A coronary vessel segmentation method and device and a storage medium

Abstract

The application relates to a coronary vessel segmentation method and device and a storage medium. The segmentation method comprises the following steps: acquiring a three-dimensional CTA image containing a coronary vessel; inputting the three-dimensional CTA image into a trained bifurcation point detection network to obtain bifurcation point information of the coronary vessel, wherein the bifurcation point detection network is constructed based on a first deep full convolution network; and inputting the three-dimensional CTA image and the bifurcation point information into a trained blood vessel segmentation network to determine a segmentation result of the coronary vessel, wherein the blood vessel segmentation network is constructed based on a second deep full convolution network. The coronary vessel segmentation method can quickly and accurately segment the coronary vessel from the three-dimensional CTA image in an end-to-end manner, and the reliability and accuracy of the blood vessel segmentation are further improved due to the bifurcation points in the coronary vessel as auxiliary information.

Description

Technical Field

[0001] This application relates to the field of medical image processing, and more specifically, to a method, apparatus, and storage medium for segmenting coronary arteries. Background Technology

[0002] Non-invasive vascular function assessment technology developed based on coronary CT imaging can provide hemodynamic information (physiological function assessment) of diseased vessels, thus providing non-invasive "one-stop" anatomical and functional evaluation without increasing radiation dose. Compared with invasive FFR (fractional flow reserve) measurement, non-invasive FFR assessment technology has more advantages.

[0003] Non-invasive functional assessment using CTA typically requires segmentation of blood vessels based on 3D CT images, known as 3D vessel segmentation. 3D vessel segmentation is a computer algorithm process that distinguishes the foreground and background of blood vessels from coronary CTA images and extracts pixel-level 3D models of the vessels, producing pixel-level segmentation results. However, accurately and reliably segmenting coronary arteries from CTA images presents certain challenges, primarily due to the variable orientation and thickness of the coronary arteries themselves, the varying noise levels in different CTA images, and the requirement for the segmentation method to rapidly process large 3D images.

[0004] Traditional image brightness-based segmentation methods, such as level sets, have been used for coronary artery segmentation in CTA in the literature. These methods have high requirements for the brightness contrast between the coronary arteries and surrounding tissues in the image, and their performance varies on images with different degrees of coronary artery brightening. They are also sensitive to noise in the image.

[0005] Another commonly used method for segmenting blood vessels is based on multi-scale second-order derivative calculations around each pixel in the image to perform tubularity analysis, highlighting areas with shapes resembling tubes and suppressing other areas. However, the tubularity value output by this method is sensitive to the thickness of the blood vessels, requiring the use of different parameters to highlight vessels of varying thicknesses in practical applications, resulting in uneven extraction of vessels of different thicknesses. Furthermore, the computational speed is very slow for large 3D CTA images due to the need to calculate the second derivative at each point. Summary of the Invention

[0006] This application addresses the aforementioned deficiencies in the prior art. A method, apparatus, and storage medium for segmenting coronary arteries are needed, which can fully utilize the bifurcation point information in the coronary arteries, enabling faster, more accurate, robust, and higher-precision segmentation results when using a vascular segmentation network for coronary artery segmentation.

[0007] According to the first aspect of this application, a method for segmenting coronary vessels is provided. The segmentation method includes: acquiring a three-dimensional CTA image containing coronary vessels; determining the bifurcation point information of the coronary vessels based on the CTA image using a trained bifurcation point detection network; and determining the segmentation result of the coronary vessels based on the CTA image and the bifurcation point information using a trained vessel segmentation network.

[0008] According to a second aspect of this application, a coronary artery segmentation apparatus is provided, the apparatus comprising: an interface configured to acquire a three-dimensional CTA image containing coronary arteries; and a processor configured to execute the coronary artery segmentation method according to various embodiments of this application.

[0009] According to a third aspect of this application, a computer-readable storage medium is provided, on which instructions are stored, wherein, when executed by a processor, the instructions perform the steps of the coronary artery segmentation method described in the various embodiments of this application.

[0010] The coronary artery segmentation methods, apparatus, and storage media provided in the various embodiments of this application first detect bifurcation points in the coronary artery structure of a 3D CTA image using a bifurcation point detection network. The bifurcation point information is then used as auxiliary information and input along with the 3D CTA image into a trained vessel segmentation network. This provides auxiliary and guiding information during vessel segmentation, resulting in higher accuracy and robustness of the determined coronary artery segmentation, and higher segmentation precision. Furthermore, since both the bifurcation point detection network and the vessel segmentation network employ fully convolutional networks, they have lower computational costs, resulting in higher training and operational efficiency when segmenting complex 3D CTA images of coronary arteries. Attached Figure Description

[0011] Figure 1 A flowchart illustrating a method for segmenting coronary arteries according to an embodiment of this application;

[0012] Figure 2 The structure and training process of a second deep fully convolutional network according to an embodiment of this application are shown;

[0013] Figure 3 This diagram illustrates the extraction of true values ​​for blood vessel bifurcation points according to an embodiment of this application; and

[0014] Figure 4 A schematic diagram showing a portion of the components of a coronary artery segmentation device according to an embodiment of this application is provided. Detailed Implementation

[0015] To enable those skilled in the art to better understand the technical solutions of this disclosure, the present application will be described in detail below with reference to the accompanying drawings and specific embodiments. The embodiments of the present application will be further described in detail below with reference to the accompanying drawings and specific examples, but these are not intended to limit the present application.

[0016] The terms “first,” “second,” and similar words used in this application do not indicate any order, quantity, or importance, but are merely used for distinction. Words such as “including” or “comprising” mean that the element preceding the word covers the element listed after the word, and do not exclude the possibility of covering other elements as well.

[0017] This application proposes a method for segmenting coronary arteries. Figure 1 A flowchart illustrating a method for segmenting coronary arteries according to an embodiment of this application is shown. Figure 1 As shown, the method begins with step 101, acquiring a three-dimensional CTA image containing coronary vessels.

[0018] The three-dimensional CTA image containing coronary vessels is a three-dimensional CTA image obtained by computed tomography angiography of the coronary arteries. In some embodiments, such as various non-invasive vascular assessment techniques developed based on three-dimensional CTA images, it is necessary to segment the vessels in the three-dimensional CTA image to distinguish the foreground and background of the vessels, extract pixel-level three-dimensional models of the vessels, and generate pixel-level vessel segmentation results.

[0019] In step 102, the 3D CTA image is input into a trained bifurcation detection network to obtain bifurcation information of the coronary arteries. This bifurcation detection network is constructed based on a first-depth fully convolutional network. It is important to note that the output of this bifurcation detection network is for bifurcation points within the anatomical structure of the coronary arteries; bifurcation points in other blood vessels or tissues outside the coronary arteries are not recognized and are excluded.

[0020] As an example, coronary arteries exhibit a tree-like structure, with one or more bifurcation points along their tapering course. These bifurcation points explicitly or implicitly contain information about the vascular anatomy, which can guide or assist in segmentation in certain situations. For instance, bifurcation point information can significantly enhance coronary artery localization and guide segmentation algorithms in extracting the correct coronary artery tree topology. For example, if bifurcation point information indicates that a certain location in the image is not a bifurcation point, then during coronary artery segmentation, the bifurcation originating from that location can be assumed to be a vein or other tissue structure, or a segmentation error caused by noise. The various ways in which bifurcation point information can be utilized are not listed here.

[0021] In some embodiments, the bifurcation point detection network is configured to be built based on a deep fully convolutional network (i.e., a deep fully convolutional neural network). Sufficient model capacity can also be achieved by adjusting the network depth (i.e., the number of layers). In some embodiments, the specific configuration of the deep fully convolutional neural network can be configured such that the bifurcation point detection network has only convolutional layers and no fully connected layers, thereby enabling it to have high training speed and low computational cost during training and prediction.

[0022] Next, in step 103, the 3D CTA image and the bifurcation point information are input into the trained vascular segmentation network to determine the segmentation result of the coronary arteries. The vascular segmentation network is constructed based on a second deep fully convolutional network. In some embodiments, the segmentation result of the coronary arteries is a pixel-level vascular segmentation mask. In some embodiments, the segmentation result of the coronary arteries may also be represented in other forms, which are not limited in this application.

[0023] Through step 102 above, information on key and important coronary artery bifurcation points can be obtained. This information, along with the 3D CTA image, is used as input to a vascular segmentation network also constructed based on a deep fully convolutional network. Thus, with the assistance or guidance of bifurcation point information, problems such as the complex anatomy of coronary arteries, large individual differences, significant differences in the thickness of proximal and distal segments, interference from veins and other tissues, and varying levels of image noise and image quality can be effectively addressed. This results in more complete, accurate, and reliable coronary artery segmentation output by the vascular segmentation network, and improvements in both the robustness and segmentation accuracy of the vascular segmentation network.

[0024] In some embodiments, the bifurcation point information of coronary arteries can be represented in different forms, such as including but not limited to bifurcation point masks, bifurcation point coordinates, spheres obtained based on bifurcation point expansion, heat maps obtained based on bifurcation point transformations, etc. The above transformations can be Gaussian transformations or other transformations, as long as they can characterize the probability, confidence level, etc., of one or more pixels belonging to the bifurcation point of the blood vessel.

[0025] In some embodiments, the bifurcation point information can be stored in the same way as the size of the 3D CTA image. For example, it can be stored in a binarized manner. The bifurcation point mask formed by marking the bifurcation points of each pixel in the 3D CTA image can be used together with the 3D CTA image to input into the trained blood vessel segmentation network. The 3D CTA image and the bifurcation point information can be concatenated (also called connection) point by point before being input into the trained blood vessel segmentation network for the prediction of the segmentation results of coronary arteries.

[0026] Figure 2The structure and training process of a second deep fully convolutional network according to an embodiment of this application are shown. Figure 2 Taking the second deep fully convolutional network 200 used in the blood vessel segmentation network as an example, the structure of the deep fully convolutional neural network is illustrated. However, it should be noted that the first deep fully convolutional network used in the bifurcation point detection network also adopts a similar structure and training method as the second deep fully convolutional network.

[0027] Figure 2 The meanings of the various symbols in a medium-depth fully convolutional neural network 200 are shown in the illustration on the right. For example... Figure 2 As shown, the second deep fully convolutional network 200 can be constructed using a network structure such as U-Net, and is a fully convolutional structure, that is, it only contains convolutional layers and does not contain fully connected layers that consume a lot of memory and computation time. Thus, its computational speed has a significant advantage when processing large 3D CTA images. Furthermore, the second deep fully convolutional network 200 is a deep fully convolutional neural network that employs multi-scale information, for example, it can sequentially contain layers with different resolutions, each having an encoder and a decoder. Figure 2 Taking a high-resolution layer 201, a medium-resolution layer 202, and a low-resolution layer 203 as an example, the network can quickly and accurately fuse image information at different resolutions in 3D CTA. In other embodiments, the second deep fully convolutional network 200 may have more or fewer layers, depending on the comprehensive requirements for performance indicators such as blood vessel segmentation accuracy and speed; this application does not limit this.

[0028] In some embodiments, the second deep fully convolutional network 200 includes an abstract path and a restoration path. Specifically, the path from high resolution to medium resolution and then to low resolution is called the abstract path. On the abstract path, image features are extracted layer by layer through a series of convolution and sampling operations. The image features extracted by the shallow layers of the network are relatively low-level and primitive, similar to traditional edge detection features, while the image features extracted by the deep layers of the network are higher-level and more abstract, closer to human cognitive classification. Therefore, as an example only, in the second deep fully convolutional network 200, the high-resolution layer 201 corresponds to the original resolution 3D CTA image, and the medium-resolution layer 202 and low-resolution layer 203 correspond to the images after the original resolution 3D CTA image has undergone the first downsampling 211 processing and the second downsampling 212 processing, respectively. By performing feature encoding processing on the inputs of different resolutions at each level, the coronary vessel features associated with the coronary vessels are extracted layer by layer from the higher-dimensional representation after the 3D CTA image and bifurcation point information are spliced ​​together. The coronary vessel structures of interest in the image are abstracted hierarchically in a large field of view (in the first deep fully convolutional network used for bifurcation point detection, the abstracted and refined structures are the bifurcation points in the coronary vessel structures), and false positive structures with similar appearances, such as veins, are excluded (in the first deep fully convolutional network used for bifurcation point detection, the excluded false positive structures include vein bifurcation points, nodules, etc.).

[0029] In some embodiments, the second deep fully convolutional network 200 restores the feature maps that are progressively reduced during the layer-by-layer abstraction process through a restoration path to obtain a fine-grained blood vessel segmentation result corresponding to the original high-resolution image. In the restoration path, the location information corresponding to the features at each resolution level saved during the abstraction process is used to accurately restore the abstracted coronary artery structure to the pixel-level precise location in the original image layer by layer. During this process, the restoration path integrates the high-resolution feature map information of the corresponding layers in the abstraction path to improve the localization accuracy of coronary artery segmentation. Specifically, as follows... Figure 2 As shown, after the feature map of the low-resolution layer 203 is processed by the second upsampling 213, it is concatenated (or added, etc.) with the feature map in the next higher-resolution layer 202. The medium-resolution layer 202 processes the concatenated feature map, and after the feature map generated by this layer is processed by the first upsampling 214, it is concatenated (or added, etc.) with the feature map in the high-resolution layer 201. In this way, along the restoration path, the coronary artery structure abstracted from the lower-resolution layer is fused with the feature map information (such as the position information) of the previous resolution layer, and finally the pixel-level accurate segmentation of the coronary artery in the original highest-resolution 3D CTA image is achieved.

[0030] It should be noted that, in order to ensure that feature maps at different resolution levels have the same size so that they can be merged by splicing (or addition, etc.), the ratios of the first downsampling 211 and the first upsampling 214, the second downsampling 212 and the second upsampling 213 are set accordingly. Furthermore, downsampling can be done using methods such as Gaussian pyramids, and upsampling can be done using methods such as skip-layer connections, or any other applicable methods, as long as the transformation of resolution and feature map size can be achieved. This application does not impose any restrictions on this.

[0031] In some embodiments, the second deep fully convolutional network 200 can be trained based on a second training dataset, wherein the second training dataset consists of a three-dimensional CTA image containing coronary vessels, ground truth values ​​of vessel segmentation of coronary vessels in the three-dimensional CTA image, and information on the corresponding bifurcation points of coronary vessels.

[0032] In some embodiments, the ground truth of the segmented vascular surface of the coronary arteries can be obtained in a semi-automatic or manual manner, such as by experienced medical personnel through manual annotation. In some embodiments, the ground truth of the segmented vascular surface can be represented by a vascular segmentation mask, such as a pixel-level binarized mask, a pixel-level grayscale mask, or a mask represented in other ways, which are not limited in this application.

[0033] In some embodiments, the bifurcation point information in the second training dataset can be the output of the first deep fully convolutional network (not shown) during training. Specifically, the bifurcation point information can be represented by a bifurcation point mask, the coordinates of the bifurcation point, etc., or it can be the output after a certain expansion (e.g., expanding to a sphere of a certain size) or a certain transformation (e.g., obtaining a heat map through Gaussian transformation). When using the bifurcation point information for training the second deep fully convolutional network 200, the bifurcation point information can be concatenated with the corresponding 3D CTA image used for training. As an example, when the pixels in the 3D CTA image are represented by CT values ​​(e.g., 12-bit grayscale), and the bifurcation point information is represented by a pixel-level binary mask, then 1-bit information of the corresponding pixel in the bifurcation point information can be concatenated into the CTA image representation, for example, represented by CT values, etc. Other concatenation methods can also be used, which are not listed one by one.

[0034] In other embodiments, training supervision for each resolution level of the second deep fully convolutional network 200 can also be provided based on the ground truth of coronary artery segmentation. Specifically, the ground truth of coronary artery segmentation, and ground truth of artery segmentation at different resolutions after downsampling the ground truth of the segmentation corresponding to the resolution of each layer, can be used respectively for training supervision of each layer in the second deep fully convolutional network. For example, in Figure 2 In this application, the ground truth values ​​of the original resolution of coronary artery segmentation, the ground truth values ​​of the medium resolution segmentation after a first downsampling of the ground truth values ​​of the segmentation, and the ground truth values ​​of the low resolution segmentation after a second downsampling of the ground truth values ​​of the segmentation can be used as the first training supervision 221 for the high-resolution layer 201, the second training supervision 222 for the medium-resolution layer 202, and the third training supervision 223 for the low-resolution layer 203, respectively. Compared with traditional U-Net networks, which only set the supervision at the highest resolution layer of the reconstruction path during training (i.e., the last layer of the network), this application sets training supervision at each resolution level. In this way, the error gradient information relied upon by all other layers in the network during training is no longer just indirectly transmitted from the last layer, but can also be introduced as an additional training supervision for the segmentation output of the current layer by using the approximate ground truth value of the image at the corresponding resolution of the current layer as a reference. This multi-layer supervision mechanism can ensure that the network converges to a reasonable state more quickly and effectively, making the second deep fully convolutional network 200 more robust, while making full use of the network capacity.

[0035] As mentioned earlier, the first deep fully convolutional network used for bifurcation point information detection can also adopt a similar network structure to the second deep fully convolutional network 200. Furthermore, to achieve fast training and prediction, it can be configured to contain only convolutional layers and no fully connected layers. The first deep fully convolutional network can be trained based on a first training dataset consisting of a 3D CTA image containing coronary vessels and ground truth values ​​of coronary vessel bifurcation point information from the 3D CTA image. The ground truth values ​​of the bifurcation point information of the coronary vessels corresponding to the 3D CTA image used for training can be obtained from the ground truth values ​​of coronary vessel segmentation.

[0036] Figure 3 This diagram illustrates the extraction of ground truth values ​​for blood vessel bifurcation points according to an embodiment of this application. Specifically, the extraction of ground truth values ​​for blood vessel bifurcation points can be achieved through the following steps: First, extract the center line of the ground truth value 301 of the blood vessel segmentation surface represented by the blood vessel segmentation mask, for example... Figure 3The centerline of the coronary artery (left coronary artery) is defined as 302. Then, based on the topology of this centerline, ground truth values ​​of the bifurcation points of the coronary artery (left coronary artery) are determined automatically, semi-automatically, or manually using methods such as image processing and morphological processing (e.g., opening operations). In some embodiments, the ground truth values ​​of the bifurcation points can be represented as bifurcation point masks, bifurcation point coordinates, etc. In other embodiments, the obtained ground truth values ​​of the bifurcation points represented by masks or coordinates can be expanded (e.g., expanded to a sphere of a certain size) or transformed (e.g., obtained by Gaussian transformation to obtain a heatmap) before being used for training the first deep fully convolutional network.

[0037] In addition, with Figure 2 Similar to the second deep fully convolutional network 200, the first deep fully convolutional network can also contain multiple layers with different resolutions, and includes abstract paths and restoration paths. At the same time, it can also perform multi-layer supervised training on each resolution layer based on the ground truth of the bifurcation point information at the original resolution and the ground truth of the bifurcation point information at the lower resolution after downsampling. Its training method is similar to that of the second deep fully convolutional network 200, and will not be elaborated here.

[0038] According to embodiments of this application, a coronary artery segmentation device is also provided. Figure 4 A schematic diagram showing partial components of a coronary artery segmentation device according to an embodiment of this application is provided. Figure 4 As shown, the coronary artery segmentation device 400 may include an interface 401 and a processor 402. In some embodiments, the interface 401 may acquire a three-dimensional CTA image containing coronary arteries from a database or from a CTA device, for example, through any suitable form such as a network interface or a hardware interface with a CTA device. The three-dimensional CTA image containing coronary arteries may be an image of the coronary artery to be segmented during the prediction phase, or it may be a training image used for training the bifurcation point detection network and / or the vessel segmentation network.

[0039] like Figure 4 As shown, the processor 402 can, for example, execute the steps of the coronary artery segmentation method according to various embodiments of this application. By way of example only, the processor 402 may include, for example, a bifurcation point detection module 4021 and a vessel segmentation module 4022. The bifurcation point detection module 4021 can obtain bifurcation point information of the coronary arteries based on a three-dimensional CTA image acquired via interface 401; the vessel segmentation module 4022 can output a coronary artery segmentation result based on both the three-dimensional CTA image acquired via interface 401 and the bifurcation point information of the coronary arteries corresponding to the image output by the bifurcation point detection module 4021.

[0040] According to embodiments of this application, a non-transitory computer-readable medium is also provided, having instructions stored thereon, wherein, when executed by a processor, the instructions perform the steps of the coronary artery segmentation method according to various embodiments of this application. In some embodiments, the computer-readable medium may include volatile or non-volatile, magnetic, semiconductor, magnetic tape, optical, removable, non-removable, or other types of computer-readable media or computer-readable storage devices. For example, the computer-readable medium may be a storage device or memory module having computer instructions stored thereon. In some embodiments, the computer-readable medium may be a disk or flash drive having computer instructions stored thereon.

[0041] Furthermore, although exemplary embodiments have been described herein, their scope includes any and all embodiments based on this disclosure that have equivalent elements, modifications, omissions, combinations (e.g., schemes involving intersections of various embodiments), adaptations, or alterations. Elements in the claims will be interpreted broadly based on the language used in the claims and are not limited to the examples described in this specification or during the implementation of this application, and such examples will be interpreted as non-exclusive. Therefore, this specification and examples are intended to be considered illustrative only, and the true scope and spirit are indicated by the following claims and the full scope of their equivalents.

[0042] The above description is intended to be illustrative and not restrictive. For example, the above examples (or one or more of them) can be used in combination with each other. Other embodiments may be used by those skilled in the art upon reading the above description. Furthermore, in the above detailed description, various features may be grouped together to simplify the application. This should not be construed as an intention that a feature of an unclaimed application is necessary for any claim. Rather, the subject matter of this application may be less than all the features of a particular disclosed embodiment. Thus, the following claims are incorporated herein by reference as examples or embodiments, wherein each claim is an independent, separate embodiment, and these embodiments are contemplated as being possible in various combinations or arrangements. The scope of the invention should be determined by reference to the appended claims and the full scope of their equivalents.

[0043] The above embodiments are merely exemplary embodiments of this application and are not intended to limit the present invention. The scope of protection of the present invention is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to the present invention within the spirit and scope of this disclosure, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of the present invention.

Claims

1. A method of segmenting a coronary vessel, characterized by, The method comprises: obtaining a three-dimensional CTA image containing a coronary vessel; inputting the three-dimensional CTA image into a trained bifurcation point detection network to obtain bifurcation point information of the coronary vessel, wherein the bifurcation point information has the same size as the three-dimensional CTA image, the bifurcation point detection network is constructed based on a first deep full convolutional network and does not contain a full connection layer, and the bifurcation point detection network only outputs bifurcation point information in the coronary vessel anatomical structure, but does not output bifurcation point information in other vessels or tissues other than the coronary vessel; inputting the three-dimensional CTA image and the bifurcation point information into a trained blood vessel segmentation network after splicing to determine a segmentation result of the coronary vessel, wherein the bifurcation point information is used to strengthen the positioning of the coronary vessel and guide the segmentation algorithm to extract the correct coronary vessel tree topology structure. The blood vessel segmentation network is constructed based on a second deep full convolutional network; the second deep full convolutional network does not contain a full connection layer and is trained based on a second training data set composed of a three-dimensional CTA image containing a coronary vessel, a blood vessel segmentation ground truth of the coronary vessel, and bifurcation point information of the coronary vessel, wherein the bifurcation point information of the coronary vessel is the output of the first deep full convolutional network in the training process.

2. The segmentation method of claim 1, wherein, The bifurcation point information includes at least one of a bifurcation point mask, coordinates of a bifurcation point, a sphere obtained based on a bifurcation point expansion, and a hotspot map obtained based on a bifurcation point transformation.

3. The segmentation method of claim 1, wherein, The blood vessel segmentation ground truth of the coronary vessel is represented by a blood vessel segmentation mask.

4. The segmentation method of claim 3, wherein, The first deep full convolutional network is trained based on a first training data set composed of a three-dimensional CTA image containing a coronary vessel and bifurcation point information ground truth of the coronary vessel.

5. The segmentation method of claim 4, wherein, The method further comprises: extracting a centerline of the blood vessel segmentation mask, and determining a bifurcation point information ground truth of the coronary vessel in an automatic, semi-automatic or manual labeling manner based on a topology structure of the centerline.

6. The segmentation method of claim 1, wherein, The second deep full convolutional network contains multiple layers, each layer of the multiple layers has a different resolution, and the training based on the second training data set composed of a three-dimensional CTA image containing a coronary vessel, a blood vessel segmentation ground truth of the coronary vessel, and bifurcation point information of the coronary vessel specifically comprises: after splicing the three-dimensional CTA image and the bifurcation point information, inputting the spliced image as an input of the second deep full convolutional network; using the blood vessel segmentation ground truth of the coronary vessel and different resolution blood vessel segmentation ground truths obtained by downsampling the blood vessel segmentation ground truth corresponding to the resolution of each layer as training supervision of each layer in the second deep full convolutional network.

7. A coronary vessel segmentation apparatus, characterized by, The method comprises: an interface configured to obtain a three-dimensional CTA image containing a coronary vessel; and a processor configured to perform the coronary vessel segmentation method of any one of claims 1-6.

8. A non-transitory computer readable medium having instructions stored thereon, wherein when executed by a processor, the instructions perform the steps of the coronary vessel segmentation method of any one of claims 1-6.

Images