A cardiovascular atherosclerosis detection system based on data analysis

Abstract

The present application relates to the field of data analysis, and more particularly to a cardiovascular atherosclerosis detection system based on data analysis, comprising: a comprehensive encoding module generates a comprehensive multi-channel vascular ultrasound encoded image by a comprehensive encoder from an intravascular ultrasound image; a backbone network feature extraction module generates multi-scale plaque features by a C2f plaque feature weighted backbone network from the comprehensive multi-channel vascular ultrasound encoded image; a neck network feature focusing module generates fused plaque features by a neck network based on a CBAM architecture from the multi-scale plaque features; a plaque class judgment module generates a cardiovascular atherosclerosis plaque class; a plaque regression recognition module generates an intravascular ultrasound annotation image with an annotated plaque recognition box; and a backbone network, a neck network, a plaque classification head, and a plaque regression recognition head constitute a plaque recognition model of a YOLOv8 framework. The present application realizes efficient detection of atherosclerotic plaques and provides more intuitive ultrasound image data for doctors.

Description

Technical Field

[0001] This invention relates to the field of data analysis, and more particularly to a cardiovascular atherosclerosis detection system based on data analysis. Background Technology

[0002] Cardiovascular disease is one of the chronic diseases with the highest morbidity and mortality rates worldwide. One of its core pathogenic factors is coronary atherosclerosis, which is the formation of lipid plaques in the blood vessel walls. As these plaques develop, they narrow the lumen, eventually leading to serious complications such as myocardial ischemia and myocardial infarction. Early and accurate identification of the type and spatial location of atherosclerotic plaques, including lipid plaques, fibrous plaques, and calcified plaques, is a key prerequisite for reducing disease risk.

[0003] Intravascular ultrasound (IVUS), a commonly used vascular imaging technique in clinical practice, allows for the insertion of a catheter probe deep into blood vessels to generate high-resolution cross-sectional images, providing a direct view of plaque morphology and vessel wall structure, making it a core tool for atherosclerosis detection. However, in clinical practice, doctors primarily rely on experience to identify and judge plaques. They must observe images frame by frame, determine plaque type, and mark location, which is not only time-consuming and labor-intensive, resulting in single examinations taking tens of minutes, but also susceptible to subjective differences in plaque boundary judgment among different doctors, especially for early-stage, small plaques, which are prone to being missed.

[0004] Therefore, the question is how to automatically analyze intravascular ultrasound image data through algorithms to achieve efficient detection of atherosclerotic plaques and provide doctors with more intuitive ultrasound image data. Summary of the Invention

[0005] To address this, the present invention provides a data analysis-based cardiovascular atherosclerosis detection system. This system encodes intravascular ultrasound images using diffusion filters, CLBP encoding, and polar coordinate position encoding. It simultaneously encodes plaque texture details, noise suppression information, and plaque spatial distribution information, improving the completeness of feature dimensions. Furthermore, through a YOLOv8 framework recognition model based on BiFPN and an improved attention mechanism, it achieves automated detection of cardiovascular atherosclerotic plaques in intravascular ultrasound image data. This enables efficient detection of atherosclerotic plaques, providing doctors with more intuitive ultrasound image data and reducing missed detections due to subjective differences in physician experience.

[0006] To achieve the above objectives, this invention proposes a data analysis-based cardiovascular atherosclerosis detection system, comprising:

[0007] The integrated coding module is used to generate integrated multi-channel vascular ultrasound coded images by passing intravascular ultrasound images through an integrated encoder based on diffusion filter, CLBP coding and polar coordinate position coding.

[0008] The backbone network feature extraction module is used to generate multi-scale plaque features by passing the integrated multi-channel vascular ultrasound encoded image through a C2f-based plaque feature weighted backbone network.

[0009] The neck network feature focusing module is used to pass the multi-scale patch features through a neck network based on CBAM and FPN architecture to generate fused patch features;

[0010] The plaque category determination module is used to pass the fused plaque features through a plaque classification head based on a convolutional block attention mechanism to generate cardiovascular atherosclerotic plaque categories;

[0011] The plaque regression and recognition module is used to pass the fused plaque features through a plaque regression and recognition head based on deformable convolution to generate an intravascular ultrasound annotation image with labeled plaque recognition boxes;

[0012] The patch feature weighted backbone network, the neck network, the patch classification head, and the patch regression recognition head constitute the patch recognition model of the YOLOv8 framework.

[0013] Furthermore, the integrated coding module includes:

[0014] A diffusion filter unit is used to pass the image gradient values ​​of the intravascular ultrasound image through a diffusion filter to generate an ultrasound image of the inner side of the blood vessel wall;

[0015] The CLBP encoding unit is used to encode the ultrasound image of the inner side of the blood vessel wall using CLBP to generate an ultrasound-encoded image of the texture inside the blood vessel wall.

[0016] A polar coordinate encoding unit is used to encode the intravascular ultrasound image using polar coordinates to generate an ultrasound-encoded image of the blood vessel location.

[0017] The encoded image stitching unit is used to stitch together the ultrasound-coded image of the intravascular texture and the ultrasound-coded image of the vascular location to generate the integrated multi-channel vascular ultrasound-coded image.

[0018] Furthermore, the diffusion filter unit includes:

[0019] An image gradient calculation subunit is used to generate image gradient values ​​by passing the intravascular ultrasound image through the Sobel operator;

[0020] A diffusion coefficient calculation subunit is used to generate an image edge diffusion coefficient by passing the image gradient value through a diffusion coefficient calculation equation.

[0021] The divergence calculation subunit is used to generate local grayscale intensity of the image by passing the image edge diffusion coefficient and the image gradient value through the divergence operator.

[0022] The inner region of the blood vessel wall is determined by a subunit, which generates an ultrasound image of the inner side of the blood vessel wall based on a comparison of the local gray intensity and threshold of the image.

[0023] The diffusion filter includes the Sobel operator, the diffusion coefficient calculation equation, and the divergence operator.

[0024] Furthermore, the CLBP encoding unit includes:

[0025] The image grayscale encoding subunit is used to encode the ultrasound image inside the blood vessel wall using grayscale CLBP based on the grayscale of the center pixel and the average grayscale of the neighborhood, so as to generate a grayscale encoded image of the center of the blood vessel wall.

[0026] The local texture coding subunit is used to encode the ultrasound image of the inner side of the blood vessel wall using local differential CLBP encoding based on the difference between the gray levels of the center pixel and the neighboring gray levels, so as to generate a fine texture coded image of the inner side of the blood vessel wall.

[0027] An amplitude grayscale encoding subunit is used to encode the ultrasound image of the inner side of the blood vessel wall using amplitude CLBP encoding based on grayscale amplitude to generate a texture-encoded image of the inner region of the blood vessel wall.

[0028] The CLBP encoding includes grayscale CLBP encoding, local difference CLBP encoding, and amplitude CLBP encoding. The intravascular wall texture ultrasound encoded image includes the intravascular wall central grayscale encoded image, the intravascular wall fine texture encoded image, and the intravascular wall regional texture encoded image.

[0029] In particular, by using diffusion filtering, CLBP encoding, and polar coordinate position encoding, noise interference from the ultrasound image inside the blood vessel wall is removed, the inner region of the blood vessel wall is accurately located, multi-dimensional texture information is extracted, and position information that conforms to the shape of the blood vessel is generated, reducing the false negatives and false positives of the YOLOv8 plaque recognition model.

[0030] Furthermore, the backbone network feature extraction module includes:

[0031] An initial convolutional unit is used to pass the integrated multichannel vascular ultrasound encoded image through an initial convolutional layer to generate a smooth vascular ultrasound feature image;

[0032] The first C2f feature extraction unit is used to extract small-scale plaque features from the smooth blood vessel ultrasound feature image through the first C2f to generate small-scale primary image plaque features;

[0033] The second C2f feature extraction unit is used to extract medium-scale patch features from the small-scale primary image patch features through the second C2f to generate medium-scale image patch features.

[0034] The third C2f feature extraction unit is used to extract global plaque vessel wall features from the mesoscale image plaque features through the third C2f to generate global image plaque features;

[0035] The multi-scale patch features include the small-scale primary image patch features, the medium-scale image patch features, and the global image patch features. The patch feature weighted backbone network includes an initial convolutional layer, a first C2f, a second C2f, and a third C2f.

[0036] Furthermore, the neck network feature focusing module includes:

[0037] The CBAM scale fusion unit is used to perform a first upsampling on the global image patch features and the mid-scale image patch features, and then perform feature addition and first FPN fusion convolution to generate temporary convolution features. The temporary convolution features are then passed through CBAM to generate mid-level attention enhancement features.

[0038] The shallow feature fusion unit is used to combine the mid-level attention enhancement features with the small-scale primary image patch features after second upsampling through feature addition and second FPN fusion convolution to generate the fused patch features;

[0039] The neck network includes the first upsampling, the first FPN fusion convolution, the CBAM, the second upsampling, and the second FPN fusion convolution.

[0040] Furthermore, the CBAM scale fusion unit includes:

[0041] The channel attention subunit is used to generate channel-focused features by passing the temporary convolutional features through a channel attention mechanism that includes global average pooling of feature maps, global max pooling of feature maps, and a multilayer perceptron.

[0042] A spatial attention subunit is used to generate the mid-level attention enhancement features by passing the temporary convolutional features through a spatial attention mechanism with channel-dimensional global average pooling and channel-dimensional global max pooling.

[0043] The CBAM includes the channel attention mechanism and the spatial attention mechanism.

[0044] Furthermore, the patch category determination module includes:

[0045] The convolutional block attention enhancement unit is used to generate category-enhanced features by passing the fused patch features through a convolutional block attention mechanism based on global average pooling, channel global average pooling, and channel global max pooling.

[0046] A classification unit is used to generate the cardiovascular atherosclerotic plaque category by passing the category enhancement features through a SiLU-based classification convolution.

[0047] The patch classification head includes the convolutional block attention mechanism and the classification convolution.

[0048] Furthermore, the patch regression identification module includes:

[0049] A feature adaptation unit is used to pass the fused patch features through a feature adaptation convolutional layer to generate adapted patch features;

[0050] A multi-scale feature enhancement unit is used to enhance the adapted patch features through deformable convolution to generate enhanced patch features.

[0051] A regression judgment unit is used to pass the enhanced plaque features through a regression convolutional layer to generate the intravascular ultrasound annotated image;

[0052] The patch regression recognition head includes the feature-adaptive convolutional layer, the deformable convolutional layer, and the regression convolutional layer.

[0053] Furthermore, the cardiovascular atherosclerosis detection system also includes:

[0054] The training optimization module is used to optimize the training of the patch recognition model using a comprehensive loss function based on the VFL loss term and the EIoU loss term.

[0055] In particular, the backbone network with initial convolution and three-level C2f units enables refined hierarchical extraction of vascular ultrasound features. The neck network combining FPN architecture and CBAM attention mechanism enables focusing on the core features of plaques. Through the output of plaque categories and precise recognition boxes, doctors are provided with intuitive diagnostic basis for cardiovascular atherosclerotic plaques, reducing their workload and subjective misjudgment.

[0056] Compared with existing technologies, the beneficial effects of this invention are that it encodes intravascular ultrasound images through diffusion filters, CLBP encoding, and polar coordinate position encoding, which can simultaneously encode the plaque texture details, noise suppression information, and plaque spatial distribution information of the image, thereby improving the integrity of feature dimensions. Furthermore, through a YOLOv8 framework recognition model based on BiFPN and an improved attention mechanism, it achieves automated detection of cardiovascular atherosclerotic plaques in intravascular ultrasound image data, realizing efficient detection of atherosclerotic plaques, providing doctors with more intuitive ultrasound image data, and reducing missed detections caused by differences in doctors' subjective experience.

[0057] In particular, this invention removes noise interference from the ultrasound image inside the blood vessel wall through diffusion filtering, CLBP encoding, and polar coordinate position encoding, accurately locates the region inside the blood vessel wall, extracts multi-dimensional texture information, and generates position information that conforms to the shape of the blood vessel, thereby reducing the missed detections and false detections of the YOLOv8 plaque recognition model.

[0058] In particular, this invention achieves refined hierarchical extraction of vascular ultrasound features through an initial convolution and a backbone network of three-level C2f units. By combining the FPN architecture with the CBAM attention mechanism in the neck network, it achieves focus on the core features of plaques. Through the output of plaque categories and precise recognition boxes, it provides doctors with intuitive diagnostic basis for cardiovascular atherosclerotic plaques, reducing doctors' workload and subjective misjudgment. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the structure of the cardiovascular atherosclerosis detection system based on data analysis according to an embodiment of the present invention;

[0060] Figure 2 This is a flowchart illustrating the integrated encoder of the data analysis-based cardiovascular atherosclerosis detection system according to an embodiment of the present invention.

[0061] Figure 3 This is a schematic diagram of the feature-weighted backbone network of the data analysis-based cardiovascular atherosclerosis detection system according to an embodiment of the present invention.

[0062] Figure 4 This is a schematic flowchart of the neck network of the data analysis-based cardiovascular atherosclerosis detection system according to an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0064] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0065] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0066] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0067] like Figures 1 to 4 As shown, this invention provides a cardiovascular atherosclerosis detection system based on data analysis. It encodes intravascular ultrasound images using diffusion filters, CLBP encoding, and polar coordinate position encoding, simultaneously encoding plaque texture details, noise suppression information, and plaque spatial distribution information, thus improving the integrity of feature dimensions. Furthermore, through a YOLOv8 framework recognition model based on BiFPN and an improved attention mechanism, it achieves automated detection of cardiovascular atherosclerotic plaques in intravascular ultrasound image data, enabling efficient detection of atherosclerotic plaques. This provides doctors with more intuitive ultrasound image data, reducing missed detections caused by differences in doctors' subjective experience.

[0068] like Figure 1 As shown, this embodiment proposes a data analysis-based cardiovascular atherosclerosis detection system, including:

[0069] The integrated coding module is used to generate integrated multi-channel vascular ultrasound coded images by passing intravascular ultrasound images through an integrated encoder based on diffusion filter, CLBP coding and polar coordinate position coding.

[0070] The backbone network feature extraction module is used to generate multi-scale plaque features by passing the integrated multi-channel vascular ultrasound encoded image through a C2f-based plaque feature weighted backbone network.

[0071] The neck network feature focusing module is used to pass the multi-scale patch features through a neck network based on CBAM and FPN architecture to generate fused patch features;

[0072] The plaque category determination module is used to pass the fused plaque features through a plaque classification head based on a convolutional block attention mechanism to generate cardiovascular atherosclerotic plaque categories;

[0073] The plaque regression and recognition module is used to pass the fused plaque features through a plaque regression and recognition head based on deformable convolution to generate an intravascular ultrasound annotation image with labeled plaque recognition boxes;

[0074] The patch feature weighted backbone network, the neck network, the patch classification head, and the patch regression recognition head constitute the patch recognition model of the YOLOv8 framework.

[0075] In particular, the YOLOv8 architecture emphasizes modularity and flexibility, breaking down the target detection process into three independent modules: the backbone network, the neck network, and the head. Each module has a clear responsibility and standardized interfaces, making it easy to replace or modify according to task requirements. Therefore, this embodiment is based on the native YOLOv8 architecture and is adjusted according to the needs of patch detection, which features subtle characteristics, irregular shapes, and small category differences, in order to achieve high-accuracy and high-efficiency patch detection.

[0076] like Figure 2 As shown, the integrated encoding module further includes: a diffusion filter unit, used to pass the image gradient values ​​of the intravascular ultrasound image through a diffusion filter to generate an intravascular ultrasound image;

[0077] The CLBP encoding unit is used to encode the ultrasound image of the inner side of the blood vessel wall using CLBP to generate an ultrasound-encoded image of the texture inside the blood vessel wall.

[0078] A polar coordinate encoding unit is used to encode the intravascular ultrasound image using polar coordinates to generate an ultrasound-encoded image of the blood vessel location.

[0079] The encoded image stitching unit is used to stitch together the ultrasound-coded image of the intravascular texture and the ultrasound-coded image of the vascular location to generate the integrated multi-channel vascular ultrasound-coded image.

[0080] In particular, the intravascular ultrasound image is a complete intravascular ultrasound (IVUS) image including both the inner and outer sides of the vessel wall. After the image is separated into noise edges and stepped vessel wall edges by a diffusion filter, it is encoded using a complete local binary mode (CLBP) encoding algorithm to encode the image based on the vascular texture information reflected by the image grayscale values ​​of the inner side of the vessel wall. Polar coordinate encoding is used to generate a positional code that better conforms to the vascular structure. The encoding vectors of the two codes are used as data for one channel of the comprehensive multi-channel vascular ultrasound encoded image input to the plaque recognition model of the YOLOv8 framework. This reduces the burden on the plaque recognition model to learn complex plaque features from the original noisy image and achieves more accurate plaque recognition.

[0081] like Figure 2 As shown, the diffusion filter unit further includes:

[0082] An image gradient calculation subunit is used to generate image gradient values ​​by passing the intravascular ultrasound image through the Sobel operator;

[0083] A diffusion coefficient calculation subunit is used to generate an image edge diffusion coefficient by passing the image gradient value through a diffusion coefficient calculation equation.

[0084] The divergence calculation subunit is used to generate local grayscale intensity of the image by passing the image edge diffusion coefficient and the image gradient value through the divergence operator.

[0085] The inner region of the blood vessel wall is determined by a subunit, which generates an ultrasound image of the inner side of the blood vessel wall based on a comparison of the local gray intensity and threshold of the image.

[0086] The diffusion filter includes the Sobel operator, the diffusion coefficient calculation equation, and the divergence operator.

[0087] In particular, the Sobel operator is used to accurately capture the gray-level gradient changes in intravascular ultrasound images, highlighting the edge contours of the vessel wall and lumen. The diffusion intensity is dynamically adjusted through the diffusion coefficient equation, which enhances diffusion in the lumen background area with small gradient to suppress noise, and weakens diffusion in the vessel wall edge area with large gradient to protect the edge. The divergence operator integrates the gray-level change trends of local areas to generate more discriminative gray-level intensity values, so as to accurately determine the inner region of the vessel wall.

[0088] Specifically, the process of generating the ultrasound image of the inner side of the blood vessel wall can be represented as follows:

[0089]

[0090] In the formula, These represent the intravascular ultrasound image, horizontal gradient map, vertical gradient map, image gradient value, mean of image gradient values, standard deviation of image gradient values, image edge diffusion coefficient, and local grayscale intensity of the image, respectively. Represents the Sobel operators for the horizontal and vertical directions. Together, they constitute the equation for calculating the diffusion coefficient, where N represents the total number of pixels in the intravascular ultrasound image. These represent the horizontal and vertical half-pixel image edge diffusion coefficients calculated using the Sobel operator and the diffusion coefficient calculation equation, respectively. These represent the image gradient values ​​of half a pixel in the horizontal and vertical directions, calculated using the Sobel operator and the diffusion coefficient calculation equation, respectively. and The mean of the calculation results substituted into the Sobel operator and the diffusion coefficient calculation equation is used as... and ,Will and The mean of the calculation results substituted into the Sobel operator and the diffusion coefficient calculation equation is used as... and ,Will and The mean of the calculation results substituted into the Sobel operator and the diffusion coefficient calculation equation is used as... and ,Will and The mean of the calculation results substituted into the Sobel operator and the diffusion coefficient calculation equation is used as... and .

[0091] Specifically, when the local grayscale intensity of the image is greater than 0, the pixel is inside the blood vessel wall; when it is equal to 0, the pixel is at the boundary of the blood vessel wall; and when it is less than 0, the pixel is outside the blood vessel wall, so as to generate the ultrasound image of the inside of the blood vessel wall.

[0092] Furthermore, the CLBP encoding unit includes:

[0093] The image grayscale encoding subunit is used to encode the ultrasound image inside the blood vessel wall using grayscale CLBP based on the grayscale of the center pixel and the average grayscale of the neighborhood, so as to generate a grayscale encoded image of the center of the blood vessel wall.

[0094] The local texture coding subunit is used to encode the ultrasound image of the inner side of the blood vessel wall using local differential CLBP encoding based on the difference between the gray levels of the center pixel and the neighboring gray levels, so as to generate a fine texture coded image of the inner side of the blood vessel wall.

[0095] An amplitude grayscale encoding subunit is used to encode the ultrasound image of the inner side of the blood vessel wall using amplitude CLBP encoding based on grayscale amplitude to generate a texture-encoded image of the inner region of the blood vessel wall.

[0096] The CLBP encoding includes grayscale CLBP encoding, local difference CLBP encoding, and amplitude CLBP encoding. The intravascular wall texture ultrasound encoded image includes the intravascular wall central grayscale encoded image, the intravascular wall fine texture encoded image, and the intravascular wall regional texture encoded image.

[0097] In particular, CLBP encoding is used to lock in grayscale distribution differences at the regional level, establishing a basic degree of differentiation between plaques and normal tissue.

[0098] Specifically, the process of generating the central grayscale encoded image of the blood vessel wall, the fine texture encoded image of the blood vessel wall, and the regional texture encoded image of the blood vessel wall can be represented as follows:

[0099]

[0100] In the formula, These represent the central grayscale encoded image, the fine texture encoded image, and the regional texture encoded image within the blood vessel wall, respectively. This represents the first threshold function, which is the gray level of the center pixel in the neighborhood of the ultrasound image inside the blood vessel wall. Greater than the average gray level of the neighborhood When the threshold function is set to 1, it is set to 0 otherwise. P is the set number of neighborhood pixels, preferably 25, that is, the neighborhood is set to a 5×5 range around the center pixel. The pixels of the ultrasound image on the inner side of the blood vessel wall are used as the center pixels for image pixel encoding. The sign function represents the gray level of the p-th pixel's neighborhood. and the grayscale of the center pixel in that neighborhood When the difference is greater than 0, the sign function takes the value 1; otherwise, it takes the value 0. This means calculating the integer code by accumulating the p-th pixel in the neighborhood according to the binary bit weights. This represents the second threshold function, which is the threshold value of the p-th pixel in the neighborhood. When the value is greater than the amplitude threshold, the value is 1; otherwise, it is 0. (Grayscale amplitude) The neighborhood gray level of the p-th pixel in the neighborhood and center pixel grayscale The absolute value of the difference.

[0101] Specifically, the process of generating ultrasound-coded images of vascular locations and synthesizing multi-channel ultrasound-coded images of vascular locations can be represented as follows:

[0102]

[0103] In the formula, represents the polar radius and polar angle of the intravascular ultrasound image encoded in polar coordinates, and x and y represent the horizontal and vertical coordinates of the pixels in the intravascular ultrasound image. The x and y coordinates of the center pixel in the intravascular ultrasound image. Represents the arctangent function. These represent the coding radius and coding polar angle of the ultrasound-coded image of the blood vessel location, respectively. Let represent the minimum and maximum polar radius of the intravascular ultrasound image encoded in polar coordinates, respectively. The encoded polar radius is normalized to the range [0, 1]. The encoded polar angle includes sine and cosine encoding to achieve a continuous table from 0 to 360 degrees. This represents a composite multi-channel vascular ultrasound encoded image. This indicates a component concatenation operation. These represent the central grayscale encoded image, the fine texture encoded image, and the regional texture encoded image within the blood vessel wall, respectively.

[0104] like Figure 3 As shown, the backbone network feature extraction module further includes:

[0105] An initial convolutional unit is used to pass the integrated multichannel vascular ultrasound encoded image through an initial convolutional layer to generate a smooth vascular ultrasound feature image;

[0106] The first C2f feature extraction unit is used to extract small-scale plaque features from the smooth blood vessel ultrasound feature image through the first C2f to generate small-scale primary image plaque features;

[0107] The second C2f feature extraction unit is used to extract medium-scale patch features from the small-scale primary image patch features through the second C2f to generate medium-scale image patch features.

[0108] The third C2f feature extraction unit is used to extract global plaque vessel wall features from the mesoscale image plaque features through the third C2f to generate global image plaque features;

[0109] The multi-scale patch features include the small-scale primary image patch features, the medium-scale image patch features, and the global image patch features. The patch feature weighted backbone network includes an initial convolutional layer, a first C2f, a second C2f, and a third C2f.

[0110] It is understandable that the backbone network of the native YOLOv8 architecture has an initial convolution, at least three C2f layers and SPPF, where SPPF is not necessary. In this embodiment, global plaque vascular wall features have been extracted through the large receptive field and weighting mechanism of the third C2f. Redundant aggregation of SPPF will lead to feature redundancy. Therefore, SPPF is removed and the global image plaque features output by the third C2f are directly used as global features. This simplifies the link while avoiding excessive compression of global information. Moreover, the feature dimension and resolution of this global feature are adapted to the CBAM and FPN of the neck network.

[0111] Specifically, small-scale primary image patch features roughly correspond to the features of early micro-lipid plaques (diameter less than 2 mm), medium-scale image patch features roughly correspond to the features of progressive fibrous plaques (diameter 2 to 5 mm), and global image patch features roughly correspond to diffuse plaques.

[0112] In particular, by unifying the feature dimensions through initial convolution and filtering out high-frequency noise, the features of early micro lipid plaques (diameter less than 2 mm), progressive fibrous plaques (diameter of 2 to 5 mm), and diffuse plaques are extracted sequentially through three levels of C2f, thus achieving the extraction of plaque features across the entire scale.

[0113] Specifically, the process of generating ultrasound-coded images of vascular locations and synthesizing multi-channel ultrasound-coded images of vascular locations using a plaque feature-weighted backbone network can be represented as follows:

[0114]

[0115] In the formula, These represent the ultrasound features of smooth blood vessels, plaque features in small-scale primary images, plaque features in medium-scale images, and plaque features in global images, respectively. Indicates the initial convolutional layer. Indicates batch normalization, express The activation function, its convolution stride ( The preferred value is 2, used to capture a wide range of low-level image features through a large convolution kernel with a 6x6 kernel size, and through a convolution stride of 2 ( Downsampling is performed to halve the spatial size, reducing the computational cost of C2f. This indicates that for the first C2f, the preferred number of Bottlenecks is 3. The second C2f is preferably represented by 3 Bottlenecks, where Used to determine the stride through a 3x3 convolution kernel. The downsampling value is 2. The third C2f is preferably represented by 3 Bottlenecks, where Used to determine the stride through a 3x3 convolution kernel. The downsampling is 2.

[0116] like Figure 4 As shown, the neck network feature focusing module further includes:

[0117] The CBAM scale fusion unit is used to perform a first upsampling on the global image patch features and the mid-scale image patch features, and then perform feature addition and first FPN fusion convolution to generate temporary convolution features. The temporary convolution features are then passed through CBAM to generate mid-level attention enhancement features.

[0118] The shallow feature fusion unit is used to combine the mid-level attention enhancement features with the small-scale primary image patch features after second upsampling through feature addition and second FPN fusion convolution to generate the fused patch features;

[0119] The neck network includes the first upsampling, the first FPN fusion convolution, the CBAM, the second upsampling, and the second FPN fusion convolution.

[0120] Understandably, the native YOLOv8 architecture's neck network is an FPN (Feature Pyramid Network) and PAN (Path Aggregation Network) architecture, achieving complementarity between high-level semantic features and low-level detailed features through top-down (FPN) and bottom-up (PAN) feature transfer. This embodiment embeds the attention mechanism CBAM into the neck network, enabling enhanced feature fusion to focus on mid-scale image patch features containing a large amount of patch information, thus adapting to the needs of patch detection, which places greater emphasis on mid-scale patches and requires higher detection accuracy.

[0121] Specifically, the process by which the neck network generates the fused patch features can be represented as follows:

[0122]

[0123] In the formula, These represent global image patch features, global image patch upsampling features, mid-scale image patch features, first additive features, temporary convolution features, mid-level attention-enhanced features, mid-level attention-enhanced upsampling features, second additive features, small-scale primary image patch features, and fused patch features, respectively. Indicates the first upsampling. This indicates the first FPN fused convolution. Indicates the second upsampling. This represents the second FPN fusion convolution. Preferably, the kernel size of both the first FPN fusion convolution and the second FPN fusion convolution is 3x3.

[0124] Furthermore, the CBAM scale fusion unit includes:

[0125] The channel attention subunit is used to generate channel-focused features by passing the temporary convolutional features through a channel attention mechanism that includes global average pooling of feature maps, global max pooling of feature maps, and a multilayer perceptron.

[0126] A spatial attention subunit is used to generate the mid-level attention enhancement features by passing the temporary convolutional features through a spatial attention mechanism with channel-dimensional global average pooling and channel-dimensional global max pooling.

[0127] The CBAM includes the channel attention mechanism and the spatial attention mechanism.

[0128] In particular, through the global, mesoscale, and small-scale reverse feature fusion links of FPN (Feature Pyramid Network), deep integration of multi-scale features is achieved. CBAM dynamically identifies and enhances key information for plaque diagnosis, weakens irrelevant interference such as normal vascular wall tissue and blood artifacts, and allows the fused features to focus on the plaque.

[0129] Specifically, the process of CBAM generating mid-level attention-enhanced features can be represented as:

[0130]

[0131] In the formula, These represent temporary convolutional features, channel focusing features, and mid-level attention enhancement features, respectively. This represents the SIGMOD activation function. A multilayer perceptron representing the channel attention mechanism is preferably a multilayer perceptron. ,in This represents the learnable weight matrix of the convolutional kernels at two levels. This represents the ReLU function, and X represents the input feature. These represent the global average pooling and global max pooling of the feature map in the channel attention mechanism, respectively. This indicates element-wise multiplication. Let $\mathbf$ represent the channel-dimensional global average pooling and channel-dimensional global max pooling of the spatial attention mechanism, respectively. This indicates a feature concatenation operation. This indicates a convolution operation with a kernel size of 7×7.

[0132] Furthermore, the patch category determination module includes:

[0133] The convolutional block attention enhancement unit is used to generate category-enhanced features by passing the fused patch features through a convolutional block attention mechanism based on global average pooling, channel global average pooling, and channel global max pooling.

[0134] A classification unit is used to generate the cardiovascular atherosclerotic plaque category by passing the category enhancement features through a SiLU-based classification convolution.

[0135] The patch classification head includes the convolutional block attention mechanism and the classification convolution.

[0136] Specifically, the cardiovascular atherosclerotic plaque categories include the characteristics of early microlipidic plaques (less than 2 mm in diameter), the characteristics of advanced fibrous plaques (2 to 5 mm in diameter), and diffuse plaques.

[0137] In particular, by using the convolutional block attention mechanism, features strongly correlated with patch categories are selected and enhanced from both global and channel dimensions, while irrelevant features are weakened, making category features more prominent and achieving accurate determination of patch types.

[0138] Specifically, the process of classifying cardiovascular atherosclerotic plaques can be represented as follows:

[0139]

[0140] In the formula, These represent category enhancement features, fusion plaque features, and cardiovascular atherosclerotic plaque categories, respectively. This represents the SIGMOD activation function. Represents the ReLU function. It represents the Hadamah accumulation. The learnable weight matrix represents the convolution. Indicates global average pooling. These represent channel-wide average pooling and channel-wide max pooling, respectively. This indicates a convolutional channel compression operation with a kernel size of 3×3. This represents a convolution mapping operation with a kernel size of 3×3. express Activation function.

[0141] Furthermore, the patch regression identification module includes:

[0142] A feature adaptation unit is used to pass the fused patch features through a feature adaptation convolutional layer to generate adapted patch features;

[0143] A multi-scale feature enhancement unit is used to enhance the adapted patch features through deformable convolution to generate enhanced patch features.

[0144] A regression judgment unit is used to pass the enhanced plaque features through a regression convolutional layer to generate the intravascular ultrasound annotated image;

[0145] The patch regression recognition head includes the feature-adaptive convolutional layer, the deformable convolutional layer, and the regression convolutional layer.

[0146] In particular, by dynamically adapting deformable convolution to the irregular shape and multiple sizes of patches, the problem of large selection deviation and incomplete coverage of small patches in traditional fixed convolution is solved.

[0147] Specifically, the process of generating intravascular ultrasound-annotated images can be represented as follows:

[0148]

[0149] In the formula, These represent adapted plaque features, fused plaque features, deformable convolution spatial offset, deformable convolution modulation weights, enhanced plaque features, and intravascular ultrasound annotated images, respectively. Indicates feature-adaptive convolutional layers, Indicates instance standardization, express Activation function This represents a convolution mapping operation with a kernel size of 1×1. This represents a convolutional layer that predicts spatial offsets and convolutional modulation weights. This represents the deformable convolution feature extraction operation. This indicates a prediction convolution operation for the distribution of bounding boxes with a kernel size of 3×3.

[0150] Understandably, the native YOLOv8 architecture's detection head has simple classification and regression branches. The classification branch generates the category of the target, and the regression branch generates the location of the target, i.e., generating a bounding box in the image. In this embodiment, the convolutional block attention enhancement and SiLU classification convolution of the plaque classification head further focus on the small texture grayscale difference between lipid plaques and progressive fibrous plaques. The deformable convolution enhancement and 1×1 convolution of the regression convolution layer of the plaque regression recognition head generate pixel-level plaque edge recognition boxes, so that doctors can more intuitively confirm the size and boundaries of the plaques and avoid misdiagnosis of plaque size due to differences in doctors' experience.

[0151] Furthermore, the cardiovascular atherosclerosis detection system also includes:

[0152] The training optimization module is used to optimize the training of the patch recognition model using a comprehensive loss function based on the VFL loss term and the EIoU loss term.

[0153] Specifically, the comprehensive loss function can be expressed as:

[0154]

[0155] In the formula, These represent the VFL loss term, EIoU loss term, and combined loss function, respectively. This represents the total number of positive samples in the sample image. This represents the total number of sample images. This represents the cardiovascular atherosclerotic plaque category generated by the plaque recognition model for the i-th sample. This indicates the true labeled cardiovascular atherosclerotic plaque category of the i-th sample. This represents two weighting coefficients, preferably 1 and 2.5.

[0156] In particular, the backbone network with initial convolution and three-level C2f units enables refined hierarchical extraction of vascular ultrasound features. The neck network combining FPN architecture and CBAM attention mechanism enables focusing on the core features of plaques. Through the output of plaque categories and precise recognition boxes, doctors are provided with intuitive diagnostic basis for cardiovascular atherosclerotic plaques, reducing their workload and subjective misjudgment.

[0157] In this embodiment, intravascular ultrasound images are encoded using diffusion filtering, CLBP encoding, and polar coordinate position encoding. This simultaneously encodes plaque texture details, noise suppression information, and plaque spatial distribution information, improving the completeness of feature dimensions. Furthermore, an improved YOLOv8 framework recognition model based on BiFPN and attention mechanisms enables automated detection of cardiovascular atherosclerotic plaques in intravascular ultrasound image data. This achieves efficient detection of atherosclerotic plaques, providing physicians with more intuitive ultrasound image data and reducing missed detections due to subjective differences in physician experience. By using diffusion filtering, CLBP encoding, and polar coordinate position encoding, noise interference from the inner wall of the vessel is removed, accurately locating the inner region of the vessel wall, extracting multidimensional texture information, and generating position information consistent with the vessel shape. This reduces missed detections and false detections in the YOLOv8 plaque recognition model. By using an initial convolutional and three-level C2f unit backbone network, refined hierarchical extraction of vascular ultrasound features is achieved. By combining the FPN architecture with the CBAM attention mechanism in the neck network, the core features of plaques are focused. Through the output of plaque categories and precise recognition boxes, doctors are provided with intuitive diagnostic basis for cardiovascular atherosclerotic plaques, reducing their workload and subjective misjudgment.

[0158] Those skilled in the art will recognize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0159] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0160] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A data analysis-based cardiovascular atherosclerosis detection system, characterized by, The method comprises the following steps: a comprehensive encoding module is used to encode the intravascular ultrasound image through a comprehensive encoder based on a diffusion filter, CLBP encoding and polar coordinate position encoding to generate a comprehensive multi-channel vascular ultrasound encoded image; a backbone network feature extraction module is used to extract the comprehensive multi-channel vascular ultrasound encoded image through a C2f-based plaque feature weighted backbone network to generate multi-scale plaque features; a neck network feature focusing module is used to focus the multi-scale plaque features through a neck network based on CBAM and FPN architecture to generate fused plaque features; a plaque category judgment module is used to judge the fused plaque features through a plaque classification head based on a convolution block attention mechanism to generate a cardiovascular atherosclerotic plaque category; a plaque regression identification module is used to identify the fused plaque features through a plaque regression identification head based on deformable convolution to generate an intravascular ultrasound annotation image with labeled plaque bounding boxes; wherein the plaque feature weighted backbone network, the neck network, the plaque classification head and the plaque regression identification head constitute a plaque identification model of a YOLOv8 framework; the comprehensive encoding module comprises: a diffusion filter unit is used to generate an intravascular ultrasound image on the inner side of the vascular wall through a diffusion filter based on the image gradient value of the intravascular ultrasound image; a CLBP encoding unit is used to generate a vascular wall texture ultrasound encoded image through CLBP encoding based on the intravascular ultrasound image on the inner side of the vascular wall; a polar coordinate encoding unit is used to generate a vascular position ultrasound encoded image through polar coordinate encoding based on the intravascular ultrasound image; an encoded image splicing unit is used to splice the vascular wall texture ultrasound encoded image and the vascular position ultrasound encoded image to generate the comprehensive multi-channel vascular ultrasound encoded image.

2. The data analysis-based cardiovascular atherosclerosis detection system of claim 1, wherein, The diffusion filter unit comprises: an image gradient calculation subunit is used to generate an image gradient value through a Sobel operator based on the intravascular ultrasound image; a diffusion coefficient calculation subunit is used to generate an image edge diffusion coefficient through a diffusion coefficient calculation equation based on the image gradient value; a divergence calculation subunit is used to generate an image local gray intensity through a divergence operator based on the image edge diffusion coefficient and the image gradient value; an intravascular wall inner side region judgment subunit is used to generate the intravascular ultrasound image on the inner side of the vascular wall based on a comparison between the image local gray intensity and a threshold value; wherein the diffusion filter comprises the Sobel operator, the diffusion coefficient calculation equation and the divergence operator.

3. The data analysis based cardiovascular atherosclerosis detection system of claim 1, wherein, The CLBP encoding unit comprises: an image gray encoding subunit is used to generate a vascular wall center gray encoded image through gray CLBP encoding based on the center pixel gray and the neighborhood average gray of the intravascular ultrasound image on the inner side of the vascular wall; a local texture encoding subunit is used to generate a vascular wall fine texture encoded image through local differential CLBP encoding based on the difference between the center pixel gray and the neighborhood gray of the intravascular ultrasound image on the inner side of the vascular wall; An amplitude gray scale coding subunit is configured to code the intravascular wall ultrasound image based on amplitude CLBP coding to generate an intravascular wall region texture coding image. The CLBP coding includes the gray scale CLBP coding, the local differential CLBP coding and the amplitude CLBP coding, and the intravascular wall texture ultrasound coding image includes the intravascular wall center gray scale coding image, the intravascular wall fine texture coding image and the intravascular wall region texture coding image.

4. The data analysis-based cardiovascular atherosclerosis detection system of claim 1, wherein, The backbone network feature extraction module includes: An initial convolution unit is configured to code the integrated multi-channel vascular ultrasound coding image through an initial convolution layer to generate a smooth vascular ultrasound feature image; A first C2f feature extraction unit is configured to code the smooth vascular ultrasound feature image through a first C2f to extract small-scale primary image plaque features to generate small-scale primary image plaque features; A second C2f feature extraction unit is configured to code the small-scale primary image plaque features through a second C2f to extract medium-scale image plaque features to generate medium-scale image plaque features; A third C2f feature extraction unit is configured to code the medium-scale image plaque features through a third C2f to extract global plaque vascular wall features to generate global image plaque features; The multi-scale plaque features include the small-scale primary image plaque features, the medium-scale image plaque features and the global image plaque features, and the plaque feature weighted backbone network includes the initial convolution layer, the first C2f, the second C2f and the third C2f.

5. The data analysis-based cardiovascular atherosclerosis detection system of claim 4, wherein, The neck network feature focusing module includes: A CBAM scale fusion unit is configured to code the global image plaque features through a first up-sampling and fuse the medium-scale image plaque features through feature addition and a first FPN fusion convolution to generate temporary convolution features, and code the temporary convolution features through a CBAM to generate a middle layer attention enhanced feature; A shallow layer feature fusion unit is configured to code the middle layer attention enhanced feature through a second up-sampling and fuse the small-scale primary image plaque features through feature addition and a second FPN fusion convolution to generate the fused plaque features; The neck network includes the first up-sampling, the first FPN fusion convolution, the CBAM, the second up-sampling and the second FPN fusion convolution.

6. The data analysis-based cardiovascular atherosclerosis detection system of claim 5, wherein, The CBAM scale fusion unit includes: A channel attention subunit is configured to code the temporary convolution features through a channel attention mechanism with a feature map global average pooling, a feature map global maximum pooling and a multi-layer perception to generate a channel focused feature; A spatial attention subunit is configured to code the temporary convolution features through a spatial attention mechanism with a channel dimension global average pooling and a channel dimension global maximum pooling to generate the middle layer attention enhanced feature; The CBAM includes the channel attention mechanism and the spatial attention mechanism.

7. The data analysis-based cardiovascular atherosclerosis detection system of claim 1, wherein, The plaque category judgment module includes: a convolution block attention enhancement unit, configured to pass the fused plaque feature through a convolution block attention mechanism based on global average pooling, channel global average pooling and channel global maximum pooling to generate a class enhancement feature; a classification unit, configured to pass the class enhancement feature through a classification convolution based on SiLU to generate the cardiovascular atherosclerotic plaque class; wherein the plaque classification head comprises the convolution block attention mechanism and the classification convolution.

8. The data analysis based cardiovascular atherosclerosis detection system of claim 1, wherein, The plaque regression identification module comprises: a feature adaptation unit, configured to pass the fused plaque feature through a feature adaptation convolution layer to generate an adapted plaque feature; a multi-scale feature enhancement unit, configured to pass the adapted plaque feature through deformable convolution for multi-scale feature enhancement to generate an enhanced plaque feature; a regression judgment unit, configured to pass the enhanced plaque feature through a regression convolution layer to generate the intravascular ultrasound labeled image; wherein the plaque regression identification head comprises the feature adaptation convolution layer, the deformable convolution and the regression convolution layer.

9. The data analysis-based cardiovascular atherosclerosis detection system according to any one of claims 1 to 8, characterized in that, Further comprising: a training optimization module, configured to perform optimal training on the plaque identification model through a comprehensive loss function based on a VFL loss term and an EIoU loss term.

Images