A method and device for segmenting an abnormal region of a blood vessel, and an electronic device

Abstract

The application discloses a method and device for segmenting an abnormal area of a blood vessel and an electronic device, wherein the method comprises: acquiring a blood vessel image to be segmented, the blood vessel image comprising a blood vessel area and an abnormal area attached to the blood vessel; acquiring a point cloud domain image comprising the blood vessel area and the abnormal area based on the blood vessel image; inputting the point cloud domain image into a pre-trained target segmentation model to acquire a segmentation image of the abnormal area, wherein the target segmentation model comprises a first segmentation module for segmenting the abnormal area and a second segmentation module for blocking a region adjacent to a boundary between the blood vessel area and the abnormal area. The application segments the point cloud domain image by using the first segmentation module to obtain a segmentation image of an initial abnormal area and determines a boundary between the blood vessel area and the initial abnormal area, which facilitates further blocking of point clouds near the boundary by using the second segmentation module, so that a segmentation image of a final abnormal area is accurately obtained.

Description

Technical Field

[0001] This invention relates to the field of image segmentation technology, and in particular to a method, apparatus, and electronic device for segmenting abnormal regions of blood vessels. Background Technology

[0002] Intracranial aneurysm (IA) is a common disease that threatens people's health, accounting for approximately 3.2%-6% of the general population and approximately 80%-85% of patients with spontaneous subarachnoid hemorrhage (SAH). Rupture of an aneurysm leads to a significantly high mortality rate, and even survivors may suffer long-term physical and psychological stress, resulting in a severe decline in their quality of life.

[0003] With advancements in science and technology and medical advancements, advanced imaging techniques have been widely applied, leading to the discovery of an increasing number of aneurysms. Existing aneurysm segmentation methods typically employ either image-domain segmentation or point-cloud-domain detection. However, image-domain-based segmentation methods not only consume significant computational resources but also fail to fully utilize the anatomical characteristics of the aneurysm, resulting in low accuracy. Point-cloud-based segmentation methods also suffer from insufficient segmentation accuracy.

[0004] Therefore, there is an urgent need for a method to segment abnormal regions of blood vessels in order to solve the problem that the segmentation of abnormal regions of blood vessels in existing technologies is not accurate enough. Summary of the Invention

[0005] In view of this, the present invention provides a method, apparatus and electronic device for segmenting abnormal regions of blood vessels, the main purpose of which is to solve the problem of insufficient accuracy in the current segmentation of abnormal regions of blood vessels.

[0006] To address the above problems, this application provides a method for segmenting abnormal regions of blood vessels, comprising:

[0007] Obtain a blood vessel image to be segmented, the blood vessel image including the blood vessel region and abnormal regions attached to the blood vessel;

[0008] Based on the blood vessel image, a point cloud domain image containing the blood vessel region and the abnormal region is obtained;

[0009] The point cloud image is input into a pre-trained target segmentation model to obtain a segmented image of the abnormal region. The target segmentation model includes a first segmentation module for segmenting the abnormal region and a second segmentation module for blocking the boundary between the adjacent blood vessel region and the abnormal region.

[0010] Optionally, obtaining a point cloud domain image containing the blood vessel region and the abnormal region based on the blood vessel image includes:

[0011] The blood vessels and abnormal regions in the blood vessel image are meshed, and at least one pixel is extracted from the mesh to obtain a set of pixels.

[0012] Convert the pixel set into a point cloud domain to obtain a point cloud domain image; or,

[0013] Extract the boundary point pixels of the blood vessel region and abnormal region within the blood vessel image to obtain a pixel set;

[0014] The set of pixels is converted into a point cloud domain to obtain a point cloud domain image.

[0015] Optionally, inputting the point cloud image into a pre-trained target segmentation model to obtain a segmented image of the abnormal region includes:

[0016] Based on the first segmentation module, an initial segmentation image of the abnormal region and the boundary between the abnormal region and the blood vessel region are obtained;

[0017] Based on the second segmentation module, the point cloud in a specific area on both sides of the boundary between the abnormal region and the blood vessel region in the initial segmentation image is blocked to obtain the segmentation image of the abnormal region. The blocking process includes classifying the point cloud located on one side of the abnormal region as the abnormal region and classifying the point cloud located on the one side of the blood vessel region as the blood vessel region.

[0018] Optionally, before inputting the point cloud image into a pre-trained target segmentation model to obtain a segmented image of the abnormal region, the method further includes:

[0019] The point cloud image is segmented into several blood vessel segment images;

[0020] The step of inputting the point cloud image into a pre-trained target segmentation model to obtain a segmented image of the abnormal region includes:

[0021] The images of the several blood vessel segments are input into a pre-trained target segmentation model to obtain several sub-segmentation images of the abnormal regions corresponding to the several blood vessel segments.

[0022] The sub-segmented images are recombined to obtain the segmented image of the abnormal region.

[0023] Optionally, the method for segmenting the abnormal region of the blood vessel further includes: obtaining a pre-trained target segmentation model, wherein obtaining the pre-trained target segmentation model includes:

[0024] Acquire a sample point cloud image containing abnormal regions and blood vessels, and a labeled point cloud image with the abnormal regions annotated.

[0025] An initial first segmentation module is trained based on the sample point cloud domain image and the label point cloud domain image, and the first segmentation module is obtained when the first preset termination condition is met.

[0026] Based on the labeled point cloud image, the boundary between the abnormal region and the blood vessel is obtained;

[0027] Based on the boundary and label point cloud images and the output image of the first segmentation module, an initial second segmentation module is trained. When a second preset termination condition is reached, the second segmentation module is acquired. The region of interest of the second segmentation module includes at least the region within a specific range on both sides of the boundary. The first preset termination condition includes the value of the loss function corresponding to the first segmentation module reaching a preset threshold or the number of training iterations of the first segmentation module reaching a preset number. The second preset termination condition includes the value of the loss function corresponding to the second segmentation module reaching a preset threshold or the number of training iterations of the second segmentation module reaching a preset number.

[0028] Optionally, the segmentation method for abnormal regions of blood vessels may further include: determining the region of interest of the second segmentation module through a nearest neighbor algorithm or a fixed threshold method.

[0029] Optionally, acquiring the labeled point cloud image of the abnormal region includes:

[0030] Obtain the boundary between blood vessels and abnormal regions in the sample point cloud image, as well as the seed points within the abnormal regions;

[0031] Based on the anatomical morphological features, boundaries, and seed points of the abnormal regions, the abnormal regions in the sample point cloud image are labeled to obtain a labeled point cloud image.

[0032] Optionally, after obtaining the segmented image of the abnormal region, the method further includes: displaying the abnormal region and the blood vessel region in the point cloud domain image using different display methods.

[0033] Optionally, the target segmentation model further includes a post-processing module for restoring the output of the second segmentation module to the image domain.

[0034] To address the aforementioned problems, this application provides a segmentation device for abnormal regions of blood vessels, comprising:

[0035] The first acquisition module is used to acquire a blood vessel image to be segmented, the blood vessel image including a blood vessel region and an abnormal region attached to the blood vessel.

[0036] The second acquisition module is used to acquire a point cloud domain image containing the blood vessel region and the abnormal region based on the blood vessel image;

[0037] The segmentation module is used to input the point cloud image into a pre-trained target segmentation model to obtain a segmented image of the abnormal region. The target segmentation model includes a first segmentation module for segmenting the abnormal region and a second segmentation module for blocking the boundary between the adjacent blood vessel region and the abnormal region.

[0038] To address the aforementioned problems, this application provides an electronic device, comprising at least a memory and a processor, wherein the memory stores a computer program, and the processor, when executing the computer program in the memory, implements the steps of the method for segmenting abnormal regions of blood vessels as described above.

[0039] The method, apparatus, and electronic device for segmenting abnormal regions of blood vessels in this application acquire a point cloud image of the blood vessel image to be segmented. They can then use the first segmentation module in the target segmentation model to detect and identify each point cloud in the point cloud image, determining whether each point cloud is a blood vessel point or an abnormal point, thus obtaining an initial segmented image of the abnormal region. The boundary between the blood vessel region and the initial abnormal region is also determined, facilitating further optimization / blocking of the point cloud near the boundary using the second segmentation module. This allows for accurate identification of abnormal points and ultimately, accurate acquisition of the final segmented image of the abnormal region.

[0040] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0041] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0042] Figure 1 This is a flowchart illustrating a method for segmenting abnormal regions of blood vessels according to an embodiment of this application;

[0043] Figure 2 For CTA images;

[0044] Figure 3 The point cloud domain image obtained by conversion;

[0045] Figure 4 This is a magnified view of a local point cloud region image;

[0046] Figure 5 (a) is the sample point cloud image before annotation;

[0047] Figure 5 (b) is the labeled sample point cloud image;

[0048] Figure 6 This is a schematic diagram illustrating the positional relationship between the boundary region and the region of interest in an embodiment of this application;

[0049] Figure 7 This is a schematic diagram of the images of each blood vessel segment after image segmentation processing of the point cloud domain image in the embodiments of this application;

[0050] Figure 8 (a) is a CAT image of an intracranial aneurysm;

[0051] Figure 8 (b) For Figure 8 (a) Aneurysm segmentation map obtained using traditional manual labeling methods;

[0052] Figure 8 (c) For Figure 8 (a) Aneurysm segmentation map obtained using the segmentation method for abnormal regions of blood vessels in this application;

[0053] Figure 8 (d) is related to Figure 8 (b) Corresponding 3D reconstruction image;

[0054] Figure 8 (e) for Figure 8 (c) Corresponding 3D reconstruction image;

[0055] Figure 9 A structural block diagram of a device for segmenting abnormal regions of blood vessels according to another embodiment of this application. Detailed Implementation

[0056] Various embodiments and features of this application are described herein with reference to the accompanying drawings.

[0057] It should be understood that various modifications can be made to the embodiments described herein. Therefore, the above description should not be considered as limiting, but merely as an example of embodiments. Other modifications within the scope and spirit of this application will be apparent to those skilled in the art.

[0058] The accompanying drawings, which are included in and form part of this specification, illustrate embodiments of the present application and, together with the general description of the present application given above and the detailed description of the embodiments given below, serve to explain the principles of the present application.

[0059] These and other features of this application will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.

[0060] It should also be understood that although this application has been described with reference to some specific examples, those skilled in the art can certainly implement many other equivalent forms of this application.

[0061] The above and other aspects, features and advantages of this application will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.

[0062] Specific embodiments of this application are described thereafter with reference to the accompanying drawings; however, it should be understood that the claimed embodiments are merely examples of this application, which can be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that could obscure the application. Therefore, the specific structural and functional details claimed herein are not intended to be limiting, but merely serve as the basis and representative basis for the claims to teach those skilled in the art to use this application in a variety of substantially any suitable detailed structures.

[0063] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in other embodiments,” all of which may refer to one or more of the same or different embodiments according to this application.

[0064] This application provides a method for segmenting abnormal regions of blood vessels, such as... Figure 1 As shown, it includes the following steps:

[0065] Step S101: Obtain the blood vessel image to be segmented, wherein the blood vessel image includes the blood vessel region and the abnormal region attached to the blood vessel.

[0066] In this step, the abnormal region refers to a lesion area outside the vascular region, such as an aneurysm. The image to be segmented in this step can be any non-point cloud image type; that is, the image to be segmented can be any modality image, such as... Figure 2 The CT angiography (CTA) image shown can also be an MRA image, a fusion of PET and CT images. The following description uses an aneurysm as an example to illustrate the embodiments of this application.

[0067] Step S102: Based on the blood vessel image, obtain a point cloud domain image containing the blood vessel region and the abnormal region;

[0068] This step involves transforming the obtained blood vessel image to obtain a point cloud image, such as... Figure 3 The point cloud domain image shown, point cloud domain Figure 3 A magnified view of a part can be seen as follows Figure 4 As shown, where Figure 4 In this context, 'x' represents the abnormal area, i.e., the aneurysm area.

[0069] In practice, this step can be implemented by extracting surface pixels from the blood vessel region of the CTA image to achieve image type conversion. Alternatively, other methods can be used, such as meshing the image domain data or removing non-boundary points from the image domain data to obtain a point cloud image.

[0070] Step S103: Input the point cloud domain image into a pre-trained target segmentation model to obtain a segmented image of the abnormal region. The target segmentation model includes a first segmentation module for segmenting the abnormal region and a second segmentation module for blocking the boundary between the adjacent blood vessel region and the abnormal region.

[0071] In this step, the first segmentation module of the target segmentation model can be used to identify and judge each point cloud / 3D point in the blood vessel image, determining whether each point cloud is a blood vessel point or an anomalous point. Based on the identification results of each point cloud, an initial abnormal region containing several anomalous points can be determined. Further, the second segmentation module is used to block the point cloud within a predetermined area on both sides of the boundary between the abnormal region and the blood vessel, to accurately determine whether the point cloud within the predetermined area on both sides of the boundary is a blood vessel point or an anomalous point, thus making the judgment of the abnormal region more accurate. The segmented image can be a point cloud domain image.

[0072] The method for segmenting abnormal regions of blood vessels in this embodiment obtains a point cloud image of the blood vessel image to be segmented. It can then use the first segmentation module in the target segmentation model to detect and identify each point cloud in the point cloud image, determine whether each point cloud is a blood vessel point or an abnormal point, and obtain an initial segmented image of the abnormal region. It also determines the boundary between the blood vessel region and the initial abnormal region, which facilitates the subsequent use of the second segmentation module to further optimize / block the point cloud near the boundary, thereby accurately identifying the abnormal points and accurately obtaining the final segmented image of the abnormal region.

[0073] Taking aneurysms as an example, the junction area occupies a small proportion of the image, resulting in less supervision from related technologies and poor junction segmentation. However, for medical image segmentation, especially aneurysm segmentation, junction segmentation is crucial, as it largely guides the location of aneurysm cutting. The target segmentation model in this application can not only segment blood vessels and aneurysms but also optimize the junction between them, thereby accurately identifying the aneurysm.

[0074] Another embodiment of this application provides a method for segmenting abnormal regions of blood vessels. Based on the above embodiments, this embodiment can specifically adopt any of the following methods when obtaining point cloud domain images based on blood vessel image conversion:

[0075] Method 1: Mesh out the blood vessels and abnormal areas in the blood vessel image, and extract at least one pixel within the mesh to obtain a pixel set; convert the pixel set into a point cloud domain to obtain a point cloud domain image.

[0076] Method 2: Extract the boundary pixel points of the blood vessel region and abnormal region within the blood vessel image to obtain a pixel set; convert the pixel set into a point cloud domain to obtain a point cloud domain image.

[0077] In this embodiment, by processing the non-point cloud domain vascular image to obtain the point cloud domain vascular image, the anatomical structural characteristics of the abnormal region can be fully utilized. The shape of the abnormal region's protrusion can be clearly shown in the point cloud domain image, laying the foundation for accurately determining the range of the abnormal region based on the protrusion features in the point cloud domain, and thus accurately segmenting the abnormal region from the point cloud domain image. The transformation of the pixel set into the point cloud domain is a coordinate transformation relationship, that is, transforming the pixel set from the image coordinate system to the point cloud coordinate system, which can be the world coordinate system.

[0078] Another embodiment of this application provides a method for segmenting abnormal regions of blood vessels. In this embodiment, when the point cloud domain image is input into a pre-trained target segmentation model to obtain a segmented image of the abnormal region, the specific process is as follows: an initial segmented image of the abnormal region is obtained based on the first segmentation module; the point cloud in a specific region on both sides of the boundary between the blood vessel and the abnormal region is blocked based on the second segmentation module to obtain a segmented image of the abnormal region. The blocking process includes classifying the point cloud located on one side of the abnormal region as the abnormal region and classifying the point cloud located on one side of the blood vessel region as the blood vessel region.

[0079] In other words, the first segmentation module in the target segmentation model can be used to extract features from each point cloud / 3D point of the blood vessel image, obtaining feature information for each point cloud. Then, based on the feature information, each point cloud can be determined to be either a blood vessel point or an anomaly point. Based on the judgment results of each point cloud, an initial abnormal region containing several anomalies can be identified, and the boundary between the initial abnormal region and the blood vessel region can be determined. It should be noted that due to insufficient attention to the point cloud at the boundary position, the output results of the first segmentation module near the boundary position may be inaccurate. Then, the second segmentation module can be used to further optimize the point cloud in the predetermined area on both sides of the boundary, thereby making a second judgment on whether the point cloud in the predetermined area on both sides of the boundary is a blood vessel point or an anomaly point. This makes the judgment of the abnormal region more accurate, providing a guarantee for the subsequent accurate segmentation of the abnormal region from the point cloud domain image.

[0080] Another embodiment of this application provides a method for segmenting abnormal regions of blood vessels. To reduce the computational load during abnormal region segmentation, before inputting the point cloud image into a pre-trained target segmentation model to obtain the segmented image of the abnormal region, the point cloud image can be segmented to obtain several blood vessel segment images, i.e., the point cloud image is segmented into several blood vessel segment images. Therefore, during abnormal region segmentation, each blood vessel segment image can be input into the pre-trained target segmentation model to obtain several sub-segmented images of the abnormal region corresponding to several blood vessel segments; finally, the several sub-segmented images are recombined to obtain the segmented image of the abnormal region.

[0081] In other words, for each blood vessel segment image, the first segmentation module in the target segmentation model extracts features from each point cloud / 3D point of the blood vessel segment image to obtain feature information of each point cloud. Then, based on the feature information of each point cloud, it is determined whether each point cloud is a blood vessel point or an anomalous point. Based on the judgment results of each point cloud, an initial abnormal region containing several anomalous points can be determined, and the boundary between the initial abnormal region and the blood vessel region can be determined. Then, the second segmentation module can be used to further optimize the point cloud in the predetermined area on both sides of the boundary, thereby making a second judgment on whether the point cloud in the predetermined area on both sides of the boundary is a blood vessel point or an anomalous point. Thus, based on the final recognition results of each point cloud, the abnormal region containing each anomalous point can be obtained. Of course, when the recognition results of each point cloud in the same blood vessel segment image are all blood vessel points, it is determined that the blood vessel segment image does not contain an abnormal region; when the recognition results of each point cloud in the same blood vessel segment image are all anomalous points, or some point clouds are anomalous points, it is determined that the blood vessel segment contains a dynamic abnormal region.

[0082] In this embodiment, after obtaining sub-segmented images of the abnormal regions corresponding to each blood vessel segment image, the blood vessel segment image containing the abnormal region can be identified as the target blood vessel segment image. Then, the first position information of the sub-segmented image within the target blood vessel segment image and the second position information of the target blood vessel segment within the point cloud domain image are determined. Based on the first and second position information, the target position information of the sub-segmented image within the point cloud domain image is determined. Therefore, each sub-segmented image is reconstructed according to its target position information to obtain the final abnormal region segmentation image. In this embodiment, since multiple abnormal regions may correspond to the same point cloud domain image, determining the target position of each sub-segmented image within the point cloud domain image makes the reconstruction of the sub-segmented images more accurate, ensuring the precise acquisition of the final segmentation image of the abnormal region.

[0083] Another embodiment of this application provides a method for segmenting abnormal regions of blood vessels, including:

[0084] Step S201: Train to obtain the target segmentation model;

[0085] In the specific implementation of this step, the training process of the target segmentation model is as follows:

[0086] Step S2011: Obtain a sample point cloud image containing abnormal regions and vascular regions, and a labeled point cloud image with the abnormal regions annotated.

[0087] In this step, several sample images can be acquired. Then, each acquired sample image is transformed to extract the surface pixels corresponding to the blood vessels in each sample image. Based on these surface pixels, a sample point cloud image corresponding to each sample image is constructed. Specifically, the blood vessels and abnormal regions within the blood vessel image are meshed, and at least one pixel within each mesh is extracted to obtain a pixel set. This pixel set is then converted into a point cloud image. Alternatively, the boundary pixels between the blood vessel region and the abnormal region within the blood vessel image can be extracted to obtain a pixel set. This pixel set is then converted into a point cloud image.

[0088] In this step, when acquiring the labeled point cloud image of the abnormal region, the specific steps are as follows: acquiring the boundary between the blood vessel and the abnormal region in the sample point cloud image, as well as the seed points within the abnormal region; based on the anatomical morphological features, boundary, and seed points of the abnormal region, labeling the abnormal region in the sample point cloud image to acquire the labeled point cloud image.

[0089] Specifically, abnormal regions can be manually labeled on point cloud images. That is, in response to user annotation requests for each sample point cloud image, the boundary between blood vessels and abnormal regions in the sample point cloud image, as well as seed points within the abnormal regions, are obtained. Then, based on the anatomical morphological features, boundaries, and seed points of the abnormal regions, the abnormal regions in the sample point cloud images are labeled, resulting in a labeled point cloud image. This annotation process is performed on point cloud data, and annotating abnormal regions on point cloud images has three advantages: First, the appearance of abnormal regions is clearer and easier to identify on point cloud images. Second, considering that an abnormal region is a convex closed-loop structure, image processing algorithms can be used to annotate only the boundary between the abnormal region and the blood vessel, and one seed point of the abnormal region. The image processing algorithm automatically generates all the annotation results, greatly reducing the workload of annotation. Third, this annotation method can not only annotate abnormal regions but also the connecting boundary regions between abnormal regions and blood vessels, making more detailed optimization of the boundaries possible. For example, for example, for... Figure 5 (a) shows the sample point cloud region image before annotation. The annotated sample point cloud region image can be seen as follows: Figure 5 As shown in (b), users only need to mark the connecting line y and a seed point x of an aneurysm in the image. Using image processing algorithms, the entire aneurysm region can be labeled. By adopting this labeling method, the labeling speed is faster, the labeled area is more accurate, more label categories are available, and the labeling of the aneurysm region and the junction of the aneurysm and the blood vessel can be obtained simultaneously.

[0090] Step S2012: Train an initial first segmentation module based on the sample point cloud domain image and the label point cloud domain image, and obtain the first segmentation module when the first preset termination condition is met;

[0091] In this step, the abnormal regions of each sample point cloud image can be segmented based on the first segmentation module in the initial segmentation model to obtain the segmentation results of each sample point cloud image. Based on the difference between the segmentation results of each sample point cloud image and the label information, the model parameters in the first segmentation module are adjusted until a first preset termination condition is met, at which point training stops, thus obtaining the first segmentation module. The first preset termination adjustment includes the loss function corresponding to the first segmentation module reaching a preset threshold, or the training count of the first segmentation module reaching a preset number of times.

[0092] In the specific implementation process of this step: First, annotation can be performed on the unsegmented sample point cloud domain data; after annotation, segmentation is performed, and each blood vessel segment is used as a training sample, that is, blood vessel segment images containing abnormal regions are used as positive samples, and blood vessel segment images not containing abnormal regions are used as negative samples. The representation form can be coordinate values, but is not limited to coordinate values, and can also be features such as normal vectors. The positive and negative samples are input into the initial first segmentation module. The initial first segmentation module is used to extract features from each three-dimensional point / point cloud of each sample blood vessel segment image to obtain the feature information of each sample point. Then, anomaly point identification is performed based on the feature information of each sample point to obtain the first identification result of each sample point, thereby obtaining the abnormal region segmentation result of each sample point cloud domain image; finally, based on the abnormal region segmentation result of each sample point cloud domain image and the label information of whether each sample point cloud domain image contains abnormal regions, the difference between the segmentation result of each sample point cloud domain image and the corresponding label information of each sample point cloud domain image is determined. Then, the parameters in the initial first segmentation module are adjusted based on the difference until the training conditions are met and training stops, thus obtaining the first segmentation module. In this step, the network structure of the first segmentation module can adopt an autoencoder network similar to Unet (Unet is a segmentation network for medical imaging proposed based on Fcn), but is not limited to it. The final first segmentation module can be obtained through optimization of the loss function.

[0093] Step S2013: Based on the labeled point cloud image, obtain the boundary between the abnormal region and the blood vessel region;

[0094] For example, targeting Figure 6 As shown, firstly, the boundary region Ω between the abnormal region and the vascular region can be determined based on the annotation results; then, the region on one side of the boundary can be represented as Ω. + The region on the other side is represented as Ω. - .

[0095] Step S2014: Based on the boundary and label point cloud domain image and the output image of the first segmentation module, train the initial second segmentation module. When the second preset termination condition is reached, obtain the second segmentation module. The region of interest of the second segmentation module includes at least the region within a specific range on both sides of the boundary.

[0096] In this step, for positive samples containing abnormal regions, the boundary region between the abnormal region and the blood vessel region in the blood vessel segment image / point cloud domain image is determined. Then, based on the boundary region, the regions of interest located on both sides of the boundary region are determined. The feature information of each sample point cloud in the region of interest is obtained; in this step, the feature information of each sample point cloud can be extracted by the first segmentation submodule. That is, the feature information of each sample point cloud in the region of interest obtained by the first segmentation module is input into the initial second segmentation module, and then the initial second segmentation module is used to perform secondary recognition on each sample point cloud in the region of interest to obtain the second recognition result of each sample point cloud. Based on the second recognition result of each point cloud in the region of interest in the same point cloud domain image / blood vessel segment image and the recognition result of each point cloud in the non-region of interest, the recognition result of the current abnormal point corresponding to the point cloud domain image / blood vessel segment image is obtained. Finally, based on the difference between the recognition result of the current abnormal point and the label information corresponding to each point cloud domain image / blood vessel segment image, the parameters in the initial second segmentation module are adjusted to obtain the final second segmentation module. The second preset termination adjustment includes the loss function corresponding to the second segmentation module reaching a preset threshold, or the training number of the second segmentation module reaching a preset number.

[0097] In practical implementation, the region of interest can be determined using the nearest neighbor algorithm and a fixed threshold (the width of the boundary). Figure 6 Ω + and Ω - Within the region, the region of interest A that is relatively close to the boundary is determined iteratively. + and Area of ​​Interest A - Of course, focus on region A. + and Area of ​​Interest A - Manual selection is also possible. Finally, secondary detection can be performed on the sample point clouds within these two regions of interest to optimize the samples that are prone to prediction errors, thereby training the second segmentation module. The specific training process is as follows: the feature information learned by the first segmentation module is used as input. The starting point for the second segmentation module is the region of interest A. + The prediction / identification results for each sample point should be consistent. Similarly, the focus should be on region A. - The prediction / recognition results for the mid-sample points should also be consistent, and the focus should be on region A. + With regard to area A - The results of the point clouds for each sample within the same region should be reversed. That is, due to the existence of the boundary region, A is blocked. + and A -The connection of the sample point cloud means that during the training process, the samples in the receptive domain must be samples of the same category, which can greatly enhance the segmentation effect. Based on the difference between the results obtained by the initial second segmentation module and the label information of each sample point cloud in the region of interest, the parameters of the initial second segmentation module are adjusted until the training conditions are met and training stops, thus obtaining the final second segmentation module.

[0098] In this step, after obtaining the first and second segmentation modules, the target segmentation model can be constructed based on these two modules.

[0099] Step S202: Obtain a point cloud domain image containing the blood vessel region and the abnormal region;

[0100] In this step, the blood vessels and abnormal regions within the vascular image can be meshed, and at least one pixel within each mesh can be extracted to obtain a pixel set. This pixel set is then converted into a point cloud domain to obtain a point cloud domain image. Alternatively, the boundary pixels between the blood vessel region and the abnormal region within the vascular image can be extracted to obtain a pixel set; this pixel set can then be converted into a point cloud domain to obtain a point cloud domain image. By extracting surface pixels of the vascular object and constructing a point cloud domain image based on these surface pixels, a point cloud domain image containing sufficient vascular surface features can be obtained, laying the foundation for subsequent segmentation of abnormal regions of the blood vessel based on this point cloud domain image.

[0101] Step S203: Divide the point cloud image into several blood vessel segment images;

[0102] In this step, the segmentation / division processing can be performed as follows: N points are uniformly sampled from the overall point cloud image. Then, using each point as the initial point, all points whose distance from the initial point to the ground is less than a certain threshold are recorded. These points, together with the initial point, form a blood vessel segment. This process can be repeated to divide an overall point cloud image into N blood vessel segments.

[0103] In this step, after image segmentation, the images of each segmented blood vessel segment can be obtained as follows: Figure 7 As shown.

[0104] Step S204: Input the images of each blood vessel segment into the pre-trained target segmentation model to obtain several sub-segmentation images of the abnormal regions corresponding to several blood vessel segments; reassemble the several sub-segmentation images to obtain the segmentation images of the abnormal regions.

[0105] In this step, for each blood vessel segment image, the first segmentation module in the target segmentation model is used to extract features from each point cloud / 3D point of the blood vessel segment image to obtain feature information of each point cloud. Then, based on the feature information of each point cloud, it is determined whether each point cloud is a blood vessel point or an anomaly point. Thus, based on the first judgment / first recognition result of each point cloud, an initial abnormal region containing several anomalies can be determined, and the boundary between the initial abnormal region and the blood vessel region can be determined. Then, the second segmentation module can be used to further optimize the point clouds in the predetermined areas on both sides of the boundary, thereby making a second judgment on whether the point clouds in the predetermined areas (areas of interest) on both sides of the boundary are blood vessel points or anomalies, that is, obtaining the second recognition result of each point cloud in the area of ​​interest. Thus, based on the final recognition result of each point cloud, the abnormal region containing each anomaly point can be obtained. Of course, when the recognition result of each point cloud in the same blood vessel segment image is a blood vessel point, it is determined that the blood vessel segment image does not contain an abnormal region; when the recognition result of each point cloud in the same blood vessel segment image is an anomaly point, or some point clouds are anomalies, it is determined that the blood vessel segment contains a dynamic abnormal region.

[0106] In this embodiment, after obtaining sub-segmented images of the abnormal regions corresponding to each blood vessel segment image, the blood vessel segment image containing the abnormal region can be identified as the target blood vessel segment image. Then, the first position information of the sub-segmented image within the target blood vessel segment image and the second position information of the target blood vessel segment within the point cloud domain image are determined. Based on the first and second position information, the target position information of the sub-segmented image within the point cloud domain image is determined. Therefore, each sub-segmented image is reconstructed according to its target position information to obtain the final abnormal region segmentation image. In this embodiment, since multiple abnormal regions may correspond to the same point cloud domain image, determining the target position of each sub-segmented image within the point cloud domain image makes the reconstruction of the sub-segmented images more accurate, ensuring the precise acquisition of the final segmentation image of the abnormal region.

[0107] Step S205: Different display methods are used to display the abnormal areas and blood vessel areas in the point cloud image.

[0108] In this step, after reconstructing each sub-segmented image, the location of the abnormal region in the point cloud image can be further determined. Then, based on this location, different display methods can be used to display the abnormal region and the blood vessel region. Alternatively, different display methods can be used for the abnormal region and the blood vessel region in each blood vessel segment image.

[0109] In this step, the results from the point cloud domain can be directly displayed. This display method can be directly output to applications that require 3D reconstructed images, such as metaverse modeling and manipulation. By determining the location of the abnormal region in the point cloud domain image and displaying it using different methods, this step allows users to more clearly understand the specific location of the aneurysm, providing assistance for subsequent diagnosis and treatment.

[0110] Furthermore, in this embodiment, after identifying the abnormal region or blood vessel region, the result in the point cloud domain can be further post-processed to restore it to the image domain, resulting in a non-point cloud domain image. That is, the target segmentation model also includes a post-processing module to restore the output of the second segmentation module to the image domain, i.e., segmenting the image into an image domain image. Based on the position coordinates of the abnormal region in the point cloud domain image, a first region corresponding to the abnormal region and a second region corresponding to the blood vessel region in the non-point cloud domain image are determined. Finally, the first and second regions are displayed based on different display methods.

[0111] The abnormal region detection method in this embodiment has higher sensitivity and lower false positives: this is because this application makes full use of the anatomical structure characteristics of the abnormal region, which can clearly show the shape change of its protrusion in the point cloud domain, so the segmentation accuracy is higher.

[0112] The abnormal region segmentation method in this application has stronger generalization: the method in this application does not require the original CTA image, but only the segmentation result of the blood vessel. Therefore, the segmentation effect of the aneurysm is not affected by the image contrast, and the imaging differences of different modalities and different manufacturers' equipment have little impact on it.

[0113] The aneurysm segmentation method in this application requires fewer computational resources: the point cloud domain data only collects points on the surface of the region of interest, and these points are downsampled during the training and application phases, so fewer sample points are actually involved in the calculation, resulting in less computational resource consumption.

[0114] The aneurysm segmentation method in this application is better suited to clinical needs: due to the function of the abnormal region boundary optimization module, the segmentation results of the abnormal region obtained by the method of this application are more accurate in segmenting the connection position with the blood vessel, which is the most important location in clinical tasks such as aneurysm resection surgery.

[0115] In this application, regarding Figure 8The intracranial aneurysm CAT image shown in (a) can be used to detect the aneurysm using the abnormal region segmentation method described in this application, resulting in the aneurysm segmentation map and its corresponding 3D reconstruction map shown in Figures (c) and (e). Figures (b) and (d) show the aneurysm segmentation map and its corresponding 3D reconstruction map obtained using the traditional manual labeling method. The comparison shows that the segmentation results obtained by this invention accurately identify the aneurysm and clearly segment its boundaries, thus meeting the needs of clinical applications and demonstrating good practicality.

[0116] Another embodiment of this application provides a device for segmenting abnormal regions of blood vessels, such as... Figure 9 As shown, it includes:

[0117] The first acquisition module 11 is used to acquire a blood vessel image to be segmented, the blood vessel image including a blood vessel region and an abnormal region attached to the blood vessel.

[0118] The second acquisition module 12 is used to acquire a point cloud domain image containing the blood vessel region and the abnormal region based on the blood vessel image;

[0119] The segmentation module 13 is used to input the point cloud domain image into a pre-trained target segmentation model to obtain a segmented image of the abnormal region. The target segmentation model includes a first segmentation module for segmenting the abnormal region and a second segmentation module for blocking the boundary between the adjacent blood vessel region and the abnormal region.

[0120] In this embodiment, the first acquisition module is specifically used to: perform gridding processing on the blood vessels and abnormal regions in the blood vessel image, and extract at least one pixel point within the grid to obtain a pixel point set; convert the pixel point set into a point cloud domain to obtain a point cloud domain image; or, extract the boundary point pixels of the blood vessel region and abnormal region in the blood vessel image to obtain a pixel point set; convert the pixel point set into a point cloud domain to obtain a point cloud domain image.

[0121] In this embodiment, the segmentation module is specifically used to: obtain an initial segmented image of the abnormal region based on the first segmentation module; and perform blocking processing on the point clouds in specific areas on both sides of the boundary between the blood vessel and the abnormal region in the initial segmented image based on the second segmentation module to obtain a segmented image of the abnormal region. The blocking processing includes classifying the point clouds located on one side of the abnormal region as the abnormal region and classifying the point clouds located on the side of the blood vessel region as the blood vessel region.

[0122] In this embodiment, the segmentation device for the abnormal region of the blood vessel further includes a segmentation module. The segmentation module is used to: segment the point cloud image into several blood vessel segment images before inputting the point cloud image into a pre-trained target segmentation model to obtain the segmented image of the abnormal region.

[0123] The segmentation module is specifically used to: input the images of the several blood vessel segments into a pre-trained target segmentation model to obtain several sub-segmentation images of the abnormal regions corresponding to the several blood vessel segments; and reassemble the several sub-segmentation images to obtain the segmentation image of the abnormal regions.

[0124] In this embodiment, the segmentation device for the abnormal region of the blood vessel further includes a training module. The training module is used to: acquire a sample point cloud image containing the abnormal region and the blood vessel, and a labeled point cloud image annotating the abnormal region; train an initial first segmentation module based on the sample point cloud image and the labeled point cloud image, and acquire the first segmentation module when a first preset termination condition is met; acquire the boundary between the abnormal region and the blood vessel based on the labeled point cloud image; train an initial second segmentation module based on the boundary, the labeled point cloud image, and the output image of the first segmentation module, and acquire the second segmentation module when a second preset termination condition is met, wherein the region of interest of the second segmentation module includes at least the region within a specific range on both sides of the boundary.

[0125] In this embodiment, the training module is specifically used to determine the region of interest of the second segmentation module through a nearest neighbor algorithm or a fixed threshold method.

[0126] In this embodiment, the training module is specifically used to: obtain the boundary between blood vessels and abnormal regions in the sample point cloud image, as well as seed points within the abnormal regions;

[0127] Based on the anatomical morphological features, boundaries, and seed points of the abnormal regions, the abnormal regions in the sample point cloud image are labeled to obtain a labeled point cloud image.

[0128] In this embodiment, the segmentation device for abnormal regions of blood vessels further includes a display module, which is used to display the abnormal regions and blood vessel regions in the point cloud image using different display methods.

[0129] The segmentation device for abnormal regions of blood vessels in this application acquires a point cloud image of the blood vessel image to be segmented. It can then use the first segmentation module in the target segmentation model to detect and identify each point cloud in the point cloud image, determine whether each point cloud is a blood vessel point or an abnormal point, thus obtaining an initial segmentation image of the abnormal region. It also determines the boundary between the blood vessel region and the initial abnormal region, which facilitates the subsequent use of the second segmentation module to further optimize / block the point cloud near the boundary, thereby accurately identifying the abnormal points and accurately obtaining the final segmentation image of the abnormal region.

[0130] Another embodiment of this application provides an electronic device, including at least a memory and a processor. The memory stores a computer program, and the processor, when executing the computer program in the memory, performs the following method steps:

[0131] Step 1: Obtain the blood vessel image to be segmented, which includes the blood vessel region and abnormal regions attached to the blood vessel.

[0132] Step 2: Based on the blood vessel image, obtain a point cloud domain image containing the blood vessel region and the abnormal region;

[0133] Step 3: Input the point cloud image into a pre-trained target segmentation model to obtain a segmented image of the abnormal region. The target segmentation model includes a first segmentation module for segmenting the abnormal region and a second segmentation module for blocking the boundary between the adjacent blood vessel region and the abnormal region.

[0134] The specific implementation process of the above method steps can be found in the embodiment of the above-mentioned method for segmenting abnormal regions of any blood vessel, which will not be repeated here.

[0135] The electronic device in this embodiment, by acquiring the point cloud domain image of the blood vessel image to be segmented, can use the first segmentation module in the target segmentation model to detect and identify each point cloud in the point cloud domain image, determine whether each point cloud is a blood vessel point or an abnormal point, that is, obtain the segmentation image of the initial abnormal region, and determine the boundary between the blood vessel region and the initial abnormal region, so as to facilitate the subsequent use of the second segmentation module to further optimize / block the point cloud near the boundary, thereby accurately determining the abnormal point, and thus accurately obtaining the final segmentation image of the abnormal region.

[0136] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method for segmenting abnormal regions of blood vessels, characterized in that, include: Obtain a blood vessel image to be segmented, the blood vessel image including the blood vessel region and abnormal regions attached to the blood vessel; Based on the blood vessel image, a point cloud domain image containing the blood vessel region and the abnormal region is obtained; The point cloud image is input into a pre-trained target segmentation model to obtain a segmented image of the abnormal region. The target segmentation model includes a first segmentation module for segmenting the abnormal region and a second segmentation module for blocking the boundary between the adjacent blood vessel region and the abnormal region. The first segmentation module is used to identify each point cloud in the point cloud domain image, determine whether each point cloud is a blood vessel point or an abnormal point, and determine an initial abnormal region containing several abnormal points and the boundary between the blood vessel region and the initial abnormal region based on the identification results of each point cloud; the second segmentation module is used to block the point cloud in a predetermined area on both sides of the boundary between the initial abnormal region and the blood vessel region, so as to determine whether the point cloud in the predetermined area on both sides of the boundary is a blood vessel point or an abnormal point.

2. The method as described in claim 1, characterized in that, The step of obtaining a point cloud domain image containing the blood vessel region and the abnormal region based on the blood vessel image includes: The blood vessels and abnormal regions in the blood vessel image are meshed, and at least one pixel is extracted from the mesh to obtain a set of pixels. Convert the pixel set into a point cloud domain to obtain a point cloud domain image; or, Extract the boundary point pixels of the blood vessel region and abnormal region within the blood vessel image to obtain a pixel set; The set of pixels is converted into a point cloud domain to obtain a point cloud domain image.

3. The method as described in claim 1, characterized in that, The step of inputting the point cloud image into a pre-trained target segmentation model to obtain a segmented image of the abnormal region includes: The initial segmentation image of the abnormal region is obtained based on the first segmentation module; Based on the second segmentation module, the point cloud in a specific area on both sides of the boundary between the abnormal region and the blood vessel region in the initial segmentation image is blocked to obtain the segmentation image of the abnormal region. The blocking process includes classifying the point cloud located on one side of the abnormal region as the abnormal region and classifying the point cloud located on the one side of the blood vessel region as the blood vessel region.

4. The method as described in claim 1, characterized in that, Before inputting the point cloud image into a pre-trained target segmentation model to obtain the segmented image of the abnormal region, the method further includes: The point cloud image is segmented into several blood vessel segment images; The step of inputting the point cloud image into a pre-trained target segmentation model to obtain a segmented image of the abnormal region includes: The images of the several blood vessel segments are input into a pre-trained target segmentation model to obtain several sub-segmentation images of the abnormal regions corresponding to the several blood vessel segments. The several sub-segmented images are recombined to obtain the segmented image of the abnormal region.

5. The method as described in claim 1, characterized in that, Also includes: Obtain a pre-trained target segmentation model, wherein obtaining the pre-trained target segmentation model includes: Acquire a sample point cloud image containing abnormal regions and blood vessels, and a labeled point cloud image with the abnormal regions annotated. An initial first segmentation module is trained based on the sample point cloud domain image and the label point cloud domain image, and the first segmentation module is obtained when the first preset termination condition is met. Based on the labeled point cloud image, the boundary between the abnormal region and the blood vessel is obtained; Based on the boundary and label point cloud images and the output image of the first segmentation module, an initial second segmentation module is trained. When a second preset termination condition is reached, the second segmentation module is acquired. The region of interest of the second segmentation module includes at least the region within a specific range on both sides of the boundary. The first preset termination condition includes the value of the loss function corresponding to the first segmentation module reaching a preset threshold or the number of training iterations of the first segmentation module reaching a preset number. The second preset termination condition includes the value of the loss function corresponding to the second segmentation module reaching a preset threshold or the number of training iterations of the second segmentation module reaching a preset number.

6. The method as described in claim 5, characterized in that, Also includes: The region of interest for the second segmentation module is determined by either the nearest neighbor algorithm or a fixed threshold method.

7. The method as described in claim 5, characterized in that, Acquiring a labeled point cloud image of the anomaly region includes: Obtain the boundary between blood vessels and abnormal regions in the sample point cloud image, as well as the seed points within the abnormal regions; Based on the anatomical morphological features, boundaries, and seed points of the abnormal regions, the abnormal regions in the sample point cloud image are labeled to obtain a labeled point cloud image.

8. The method according to claim 1, characterized in that, The target segmentation model also includes a post-processing module for restoring the output of the second segmentation module to the image domain.

9. A device for segmenting abnormal regions of blood vessels, characterized in that, include: The first acquisition module is used to acquire a blood vessel image to be segmented, the blood vessel image including a blood vessel region and an abnormal region attached to the blood vessel. The second acquisition module is used to acquire a point cloud domain image containing the blood vessel region and the abnormal region based on the blood vessel image; The segmentation module is used to input the point cloud image into a pre-trained target segmentation model to obtain a segmented image of the abnormal region. The target segmentation model includes a first segmentation module for segmenting the abnormal region and a second segmentation module for blocking the boundary between the adjacent blood vessel region and the abnormal region. The first segmentation module is used to identify each point cloud in the point cloud domain image, determine whether each point cloud is a blood vessel point or an abnormal point, and determine an initial abnormal region containing several abnormal points and the boundary between the blood vessel region and the initial abnormal region based on the identification results of each point cloud; the second segmentation module is used to block the point cloud in a predetermined area on both sides of the boundary between the initial abnormal region and the blood vessel region, so as to determine whether the point cloud in the predetermined area on both sides of the boundary is a blood vessel point or an abnormal point.

10. An electronic device, characterized in that, It includes at least a memory and a processor, wherein the memory stores a computer program, and the processor, when executing the computer program in the memory, implements the steps of the method for segmenting abnormal regions of blood vessels according to any one of claims 1-8.

Images