A method for carotid stenosis detection and quantification based on digital subtraction angiography

Abstract

The application discloses a carotid artery stenosis detection and quantification method based on digital subtraction angiography, comprising: detecting the stenosis blood vessels and normal blood vessels of the carotid artery in the image of the digital subtraction angiography by using the classification branch of the target detection model; wherein the detection head of the target detection model comprises a classification branch, a Match-ness branch and a regression branch; the matching degree between multiple normal blood vessels and a single stenosis blood vessel is respectively predicted by using the Match-ness branch in the detection head; the stenosis blood vessel and the normal blood vessel with the highest matching degree with the stenosis blood vessel are intercepted and input into the regression model to predict the morphological indexes of the carotid artery stenosis, and the quantitative analysis result of the carotid artery stenosis is obtained. The matching relationship between the carotid artery stenosis and multiple normal blood vessel segments in the blood vessel tree is obtained by the target detection model, the stenosis blood vessel and the normal blood vessel suitable for predicting the reference blood vessel diameter of the stenosis are accurately detected, and the accurate quantification of the stenosis morphological indexes is further carried out.

Description

Technical Field

[0001] This invention relates to the fields of medical image analysis and computer vision technology, and in particular to a method for detecting and quantifying carotid artery stenosis based on digital subtraction angiography. Background Technology

[0002] Carotid artery stenosis is a vascular disease that reduces blood flow to the brain and is a leading cause of stroke. The morbidity and mortality rates of stroke are very high in cases of severe carotid artery stenosis. When patients are diagnosed with high-grade carotid artery stenosis—for example, symptomatic patients with stenosis exceeding 50% or asymptomatic patients with stenosis exceeding 70%—they require aggressive endovascular treatment or carotid endarterectomy for carotid revascularization. Therefore, accurately quantifying the degree of carotid artery stenosis is crucial for timely intervention and stroke prevention. Traditional manual screening is a tedious and error-prone process, characterized by high labor costs and slow screening speed. With the accumulation of imaging data and the improvement of computing power, computer-aided diagnosis of carotid artery stenosis using deep learning methods has shown great potential.

[0003] In clinical practice, digital subtraction angiography (DSA) is one of the main imaging methods for guiding diagnosis and treatment by visualizing the morphology of the carotid arteries. However, in current DSA-based research, few studies use deep learning to detect carotid artery stenosis and perform direct quantitative analysis. Most studies focus primarily on the grading of arterial lesions at the image level, providing a rough description of the degree of stenosis rather than precise localization and quantification. These studies can only provide a preliminary diagnosis of the degree of stenosis and cannot obtain morphological indicators of stenosis such as reference vessel diameter and minimum lumen diameter for further guidance in clinical diagnosis and decision-making. Summary of the Invention

[0004] This invention provides a method for detecting and quantifying carotid artery stenosis based on digital subtraction angiography. By using a target detection model, the matching relationship between carotid artery stenosis and multiple normal vessel segments in the vascular tree is obtained, and the stenotic vessels and normal vessels with reference vessel diameters suitable for predicting stenosis are accurately detected. Furthermore, the morphological indicators of stenosis are accurately quantified.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for detecting and quantifying carotid artery stenosis based on digital subtraction angiography, comprising the following steps:

[0007] S1, using the classification branch of the target detection model to detect stenotic and normal vessels in the carotid artery in digital subtraction angiography images; the detection head of the target detection model includes a classification branch, a match-ness branch, and a regression branch;

[0008] S2, using the Match-ness branch in the detection head to predict the degree of matching between multiple normal blood vessels and a single stenotic blood vessel, and to determine the normal blood vessel with the highest degree of matching with the stenotic blood vessel;

[0009] S3: Extract the stenotic vessel and the normal vessel with the highest matching degree to the stenotic vessel, and input them into the regression model to predict the morphological indicators of carotid artery stenosis, and obtain the quantitative analysis results of carotid artery stenosis.

[0010] Further, S1, utilizing the classification branch of the target detection model to detect stenotic and normal vessels in the carotid artery in digital subtraction angiography images, specifically includes:

[0011] Based on digital subtraction angiography images, the stenotic and normal vessels of the carotid artery are identified according to morphological features, and the ground truth bounding boxes of the target detection are labeled according to the stenosis class and the normal class.

[0012] Each digital subtraction angiography image is labeled with a stenosis class ground truth bounding box and multiple normal class ground truth bounding boxes. When training the target detection model, the normal class ground truth bounding boxes are additionally labeled with the ground truth value of the matching degree with the stenotic blood vessels in the current image. The matching degree of normal blood vessels suitable for predicting the reference blood vessel diameter of stenosis is labeled as 1, and the matching degree of other normal blood vessels is labeled as 0.

[0013] Further, in step S2, the Match-ness branch in the detection head is used to predict the degree of matching between multiple normal blood vessels and a single stenotic blood vessel, and the normal blood vessel with the highest degree of matching with the stenotic blood vessel is determined, specifically including:

[0014] When training the object detection model, the matching score output by the match-ness branch is adjusted to between 0 and 1 using the sigmoid activation function, and the loss is calculated by comparing it with the ground truth value of the matching degree of the label.

[0015] Using the Match-ness branch in the trained object detection model, a matching score is predicted for each normal class prediction box based on the currently predicted normal blood vessel. The normal blood vessel with the highest matching score is taken as the normal blood vessel with the highest degree of matching with the stenotic blood vessel.

[0016] Further, in step S3, the stenotic vessel and the normal vessel with the highest matching degree to the stenotic vessel are intercepted and input into a regression model to predict the morphological indicators of carotid artery stenosis, thereby obtaining the quantitative analysis results of carotid artery stenosis, specifically including:

[0017] From digital subtraction angiography images, the stenotic vessels predicted by the target detection model and the normal vessels that best match the stenotic vessels are extracted and fed into a regression model to predict the morphological indicators of carotid artery stenosis. The morphological indicators include the minimum lumen diameter of the stenotic vessels and the reference vessel diameter of the normal vessels that best match the stenotic vessels.

[0018] The degree of stenosis of the narrowed vessel is obtained based on the minimum lumen diameter of the narrowed vessel and the reference vessel diameter, enabling quantitative analysis.

[0019] Furthermore, S3 also includes:

[0020] The image of digital subtraction angiography was denoised using a nonlocal mean algorithm, and then the image was filled to a 1:1 aspect ratio using a gray value similar to the background of the digital subtraction angiography image.

[0021] A deep convolutional neural network is used to extract features from images, and finally a multi-layer fully connected layer is used to predict morphological indicators. The mean squared error loss function is used to train the object detection model.

[0022] Furthermore, step S2 also includes: during the training of the target detection model, introducing prior knowledge based on blood vessel trends to improve the performance of the target detection model during the calculation of the regression loss, specifically:

[0023] The direction of blood vessel movement is obtained by connecting the center points of the narrow ground truth bounding box and the normal ground truth bounding box;

[0024] During the network training process of the object detection model, the second direction is obtained by connecting the center points of the normal class ground truth bounding boxes and the normal class predicted boxes;

[0025] A penalty term is added to the D-IoU loss of the target detection model. The penalty term is maximized when the angle between the direction of blood vessel movement and the second direction is close to 0 degrees, and minimized when it is close to 90 degrees.

[0026] Furthermore, the penalty items include:

[0027] The dot product of vectors is used to calculate the angle between the direction of blood vessel movement and the second direction, which is denoted as Angle;

[0028] The penalty term is obtained by calculating the sine of the included angle. The formula for the penalty term P is as follows:

[0029] P = a * (1 - sin(Angle)) (4)

[0030] Where α represents the hyperparameter for weighting the penalty term;

[0031] The final bounding box regression loss function L 2DIoU loss The formula is shown below:

[0032] L 2DIoU loss =L DIoU +P=1-IoU+R(B pd B gt )+a*(1-sin(Angle)) (5)

[0033] Among them, L DIoU The D-IoU loss function is shown in the following formula:

[0034] L DIoU =1-IoU+R(B pd B gt (6)

[0035]

[0036] Among them, B pd It is a prediction box, B gt It is the ground truth bounding box; b pd and b gt These are the center coordinates of the predicted bounding box and the ground truth bounding box, respectively, d 2 (b pd b gt R(B) represents the Euclidean distance between the centers of the two bounding boxes; c is the diagonal length of the smallest bounding rectangle containing the predicted box and the ground truth bounding box; pd B gt ) is a penalty term added to the IoU loss.

[0037] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The carotid artery stenosis detection and quantification method based on digital subtraction angiography provided by the present invention decouples the carotid artery detection task into the separate detection of stenosis and normal vessel segments suitable for predicting reference vessel diameter, further providing accurate direct prediction of morphological indicators; it uses matching degree labels of normal vessel classes to model and obtain the matching relationship between stenotic vessels and multiple normal vessels, thereby introducing prior knowledge of the reference vessel diameter for measuring stenosis; it uses the Match-ness branch in the target detection head to predict the matching degree between normal vessels and stenosis, and performs effective supervised learning together with the classification branch and regression branch; it uses prior knowledge based on vessel trend to improve the regression loss, making the model pay more attention to the bounding box regression along the vessel trend direction, thus improving the accuracy of the target detection model. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a schematic flowchart of the method for detecting and quantifying carotid artery stenosis based on digital subtraction angiography according to the present invention.

[0040] Figure 2 This is a schematic diagram illustrating the design and use of the Match-ness branch provided in an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of the overall structure of the target detection model under two different working stages provided in the embodiments of the present invention;

[0042] Figure 4 This is a schematic diagram of the improved IoU loss provided in an embodiment of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] This invention proposes a deep learning-assisted diagnostic scheme for carotid artery stenosis based on DSA images. First, a target detection model is used to detect stenosis and matching normal blood vessels, allowing the model to focus on the local area. Then, a regression model is used to predict the reference vessel diameter and minimum lumen diameter, achieving for the first time accurate detection and quantitative analysis of carotid artery stenosis in DSA images.

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] like Figure 1 As shown, the method for detecting and quantifying carotid artery stenosis based on digital subtraction angiography provided by this invention includes the following steps:

[0047] S1, the classification branch of the target detection model is used to detect stenotic and normal vessels in the carotid artery in digital subtraction angiography images. The target detection model is built using the classic anchor-based one-stage target detection framework. Specifically, firstly, a general 2D image backbone is used to perform preliminary feature extraction on the input carotid artery digital subtraction angiography image to obtain a multi-scale feature map. Then, a feature pyramid structure is used for feature enhancement and interactive learning between multi-scale features. The output feature map is fed into the detection head to predict the target's category, bounding box center point, and size. The detection head includes a classification branch, a match-ness branch, and a regression branch. Furthermore, the match-ness branch is parallel to the classification and regression branches.

[0048] like Figure 2 As shown, the data annotation for carotid artery stenosis detection is divided into two categories: stenosis and normal vessels. The ground truth bounding boxes for target detection are also labeled according to these categories. The classification of stenotic and normal vessels is determined by the morphological features of the carotid artery digital subtraction angiography (DSA) images. Specifically, from the DSA images, stenotic vessels exhibit an image feature that progresses from thick to thin and back to thick, while normal vessels have uniform, parallel, or nearly parallel vessel walls. The normal vessel category represents a normal vessel segment with a potential reference vessel diameter suitable for stenosis. Each DSA image is labeled with one stenosis bounding box and multiple normal vessel bounding boxes. The normal vessel bounding boxes are further annotated with their degree of matching with the stenotic vessels in the current image, denoted as CT. Ms∈[0,1], where the ground truth of the matching degree is obtained from clinical experience. That is, the matching degree of the normal blood vessel that is most suitable for predicting the reference blood vessel diameter for stenosis is labeled as 1, and the matching degree of other normal blood vessels is labeled as 0. During the training of the target detection model, the output of the match-ness branch will use the sigmoid activation function to adjust the output value to between 0 and 1, and calculate the loss with the ground truth of the labeled matching degree. The formula for labeling the matching degree is as follows:

[0049]

[0050] S2, using the Match-ness branch in the detection head to predict the degree of matching between multiple normal blood vessels and a single stenotic blood vessel, and to determine the normal blood vessel with the highest degree of matching with the stenotic blood vessel;

[0051] like Figure 2 As shown, a Match-ness branch is added to the detection head of the target detection model to predict a matching score for each normal blood vessel prediction box. The matching score on the prediction box is the probability prediction value of the prediction box most likely belonging to a certain category. The matching score is a value between 0 and 1. The larger the value, the more suitable the normal blood vessel detection box is for predicting the reference diameter of the stenotic vessel, that is, the higher the matching degree between the normal vessel and the stenotic vessel. Specifically, a Match-ness branch is added to the detection head in a manner parallel to the classification branch and the regression branch, and a Sigmoid activation function is applied to constrain the output to the range [0,1]. The score m predicted by the Match-ness branch indicates the matching degree between the normal vessel and the stenotic vessel. The higher the m, the more suitable the region is for predicting the reference diameter of the stenotic vessel. The Match-ness branch is trained in parallel with the classification branch and the regression branch in the detection head, and is supervised using the ground truth of the matching degree marked in S1 and the Match-ness branch loss. The Match-ness branch loss includes:

[0052] The output m of the Match-ness branch is between 0 and 1, and can be conveniently trained using the Binary Cross-Entropy (BCE) loss. The loss formula is shown below:

[0053]

[0054] Where N represents the number of samples, m i This represents the output of the Match-ness branch for the i-th sample. This represents the true value of the matching degree of the i-th sample.

[0055] S3. The stenotic vessel and the normal vessel with the highest matching degree to the stenotic vessel were cut off, and the morphological indicators of carotid artery stenosis were predicted by regression model to obtain the quantitative analysis results of carotid artery stenosis.

[0056] Figure 3 This diagram illustrates the workflow of the target detection model in the detection and quantitative analysis of carotid artery stenosis. The model extracts the stenotic vessels predicted by the target detection model and the normal vessels with the highest matching scores from the DSA image. These are then fed into a regression model to predict the vessel diameter, i.e., the morphological index of stenosis. The extracted stenosis prediction results are used to predict the minimum lumen diameter of the carotid artery, while the normal vessel prediction results are used to predict the reference vessel diameter for carotid artery stenosis. Specifically, a nonlocal means algorithm is used to denoise the image, and then the image is padded to a 1:1 aspect ratio using grayscale values ​​similar to the DSA image background. A deep CNN is then used to extract image features, and finally, multiple fully connected layers are used to predict the quantitative index. The mean squared error (MSE) loss function is used to train the model. After obtaining accurate quantitative results of the carotid artery stenosis morphological index, the degree of stenosis can be directly calculated by combining the minimum lumen diameter and the reference vessel diameter. The loss function of the regression model is shown below:

[0057]

[0058] Among them, y i This represents the regression model's prediction of the blood vessel diameter for the i-th sample. The label represents the diameter of the blood vessel in the i-th sample, and N represents the number of samples.

[0059] Step S2 further includes: during the training of the target detection model, incorporating prior knowledge based on blood vessel trends to improve the performance of the target detection model during the calculation of the regression loss, specifically:

[0060] The rectangular bounding box annotations of object detection cannot perfectly fit the carotid artery structure in DSA images. Therefore, prior knowledge based on the vessel trend is introduced into the regression loss calculation of the object detection model, making the model focus more on the bounding box regression along the vessel trend direction. For example... Figure 4As shown, the direction of blood vessel movement is obtained by connecting the center points of the ground truth bounding boxes of the narrow class and the normal class. During network training, a second direction is obtained by connecting the center points of the ground truth bounding boxes of the normal class and the predicted boxes of the normal class. A penalty term is added to the D-IoU loss of object detection. This penalty term is maximized when the angle between the blood vessel movement direction and the second direction is close to 0 degrees and minimized when it is close to 90 degrees. The introduced penalty term increases the loss weight of bounding box regression along the blood vessel movement direction, making the model pay more attention to bounding box regression along the blood vessel movement direction, thus improving the model's performance.

[0061] The penalties include:

[0062] The dot product of vectors is used to calculate the angle between the direction of blood vessel movement and the second direction, which is denoted as Angle;

[0063] The penalty term is obtained by calculating the sine of the included angle. The formula for the penalty term P is as follows:

[0064] P = a * (1 - sin(Angle)) (4)

[0065] Where α represents the hyperparameter for weighting the penalty term;

[0066] The final bounding box regression loss function L 2DIoU loss The formula is shown below:

[0067] L 2DIoU loss =L DIoU +P=1-IoU+R(B pd B gt )+a*(1-sin(Angle)) (5)

[0068] Among them, L DIoU The D-IoU loss function is shown in the following formula:

[0069] L DIoU =1-IoU+R(B pd B gt (6)

[0070]

[0071] Among them, B pd It is a prediction box, B gt It is the ground truth bounding box; b pd and b gt These are the center coordinates of the predicted bounding box and the ground truth bounding box, respectively, d 2 (bpd b gt R(B) represents the Euclidean distance between the centers of the two bounding boxes; c is the diagonal length of the smallest bounding rectangle containing the predicted box and the ground truth bounding box; pd B gt ) is a penalty term added to the IoU loss.

[0072] Example 1

[0073] In this embodiment, digital silhouette angiography images of 1268 patients with carotid artery stenosis were collected to validate the effectiveness of a digital silhouette angiography-based method for detecting and quantifying carotid artery stenosis. All patients underwent DSA examination using single-plane X-ray systems from three different manufacturers: Philips Allura Xper FD 20 / 20 (UK), GE Healthcare Innova (USA), and Siemens AXIOMArtis (Germany). Angiographic images were captured at 4 frames per second, and two neuroradiologists reviewed the data using a Radiant DICOM viewer. To select appropriate angiographic sequences for analysis, the reviewers chose the sequence with the largest angle between the internal carotid arteries. If the lesion site was obstructed in that sequence, other sequences where the lesion site was not obstructed were selected. The neuroradiologists then selected images of contrast agent passing through the stenosis as keyframes for subsequent analysis. The degree of stenosis was measured according to the clinical trial criteria of the North American Symptomatic Carotid Endarterectomy Trial (NASCET), and the measured minimum luminal diameter (MLD) and reference vessel diameter (RVD) were used as markers for the regression model. The keyframes were then exported as JPGs for further analysis, with the maximum image resolution set to 960×960. This retrospective observational study was approved by the institutional review board.

[0074] Then, neuroradiologists labeled the ground truth bounding boxes for stenosis and normal blood vessels, resulting in 1205 ground truth bounding boxes for stenosis and 4031 for normal blood vessels. The object detection model was trained using 10x cross-validation with a batch size of 8. The Adam optimizer was used for parameter optimization, with an initial learning rate of 0.0001, which was then reduced to 0.00001 based on the validation set results. The entire training set was trained for 32 epochs. The dataset was divided into three parts: training, validation, and test sets, in a ratio of 8:1:1. The evaluation metric for the object detection model was mAP, and the evaluation metric for the regression model was mean absolute error.

[0075] Furthermore, the target detection model for carotid artery stenosis and its matching normal blood vessels in this embodiment is denoted as StenosisDet, and the detailed structure of the network is as follows: Figure 2 As shown in Table 1, to demonstrate the effectiveness of the detection strategy and model proposed in this embodiment in detecting carotid artery stenosis, some classic single-stage object detectors were selected for comparative experiments. These detectors have similar parameter counts to StenosisDet and employ the same training strategy.

[0076] Table 1 Comparison between StenosisDet and classic one-stage target detection models

[0077]

[0078] Table 1 shows the effectiveness of StenosisDet. StenosisDet significantly improves the detection performance of normal blood vessels. Since the "coarse-to-fine, fine-to-coarse" characteristics of stenosis are easily distinguishable, general target detection models have already achieved high AP values, so the improvement is not significant.

[0079] To illustrate the impact of introducing prior knowledge based on blood vessel trajectory on the performance improvement of the target detection model, the proposed improved IoU loss is compared with other target detection regression loss functions. As shown in Table 2, for the target detection regression loss, firstly, replacing the L1 loss of StenosisDet with DIoU loss does not significantly improve model performance. Then, the performance of the proposed improved IoU loss, denoted as 2DIoU loss, is evaluated. It improves the detection performance of normal blood vessel classes without reducing the accuracy of bounding box regression.

[0080] Table 2 compares the proposed improved IoU loss with other regression loss functions.

[0081]

[0082] To demonstrate the accuracy of the quantitative analysis and prediction of carotid artery stenosis, the prediction accuracy of morphological indices of stenosis was evaluated. Specifically, the stenotic portion and the matched normal vessel portion from the original image were input into the regression model to directly predict the minimum lumen diameter (MLD) and reference vessel diameter (RVD) of the stenotic portion. Two training strategies were employed for the regression model: 1) training the two classes of data separately; 2) merging the two classes of data into a single dataset for training. Furthermore, the performance of several variants of ResNet as the model backbone was compared, and the effectiveness of the SE module was evaluated. Experimental results are shown in Table 3.

[0083] Table 3. Performance of the regression of stenosis morphology indicators.

[0084]

[0085] In Table 3, “Average” represents the average MAE of RVD and MLD.

[0086] The method proposed in this invention achieves accurate estimation of the stenosis index, as shown in the last row of Table 3. The MAE values ​​for RVD, MLD, and SD are 0.378, 0.221, and 4.9, respectively, with an average MAE of 0.2998 mm for RVD and MLD. Experimental results show that using a deeper network as the backbone can slightly reduce the MAE, while the channel attention mechanism of the SE module can further improve performance. Meanwhile, quantitative regression of the stenosis class achieves a lower MAE compared to the normal vessel class. Furthermore, to avoid uncertainties introduced by slight variations in pixel density in the images of the dataset, the following formula is used to calculate the degree of stenosis and evaluate its MAE.

[0087]

[0088] In summary, the method in this embodiment first detects and locates stenosis and normal vessels used to predict reference vessel diameter, and then feeds them into a regression model to predict vessel diameter, thus successfully solving the problem of predicting morphological indicators of stenosis. Furthermore, by introducing prior knowledge based on vessel trajectory and proposing a corresponding improved regression loss function, the predictive performance of the model is enhanced.

[0089] The present invention also discloses an electronic device comprising one or more processors; a memory; and one or more application programs, wherein the one or more application programs are stored in the memory and configured to be executed by the one or more processors, and the one or more application programs are configured to perform the carotid artery stenosis detection and quantification method based on digital subtraction angiography as described above.

[0090] Those skilled in the art will recognize that the units and algorithm steps of the various 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 implement the described functions using different methods for each specific application, but such implementation should not exceed the scope of this invention.

[0091] The disclosed systems, modules, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of the units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0092] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0093] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, ROM, RAM, etc.

[0094] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0095] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of these embodiments are merely for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for carotid artery image detection and quantification based on digital subtraction angiography, characterized in that, Includes the following steps: S1, using the classification branch of the target detection model to detect stenotic and normal vessels in the carotid artery in digital subtraction angiography images; wherein, the detection head of the target detection model includes a classification branch, a match-ness branch, and a regression branch; S2, using the Match-ness branch in the detection head, predicts the matching degree between multiple normal blood vessels and a single stenotic blood vessel, and identifies the normal blood vessel with the highest matching degree to the stenotic blood vessel, specifically including: When training the object detection model, the matching score output by the match-ness branch is adjusted to between 0 and 1 using the sigmoid activation function, and the loss is calculated by comparing it with the ground truth value of the matching degree of the annotation. Using the Match-ness branch in the trained target detection model, a matching score is predicted for each normal class prediction box based on the currently predicted normal blood vessel. The normal blood vessel with the highest matching score is taken as the normal blood vessel with the highest degree of matching with the stenotic blood vessel. S3 involves cutting out the stenotic vessel and the normal vessel with the highest matching degree to the stenotic vessel, and inputting them into a regression model to predict the morphological indicators of carotid artery stenosis, thereby obtaining the quantitative analysis results of carotid artery stenosis.

2. The method for carotid artery image detection and quantification based on digital subtraction angiography according to claim 1, characterized in that, S1 utilizes the classification branch of the target detection model to detect stenotic and normal vessels in the carotid artery in digital subtraction angiography images, specifically including: Based on digital subtraction angiography images, the stenotic and normal vessels of the carotid artery are identified according to morphological features, and the ground truth bounding boxes of the target detection are labeled according to the stenosis class and the normal class. Each digital subtraction angiography image is labeled with a stenosis-class ground truth bounding box and multiple normal-class ground truth bounding boxes. When training the target detection model, the normal-class ground truth bounding boxes are additionally labeled with the ground truth value of the matching degree with the stenotic vessels in the current image. The matching degree of normal vessels suitable for predicting the diameter of the reference vessel for stenosis is labeled as 1, and the matching degree of other normal vessels is labeled as 0.

3. The method for carotid artery image detection and quantification based on digital subtraction angiography according to claim 1, characterized in that, S3 involves cutting out the stenotic vessel and the normal vessel with the highest matching degree to the stenotic vessel, and using a regression model to predict the morphological indicators of carotid artery stenosis to obtain quantitative analysis results of carotid artery stenosis, specifically including: From digital subtraction angiography images, the stenotic vessels predicted by the target detection model and the normal vessels that best match the stenotic vessels are extracted and fed into a regression model to predict the morphological indicators of carotid artery stenosis. The morphological indicators include the minimum lumen diameter of the stenotic vessels and the reference vessel diameter of the normal vessels that best match the stenotic vessels. The degree of stenosis of the narrowed vessel is obtained based on the minimum lumen diameter of the narrowed vessel and the reference vessel diameter, enabling quantitative analysis.

4. The method for carotid artery image detection and quantification based on digital subtraction angiography according to claim 3, characterized in that, S3 also includes: The image of digital subtraction angiography was denoised using a nonlocal mean algorithm, and then the image was filled to a 1:1 aspect ratio using a gray value similar to the background of the digital subtraction angiography image. A deep convolutional neural network is used to extract features from images, and finally a multi-layer fully connected layer is used to predict morphological indicators. The mean squared error loss function is used to train the object detection model.

5. The method for carotid artery image detection and quantification based on digital subtraction angiography according to claim 1, characterized in that, The S2 further includes: during the training of the target detection model, incorporating prior knowledge based on blood vessel trends into the calculation of the regression loss to improve the performance of the target detection model, specifically: The direction of blood vessel movement is obtained by connecting the center points of the narrow ground truth bounding box and the normal ground truth bounding box; During the network training of the object detection model, the second direction is obtained by connecting the center points of the normal class ground truth bounding boxes and the normal class predicted boxes; A penalty term is added to the D-IoU loss of the target detection model. The penalty term is maximized when the angle between the direction of blood vessel movement and the second direction is close to 0 degrees, and minimized when it is close to 90 degrees.

6. The method for carotid artery image detection and quantification based on digital subtraction angiography according to claim 5, characterized in that, The penalties include: The dot product of vectors is used to calculate the angle between the direction of blood vessel movement and the second direction, which is denoted as Angle; The penalty term is obtained by calculating the sine of the included angle. The formula for the penalty term P is as follows: (4)， Where α represents the hyperparameter for weighting the penalty term; The final bounding box regression loss function The formula is shown below: (5)， in, The D-IoU loss function is shown in the following formula: (6)， (7)， in, It is a prediction box. It is the ground truth bounding box; These are the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. The Euclidean distance between the centers of the two bounding boxes is represented; c is the diagonal length of the smallest bounding rectangle containing the prediction box and the ground truth bounding box. It is a penalty added to the IoU loss.

Images