Coronary artery image segmentation method and device, storage medium and terminal

Abstract

The present invention discloses a coronary artery image segmentation method and device, storage medium, and terminal. The method comprises obtaining a coronary artery image; obtaining a cardiac sub-organ mask group, a coronary artery coarse segmentation result, and a cropped coronary artery original image result; then obtaining a coronary artery point set and a cardiac sub-organ point set group; obtaining a distance field feature map corresponding to each cardiac sub-organ point set, and then concatenating all the distance field feature maps to obtain a five-channel distance field feature map; obtaining a coronary artery merged map; inputting the coronary artery merged map into a trained hierarchical topology learning model to obtain a cube connectivity map, and obtaining a binary coronary artery mask based on the cube connectivity map. The method of the present invention can reduce the interference of background information on the segmentation target, improve segmentation performance, and accurately achieve coronary artery segmentation.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data analysis, and in particular to a coronary artery image segmentation method and device, a storage medium and a terminal. Background Art

[0002] The diagnosis of coronary artery disease (CAD) relies on accurate segmentation of the coronary arteries from coronary computed tomography angiography (CCTA). Therefore, measuring the diameter of the narrowed arterial lumen is an important criterion for grading and quantifying coronary stenosis in clinical practice. However, coronary artery images still face several challenges, making accurate segmentation of coronary arteries from CCTA images a challenging task. These challenges include: First, the small volume ratio of the coronary arteries in the angiographic images makes it difficult to balance the image background and the coronary arteries; second, the coronary arteries are thin structures with diameters of 1mm-5mm, resulting in many difficult-to-segment regions and discontinuous segmentations; and finally, the patient-specific spatial distribution with various shapes and positions hinders the network from accurately capturing the vascular anatomy.

[0003] Currently, medical image segmentation networks all use the original U-shaped structure and its variants. While the network input size is fixed, multiple downsampling operations are used as the network depth increases to reduce the computational effort. In CCTA coronary images, background information occupies a much larger number of voxels than the target vessels, with the number of vascular voxels accounting for only one thousandth of the total image voxels. Therefore, directly using the classic U-shaped structure network to segment coronary arteries will lose most of the detailed information during the multiple downsampling processes. It is also difficult to fully complete the information through skip connections and upsampling operations during the decoding stage. As a result, segmentation of small vascular branches is often broken, incomplete, or lost.

[0004] However, various existing deep learning-based coronary artery segmentation methods only perform partial topological constraints and therefore still cannot fully describe the entire topological structure of the coronary arteries. Summary of the Invention

[0005] The technical problem to be solved by the present invention is that the existing coronary artery segmentation often causes the phenomenon of segmentation of small vascular branches being broken, incomplete, or lost, and the entire topological structure of the coronary artery cannot be obtained.

[0006] In order to solve the above technical problems, the present invention provides a coronary artery image segmentation method, comprising:

[0007] Acquire coronary artery vascular images;

[0008] Segmenting the coronary artery image using a first U-shaped residual network to obtain a cardiac sub-organ mask group, wherein the cardiac sub-organ mask group includes a left atrium mask, a right atrium mask, a left ventricle mask, a right ventricle mask, and an aorta mask;

[0009] Segmenting the coronary artery image using a second U-shaped residual network to obtain a coronary artery rough segmentation result, and cropping the coronary artery image based on the coronary artery rough segmentation result to obtain a coronary artery original image result;

[0010] performing three-dimensional coordinate sampling on the coronary vessel rough segmentation result to obtain a coronary vessel point set, and performing three-dimensional coordinate sampling on all cardiac sub-organ masks in the cardiac sub-organ mask group to obtain a cardiac sub-organ point set group;

[0011] Calculating the minimum distance from each sampling point in the coronary vessel point set to each cardiac sub-organ point set in the cardiac sub-organ point set group to obtain a distance field feature map corresponding to each cardiac sub-organ point set, and then concatenating all the distance field feature maps to obtain a five-channel distance field feature map;

[0012] Merging the coronary artery original image result and the five-channel distance field feature map to obtain a coronary artery merged image;

[0013] The coronary vessel merged map is input into a trained hierarchical topology learning model to obtain a cube connectivity map, and a binary coronary vessel mask is obtained based on the cube connectivity map.

[0014] Preferably, the trained hierarchical topology learning model includes a common encoder and a first decoder, a second decoder and a third decoder respectively connected to the common encoder;

[0015] The public encoder is used to extract features from the combined coronary artery image to obtain coronary artery features;

[0016] The first decoder is configured to identify bifurcation points and short points of the coronary arteries in the coronary artery vascular features to obtain key point topology information;

[0017] The second decoder is configured to identify a centerline of a coronary artery in the coronary artery vascular feature to obtain centerline topology information;

[0018] The third decoder is used to preset connectivity relationships between adjacent vascular points in the coronary artery vascular features to obtain connectivity topology information.

[0019] Preferably, the loss function in the training process of the hierarchical topology learning model is expressed as:

[0020] S=λA+βB+C

[0021] Wherein, S represents the loss function, A represents the key point loss value of the first decoder during the training process, A represents the centerline loss value of the second decoder during the training process, C represents the cube connectivity loss value of the third decoder during the training process, λ represents the key point loss value weight, and β represents the centerline loss value weight.

[0022] Preferably, when the first decoder is trained, the Gaussian heat map of the key points is used as a label for training; and the Gaussian heat map of the key points is generated by a key point filtering unit.

[0023] Preferably, when training, the second decoder uses a centerline heat map as a label for training; and the centerline heat map is generated by a 3D adaptive Gaussian filter.

[0024] Preferably, when training the second U-shaped residual network, the expanded coronary artery mask is used as a label for training.

[0025] Preferably, obtaining a binary coronary vessel mask based on the cube connectivity graph comprises:

[0026] Based on the cube connectivity graph, a binary coronary vessel mask is obtained by using the maximum voting principle.

[0027] In order to solve the above technical problems, the present invention provides a coronary artery image segmentation device, which is characterized by comprising a blood vessel image acquisition module, a heart mask acquisition module, a blood vessel coarse segmentation module, a point set acquisition module, a distance field feature map acquisition module, an image merging module and a coronary artery mask acquisition module;

[0028] The blood vessel image acquisition module is used to acquire coronary artery blood vessel images;

[0029] The heart mask acquisition module is configured to segment the coronary artery image using a first U-shaped residual network to acquire a heart sub-organ mask group, wherein the heart sub-organ mask group includes a left atrium mask, a right atrium mask, a left ventricle mask, a right ventricle mask, and an aorta mask;

[0030] The coarse vessel segmentation module is configured to segment the coronary artery image using a second U-shaped residual network to obtain a coronary artery coarse segmentation result, and to crop the coronary artery image based on the coronary artery coarse segmentation result to obtain a coronary artery original image result;

[0031] The point set acquisition module is configured to acquire a coronary vessel point set by performing three-dimensional coordinate sampling on the coronary vessel coarse segmentation result, and acquire a cardiac sub-organ point set group by performing three-dimensional coordinate sampling on all cardiac sub-organ masks in the cardiac sub-organ mask group respectively;

[0032] The distance field feature map acquisition module is used to calculate the minimum distance from each sampling point in the coronary vessel point set to each cardiac sub-organ point set in the cardiac sub-organ point set group to obtain a distance field feature map corresponding to each cardiac sub-organ point set, and then connect all the distance field feature maps in series to obtain a five-channel distance field feature map;

[0033] The image merging module is configured to merge the original coronary artery image result and the five-channel distance field feature map to obtain a merged coronary artery image;

[0034] The coronary vessel mask acquisition module is used to input the coronary vessel merged map into the trained hierarchical topology learning model to obtain a cube connectivity map, and obtain a binary coronary vessel mask based on the cube connectivity map.

[0035] In order to solve the above technical problem, the present invention provides a storage medium having a computer program stored thereon, which implements a coronary artery image segmentation method when executed by a processor.

[0036] In order to solve the above technical problem, the present invention provides a terminal, comprising: a processor and a memory, wherein the memory is communicatively connected to the processor;

[0037] The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the terminal performs, for example, a coronary artery image segmentation method.

[0038] Compared with the prior art, one or more embodiments of the above solutions may have the following advantages or beneficial effects:

[0039] The coronary artery image segmentation method provided by an embodiment of the present invention is applied to obtain a coarse segmentation result by performing a coarse segmentation on the coronary arteries, and to capture the spatial relationship between the coronary arteries and various sub-components of the heart through a distance field feature map to reduce the imbalance between the blood vessels and the background. At the same time, the distance field feature map can also be considered as an anatomical dependency. A hierarchical topological learning model is then designed to improve the coarse segmentation result by learning three levels of structural representation of the coronary arteries. That is, by adding key points, centerlines, and connectivity prediction branches, the problems of discontinuity and patient-specific distribution differences in small blood vessel branches can be significantly alleviated. The method of the present invention can reduce the interference of background information on the segmentation target, improve segmentation performance, and accurately achieve coronary artery segmentation.

[0040] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purposes and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0042] Figure 1 FIG2 shows a flow chart of a method for segmenting a coronary artery image according to a first embodiment of the present invention;

[0043] Figure 2 FIG2 shows a framework representation of a coronary artery image segmentation method according to a first embodiment of the present invention;

[0044] Figure 3 Schematic diagram showing the results of segmentation of coronary arteries of three typical subjects using five different methods in Example 1 of the present invention;

[0045] Figure 4 The present invention shows a coronary artery image segmentation device according to a second embodiment of the present invention;

[0046] Figure 5 It shows a schematic structural diagram of a terminal according to a fourth embodiment of the present invention. DETAILED DESCRIPTION

[0047] The following describes the embodiments of the present invention in detail with reference to the accompanying drawings and examples, so that the present invention can fully understand how to apply technical means to solve technical problems and achieve technical effects, and thus implement the invention accordingly. It should be noted that, as long as no conflict exists, the various embodiments of the present invention and the various features of the embodiments can be combined with each other, and the resulting technical solutions are all within the scope of protection of the present invention.

[0048] Coronary Artery Disease (CAD) is one of the most common cardiovascular diseases, accounting for one-third of all cardiovascular diseases, and is a very large group of people. The cause of coronary artery disease is usually due to the formation of plaques in the surrounding tissue lesions, which causes stenosis or blockage of the blood vessel lumen, resulting in insufficient blood supply to the coronary arteries and inducing the disease. Coronary Computed Tomography Angiography (CCTA) is a non-invasive imaging technology that is commonly used to show the location of coronary artery lesions. Accurate diagnosis of CAD depends on accurate segmentation of the coronary arteries from CCTA images, so measuring the diameter of the narrowed arterial lumen is an important criterion for grading and quantifying coronary stenosis in clinical practice.

[0049] Various deep learning-based methods have been proposed for medical image segmentation and have demonstrated excellent performance. General models such as UNet and its variants fail to achieve good performance in coronary artery segmentation due to the extremely complex anatomy and patient-specific spatial distribution of coronary arteries. Several existing studies have attempted to address these issues by designing customized networks. For example, a modified PSPNet with a pyramid pooling module was proposed to learn multi-scale global information, thereby mitigating the negative impact of small volume ratios. Furthermore, attention-based methods have been proposed to learn finer-grained feature representations for coronary artery segmentation. However, these methods ignore important coronary artery anatomy and topology, resulting in limited segmentation performance. Other methods employ partial topological representations as loss constraints for training. A contour-constrained weighted Hausdorff distance loss was proposed, which focuses on learning boundary continuity and connectivity. Recently, Shit et al. introduced a similarity measure called centerline Dice, which achieves good segmentation performance by constraining the intersection of the segmentation mask and the skeleton. However, the topological constraints in these methods are only partial, not complete, and fail to fully describe the entire topological structure of the coronary arteries.

[0050] Example 1

[0051] To solve the technical problems existing in the prior art, an embodiment of the present invention provides a coronary artery image segmentation method.

[0052] Figure 1 FIG2 shows a flow chart of a method for segmenting a coronary artery image according to a first embodiment of the present invention; Figure 2 FIG1 shows a schematic diagram of a coronary artery image segmentation method according to an embodiment of the present invention; FIG2 shows a schematic diagram of a coronary artery image segmentation method according to an embodiment of the present invention; Figure 1 and Figure 2As shown, the coronary artery image segmentation method according to the embodiment of the present invention includes the following steps.

[0053] Step S101: Acquire a coronary artery image.

[0054] Specifically, the image to be segmented for the coronary artery is used as the coronary artery image. The coronary artery image may be a coronary artery computed tomography angiography image or other suitable coronary artery image. The form of expression of the coronary artery image is not specifically limited herein.

[0055] Step S102 : segmenting the coronary artery image using a first U-shaped residual network to obtain a heart sub-organ mask group.

[0056] Specifically, the first U-shaped residual network is used to segment the various organs of the heart in the coronary artery image to obtain a left atrium mask, a right atrium mask, a left ventricle mask, a right ventricle mask, and an aorta mask. The left atrium mask, the right atrium mask, the left ventricle mask, the right ventricle mask, and the aorta mask form a cardiac sub-organ mask group. Preferably, the first U-shaped residual network is a ResUNet network structure.

[0057] Step S103 , segmenting the coronary artery image using a second U-shaped residual network to obtain a coronary artery rough segmentation result, and cropping the coronary artery image based on the coronary artery rough segmentation result to obtain a coronary artery original image result.

[0058] Specifically, the coronary arteries in the coronary artery image are segmented by a second U-shaped residual network to obtain a coarse segmentation result of the coronary artery. Preferably, the second U-shaped residual network is a ResUNet network structure. It should be noted that since the heart is a large organ, satisfactory results can be achieved by segmenting it using the first U-shaped residual network. However, due to the delicate structure of the coronary arteries, it is difficult to obtain good segmentation results directly using the U-shaped residual network, and there will be a large number of under-segmentation phenomena. Therefore, in order to obtain all voxel points of the coronary arteries when calculating the distance field, when training the second U-shaped residual network, the labeled coronary artery mask needs to be expanded, and the expanded labeled coronary artery mask is used as the supervision signal of the second U-shaped residual network. The second U-shaped residual network can output a coarse segmentation result. Preferably, the expansion kernel can be set to (5,5,5).

[0059] Step S104 , performing three-dimensional coordinate sampling on the coronary vessel rough segmentation result to obtain a coronary vessel point set, and performing three-dimensional coordinate sampling on all cardiac sub-organ masks in the cardiac sub-organ mask group to obtain a cardiac sub-organ point set group.

[0060] Specifically, all coronary vessels in the coronary vessel rough segmentation result are fully sampled in three-dimensional coordinates to obtain a coronary vessel point set P = [p1, p2, ····, p N ]. A voxel point on the coronary artery is a point in the three-dimensional coordinates. The dimension of the coronary artery point set is N×3, where n∈[13000,20000]. The number of vascular voxel points varies between individuals, so the number of n is not fixed. Similarly, the left atrium mask, right atrium mask, left ventricle mask, right ventricle mask and aorta mask in the cardiac sub-organ mask group are sampled in three-dimensional coordinates to obtain the left atrium point set, right atrium point set, left ventricle point set, right ventricle point set and aorta point set. The left atrium point set, right atrium point set, left ventricle point set, right ventricle point set and aorta point set form the cardiac sub-organ point set group. It should be noted that in order to reduce the computational cost, all cardiac sub-organ masks in the cardiac sub-organ mask group are sampled using a three-dimensional coordinate downsampling method. For example, each point set of the heart sub-organ mask may be uniformly downsampled to 8000 points, and the heart sub-organ point set group may be recorded as {O1, O2, . . . , O5}.

[0061] Step S105 , calculating the minimum distance from each sampling point in the coronary vessel point set to each cardiac sub-organ point set in the cardiac sub-organ point set group to obtain a distance field feature map corresponding to each cardiac sub-organ point set, and then concatenating all the distance field feature maps to obtain a five-channel distance field feature map.

[0062] Specifically, the minimum distance between each sampling point in the coronary vessel point set and the left atrium point set is calculated by the following formula (1) to obtain the distance field feature map corresponding to the left atrium.

[0063]

[0064] Among them, S n,m Indicates the nth point p in the coronary artery point set n To any heart sub-organ point o m The shortest distance, where O M represents any cardiac sub-organ point set, and (x, y, z) represents the three-dimensional coordinates of the points in the point set. Similarly, the distance field feature map corresponding to the right atrium, the distance field feature map corresponding to the left ventricle, the distance field feature map corresponding to the right ventricle, and the distance field feature map corresponding to the aorta are calculated by equation (1).

[0065] The five distance field feature maps are then normalized using the method shown in equation (2) to obtain five normalized distance field feature maps. Normalizing the distance field feature maps can reduce the interference of outliers in the distance field feature maps. Finally, the five normalized distance field feature maps are connected in series to form a five-channel distance field feature map, that is, a five-channel distance field feature map Fa = {(dn,1, dn,2····, dn,5)}.

[0066]

[0067] Among them, d n,m For S n,m The normalized representation of τ is an empirical value that can be set to 50.

[0068] Step S106 , merging the coronary artery original image result and the five-channel distance field feature map to obtain a coronary artery merged image.

[0069] Specifically, the original coronary image results and the five-channel distance feature field map are connected in series to form a six-channel merged map. The six-channel merged map is the coronary artery merged map, which is used as input into the subsequent hierarchical topology learning model.

[0070] Step S107 : inputting the coronary vessel merged map into the trained hierarchical topology learning model to obtain a cube connectivity map, and obtaining a binary coronary vessel mask based on the cube connectivity map.

[0071] Specifically, the coronary artery merged map is input into the trained hierarchical topological learning model to obtain 27-channel results corresponding to each coronary vessel point in the coronary artery merged map, that is, each vessel point has 27 prediction results. Then, the maximum voting principle is used to fuse the 27 results of each vessel point into a binary image to obtain the binary segmentation mask of the coronary artery.

[0072] Since the final desired outcome is a binary coronary artery mask, the cube connectivity graph output by the trained hierarchical topology learning model must be fused to obtain the final binary vascular mask. Specifically, following the inverse process of cube connectivity construction, the network outputs 27 channels, generating 27 predictions at each voxel. The maximum voting principle is then used to select the best result to obtain the final binary coronary artery mask.

[0073] The specific hierarchical topology learning model includes a common encoder and a first decoder, a second decoder, and a third decoder respectively connected to the common encoder; the common encoder is mainly used to extract features from the coronary artery vascular merge map to obtain coronary artery vascular features; the coronary artery vascular features obtained during the training phase are transmitted to the first decoder, the second decoder, and the third decoder respectively to learn key point topology information, centerline topology information, and connectivity topology information.

[0074] The first decoder is mainly used to identify the bifurcation points and short points of the coronary arteries in the coronary artery vascular features to obtain the key point topological information of the coronary arteries. Furthermore, the bifurcation points and endpoints of each coronary artery branch can provide a strong reference for the vascular tree structure, so a third U-shaped residual network (ResUNet decoder structure) is added as the first decoder to assist in vascular segmentation. In the key point detection task, compared with the case of directly regressing coordinate points, the key point heat map regression focuses on regressing a distribution representation, which is more suitable for the key point detection task and can produce more robust results. Therefore, when the first decoder is trained, the Gaussian heat map of the key point is used as a label for training; and the Gaussian heat map of the key point is generated using the key point filter unit.

[0075] The second decoder is mainly used to identify the centerline of the coronary artery in the coronary artery vascular features to obtain the centerline topology information of the coronary artery. Furthermore, since the voxel points on the centerline only cover some discrete voxel points on each blood vessel cross section, if the positions of these central voxels are directly predicted, a small position deviation will often lead to a huge error in the centerline prediction result. Therefore, the fourth U-shaped residual network (ResUNet decoder structure) is used as the second decoder to obtain the centerline heat map regression instead of directly predicting the voxel points on the centerline. Therefore, when the second decoder is trained, the centerline heat map is used as the label for training; and the centerline heat map is generated by a 3D adaptive Gaussian filter. The use of centerline heat map regression can pay more attention to the topological connection of the coronary artery and prevent the phenomenon of local blood vessel breakage during the segmentation process. The 3D adaptive Gaussian filter generates a local normalized centerline heat map regression process as shown in Equations (3) and (4).

[0076]

[0077]

[0078] Among them, σ trepresents the standard deviation of the coronary artery mask in three dimensions, which is proportional to the width, height, and depth of a single coronary artery. The design of formula (4) can ensure that the centerline heat map of thick vessels has a smooth value change, and can also significantly enhance the centerline heat map of thin vessels. t represents the three-dimensional coordinates of the t-th centerline point of the coronary artery, P t l Represents the three-dimensional coordinates of the lth blood vessel point around the tth center line point. There are L points in total, and l represents the blood vessel point around the center line point.

[0079] The third decoder is mainly used to preset the connectivity relationship between adjacent vascular points in the coronary artery features to obtain the cube connectivity map of the coronary artery. Furthermore, the cube connectivity is defined as: between any vascular point P on the coronary artery and its 26 nearest adjacent points Q, if the vascular point P and the connected points Q i (i=1、2……26) are all points on the coronary artery, then define (P,Q i ) is connected, at this time the label on the i-th channel of the blood vessel point P is marked as 1, otherwise the label on the i-th channel of the blood vessel point P is marked as 0, indicating that the blood vessel points P and Q i The connections between them are disconnected, and the 27-channel result corresponding to the vessel point P is obtained. The design of cube connectivity can encode the topological relationship of all coronary artery voxels, so that all vascular voxel points in the image are connected in one space, while all background voxels are disconnected, thereby reducing background interference during network learning.

[0080] It should be noted that during the training phase, the coronary artery features obtained by the public encoder of the hierarchical topology learning model are then transmitted to the first decoder, the second decoder, and the third decoder to learn keypoint topology information, centerline topology information, and connectivity topology information, respectively. However, during the actual measurement phase, after the public encoder of the trained hierarchical topology learning model encodes the coronary artery features, it transmits them to the third decoder to obtain the cube connectivity map.

[0081] And no matter in the training process or in the actual measurement stage, the public encoder encoding and the third decoder both work in the above working manner.

[0082] Furthermore, the loss function in the training process of the hierarchical topology learning model is expressed as:

[0083] S=λA+βB+C

[0084] Where S represents the loss function, A represents the keypoint loss value of the first decoder during training, A represents the centerline loss value of the second decoder during training, C represents the cube connectivity loss value of the third decoder during training, λ represents the keypoint loss weight, and β represents the centerline loss weight. The keypoint loss value of the first decoder and the centerline loss value of the second decoder are both obtained using the mean squared error loss function, while the cube connectivity loss value of the third decoder is obtained using the average Dice loss function. Both λ and β are preferably 0.2.

[0085] To illustrate the accuracy of the coronary artery image segmentation method of the present invention, this example compares the method of the present invention with the specific implementation data of existing deep learning-based blood vessel segmentation methods, including ResUNet, improved PSPNet, CS2-Net, and clDice. The quantitative results are shown in Table 1, and the qualitative results are shown in Table 2. Figure 3 shown.

[0086]

[0087] Table 1 summarizes the indicator results of five different methods in the coronary artery segmentation task, and it can be seen that the method proposed in this paper significantly outperforms the methods proposed in other literatures. Specifically, compared with the baseline ResUNet, our method has achieved significant improvements in most indicators. At the same time, our method achieves better performance than clDice (for example, increasing the segmentation accuracy (Dice) from 76.72% to 80.36%). It is worth noting that all these compared methods pay little attention to the dependence of coronary vessels on the heart and the coronary artery topology, which often leads to false positives and vascular discontinuities in the segmentation results. These results fully demonstrate the effectiveness of the framework proposed in this paper.

[0088] Figure 3 Schematic diagram showing the results of segmenting the coronary arteries of three typical subjects using five different coronary artery segmentation methods in Example 1 of the present invention; Figure 3 As shown in Figure 2, for simpler segmentation cases (such as the first row), the five methods can produce satisfactory results on the overall structure of the blood vessels, but it is difficult for these contrasting methods to maintain fine details in local areas (such as bifurcations) that are difficult to segment. Figure 3The area in the middle circle. For the more difficult segmentation cases (the last two rows), the visual display results show that our method is significantly better than other comparison methods. This is because most of the background of the original data is eliminated in the spatial anatomical dependency module, and background mis-segmentation rarely appears in our results. In addition, through hierarchical topological constraints, our results maintain good topological continuity in each coronary artery branch, which is different from other methods with a large number of discontinuities. Qualitative results further verify the effectiveness of the coronary artery segmentation method proposed in this paper

[0089] The coronary artery image segmentation method provided by an embodiment of the present invention obtains a coarse segmentation result by performing a coarse segmentation on the coronary vessels, and captures the spatial relationship between the coronary vessels and various sub-components of the heart through a distance field feature map to reduce the imbalance between the vessels and the background. The distance field feature map can also be considered as an anatomical dependency. A hierarchical topological learning model is then designed to improve the coarse segmentation result by learning three levels of structural representation of the coronary vessels. Specifically, by adding key points, centerlines, and connectivity prediction branches, the discontinuity and patient-specific distribution differences in small vascular branches can be significantly alleviated. The method of the present invention can reduce the interference of background information on the segmentation target, improve segmentation performance, and accurately achieve coronary artery segmentation.

[0090] Example 2

[0091] In order to solve the above technical problems existing in the prior art, an embodiment of the present invention further provides a coronary artery image segmentation device.

[0092] Figure 4 The coronary artery image segmentation device of the second embodiment of the present invention is shown; Figure 4 As shown, the coronary artery image segmentation device according to an embodiment of the present invention includes a blood vessel image acquisition module, a heart mask acquisition module, a blood vessel coarse segmentation module, a point set acquisition module, a distance field feature map acquisition module, an image merging module and a coronary artery mask acquisition module.

[0093] The blood vessel image acquisition module is used to acquire coronary artery blood vessel images.

[0094] The cardiac mask acquisition module is used to segment the coronary artery image through a first residual U-shaped network to obtain a cardiac sub-organ mask group, which includes a left atrium mask, a right atrium mask, a left ventricle mask, a right ventricle mask and an aorta mask.

[0095] The coarse vessel segmentation module is used to segment the coronary artery image through the second residual U-shaped network to obtain the coronary artery coarse segmentation result, and to crop the coronary artery image based on the coronary artery coarse segmentation result to obtain the coronary artery original image result.

[0096] The point set acquisition module is used to obtain a coronary vessel point set by performing three-dimensional coordinate sampling on the coronary vessel rough segmentation result, and to obtain a cardiac sub-organ point set group by performing three-dimensional coordinate sampling on all cardiac sub-organ masks in the cardiac sub-organ mask group.

[0097] The distance field feature map acquisition module is used to calculate the minimum distance from each sampling point in the coronary artery point set to each cardiac sub-organ point set in the cardiac sub-organ point set group to obtain the distance field feature map corresponding to each cardiac sub-organ point set, and then concatenate all the distance field feature maps to obtain a five-channel distance field feature map.

[0098] The image merging module is used to merge the original coronary artery image results and the five-channel distance field feature map to obtain a coronary artery merged image.

[0099] The coronary vessel mask acquisition module is used to input the coronary vessel merged map into the trained hierarchical topology learning model to obtain a cube connectivity map, and obtain a binary coronary vessel mask based on the cube connectivity map.

[0100] The coronary artery image segmentation device provided by an embodiment of the present invention obtains a coarse segmentation result by performing a coarse segmentation on the coronary arteries, and captures the spatial relationship between the coronary arteries and various sub-components of the heart through a distance field feature map to reduce the imbalance between the vessels and the background. The distance field feature map can also be considered as an anatomical dependency. A hierarchical topological learning model is then designed to improve the coarse segmentation result by learning three levels of structural representation of the coronary arteries. Specifically, by adding key points, centerlines, and connectivity prediction branches, the discontinuity and patient-specific distribution differences in small vascular branches can be significantly alleviated. The device of the present invention can reduce the interference of background information on the segmentation target, improve segmentation performance, and accurately achieve coronary artery segmentation.

[0101] Example 3

[0102] To solve the above technical problems existing in the prior art, an embodiment of the present invention further provides a storage medium storing a computer program. When the computer program is executed by a processor, all steps in the coronary artery image segmentation method in the first embodiment can be implemented.

[0103] The specific steps of the coronary artery image segmentation method and the beneficial effects obtained by applying the readable storage medium provided by the embodiment of the present invention are the same as those in the first embodiment and will not be described in detail here.

[0104] It should be noted that the storage medium includes various media that can store program codes, such as ROM, RAM, magnetic disk or optical disk.

[0105] Example 4

[0106] In order to solve the above technical problems existing in the prior art, an embodiment of the present invention further provides a terminal.

[0107] Figure 5 The schematic diagram of the terminal structure of the fourth embodiment of the present invention is shown. Figure 5 In this embodiment, the terminal includes a processor and a memory connected to each other; the memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory, so that when the terminal executes the computer programs, all steps in the coronary artery image segmentation method of embodiment 1 can be implemented.

[0108] The specific steps of the coronary artery image segmentation method and the beneficial effects obtained by applying the terminal provided by the embodiment of the present invention are the same as those in the first embodiment, and will not be described in detail here.

[0109] It should be noted that the memory may include random access memory (RAM) and may also include non-volatile memory (non-volatile memory), such as at least one disk storage. Similarly, the processor may also be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0110] Although the embodiments disclosed herein are as described above, the contents described herein are merely embodiments for facilitating understanding of the present invention and are not intended to limit the present invention. Any person skilled in the art may make any modifications and variations in the form and details of the embodiments without departing from the spirit and scope of the present invention. However, the scope of protection of the present invention shall remain subject to the scope defined by the appended claims.

Claims

1. A coronary artery image segmentation method, comprising: Acquire coronary artery vascular images; Segmenting the coronary artery image using a first U-shaped residual network to obtain a cardiac sub-organ mask group, wherein the cardiac sub-organ mask group includes a left atrium mask, a right atrium mask, a left ventricle mask, a right ventricle mask, and an aorta mask; Segmenting the coronary artery image using a second U-shaped residual network to obtain a coronary artery rough segmentation result, and cropping the coronary artery image based on the coronary artery rough segmentation result to obtain a coronary artery original image result; performing three-dimensional coordinate sampling on the coronary vessel rough segmentation result to obtain a coronary vessel point set, and performing three-dimensional coordinate sampling on all cardiac sub-organ masks in the cardiac sub-organ mask group to obtain a cardiac sub-organ point set group; Calculating the minimum distance from each sampling point in the coronary vessel point set to each cardiac sub-organ point set in the cardiac sub-organ point set group to obtain a distance field feature map corresponding to each cardiac sub-organ point set, and then concatenating all the distance field feature maps to obtain a five-channel distance field feature map; Merging the coronary artery original image result and the five-channel distance field feature map to obtain a coronary artery merged image; The coronary vessel merged map is input into a trained hierarchical topology learning model to obtain a cube connectivity map, and a binary coronary vessel mask is obtained based on the cube connectivity map.

2. The method according to claim 1, characterized in that The trained hierarchical topology learning model includes a common encoder and a first decoder, a second decoder and a third decoder respectively connected to the common encoder; The public encoder is used to extract features from the combined coronary artery image to obtain coronary artery features; The first decoder is configured to identify bifurcation points and short points of the coronary arteries in the coronary artery vascular features to obtain key point topology information; The second decoder is configured to identify a centerline of a coronary artery in the coronary artery vascular feature to obtain centerline topology information; The third decoder is used to preset connectivity relationships between adjacent vascular points in the coronary artery vascular features to obtain connectivity topology information.

3. The method according to claim 2, characterized in that The loss function in the training process of the hierarchical topology learning model is expressed as: S=λA+βB+C Wherein, S represents the loss function, A represents the key point loss value of the first decoder during the training process, A represents the centerline loss value of the second decoder during the training process, C represents the cube connectivity loss value of the third decoder during the training process, λ represents the key point loss value weight, and β represents the centerline loss value weight.

4. The method according to claim 2, characterized in that When the first decoder is trained, the Gaussian heat map of the key point is used as a label for training; and the Gaussian heat map of the key point is generated by a key point filtering unit.

5. The method according to claim 2, characterized in that When training, the second decoder uses the centerline heat map as a label for training; and the centerline heat map is generated by a 3D adaptive Gaussian filter.

6. The method according to claim 1, wherein When training the second U-shaped residual network, the expanded coronary artery mask is used as a label for training.

7. The method according to claim 1, characterized in that Obtaining a binary coronary vessel mask based on the cube connectivity graph includes: Based on the cube connectivity graph, a binary coronary vessel mask is obtained by using the maximum voting principle.

8. A coronary artery image segmentation device, characterized in that: It includes a blood vessel image acquisition module, a heart mask acquisition module, a blood vessel coarse segmentation module, a point set acquisition module, a distance field feature map acquisition module, an image merging module and a coronary vessel mask acquisition module; The blood vessel image acquisition module is used to acquire coronary artery blood vessel images; The heart mask acquisition module is configured to segment the coronary artery image using a first U-shaped residual network to acquire a heart sub-organ mask group, wherein the heart sub-organ mask group includes a left atrium mask, a right atrium mask, a left ventricle mask, a right ventricle mask, and an aorta mask; The coarse vessel segmentation module is configured to segment the coronary artery image using a second U-shaped residual network to obtain a coronary artery coarse segmentation result, and to crop the coronary artery image based on the coronary artery coarse segmentation result to obtain a coronary artery original image result; The point set acquisition module is configured to acquire a coronary vessel point set by performing three-dimensional coordinate sampling on the coronary vessel coarse segmentation result, and acquire a cardiac sub-organ point set group by performing three-dimensional coordinate sampling on all cardiac sub-organ masks in the cardiac sub-organ mask group respectively; The distance field feature map acquisition module is used to calculate the minimum distance from each sampling point in the coronary vessel point set to each cardiac sub-organ point set in the cardiac sub-organ point set group to obtain a distance field feature map corresponding to each cardiac sub-organ point set, and then connect all the distance field feature maps in series to obtain a five-channel distance field feature map; The image merging module is configured to merge the original coronary artery image result and the five-channel distance field feature map to obtain a merged coronary artery image; The coronary vessel mask acquisition module is used to input the coronary vessel merged map into the trained hierarchical topology learning model to obtain a cube connectivity map, and obtain a binary coronary vessel mask based on the cube connectivity map.

9. A storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the coronary artery image segmentation method according to any one of claims 1 to 7 is implemented.

10. A terminal, characterized in that: include: a processor and a memory, wherein the memory is communicatively connected to the processor; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the terminal performs the coronary artery image segmentation method according to any one of claims 1 to 7.

Images