An automatic measurement method for intima-media thickness based on ultrasonic image blood vessel diameter detection

Abstract

The application discloses an automatic intima-media thickness measurement method based on ultrasonic image blood vessel diameter detection, relates to the technical field of medical image processing, and comprises the following steps: ultrasonic image enhancement, noise suppression and texture reservation are balanced through weighted fusion and a residual enhancement unit; adaptive blood vessel wall edge optimization, a blood vessel is segmented by using an improved U-Net network and an edge is dynamically corrected; reference line positioning, a local coordinate system is constructed based on a segmentation result and a reference line is optimized; parameter measurement and result output, a blood vessel diameter and an intima-media thickness are measured along the reference line, and a labeled image and a data report are generated; by means of multi-scale feature fusion, dynamic edge correction and a multi-level verification mechanism, the application remarkably improves blood vessel wall segmentation precision and positioning precision of the measurement reference line, and meanwhile, in combination with a natural language generation technology, structured annotation is generated, thereby providing an efficient and accurate tool for clinical diagnosis.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to an automatic measurement method for intima-media thickness based on ultrasound image-based blood vessel diameter detection. Background Technology

[0002] In the field of ultrasound imaging for vascular diameter detection and intima-media thickness measurement, existing technologies have several shortcomings that urgently need to be addressed. First, traditional edge detection algorithms, such as the Canny operator, exhibit significant noise sensitivity when processing ultrasound images. Ultrasound images typically contain speckle noise, which interferes with the accuracy of edge detection algorithms, leading to misjudgments of vessel wall boundaries. Such misjudgments not only affect the visualization of vascular structures but may also lead to inaccurate subsequent measurements, thereby impacting the reliability of clinical diagnosis.

[0003] Secondly, existing single-scale convolutional neural networks struggle to simultaneously capture both the macroscopic morphological features and microscopic texture features of the vessel wall when processing vascular images. This insufficient feature fusion limits the network's comprehensive understanding of the vascular structure, thereby affecting the accuracy and continuity of vessel wall segmentation.

[0004] Finally, subjective bias exists when manually setting the region of interest (ROI), which directly affects the positioning accuracy of the baseline for measuring vessel diameter. Due to inconsistencies in manual operation, the repeatability and accuracy of measurement results are difficult to guarantee, thus limiting the application scope and effectiveness of ultrasound images in the diagnosis of vascular diseases.

[0005] To address the aforementioned shortcomings, a technical solution is provided. Summary of the Invention

[0006] The purpose of this invention is to solve the problems of traditional edge detection algorithms being sensitive to noise, insufficient feature fusion of single-scale convolutional neural networks, and subjective bias in manually set measurement baselines in the prior art. In response, this invention proposes an automatic measurement method for intima-media thickness based on ultrasound image vascular diameter detection.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] An automatic method for measuring intima-media thickness based on ultrasound imaging for detecting blood vessel diameter includes the following steps:

[0009] T1: Ultrasound image enhancement, which processes the original ultrasound image through weighted fusion and residual enhancement units to balance noise suppression and texture preservation, and extract deep vascular features;

[0010] T2: Adaptive vessel wall edge optimization, constructing an improved U-Net network to segment vessels, dynamically correcting edges, and generating smooth and continuous vessel intima-endothelium boundaries;

[0011] T3: Baseline positioning, calculate relevant data based on segmentation results, construct a local coordinate system, and optimize the baseline through a multi-level verification mechanism;

[0012] T4: Parameter measurement and result output, measure vessel diameter and intima-media thickness along the baseline, and generate labeled images and data reports;

[0013] T5: Annotation generation and data association, detection of abnormal regions, generation of semantic annotations, implementation of multi-terminal adaptive output, and support for model iteration.

[0014] Furthermore, the specific steps of T1 are as follows:

[0015] Ultrasound image enhancement processes the original ultrasound image through weighted fusion and residual enhancement units to balance noise suppression and texture preservation, and extract deep vascular features;

[0016] First, ultrasound image enhancement processing is performed. A dual-channel input layer is constructed, and the original ultrasound image and the preprocessed image after nonlocal mean filtering are weighted and fused with weight coefficients α=0.6 and β=0.4 to generate an initial feature map to balance noise suppression and texture preservation. Here, α is the fusion weight coefficient of the original ultrasound image channel, and β is the fusion weight coefficient of the preprocessed image channel after nonlocal mean filtering.

[0017] Subsequently, a residual enhancement unit was designed to extract deep vascular features step by step through 5 cascaded residual blocks. Each residual block consists of a 3×3 convolutional layer, a batch normalization layer, and a LeakyReLU activation function, and residual learning is achieved through skip connections.

[0018] To fuse multi-scale features, a dilated convolutional layer with a dilation rate of 2 is inserted after the second residual block to capture the macroscopic morphology of blood vessels. Specifically, a 7×7 receptive field is used to capture the macroscopic shape. A dilated convolutional layer with a dilation rate of 4 is inserted after the fourth residual block to enhance the microscopic edge response. Specifically, a 15×15 receptive field is used to enhance the microscopic edges. Finally, a 1×1 convolutional layer compresses the stitched multi-scale features into a single channel, outputting a denoised and enhanced image, improving the peak signal-to-noise ratio (PSNR) by ≥4dB. PSNR is an evaluation metric for image enhancement quality, and the formula is: ,in is the maximum pixel value of the image, and MSE is the mean square error.

[0019] Furthermore, the specific operation steps of T2 are as follows:

[0020] An improved U-Net segmentation network was constructed, with a spatial pyramid pooling layer added at the end of the encoder. 1×1, 3×3, 5×5 and global pooling features were fused to enhance context awareness. A channel attention mechanism was introduced in the decoder to optimize skip connection feature fusion, and a weighting mechanism was used to improve the segmentation accuracy of blood vessel wall edges.

[0021] Dynamic edge correction is performed on the preliminary segmentation results. First, an adaptive morphological closing operation is used. The size of the structuring element is dynamically calculated according to the local blood vessel diameter, using the formula: k=2×⌊d / 3⌋+1. The process is iterated twice to fill the holes and smooth the burrs, where d is the local blood vessel diameter and k is the size of the structuring element in the morphological closing operation.

[0022] Subsequently, an energy function was established along the blood vessel axis. ,in This is the weight of the gradient strength term, reflecting the contribution of the edge gradient magnitude to the energy function, and its value is 0.5; The weight of the curvature constraint term penalizes abrupt changes in boundary curvature, with a value of 0.3; The weight of the distance penalty term constrains the rate of change of the distance between adjacent boundary points, with a value of 0.2. These are gradient strength, curvature constraint, and distance penalty term, respectively. The minimum energy path is solved by combining the graph cut algorithm, and finally, the discrete boundary points are fitted by cubic B-spline curves to generate smooth and continuous inocular and endovascular boundaries.

[0023] Furthermore, the weighting mechanism in T2 is as follows:

[0024] b1: Determine the importance weight of different vascular structures, analyze the criticality of different vascular structures in detecting intima-media thickness and diagnosing diseases, and set corresponding importance weights for different structures in combination with clinical needs;

[0025] b2: Encoder feature map weighting. In the encoder, starting from the input layer, each convolutional or pooling layer outputs feature maps of different levels. These feature maps are weighted according to the determined importance weights of different blood vessel structures.

[0026] b3: Weighted processing of skip connection feature maps in the decoder. Skip connections exist in the decoder, which fuse the feature map of the corresponding level of the encoder with the feature map of the current level of the decoder. Before fusion, the feature maps from the skip connections are weighted according to the importance of different blood vessel structures.

[0027] b4: Weighted summation and feature fusion: After weighting the feature maps of each level of the encoder and the skip connection feature maps of the decoder, when fusing pooling features of different scales at the spatial pyramid pooling layer at the end of the encoder, the weighted feature maps are weighted and summed; in the decoder, the skip connection feature maps optimized by the channel attention mechanism are also weighted and summed.

[0028] Furthermore, the specific operation steps of T3 are as follows:

[0029] Based on the segmentation results, a distance transformation map is calculated. The center line of a single-pixel wide blood vessel is extracted using the Zhang-Suen thinning algorithm. The tangent direction is sampled every 0.1 mm along the center line and calculated. A local coordinate system that can be rotated in the normal direction is constructed to ensure that the baseline is perpendicular to the direction of the blood vessel.

[0030] The baseline is optimized through a multi-level verification mechanism. Primary verification removes outliers where the angle between the normal and tangent is less than 85°. Secondary verification calculates the coefficient of variation (CV) of the distance between adjacent normals. When CV≥15%, interpolation point supplementation operation is triggered. The three-level verification checks the symmetry of the baseline on both sides, i.e., the error <8%. For the verification failure area, the normal direction is recalculated using a 3×3 sliding window. Finally, the measurement baseline that matches the blood vessel geometry is output.

[0031] Furthermore, the specific operation steps of T4 are as follows:

[0032] Perform parameter measurements and output results, scan and locate the intersection of the intima and epicardium of the blood vessel along the baseline normal direction, and calculate the average diameter. ,in Let i be the coordinates of the outer membrane boundary of the i-th sampling point; Let be the coordinates of the intima boundary at the i-th sampling point; N is the total number of sampling points along the baseline; and D is the average diameter of the blood vessel.

[0033] Grayscale profiles are extracted from areas extending 1 mm on both sides of the baseline. The intima-media interface is determined by the dynamic thresholding method, and the average intima-media thickness (IMT) is calculated. Finally, an annotated image of the superimposed vessel boundary, baseline, and measurement points is generated, and a data report containing diameter, IMT statistics, and distribution curves is output.

[0034] Furthermore, the specific operation steps of T5 are as follows:

[0035] q1: Abnormal region detection and annotation triggering, based on clinical guidelines defining abnormal thresholds, including: IMT abnormalities: maximum value >1.0mm or local thickness mutation >0.2mm; diameter abnormalities: difference between adjacent points >15% or symmetry error at bifurcation >10%;

[0036] During the parameter measurement phase, data is scanned synchronously, and triggering conditions include: IMT value exceeding the threshold triggering thickening risk annotation; diameter mutation triggering stenosis warning annotation; plaque detection triggering plaque warning annotation.

[0037] q2: Semantic annotation generation and association: Utilizing a natural language generation model, inputs include: anomaly type, location coordinates, measurement values, and historical data comparison; output is: structured annotation text.

[0038] In the visualization interface, clicking on the annotation text will display the following linked images: Related Images: Locate the abnormal area and highlight it, specifically with a flashing red box; Related Charts: Pop up the IMT thickness distribution curve of the area and compare it with historical data; External Knowledge Links: Automatically retrieve relevant clinical guidelines and display the corresponding entries.

[0039] Q3: Feedback-driven annotation optimization: Add an accuracy rating button next to each annotation, ranging from 1 to 5 stars, to support doctors in manually modifying annotation content;

[0040] Model iteration mechanism: Collect feedback data, including scores and modification records, build a training set and fine-tune the NLG model, including: initialization using BERT-base and adjustment in combination with medical corpus; loss function: cross-entropy loss and human rating weights;

[0041] Q4: Multi-terminal adaptive output: Desktop: Three-column layout, including images, annotations, and charts, with drag-and-drop panel resizing support; Mobile: Collapsible design, click on annotations to expand details, with gesture-based image zooming support; Print adaptation includes: automatically generating simplified annotations, including bold keywords and arrows indicating image positions; matching A4 paper column layout.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] (1) This invention processes the original image through ultrasound image enhancement and residual enhancement units, effectively balancing noise suppression and texture preservation, and solving the problem of misjudgment of blood vessel wall boundaries caused by noise interference in traditional methods. At the same time, through multi-scale feature fusion, it can simultaneously extract the macroscopic morphology and microscopic texture features of the blood vessel wall, overcoming the shortcomings of single-scale convolutional neural networks in feature fusion, and significantly improving the accuracy and continuity of blood vessel wall segmentation; the adaptive blood vessel wall edge optimization and multi-level verification mechanism are used to optimize the baseline, effectively avoiding subjective bias when manually setting the ROI region, improving the positioning accuracy of the measurement baseline, and thus providing a reliable guarantee for the accurate measurement of blood vessel diameter and intima-media thickness;

[0044] (2) In the parameter measurement and result output stage, this invention locates the intersection of the intima and media by scanning along the normal direction of the baseline, and determines the intima-media interface by combining the dynamic threshold method. This enables accurate calculation of the vessel diameter and intima-media thickness, and generates labeled images and data reports. Simultaneously, it also features abnormal region detection and semantic annotation functions, automatically triggering abnormal annotations based on clinical guidelines and generating structured annotation text through a natural language generation model, providing doctors with intuitive diagnostic references. The annotation optimization mechanism allows doctors to manually modify the annotation content and continuously optimize the natural language generation model through a feedback-driven model iteration mechanism, further improving the accuracy and practicality of the annotations.

[0045] (3) This invention supports multi-terminal adaptive output, including desktop, mobile and print adaptation, which can meet the needs of different usage scenarios; the desktop adopts a three-column layout and supports dragging to adjust the panel size, making it convenient for doctors to view images, annotations and charts; the mobile adopts a folding design and supports gesture zooming of images, making it convenient for doctors to view key information anytime and anywhere; the print adaptation function automatically generates simplified annotations to ensure that the text and images correspond correctly and meet the needs of clinical report printing. Attached Figure Description

[0046] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0047] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0048] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] It should be understood that the terms “comprising” and “including” used in this disclosure and claims indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0050] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this disclosure. As used in this disclosure and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this disclosure and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0051] like Figure 1 As shown, an automatic method for measuring intima-media thickness based on ultrasound image-based blood vessel diameter detection includes the following steps:

[0052] Ultrasound image enhancement processes the original ultrasound image through weighted fusion and residual enhancement units to balance noise suppression and texture preservation, and extract deep vascular features;

[0053] First, ultrasound image enhancement processing is performed. A dual-channel input layer is constructed, and the original ultrasound image and the preprocessed image after nonlocal mean filtering (search window 21×21, similarity window 5×5) are weighted and fused with weight coefficients α=0.6 and β=0.4 to generate an initial feature map to balance noise suppression and texture preservation. Here, α is the fusion weight coefficient of the original ultrasound image channel; β is the fusion weight coefficient of the preprocessed image channel after nonlocal mean filtering (NL-Means). Then, a residual enhancement unit is designed to extract deep blood vessel features step by step through 5 cascaded residual blocks. Each residual block consists of a 3×3 convolutional layer, a batch normalization layer, and a LeakyReLU activation function (negative slope 0.2, where 0.2 is the slope parameter of the negative value region in the activation function, controlling the proportion of negative value information retained). Residual learning is achieved through skip connections.

[0054] To fuse multi-scale features, a dilated convolutional layer with a dilation rate of 2 is inserted after the second residual block to capture the macroscopic morphology of blood vessels (7×7 receptive field, capturing macroscopic morphology). A dilated convolutional layer with a dilation rate of 4 is inserted after the fourth residual block to enhance the microscopic edge response (15×15 receptive field, enhancing microscopic edges). Finally, a 1×1 convolutional layer compresses the stitched multi-scale features into a single channel, outputting a denoised and enhanced image, improving the peak signal-to-noise ratio (PSNR) by ≥4dB. PSNR is an evaluation metric for image enhancement quality; a higher value indicates better denoising performance. The formula is: ,in is the maximum pixel value of the image, and MSE is the mean square error.

[0055] Adaptive vessel wall edge optimization: An improved U-Net network is constructed to segment blood vessels, dynamically correct edges, and generate smooth and continuous intima-endothelium boundaries.

[0056] An improved U-Net segmentation network was constructed, with a spatial pyramid pooling layer (SPP) added at the end of the encoder. This layer fuses 1×1, 3×3, 5×5, and global pooling features to enhance context awareness. A channel attention mechanism (SEBlock) is introduced in the decoder to optimize skip connection feature fusion. A weighted mechanism is used to improve the accuracy of blood vessel wall edge segmentation; the specific process is as follows:

[0057] Determining the importance weights of different vascular structures: Analyzing the criticality of different vascular structures (such as the intima, media, and adventitia) in detecting intima-media thickness and diagnosing diseases, and combining this with clinical needs, assigning corresponding importance weights to different structures. For example, since intima-media thickness is highly correlated with disease, features related to the intima-media are given higher weights; adventitia features are relatively less critical and are given lower weights. Encoder feature map weighting: In the encoder, starting from the input layer, each convolutional or pooling layer outputs feature maps of different levels. Based on the determined importance weights of different vascular structures, these feature maps are weighted. For example, if a certain level of feature map contains more intima features, and the intima features have a higher weight, a larger weighting coefficient is assigned to that feature map. The weighting coefficient is multiplied by the feature map elements to complete the weighting processing of the encoder's feature maps at each level. Decoder skip connection feature map weighting: The decoder contains skip connections, which fuse the feature maps of the corresponding levels of the encoder with the feature maps of the current level of the decoder. Before fusion, the feature maps from skip connections are weighted according to the importance of different vascular structures. Taking a skip connection near the decoder output as an example, if the feature map of this connection mainly contains features of the outer membrane and part of the middle membrane, appropriate weighting coefficients are assigned to the outer membrane and middle membrane features according to pre-set weights to weight the feature map elements. Weighted summation and feature fusion: After weighting the feature maps of each level of the encoder and the skip connection feature maps of the decoder, when fusing pooling features of different scales at the spatial pyramid pooling layer at the end of the encoder, the weighted feature maps are weighted and summed. In the decoder, the skip connection feature maps optimized by the channel attention mechanism are also weighted and summed to achieve feature fusion, making the network more focused on key blood vessel wall edge information and improving segmentation accuracy. For example, for the weighted feature map Ac of a certain level of the encoder and the weighted feature map Bc of another level, as well as the weighted feature map Cc from the skip connection in the decoder, their fusion weighting coefficients are set to rt1, rt2, and rt3, respectively. They are weighted and summed using the formula: RY=Ac×rt1+Bc×rt2+Cc×rt3, where RY is the fused feature map.

[0058] Dynamic edge correction is performed on the preliminary segmentation results. First, an adaptive morphological closing operation is used, with the structuring element size dynamically calculated based on the local vessel diameter using the formula: k = 2 × ⌊d / 3⌋ + 1. This is iterated twice to fill holes and smooth burrs, where d is the local vessel diameter and k is the size of the structuring element in the morphological closing operation. Subsequently, an energy function is established along the vessel axis. ,in This is the weight of the gradient strength term, reflecting the contribution of the edge gradient magnitude to the energy function, and its value is 0.5; The weight of the curvature constraint term penalizes abrupt changes in boundary curvature, with a value of 0.3; The weight of the distance penalty term constrains the rate of change of the distance between adjacent boundary points, with a value of 0.2. These are gradient strength, curvature constraint, and distance penalty term, respectively. The minimum energy path is solved by combining the graph cut algorithm, and finally, the discrete boundary points are fitted by cubic B-spline curves to generate smooth and continuous inocular and endovascular boundaries.

[0059] Baseline positioning involves calculating relevant data based on segmentation results, constructing a local coordinate system, and optimizing the baseline through a multi-level verification mechanism.

[0060] Based on the segmentation results, a distance transformation map is calculated. The Zhang-Suen thinning algorithm is used to extract the centerline of a single-pixel wide blood vessel. Samples are taken every 0.1 mm along the centerline, and the tangent direction is calculated. A rotatable local coordinate system is constructed to ensure the baseline is perpendicular to the vessel orientation. A multi-level validation mechanism is used to optimize the baseline: primary validation removes outliers with an angle <85° between the normal and tangent; secondary validation calculates the coefficient of variation (CV) of the distance between adjacent normals. (Interpolation point supplementation operation is triggered when CV≥15%), three-level verification checks the symmetry of the baseline on both sides (error<8%), and a 3×3 sliding window is used to recalculate the normal direction for the verification failure area, and finally outputs the measurement baseline that strictly matches the vascular geometry.

[0061] Parameter measurement and result output: Measure blood vessel diameter and intima-media thickness along the baseline, and generate labeled images and data reports;

[0062] Perform parameter measurements and output results, scan and locate the intersection of the intima and epicardium of the blood vessel along the baseline normal direction, and calculate the average diameter. ,in Let i be the coordinates of the outer membrane boundary of the i-th sampling point; Let be the coordinates of the intima boundary at the i-th sampling point; N be the total number of sampling points along the baseline; D be the average diameter of the vessel; extract grayscale profiles in areas extending 1 mm on both sides of the baseline, determine the intima-media interface using the dynamic thresholding method, and calculate the average intima-media thickness (IMT). Finally, generate an annotated image of the vessel boundary, baseline, and measurement points, and output a data report containing diameter, IMT statistics, and distribution curves.

[0063] Annotation generation is linked to data, abnormal regions are detected, semantic annotations are generated, multi-terminal adaptive output is achieved, and model iteration is supported;

[0064] Abnormal area detection and annotation triggering: Abnormal thresholds are defined based on clinical guidelines. IMT abnormality: maximum value >1.0mm or local thickness mutation >0.2mm; Diameter abnormality: difference between adjacent points >15% or symmetry error at bifurcation >10%; Data is scanned synchronously during the parameter measurement phase. Triggering conditions include: IMT value exceeding the threshold → triggering "thickening risk" annotation; diameter mutation → triggering "stenosis indication" annotation; plaque detection → triggering "plaque warning" annotation.

[0065] Semantic annotation generation and association: Utilizing a Natural Language Generation (NLG) model, the inputs are: abnormality type, location coordinates, measurement value, and historical data comparison; the output is: structured annotation text, such as: Thickening risk: "The maximum IMT at the axial position of 12-15mm is 1.2mm (exceeding the threshold by 0.3mm), and it is recommended to combine it with hemodynamic assessment." Plaque warning: "A calcified plaque (size 3.2×1.5mm) was detected at the bifurcation of the left carotid artery, with an IMT compensation value of 0.8mm." In the visualization interface, clicking on the annotation text will display the following in conjunction: Related image: Locating and highlighting the abnormal area (red flashing box); Related chart: Popping up the IMT thickness distribution curve of the area and a historical comparison chart; External knowledge links: Automatically retrieving relevant clinical guidelines (such as the "ESC Guidelines for the Management of Atherosclerosis") and displaying the corresponding entries;

[0066] Annotation optimization: Add an "Accuracy Rating" button (1 to 5 stars) next to each annotation to allow doctors to manually modify the annotation content (such as adding diagnostic opinions); Model iteration mechanism: Collect feedback data (ratings, modification records) to build a training set and fine-tune the NLG model, including: initialization using BERT-base and fine-tuning with medical corpus (PubMed summaries); Loss function: Cross-entropy loss + human rating weight (low-rated sample weight × 3);

[0067] Multi-terminal adaptive output: Desktop: Three-column layout (images, annotations, charts), supports drag-and-drop panel resizing; Mobile: Collapsible design, click on annotations to expand details, supports gesture zooming; Print adaptation includes: automatically generating simplified annotations (keywords in bold + arrows indicating image positions); matching A4 paper column layout to ensure accurate text and image correspondence.

[0068] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An automatic method for measuring intima-media thickness based on ultrasound imaging for detecting blood vessel diameter, characterized in that, Includes the following steps: T1: Ultrasound image enhancement, which processes the original ultrasound image through weighted fusion and residual enhancement units to balance noise suppression and texture preservation, and extract deep vascular features; T2: Adaptive vessel wall edge optimization, constructing an improved U-Net network to segment vessels, dynamically correcting edges, and generating smooth and continuous vessel intima-endothelium boundaries; T3: Baseline positioning, calculate relevant data based on segmentation results, construct a local coordinate system, and optimize the baseline through a multi-level verification mechanism; T4: Parameter measurement and result output, measure vessel diameter and intima-media thickness along the baseline, and generate labeled images and data reports; T5: Annotation generation and data association, detection of abnormal regions, generation of semantic annotations, implementation of multi-terminal adaptive output, and support for model iteration; The specific steps of T1 are as follows: Ultrasound image enhancement processes the original ultrasound image through weighted fusion and residual enhancement units to balance noise suppression and texture preservation, and extract deep vascular features; First, ultrasound image enhancement processing is performed. A dual-channel input layer is constructed, and the original ultrasound image and the preprocessed image after nonlocal mean filtering are weighted and fused with weight coefficients α=0.6 and β=0.4 to generate an initial feature map to balance noise suppression and texture preservation. Here, α is the fusion weight coefficient of the original ultrasound image channel, and β is the fusion weight coefficient of the preprocessed image channel after nonlocal mean filtering. Subsequently, a residual enhancement unit was designed to extract deep vascular features step by step through 5 cascaded residual blocks. Each residual block consists of a 3×3 convolutional layer, a batch normalization layer, and a LeakyReLU activation function, and residual learning is achieved through skip connections. To fuse multi-scale features, a dilated convolutional layer with a dilation rate of 2 is inserted after the second residual block to capture the macroscopic morphology of blood vessels. Specifically, a 7×7 receptive field is used to capture the macroscopic shape. A dilated convolutional layer with a dilation rate of 4 is inserted after the fourth residual block to enhance the microscopic edge response. Specifically, a 15×15 receptive field is used to enhance the microscopic edges. Finally, a 1×1 convolutional layer compresses the stitched multi-scale features into a single channel, outputting a denoised and enhanced image, improving the peak signal-to-noise ratio (PSNR) by ≥4dB. PSNR is an evaluation metric for image enhancement quality, and the formula is: ,in The maximum pixel value of the image is given by , and MSE is the mean square error. The specific operation steps of T2 are as follows: An improved U-Net segmentation network was constructed, with a spatial pyramid pooling layer added at the end of the encoder. 1×1, 3×3, 5×5 and global pooling features were fused to enhance context awareness. A channel attention mechanism was introduced in the decoder to optimize skip connection feature fusion, and a weighting mechanism was used to improve the segmentation accuracy of blood vessel wall edges. Dynamic edge correction is performed on the preliminary segmentation results. First, an adaptive morphological closing operation is used. The size of the structuring element is dynamically calculated according to the local blood vessel diameter, using the formula: k=2×⌊d / 3⌋+1. The process is iterated twice to fill the holes and smooth the burrs, where d is the local blood vessel diameter and k is the size of the structuring element in the morphological closing operation. Subsequently, an energy function was established along the blood vessel axis. ,in This is the weight of the gradient strength term, reflecting the contribution of the edge gradient magnitude to the energy function, and its value is 0.5; The weight of the curvature constraint term penalizes abrupt changes in boundary curvature, with a value of 0.3; The weight of the distance penalty term constrains the rate of change of the distance between adjacent boundary points, with a value of 0.

2. These are gradient strength, curvature constraint, and distance penalty term, respectively. The minimum energy path is solved by combining the graph cut algorithm, and finally, the discrete boundary points are fitted by cubic B-spline curves to generate smooth and continuous inocular and endovascular boundaries.

2. The automatic measurement method for intima-media thickness based on ultrasound image for detecting blood vessel diameter according to claim 1, characterized in that, The weighting mechanism in T2 is as follows: b1: Determine the importance weight of different vascular structures, analyze the criticality of different vascular structures in detecting intima-media thickness and diagnosing diseases, and set corresponding importance weights for different structures in combination with clinical needs; b2: Encoder feature map weighting. In the encoder, starting from the input layer, each convolutional or pooling layer outputs feature maps of different levels. These feature maps are weighted according to the determined importance weights of different blood vessel structures. b3: Weighted processing of skip connection feature maps in the decoder. Skip connections exist in the decoder, which fuse the feature map of the corresponding level of the encoder with the feature map of the current level of the decoder. Before fusion, the feature maps from the skip connections are weighted according to the importance of different blood vessel structures. b4: Weighted summation and feature fusion: After weighting the feature maps of each level of the encoder and the skip connection feature maps of the decoder, when fusing pooling features of different scales at the spatial pyramid pooling layer at the end of the encoder, the weighted feature maps are weighted and summed; in the decoder, the skip connection feature maps optimized by the channel attention mechanism are also weighted and summed.

3. The automatic measurement method for intima-media thickness based on ultrasound image for detecting blood vessel diameter according to claim 1, characterized in that, The specific operation steps of T3 are as follows: Based on the segmentation results, a distance transformation map is calculated. The center line of a single-pixel wide blood vessel is extracted using the Zhang-Suen thinning algorithm. The tangent direction is sampled every 0.1 mm along the center line and calculated. A local coordinate system that can be rotated in the normal direction is constructed to ensure that the baseline is perpendicular to the direction of the blood vessel. The baseline is optimized through a multi-level verification mechanism. Primary verification removes outliers where the angle between the normal and tangent is less than 85°. Secondary verification calculates the coefficient of variation (CV) of the distance between adjacent normals. When CV≥15%, interpolation point supplementation operation is triggered. The three-level verification checks the symmetry of the baseline on both sides, i.e., the error <8%. For the verification failure area, the normal direction is recalculated using a 3×3 sliding window. Finally, the measurement baseline that matches the blood vessel geometry is output.

4. The automatic measurement method for intima-media thickness based on ultrasound image for detecting blood vessel diameter according to claim 1, characterized in that, The specific operation steps of T4 are as follows: Perform parameter measurements and output results, scan and locate the intersection of the intima and epicardium of the blood vessel along the baseline normal direction, and calculate the average diameter. ,in Let i be the coordinates of the outer membrane boundary of the i-th sampling point; Let be the coordinates of the intima boundary at the i-th sampling point; N is the total number of sampling points along the baseline; and D is the average diameter of the blood vessel. Grayscale profiles are extracted from areas extending 1 mm on both sides of the baseline. The intima-media interface is determined by the dynamic thresholding method, and the average intima-media thickness (IMT) is calculated. Finally, an annotated image of the superimposed vessel boundary, baseline, and measurement points is generated, and a data report containing diameter, IMT statistics, and distribution curves is output.

5. The automatic measurement method for intima-media thickness based on ultrasound image for detecting blood vessel diameter according to claim 1, characterized in that, The specific operating steps of T5 are as follows: q1: Abnormal region detection and annotation triggering, based on clinical guidelines defining abnormal thresholds, including: IMT abnormalities: maximum value >1.0mm or local thickness mutation >0.2mm; diameter abnormalities: difference between adjacent points >15% or symmetry error at bifurcation >10%; During the parameter measurement phase, data is scanned synchronously, and triggering conditions include: IMT value exceeding the threshold triggering thickening risk annotation; diameter mutation triggering stenosis warning annotation; plaque detection triggering plaque warning annotation. q2: Semantic annotation generation and association: Utilizing a natural language generation model, inputs include: anomaly type, location coordinates, measurement values, and historical data comparison; output is: structured annotation text. In the visualization interface, clicking on the annotation text will display the following linked images: Related Images: Locate the abnormal area and highlight it, specifically with a flashing red box; Related Charts: Pop up the IMT thickness distribution curve of the area and compare it with historical data; External Knowledge Links: Automatically retrieve relevant clinical guidelines and display the corresponding entries. Q3: Feedback-driven annotation optimization: Add an accuracy rating button next to each annotation, ranging from 1 to 5 stars, to support doctors in manually modifying annotation content; Model iteration mechanism: Collect feedback data, including scores and modification records, build a training set and fine-tune the NLG model, including: initialization using BERT-base and adjustment in combination with medical corpus; loss function: cross-entropy loss and human rating weights; Q4: Multi-terminal adaptive output: Desktop: Three-column layout, including images, annotations, and charts, with drag-and-drop panel resizing support; Mobile: Collapsible design, click on annotations to expand details, with gesture-based image zooming support; Print adaptation includes: automatically generating simplified annotations, including bold keywords and arrows indicating image positions; matching A4 paper column layout.

Images