] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0121] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein are implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for identifying a vascular structure feature based on an intravascular ultrasound image, characterized by, Includes the following steps: Step S1: An ultrasound catheter is introduced into the target blood vessel and a uniform retraction scan is performed along the vascular axis to obtain a standard ultrasound image sequence; Step S2: Perform pixel-level segmentation and frame-by-frame contour feature calculation based on standard ultrasound image sequences to generate contour structure feature parameters for different frames of images; Step S3: Analyze the axial position of blood vessels in the standard ultrasound image sequence to generate three-dimensional pose information of the blood vessel contour in each frame; Step S4: Based on the 3D pose information, identify key contour feature points and model discrete point clouds to construct a 3D mesh model; Step S5: Based on the contour structure feature parameters, perform surface smoothing filtering and parameter constraint adjustment on the 3D mesh model to construct a global 3D model; Step S6: Display the global 3D model with layered coloring to construct a visualized vascular structure model.

2. The intravascular ultrasound image-based blood vessel structure feature recognition method according to claim 1, characterized by, The specific steps of step S1 are as follows: An ultrasound catheter is introduced into the target blood vessel and performs a uniform retraction scan along the vessel axis to continuously acquire ultrasound image sequences of the vessel cross-section and acquisition timestamps. The ultrasound image sequence is time-synchronized and corrected according to the acquisition timestamp to generate a first image sequence; Identify speckle noise and ring artifacts in the first image sequence; perform dynamic filtering on the speckle noise and ring artifacts to generate a second image sequence; Gray-level distribution recognition is performed on the second image sequence to obtain gray-level distribution features; Based on the grayscale distribution characteristics, the guidewire-occluded area is determined and removed to generate a third image sequence; The original polar coordinates of the third image sequence are transformed into Cartesian coordinates to output a standard ultrasound image sequence.

3. The intravascular ultrasound image-based blood vessel structure feature recognition method according to claim 2, characterized by, The specific steps of step S2 are as follows: Pixel-level segmentation is performed based on standard ultrasound image sequences, and the pixel-level segmentation results are extracted. Visual contour recognition is performed based on pixel-level segmentation results to mark the vascular lumen boundary curve and the adventitia boundary curve. The vascular lumen boundary curve and the adventitia boundary curve are smoothed, and the contour features are calculated frame by frame to generate contour structure feature parameters of different frames.

4. The intravascular ultrasound image-based blood vessel structure feature recognition method according to claim 3, characterized by, The pixel-level segmentation results include the lumen region, the external elastic membrane region, and the patch region; the contour structure feature parameters include the lumen area, the equivalent diameter, and the degree of local narrowing.

5. The intravascular ultrasound image-based vascular structure feature recognition method of claim 3, wherein, Step S3 is as follows: Calculate the retraction speed information of the uniform retraction scan; Based on the retraction speed information, the axial position of blood vessels in the standard ultrasound image sequence is analyzed to obtain the initial spatial position relationship. Image registration processing is performed on the standard ultrasound image sequence based on the initial spatial positional relationship; After image registration processing, the center point of the blood vessel lumen of the standard ultrasound image sequence is extracted; Three-dimensional analysis of blood vessel contours is performed on multiple frames of images based on the center point of the blood vessel lumen to generate three-dimensional pose information of the blood vessel contours in each frame.

6. The intravascular ultrasound image-based vascular structure feature recognition method of claim 5, wherein, The specific steps of step S4 are as follows: Key contour feature points are identified and marked based on the vascular lumen boundary curve and the adventitia boundary curve. Based on three-dimensional pose information, multiple contour feature points are uniformly mapped to generate three-dimensional point cloud data of blood vessel structure. Based on the three-dimensional point cloud data, adjacent contour sections are connected in sequence according to the axial direction of the blood vessels to construct a continuous tubular structure. Discrete point cloud modeling is performed on continuous tubular structures to construct a three-dimensional mesh model.

7. The method for identifying vascular structure features based on intravascular ultrasound images according to claim 6, characterized in that, The specific steps of step S5 are as follows: The surface of the 3D mesh model is smoothed by filtering to obtain the first optimized 3D model. The filtered mesh model is subjected to structural anomaly identification, and anomaly points are marked; the anomaly points include holes, overlaps, and topological errors. Based on the anomalies, automatic repair and mesh reconstruction are performed to construct a second optimized 3D model. Based on the contour structure feature parameters, the local deviation of the three-dimensional parameters of the second optimized three-dimensional model is identified, and the location of the structural deviation is marked. Consistency parameter constraints are adjusted to address structural deviations, and a global 3D model is constructed.

8. The method for identifying vascular structural features based on intravascular ultrasound images according to claim 7, characterized in that, The specific steps of step S6 are as follows: Based on pixel-level segmentation results, the global 3D model is subjected to region identification and layered coloring display to generate a layered coloring model. Based on the contour structure feature parameters, the structural parameters of the layered coloring model are located, and attribute visualization mapping is performed to construct a visualized vascular structure model.

9. A device for recognizing vascular structural features based on intravascular ultrasound images, characterized in that, The method for performing vascular structure feature recognition based on intravascular ultrasound images as described in claim 1 includes: The scanning imaging module is used to introduce an ultrasound catheter into the target blood vessel and perform a uniform retraction scan along the vascular axis to obtain a standard ultrasound image sequence. The feature calculation module is used to perform pixel-level segmentation and frame-by-frame contour feature calculation based on standard ultrasound image sequences, and generate contour structure feature parameters for different frames of images. The position analysis module is used to perform axial position analysis of blood vessels in standard ultrasound image sequences and generate three-dimensional pose information of blood vessel contours for each frame. The modeling module is used to identify key contour feature points and model discrete point clouds based on 3D pose information, and to construct a 3D mesh model. The constraint adjustment module is used to perform surface smoothing filtering and parameter constraint adjustment on the three-dimensional mesh model based on the contour structure feature parameters, and to construct a global three-dimensional model. The visualization module is used to display the global 3D model in layers with coloring, and to build a visualized vascular structure model.

Images