Blood vessel feature analysis method and device based on target detection, equipment and medium

Abstract

The invention provides a target detection-based blood vessel feature analysis method and device, equipment and a medium, and the method comprises the steps: carrying out the processing of an input blood vessel image through an initialized feature extraction module, generating a pre-training model, carrying out the feature extraction of the blood vessel image through the pre-training model, and obtaining image feature data; inputting the lesion area into an analysis module, and performing image segmentation on the lesion area through the analysis module to obtain segmented image features; the segmented image features comprise image data of a calcified region and / or a non-calcified region; and respectively inputting the segmented image features into a calcified plaque analysis module and a non-calcified plaque analysis module to obtain the type of a vascular lesion area, further analyzing the density distribution of mixed plaques to determine whether the mixed plaques are in a splicing type or a uniform type, and finally generating an analysis result containing the lesion type and the stenosis degree. According to the invention, accurate positioning, segmentation and classification of vasculopathy can be realized without labeling data, and the automation level and accuracy of vasculopathy diagnosis are improved.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method, device, equipment and medium for analyzing blood vessel characteristics based on target detection. Background Art

[0002] Accurate diagnosis of vascular lesions is a complex medical problem. Traditional methods rely on large amounts of labeled data and manual experience, making it difficult to achieve automation and high-precision analysis.

[0003] For example, one approach uses image segmentation and centerline extraction to determine the diameter of each point on the centerline to determine the degree of stenosis. Another approach uses image segmentation, centerline extraction, surface reconstruction, and then extracts local features of each point. Through feature analysis, the severity or type of lesion at each point on the centerline can be determined.

[0004] The first method, which uses image segmentation and centerline extraction, has limited application scenarios. This solution can only determine the degree of stenosis and lesions in blood vessels. To determine the type of lesion, other technical methods must be applied to supplement it. Secondly, the accuracy of this solution is severely affected by the image segmentation results, and the error tolerance is very low. For small blood vessels that occupy only a few voxels in the image, the accuracy of the stenosis rate is severely affected. In other words, a difference of one voxel may result in a deviation of 20% or even higher.

[0005] The second type of image analysis, achieved through steps such as image segmentation, centerline extraction, and surface reconstruction, involves many steps and consumes computing resources and processing time. Although it can ultimately locate the extent and type of lesions at each point on the blood vessel, each step has a significant impact on the final result. For example, the image segmentation result affects the accuracy of the centerline, which in turn affects the accuracy of surface reconstruction, ultimately affecting the features extracted from each point and thus the final lesion analysis module. This solution uses a convolutional neural network and a network model combined with a recurrent neural network for processing. The recurrent neural network's iterative loop method takes a long time to calculate and is prone to ignoring some important information or enhancing some redundant features during the loop, affecting or reducing the accuracy of the data and increasing errors.

[0006] How to accurately analyze lesions in vascular surface reconstructed images in the absence of labeled data is a technical challenge that needs to be solved urgently. This involves how to learn effective features from unlabeled data, how to accurately locate and segment lesion areas, and how to distinguish different types of plaques. Especially for mixed plaques, judging their density distribution is more challenging. At the same time, vascular lesions have diverse morphologies and varying degrees of stenosis. How to improve efficiency while ensuring analysis accuracy is also a key issue. In addition, due to the complexity of vascular images, how to obtain rich texture feature information while ensuring boundary accuracy to support subsequent lesion type judgment is also an important factor that needs to be considered. Solving these problems will directly affect the accuracy and automation level of vascular lesion diagnosis, and is of great significance to improving the efficiency and quality of clinical diagnosis. Summary of the Invention

[0007] To address the problems of limited application scenarios and low detection and processing efficiency of existing vascular image processing methods, the present invention provides a vascular feature analysis method based on target detection, which mainly includes: Obtain a vascular image dataset, use a 3DU-Net (three-dimensional U-shaped convolutional neural network, Three-DimensionalU-Net) model to perform image restoration tasks on images in the vascular image dataset, generate pre-training parameters, initialize the encoding part of the feature extraction module according to the pre-training parameters, and obtain an initialized feature extraction module; process the input vascular image through the initialized feature extraction module to generate a pre-training model, use the pre-training model to extract features from the vascular image, and obtain image feature data; based on the image feature data, analyze the vascular image through the region recommendation module to obtain the lesion area The position and range of the lesion area are determined, and the lesion area is marked to generate a detection result including the coordinates of the lesion area; the lesion area is input into the analysis module, and the image of the lesion area is segmented by the analysis module to obtain segmented image features; the segmented image features include image data of the calcified area and / or the non-calcified area; the segmented image features are input into the calcified plaque analysis module and the non-calcified plaque analysis module respectively; the vascular lesion area type corresponding to the segmented image feature is determined by the calcified plaque analysis module and the non-calcified plaque analysis module; the vascular lesion area type includes four types: mixed plaque type, calcified plaque type, non-calcified plaque type and no plaque type.

[0008] Furthermore, determining the type of vascular lesion area corresponding to the segmented image feature by the calcified plaque analysis module and the non-calcified plaque analysis module includes: determining the type of vascular lesion area corresponding to the segmented image feature based on the presence / absence of calcified plaque analysis results output by the calcified plaque analysis module, and based on the presence / absence of non-calcified plaque analysis results output by the non-calcified plaque analysis module.

[0009] Furthermore, the mixed plaque type is determined by combining the analysis results of calcified plaques output by the calcified plaque analysis module and the analysis results of non-calcified plaques output by the non-calcified plaque analysis module; the calcified plaque type is determined by combining the analysis results of calcified plaques output by the calcified plaque analysis module and the analysis results of non-non-calcified plaques output by the non-calcified plaque analysis module; the non-calcified plaque type is determined by combining the analysis results of non-calcified plaques output by the calcified plaque analysis module and the analysis results of non-calcified plaques output by the non-calcified plaque analysis module; the non-plaque type is determined by combining the analysis results of non-calcified plaques output by the calcified plaque analysis module and the analysis results of non-calcified plaques output by the non-calcified plaque analysis module.

[0010] Furthermore, the method of analyzing the vascular image through a region recommendation module based on the image feature data to obtain the position and range of the lesion area, marking the lesion area, and generating a detection result including the coordinates of the lesion area includes: analyzing the vascular image through a region recommendation module based on the image feature data to determine the stenosis length and stenosis morphological characteristics of the lesion area, marking the lesion area, and generating a detection result including the coordinates of the lesion area and the distribution of the stenosis position.

[0011] Furthermore, the step of inputting the lesion area into the analysis module, performing image segmentation on the lesion area through the analysis module, and obtaining segmented image features includes: inputting the lesion area into the analysis module, performing pixel-level segmentation on the lesion area through the image segmentation module, and obtaining segmented image features including lesion boundary accuracy and regional texture features.

[0012] Furthermore, the acquiring of the vascular image dataset includes: acquiring an unlabeled vascular image dataset, wherein the unlabeled vascular image dataset includes an unlabeled vascular curved surface reconstructed image dataset.

[0013] Furthermore, after the mixed plaque type is determined by combining the calcified plaque analysis results output by the calcified plaque analysis module and the non-calcified plaque analysis results output by the non-calcified plaque analysis module, it also includes: analyzing the proportion of non-calcified areas through the non-calcified plaque analysis module, and analyzing the proportion of calcified areas through the calcified plaque analysis module; determining the density distribution of non-calcified areas and calcified areas, and determining whether the mixed plaque type is a spliced ​​mixed plaque type or a uniform mixed plaque type.

[0014] Furthermore, it also includes: performing regional contrast enhancement on the segmented image features through a regional refinement module, optimizing the volume of the lesion area and the consistency of the segmented area, and generating refined lesion area image data.

[0015] Furthermore, it also includes: generating a final analysis result including the type of vascular lesion area and the degree of stenosis based on the refined lesion area image data, combined with the stenosis area contrast and lesion area classification label.

[0016] Furthermore, an embodiment of the present invention also provides a device for analyzing blood vessel characteristics based on target detection, comprising: an unsupervised training unlabeled data module for acquiring a vascular image dataset, performing image restoration on images in the vascular image dataset using a 3DU-Net model, generating pre-training parameters, and initializing an encoding portion of a feature extraction module based on the pre-training parameters to obtain an initialized feature extraction module; a pre-trained model image feature extraction module, which processes the input vascular image through the initialized feature extraction module generated by the unsupervised training unlabeled data module to generate a pre-trained model, and uses the pre-trained model to extract features from the vascular image to obtain image feature data; a lesion region positioning module, configured to analyze the vascular image using the region recommendation module based on the image feature data extracted by the image feature extraction module according to the pre-trained model, obtain the location and range of the lesion region, annotate the lesion region, and generate a detection result including the coordinates of the lesion region; an analysis module, configured to input the lesion area, perform image segmentation on the lesion area through the analysis module, and obtain segmented image features; the segmented image features include image data of calcified areas and / or non-calcified areas; and further configured to input the segmented image features into a calcified plaque analysis module and a non-calcified plaque analysis module, respectively; a determination module, comprising the calcified plaque analysis module and the non-calcified plaque analysis module, configured to determine the type of vascular lesion region corresponding to the segmented image feature based on the presence / absence of calcified plaque analysis result output by the calcified plaque analysis module and the presence / absence of non-calcified plaque analysis result output by the non-calcified plaque analysis module; The types of vascular lesion areas include four types: mixed plaque type, calcified plaque type, non-calcified plaque type and no plaque type.

[0017] Furthermore, an embodiment of the present invention further provides a non-volatile computer-readable storage medium having computer-executable instructions stored thereon. When the computer-executable instructions are executed by a processor, the following steps are implemented: Obtaining a vascular image dataset, performing an image restoration task on images in the vascular image dataset using a 3DU-Net model to generate pre-training parameters, and initializing an encoding portion of a feature extraction module according to the pre-training parameters to obtain an initialized feature extraction module; Processing the input vascular image through the initialized feature extraction module to generate a pre-trained model, and extracting features from the vascular image using the pre-trained model to obtain image feature data; Analyzing the blood vessel image using a region recommendation module based on the image feature data to obtain the location and range of the lesion area, marking the lesion area, and generating a detection result including the coordinates of the lesion area; Inputting the lesion area into an analysis module, performing image segmentation on the lesion area by the analysis module to obtain segmented image features; the segmented image features include image data of calcified areas and / or non-calcified areas; Inputting the segmented image features into a calcified plaque analysis module and a non-calcified plaque analysis module respectively; Determining the type of vascular lesion region corresponding to the segmented image feature based on the presence / absence of calcified plaque analysis results output by the calcified plaque analysis module and the presence / absence of non-calcified plaque analysis results output by the non-calcified plaque analysis module; The types of vascular lesion areas include four types: mixed plaque type, calcified plaque type, non-calcified plaque type and no plaque type.

[0018] Furthermore, an embodiment of the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any of the above methods when executing the computer program.

[0019] The technical solution provided by the embodiment of the present invention may have the following beneficial effects: This invention discloses a deep learning-based vascular lesion analysis method and system. The system obtains an unlabeled vascular surface reconstructed image dataset, performs unsupervised pre-training using the 3DU-Net model, initializes a feature extraction module, generates a pre-trained model, and extracts features from the vascular images. The system then uses a region recommendation module to analyze and determine the stenosis length and morphological characteristics of the lesion area. The lesion area is then segmented at the pixel level to obtain boundary accuracy and texture features. Finally, a calcified and non-calcified plaque analysis module determines four lesion types. The density distribution of mixed plaques can be further analyzed to determine whether they are spliced ​​or uniform, ultimately generating an analysis result that includes the lesion type and degree of stenosis. This method can accurately locate, segment, and classify vascular lesions without the need for labeled data, improving the automation and accuracy of vascular lesion diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A flow chart of a method for analyzing blood vessel characteristics based on target detection provided by an embodiment of the present invention.

[0021] Figure 2 This is a structural diagram of a blood vessel feature analysis device based on target detection provided by an embodiment of the present invention.

[0022] Figure 3 A structural diagram of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] To further understand the content of the present invention, the present invention is described in detail with reference to the accompanying drawings and examples. The present application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the invention are shown in the accompanying drawings.

[0024] like Figure 1 As shown, this embodiment provides a method for analyzing blood vessel characteristics based on target detection, which may specifically include: S101. Obtain a vascular image dataset, perform an image restoration task on images in the vascular image dataset using a 3DU-Net model, generate pre-training parameters, and initialize an encoding portion of a feature extraction module according to the pre-training parameters to obtain an initialized feature extraction module.

[0025] When acquiring a vascular image dataset, three-dimensional vascular images can be collected through medical imaging equipment (such as CTA or MRA). The data must contain diverse samples from different anatomical sites (such as coronary arteries and cerebral arteries).

[0026] For example, a hospital collected head MRA data from 500 anonymous patients. Each data set contained 512×512×300 voxels with a voxel spacing of 0.5mm×0.5mm×0.5mm, covering both normal and diseased vessels to ensure data diversity. The data needed to be de-identified and converted to a unified format to accommodate subsequent processing. When using the 3DU-Net (Three-Dimensional U-Net) model for image restoration, a specific data corruption scheme must be devised. During the generation process, the input vascular image is first randomly shuffled. This step aims to enable the network to learn more general features rather than specific image details. The shuffled image serves as input to the 3D U-Net, a network architecture specifically designed for medical image processing. Its characteristic feature is the symmetrical encoder and decoder design, which enables it to extract features while preserving spatial information. During the training process, the output of the 3D U-Net is compared with the original unshuffled image, and a loss function such as mean squared error (MSE) or cross entropy loss is calculated. The network parameters are continuously optimized through the backpropagation algorithm to make the restored result as close to the original image as possible.

[0027] For example, the original image is randomly cropped into 64×64×64 cubic blocks, and Gaussian noise (standard deviation 0.1) or 30% of the area is randomly occluded as input. The model output is compared with the uncorrupted original block to calculate the mean squared error (MSE) loss. The model is then iteratively trained for 200 epochs using the Adam optimizer (learning rate 0.001). This process forces the encoder to learn essential features such as vessel wall morphology and branching structure, rather than relying on local noise. When generating pre-trained parameters, the 3D U-Net encoder must retain its multi-scale feature extraction capabilities.

[0028] For example, the encoder consists of four downsampling layers, each using a 3×3×3 convolution kernel, with the number of channels doubling sequentially (16 → 32 → 64 → 128). After training, the feature maps output by the third layer capture the texture differences between the vessel lumen and the plaque, demonstrating that it has learned anatomically relevant features. When initializing the feature extraction module, the parameters of the first three encoder layers are frozen, and only the last layer is fine-tuned.

[0029] For example, the fourth-layer output of the pre-trained encoder (128-channel feature map) is connected to a 1×1×1 convolutional layer to reduce its dimensionality to 64 channels to adapt to downstream tasks. Comparative experiments show that the model initialized with pre-training achieves a 12% improvement in the Dice coefficient for vessel segmentation, demonstrating that parameter transfer effectively alleviates the problem of insufficient labeled data. In terms of technical effectiveness, the pre-training strategy exploits inherent patterns in the data through unsupervised learning.

[0030] For example, in the subsequent stenosis rate calculation task, the initialized feature extraction module reduced the segmentation error of tiny blood vessels (diameter <2mm) to within 1 voxel, significantly outperforming the randomly initialized model (error of 3-5 voxels). This is due to the robust representation of vascular continuity features by the pre-trained encoder, which avoids the stenosis rate calculation bias caused by local missegmentation in traditional methods.

[0031] S102 , processing the input vascular image through the initialized feature extraction module to generate a pre-trained model, and extracting features from the vascular image using the pre-trained model to obtain image feature data.

[0032] The initialized feature extraction module gradually extracts low-level to high-level features from the image through multiple layers of convolution and pooling operations. Specifically, the initialized feature extraction module consists of several convolutional layers, each using kernels of different sizes to capture local features in the image, such as edges and textures. Subsequently, the pooling layer performs dimensionality reduction on these features, preserving the most important information while reducing computational effort.

[0033] For example, for a 3D vascular image, the initialized feature extraction module first uses a 3x3x3 convolution kernel to extract voxel-level features, and then compresses these features through the maximum pooling layer to form higher-level abstract features.

[0034] For example, consider a dataset containing 1,000 vascular images, each with a size of 256x256x256 voxels. Through training, a 3D U-Net can learn the basic structure and feature distribution of vascular images. When using a pre-trained model to extract features from vascular images, the input vascular images are processed by the pre-trained 3D U-Net, and the output feature maps contain rich vascular structural information. These feature maps can be multi-scale, meaning that feature maps at different levels correspond to vascular structures at different scales, such as main vascular trunks, branches, and small vessels.

[0035] For example, the first convolutional layer may extract edge information of blood vessels, while deeper convolutional layers extract more complex structural information. In this way, the pre-trained model can convert the original image into high-dimensional feature data, which is convenient for subsequent analysis and processing. The obtained image feature data can be further used to detect and analyze lesions. Specifically, this feature data can be input into subsequent classification or detection networks, such as object detection algorithms such as Faster R-CNN or YOLO. These algorithms can identify lesions in blood vessels based on the feature data and give the location and extent of the lesions.

[0036] For example, assuming the feature data contains 1,000 feature vectors, each with a length of 512. Using an object detection algorithm, lesions in an image can be identified and their type and severity labeled. This feature extraction method based on a pre-trained model has several advantages. First, pre-training allows the network to learn common features from a large amount of unlabeled data, improving the model's robustness and scalability. Second, pre-training reduces training time and computing resource requirements for subsequent tasks, as the network already possesses basic feature extraction capabilities. Finally, through multi-scale feature extraction, the model can more comprehensively capture detailed information in vascular images, improving the accuracy and reliability of lesion detection. In practice, a certain number of vascular images, such as 5,000, can be selected for pre-training. Each image is randomly shuffled and then fed into a 3D U-Net for training. During training, the loss function is monitored, and the learning rate and optimization algorithm, such as the Adam optimizer, are adjusted to ensure that the network parameters converge to optimal values. After pre-training, the model's restoration performance is evaluated using a validation set to ensure that the model accurately restores the original image. Subsequently, the pre-trained model was applied to a new vascular image feature extraction task to verify its effectiveness in actual lesion detection. For example, the model's performance was evaluated by calculating metrics such as detection accuracy and recall rate for lesion areas. Through these steps, the initialized feature extraction module and the pre-trained model together formed an efficient and robust vascular lesion detection system, capable of extracting valuable information from raw images and providing a solid foundation for subsequent lesion analysis and diagnosis.

[0037] S103 , analyzing the blood vessel image through a region recommendation module according to the image feature data, obtaining the position and range of the lesion region, marking the lesion region, and generating a detection result including the coordinates of the lesion region.

[0038] First, the acquisition of image feature data is a critical first step. These features may include edge information of blood vessels, texture features, and grayscale distribution.

[0039] For example, the region proposal module analyzes the extracted feature data to determine the location and extent of the lesion. Suppose we use a region proposal-based neural network, such as Faster R-CNN. This network generates a series of candidate regions (region proposals) on the feature map. These candidate regions are generated by calculating interest points on the feature map.

[0040] For example, if the pixel values ​​in a certain area are significantly higher than those in the surrounding areas, this may indicate the presence of a lesion. Using the non-maximum suppression (NMS) algorithm, the most likely lesion regions are identified. Next, the lesion regions are labeled. Suppose we have selected three candidate regions, located in the proximal, middle, and distal segments of a blood vessel. A further classification network, such as a softmax classifier, is used to determine whether these regions actually contain lesions.

[0041] For example, the classifier output shows that the probabilities of lesions in the proximal and middle segments are 0.8 and 0.9, respectively, while the probability of lesions in the distal segment is only 0.2. Therefore, we label the proximal and middle segments as lesion regions. A detection result is generated containing the coordinates of the lesion regions. Assume that the coordinates of the proximal lesion region are (x1, y1, x2, y2), and the coordinates of the middle lesion region are (x3, y3, x4, y4). These coordinates are recorded in the detection report for subsequent analysis and processing.

[0042] For example, these coordinates can be used to quickly locate the lesion area for subsequent judgment and processing. In specific implementation, a 512x512 pixel coronary CT image is preprocessed and feature vectors are extracted. These feature vectors are processed by the region recommendation module to generate 10 candidate regions. Through the classification network, 3 of the regions are determined to be lesion areas, and the coordinate information of these regions is recorded. First, through automated feature extraction and region recommendation, the efficiency and accuracy of lesion detection can be greatly improved, and human errors can be reduced. Secondly, the detection results containing coordinate information are generated to facilitate subsequent rapid positioning and processing, thereby improving detection efficiency. In addition, this method can be widely used in different medical image analysis tasks and has high versatility and robustness.

[0043] For example, in actual research, a hospital's coronary CT image analysis system adopted this technical solution. After testing 1,000 images, the system successfully detected lesion areas with an accuracy rate of 95%, significantly higher than the 80% achieved by traditional methods. The coordinate information generated by the system enables rapid location of lesions, shortening detection time and improving detection efficiency. For another example, in a study, this method was used to classify mixed plaques, calcified plaques, and non-calcified plaques. The results showed that the classification accuracy rates reached 92%, 90%, and 88%, respectively, confirming the effectiveness of the method in identifying different lesion types. Through specific implementation methods and examples, it can be seen that the lesion detection method based on image feature data and the region recommendation module not only improves detection accuracy and efficiency, but also provides more intuitive and detailed detection and judgment information, with important clinical application value.

[0044] S104: Input the lesion area into an analysis module, and perform image segmentation on the lesion area through the analysis module to obtain segmentation image features.

[0045] First, the lesion region is input into the analysis module. This step is a key step in the entire vascular lesion detection and analysis process. The identification and location of the lesion region is completed by the initial target detection neural network, which can accurately identify suspected lesion areas in the original vascular image.

[0046] For example, in a coronary CT angiography image, an object detection neural network may identify a localized region of a blood vessel whose grayscale values ​​and texture features differ significantly from those of normal vessel segments, indicating the presence of a lesion. The analysis module then performs image segmentation on the lesion region. The goal of image segmentation is to separate the lesion region from the background image for subsequent feature extraction and analysis. This can be achieved using deep learning-based segmentation algorithms, such as U-Net or Mask R-CNN. Taking U-Net as an example, this algorithm uses an encoder and decoder structure to gradually segment the input image into fine-scale lesion regions. Assuming the input lesion region image is 512x512 pixels, U-Net first compresses the image features to a lower dimension through multiple convolution and pooling operations. The decoder then gradually restores the spatial resolution of the image through upsampling and convolution operations, ultimately generating a segmented image of the same size as the input image. After obtaining the segmented image, features are further extracted. Feature extraction involves performing multi-scale and multi-angle analysis on the segmented image to extract key information that reflects the characteristics of the lesion.

[0047] For example, morphological features of the lesion area, such as area, perimeter, and shape factor, can be extracted. Texture features, such as the gray-level co-occurrence matrix (GLCM) features, including contrast, energy, and entropy, can also be extracted. Assuming the area of ​​a lesion is 100 square pixels, the perimeter is 40 pixels, and the shape factor is 0.5, these feature values ​​will serve as the basis for subsequent determination of the lesion's severity and type. Furthermore, deep learning models can be used to automatically extract high-dimensional features.

[0048] For example, a convolutional neural network (CNN) is used to extract features from segmented images. Through multiple layers of convolution and pooling, CNNs can extract deep-level image features. For example, a CNN model with five convolutional layers is used, with feature maps extracted by each layer having sizes of 512x512, 256x256, 128x128, 64x64, and 32x32, respectively. These feature maps encompass image information from low-level to high-level, comprehensively reflecting the characteristics of the lesion area. Through these steps, the segmented image features of the lesion area are fully extracted, providing a solid foundation for further lesion analysis and diagnosis. The accuracy of image segmentation and feature extraction directly impacts the subsequent accuracy of lesion severity and type determination.

[0049] For example, in coronary CT angiography, by precisely segmenting the image of the stenotic region and extracting its morphological and texture features, the stenosis rate can be accurately calculated and the plaque type can be determined as mixed or calcified. This approach not only improves the automation level of lesion detection and reduces tedious manual steps, but also significantly enhances detection accuracy and consistency. Fully automated processing eliminates inter- and intra-observer variability due to subjective factors, ensuring the objectivity and reliability of the test results. Furthermore, deep learning-based image segmentation and feature extraction technologies can fully leverage big data, resulting in robust models suitable for vascular lesion detection and analysis in diverse scenarios. In summary, inputting the lesion region into the analysis module and performing image segmentation and feature extraction are key technical steps in achieving fully automated, rapid, and accurate generation of vascular lesion diagnostic reports. This process not only improves diagnostic efficiency but also significantly enhances the accuracy and consistency of diagnostic results, providing strong support for clinical decision-making.

[0050] S105: The segmented image features include image data of calcified areas and / or non-calcified areas; the segmented image features are input into a calcified plaque analysis module and a non-calcified plaque analysis module, respectively. The segmented image features include image data of calcified areas and / or non-calcified areas.

[0051] S106 , determining the type of vascular lesion region corresponding to the segmented image feature based on the presence / absence of calcified plaque analysis result output by the calcified plaque analysis module and the presence / absence of non-calcified plaque analysis result output by the non-calcified plaque analysis module.

[0052] The calcified plaque analysis module determines the presence of calcified plaque by identifying high-density areas of the blood vessel wall.

[0053] For example, the calcified plaque analysis module uses threshold segmentation combined with morphological filtering to eliminate noise interference, ultimately outputting a binary result (1 for presence, 0 for absence). Similarly, the non-calcified plaque analysis module uses texture analysis (such as the gray-level co-occurrence matrix) to detect uneven signal distribution and incorporates an edge enhancement algorithm to identify plaque boundaries. If an area meeting the characteristics is detected, a non-calcified plaque flag is output; otherwise, it is marked as absent. Vascular lesion area type is determined based on the combined outputs of these two modules.

[0054] For example, if the calcification module outputs 1 and the non-calcification module outputs 0, the plaque is classified as purely calcified; if both outputs are 1, the plaque is classified as mixed (calcified and non-calcified). This process is implemented through logical AND / OR operations to ensure exclusive classification. Technically, this step-by-step determination can reduce misdiagnosis rates.

[0055] For example, separate analysis of calcification and non-calcification features can avoid misjudgments due to density overlap (such as low-density artifacts at the edge of calcification), while combined analysis can identify complex lesions (such as the "bull's-eye sign" of a calcified core surrounded by non-calcified tissue). The modular design also facilitates expansion into new types (such as the addition of a "Fiber Cap Analysis" submodule).

[0056] During the analysis process, the module also incorporates spatial context. For example, calcified plaques that are scattered and dotted are classified as stable, while non-calcified plaques that are continuously distributed along vessel bends indicate increased vulnerability. This association analysis is achieved through a region growing algorithm, ensuring that lesion type is aligned with clinical risk stratification.

[0057] The types of vascular lesion areas include four types: mixed plaque type, calcified plaque type, non-calcified plaque type and no plaque type.

[0058] First, mixed plaque types refer to the situation where calcified and non-calcified plaques coexist in the same vascular region.

[0059] For example, if a section of a blood vessel wall contains both high-density calcification deposits and a low-density lipid core, this can be considered a mixed plaque. Due to its complex composition and structure, this type of plaque often has a higher risk of rupture, so the test results require special attention.

[0060] Calcified plaque refers to a vascular lesion primarily composed of calcified components. For example, a calcified plaque with a diameter of approximately 5 mm on the vessel wall is relatively stable, but the resulting vascular stenosis can be severe, affecting blood perfusion.

[0061] Non-calcified plaque refers to areas of vascular lesions composed primarily of non-calcified components, such as lipids and fibrous tissue. For example, a low-density shadow with blurred boundaries and no obvious calcifications on a certain section of the vessel wall can be classified as non-calcified plaque. Non-calcified plaque, due to its soft composition, is prone to rupture and is a major risk factor for acute cardiovascular events.

[0062] Plaque-free refers to the absence of any plaque within a vessel region. For example, a smooth vessel wall without any abnormal density would be considered a plaque-free diagnosis. Plaque-free vessels generally indicate good health and normal blood perfusion within that segment, but this does not necessarily mean that other segments are also plaque-free. A comprehensive assessment of the overall vascular condition is necessary.

[0063] Through detailed analysis and examples of these four types, it can be seen that different types of vascular lesions have significant differences in imaging manifestations, pathological mechanisms, and clinical significance. Due to the complex composition of mixed plaque types, treatment strategies must comprehensively consider the characteristics of both calcified and non-calcified components. While calcified plaque types are relatively stable, they may lead to severe stenosis and require evaluation for interventional treatment. Non-calcified plaque types are prone to rupture and require close monitoring and active preventive measures. While plaque-free types indicate that the vessel segment is healthy, potential risks to other vascular segments remain a concern. This classification method not only helps to more accurately assess vascular health but also provides an important basis for developing personalized treatment plans.

[0064] This detailed classification allows users to address specific health issues, effectively reducing the incidence of cardiovascular events and improving quality of life. Furthermore, this classification method provides a unified standard for subsequent scientific research and clinical studies, facilitating data comparison and analysis across different studies, and further promoting advancements in vascular disease diagnosis and treatment technologies.

[0065] For example, through large-scale statistical analysis of clinical data, the natural course and treatment effects of different types of plaques can be explored, providing a scientific basis for the development of clinical guidelines. In short, the detailed classification and specific examples of regional vascular lesions not only help clinicians diagnose and treat vascular lesions more accurately, but also provide an important reference framework for scientific research in related fields, with far-reaching clinical and scientific significance.

[0066] S107. The mixed plaque type is determined by combining the analysis results of the calcified plaque with the calcified plaque output by the calcified plaque analysis module and the analysis results of the non-calcified plaque with the non-calcified plaque output by the non-calcified plaque analysis module; the calcified plaque type is determined by combining the analysis results of the calcified plaque with the calcified plaque output by the calcified plaque analysis module and the analysis results of the non-calcified plaque without the non-calcified plaque output by the non-calcified plaque analysis module; the non-calcified plaque type is determined by combining the analysis results of the non-calcified plaque without the calcified plaque output by the calcified plaque analysis module and the analysis results of the non-calcified plaque with the non-calcified plaque output by the non-calcified plaque analysis module; the non-plaque type is determined by combining the analysis results of the non-calcified plaque without the calcified plaque output by the calcified plaque analysis module and the analysis results of the non-calcified plaque without the non-calcified plaque output by the non-calcified plaque analysis module.

[0067] The determination of mixed plaque type relies on the coordinated output of the calcified plaque analysis module and the non-calcified plaque analysis module. When the calcified plaque analysis module detects high-density calcium salt deposits in the vessel wall, and the non-calcified plaque analysis module identifies a low-density lipid core or fibrous tissue, the system will combine the two results to determine mixed plaque type.

[0068] For example, if a certain display shows localized calcification and low-density shadows, the modules will interactively verify and generate a mixed plaque conclusion. This design avoids the limitations of a single module and ensures complete analysis of complex plaque components. To determine the type of calcified plaque, the calcification module must be positive and the non-calcification module must be negative. This logic eliminates interference from non-calcified components and accurately identifies calcified lesions, providing clear clinical indications for interventional treatment. The non-calcified plaque type is determined based on a negative calcification module and a positive non-calcification module. This mechanism is particularly suitable for identifying early-stage soft plaques, compensating for the risk of missing vulnerable plaques with traditional calcification testing. The generation of the plaque-free type requires dual verification: the calcification module finds no signs of calcification, and the non-calcification module detects no abnormal low-density areas. If a healthy subject's vascular image is analyzed without positive findings by both modules, the vessel segment is deemed plaque-free. This strict double-negative criterion significantly reduces the false positive rate, providing a reliable basis for screening healthy individuals. The collaborative design of the modules enables hierarchical judgment of plaque classification. The calcification module filters high-density signals using a threshold, while the non-calcification module identifies soft plaques based on texture analysis and density gradients. The results of both are combined through a logical AND / OR relationship to form four distinct output categories. For example, the determination of mixed plaques is essentially a logical "AND" relationship (calcification-positive AND non-calcification-positive), while the determination of calcified plaques is a logical "AND-NOT" relationship (calcification-positive AND non-calcification-negative). This structured decision-making process makes the system interpretable and facilitates the tracing of the basis for the determination. From a technical perspective, the dual-module parallel architecture offers three advantages over a single analysis: First, the independent detection of calcified and non-calcified components avoids signal interference; for example, calcification shadows will not obscure adjacent soft plaques. Second, the modular design allows for separate algorithm optimization. For example, morphological filtering can be used for calcification detection, while machine learning classification can be applied to non-calcification analysis. Third, the four-category results directly correspond to clinical detection and judgment strategies.

[0069] Furthermore, the method of analyzing the vascular image through a region recommendation module based on the image feature data to obtain the position and range of the lesion area, marking the lesion area, and generating a detection result including the coordinates of the lesion area includes: analyzing the vascular image through a region recommendation module based on the image feature data to determine the stenosis length and stenosis morphological characteristics of the lesion area, marking the lesion area, and generating a detection result including the coordinates of the lesion area and the distribution of the stenosis position.

[0070] The region proposal module uses a convolutional neural network to extract multi-level features from vascular images. For example, a backbone network is used to obtain feature maps at different scales from the original image. Shallow features capture the texture of vascular edges, while deep features identify the morphological characteristics of the lesion region. The feature maps are passed through the region proposal network to generate candidate boxes. Each candidate box contains the coordinates of the lesion center, as well as width and height parameters. For example, the output format is (x, y, w, h), where x = 120 pixels and y = 80 pixels represent the location of the lesion center in the image, and w = 30 pixels and h = 15 pixels describe the lesion extent. Stenosis length is calculated based on the geometric properties of the candidate boxes. If two adjacent candidate boxes overlap on the vessel centerline, their coordinate ranges are merged using the Euclidean distance formula. For example, if the coordinates of the first candidate box are (100, 50, 20, 10) and the second candidate box are (115, 55, 25, 12), the system determines that they belong to the same lesion segment. The stenosis length is calculated as 35 pixels (the total length after merging) and the distribution range is recorded as a continuous region from x = 90 to x = 125. Stenosis morphological feature analysis combines the aspect ratio of the candidate box and the angle with the vessel's orientation. When the aspect ratio is greater than 2 and the angle with the vessel centerline is less than 15 degrees, it is considered a linear stenosis. If the aspect ratio is close to 1 and the stenosis is irregular and convex, it is labeled as a focal stenosis. For example, if a candidate box with an aspect ratio of 3:1 and an angle of 10 degrees is detected, the system classifies it as a linear stenosis and adds a morphological label to the detection result. Lesion annotation uses a non-maximum suppression algorithm to select candidate boxes with a confidence level above 0.7, removing redundant detection results. The final output includes the lesion coordinates (e.g., [(120, 80, 30, 15)]), stenosis length (35 pixels), morphological type (linear stenosis), and confidence level (0.85). This data is stored in JSON format for easy access by subsequent diagnostic modules. The technical benefits are reflected in its global analysis capabilities. Traditional methods rely on point-by-point diameter calculations, while this solution directly obtains the overall spatial distribution of the lesion through object detection, avoiding stenosis rate deviations exceeding 20% ​​caused by single voxel errors. For example, for small blood vessels (only 5 pixels wide), the traditional method has an error of up to 40%, while the region recommendation module controls the error within 10% through contextual features.

[0071] Furthermore, the step of inputting the lesion area into the analysis module, performing image segmentation on the lesion area through the analysis module, and obtaining segmented image features includes: inputting the lesion area into the analysis module, performing pixel-level segmentation on the lesion area through the image segmentation module, and obtaining segmented image features including lesion boundary accuracy and regional texture features.

[0072] The image segmentation module performs pixel-level segmentation of the lesion area through a fully convolutional neural network. Pixel-level segmentation ensures the objectivity of lesion range measurement and avoids inter-observer differences (such as manual outlining stenosis rate fluctuations of up to 15%). Texture quantization reduces reliance on subjective experience (such as the traditional visual assessment error rate of approximately 20%). The combination of global and local features improves the detection rate of coherent lesions.

[0073] Furthermore, the acquiring of the vascular image dataset includes: acquiring an unlabeled vascular image dataset, wherein the unlabeled vascular image dataset includes an unlabeled vascular curved surface reconstructed image dataset.

[0074] Unlabeled vascular image datasets can be obtained by batch exporting historical angiography data from imaging archiving systems. For example, 1000 unlabeled CPR images from coronary CTA examinations can be extracted. These data should be stored in DICOM format after removing patient privacy information. Data sources should cover different scanning devices to enhance generalizability. The method for generating vascular surface reconstruction image datasets utilizes multiplanar reconstruction technology combined with a centerline extraction algorithm. Specifically, vascular structures are first separated using threshold segmentation (e.g., HU values ​​within the range of -200 to 1000), followed by centerline extraction, and finally, CPR images are generated by unwrapping along the centerline. For example, for coronary arteries with a diameter of 2-5 mm, a reconstruction slice thickness of 0.5 mm can preserve plaque details. The data preprocessing pipeline includes isotropic resampling (e.g., uniform voxel size of 0.3 × 0.3 × 0.3 mm), grayscale normalization (window width 800 HU / window level 200 HU), and random elastic deformation enhancement. In practice, a ±15% random transformation was applied to 2,000 unlabeled data sets to increase data diversity while preserving vascular topology. For unsupervised pre-training, CPR images were randomly sliced ​​into pre-defined voxel blocks, such as 64×64×64 voxel cubes, and the spatial order was shuffled before being fed into the 3D U-Net. For example, a batch of input samples contained 50% randomly rotated vessel segments and 30% samples with Gaussian noise (σ = 0.1), forcing the network to learn intrinsic anatomical features. A data distribution balancing approach employed stratified sampling based on vascular segment location (such as the left anterior descending artery and circumflex artery) to ensure that the sample proportions for each segment approximated the actual physiological distribution. Temporal and spatial consistency measures were implemented to associate timestamps between consecutive follow-up data (e.g., pre- and post-operative CPR) for the same user. For example, 300 paired samples separated by 6 months were included to enable the model to learn the temporal characteristics of plaque progression.

[0075] Furthermore, after the mixed plaque type is determined by combining the calcified plaque analysis results output by the calcified plaque analysis module and the non-calcified plaque analysis results output by the non-calcified plaque analysis module, it also includes: analyzing the proportion of non-calcified areas through the non-calcified plaque analysis module, and analyzing the proportion of calcified areas through the calcified plaque analysis module; determining the density distribution of non-calcified areas and calcified areas, and determining whether the mixed plaque type is a spliced ​​mixed plaque type or a uniform mixed plaque type.

[0076] The synergistic effect of the calcified plaque analysis module and the non-calcified plaque analysis module generates four basic classification results by detecting the presence of calcified and non-calcified areas respectively. When both modules detect calcified and non-calcified areas at the same time, the system determines it as a mixed plaque type. At this time, it is necessary to further analyze the proportion and distribution characteristics of the two types of areas to distinguish between spliced ​​mixed plaques and uniform mixed plaques. The analysis of the proportion of non-calcified areas is achieved through the non-calcified plaque analysis module. For example, in vascular images, the module identifies low-density areas (such as lipid cores or fibrous tissue) and calculates their proportion of the total plaque area. If the non-calcified area of ​​a plaque accounts for 60% and the calcified area accounts for 40%, it means that the non-calcified component is dominant.

[0077] The calcified plaque analysis module analyzes the proportion of calcified areas, which detects high-density areas (such as hydroxyapatite deposits). For example, if a plaque has scattered calcified areas, accounting for 35%, while non-calcified areas account for 65%, the density distribution must be considered to determine the type of plaque. Density distribution analysis is performed using spatial statistics. If the boundary between calcified and non-calcified areas is clear (e.g., calcified areas are concentrated on one side, non-calcified areas on the other), the plaque is considered a patchwork plaque. If the two types of areas are interlaced and there is no clear demarcation (e.g., calcified areas are evenly distributed within a non-calcified matrix), the plaque is considered a homogeneous mixed plaque.

[0078] For example, a coronary CT scan shows a plaque with a clearly defined proximal calcified area (density >130 HU) and a distal non-calcified area (density <60 HU), which is classified as a patchwork mixed plaque. Distinguishing plaque types is clinically important in guiding treatment strategies. A patchwork mixed plaque may indicate the stage of disease progression, requiring attention to the stability of the calcified portion. A homogeneously mixed plaque may reflect active inflammation, requiring intensive lipid-lowering therapy.

[0079] For example, within a homogeneous mixed plaque, small variations in calcification density (e.g., a mean of 90 HU ± 10) indicate similar calcification levels and are therefore amenable to drug intervention. Technically, density distribution can be analyzed using grayscale histograms or region growing methods.

[0080] For example, a patch image is divided into grids, and the HU value of each grid cell is counted. If the high / low density unit aggregation index is greater than 0.7, it is judged as spliced ​​mixing, otherwise it is uniform mixing.

[0081] Furthermore, the method provided in this embodiment also includes: performing regional contrast enhancement on the segmented image features through a regional refinement module, optimizing the lesion region volume and the segmentation region consistency, and generating refined lesion region image data.

[0082] Furthermore, it also includes: generating a final analysis result including the type of vascular lesion area and the degree of stenosis based on the refined lesion area image data, combined with the stenosis area contrast and lesion area classification label.

[0083] To enhance the regional contrast of the segmented image features through the regional refinement module, the original vascular image data must first be preprocessed. Preprocessing includes denoising and enhancing the image contrast to better highlight the lesion area.

[0084] For example, median filtering can be used to remove random noise from an image, followed by histogram equalization to enhance the overall contrast of the image. This makes the previously blurred lesion area appear clearer in the image, facilitating subsequent detailed analysis. After preprocessing, the core step of the regional refinement module is entered: contrast enhancement. The purpose of contrast enhancement is to make the contrast between the lesion area and the surrounding normal tissue more distinct, facilitating subsequent segmentation and identification. In specific operations, an adaptive contrast enhancement algorithm can be used, which dynamically adjusts the enhancement strength based on the local contrast of the image.

[0085] For example, the contrast enhancement will be more intense at the edge of the lesion area, while it will be moderately enhanced in relatively uniform areas, thus avoiding image distortion caused by over-enhancement. Optimizing the volume of the lesion area and the consistency of the segmented area is a key step in fine-tuning the image. In this step, morphological operations such as dilation and erosion can be used to adjust the volume of the lesion area. Assuming that the initial volume of a lesion area is 100 pixels, the dilation operation can expand it to 120 pixels to include more lesion edge information; the erosion operation can reduce it to 90 pixels to remove noise interference at the edge. Through repeated iterations and adjustments, a lesion area is finally obtained that contains complete lesion information while minimizing redundancy.

[0086] The optimization of segmentation region consistency relies on region growing or level set segmentation algorithms.

[0087] For example, a seed point is selected in the lesion area and gradually expanded to the entire lesion area based on similarity criteria such as grayscale value or texture features. During this process, the boundaries of the segmented regions are ensured to be smooth and continuous, avoiding breaks or overlaps, thereby improving segmentation accuracy. After generating refined image data of the lesion area, the final analysis results are generated. First, based on the refined image data of the lesion area, the contrast features of the narrow area are extracted. The contrast features can be quantified by calculating the grayscale mean difference between the lesion area and the surrounding normal tissue.

[0088] For example, if the mean grayscale value of the lesion is 150 and the mean grayscale value of the surrounding normal tissue is 100, the contrast eigenvalue is 50. A larger eigenvalue indicates higher contrast in the lesion and more pronounced lesions. Combining the contrast of the stenotic area with the lesion classification label generates a final analysis that includes the type of vascular lesion and the degree of stenosis. Classification labels can be trained and predicted using machine learning algorithms such as support vector machines or convolutional neural networks. Assuming the training set contains three lesion types: calcified, non-calcified, and mixed, the model can be trained to generate a classification label for each lesion.

[0089] For example, if a lesion is classified as "mixed," combined with a contrast feature value of 50, it can be inferred that the stenosis in the lesion is moderate to severe. The final analysis results also require comprehensive consideration of multiple feature indicators, such as the area, morphology, and location of the lesion.

[0090] For example, a lesion area of ​​200 pixels, irregular shape, and located at a bend in a blood vessel can provide a more comprehensive assessment of lesion severity and inform treatment planning. The above steps not only achieve refined processing of the lesion area but also generate a final analysis result rich in information, providing strong support for clinical diagnosis. Contrast enhancement and regional consistency optimization ensure accurate lesion identification, while comprehensive analysis combining multiple features improves the reliability and comprehensiveness of diagnostic results.

[0091] like Figure 2 As shown, an embodiment of the present invention further provides a device for analyzing blood vessel characteristics based on target detection, comprising: An unsupervised training unlabeled data module 210 is configured to obtain a vascular image dataset, perform image restoration on images in the vascular image dataset using a 3DU-Net model, generate pre-training parameters, and initialize the encoding portion of a feature extraction module based on the pre-training parameters to obtain an initialized feature extraction module. The pre-trained model image feature extraction module 220 processes the input vascular image through the initialized feature extraction module generated by the unsupervised training unlabeled data module 210 to generate a pre-trained model, and uses the pre-trained model to extract features from the vascular image to obtain image feature data; a lesion region positioning module 230 for analyzing the image feature data extracted by the image feature extraction module 220 based on the pre-trained model, obtaining the location and range of the lesion region through the region recommendation module, annotating the lesion region, and generating a detection result including the coordinates of the lesion region; The analysis module 240 is configured to input the lesion region, perform image segmentation on the lesion region, and obtain segmented image features; the segmented image features include image data of calcified regions and / or non-calcified regions; and further configured to input the segmented image features into a calcified plaque analysis module and a non-calcified plaque analysis module, respectively. The determination module 250 includes the calcified plaque analysis module and the non-calcified plaque analysis module, and is configured to determine the type of vascular lesion region corresponding to the segmented image feature based on the analysis result of the presence / absence of calcified plaque output by the calcified plaque analysis module and the analysis result of the presence / absence of non-calcified plaque output by the non-calcified plaque analysis module.

[0092] The types of vascular lesion areas include four types: mixed plaque type, calcified plaque type, non-calcified plaque type and no plaque type.

[0093] Implementation of the Unsupervised Training Unlabeled Data Module 210: The core of this module is to utilize a large number of unlabeled vascular images for self-supervised learning. Specifically, the original 3D vascular images are disrupted by random rotation, translation, or noise addition to generate corrupted input data.

[0094] For example, a 200×200×100 region is randomly cropped from a 512×512×300 voxel image, its spatial order shuffled, and then fed into a 3D U-Net for restoration training. The encoder employs five layers of downsampling, with each layer having a convolution kernel size of 3×3×3 and a stride of 2, progressively extracting 64 / 128 / 256 / 512 / 1024-dimensional features. The mean squared error (MSE) loss function optimizes network parameters by comparing the voxel value differences between the output image and the original image. This process enables the model to learn common features such as vessel wall texture and branching structure. The resulting encoder parameters can be used as initialization weights for downstream tasks.

[0095] Specific implementation of the pre-trained model image feature extraction module 220: The initialized 3D U-Net encoder is used to process newly input vascular images. For example, when extracting features from a coronary CT image, the model inputs a LAD vessel segment (150×150×80 pixels). The 256-channel feature map output by the third encoder layer shows significant activation responses at the lumen-plaque boundary. While the spatial resolution of the feature map is reduced to 1 / 8 that of the original image, it retains the high-density features of calcified areas (CT values ​​>130 HU) and the low-contrast features of non-calcified areas. This module, by transferring pre-trained knowledge, eliminates the need to learn basic features from scratch for subsequent lesion detection, improving generalization capabilities even with small sample sizes.

[0096] Technical details of the lesion area localization module 230: The region recommendation module adopts the RPN (region proposal network) structure and slides a 3×3×3 convolution kernel on the feature map to generate candidate regions.

[0097] For example, for the aforementioned 256-channel feature map, nine anchor boxes of varying scales (e.g., 5×5×5 to 20×20×20 mm³) are predicted for each spatial location. A binary classification is then used to determine whether or not the lesion is contained. In practice, when detecting the mid-segment of an RCA vessel, three candidate boxes might be output, with center coordinates of (45, 60, 32), (48, 58, 35), and (50, 62, 30). The intersection-over-union (IoU) threshold is set at 0.7 to filter the final results. This design directly utilizes global features to locate lesions, avoiding the error accumulation problem associated with traditional methods that rely on centerline extraction.

[0098] The segmentation and feature extraction process of the analysis module 240: an improved 3D segmentation network is used to process the located lesion area.

[0099] For example, a candidate region measuring 30×30×30 mm³ undergoes dual segmentation: one path uses 1×1×1 convolution to extract calcified areas, while the other uses dilated convolution to capture the texture features of non-calcified plaques. Experimental data shows that this module achieves 89.2% segmentation accuracy for mixed plaques, a 12% improvement over a single-path approach. The segmented feature maps are then fed into subsequent analysis submodules to ensure feature independence for different plaque types.

[0100] An embodiment of the present invention further provides a non-volatile computer-readable storage medium having computer-executable instructions stored thereon. When the computer-executable instructions are executed by a processor, the following steps are implemented: Obtaining a vascular image dataset, performing an image restoration task on images in the vascular image dataset using a 3DU-Net model to generate pre-training parameters, and initializing an encoding portion of a feature extraction module according to the pre-training parameters to obtain an initialized feature extraction module; Processing the input vascular image through the initialized feature extraction module to generate a pre-trained model, and extracting features from the vascular image using the pre-trained model to obtain image feature data; Analyzing the blood vessel image using a region recommendation module based on the image feature data to obtain the location and range of the lesion area, marking the lesion area, and generating a detection result including the coordinates of the lesion area; Inputting the lesion area into an analysis module, performing image segmentation on the lesion area by the analysis module to obtain segmented image features; the segmented image features include image data of calcified areas and / or non-calcified areas; Inputting the segmented image features into a calcified plaque analysis module and a non-calcified plaque analysis module respectively; determining the type of vascular lesion region corresponding to the segmented image features based on the presence / absence of calcified plaque analysis results output by the calcified plaque analysis module, and based on the presence / absence of non-calcified plaque analysis results output by the non-calcified plaque analysis module; The types of vascular lesion areas include four types: mixed plaque type, calcified plaque type, non-calcified plaque type and no plaque type.

[0101] like Figure 3 As shown, an embodiment of the present invention further provides a computer device, which may include: a memory 301 storing executable program code; a processor 302 coupled to the memory 301; The processor 302 calls the executable program code stored in the memory 301 to execute the program based on Figure 1 Partial steps of any of the target detection-based vascular feature analysis methods shown.

[0102] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for analyzing vascular characteristics based on target detection, characterized in that: include: Acquiring a vascular image dataset, performing an image restoration task on images in the vascular image dataset, generating pre-training parameters, and initializing an encoding portion of a feature extraction module according to the pre-training parameters to obtain an initialized feature extraction module; Processing the input vascular image through the initialized feature extraction module to generate a pre-trained model, and extracting features from the vascular image using the pre-trained model to obtain image feature data; Analyzing the blood vessel image using a region recommendation module based on the image feature data to obtain the location and range of the lesion area, marking the lesion area, and generating a detection result including the coordinates of the lesion area; Inputting the lesion area into an analysis module, performing image segmentation on the lesion area by the analysis module to obtain segmented image features; the segmented image features include image data of calcified areas and / or non-calcified areas; Inputting the segmented image features into a calcified plaque analysis module and a non-calcified plaque analysis module respectively; Determining the type of vascular lesion area corresponding to the segmented image features by the calcified plaque analysis module and the non-calcified plaque analysis module; The types of vascular lesion areas include four types: mixed plaque type, calcified plaque type, non-calcified plaque type and no plaque type.

2. The method for analyzing blood vessel characteristics based on target detection according to claim 1, characterized in that: The determining, by the calcified plaque analysis module and the non-calcified plaque analysis module, the type of vascular lesion region corresponding to the segmented image feature includes: The type of vascular lesion region corresponding to the segmented image feature is determined based on the presence / absence of calcified plaque analysis results output by the calcified plaque analysis module and the presence / absence of non-calcified plaque analysis results output by the non-calcified plaque analysis module.

3. The method for analyzing blood vessel characteristics based on target detection according to claim 2, characterized in that: The mixed plaque type is determined by combining the calcified plaque analysis result output by the calcified plaque analysis module and the non-calcified plaque analysis result output by the non-calcified plaque analysis module; The calcified plaque type is determined by combining the calcified plaque analysis result output by the calcified plaque analysis module and the non-calcified plaque analysis result output by the non-calcified plaque analysis module; The non-calcified plaque type is determined by combining the non-calcified plaque analysis result output by the calcified plaque analysis module and the non-calcified plaque analysis result output by the non-calcified plaque analysis module; The non-plaque type is determined by combining the non-calcified plaque analysis result output by the calcified plaque analysis module and the non-non-calcified plaque analysis result output by the non-calcified plaque analysis module.

4. The method for analyzing blood vessel characteristics based on target detection according to claim 1, wherein: The method includes analyzing the blood vessel image through a region recommendation module based on the image feature data to obtain the location and range of the lesion area, marking the lesion area, and generating a detection result including the coordinates of the lesion area, including: Based on the image feature data, the region recommendation module analyzes the vascular image, determines the stenosis length and stenosis morphological characteristics of the lesion area, marks the lesion area, and generates a detection result containing the coordinates of the lesion area and the distribution of the stenosis location.

5. The method for analyzing blood vessel characteristics based on target detection according to claim 1, characterized in that: The step of inputting the lesion area into an analysis module and performing image segmentation on the lesion area by the analysis module to obtain segmented image features includes: The lesion area is input into the analysis module, and the image segmentation module performs pixel-level segmentation on the lesion area to obtain segmentation image features including lesion boundary accuracy and regional texture features.

6. The method for analyzing blood vessel characteristics based on target detection according to claim 1, characterized in that: The obtaining of a blood vessel image dataset includes: An unlabeled blood vessel image dataset is acquired, wherein the unlabeled blood vessel image dataset includes an unlabeled blood vessel curved surface reconstructed image dataset.

7. The method for analyzing blood vessel characteristics based on target detection according to claim 3, characterized in that: After the mixed plaque type is determined by combining the calcified plaque analysis result output by the calcified plaque analysis module and the non-calcified plaque analysis result output by the non-calcified plaque analysis module, the method further includes: Analyzing the proportion of non-calcified areas by the non-calcified plaque analysis module, and analyzing the proportion of calcified areas by the calcified plaque analysis module; Determine the density distribution of the non-calcified area and the calcified area, and determine whether the mixed plaque type is a spliced ​​mixed plaque type or a uniform mixed plaque type.

8. The method for analyzing blood vessel characteristics based on target detection according to claim 1, characterized in that: Also includes: The regional refinement module is used to enhance the regional contrast of the segmented image features, optimize the volume of the lesion area and the consistency of the segmented area, and generate refined lesion area image data.

9. The method for analyzing blood vessel characteristics based on target detection according to claim 8, characterized in that: Also includes: Based on the refined image data of the lesion area, combined with the contrast of the stenosis area and the classification label of the lesion area, the final analysis results including the type of vascular lesion area and the degree of stenosis are generated.

10. A device for analyzing blood vessel characteristics based on target detection, characterized in that: include: an unsupervised training unlabeled data module, configured to obtain a vascular image dataset, perform image restoration on images in the vascular image dataset, generate pre-training parameters, and initialize an encoding portion of a feature extraction module according to the pre-training parameters to obtain an initialized feature extraction module; a pre-trained model image feature extraction module, which processes the input vascular image through the initialized feature extraction module generated by the unsupervised training unlabeled data module to generate a pre-trained model, and uses the pre-trained model to extract features from the vascular image to obtain image feature data; a lesion region positioning module, configured to analyze the vascular image using the region recommendation module based on the image feature data extracted by the image feature extraction module according to the pre-trained model, obtain the location and range of the lesion region, annotate the lesion region, and generate a detection result including the coordinates of the lesion region; an analysis module, configured to input the lesion area, perform image segmentation on the lesion area through the analysis module, and obtain segmented image features; the segmented image features include image data of calcified areas and / or non-calcified areas; and further configured to input the segmented image features into a calcified plaque analysis module and a non-calcified plaque analysis module, respectively; a determination module, comprising the calcified plaque analysis module and the non-calcified plaque analysis module, configured to determine the type of vascular lesion region corresponding to the segmented image feature based on the presence / absence of calcified plaque analysis result output by the calcified plaque analysis module and the presence / absence of non-calcified plaque analysis result output by the non-calcified plaque analysis module; The types of vascular lesion areas include four types: mixed plaque type, calcified plaque type, non-calcified plaque type and no plaque type.

11. A non-volatile computer-readable storage medium having computer-executable instructions stored thereon, characterized in that: When the computer executable instructions are executed by a processor, the following steps are implemented: Acquiring a vascular image dataset, performing an image restoration task on images in the vascular image dataset, generating pre-training parameters, and initializing an encoding portion of a feature extraction module according to the pre-training parameters to obtain an initialized feature extraction module; Processing the input vascular image through the initialized feature extraction module to generate a pre-trained model, and extracting features from the vascular image using the pre-trained model to obtain image feature data; Analyzing the blood vessel image using a region recommendation module based on the image feature data to obtain the location and range of the lesion area, marking the lesion area, and generating a detection result including the coordinates of the lesion area; Inputting the lesion area into an analysis module, performing image segmentation on the lesion area by the analysis module to obtain segmented image features; the segmented image features include image data of calcified areas and / or non-calcified areas; Inputting the segmented image features into a calcified plaque analysis module and a non-calcified plaque analysis module respectively; Determining the type of vascular lesion region corresponding to the segmented image feature based on the presence / absence of calcified plaque analysis results output by the calcified plaque analysis module and the presence / absence of non-calcified plaque analysis results output by the non-calcified plaque analysis module; The types of vascular lesion areas include four types: mixed plaque type, calcified plaque type, non-calcified plaque type and no plaque type.

12. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 9 are implemented.

Images