Method for tracking a blood vessel centerline and related products

Abstract

The disclosure discloses a method for tracking a blood vessel center line and related products, which comprises: taking a starting point of a blood vessel contained in target angiography data as a center point, and intercepting an image block of a preset size; inputting the image block into a trained blood vessel center line tracking model for center line tracking to output a next center point of the center point; taking the next center point as a new center point, intercepting a new image block of a preset size; based on the new image block, proceeding to input the image block into the trained blood vessel center line tracking model for center line tracking to output a next center point of the center point, and continuing to execute until no new center point is output; and determining a center line of the blood vessel based on all the center points. By using the scheme of the disclosure, the influence of a large aneurysm on blood vessel center line extraction can be avoided, and the accuracy of blood vessel center line tracking is improved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates generally to the technical field of computer vision. More particularly, the present disclosure relates to a method, an electronic device and a computer readable storage medium for training a blood vessel centerline tracking model. Further, the present disclosure also relates to a method, an electronic device and a computer readable storage medium for tracking a blood vessel centerline. BACKGROUND

[0002] In the field of blood vessel disease diagnosis, Digital Subtraction Angiography (DSA) has long been regarded as the gold standard for screening and diagnosis of various blood vessel diseases, especially in the diagnosis and accurate measurement of aneurysm. In addition, CT Angiography (CTA) and Magnetic Resonance Angiography (MRA) techniques have also been widely used, which can be used for preliminary screening of blood vessel diseases, detection of aneurysm and disease follow-up, etc., providing diversified imaging basis for clinical diagnosis.

[0003] Current blood vessel centerline extraction techniques usually follow a step-by-step process, i.e., first performing blood vessel segmentation on medical images, and then extracting the centerline by means of skeletonization algorithm. However, this method has obvious defects. In actual application, when the segmented blood vessel image is processed by skeletonization, the centerline is prone to be disconnected and incorrectly connected. In order to correct these problems, complex post-processing algorithms need to be introduced, but even so, it is still difficult to completely guarantee the integrity of the blood vessel topological structure. Especially when facing the case where there is an aneurysm in the blood vessel, the existence of a large aneurysm will interfere with the accurate extraction of the blood vessel centerline, which significantly affects the stability of the whole system, limiting its efficient application in clinical practice.

[0004] Therefore, it is urgent to provide a scheme capable of tracking the blood vessel tree structure and ensuring the integrity of the blood vessel topological structure, so as to realize accurate tracking of the blood vessel centerline, effectively avoid the adverse interference of a large aneurysm on the tracking result, and improve the extraction accuracy of the blood vessel centerline. SUMMARY

[0005] In order to at least solve one or more technical problems as mentioned above, the present disclosure proposes a scheme for tracking a blood vessel centerline in the following aspects.

[0006] In a first aspect, the present disclosure provides a method for training a blood vessel centerline tracking model, the method comprising: performing blood vessel and aneurysm segmentation on angiography data to obtain a segmentation result; performing centerline extraction on the blood vessels in the segmentation result to obtain a centerline of the blood vessels; starting from a blood vessel starting point on the centerline, progressively slicing the angiography data to obtain a plurality of image blocks distributed along the centerline, a center point of each image block being located on the centerline, and the blood vessels included in each image block constituting the blood vessels; performing information labeling on each image block according to the centerline and the segmentation result to obtain labeling information of each image block, the labeling information including a block segmentation result, a blood vessel radius at a center point, a class of the center point, and a next center point of the center point; and inputting the plurality of image blocks and the labeling information of the plurality of image blocks as training data into the blood vessel centerline tracking model to train the blood vessel centerline tracking model.

[0007] In some embodiments, inputting the plurality of image blocks and the labeling information of the plurality of image blocks as training data into the blood vessel centerline tracking model to train the blood vessel centerline tracking model comprises: determining an image block with the blood vessel starting point as a center point as a starting image block; and inputting the plurality of image blocks and the labeling information of the plurality of image blocks as training data into the blood vessel centerline tracking model to train the blood vessel centerline tracking model in sequence along the centerline starting from the starting image block.

[0008] In some embodiments, the blood vessel centerline tracking model comprises an image encoder, a classifier, and a tracking decoder, and inputting the image blocks and the labeling information of the image blocks as training data into the blood vessel centerline tracking model to train the blood vessel centerline tracking model comprises: inputting the image blocks into the image encoder to perform feature extraction to obtain image features; inputting the image features into the classifier to perform class prediction to obtain a predicted class of a center point of the image blocks; inputting the image features into the tracking decoder to perform tracking prediction to obtain a predicted blood vessel radius at the center point of the image blocks and a predicted next center point of the center point; determining a loss value based on the class of the center point, the next center point of the center point, the predicted class, and the predicted next center point of the center point, and updating parameters of the image encoder, the classifier, and the tracking decoder based on the loss value.

[0009] In some embodiments, the blood vessel centerline tracking model further comprises a segmentation decoder, and the method further comprises: after obtaining the image features, inputting the image features into the segmentation decoder for segmentation to obtain a predicted block segmentation result; determining a loss value based on the block segmentation result, the category of the center point, the next center point of the center point, the predicted block segmentation result, the predicted category, and the predicted next center point, and updating parameters of the image encoder, the classifier, the tracking decoder, and the tracking decoder based on the loss value.

[0010] In some embodiments, the categories include a termination point, a connection point, and a bifurcation point; and in a case where the category of the center point is a termination point, the center point has no next center point; in a case where the category of the center point is a connection point, the number of next center points of the center point is 1; and in a case where the category is a bifurcation point, the number of next center points of the center point is the number of branches emitted by the bifurcation point.

[0011] In a second aspect, the present disclosure provides a method for tracking a blood vessel centerline, comprising: taking a starting point of a blood vessel contained in target angiography data as a center point, and cutting an image block of a preset size; inputting the image block into a blood vessel centerline tracking model trained according to the method of the first aspect and the plurality of embodiments thereof to perform centerline tracking, to output a next center point of the center point; taking the next center point as a new center point, and cutting a new image block of the preset size; based on the new image block, proceeding to input the image block into the blood vessel centerline tracking model trained according to the method of the first aspect and the plurality of embodiments thereof to perform centerline tracking, to output a next center point of the center point, and continuing to perform until no new center point is outputted; and determining a centerline of the blood vessel based on all the center points.

[0012] In some embodiments, the blood vessel centerline tracking model comprises an image encoder, a classifier, and a tracking decoder; and inputting the image block into the blood vessel centerline tracking model to perform centerline tracking to output a next center point of the center point comprises: inputting the image block into the image encoder to perform feature extraction to obtain image features; inputting the image features into the classifier to perform category prediction to obtain a category of the center point of the image block; and inputting the image features into the tracking decoder to perform tracking prediction to obtain the next center point of the center point.

[0013] In some embodiments, the categories include termination points, connection points, and branch points; and in the case that the category of the center point is a termination point, the center point has no next center point; in the case that the category of the center point is a connection point, the number of next center points of the center point is 1; in the case that the category of the center point is a branch point, the number of next center points of the center point is the number of branches emitted by the branch point.

[0014] In a third aspect, the present disclosure provides an electronic device, comprising: a processor; and a memory storing program instructions for training a blood vessel centerline tracking model and / or for tracking a blood vessel centerline, which when executed by the processor, cause the implementation of the method in the foregoing first aspect and its multiple embodiments and / or the method in the foregoing second aspect.

[0015] In a fourth aspect, the present disclosure provides a computer-readable storage medium storing program instructions for training a blood vessel centerline tracking model and / or for tracking a blood vessel centerline, which when executed by a processor, cause the implementation of the method in the foregoing first aspect and its multiple embodiments and / or the method in the foregoing second aspect and its multiple embodiments.

[0016] As the above-provided scheme for tracking a blood vessel centerline, in the training phase of the blood vessel centerline tracking model, by accurately labeling the center point, blood vessel radius, next center point, and the like information of each image block in the training data, the model can learn more detailed blood vessel morphological features, so as to more accurately track the centerline of the blood vessel in actual application, reduce the deviation and error connection, and help to improve the accuracy of the blood vessel centerline tracking. In addition, the training data contains the segmentation information of the blood vessel and aneurysm, and the image blocks are labeled accordingly. The model can learn how to accurately extract the blood vessel centerline in the presence of aneurysm in the training process. This avoids the influence of large aneurysm on the extraction of the blood vessel centerline, so that the model can still stably output reliable centerline when facing diseased blood vessels, improving the overall stability of the system and the robustness to aneurysm and other abnormal conditions.

[0017] In the application stage of the blood vessel centerline tracking model, the image block is intercepted with the blood vessel starting point as the center point, and the tracking range is gradually expanded according to the prediction of the model, so it can adapt to blood vessel structures of different sizes and shapes. Whether it is a relatively thick blood vessel or a small branch blood vessel, the centerline can be accurately extracted through recursive tracking, improving the universality and applicability of the method. In addition, since each tracking is iterated with the next center point of the previous center point as the new center point, this method can ensure the connectivity of the blood vessel centerline and avoid the occurrence of disconnected or incorrect connection. At the same time, based on the prediction ability of the training model, the topological structure integrity of the blood vessel can be better maintained. BRIEF DESCRIPTION OF DRAWINGS

[0018] The above and other objects, features and advantages of the disclosed example embodiments will be more apparent from the following detailed description read in conjunction with the accompanying drawings, in which several disclosed example embodiments are shown by way of example, and wherein like or corresponding elements show like or corresponding parts, wherein:

[0019] Figure 1 An exemplary structural diagram of a blood vessel centerline tracking model according to an embodiment of the present disclosure is shown;

[0020] Figure 2 An exemplary flowchart of a method for training a blood vessel centerline tracking model according to an embodiment of the present disclosure is shown;

[0021] Figure 3 An exemplary flowchart of a processing process of an image block in a blood vessel centerline tracking model according to an embodiment of the present disclosure is shown;

[0022] Figure 4 An exemplary flowchart of a method for tracking a blood vessel centerline according to an embodiment of the present disclosure is shown;

[0023] Figure 5 An exemplary structural block diagram of an electronic device according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present disclosure will be described clearly and completely below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, rather than all the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present disclosure.

[0025] It should be understood that the terms “comprising” and “including” used in this disclosure and claims indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0026] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this disclosure. As used in this disclosure and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this disclosure and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0027] As used in this specification and claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0028] The specific embodiments disclosed herein will now be described in detail with reference to the accompanying drawings.

[0029] Figure 1 An exemplary structural diagram of a vascular centerline tracking model according to an embodiment of this disclosure is shown. Figure 1 As shown, the blood vessel centerline tracking model disclosed herein may include an image encoder 101, a classifier 102, a tracking decoder 103, and a segmentation decoder 104. Here, the model architecture of the tracking decoder 103 is preferably a Transformer architecture. The image encoder 101 and the segmentation decoder 104 may adopt a UNet architecture, where the image encoder 101 is responsible for performing the downsampling process, and the segmentation decoder 104 corresponds to the upsampling process.

[0030] The aforementioned Transformer architecture is a deep learning architecture based on self-attention mechanism. In computer vision, Transformers can be used to process sequences of image patches, capturing global dependencies between different regions in an image, thus improving the performance of the model. The aforementioned Unet architecture is a classic convolutional neural network architecture widely used in image segmentation tasks, especially in medical image analysis. It consists of an encoder and a decoder, forming a U-shaped structure. The encoder extracts high-level features of the image through convolution and pooling operations and gradually reduces the spatial resolution, while the decoder gradually restores the spatial resolution through deconvolution or upsampling operations, combines high-level features with low-level features, and finally generates a segmentation result with the same size as the input image.

[0031] The input of the blood vessel centerline tracking model is an image patch in three-dimensional angiography data. The blood vessel can be an intracranial blood vessel, and the angiography data can be any one of digital subtraction angiography (DSA), computed tomography angiography (CTA), and magnetic resonance angiography (MRA). After the image patch is input into the blood vessel centerline tracking model, it is first subjected to feature extraction by the image encoder 101 to obtain image features. Thereafter, the image features are respectively subjected to the classifier 102, the tracking decoder 103, and the segmentation decoder 104, and perform corresponding operations.

[0032] In the segmentation decoder 104, the blood vessels and aneurysms contained in the image patch can be segmented based on the image features to obtain a segmentation result. In the classifier 102, the class of the center point (for the sake of distinction, it can be referred to as the current center point) of the image patch can be predicted based on the image features to obtain the class of the current center point. The class can include but is not limited to a termination point, a connection point, and a bifurcation point. Then, in the tracking decoder, the radius of the blood vessel at the current center point and the next center point of the current center point can be predicted based on the image features. In the embodiments of the present disclosure, the next center point of the current center point refers to one or more next center points connected to the current center point. It can be understood that the number and distribution of the next center points of the current center point will be different depending on the class of the current center point.

[0033] Specifically, in the case where the class of the current center point is a termination point, the current center point has no next center point, i.e., the number of next center points is 0; in the case where the class of the current center point is a connection point, the next center point of the current center point is located on the branch where the current center point is located, and is located between the current center point and the termination point of the branch, and the number of next center points is 1; in the case where the class of the current center point is a bifurcation point, the next center points of the current center point are located on the branches emitted by the bifurcation point, and the number of next center points is equal to the number of branches emitted by the bifurcation point, and the next center point on each branch is located between the bifurcation point and the termination point of each branch.

[0034] Since the current center point has no next center point when its category is termination point, there is no need to input the image features of the image patch into the tracking decoder to predict the next center point. Conversely, if the current center point has a next center point when its category is connection point or bifurcation point, then the image features of the image patch need to be input into the tracking decoder to predict the next center point.

[0035] Figure 2 An exemplary flowchart of a method 200 for training a vascular centerline tracking model according to an embodiment of this disclosure is shown. It is understood that method 200 can be performed by any suitable device with data processing capabilities, such as, but not limited to, terminal devices, processors, and servers.

[0036] Based on this, such as Figure 2 As shown, in step S201, method 200 can segment the angiography data into vessels and aneurysms to obtain segmentation results. In step S202, method 200 can extract the centerline of the vessels in the segmentation results using a skeletonization method to obtain the centerline of the vessels. In step S203, method 200 can start from the vessel starting point on the centerline and progressively segment the angiography data into blocks to obtain multiple image blocks distributed along the centerline. The center point of each image block is located on the centerline, and the partial vessels contained in each of the multiple image blocks constitute the complete vessels in the angiography data. Next, in step S204, method 200 can annotate each image block according to the centerline and the segmentation results to obtain the annotation information of each image block. This annotation information includes the block segmentation result, the vessel radius at the center point, the category of the center point, and the next center point of the center point. Finally, in step S205, method 200 can input the multiple image blocks and their annotation information as training data into the vessel centerline tracking model for training.

[0037] In step S201, method 200 can employ any one of the following segmentation methods—machine learning-based methods, threshold segmentation, and region growing—to segment the angiography data into vessels and aneurysms, thereby obtaining segmentation results. Furthermore, each part of the segmentation result can be labeled; for example, the background can be labeled 0, the vessels 1, and the aneurysm 2. Different labels enable the model to clearly distinguish between vessels and aneurysms, learning their respective characteristics and differences. For instance, vessels typically have specific morphology, orientation, and grayscale features, while aneurysms may exhibit localized abnormal enlargement or different textures. Through labeling, the model can specifically learn these features, thereby improving its ability to recognize different structures.

[0038] At the foregoing step S202, after obtaining the centerline of the blood vessel in a skeletonization manner, the centerline can also be manually corrected to ensure the accuracy and connectivity of the centerline, so as to provide accurate training data for model training, and enable the model to learn more accurate blood vessel centerline features and position information. The centerline with complete tree topology contains detailed information of the blood vessel branch and connection relationship. The model can learn these topological features during the training process, so as to better understand the structure of the blood vessel network, and help to improve the recognition and segmentation ability of the model for complex blood vessel structures.

[0039] At the foregoing step S203, the method 200 can use a preset size to gradually slice the angiography data to obtain a plurality of image blocks with the same size. It can be understood that the preset size can be selected by the person skilled in the art according to actual needs, and the disclosure does not make specific limitations on the value of the preset size, and preferably, the preset size can be (64*64*64). After the slicing of the angiography data is completed, each image block contains part of the blood vessel (for the sake of brevity, it can be referred to as a partial blood vessel). By slicing the segmentation result in the same way, the segmentation result corresponding to each image block (for the sake of brevity, it can be referred to as a block segmentation result) can be obtained. Each block segmentation result contains part of the centerline (for the sake of brevity, it can be referred to as a partial centerline).

[0040] For the sake of convenience, the operations performed by the foregoing steps S201 to S204 are exemplarily described in combination with an example. The complete angiography data DSA image is denoted as image, the blood vessels and aneurysms contained in the segmented DSA image are denoted as seg_label, the centerline is generated in a skeletonization manner, and after being manually corrected, the centerline with complete tree topology is denoted as L. Then, according to the starting point of the blood vessel of L, the DSA image is sliced according to a preset size (64*64*64), each image block is denoted as image1, the corresponding block segmentation result is denoted as seg_label1, and the size of the block segmentation result is also (64*64*64). Then, the center point of the image block is labeled, and the main labeled items include the class n of the center point, the radius r of the blood vessel at the center point, and the next center point p of the center point. If the center point of the image block is a terminal point on the centerline, n=0 can be labeled; if the center point of the image block is a connection point on the centerline, n=1 can be labeled; if the center point of the image block is a bifurcation point on the centerline, n=2 can be labeled.

[0041] In the embodiments of the present disclosure, the blood vessel refers to an intracranial blood vessel (also referred to as a cerebral blood vessel), and the maximum number of branches at the bifurcation point of the blood vessel is usually 2. In order to meet this characteristic of the cerebral blood vessel, the next center point p of the center point can be set as a list with a length of 2. In the case that the category of the center point is a terminal point, i.e., n = 0, the 0th element and the 1st element of p are both equal to 0; in the case that the category of the center point is a connection point, i.e., n = 1, the 0th element of p records the next center point, and the 1st element of p is equal to 0; in the case that the category of the center point is a bifurcation point, i.e., n = 2, the 0th element and the 1st element of p respectively record the next center points on the two branches emitted by the bifurcation point. In actual operation, the elements in p are used to record the positions of the next center points.

[0042] After completing the information labeling of each image block, the method 200 can perform the aforementioned step S205, so as to input the plurality of image blocks and the labeled information of the plurality of image blocks into the blood vessel centerline tracking model as training data, so as to train the blood vessel centerline tracking model. Specifically, the image block with the blood vessel starting point as the center point can be determined as a starting image block. Then, starting from the starting image block, the plurality of image blocks and the labeled information of the plurality of image blocks can be sequentially input into the blood vessel centerline tracking model as training data in the order of the centerline from the blood vessel starting point to the blood vessel terminal point of each image block, so as to train the blood vessel centerline tracking model.

[0043] The above is described in combination with Figure 2 The method for training the blood vessel centerline tracking model is described. By accurately labeling the center point, blood vessel radius, next center point and other information of each image block in the training data, the model can learn more detailed blood vessel morphological features, so as to more accurately track the centerline of the blood vessel in actual application, reduce the deviation and error connection, help to improve the accuracy of the blood vessel centerline tracking, in addition, the segmentation information of the blood vessel and aneurysm is contained in the training data, and the image block is labeled accordingly, the model can learn how to accurately extract the blood vessel centerline in the presence of aneurysm in the training process. This avoids the influence of large aneurysm on the extraction of the blood vessel centerline, so that the model can still stably output reliable centerline when facing diseased blood vessels, improving the overall stability of the system and the robustness to aneurysm and other abnormal conditions.

[0044] In the embodiments of the present disclosure, in order to improve the training efficiency of the blood vessel centerline tracking model and save time cost, before the angiography data is segmented into blood vessels and aneurysms, the angiography data can be first converted in format and / or adjusted in size, so as to obtain angiography data with unified data format and image size. On the premise of ensuring sufficient information, unnecessary calculation amount can be reduced, and computing resources can be more efficiently utilized to speed up the entire training process.

[0045] Furthermore, to improve the generalization ability of the vascular centerline tracking model, mathematical enhancement operations such as rotation, flipping, scaling, and noise addition can be performed on the angiography data before segmenting the vessels and aneurysms. This generates more angiography data with different shapes and conditions. This helps the vascular centerline tracking model learn more diverse features during training, thus enabling it to better handle various complex angiography data in practical applications and improving its adaptability and generalization ability to data collected from different patients and devices. In this way, when actually tracking the vascular centerline, the vascular centerline tracking model can work more stably, less affected by noise and blurring, thereby improving the reliability of the tracking results.

[0046] As described above, the vessel centerline tracking module can include an image encoder, a classifier, a tracking decoder, and a segmentation decoder. After an image patch is input into the vessel centerline tracking model, it passes through the image encoder, classifier, tracking decoder, and segmentation decoder sequentially, each performing its corresponding operation. To facilitate understanding the processing of image patches within the vessel centerline model, the following section will combine... Figure 3 The processing of image patches in a blood vessel centerline tracking model is illustrated by an example.

[0047] like Figure 3 As shown, in step S301, the image block can be input to the image encoder for feature extraction to obtain image features. Next, in step S302, the image features can be input to the classifier for category prediction to obtain the predicted category of the center point of the image block. In step S303, the image features can be input to the segmentation decoder for segmentation to obtain the predicted block segmentation result. Then, in step S304, the image features can be input to the tracking decoder for tracking prediction to obtain the predicted vessel radius at the center point of the image block and the predicted next center point. Finally, in step S305, a loss value can be determined based on the block segmentation result, the category of the center point, the next center point of the center point, the predicted block segmentation result, the predicted category, and the predicted next center point. Based on the loss value, the parameters of the image encoder, classifier, tracking decoder, and tracking decoder are updated through backpropagation.

[0048] In the embodiments of the present disclosure, based on the block segmentation result, the category of the center point, the next center point of the center point, the predicted block segmentation result, the predicted category and the predicted next center point, the loss value is determined, the Dice value of the block segmentation result and the predicted block segmentation result can be determined first, then the deviation values between the category of the center point and the predicted category and between the next center point of the center point and the predicted next center point can be determined, and then the difference value obtained by subtracting the Dice value from the deviation values is determined as the loss value. Here, the Dice value is a kind of similarity measure between the block segmentation result and the predicted block segmentation result, and the value range is between 0 and 1, and the greater the value, the higher the similarity between the block segmentation result and the predicted block segmentation result. The first deviation value between the category of the center point and the predicted category can be measured by cross entropy (Cross Entropy), and the second deviation value between the next center point of the center point and the predicted next center point can be measured by mean absolute error (Mean Absolute Error, MAE), and the sum of the first deviation value and the second deviation value is the aforementioned deviation value.

[0049] Based on the operations performed in steps S301 to S305, the simultaneous training of the tracking decoder and the segmentation decoder is realized, and the model can learn how to accurately extract the blood vessel centerline in the presence of an aneurysm during the training process. This avoids the influence of large aneurysms on blood vessel centerline extraction and radius calculation, so that the model can still stably output reliable centerlines when facing diseased blood vessels, improving the overall stability of the system and improving the robustness to abnormal conditions such as aneurysms.

[0050] For ease of understanding, the operations performed by the foregoing steps S301 to S305 are exemplarily described in combination with an example. An image encoder is denoted as Encoder, an image block is denoted as image1, an image feature obtained by the image encoder performing feature extraction on the image block is feature = Encoder(image1), a segmentation decoder is denoted as decoder_seg, a prediction block segmentation result is seg_pred = decoder_seg(feature), a classifier is denoted as clf, a prediction category is n_pred = clf(feature), a tracking decoder is denoted as decoder_track, and a prediction vessel radius and a prediction next center point (r_pred, p_pred) = decoder_track(feature) are obtained, where r_pred is a prediction vessel radius, and p_pred is a list with a length of 2; if n_pred = 0, the 0th and 1st elements of p_pred are both 0; if n_pred = 1, the 0th element of p_pred is a next center point, and the value of the 1st element is 0; if n_pred = 2, p_pred[0] and p_pred[1] are next center points on two branches, respectively. In actual operation, the elements in p record the positions of the next center points, such as coordinate values.

[0051] Next, a loss value can be determined according to the prediction result output by the vessel centerline tracking model and the annotation information. Specifically, the following formula can be used to determine the loss value Loss:

[0052] Loss1 = - Dice(seg_label1, seg_pred);

[0053] Loss2 = CrossEntroy(n, n_pred);

[0054] Loss3 = MeanAbsoluteError(p, p_pred);

[0055] Loss = Loss1 + Loss2 + Loss3.

[0056] Thereafter, based on the loss value Loss, the parameters of each module in the vessel centerline tracking model are updated by back propagation, so that one training of the model is realized. By repeating the foregoing steps S301 to S305, the model is trained multiple times until the loss value reaches an expected minimization target, and the training of the vessel centerline tracking model is determined to be completed.

[0057] It can be understood that the more components of the blood vessel centerline tracking model, the higher the time cost and computational cost required for training the model. Based on this, in some implementation scenarios, in order to reduce the training cost, the segmentation decoder described above can not be trained in the case of no aneurysm in the angiography data, only blood vessel stenosis. In this way, at the aforementioned step S305, the parameter updating operation does not need to consider the loss value caused by the block segmentation result and the predicted block segmentation result, and the loss value can be determined based on the category of the center point, the next center point of the center point, the predicted category and the predicted next center point, and the parameters of the image encoder, the classifier and the tracking decoder are updated through back propagation based on the loss value. In this implementation scenario, the following formula can be used to calculate the loss value Loss:

[0058] Loss2 = CrossEntroy (n, n_pred);

[0059] Loss3 = MeanAbsoluteError (p, p_pred);

[0060] Loss = Loss2 + Loss3.

[0061] The above describes a method for training a blood vessel centerline tracking model. Figure 2 and Figure 3 The above describes a method for training a blood vessel centerline tracking model. Figure 4 which shows an exemplary flowchart of a method 400 for tracking a blood vessel centerline according to an embodiment of the present disclosure. It can be understood that the method 400 can be performed by any appropriate device with data processing capability, for example, can include but is not limited to a processor, a terminal device and a server, etc.

[0062] As shown in Figure 4 , at step S401, the method 400 can take the starting point of the blood vessel contained in the target angiography data as a center point, and cut an image block of a preset size. At step S402, the method 400 can input the image block into the trained blood vessel centerline tracking model for centerline tracking to output the next center point of the center point. Then, at step S403, the method 400 can take the next center point as a new center point, and cut a new image block of a preset size. Further, based on the new image block, the method 400 can proceed to step S402 for continuous execution until the blood vessel contained in the target angiography data is completely traversed, i.e., no new center point is output. Finally, at step S404, the method 400 can determine the centerline of the blood vessel based on all the center points.

[0063] The foregoing step S402 can determine whether the center point of the image block is a termination point, a connection point or a bifurcation point on the center line, and can output the next center point of the connection point or the next center points on the multiple branches of the bifurcation point. In this way, the foregoing step S403 can obtain a new center point and intercept a new image block to track the complete center line along the new center point.

[0064] As described above, the blood vessel center line tracking model can include an image encoder, a classifier, a segmentation decoder and a tracking decoder. Thus, in the foregoing step S402, the image block is input into the blood vessel center line tracking model for center line tracking to output the next center point of the center point, which can specifically include the following operations: the image block is input into the image encoder for feature extraction to obtain image features; the image features are input into the classifier for category prediction to obtain the category of the center point of the image block; and the image features are input into the tracking decoder for tracking prediction to obtain the next center point of the center point. The category of the center point and the next center point of the center point are described above and will not be repeated here.

[0065] In addition, when the target angiography data contains an aneurysm, the method 400 can further perform the following operations to obtain the blood vessel and aneurysm segmentation result: input the image features into the segmentation decoder to segment the blood vessel and the aneurysm contained in the image block to obtain the blood vessel and aneurysm segmentation result. Thereafter, based on the blood vessel and aneurysm segmentation results of all image blocks, the complete blood vessel and aneurysm segmentation result of the target angiography data can be obtained.

[0066] For ease of understanding, the operations performed by the foregoing steps S401 to S404 are exemplarily described in conjunction with an example. For the target angiography data image_new, first, the starting point P0 of the blood vessel contained in the target angiography data is specified, and then the image block image1 of 64*64*64 is intercepted with the starting point P0 as the center. Next, the image block image1 is input into the image encoder Encoder to obtain the image features feature = Encoder(image1). Then, the image features pass through the segmentation decoder decoder_seg to obtain the blood vessel and aneurysm segmentation result seg_pred = decoder_seg(feature); the image features feature pass through the classifier clf to obtain the category n_pred = clf(feature); and the image features feature pass through the tracking decoder decoder_track to obtain the blood vessel radius at the center point and the next center point of the center point (r_pred, p_pred) = decoder_track(feature).

[0067] In actual operation, before the first bifurcation point appears, the branch can be recorded as Branch0, Branch0 has m center points, which can be recorded as P_0~P_m in turn, and the vessel radius at each center point can be recorded as r_0~r_m. The branches of Branch0 can be recorded as Branch0-0 and Branch0-1, Branch0-0 has n center points, which can be recorded as P_0_0~P_0_n in turn, and the vessel radius at each center point can be recorded as r_0_0~r_0_n, and the like can be obtained. Complete tree structure.

[0068] The above Figure 4 A method for tracking a blood vessel center line is described, an image block is intercepted with a blood vessel starting point as a center point, and the tracking range is gradually expanded according to the prediction of the model, so it can adapt to blood vessel structures of different sizes and shapes. Whether it is a relatively thick blood vessel or a small branch blood vessel, the center line can be accurately extracted by recursive tracking, improving the universality and applicability of the method. In addition, since each tracking iteration takes the next center point of the previous center point as a new center point, this method can ensure the connectivity of the blood vessel center line and avoid disconnected or incorrect connections. At the same time, based on the prediction ability of the training model, the topological structure integrity of the blood vessel can be better maintained.

[0069] Next, the Figure 5 An electronic device 500 provided by an embodiment of the present application is exemplarily introduced. As shown in the Figure 5 The electronic device 500 of the embodiment of the present application can include a processor 501, a memory 502, and a communication bus 503.

[0070] In the process of the specific embodiment, the above-mentioned processor 501 can be at least one of an application specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing image processing device (DSPD), a programmable logic image processing device (PLD), a field programmable gate array (FPGA), a CPU, a controller, a microcontroller, and a microprocessor. It can be understood that for different devices, the electronic devices used to realize the functions of the above-mentioned processor can also be others, and the present embodiment does not make specific limitations.

[0071] In the embodiments of the present application, the communication bus 503 is used to realize the connection communication between the processor 501 and the memory 502; the memory 502 stores program instructions for training a blood vessel centerline tracking model and / or for tracking a blood vessel centerline; and the processor 501 realizes the method for training a blood vessel centerline tracking model and / or the method for tracking a blood vessel centerline described in combination with Figure 2 and Figure 3 the description of the present application. Figure 4 the description of the present application.

[0072] The electronic device described above in combination with Figure 5 may be used to execute the program instructions for training a blood vessel centerline tracking model and / or for tracking a blood vessel centerline of the present application. It should be understood that the device structure or architecture herein is only exemplary, and the implementation manner and implementation entity of the present application are not limited thereto, but can be changed without departing from the spirit of the present application. It can be understood that the description of various embodiments of the present disclosure emphasizes the differences between various embodiments, and the same or corresponding parts can be referred to each other. For the purpose of brevity, the present disclosure will not be described one by one.

[0073] According to the above description in combination with the drawings, those skilled in the art can also understand that the embodiments of the present application can also be realized by a software program. Therefore, the present application also provides a computer readable storage medium. The computer readable storage medium stores program instructions for training a blood vessel centerline tracking model and / or for tracking a blood vessel centerline, which can be used to realize the method for training a blood vessel centerline tracking model and / or the method for tracking a blood vessel centerline described in combination with Figure 2 and Figure 3 the description of the present application. Figure 4 the description of the present application.

[0074] It should be noted that although the operations of the method of the present application are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all of the shown operations must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart can change the order of execution. Additionally or alternatively, some steps can be omitted, a plurality of steps can be combined into one step, and / or one step can be divided into a plurality of steps.

[0075] While several embodiments of the application have been shown and described herein, it will be obvious to those skilled in the art that many changes, modifications, and substitutions can be made to the embodiments without departing from the spirit and scope of the application. It is to be understood that various alternatives to the embodiments of the application described herein can be employed in practicing the application. The appended claims are intended to cover all such alternatives as would be included within the spirit and scope of the application.

Claims

1. A method for training a blood vessel centerline tracking model, the method comprising: segmenting blood vessels and aneurysms from angiography data to obtain a segmentation result; extracting a centerline from the blood vessels in the segmentation result to obtain a centerline of the blood vessels; stepwise slicing the angiography data from a starting point of the blood vessels on the centerline to obtain a plurality of image blocks distributed along the centerline, a center point of each image block being on the centerline, and the blood vessels included in each of the plurality of image blocks constituting the blood vessels; annotating information of each image block according to the centerline and the segmentation result to obtain annotation information of each image block, the annotation information including a block segmentation result, a blood vessel radius at the center point, a category of the center point, and a next center point of the center point; inputting the plurality of image blocks and the annotation information of the plurality of image blocks as training data into the blood vessel centerline tracking model to train the blood vessel centerline tracking model; wherein the blood vessel centerline tracking model comprises an image encoder, a classifier, and a tracking decoder, and inputting the image blocks and the annotation information of the image blocks as training data into the blood vessel centerline tracking model to train the blood vessel centerline tracking model comprises: inputting the image blocks into the image encoder to extract image features to obtain image features; inputting the image features into the classifier to predict a category of the center point of the image blocks to obtain a predicted category of the center point of the image blocks; inputting the image features into the tracking decoder to predict a blood vessel radius at the center point of the image blocks and a predicted next center point of the center point to obtain a predicted blood vessel radius at the center point of the image blocks and a predicted next center point of the center point; determining a loss value based on the category of the center point, the next center point of the center point, the predicted category, and the predicted next center point, and updating parameters of the image encoder, the classifier, and the tracking decoder based on the loss value.

2. The method of claim 1, wherein, inputting the plurality of image blocks and the annotation information of the plurality of image blocks as training data into the blood vessel centerline tracking model to train the blood vessel centerline tracking model comprises: determining an image block with the starting point of the blood vessels as the center point as a starting image block; starting from the starting image block, sequentially inputting the plurality of image blocks and the annotation information of the plurality of image blocks as training data into the blood vessel centerline tracking model along the centerline to train the blood vessel centerline tracking model.

3. The method of claim 1, wherein, The blood vessel centerline tracking model further comprises a segmentation decoder, and the method further comprises: after obtaining the image features, inputting the image features into the segmentation decoder to perform segmentation to obtain a predicted block segmentation result; determining a loss value based on the block segmentation result, the category of the center point, the next center point of the center point, the predicted block segmentation result, the predicted category, and the predicted next center point, and updating parameters of the image encoder, the classifier, the tracking decoder, and the tracking decoder based on the loss value.

4. The method of any of claims 1-3, wherein, The categories include a termination point, a connection point, and a bifurcation point; and in a case where the category of the center point is a termination point, the center point has no next center point. In a case where the category of the center point is a connection point, the number of next center points of the center point is 1; In a case where the category of the center point is a branch point, the number of next center points of the center point is the number of branches emitted by the branch point. 5.A method for tracking a blood vessel centerline, comprising: taking an image block of a preset size with a starting point of a blood vessel contained in target angiography data as a center point; inputting the image block into a blood vessel centerline tracking model trained according to the method of any one of claims 1-4 to track a centerline, to output a next center point of the center point; taking the next center point as a new center point, and taking a new image block of the preset size; continuing to input the new image block into the blood vessel centerline tracking model trained according to the method of any one of claims 1-4 to track a centerline, to output a next center point of the center point, until no new center point is outputted; determining a centerline of the blood vessel based on all the center points.

6. The method of claim 5, wherein, The blood vessel centerline tracking model comprises an image encoder, a classifier and a tracking decoder; inputting the image block into the blood vessel centerline tracking model to track a centerline, to output a next center point of the center point, comprises: inputting the image block into the image encoder to extract features, to obtain image features; inputting the image features into the classifier to predict a category, to obtain a category of the center point of the image block; inputting the image features into the tracking decoder to predict a next center point, to obtain a next center point of the center point.

7. The method of claim 6, wherein, The category comprises a termination point, a connection point and a branch point; and In a case where the category of the center point is a termination point, the center point has no next center point; In a case where the category of the center point is a connection point, the number of next center points of the center point is 1; In a case where the category of the center point is a branch point, the number of next center points of the center point is the number of branches emitted by the branch point. 8.An electronic device, comprising: a processor; and a memory storing program instructions for training a blood vessel centerline tracking model and / or for tracking a blood vessel centerline, which when executed by the processor, cause the implementation of the method of any one of claims 1-4 and / or the method of any one of claims 5-7. 9.A computer readable storage medium storing program instructions for training a blood vessel centerline tracking model and / or for tracking a blood vessel centerline, which when executed by a processor, cause the implementation of the method of any one of claims 1-4 and / or the method of any one of claims 5-7.

Images