Coronary centerline segmentation and naming method and device based on contrast image segmentation

Abstract

The application provides a coronary centerline segmentation naming method and device based on contrast image segmentation, and relates to the technical field of medical image processing. The method comprises the following steps: acquiring an angiogram of a coronary vessel; inputting the angiogram into a pre-trained joint segmentation model to output a vessel segmentation result and a segmentation naming result; generating an ordered centerline of the coronary vessel based on the vessel segmentation result and the segmentation naming result; wherein the ordered centerline is a centerline with a clear starting point, ending point and direction for each segment in the skeleton graph of the coronary vessel; and generating a coronary centerline segmentation naming result of the coronary vessel in accordance with anatomical rules based on the ordered centerline and the segmentation naming result. Through the method provided by the application, the joint segmentation model synchronously completes the double tasks, avoids error accumulation, and improves the overall analysis efficiency and accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, and in particular to a coronary centerline segmentation naming method and device based on contrast image segmentation. BACKGROUND

[0002] In clinical practice, doctors usually need to anatomically segment the main branches of the coronary artery (such as the left anterior descending branch, the circumflex branch, the right coronary artery, etc.) and their secondary branches according to the angiogram, and accurately name them according to the standardized naming specification. This process not only directly relates to the accurate positioning of the lesion location, but also is an important basis for developing interventional treatment plans.

[0003] In the prior art, coronary artery analysis usually adopts a step-by-step processing method: first, the vessel contour is extracted by an image segmentation algorithm, and then the skeleton is extracted and the branch is named based on the segmentation result. This method has the following typical schemes: the segmentation model and the naming model are trained independently, which causes the error to be transmitted step by step, and the overall accuracy is limited; the loss weights of segmentation and segmentation tasks are fixed during the training process, which cannot adapt to the needs of different training stages; after the skeleton is extracted, there is a lack of anatomical rule constraints, and noise branches and topological errors affect the generation of the centerline; the integration of segmentation naming and centerline relies on manual correction, which is low in efficiency and easy to introduce subjective bias.

[0004] How to simultaneously complete the dual tasks of segmentation and segmentation, avoid error accumulation, and improve the overall analysis efficiency and accuracy is a technical problem that needs to be solved at present. SUMMARY

[0005] The present application provides a coronary centerline segmentation naming method and device based on contrast image segmentation to solve the defects in the prior art.

[0006] The present application provides a coronary centerline segmentation naming method based on contrast image segmentation, comprising the following steps:

[0007] Obtaining an angiogram image of a coronary artery; wherein the angiogram image contains: image data of at least one target shooting body position;

[0008] Inputting the angiogram image into a pre-trained joint segmentation model to output a vessel segmentation result and a segmentation naming result; wherein the joint segmentation model is obtained by training a multi-task deep learning model based on a coronary artery angiogram image dataset;

[0009] Based on the vessel segmentation result and the segmentation naming result, generating an ordered centerline of the coronary artery; wherein the ordered centerline is a centerline that each segment in the skeleton graph of the coronary artery has a clear starting point, end point and direction;

[0010] Based on the ordered centerline and the segment naming result, a coronary centerline segment naming result of the coronary vessel conforming to anatomical rules is generated.

[0011] According to the coronary centerline segment naming method based on contrast image segmentation provided by the application, the joint segmentation and segmentation model comprises an image encoding layer, a segmentation image decoding layer and a segment image decoding layer, and the image encoding layer, the segmentation image decoding layer and the segment image decoding layer constitute a coronary segmentation task and a coronary segment naming task.

[0012] The method comprises the following steps:

[0013] The method comprises the following steps:

[0014] The coronary segmentation task is used for extracting a filling contour of the coronary artery, and each pixel in the contrast image is assigned to a single category, and the coronary segment naming task is used for anatomical segmentation of the main branch and the secondary branch of the coronary artery on the basis of the coronary artery segmentation, and pixels of different segments are assigned to different categories.

[0015] According to the coronary centerline segment naming method based on contrast image segmentation provided by the application, the training process of the joint segmentation and segmentation model comprises the following steps:

[0016] The deep learning model is trained based on a pre-constructed multi-task learning framework, so that the deep learning model simultaneously learns multiple tasks; each task in the multiple tasks has a corresponding loss function, and a weighted sum of the loss functions of each task in the multiple tasks constitutes a total loss function; the total loss function adjusts the relevance of the segmentation loss and the segmentation loss through a consistency index, and the consistency index is calculated based on the number of overlapping region pixels of a segmentation error map and a segmentation error map.

[0017] The weights of the deep learning model are dynamically adjusted by minimizing the total loss function, and the joint segmentation and segmentation model is obtained.

[0018] According to the coronary centerline segment naming method based on contrast image segmentation provided by the application, the weights of the deep learning model are dynamically adjusted by minimizing the total loss function, and the weights of the deep learning model are dynamically adjusted by minimizing the total loss function, and the weights of the deep learning model are dynamically adjusted by minimizing the total loss function.

[0019] In the early training stage, the segmentation loss weight is higher than the segmentation loss weight.

[0020] In the middle of training, the segmentation loss weight is higher than the segmentation loss weight;

[0021] In the later stage of training, the segmentation loss weight is balanced with the segmentation loss weight.

[0022] According to the coronary centerline segmentation naming method based on the contrast image segmentation provided by the application, the ordered centerline of the coronary artery is generated based on the blood vessel segmentation result and the segmentation naming result, which comprises:

[0023] The blood vessel segmentation result is extracted to obtain the initial skeleton graph of the coronary artery, and the noise branch in the initial skeleton graph is filtered according to the pre-set branch pixel length threshold to obtain the skeleton graph of the coronary artery;

[0024] According to the anatomical rule of coronary artery tree 16 segmentation method, the target segmentation is determined in the skeleton graph of the coronary artery based on the bifurcation point and the end point;

[0025] According to the segmentation naming result, the starting point of the coronary artery is determined;

[0026] Based on the skeleton graph of the coronary artery, the target segmentation and the starting point of the coronary artery, the ordered centerline of the coronary artery is generated.

[0027] According to the coronary centerline segmentation naming method based on the contrast image segmentation provided by the application, the coronary centerline segmentation naming result of the coronary artery is generated based on the ordered centerline and the segmentation naming result, which comprises:

[0028] According to the ordered centerline and the segmentation naming result, the coordinate points on the ordered centerline are sequentially segmented and classified, and the class of each coordinate point is determined;

[0029] Based on the determined class of each coordinate point, each ordered centerline is respectively judged by a dynamic class correction strategy, and the coronary centerline segmentation naming result of the coronary artery which meets the anatomical rule is output.

[0030] According to the coronary centerline segmentation naming method based on the contrast image segmentation provided by the application, the coronary centerline segmentation naming result of the coronary artery which meets the anatomical rule is output based on the determined class of each coordinate point, and the coronary centerline segmentation naming result of the coronary artery which meets the anatomical rule is output.

[0031] In the case that there are multiple segmentation naming results in a segment of ordered centerline, if the pixel proportion of any class in the ordered centerline is lower than the pre-set threshold and the distribution is discontinuous, the class is determined as noise and filtered out;

[0032] In a case where multiple segment naming results exist in an ordered centerline, if a proportion of pixels of any category in the ordered centerline is higher than a preset threshold and is continuously distributed, the category is reserved and is mapped and matched with the segment naming result, and a coronary centerline segment naming result of the coronary vessel in line with an anatomical rule is output.

[0033] The application further provides a coronary centerline segment naming device based on contrast image segmentation, comprising the following modules.

[0034] An acquisition module is configured to acquire a coronary angiography image of a coronary vessel, wherein the coronary angiography image comprises image data of at least one target shooting body position.

[0035] A segment naming module is configured to input the coronary angiography image into a pre-trained joint segmentation and segment model, and output a vessel segmentation result and a segment naming result, wherein the joint segmentation and segment model is obtained by training a multi-task deep learning model based on a coronary angiography image data set.

[0036] A centerline extraction module is configured to generate an ordered centerline of the coronary vessel based on the vessel segmentation result and the segment naming result, wherein the ordered centerline is a centerline having a clear starting point, ending point and direction for each segment in a skeleton graph of the coronary vessel.

[0037] A centerline and segment naming fusion module is configured to generate a coronary centerline segment naming result of the coronary vessel in line with an anatomical rule based on the ordered centerline and the segment naming result.

[0038] The application further provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the coronary centerline segment naming method based on contrast image segmentation according to any one of the above when executing the computer program.

[0039] The application further provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the coronary centerline segment naming method based on contrast image segmentation according to any one of the above.

[0040] The application further provides a computer program product comprising a computer program, wherein the computer program is executable by a processor to implement the coronary centerline segment naming method based on contrast image segmentation according to any one of the above.

[0041] The application provides a coronary center line segmentation naming method and device based on contrast image segmentation. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort based on these drawings.

[0043] Figure 1 FIG. 1 is an application scenario diagram of the coronary center line segmentation naming method based on contrast image segmentation provided by the present application.

[0044] Figure 2 FIG. 2 is a flowchart of the coronary center line segmentation naming method based on contrast image segmentation provided by the present application.

[0045] Figure 3 FIG. 3 is a structural diagram of the coronary center line segmentation naming device based on contrast image segmentation provided by the present application.

[0046] Figure 4 FIG. 4 is a structural diagram of the electronic device provided by the present application. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely in the following with reference to the drawings in the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort belong to the protection scope of the present application.

[0048] The application is described below Figures 1-4 A coronary centerline segmentation and naming method and device based on contrast image segmentation are provided.

[0049] It should be noted that a single complete coronary segmentation and naming by a human takes a long time, and the consistency of interpretation between different doctors is poor. At the same time, complex vascular variations and lesion characteristics are easy to cause misjudgment. The segmentation model and the naming model are independently trained, which causes error to be transmitted step by step, the overall accuracy is limited, and the model is unstable in complex scenes (such as bifurcation and stenosis); after the skeleton is extracted, there is a lack of anatomical rule constraint, and noise branches and topological errors affect the generation of the centerline; the integration of segmentation and naming and the centerline lacks a dynamic correction mechanism, and algorithm noise easily affects the explainability of the final result. Based on this, the application provides a coronary centerline segmentation and naming method based on contrast image segmentation to solve at least one of the above problems.

[0050] Figure 1 An application scenario diagram of the coronary centerline segmentation and naming method based on contrast image segmentation provided by the application is shown in FIG. 1. Figure 1 As shown in FIG. 1, the system is composed of an angiography device 11 and a processing electronic device 12. In the operation process, first, a catheter is inserted from the femoral artery or other peripheral arteries of a patient and guided to the ascending aorta, and then inserted into the left or right coronary artery orifice. After the contrast agent is injected, the angiography device 11 shoots the developed contrast image and transmits the image data to the electronic device 12. The electronic device 12 performs segmentation and processing on the contrast image to generate a clear cardiovascular segmentation map, providing a basis for subsequent analysis.

[0051] The specific implementation form of the electronic device 12 is various, which can be an examination terminal equipped with a medical X-ray examination device, a doctor's office computer, a hospital information management system (HIS) management server of a hospital, or a cloud server, and the specific selection can be flexibly adjusted according to the actual application scene.

[0052] In a clinical scene, when a patient has a blood vessel related discomfort symptom, a blood vessel static examination (such as CT, MR) and a dynamic examination (such as DSA) are usually needed. After the examination is completed, the system scores the lesion area based on the image data to provide scientific and reliable data reference for the doctor's diagnosis and treatment.

[0053] Figure 2 A flowchart of the coronary centerline segmentation and naming method based on contrast image segmentation provided by the application is shown in FIG. 2. Figure 2 As shown in FIG. 2, the method comprises the following steps.

[0054] Step 100, an angiography image of a coronary vessel is acquired; wherein the angiography image comprises image data of at least one target shooting body position.

[0055] Specifically, the angiography image at least contains image data of a usable shooting position, which includes but is not limited to a left coronary left shoulder position, a left coronary foot position, a right coronary left anterior oblique position and a right coronary head position. These positions can ensure the integrity of part of the vascular structure. It should be noted that the image frame filled with contrast agent is selected as the effective image. In the image size selection, the target image size can be set as a 2 multiple size with the same length and width, for example, 512*512 pixels. The specific size can be adjusted according to the actual application scene, which is not limited herein.

[0056] Step 200, inputting the angiography image into a pre-trained joint segmentation and segmentation model to output a blood vessel segmentation result and a segmentation naming result; wherein the joint segmentation and segmentation model is obtained by training a multi-task deep learning model based on a coronary angiography image data set.

[0057] It should be noted that the joint segmentation and segmentation model includes an image encoding layer, a segmentation image decoding layer and a segmentation image decoding layer, and the image encoding layer, the segmentation image decoding layer and the segmentation image decoding layer constitute a coronary artery segmentation task and a coronary artery segmentation naming task.

[0058] Step 200 inputs the angiography image into the pre-trained joint segmentation and segmentation model to output the blood vessel segmentation result and the segmentation naming result, which includes:

[0059] The angiography image is input into the joint segmentation and segmentation model, and the coronary artery segmentation task and the coronary artery segmentation naming task are performed respectively to output the blood vessel segmentation result and the segmentation naming result.

[0060] The coronary artery segmentation task is used to extract the filling contour of the coronary artery, and each pixel in the angiography image is assigned to a single class. The coronary artery segmentation naming task is used to anatomically segment the main branches (such as the left anterior descending branch, the circumflex branch and the right coronary artery) and the secondary branches of the coronary artery based on the coronary artery segmentation, and the pixels of different segments are assigned to different classes.

[0061] Further, the training process of the joint segmentation and segmentation model provided in the embodiment is described, which specifically includes:

[0062] Step 210, training a deep learning model based on a pre-constructed multi-task learning framework to make the deep learning model learn multiple tasks simultaneously; wherein each task in the multiple tasks has a corresponding loss function, and the weighted sum of the loss functions of each task in the multiple tasks constitutes a total loss function; the total loss function adjusts the relevance of the segmentation loss and the segmentation loss through a consistency index, and the consistency index is calculated based on the number of overlapping area pixels of the segmentation error map and the segmentation error map.

[0063] Step 220, dynamically adjusting the weights of the deep learning model by minimizing the total loss function, to obtain the joint segmentation and sectioning model.

[0064] Specifically, the image encoding layer is implemented by a feature extraction network, including but not limited to a convolutional neural network (CNN) based model and a Transformer based model. After the coronary image is encoded by the feature extraction network, a multi-scale feature matrix of the image is obtained for use by the decoding layers of different tasks.

[0065] Specifically, the image segmentation part and the section naming part are both provided with a forward propagable loss function, including but not limited to mean square error loss, cross entropy loss and logarithmic loss. The present application sets a multi-task learning framework in the model, and the model learns multiple tasks simultaneously, each task being provided with an independent loss function, which are combined into a total loss function. The model updates the weights by minimizing the total loss function, thereby realizing the collaborative optimization of the segmentation and sectioning tasks.

[0066] In some embodiments, the total loss function can be calculated by the following formula:

[0067] (1)

[0068] wherein, is the total loss of the model, is the segmentation loss, is the section naming loss, and are preset parameters, respectively.

[0069] In some embodiments, the total loss function can also be calculated by the following formula:

[0070] (2)

[0071] The consistency index k proposed in this embodiment is used to quantify the correlation between the segmentation task and the sectioning task, and the calculation method includes the following steps:

[0072] Error map generation: the model generates a pixel-level difference map with the true value through the segmentation task and the sectioning task, i.e. a segmentation error map and a sectioning error map.

[0073] Consistency pixel definition: when there is a spatial overlap area between the segmentation error map and the sectioning error map, if a pixel point simultaneously shows significant difference in the two error maps, the pixel point is defined as a consistency pixel. These consistency pixels represent the sectioning error area caused by segmentation error.

[0074] Index calculation: the mathematical expression of consistency index k is the ratio of the total number of consistent pixels to the total number of significant difference pixels in the segmentation error map, and the specific formula is as follows:

[0075] (3)

[0076] When the segmentation error induced segmentation error is more significant, the system will generate more consistent pixels, so that the value of k tends to 1. This index can effectively reflect the influence degree of the segmentation task on the segmentation task, and provide quantitative basis for model optimization.

[0077] In the embodiment of the application, in addition to the aforementioned loss function fusion mode, another method for optimizing image segmentation results is also provided. According to the prediction probability of each pixel, binary processing is performed by setting a threshold to distinguish the foreground (blood vessel region) and the background part. The specific implementation manner is as follows: the segmentation loss and the segmentation loss of the image are respectively calculated to obtain the loss function, and the network parameters are updated independently; the blood vessel segmentation probability is determined according to the segmentation loss, and the blood vessel segmentation image is updated; when the segmentation loss is calculated, the intersection of the segmentation result and the predicted segmentation result is taken, and only the pixels in the predicted segmentation area are calculated for the segmentation and segmentation loss; for the region of the difference between the predicted union and the intersection, only the segmentation loss is calculated, so as to ensure the accuracy of the segmentation result.

[0078] In the embodiment of the application, the loss weight is dynamically adjusted according to the training stage. In the early training stage, the proportion of the segmentation loss is adjusted to be high, and the proportion of the segmentation loss is adjusted to be low, so that the overall blood vessel contour is preferentially trained; in the middle training stage, the proportion of the segmentation loss is adjusted to be high, and the proportion of the segmentation loss is adjusted to be low, so that the segmentation and segmentation naming categories in the contour are optimized on the basis of stabilizing the contour; in the late training stage, the segmentation loss and the segmentation loss are balanced, the model performance is optimized as a whole, and the segmentation and segmentation tasks are cooperatively converged.

[0079] When evaluating the performance of the semantic segmentation and instance segmentation multi-task model on the validation set, the following steps are adopted: inputting the validation image not participating in the training into the trained model to obtain the prediction result; according to the intersection over union (IoU) and the Dice coefficient index, the performance of the model in the segmentation and segmentation tasks is quantitatively evaluated to ensure the accuracy and robustness of the output result. The calculation method is as follows:

[0080] (4)

[0081] (5)

[0082] wherein, is the predicted pixel region, is the real pixel region, represents the intersection area of the predicted pixel region and the real pixel region.

[0083] The embodiment of the application constructs a joint segmentation and segmentation model, and synchronously completes the coronary artery segmentation and segmentation naming tasks through a multi-task learning framework. The model extracts multi-scale features through a shared encoding layer, and designs independent decoding layers (segmentation decoding layer and segmentation decoding layer) to process different tasks. The segmentation task generates a blood vessel contour through a U-Net structure, and the segmentation task names the branch based on an anatomical classification. The feature fusion is strengthened through a skip connection and an attention mechanism, so that the consistency of the segmentation and segmentation results at the pixel level is ensured. The joint model avoids the error accumulation problem in the traditional step-by-step processing, and significantly improves the efficiency and accuracy of the coronary artery analysis. In the model training, a dynamic loss weight distribution strategy is introduced, the segmentation loss is dominated in the early training stage, and the accurate blood vessel contour is established in priority; the segmentation loss is turned in the middle training stage, and the branch naming classification is refined; and the two are balanced in the later training stage to globally optimize. The consistency index is used to quantify the correlation between the tasks, and the double-task errors are fused in the total loss function. This mechanism ensures that the model can automatically identify the segmentation deviation caused by the segmentation error, optimize the parameters in a targeted manner, and improve the robustness of the model.

[0084] Step 300, based on the blood vessel segmentation result and the segmentation naming result, generating an ordered center line of the coronary artery; wherein the ordered center line is a center line with a clear starting point, ending point and direction for each segment in the skeleton graph of the coronary artery.

[0085] Step 300, based on the blood vessel segmentation result and the segmentation naming result, generating an ordered center line of the coronary artery, comprising:

[0086] Step 310, extracting the blood vessel segmentation result to obtain an initial skeleton graph of the coronary artery, filtering noise branches in the initial skeleton graph according to a pre-set branch pixel length threshold, and obtaining a skeleton graph of the coronary artery.

[0087] Step 320, according to the anatomical rules of the coronary artery tree 16 segmentation method, determining a target segment in the skeleton graph of the coronary artery based on a bifurcation point and an end point.

[0088] Step 330, according to the segmentation naming result, determining a starting point of the coronary artery.

[0089] Step 340, based on the skeleton graph of the coronary artery, the target segment and the starting point of the coronary artery, generating an ordered center line of the coronary artery.

[0090] Specifically, first, a morphological thinning algorithm or a distance transform method is used to extract an initial skeleton graph of the coronary vessel from the blood vessel segmentation result, and the initial skeleton graph mainly includes the main branch skeleton of the blood vessel, which is used for subsequent segmentation processing. Then, the noise branches in the initial skeleton graph are filtered to obtain a skeleton graph, which specifically includes removing the skeleton parts with too short branch length, setting a branch pixel length threshold, and regarding the branches with length less than the threshold as invalid segments; at the same time, the skeleton graph is divided into the following segment types based on the pixel meaning segmentation definition: a starting point to a bifurcation point is a segment, a starting point to a branch end is a segment, and a bifurcation point to a branch end is a segment. Then, according to the segment naming result, the starting position of the coronary vessel is determined in combination with the coronary artery tree 16 segmentation method. Specifically, the starting segment of the left coronary is No. 5 segment, and the starting segment of the right coronary is No. 1 segment. Based on the segment categories of the left and right coronaries, the spatial position of the starting point is determined by mapping to the skeleton graph. Finally, based on the skeleton graph and the starting point position, an ordered center line is generated. The ordered center line is a center line with a clear starting point, end point and direction for each segment in the skeleton graph, which ensures that the output result conforms to the anatomical structure characteristics of the coronary tree and provides structured data support for clinical diagnosis.

[0091] The embodiment of the present application extracts a blood vessel skeleton graph from the segmentation result by using a morphological thinning algorithm, and filters noise in combination with anatomical rules. A branch length threshold is defined, and branches that are too short are regarded as noise and invalid branches are removed. The skeleton segments are divided based on bifurcation points and end points to ensure that the topological structure conforms to the characteristics of the coronary tree. The starting point is determined by the anatomical segment naming result, and an ordered center line is generated, which gives each segment a clear starting point, end point and direction, thereby providing structured data support for clinical diagnosis.

[0092] Step 400, based on the ordered center line and the segment naming result, generates a coronary center line segment naming result of the coronary vessel in accordance with anatomical rules.

[0093] Step 400 generates a coronary center line segment naming result of the coronary vessel based on the ordered center line and the segment naming result, including:

[0094] Step 410, according to the ordered center line and the segment naming result, sequentially assigns segment categories to the coordinate points on the ordered center line to determine the category of each coordinate point.

[0095] Step 420, based on the determined category of each coordinate point, respectively judges each ordered center line by a dynamic category correction strategy, and outputs a coronary center line segment naming result of the coronary vessel in accordance with anatomical rules.

[0096] Step 420 specifically includes:

[0097] Step 421, in the case that there are multiple segment naming results in a segment of the ordered center line, if the proportion of any category in the pixels of the ordered center line is lower than the preset threshold and the distribution is discontinuous, it is determined that the category is a noise and is filtered out.

[0098] Step 422, in the case that there are multiple segment naming results in a segment of the ordered center line, if the proportion of any category in the pixels of the ordered center line is higher than the preset threshold and the distribution is continuous, the category is retained and is mapped and matched with the segment naming result, and a coronary center line segment naming result of the coronary vessel in line with the anatomical rule is output.

[0099] Specifically, first, according to the ordered center line and the segment naming result in line with the anatomy, the coordinate points on the ordered center line are sequentially assigned segment categories, so that each coordinate point has a clear segment label. For the case that there are multiple segment naming results in a segment of the center line, if the proportion of a single category in the whole segment is low and the distribution is discontinuous, the category is excluded as a noise, and the whole segment is uniformly assigned to the segment naming result in line with the anatomy. For the case that there are multiple segment naming results in a segment of the center line, if the proportion of a single category in the whole segment is high and the distribution is continuous, the category is retained, and is regarded as the difference between the pixel significance segment and the anatomy segment. The anatomical segment naming is used as a reference to ensure that the output result meets the clinical diagnosis standard.

[0100] The embodiment of the present application proposes a dynamic category correction strategy for the integration of segment naming and center line. The abnormal classification pixels with low proportion and dispersion are directly excluded, and the continuous and dominant segment difference is preferentially corrected according to the anatomical naming. This mechanism effectively balances the conflict between pixel-level prediction noise and anatomical rules, and ensures that the output result has both algorithm accuracy and clinical interpretability.

[0101] The above is a step description of the coronary centerline segmentation naming method based on contrast image segmentation provided by the present application. As can be seen from the description of the above steps, according to the coronary centerline segmentation naming method based on contrast image segmentation provided by the present application, a contrast angiogram of a coronary vessel is obtained; wherein the contrast angiogram contains image data of at least one target shooting body position; the contrast angiogram is input into a pre-trained joint segmentation segmentation model to output a blood vessel segmentation result and a segmentation naming result; wherein the joint segmentation segmentation model is obtained by training a multi-task deep learning model based on a coronary vessel contrast image data set; based on the blood vessel segmentation result and the segmentation naming result, an ordered centerline of the coronary vessel is generated; wherein the ordered centerline is a centerline that each segment in the skeleton graph of the coronary vessel has a clear starting point, end point and direction; based on the ordered centerline and the segmentation naming result, a coronary centerline segmentation naming result of the coronary vessel in accordance with anatomical rules is generated. Therefore, the present application synchronously completes two tasks through a joint segmentation segmentation model, avoids error accumulation, and improves overall analysis efficiency and accuracy.

[0102] The coronary centerline segmentation naming device based on contrast image segmentation provided by the present application is described below, and the coronary centerline segmentation naming device based on contrast image segmentation described below can be correspondingly referred to the coronary centerline segmentation naming method based on contrast image segmentation described above.

[0103] Figure 3 FIG. 1 is a structural schematic diagram of the coronary centerline segmentation naming device based on contrast image segmentation provided by the present application, as shown in FIG. 1, the coronary centerline segmentation naming device based on contrast image segmentation provided by the present application comprises: Figure 3

[0104] The acquisition module 301 is configured to acquire a contrast angiogram of a coronary vessel; wherein the contrast angiogram contains image data of at least one target shooting body position;

[0105] The segmentation naming module 302 is configured to input the contrast angiogram into a pre-trained joint segmentation segmentation model to output a blood vessel segmentation result and a segmentation naming result; wherein the joint segmentation segmentation model is obtained by training a multi-task deep learning model based on a coronary vessel contrast image data set;

[0106] The centerline extraction module 303 is configured to generate an ordered centerline of the coronary vessel based on the blood vessel segmentation result and the segmentation naming result; wherein the ordered centerline is a centerline that each segment in the skeleton graph of the coronary vessel has a clear starting point, end point and direction;

[0107] ​The center line and the segment naming fusion module 304 is configured to generate a coronary center line segment naming result of the coronary vessel in accordance with the anatomical rules based on the ordered center line and the segment naming result.

[0108] The coronary center line segment naming device based on contrast image segmentation provided by the application obtains a angiography image of a coronary vessel, wherein the angiography image comprises image data of at least one target shooting body position; the angiography image is input into a pre-trained joint segmentation and segmentation model to output a vessel segmentation result and a segment naming result; the joint segmentation and segmentation model is obtained by training a multi-task deep learning model based on a coronary angiography image data set; an ordered center line of the coronary vessel is generated based on the vessel segmentation result and the segment naming result; the ordered center line is a center line with a clear starting point, ending point and direction for each segment in the skeleton graph of the coronary vessel; and a coronary center line segment naming result of the coronary vessel in accordance with the anatomical rules is generated based on the ordered center line and the segment naming result. Therefore, the joint segmentation and segmentation model is used to complete two tasks simultaneously, error accumulation is avoided, and the overall analysis efficiency and accuracy are improved.

[0109] In the embodiment, the joint segmentation and segmentation model comprises an image encoding layer, a segmentation image decoding layer and a segment image decoding layer, and the image encoding layer, the segmentation image decoding layer and the segment image decoding layer constitute a coronary segmentation task and a coronary segment naming task.

[0110] The segment naming module 302 is specifically configured to:

[0111] The angiography image is input into the joint segmentation and segmentation model to perform the coronary segmentation task and the coronary segment naming task respectively, and the vessel segmentation result and the segment naming result are output.

[0112] The coronary segmentation task is configured to extract a filled contour of the coronary artery and assign each pixel in the angiography image to a single category, and the coronary segment naming task is configured to anatomically segment the main branch and the secondary branch of the coronary artery based on the coronary segmentation and assign the pixels of different segments to different categories.

[0113] In the embodiment, the device further comprises a training module, which is specifically configured to:

[0114] training a deep learning model based on a pre-constructed multi-task learning framework, so that the deep learning model learns multiple tasks simultaneously; wherein each of the multiple tasks has a corresponding loss function, and a weighted sum of the loss functions of each of the multiple tasks constitutes a total loss function; the total loss function adjusts the relevance of segmentation loss and segmentation loss through a consistency index, and the consistency index is calculated based on the number of overlapping area pixels of segmentation error map and segmentation error map;

[0115] The weights of the deep learning model are dynamically adjusted by minimizing the total loss function to obtain the joint segmentation and segmentation model.

[0116] Based on the above embodiment, in this embodiment, the device further comprises a dynamic adjustment module, specifically for:

[0117] In the early training stage, the segmentation loss weight is higher than the segmentation loss weight;

[0118] In the middle training stage, the segmentation loss weight is higher than the segmentation loss weight;

[0119] In the later training stage, the segmentation loss weight and the segmentation loss weight are balanced.

[0120] Based on the above embodiment, in this embodiment, the centerline extraction module 303 is specifically configured to:

[0121] The blood vessel segmentation result is extracted to obtain an initial skeleton graph of the coronary blood vessel, and noise branches in the initial skeleton graph are filtered according to a pre-set branch pixel length threshold to obtain a skeleton graph of the coronary blood vessel;

[0122] According to the anatomical rules of the coronary artery tree 16 segmentation method, the target segmentation is determined in the skeleton graph of the coronary blood vessel based on the bifurcation point and the end point;

[0123] According to the segmentation naming result, the starting point of the coronary blood vessel is determined;

[0124] Based on the skeleton graph of the coronary blood vessel, the target segmentation and the starting point of the coronary blood vessel, an ordered centerline of the coronary blood vessel is generated.

[0125] Based on the above embodiment, in this embodiment, the centerline and segmentation naming fusion module 304 is specifically configured to:

[0126] According to the ordered centerline and the segmentation naming result, the coordinate points on the ordered centerline are sequentially assigned with segmentation categories to determine the category of each coordinate point;

[0127] Based on the determined category of each coordinate point, each ordered center line is judged respectively by a dynamic category correction strategy, and the coronary center line segmentation naming result of the coronary blood vessel meeting the anatomical rules is output.

[0128] Based on the above embodiment, in this embodiment, the center line and segmentation naming fusion module 304 is specifically further used for:

[0129] In the case where there are multiple segmentation naming results in an ordered center line, if the pixel proportion of any category in the ordered center line is lower than a preset threshold and the distribution is discontinuous, the category is determined as a noise point and is filtered out;

[0130] In the case where there are multiple segmentation naming results in an ordered center line, if the pixel proportion of any category in the ordered center line is higher than a preset threshold and the distribution is continuous, the category is reserved and is mapped and matched with the segmentation naming result, and the coronary center line segmentation naming result of the coronary blood vessel meeting the anatomical rules is output.

[0131] Figure 4 An example of a schematic diagram of the physical structure of an electronic device is shown in Figure 4 The electronic device can be a robot or other electronic device. The electronic device can include a processor 410, a communications interface 420, a memory 430, and a communications bus 440. The processor 410, the communications interface 420, and the memory 430 can communicate with each other through the communications bus 440. The processor 410 can invoke logical instructions in the memory 430 to execute a coronary center line segmentation naming method based on contrast image segmentation, including:

[0132] Obtaining an angiogram image of a coronary blood vessel; wherein the angiogram image contains image data of at least one target shooting body position;

[0133] Inputting the angiogram image into a pre-trained joint segmentation and segmentation model to output a vessel segmentation result and a segmentation naming result; wherein the joint segmentation and segmentation model is obtained by training a multi-task deep learning model based on a coronary angiogram image dataset;

[0134] Based on the vessel segmentation result and the segmentation naming result, an ordered center line of the coronary blood vessel is generated; wherein the ordered center line is a center line with a clear starting point, ending point, and direction for each segment in the skeleton graph of the coronary blood vessel;

[0135] Based on the ordered center line and the segmentation naming result, a coronary center line segmentation naming result of the coronary blood vessel meeting the anatomical rules is generated.

[0136] Further, the logic instructions in the memory 430 described above can be implemented in the form of software functional units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0137] On the other hand, the present application also provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the coronary centerline segmentation naming method based on contrast image segmentation provided by the above-mentioned methods, comprising:

[0138] Obtaining an angiogram image of a coronary vessel; wherein the angiogram image contains image data of at least one target shooting body position;

[0139] Inputting the angiogram image into a pre-trained joint segmentation segmentation model to output a vessel segmentation result and a segmentation naming result; wherein the joint segmentation segmentation model is obtained by training a multi-task deep learning model based on a coronary angiogram image data set;

[0140] Based on the vessel segmentation result and the segmentation naming result, generating an ordered centerline of the coronary vessel; wherein the ordered centerline is a centerline that each segment in the skeleton graph of the coronary vessel has a clear starting point, ending point and direction;

[0141] Based on the ordered centerline and the segmentation naming result, generating a coronary centerline segmentation naming result of the coronary vessel that conforms to anatomical rules.

[0142] In another aspect, the present application also provides a non-transitory computer readable storage medium having a computer program stored thereon, which is executed by a processor to implement the coronary centerline segmentation naming method based on contrast image segmentation provided by the above-mentioned methods, comprising:

[0143] obtaining an angiography image of a coronary vessel; wherein the angiography image comprises image data of at least one target shooting body position;

[0144] inputting the angiography image into a pre-trained joint segmentation model to output a vessel segmentation result and a segmentation naming result; wherein the joint segmentation model is obtained by training a multi-task deep learning model based on a coronary angiography image data set;

[0145] generating an ordered center line of the coronary vessel based on the vessel segmentation result and the segmentation naming result; wherein the ordered center line is a center line with a clear starting point, ending point and direction for each segment in a skeleton graph of the coronary vessel;

[0146] generating a coronary center line segmentation naming result of the coronary vessel in accordance with anatomical rules based on the ordered center line and the segmentation naming result.

[0147] The device embodiments described above are only schematic, wherein the units shown as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment scheme according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0148] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary universal hardware platforms, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, which can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and include a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0149] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for segmenting and naming coronary artery centerlines based on angiographic image segmentation, characterized in that: include: Acquire an angiographic image of a coronary vessel; wherein the angiographic image comprises: image data of at least one target photographing position; Inputting the angiography image into a pre-trained joint segmentation model to output a vessel segmentation result and a segmentation naming result; wherein the joint segmentation model is obtained by training a multi-task deep learning model based on a coronary angiography image dataset; Based on the vessel segmentation result and the segment naming result, an ordered centerline of the coronary vessel is generated; wherein the ordered centerline is a centerline having a clear starting point, an end point, and a direction for each segment in the skeleton image of the coronary vessel; generating a coronary centerline segmentation naming result of the coronary vessel that complies with anatomical rules based on the ordered centerline and the segmentation naming result; The joint segmentation and segmentation model includes: a shared image coding layer, an independent segmentation image decoding layer and an independent segmentation image decoding layer, wherein the image coding layer, the segmentation image decoding layer and the segmentation image decoding layer constitute a coronary artery segmentation task and a coronary artery segmentation naming task; The step of inputting the angiography image into a pre-trained joint segmentation model and outputting a blood vessel segmentation result and a segmentation naming result includes: Inputting the angiography image into the joint segmentation model, performing the coronary artery segmentation task and the coronary artery segmentation naming task respectively, and outputting the blood vessel segmentation result and the segmentation naming result; Among them, the coronary artery segmentation task is used to extract the filled contour of the coronary artery and assign each pixel in the angiography image to a single category. The coronary artery segmentation and naming task is used to anatomically segment the main branches of the coronary artery and their secondary branches based on the coronary artery segmentation, and assign pixels of different segments to different categories.

2. The method for segmenting and naming coronary artery centerlines based on angiographic image segmentation according to claim 1, characterized in that: The training process of the joint segmentation model includes: A deep learning model is trained based on a pre-built multi-task learning framework so that the deep learning model learns multiple tasks simultaneously; each of the multiple tasks has a corresponding loss function, and the weighted sum of the loss functions of each of the multiple tasks constitutes a total loss function; the total loss function adjusts the correlation between the segmentation loss and the segmentation loss using a consistency index, where the consistency index is calculated based on the number of pixels in the overlapping area of ​​the segmentation error map and the segmentation error map; The weights of the deep learning model are dynamically adjusted by minimizing the total loss function to obtain the joint segmentation model.

3. The method for segmenting and naming coronary artery centerlines based on angiographic image segmentation according to claim 2, characterized in that: Dynamically adjusting the weight of the deep learning model by minimizing the total loss function includes: In the early stage of training, the segmentation loss weight is higher than the segmentation loss weight; In the middle of training, the segmentation loss weight is higher than the segmentation loss weight; In the later stages of training, the segmentation loss weight is balanced with the fragmentation loss weight.

4. The method for segmenting and naming coronary artery centerlines based on angiographic image segmentation according to claim 1, characterized in that: Generating the ordered centerline of the coronary vessels based on the vessel segmentation result and the segment naming result includes: Extracting the blood vessel segmentation result to obtain an initial skeleton image of the coronary artery, filtering noise branches in the initial skeleton image according to a preset branch pixel length threshold to obtain a skeleton image of the coronary artery; According to the anatomical rules of the coronary artery tree 16 segmentation method, a target segment is determined based on bifurcation points and end points in the skeleton image of the coronary vessels; determining the starting point of the coronary artery according to the segmented naming result; An ordered centerline of the coronary vessel is generated based on the skeleton image of the coronary vessel, the target segment, and the starting point of the coronary vessel.

5. The method for segmenting and naming coronary artery centerlines based on angiographic image segmentation according to claim 1, characterized in that: Generating the segmented naming result of the coronary centerline of the coronary vessel based on the ordered centerline and the segmented naming result includes: According to the ordered center line and the segment naming result, the coordinate points on the ordered center line are assigned segment categories in sequence to determine the category of each coordinate point; Based on the determined category of each coordinate point, each ordered centerline is judged separately through a dynamic category correction strategy, and the segmented naming result of the coronary centerline of the coronary artery that conforms to the anatomical rules is output.

6. The method for segmenting and naming coronary artery centerlines based on angiographic image segmentation according to claim 5, characterized in that: Based on the determined category of each coordinate point, each ordered centerline is judged separately through a dynamic category correction strategy, and a segmented naming result of the coronary centerline of the coronary artery that conforms to anatomical rules is output, including: In the case where there are multiple segmented naming results in an ordered centerline, if the pixel ratio of any category in the ordered centerline is lower than a preset threshold and the distribution is discontinuous, the category is determined to be noise and filtered out; When there are multiple segmentation naming results in an ordered centerline, if the pixel ratio of any category in the ordered centerline is higher than a preset threshold and is continuously distributed, the category is retained and mapped and matched with the segmentation naming result, and the coronary centerline segmentation naming result of the coronary artery that conforms to the anatomical rules is output.

7. A device for segmenting and naming coronary artery centerlines based on angiographic image segmentation, characterized in that: include: An acquisition module is used to acquire an angiographic image of a coronary vessel; wherein the angiographic image includes: image data of at least one target shooting position; a segmentation and naming module, configured to input the angiography image into a pre-trained joint segmentation model and output a vessel segmentation result and a segmentation and naming result; wherein the joint segmentation model is obtained by training a multi-task deep learning model based on a coronary angiography image dataset; a centerline extraction module, configured to generate an ordered centerline of the coronary vessel based on the vessel segmentation result and the segment naming result; wherein the ordered centerline is a centerline having a clear starting point, an ending point, and a direction for each segment in the skeleton image of the coronary vessel; a centerline and segment naming fusion module, configured to generate a coronary centerline segment naming result of the coronary vessel that complies with anatomical rules based on the ordered centerline and the segment naming result; The joint segmentation and segmentation model includes: a shared image coding layer, an independent segmentation image decoding layer and an independent segmentation image decoding layer, wherein the image coding layer, the segmentation image decoding layer and the segmentation image decoding layer constitute a coronary artery segmentation task and a coronary artery segmentation naming task; The segment naming module is specifically used to: Inputting the angiography image into the joint segmentation model, performing the coronary artery segmentation task and the coronary artery segmentation naming task respectively, and outputting the blood vessel segmentation result and the segmentation naming result; Among them, the coronary artery segmentation task is used to extract the filled contour of the coronary artery and assign each pixel in the angiography image to a single category. The coronary artery segmentation and naming task is used to anatomically segment the main branches of the coronary artery and their secondary branches based on the coronary artery segmentation, and assign pixels of different segments to different categories.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for naming coronary artery centerline segments based on angiography image segmentation as described in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for naming coronary artery centerline segments based on angiography image segmentation as claimed in any one of claims 1 to 6 is implemented.

Images