Method and device for detecting anatomical key points in three-dimensional angiography images

Through a multi-task deep learning network combined with spatial prior information, explicitly modeling of vascular topological variant types, solving the complexity and variability problems of anatomical key point detection in three-dimensional angiography images, and achieving better detection performance.

CN114943682BActive Publication Date: 2024-11-26TSINGHUA UNIVERSITY +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210179800.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-02-25
Publication Date
2024-11-26
Estimated Expiration
2042-02-25

AI Technical Summary

Technical Problem

The detection task of anatomical key points in three-dimensional angiography images faces detection difficulties caused by complex vascular morphology, diverse topological variations and lack of them. Especially the physiological variation in the Willis ring area in intracranial blood vessels makes it more difficult to detect key points automatically.

Method used

Using a multi-task deep learning network combined with spatial prior information, the training data set is generated through preprocessing and annotation, so as to realize the synergistic effects of anatomical key points, semantic segment segment segment segment loss classification and local fork feature classification, and explicitly constructed Modular vascular topological variant type, for detection.

Benefits of technology

Improve the accuracy and robustness of anatomical key point detection, can effectively deal with the complexity and variability of vascular topology, and enhance detection performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114943682B_ABST
    Figure CN114943682B_ABST
Patent Text Reader

Abstract

The present application proposes a method for detecting anatomical key points in three-dimensional angiography images, which relates to the field of medical image processing technology, wherein the method comprises: obtaining a three-dimensional angiography image as a test image; preprocessing the test image, and inputting the preprocessed image into a pre-trained multi-task deep learning network, and outputting an anatomical key point prediction probability map, wherein the multi-task deep learning model is trained using a three-dimensional angiography training image containing the same vascular type as the test image and the annotation results of the three-dimensional angiography training image as a training data set; and generating the detection results of the anatomical key points according to the predicted probability of the voxel position in the anatomical key point prediction probability map. The present application adopting the above scheme can make full use of the synergy between different tasks, explicitly model the type of vascular topology variation, and combine spatial prior information to achieve good detection performance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of medical image processing, and in particular to a method and device for detecting anatomical key points in three-dimensional angiography images. Background Art

[0002] Three-dimensional angiography technology includes magnetic resonance angiography (MRA), computed tomography angiography (CTA), digital subtraction angiography (DSA), etc., which can use the imaging characteristics of blood flow to clearly and three-dimensionally display the characteristics of blood vessels and blood flow signals in the body. Three-dimensional angiography covers a variety of vascular structures such as intracranial blood vessels, coronary arteries, carotid arteries, aortas, etc. Inspection and analysis of these structures are important auxiliary means for diagnosing and treating related diseases. Taking intracranial vascular magnetic resonance angiography images as an example, it can reflect whether there are malformations in intracranial arteries and veins, and non-invasively, safely and clearly display the morphology of intracranial aneurysms and the morphology of the parent arteries. It has become the preferred method for diagnosing intracranial aneurysms. In recent years, medical image intelligent analysis technology based on computer theory has made great progress. Tasks such as automatic extraction of blood vessels, lesion location, and lesion measurement in three-dimensional angiography images have been widely studied and clinically applied.

[0003] The task of detecting anatomical key points in three-dimensional angiography images focuses on the bifurcation points of blood vessels at all levels. They are located at the bifurcation positions of the vascular segments at all levels and have unique and important anatomical significance. Taking the detection of intracranial vascular key points as an example, according to the definition of cerebral vascular topology, the complete Willis circle area can be divided into 20 vascular segments with independent anatomical names (excluding ICA-C4 and outer segments here), and a total of 19 key points can be defined at the junction of each vascular segment. Therefore, the anatomical key points explicitly model the overall topological structure of the blood vessels, which can provide rich semantic information for vascular semantic segmentation, lesion localization and disease diagnosis. In addition, anatomical key point detection is an important enabling means for subsequent tasks of intelligent analysis of medical images. It can provide initialization conditions for vascular tracking and centerline extraction, and assist in the realization of vascular tree registration of multiple images of the same patient or different patients. However, due to the slender and curved morphology of blood vessels, complex structural distribution, and diverse changes between different individuals, and the fact that some vascular segments need to rely on the position of surrounding tissues, the local appearance and grayscale distribution of the image may be affected by pathology, the anatomical key point detection task faces great challenges.

[0004] On the other hand, unlike other anatomical tree structures such as the trachea and aorta, some vascular structures such as intracranial blood vessels and coronary arteries may have topological changes. Taking the Willis circle region in intracranial blood vessels as an example, according to relevant studies, only about 52% of individuals have a complete Willis circle, and physiological variations represented by unilateral or bilateral PCoA deletion and PCA-P1 segment deletion (embryonic posterior communication) are widely present. Studies have pointed out that these physiological variations may be associated with potential disease risks, and how to model these types of variations is one of the key points of intracranial vascular analysis. It is worth noting that the loss of some vascular segments will cause the relevant key points to lose their local bifurcation characteristics and become difficult to identify. For example, when one PCoA exists, the key point PCoA-A (the bifurcation point between the vascular segment PCoA and ICA) is located at the bifurcation position; when the PCoA is missing, the key point will be located on the smooth ICA vascular segment and will not have local bifurcation characteristics. In clinical annotation, the positions of these key points often need to be determined based on the physician's experience and spatial symmetry, which makes automatic detection of key points particularly difficult. Summary of the invention

[0005] The present application aims to solve one of the technical problems in the related art at least to some extent.

[0006] To this end, the first purpose of this application is to propose a method for detecting anatomical key points in three-dimensional angiography images. This application adopting the above scheme can make full use of the synergy between different tasks, explicitly model the vascular topology variation types, and combine spatial prior information to achieve good detection performance.

[0007] The second objective of the present application is to provide a device for detecting anatomical key points in three-dimensional angiography images.

[0008] To achieve the above-mentioned purpose, the first aspect of the present application proposes a method for detecting anatomical key points in a three-dimensional angiography image, comprising: obtaining a three-dimensional angiography image as a test image; preprocessing the test image, and inputting the preprocessed image into a pre-trained multi-task deep learning network, and outputting an anatomical key point prediction probability map, wherein the multi-task deep learning model is trained using a three-dimensional angiography training image containing the same blood vessel type as the test image and the annotation results of the three-dimensional angiography training image as a training data set; generating a detection result of the anatomical key point based on the predicted probability of the voxel position in the anatomical key point prediction probability map.

[0009] The method for detecting anatomical key points in three-dimensional angiography images of the embodiment of the present application obtains three-dimensional angiography image data containing specific vascular structures and preprocesses them in the offline stage, manually annotates the images, and generates corresponding prediction targets for anatomical key point detection, vascular segment semantic segmentation, vascular segment missing classification, and key point local branch feature classification tasks, which together constitute a training data set to train a multi-task deep learning network; then in the online stage, the trained network model is used to output the key point probability heat map prediction result from an image of the same type, and the final anatomical key point detection position is obtained from it. The present application explicitly introduces structural prior knowledge and models spatial semantic information, which can achieve good detection performance.

[0010] Optionally, in one embodiment of the present application, the test image is preprocessed, including unifying the resolution, cropping to a preset size, and normalizing the voxel grayscale values.

[0011] Optionally, in one embodiment of the present application, pre-training a multi-task deep learning network includes:

[0012] Acquire a three-dimensional angiography image containing the same blood vessel type as the test image as an original data set;

[0013] Preprocessing the original data set and obtaining the annotation results corresponding to the preprocessed data set, wherein the annotation results include vascular anatomical key point annotation results, vascular binary segmentation annotation results, and vascular segment semantic segmentation annotation results;

[0014] Generate a training data set based on the preprocessed data set and the corresponding annotation results;

[0015] Construct a multi-task deep learning network, and use the training data set to train the multi-task deep learning network to obtain the trained multi-task deep learning network.

[0016] Optionally, in one embodiment of the present application, obtaining a labeling result corresponding to the preprocessed data set includes:

[0017] Using medical image processing software, each image in the preprocessed data set is manually annotated with predefined vascular anatomical key points and vascular binary segmentation. The vascular anatomical key point annotation results are the three-dimensional coordinates corresponding to each key point, and the vascular binary segmentation annotation results are voxel-by-voxel binary images of the same size as the image.

[0018] Based on the annotation results of vascular anatomical key points and vascular binary segmentation, an automated method is used to generate the semantic segmentation annotation results of the vascular segments corresponding to each image in the dataset.

[0019] Optionally, in one embodiment of the present application, based on the vascular anatomical key points and the vascular binary segmentation annotation results, an automated method is used to generate a vascular segment semantic segmentation annotation result corresponding to each image in the data set, including:

[0020] The corresponding lumen centerline is obtained from the binary segmentation and annotation results of the blood vessels using a thinning algorithm, and the centerline is divided into different semantic segments based on the anatomical key point annotations;

[0021] For each vascular voxel in the vascular binary segmentation annotation, its semantic label is determined based on the nearest centerline voxel;

[0022] The automatic semantic segmentation annotation results of each image are manually corrected in the medical image processing software to obtain the final semantic segmentation annotation results, wherein the automatic semantic segmentation annotation results include semantic segments and semantic labels.

[0023] Optionally, in one embodiment of the present application, generating a training data set according to the preprocessed data set and the corresponding annotation results includes:

[0024] According to the annotation results, each image in the preprocessed data set is processed to obtain the multi-channel probability heat map of anatomical key points, the multi-channel probability map of vascular segment semantic segmentation, the classification vector of vascular segment missing, and the classification vector of key point local bifurcation features as the prediction target corresponding to the image;

[0025] Each image in the preprocessed data set and the corresponding prediction target constitute a training data pair, and all training data pairs together constitute the training data set.

[0026] Optionally, in one embodiment of the present application, each image in the preprocessed data set is processed according to the annotation result to obtain a multi-channel probability heat map of anatomical key points, a multi-channel probability map of vascular segment semantic segmentation, a vascular segment missing classification vector, and a key point local bifurcation feature classification vector as a prediction target corresponding to the image, including:

[0027] For each preprocessed image in the dataset, a multi-channel probability heat map of anatomical key points with the same size as the input image is output for each predefined key point according to the annotation results of the vascular anatomical key points. For each target key point, the corresponding probability heat map presents a three-dimensional Gaussian distribution with the key point as the center.

[0028] Generate a multi-channel probability map of the semantic segmentation of the blood vessel segments according to the semantic segmentation annotation results of the blood vessel segments, wherein the last channel of the multi-channel probability map of the semantic segmentation of the blood vessel segments is a background channel, and the remaining channels respectively reflect the position distribution of each blood vessel segment in the input image;

[0029] According to the annotation results of vascular segment semantic segmentation, the vascular segment missing classification vector and the key point local bifurcation feature classification vector are obtained. When a vascular segment is missing in the vascular segment semantic segmentation annotation results, the anatomical key points at both ends of the vascular segment lose the local bifurcation feature, otherwise the anatomical key points at both ends of the vascular segment have the local bifurcation feature.

[0030] Optionally, in one embodiment of the present application, the multi-task deep learning network includes a trunk part and four branch parts, wherein:

[0031] The main part is used to extract features from the input image and output feature maps;

[0032] The first branch is used to process the feature map and generate the prediction results of the multi-channel probability heat map of the anatomical key points;

[0033] The second branch is used to process the feature map and generate the prediction result of the multi-channel probability map of the semantic segmentation of the blood vessel segment;

[0034] The third branch is used to process the feature map and generate the prediction result of the missing classification vector of the blood vessel segment;

[0035] The fourth branch is used to process the feature map and generate the prediction result of the key point local fork feature classification vector.

[0036] Optionally, in one embodiment of the present application, using a training data set to train the initialized network includes:

[0037] Step S1: randomly selecting a training data pair from the training data set, inputting the preprocessed three-dimensional angiography image in the training data pair into the constructed multi-task deep learning network, and obtaining the output results of each branch of the network as the prediction result;

[0038] Step S2: input the prediction result and the prediction target in the training data pair into the loss function to obtain the loss function value;

[0039] Step S3: Based on the calculated loss function value, the loss function is minimized using the gradient descent method and the network parameters are adjusted;

[0040] Step S4: Repeat steps S1, S2, S3, and S4, and continuously adjust the network parameters. When the number of training times exceeds the set upper limit, the training is completed, the multi-task deep learning network parameters are determined, and a trained multi-task deep learning network is obtained.

[0041] To achieve the above-mentioned purpose, the second embodiment of the present invention proposes a device for detecting anatomical key points in a three-dimensional angiography image, comprising an acquisition module, a processing module, and a result generation module, wherein:

[0042] An acquisition module, used for acquiring a three-dimensional angiography image as a test image;

[0043] A processing module is used to preprocess the test image and input the preprocessed image into a pre-trained multi-task deep learning network to output an anatomical key point prediction probability map, wherein the multi-task deep learning model is trained using a three-dimensional angiography training image containing the same blood vessel type as the test image and the annotation results of the three-dimensional angiography training image as a training data set;

[0044] The result generation module is used to generate the detection result of the anatomical key point according to the predicted probability of the voxel position in the anatomical key point prediction probability map.

[0045] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0047] Figure 1 A schematic diagram of a flow chart of a method for detecting anatomical key points in a three-dimensional angiography image provided in Example 1 of the present application;

[0048] Figure 2 This is an overall flow chart of the method of the embodiment of the present application;

[0049] Figure 3 This is a schematic diagram of intracranial blood vessel annotation and data generation results in an embodiment of the present application;

[0050] Figure 4 A schematic diagram of the corresponding relationship between the missing blood vessel segments of the intracranial blood vessels and the change of the local bifurcation characteristics of the key points in the embodiment of the present application;

[0051] Figure 5 This is a schematic diagram of a multi-task deep learning network in the offline phase of an embodiment of the present application;

[0052] Figure 6 A schematic diagram of a multi-task deep learning network in the online stage of an embodiment of the present application;

[0053] Figure 7 A detection result diagram of anatomical key points in the intracranial vascular magnetic resonance angiography image according to an embodiment of the present application;

[0054] Figure 8A schematic diagram of the structure of a device for detecting anatomical key points in a three-dimensional angiography image provided in an embodiment of the present application. DETAILED DESCRIPTION

[0055] Embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0056] If the attribute of whether each vascular segment is missing or not is introduced into the detection algorithm of vascular anatomical key points, it can help the algorithm to explicitly model the vascular variation type and achieve better detection performance. In addition, the anatomical key points are located at the two end points of the corresponding vascular segment and have very distinct structural characteristics. Considering that the position distribution of specific vascular segments often has strong regularity and consistency, introducing the auxiliary task of segmenting each vascular segment in the key point detection algorithm (i.e., realizing the semantic segmentation of blood vessels) can introduce structural prior information, which provides a guarantee for further improving the detection accuracy. Therefore, this application proposes an anatomical key point detection method that explicitly combines the vascular topological structure variation type.

[0057] The following describes the method and device for detecting anatomical key points in a three-dimensional angiography image according to an embodiment of the present application with reference to the accompanying drawings.

[0058] Figure 1 A schematic flow chart of a method for detecting anatomical key points in a three-dimensional angiography image provided in Example 1 of the present application.

[0059] like Figure 1 As shown, the method for detecting anatomical key points in the three-dimensional angiography image includes the following steps:

[0060] Step 101, obtaining a three-dimensional angiography image as a test image;

[0061] Step 102, preprocessing the test image, and inputting the preprocessed image into a pre-trained multi-task deep learning network, and outputting an anatomical key point prediction probability map, wherein the multi-task deep learning model is trained using a three-dimensional angiography training image containing the same blood vessel type as the test image and the annotation results of the three-dimensional angiography training image as a training data set;

[0062] Step 103: Generate detection results of anatomical key points according to the predicted probabilities of voxel positions in the anatomical key point prediction probability map.

[0063] The method for detecting anatomical key points in three-dimensional angiography images of the embodiment of the present application obtains three-dimensional angiography image data containing specific vascular structures and preprocesses them in the offline stage, manually annotates the images, and generates corresponding prediction targets for anatomical key point detection, vascular segment semantic segmentation, vascular segment missing classification, and key point local branch feature classification tasks, which together constitute a training data set to train a multi-task deep learning network; then in the online stage, the trained network model is used to output the key point probability heat map prediction result from an image of the same type, and the final anatomical key point detection position is obtained from it. The present application explicitly introduces structural prior knowledge and models spatial semantic information, which can achieve good detection performance.

[0064] The method for detecting anatomical key points in three-dimensional angiography images of the embodiment of the present application obtains three-dimensional angiography image data containing specific vascular structures and preprocesses them in the offline stage, manually annotates the images, and generates corresponding prediction targets for anatomical key point detection, vascular segment semantic segmentation, vascular segment missing classification, and key point local branch feature classification tasks, which together constitute a training data set to train a multi-task deep learning network; then in the online stage, the trained network model is used to output the key point probability heat map prediction result from an image of the same type, and the final anatomical key point detection position is obtained from it. The present application explicitly introduces structural prior knowledge and models spatial semantic information, which can achieve good detection performance.

[0065] Optionally, in one embodiment of the present application, the test image is preprocessed, including unifying the resolution, cropping to a preset size, and normalizing the voxel grayscale values.

[0066] Optionally, in one embodiment of the present application, pre-training a multi-task deep learning network includes:

[0067] Acquire a three-dimensional angiography image containing the same blood vessel type as the test image as an original data set;

[0068] Preprocessing the original data set and obtaining the annotation results corresponding to the preprocessed data set, wherein the annotation results include vascular anatomical key point annotation results, vascular binary segmentation annotation results, and vascular segment semantic segmentation annotation results;

[0069] Generate a training data set based on the preprocessed data set and the corresponding annotation results;

[0070] Construct a multi-task deep learning network, and use the training data set to train the multi-task deep learning network to obtain the trained multi-task deep learning network.

[0071] Optionally, in one embodiment of the present application, obtaining a labeling result corresponding to the preprocessed data set includes:

[0072] Using medical image processing software, each image in the preprocessed data set is manually annotated with predefined vascular anatomical key points and vascular binary segmentation. The vascular anatomical key point annotation results are the three-dimensional coordinates corresponding to each key point, and the vascular binary segmentation annotation results are voxel-by-voxel binary images of the same size as the image.

[0073] Based on the annotation results of vascular anatomical key points and vascular binary segmentation, an automated method is used to generate the semantic segmentation annotation results of the vascular segments corresponding to each image in the dataset.

[0074] Optionally, in one embodiment of the present application, based on the vascular anatomical key points and the vascular binary segmentation annotation results, an automated method is used to generate a vascular segment semantic segmentation annotation result corresponding to each image in the data set, including:

[0075] The corresponding lumen centerline is obtained from the binary segmentation and annotation results of the blood vessels using a thinning algorithm, and the centerline is divided into different semantic segments based on the anatomical key point annotations;

[0076] For each vascular voxel in the vascular binary segmentation annotation, its semantic label is determined based on the nearest centerline voxel;

[0077] The automatic semantic segmentation annotation results of each image are manually corrected in the medical image processing software to obtain the final semantic segmentation annotation results, wherein the automatic semantic segmentation annotation results include semantic segments and semantic labels.

[0078] Optionally, in one embodiment of the present application, generating a training data set according to the preprocessed data set and the corresponding annotation results includes:

[0079] According to the annotation results, each image in the preprocessed data set is processed to obtain the multi-channel probability heat map of anatomical key points, the multi-channel probability map of vascular segment semantic segmentation, the classification vector of vascular segment missing, and the classification vector of key point local bifurcation features as the prediction target corresponding to the image;

[0080] Each image in the preprocessed data set and the corresponding prediction target constitute a training data pair, and all training data pairs together constitute the training data set.

[0081] Optionally, in one embodiment of the present application, each image in the preprocessed data set is processed according to the annotation result to obtain a multi-channel probability heat map of anatomical key points, a multi-channel probability map of vascular segment semantic segmentation, a vascular segment missing classification vector, and a key point local bifurcation feature classification vector as a prediction target corresponding to the image, including:

[0082] For each preprocessed image in the dataset, a multi-channel probability heat map of anatomical key points with the same size as the input image is output for each predefined key point according to the annotation results of the vascular anatomical key points. For each target key point, the corresponding probability heat map presents a three-dimensional Gaussian distribution with the key point as the center.

[0083] Generate a multi-channel probability map of the semantic segmentation of the blood vessel segments according to the semantic segmentation annotation results of the blood vessel segments, wherein the last channel of the multi-channel probability map of the semantic segmentation of the blood vessel segments is a background channel, and the remaining channels respectively reflect the position distribution of each blood vessel segment in the input image;

[0084] According to the annotation results of vascular segment semantic segmentation, the vascular segment missing classification vector and the key point local bifurcation feature classification vector are obtained. When a vascular segment is missing in the vascular segment semantic segmentation annotation results, the anatomical key points at both ends of the vascular segment lose the local bifurcation feature, otherwise the anatomical key points at both ends of the vascular segment have the local bifurcation feature.

[0085] Optionally, in one embodiment of the present application, the multi-task deep learning network includes a trunk part and four branch parts, wherein:

[0086] The main part is used to extract features from the input image and output feature maps;

[0087] The first branch is used to process the feature map and generate the prediction results of the multi-channel probability heat map of the anatomical key points;

[0088] The second branch is used to process the feature map and generate the prediction result of the multi-channel probability map of the semantic segmentation of the blood vessel segment;

[0089] The third branch is used to process the feature map and generate the prediction result of the missing classification vector of the blood vessel segment;

[0090] The fourth branch is used to process the feature map and generate the prediction result of the key point local fork feature classification vector.

[0091] Optionally, in one embodiment of the present application, using a training data set to train the initialized network includes:

[0092] Step S1: randomly selecting a training data pair from the training data set, inputting the preprocessed three-dimensional angiography image in the training data pair into the constructed multi-task deep learning network, and obtaining the output results of each branch of the network as the prediction result;

[0093] Step S2: input the prediction result and the prediction target in the training data pair into the loss function to obtain the loss function value;

[0094] Step S3: Based on the calculated loss function value, the loss function is minimized using the gradient descent method and the network parameters are adjusted;

[0095] Step S4: Repeat steps S1, S2, S3, and S4, and continuously adjust the network parameters. When the number of training times exceeds the set upper limit, the training is completed, the multi-task deep learning network parameters are determined, and a trained multi-task deep learning network is obtained.

[0096] The present application is dedicated to adapting to the complex variation of vascular topological structure, and models whether a vascular segment is missing, and the change of local bifurcation characteristics of key points caused by the missing vascular segment (i.e., whether the key point has a local bifurcation feature) as additional attributes of each vascular segment and bifurcation point, and requires the algorithm to classify and predict the attribute. The method of the present invention is based on a deep learning network, with a multi-task model as the framework, and simultaneously completes four subtasks: anatomical key point detection, vascular segment semantic segmentation, vascular segment missing classification, and key point local bifurcation feature classification. Each subtask is highly related and shares the spatial semantic features extracted from the backbone of the network, making full use of the synergy between tasks, and explicitly modeling vascular variation types and structural prior information. The present application can be widely used in the detection tasks of various vascular anatomical key points, such as intracranial blood vessels, coronary arteries, etc., and can achieve good detection performance.

[0097] The method of the embodiment of the present invention is described in detail below with a specific embodiment.

[0098] The anatomical key point detection method proposed in this application is applied to the detection of some key points in the intracranial vascular magnetic resonance angiography images. The overall process is as follows: Figure 2 As shown, it includes an offline phase and an online phase.

[0099] (1) Offline stage

[0100] (1-1) Obtain the original data set and perform preprocessing;

[0101] A large number of three-dimensional angiography images containing the same blood vessel type are used as the original data set (intracranial vascular magnetic resonance angiography images are used in this embodiment). The images can come from public data sets or cooperative hospitals, and the number should be no less than 50. Preprocessing is performed on each image in the original data set, including unifying the resolution, cropping to the same size, and normalizing the voxel grayscale values. Among them, the present invention has no special requirements for the specific values ​​of the resolution and the size after cropping (the resolution is set to 0.5×0.5×0.8mm in this embodiment. 3 , the cropped size is 192×160×60); the cropped image should include the overall vascular structure to be detected (such as the Willis ring area required to be included in this embodiment), the cropping process can remove the interference of noise such as bones and other irrelevant tissues, and the size of the cropped area can be determined based on the average statistical distribution of the vascular structure.

[0102] The preprocessed magnetic resonance angiography image in this embodiment is as follows Figure 3 As shown in (A) in .

[0103] (1-2) Label the preprocessed data set;

[0104] Use medical image processing software (3D Slicer software is used in this embodiment) to manually annotate each image in the preprocessed data set, and it is necessary to annotate the predefined vascular anatomical key points and vascular binary segmentation. The annotated results of the anatomical key points are the three-dimensional coordinates corresponding to each key point, and the annotated results of the vascular binary segmentation are voxel-by-voxel binary images of the same size as the image (wherein the voxel value belonging to the vascular area is 1, and the voxel value of the remaining background area is 0).

[0105] Based on the anatomical key points and the binary segmentation annotations of the blood vessels, an automated method can be used to generate the semantic segmentation annotations of the blood vessel segments corresponding to each image in the data set. Specifically, a refinement algorithm is used to obtain the corresponding lumen centerline from the binary segmentation annotations of the blood vessels, and the centerline is divided into different semantic segments according to the anatomical key point annotations (for example, in this embodiment, the centerline portion between the key points PCoA-A and PCoA-P is the PCoA semantic segment). For each blood vessel voxel (voxel with a voxel value of 1) in the binary segmentation annotations of the blood vessels, its semantic label is determined according to the nearest centerline voxel. In particular, at the end of the peripheral blood vessel segment (such as the outer end of the MCA-M1 segment in this embodiment), the semantic segmentation annotation is cut so that the section is perpendicular to the centerline direction at that location. Thereafter, the semantic segmentation automated annotation results obtained for each image are manually corrected in the medical image processing software to obtain the final semantic segmentation annotation.

[0106] In this embodiment, the binary segmentation annotation of blood vessels in the magnetic resonance angiography image is as follows: Figure 3As shown in (B) in the figure, the anatomical key points corresponding to the image are marked as follows Figure 3 As shown in (C), the numbers in the figure are the serial numbers of the predefined 19 key points; the semantic segmentation annotation corresponding to the image is as follows Figure 3 As shown in (D), the areas with different gray levels in the figure represent different vascular segments (ie, different semantic labels), and the English abbreviation is the anatomical name of the vascular segment.

[0107] (1-3) Prepare training data set;

[0108] The preparation of the training data set is completed using the original data set preprocessed by step (1-1) and the manual annotation results obtained by step (1-2). That is, for each image in the data set, the prediction targets of the four subtasks of anatomical key point detection, vascular segmentation, vascular segment missing classification, and key point local branch feature classification in the multi-task network are obtained.

[0109] (1-3-1) Anatomical key point detection and prediction target generation;

[0110] In this application, the anatomical key point detection target is modeled as a multi-channel Gaussian heat map regression task. Specifically, for each preprocessed image in the data set, the network is required to output a probability heat map of the same size as the input image for each predefined key point. Among them, for each target key point, the corresponding probability heat map is a three-dimensional Gaussian distribution with the key point as the center, and the value of each voxel reflects the probability that the voxel belongs to the target key point. The probability value is determined by the Euclidean distance from the voxel position to the target key point, decreasing from 1 to 0 from the voxel position outward, and the decreasing rate is determined by the standard deviation δ of the Gaussian distribution. Specifically, for any preprocessed image in the data set, let the spatial coordinates of its ith key point be x i , then the probability value G of the corresponding probability heat map at any voxel space position x i (x) can be defined as:

[0111]

[0112] Where N is the total number of predefined anatomical key points in each image. In this embodiment, the heat map generated by each anatomical key point in the magnetic resonance angiography image is as follows: Figure 3 As shown in (E) in . For easy observation, the three-dimensional heat maps of all key points are projected into the same plane.

[0113] (1-3-2) Generate target for semantic segmentation prediction of blood vessel segments;

[0114] In this application, the task of semantic segmentation of vascular segments is modeled as a multi-channel single vascular segment binary segmentation task, that is, the prediction target is a multi-channel probability map generated by the semantic segmentation annotation of the vascular segments. For S predefined vascular segments (i.e., S semantic classes), the prediction target should contain S+1 channels, where the first S channels respectively reflect the position distribution of each vascular segment in the input image (i.e., whether each voxel in the input image belongs to each vascular segment, when a voxel belongs to the i-th vascular segment, the value of the i-th channel at the voxel position is 1, and the values ​​of the other channels are 0), and the S+1-th channel is the background channel (i.e., reflecting whether each voxel in the input image belongs to the background class, when a voxel does not belong to any vascular segment, the value of the background channel at the voxel position is 1, and the values ​​of the other channels are 0).

[0115] In this embodiment, the semantic segmentation prediction target of the blood vessel segment in the magnetic resonance angiography image is as follows: Figure 3 As shown in (F) in the figure, the channel corresponding to the MCA-M1 vascular segment in the predicted target is shown.

[0116] (1-3-3) Classification of missing blood vessel segments and classification of key point local bifurcation features to predict target generation;

[0117] Whether a blood vessel segment is missing corresponds to whether the key points at its two ends have local bifurcation features, which can be obtained from the manual annotation results of the blood vessel segment semantic segmentation. Specifically, when a blood vessel segment is missing in the blood vessel segment semantic segmentation annotation (that is, the number of voxels belonging to the blood vessel segment in the annotation is 0), the anatomical key points at both ends of the blood vessel segment lose the local bifurcation features; when a blood vessel segment exists (that is, the number of voxels belonging to the blood vessel segment in the annotation is greater than 0), the anatomical key points at both ends of the blood vessel segment have local bifurcation features. The above correspondence can be obtained by Figure 4 Intuitive instructions.

[0118] In this application, the classification of missing blood vessel segments and the classification of local branch features of key points are processed as multiple independent binary classification tasks. For the predefined N anatomical key points and S blood vessel segments, the prediction targets of the classification of local branch features of key points and the classification of missing blood vessel segments are vectors y with lengths of N and S respectively. N and S , the value of each element in the vector reflects whether each key point has a local bifurcation feature and whether each blood vessel segment exists (for any input image, when the i-th blood vessel segment is missing, the vector y S The value of the i-th element in is 0, otherwise it is 1; similarly, when the i-th anatomical key point does not have a local bifurcation feature, the vector y N The value of the i-th element in is 0, otherwise it is 1).

[0119] (1-3-4) Each preprocessed image and the corresponding manually annotated anatomical key point multi-channel probability heat map, vascular segment semantic segmentation multi-channel probability map, vascular segment missing classification vector, and key point local bifurcation feature vector constitute a training data pair. All training data pairs together constitute the training data set.

[0120] (1-4) Build a multi-task deep learning network;

[0121] The input of the multi-task deep learning network is a single preprocessed 3D angiography image. The input image is required to have a uniform size and resolution, but there is no restriction on the specific values ​​(the input image size used in this embodiment is 192 × 160 × 60 and the resolution is 0.5 × 0.5 × 0.8 mm). 3 ). The network consists of a main part and four branches. Figure 4 As shown in the figure. The backbone is improved from the classic U-Net model for medical image processing, which includes a symmetrical encoder and decoder structure. The encoder contains 5 residual modules and 4 maximum pooling layers, which are distributed between two residual modules in sequence. The residual module does not change the size of the input feature map. It contains two convolutional layers with a convolution kernel size of 3×3×3 and a short connection structure between the input and output of a module to avoid the gradient disappearance problem that may occur during the training of the deep learning network. Each maximum pooling layer will reduce the size of each dimension of the feature map to 1 / 2 of the original. The decoder contains 3 residual modules and 4 deconvolution layers, which are distributed between two deconvolution layers in sequence. The decoder has the same residual module structure as the encoder, and each deconvolution layer will expand the size of each dimension of the feature map to twice the original size. The output of the fifth residual module in the encoder is connected to the input of the first deconvolution layer in the decoder, and the number of maximum pooling layers in the encoder is kept consistent with the number of deconvolution layers in the decoder to ensure that the input and output of the backbone have the same size.

[0122] In addition, in order to fuse low-dimensional local spatial features and high-dimensional global semantic information, a jump structure is added between the symmetrical layers in the encoder and decoder structures. Specifically, the output feature map of the 4th residual module in the encoder and the output feature map of the 1st deconvolution layer in the decoder are concatenated as the input of the 1st residual module in the decoder; the output feature map of the 3rd residual module in the encoder and the output feature map of the 2nd deconvolution layer in the decoder are concatenated as the input of the 2nd residual module in the decoder; the output feature map of the 2nd residual module in the encoder and the output feature map of the 3rd deconvolution layer in the decoder are concatenated as the input of the 3rd residual module in the decoder; the output feature map of the 1st residual module in the encoder and the output feature map of the 4th deconvolution layer in the decoder are concatenated as the output feature map of the backbone. Afterwards, the output feature map of the backbone is simultaneously passed to the four branches of the network.

[0123] The four branches of the network are composed of a residual module and a convolution layer with a convolution kernel size of 1×1×1. The outputs of the four branches correspond to the prediction results of the multi-channel probability heat map of anatomical key points, the multi-channel probability map of semantic segmentation of vascular segments, the classification vector of missing vascular segments, and the local bifurcation feature vector of key points. The size and resolution of the prediction results of the first two branches should be consistent with the input image, and the length of the prediction result vector of the last two branches should be consistent with the predefined number of vascular segments and key points.

[0124] The multi-task deep learning network constructed in this embodiment is as follows Figure 5 As shown, taking the intracranial vascular magnetic resonance angiography image as an example, it is noted that the numerical values ​​in the figure are only for example and other numerical values ​​may be used in practice.

[0125] (1-5) Apply the training data set generated in step (1-3) to offline train the multi-task deep learning network constructed in step (1-4). The offline training includes the following steps:

[0126] (1-5-1) Randomly initialize the parameters of the multi-task deep learning network constructed in step (1-4).

[0127] (1-5-2) Randomly select a training data pair from the training data set generated in step (1-3), input the preprocessed three-dimensional angiography image into the multi-task deep learning network constructed in step (1-4), and obtain the output results of each branch of the network, which is the prediction result of each subtask. The prediction result of each subtask and the prediction target of each subtask in the training data pair are respectively input into the corresponding loss function to obtain the corresponding loss function value. Specifically, the present invention uses the L2 loss function in the anatomical key point detection task, the Dice loss function in the vascular segment semantic segmentation task, and the cross entropy loss function in the vascular segment missing classification and key point local branch feature classification tasks. In order to avoid the network from being difficult to converge due to serious class imbalance during training, the loss functions of the anatomical key point detection and vascular segment semantic segmentation tasks are weighted, and the weights are the ratios of the number of voxels in the input image to the number of voxels in the Gaussian hot spot and the area where each vascular segment is located.

[0128] In addition, considering the one-to-one correspondence between whether the vascular segment is missing and whether there is a local bifurcation feature at the key point, in order to ensure that the prediction results of the two classification tasks are consistent with the observation law, the consistency loss function L is introduced self Specifically, the prediction results of the missing blood vessel segment classification Derive the corresponding key point local bifurcation feature category It is required to have a local fork feature classification prediction result with the key point actual output of the network Keep consistent. The consistency loss function can be defined using the cross entropy loss function:

[0129]

[0130] Wherein, the superscript i represents the i-th element in the vector (i.e., corresponding to the i-th anatomical key point), and Θ is a set of key point serial numbers of all key points that may cause changes in local bifurcation characteristics due to vascular variations (e.g., for the intracranial vascular magnetic resonance angiography image used in this embodiment, among the predefined anatomical key points of the Willis circle, common key points that may cause changes in local bifurcation characteristics due to physiological variations include the PCoA on both sides, ACoA, PCA-P1, and the endpoints of both sides of ACA-A1, etc.).

[0131] The total loss function of network training is obtained by linear combination of the above loss functions:

[0132] L=L1+αL2+β(L3+L4)+γL self (0<α,β<1)

[0133] Among them, L1, L2, L3, and L4 are the loss functions of anatomical key point detection, vascular segment semantic segmentation, vascular segment missing classification, and key point local fork feature classification tasks, respectively. self is the consistency loss function. The hyperparameters α, β, and γ can be flexibly adjusted in actual scenarios to make each loss function at the same level.

[0134] (1-5-3) The above training steps are executed repeatedly. In each training, the gradient descent method is used to minimize the loss function based on the calculated total loss function value, and the network parameters are continuously adjusted. When the number of training times exceeds the set upper limit (the upper limit is generally not less than 5000 times), the training is completed and the multi-task deep learning network parameters are obtained.

[0135] (2) Online stage;

[0136] (2-1) Obtain a three-dimensional angiography image containing the same blood vessel type as the original data set in step (1-1) as a test image.

[0137] (2-2) Perform preprocessing on the test image obtained in step (2-1). The parameters such as image resolution and cropped image size in the preprocessing operation should be consistent with the preprocessing step in step (1-1).

[0138] (2-3) The preprocessed 3D angiography image obtained in step (2-2) is input into the multi-task deep learning network trained in the offline stage to obtain the output anatomical key point prediction probability map. The voxel position with the largest prediction probability in each prediction probability map is selected as the final detection result of the key point corresponding to the heat map. The multi-task deep learning network used in this step is as follows: Figure 6 Note that the values ​​in the figure are only for example, and other values ​​may be used in practice.

[0139] Using the anatomical key point detection method proposed in the present invention, the detection results of some key points of the intracranial vascular magnetic resonance angiography image in this embodiment are as follows: Figure 7 shown.

[0140] In order to implement the above embodiment, the present application also proposes a device for detecting anatomical key points in a three-dimensional angiography image.

[0141] Figure 8 A schematic diagram of the structure of a device for detecting anatomical key points in a three-dimensional angiography image provided in an embodiment of the present application.

[0142] like Figure 8 As shown, the device for detecting anatomical key points in a three-dimensional angiography image includes an acquisition module, a processing module, and a result generation module, wherein:

[0143] An acquisition module, used for acquiring a three-dimensional angiography image as a test image;

[0144] A processing module is used to preprocess the test image and input the preprocessed image into a pre-trained multi-task deep learning network to output an anatomical key point prediction probability map, wherein the multi-task deep learning model is trained using a three-dimensional angiography training image containing the same blood vessel type as the test image and the annotation results of the three-dimensional angiography training image as a training data set;

[0145] The result generation module is used to generate the detection result of the anatomical key point according to the predicted probability of the voxel position in the anatomical key point prediction probability map.

[0146] It should be noted that the above explanation of the embodiment of the method for detecting anatomical key points in three-dimensional angiography images is also applicable to the device for detecting anatomical key points in three-dimensional angiography images of this embodiment, and will not be repeated here.

[0147] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0148] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0149] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0150] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute the instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0151] It should be understood that the various parts of the present application can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0152] A person skilled in the art may understand that all or part of the steps in the above-mentioned embodiment method may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps in the method embodiment.

[0153] In addition, each functional unit in each embodiment of the present application may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0154] The storage medium mentioned above may be a read-only memory, a disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present application. A person of ordinary skill in the art may change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A method for detecting anatomical key points in a three-dimensional angiography image, characterized in that: The following steps are involved: acquiring a three-dimensional angiography image as a test image; Preprocessing the test image, and inputting the preprocessed image into a pre-trained multi-task deep learning network, outputting an anatomical key point prediction probability map, a vascular segment semantic segmentation multi-channel probability map, a vascular segment missing classification vector, and a key point local bifurcation feature classification vector, wherein the multi-task deep learning network is trained using a three-dimensional angiography training image containing the same vascular type as the test image and the annotation results of the three-dimensional angiography training image as a training data set; Generating a detection result of an anatomical key point according to the predicted probability of the voxel position in the anatomical key point prediction probability map; Among them, whether a blood vessel segment is missing corresponds to whether the key points at both ends of the segment have local bifurcation features. When a blood vessel segment is missing in the semantic segmentation annotation of the blood vessel segment, the anatomical key points at both ends of the blood vessel segment lose the local bifurcation features; when a blood vessel segment exists, the anatomical key points at both ends of the blood vessel segment have local bifurcation features. During training, the generation process of prediction targets for vessel segment missing classification and key point local bifurcation feature classification includes: The classification of missing blood vessel segments and the classification of local branch features of key points are processed into multiple independent binary classification tasks. For the predefined N anatomical key points and S blood vessel segments, the prediction targets of the classification of local branch features of key points and the classification of missing blood vessel segments are vectors y with lengths of N and S, respectively. N and S ,The value of each element in the vector reflects whether each key point has local bifurcation features and whether each vessel segment exists; The loss function of the multi-task deep learning network is expressed as: L=L1+αL2+β(L3+L4)+γL self (0<α,β<1) Among them, L1, L2, L3, and L4 are the loss functions of anatomical key point detection, vascular segment semantic segmentation, vascular segment missing classification, and key point local fork feature classification tasks, respectively. L3 and L4 are cross entropy loss functions. α, β, and γ are hyperparameters. self is the consistency loss function, Where, the superscript i represents the i-th element in the vector, and θ is the set of key point numbers that may cause local bifurcation feature changes due to vascular variation. Represents the prediction results classified by missing blood vessel segments The corresponding key point local bifurcation feature category is derived, The prediction results of the local fork feature classification for the key points.

2. The method according to claim 1, characterized in that The preprocessing of the test image includes unifying the resolution, cropping to a preset size, and normalizing the voxel grayscale values.

3. The method according to claim 1, characterized in that Pre-training the multi-task deep learning network includes: Acquire a three-dimensional angiography image containing the same blood vessel type as the test image as an original data set; Performing the preprocessing on the original data set, and obtaining the annotation results corresponding to the preprocessed data set, wherein the annotation results include vascular anatomical key point annotation results, vascular binary segmentation annotation results, and vascular segment semantic segmentation annotation results; Generate a training data set based on the preprocessed data set and the corresponding annotation results; Construct the multi-task deep learning network, and use the training data set to train the multi-task deep learning network to obtain a trained multi-task deep learning network.

4. The method according to claim 3, characterized in that The obtaining of the annotation result corresponding to the preprocessed data set includes: Using medical image processing software, each image in the preprocessed data set is manually annotated with predefined vascular anatomical key points and vascular binary segmentation. The vascular anatomical key point annotation results are the three-dimensional coordinates corresponding to each key point, and the vascular binary segmentation annotation results are voxel-by-voxel binary images of the same size as the image. Based on the annotation results of the vascular anatomical key points and the vascular binary segmentation, an automated method is used to generate the semantic segmentation annotation results of the vascular segments corresponding to each image in the dataset.

5. The method according to claim 4, characterized in that The method of using an automated method to generate a semantic segmentation annotation result of a blood vessel segment corresponding to each image in the data set based on the blood vessel anatomical key points and the blood vessel binary segmentation annotation result includes: Using a thinning algorithm to obtain the corresponding lumen centerline from the blood vessel binary segmentation and annotation results, and dividing the centerline into different semantic segments according to the anatomical key point annotations; For each vascular voxel in the vascular binary segmentation annotation, its semantic label is determined based on the nearest centerline voxel; Manually correct the automatic semantic segmentation annotation results obtained for each image in the medical image processing software to obtain a final semantic segmentation annotation result, wherein the automatic semantic segmentation annotation result includes the semantic segment and the semantic label.

6. The method according to claim 3, characterized in that The generating of the training data set according to the preprocessed data set and the corresponding annotation results includes: According to the annotation results, each image in the preprocessed data set is processed to obtain the multi-channel probability heat map of anatomical key points, the multi-channel probability map of vascular segment semantic segmentation, the classification vector of vascular segment missing, and the classification vector of key point local bifurcation features as the prediction target corresponding to the image; Each image in the preprocessed data set and the corresponding prediction target constitute a training data pair, and all training data pairs together constitute the training data set.

7. The method according to claim 6, characterized in that According to the labeling results, each image in the preprocessed data set is processed to obtain a multi-channel probability heat map of anatomical key points, a multi-channel probability map of vascular segment semantic segmentation, a vascular segment missing classification vector, and a key point local bifurcation feature classification vector as the prediction target corresponding to the image, including: For each preprocessed image in the data set, a multi-channel probability heat map of anatomical key points with the same size as the input image is output for each predefined key point according to the annotation results of the vascular anatomical key points, wherein for each target key point, the corresponding probability heat map presents a three-dimensional Gaussian distribution with the key point as the center; generating a multi-channel probability map of the semantic segmentation of the blood vessel segments according to the semantic segmentation and annotation results of the blood vessel segments, wherein the last channel of the multi-channel probability map of the semantic segmentation of the blood vessel segments is a background channel, and the remaining channels respectively reflect the position distribution of each blood vessel segment in the input image; According to the annotation results of vascular segment semantic segmentation, the vascular segment missing classification vector and the key point local bifurcation feature classification vector are obtained. When a vascular segment is missing in the vascular segment semantic segmentation annotation results, the anatomical key points at both ends of the vascular segment lose the local bifurcation feature, otherwise the anatomical key points at both ends of the vascular segment have the local bifurcation feature.

8. The method according to claim 3, characterized in that The multi-task deep learning network includes a trunk part and four branch parts, wherein: The backbone part is used to extract features from the input image and output a feature map; The first branch is used to process the feature map to generate a prediction result of a multi-channel probability heat map of anatomical key points; The second branch is used to process the feature map to generate a prediction result of a multi-channel probability map of the semantic segmentation of the blood vessel segment; The third branch is used to process the feature map to generate a prediction result of a blood vessel segment missing classification vector; The fourth branch is used to process the feature map to generate a prediction result of a key point local fork feature classification vector.

9. The method according to claim 6, characterized in that The using the training data set to train the initialized network includes: Step S1: randomly selecting a training data pair from the training data set, inputting the preprocessed three-dimensional angiography image in the training data pair into the constructed multi-task deep learning network, and obtaining the output results of each branch of the network as the prediction result; Step S2: inputting the prediction result and the prediction target in the training data pair into a loss function to obtain a loss function value; Step S3: Based on the calculated loss function value, the loss function is minimized using the gradient descent method and the network parameters are adjusted; Step S4: Repeat steps S1, S2, S3, and S4, and continuously adjust the network parameters. When the number of training times exceeds the set upper limit, the training is completed, the multi-task deep learning network parameters are determined, and a trained multi-task deep learning network is obtained.

10. A device for detecting anatomical key points in a three-dimensional angiography image, characterized in that: It includes an acquisition module, a processing module, and a result generation module, among which: The acquisition module is used to acquire a three-dimensional angiography image as a test image; The processing module is used to preprocess the test image, and input the preprocessed image into a pre-trained multi-task deep learning network, and output an anatomical key point prediction probability map, a vascular segment semantic segmentation multi-channel probability map, a vascular segment missing classification vector, and a key point local bifurcation feature classification vector, wherein the multi-task deep learning network is trained using a three-dimensional angiography training image containing the same vascular type as the test image and the annotation results of the three-dimensional angiography training image as a training data set; The result generating module is used to generate the detection result of the anatomical key point according to the prediction probability of the voxel position in the anatomical key point prediction probability map; Among them, whether a blood vessel segment is missing corresponds to whether the key points at both ends of the segment have local bifurcation features. When a blood vessel segment is missing in the semantic segmentation annotation of the blood vessel segment, the anatomical key points at both ends of the blood vessel segment lose the local bifurcation features; when a blood vessel segment exists, the anatomical key points at both ends of the blood vessel segment have local bifurcation features. During training, the generation process of prediction targets for vessel segment missing classification and key point local bifurcation feature classification includes: The classification of missing blood vessel segments and the classification of local branch features of key points are processed into multiple independent binary classification tasks. For the predefined N anatomical key points and S blood vessel segments, the prediction targets of the classification of local branch features of key points and the classification of missing blood vessel segments are vectors y with lengths of N and S, respectively. N and S ,The value of each element in the vector reflects whether each key point has local bifurcation features and whether each vessel segment exists; The loss function of the multi-task deep learning network is expressed as: L=L1+αL2+β(L3+L4)+γL self (0<α,β<1) Among them, L1, L2, L3, and L4 are the loss functions of anatomical key point detection, vascular segment semantic segmentation, vascular segment missing classification, and key point local fork feature classification tasks, respectively. L3 and L4 are cross entropy loss functions. α, β, and γ are hyperparameters. self is the consistency loss function, Where, the superscript i represents the i-th element in the vector, and θ is the set of key point numbers that may cause local bifurcation feature changes due to vascular variation. Represents the prediction results classified by vessel segment missing The corresponding key point local bifurcation feature category is derived, The prediction results of the local fork feature classification for the key points.