Blood vessel segmentation model, model training method and device, equipment and storage medium

Abstract

The invention relates to the technical field of medical image processing, and discloses a blood vessel segmentation model, a model training method and device, equipment and a storage medium, and the blood vessel segmentation model comprises an encoder and a decoder. The encoder comprises a plurality of stages of down-sampling modules, and the decoder correspondingly comprises a plurality of stages of up-sampling modules; each level of down-sampling module comprises a multi-scale local form attention module; the multi-scale local form attention module comprises a plurality of different convolution modules, and the different convolution modules have different convolution kernel sizes and are used for respectively extracting blood vessel form characteristics under different scales; the multi-scale local form attention module further comprises a splicing module which is used for splicing the blood vessel form characteristics under different scales. And an output layer in the decoder carries out blood vessel segmentation based on the blood vessel morphological characteristics under different scales. According to the invention, the segmentation accuracy of the blood vessel with a complex morphological structure can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a blood vessel segmentation model, a model training method, a device, equipment and a storage medium. Background Art

[0002] At present, the vascular segmentation technology in medical image analysis has made significant progress and has been widely used in disease diagnosis, treatment planning, and vascular imaging research. Traditional vascular segmentation methods mainly rely on manually designed features or threshold processing, but these methods are difficult to adapt to the complex changes of different vascular morphologies and are easily affected by noise and image blur. In recent years, deep learning technology has brought new ideas and methods to vascular segmentation. Vascular segmentation methods based on convolutional neural networks, such as U-Net and improved network structures, can learn richer and more detailed image features and achieve higher segmentation accuracy. However, due to the complex and changeable vascular structure, the accuracy of vascular segmentation based on deep learning technology is still not high. Summary of the invention

[0003] In view of this, the present invention provides a blood vessel segmentation model, a model training method, an apparatus, a device and a storage medium to solve the problem of low blood vessel segmentation accuracy.

[0004] In a first aspect, the present invention provides a blood vessel segmentation model, comprising: an encoder and a decoder;

[0005] The encoder includes a multi-stage downsampling module, and the decoder correspondingly includes a multi-stage upsampling module;

[0006] Each level of the downsampling module includes a multi-scale local morphology attention module;

[0007] The multi-scale local morphological attention module includes a plurality of different convolution modules, and the convolution kernels of different convolution modules are different in size, so as to respectively extract the vascular morphological features at different scales;

[0008] The multi-scale local morphological attention module further includes a splicing module for splicing the vascular morphological features at different scales;

[0009] The output layer in the decoder performs blood vessel segmentation based on the blood vessel morphological features at different scales.

[0010] In an optional embodiment, the multi-scale local morphology attention module further includes a depthwise separable convolution module, and the output of the depthwise separable convolution module is used as the output of the multiple different convolution modules;

[0011] and / or,

[0012] The multi-scale local morphology attention module also includes a normalization layer and a multi-layer perceptron, and the splicing module, the normalization layer and the multi-layer perceptron are connected in sequence.

[0013] In a second aspect, the present invention provides a method for training a blood vessel segmentation model, the method comprising:

[0014] Obtain blood vessel segmentation training samples and corresponding true labels;

[0015] The pre-constructed initial blood vessel segmentation model is trained using the blood vessel segmentation training samples and the true labels to obtain the blood vessel segmentation model of the first aspect or any corresponding embodiment thereof.

[0016] In an optional implementation, the using the blood vessel segmentation training samples and the true labels to train a pre-built initial blood vessel segmentation model includes:

[0017] During the training process, the predicted blood vessel segmentation labels predicted by the initial blood vessel segmentation model and the corresponding real blood vessel segmentation labels are obtained;

[0018] Extracting a first blood vessel skeleton of a target blood vessel branch from the predicted blood vessel segmentation label, and extracting a second blood vessel skeleton of the target blood vessel branch from the real blood vessel segmentation label;

[0019] Determining a global morphology preservation loss based on the first blood vessel skeleton, the second blood vessel skeleton, the predicted blood vessel segmentation label, and the true blood vessel segmentation label;

[0020] Obtaining the segmentation prediction loss of the initial blood vessel segmentation model;

[0021] Based on the global morphology preservation loss and the segmentation prediction loss, a total loss of the initial blood vessel segmentation model is determined.

[0022] In an optional implementation, determining the global morphology preservation loss based on the first blood vessel skeleton, the second blood vessel skeleton, the predicted blood vessel segmentation label, and the real blood vessel segmentation label includes:

[0023] Determining the topological accuracy of the target blood vessel branch based on the real blood vessel segmentation label and the first blood vessel skeleton;

[0024] Determining a topological sensitivity of the target blood vessel branch based on the predicted blood vessel segmentation label and the second blood vessel skeleton;

[0025] The global morphology preserving loss is determined based on the topology accuracy and the topology sensitivity.

[0026] In an optional implementation, the topological accuracy is the proportion of the portion of the first blood vessel skeleton located within the real blood vessel segmentation label to the first blood vessel skeleton;

[0027] And / or, the topological sensitivity is the proportion of the portion of the second blood vessel skeleton located within the predicted blood vessel segmentation label to the second blood vessel skeleton.

[0028] In an optional implementation, determining the global morphology preservation loss based on the topology accuracy and the topology sensitivity includes:

[0029] Obtaining a harmonic mean of the topology accuracy and the topology sensitivity;

[0030] The global shape preserving loss is determined based on the harmonic mean.

[0031] In an optional implementation, the real label includes a precise label and a noise label; the labeling quality of the precise label is higher than the labeling quality of the noise label;

[0032] The using the blood vessel segmentation training samples and the true labels to train the pre-built initial blood vessel segmentation model includes:

[0033] Acquire the initial blood vessel segmentation model established in advance based on the average teacher framework, wherein the initial blood vessel segmentation model includes a teacher model and a student model;

[0034] Input the noise label and the corresponding blood vessel segmentation training sample into the teacher model and the student model respectively, and obtain a first blood vessel segmentation prediction label output by the teacher and a second blood vessel segmentation prediction label output by the student model;

[0035] Based on the first blood vessel segmentation prediction label and the second blood vessel segmentation prediction label, optimizing the noise label to obtain an optimized label;

[0036] Calculating the segmentation prediction loss of the student model based on the optimized label and the second blood vessel segmentation prediction label;

[0037] Adjusting a weight parameter of the student model based on the segmentation prediction loss of the student model;

[0038] Based on the adjusted weight parameters of the student model, the weight parameters of the teacher model are adjusted.

[0039] In an optional implementation, the optimizing the noise label based on the first blood vessel segmentation prediction label and the second blood vessel segmentation prediction label to obtain an optimized label includes:

[0040] Obtaining pixel-by-pixel KL divergence of the first blood vessel segmentation prediction label and the second blood vessel segmentation prediction label on two categories to obtain a segmented blood vessel uncertainty map and a background uncertainty map, wherein the two categories include segmented blood vessels and background;

[0041] The optimized label is obtained based on the segmented blood vessel uncertainty map and the background uncertainty map, as well as the noise label.

[0042] In an optional implementation, obtaining the optimized label based on the segmented blood vessel uncertainty map and the background uncertainty map, and the noise label, includes:

[0043] Obtaining a reliable label for each pixel based on the segmented blood vessel uncertainty map and the background uncertainty map;

[0044] The optimized label is obtained based on the reliable label and the noise label.

[0045] In an optional implementation, obtaining a reliable label of each pixel based on the segmented blood vessel uncertainty map and the background uncertainty map includes:

[0046] Using a negative exponential function, converting the segmented blood vessel uncertainty map and the background uncertainty map into corresponding score values;

[0047] The reliable label is determined using the score value.

[0048] In an optional implementation, the calculating the segmentation prediction loss of the student model based on the optimized label and the second blood vessel segmentation prediction label includes:

[0049] Acquire an uncertainty weight value and / or a regularization value of each pixel, wherein the uncertainty weight value and the regularization value are determined based on the segmented blood vessel uncertainty map;

[0050] Based on the uncertainty weight value and / or the regularization value, the segmentation prediction loss of each pixel of the student model is determined.

[0051] In a third aspect, the present invention provides a training device for a blood vessel segmentation model, the device comprising:

[0052] A sample data acquisition module is used to obtain blood vessel segmentation training samples and corresponding true labels;

[0053] A training module is used to train a pre-constructed initial blood vessel segmentation model using the blood vessel segmentation training samples and the true labels to obtain the blood vessel segmentation model of the first aspect or any corresponding embodiment thereof.

[0054] In a fourth aspect, the present invention provides a computer device, comprising: a memory and a processor, the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the training method of the blood vessel segmentation model of the above-mentioned second aspect or any corresponding embodiment thereof by executing the computer instructions.

[0055] In a fifth aspect, the present invention provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to enable a computer to execute the training method for a blood vessel segmentation model of the second aspect or any corresponding embodiment thereof.

[0056] In a sixth aspect, the present invention provides a computer program product, comprising computer instructions for causing a computer to execute the method for training a blood vessel segmentation model according to the second aspect or any corresponding embodiment thereof.

[0057] The embodiments of the present invention provide a blood vessel segmentation model, a model training method, an apparatus, a device and a storage medium, wherein a multi-scale local morphological attention module in the blood vessel segmentation model can establish a connection between local shape features and global context, which helps the model understand the relationship between pixels and their surroundings, thereby better capturing the blood vessel structure and improving the accuracy of blood vessel segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the related technologies, the drawings required for use in the specific embodiments or the related technical descriptions will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0059] Figure 1 It is a schematic diagram of high-quality labeled data;

[0060] Figure 2 is a schematic diagram of data with noisy annotations;

[0061] Figure 3 is a schematic diagram of the architecture of a blood vessel segmentation model according to an embodiment of the present invention;

[0062] Figure 4 is a schematic diagram of the architecture of a multi-scale local morphological attention module according to an embodiment of the present invention;

[0063] Figure 5 is a flow chart of a method for training a blood vessel segmentation model according to an embodiment of the present invention;

[0064] Figure 6is a flowchart of another model training method according to an embodiment of the present invention;

[0065] Figure 7 is a structural block diagram of a training device for a blood vessel segmentation model according to an embodiment of the present invention;

[0066] Figure 8 It is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0068] Early vascular segmentation methods mainly used classical image analysis techniques, such as using filtering techniques to enhance vascular regions, and adopting methods such as region growing algorithms and graph cuts to segment blood vessels. However, these methods require complex parameter adjustments and are not robust enough when dealing with images with high noise levels or non-uniform vascular intensity. Recently, deep learning has made significant progress in the field of medical image segmentation. Existing deep learning methods for vascular segmentation can be roughly divided into three categories: network structure-based methods, feature extraction module-based methods, and loss function-based methods.

[0069] The network structure-based methods design specific network architectures according to the characteristics of tubular structures to guide the model to focus on key features. FR-UNet extracts multi-scale contextual information for vascular segmentation through multi-resolution convolution interactions and feature aggregation modules. Some related technologies introduce structural priors into the vascular segmentation network to improve segmentation accuracy through shared encoder and dual decoder architectures. Some related technologies propose the Examinee-Examiner Network (EE-Net) for automatic coronary artery lumen segmentation. EE-Net solves the segmentation interruption problem caused by stenosis by combining the semantic features of the lumen and the geometric constraints of the centerline obtained from the continuous. Some related technologies propose a cross transformer network (CTN) to optimize vascular segmentation, aiming to solve the problem of discontinuous segmentation results of small tubular structures common in the U-Net method. MTCL uses a dual-branch network based on an average teacher model, aiming to solve the problem of high-quality annotated (also called labeled) data (for example, such as Figure 1 The 3DIRCADb dataset shown in the figure has a high-quality labeled data set) and low-quality noisy labeled data (e.g. Figure 2However, due to the small proportion of tubular structures, the network may inevitably lose the perception of these structures.

[0070] Feature extraction-based methods enhance the understanding of specific geometric and topological features of tubular structures by supplementing the model with additional feature representations. Ce-Net uses dense atrous convolution layers to expand the receptive field. Inspired by deformable convolutions, some related techniques introduce deformable offsets into traditional convolutions, named dynamic snake convolution (DSConv). DSConv accurately captures the characteristics of tubular structures by adaptively focusing on curved and tortuous local structures. However, some redundant feature representations increase the computational burden and do not have a positive impact on the model.

[0071] The method of improving the loss function adds specific quantization term constraints during the training process through the loss function. Some related technologies use the inverse value of the normalized distance map to weight the voxels in the reference segmentation, and propose the radial distance loss (RDLoss), which successfully improves the segmentation performance of smaller blood vessels in the pulmonary vasculature. Some related technologies propose the clDiceLoss loss function to improve the connectivity problem in tubular structure segmentation. clDice includes a loss function and an indicator for measuring the connectivity of thin structure segmentation. It effectively incorporates the segmented skeleton into the dice calculation. While achieving architecture-independent, topology-aware segmentation of thin tubular structures, it also brings significant computational overhead.

[0072] Although these deep learning-based vascular segmentation methods have made significant progress, they still face many challenges in practical applications. In particular, the performance of existing methods is still insufficient when dealing with complex vascular morphology and noisy labels with low-quality annotations. These challenges are specifically reflected in the following two aspects: For vascular segmentation tasks, the current deep learning-based vascular segmentation methods face two major challenges: 1. Multiple morphological features of blood vessels. The local morphological features of blood vessels show three modes at different scales: large-scale serpentine, medium-scale tubular, and small-scale block. In terms of global morphological features, blood vessels present a tree-like structure with different diameters. The size of tree-like blood vessels varies significantly, and the contrast of small blood vessels is low, and traditional methods are prone to lose their spatial information. The lack of small blood vessels in the segmentation results will change the tree-like structure of blood vessels, which is a major challenge in measuring segmentation results using traditional Dice indicators. 2. Noise labels. Noise labels are defined in the embodiments of the present invention as inaccurate or incomplete labels caused by uncertainty or human error in the annotation process. Most deep learning-based segmentation methods rely on a large amount of high-quality annotation data. However, the complex morphological structure makes pixel-level annotation of blood vessels time-consuming and requires professional knowledge. In practice, it is often unaffordable to obtain a large amount of high-quality annotated data from experts. To this end, additional data with different label qualities is often collected for training, such as crowdsourcing from non-experts or using labels generated by semi-automatic methods without quality control. However, these labels are usually incomplete, which are referred to as noisy labels in the embodiments of the present invention. Figure 2 As shown in Figure 2, directly using additional data with noisy labels may mislead network optimization. This is because the high gradient of the noise during back propagation hinders the network from converging to the optimal solution.

[0073] The embodiment of the present invention provides a blood vessel segmentation model, including: an encoder and a decoder;

[0074] The encoder includes a multi-stage downsampling module (eg Figure 3 The decoder includes a four-stage downsampling module used in the downsampling stages 1 to 4 shown in FIG. 1 , and the decoder correspondingly includes a multi-stage upsampling module;

[0075] Each level of the downsampling module includes a multi-scale local morphology attention module;

[0076] The multi-scale local morphological attention module includes a plurality of different convolution modules, and the convolution kernels of different convolution modules are different in size, and are used to extract the vascular morphological features at different scales respectively; specifically, Figure 4As shown in the figure, there are three different convolution modules: snake convolution module, linear convolution module and block convolution module. In addition, different convolution modules can have different forms in addition to different convolution kernel sizes, such as snake convolution, dilated convolution and conventional matrix convolution.

[0077] The multi-scale local morphological attention module further includes a splicing module for splicing the vascular morphological features at different scales;

[0078] The output layer in the decoder performs blood vessel segmentation based on the blood vessel morphological features at different scales.

[0079] The vascular segmentation model provided by the embodiment of the present invention is a Hierarchical Vascular Morphology-Aware Net (HVMA-Net), which adopts an encoder-decoder architecture.

[0080] Given that, at the branch level, the local morphology of blood vessels is serpentine; at a smaller scale, it is linear; at a very small scale, it is blocky, the embodiment of the present invention specifically designs a multi-scale local morphology attention module according to the morphology of blood vessels. Specifically, three different convolution schemes are used according to the three different morphological modes of block, linear and serpentine in blood vessels. This corresponds to the morphological features of blood vessels at three different scales. This special multi-scale design enables the model to better extract vascular structures with different morphologies.

[0081] The multi-scale local morphological attention module, the output is divided into multiple heads along the channel dimension, and the features are decomposed into three parallel branches operating along the channel dimension to enhance the representation ability of the model.

[0082] In the blood vessel segmentation model provided by the embodiment of the present invention, the multi-scale local morphological attention module integrates multi-mode convolution blocks to capture complex multi-scale local blood vessels, and can establish connections between local shape features and global context, which helps the model understand the relationship between pixels and their surroundings, thereby better capturing the blood vessel structure and improving the accuracy of blood vessel segmentation.

[0083] In some optional specific implementations, the multi-scale local morphology attention module further includes a depthwise separable convolution (DWConv) module, and the output of the depthwise separable convolution module is used as the output of the multiple different convolution modules;

[0084] and / or,

[0085] The multi-scale local morphology attention module also includes a normalization layer and a multi-layer perceptron, and the splicing module, the normalization layer and the multi-layer perceptron are connected in sequence.

[0086] Specifically, Figure 4 As shown, the embodiment of the present invention first embeds the input feature map through a deep separable convolution. Then, the output of the deep separable convolution (DWConv) module is divided into three groups along the channel dimension: Where X represents the input of the multi-scale local morphological attention module, X block , X linear , X snake Represent the inputs of different vascular morphological feature extraction branches, and c represents the number of channels of each branch. The grouped inputs are then processed by three parallel morphological pattern extraction branches with three different convolutions, including block, linear, and serpentine. These different convolution methods enable capturing local features across different ranges:

[0087]

[0088] Finally, the concatenation module is used to connect the output of each branch to form the spatial attention weight, that is, X' = Concat(X' block ,X' linear ,X' snake ), which is used to weight different regions of the input feature map. These weighted features are then normalized and multilayer perceptron (MLP) to fuse the three morphological features to produce the final output of the multi-scale local morphological attention module. By doing so, the multi-scale local morphological attention module can establish a connection between local shape features and global context. This helps the vessel segmentation model understand the relationship between pixels and their surroundings, thereby better capturing the vessel structure.

[0089] According to an embodiment of the present invention, an embodiment of a training method for a blood vessel segmentation model is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of executable computer instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0090] In this embodiment, a training method for a blood vessel segmentation model is provided, which can be used in various computer devices. Figure 5 FIG. 4 is a flow chart of a method for training a blood vessel segmentation model according to an embodiment of the present invention. Figure 5 As shown, the process includes the following steps:

[0091] Step S501, obtaining blood vessel segmentation training samples and corresponding true labels.

[0092] Specifically, the blood vessel segmentation training sample is a medical image containing a blood vessel structure, such as a computed tomography (CT) image. The blood vessel segmentation training sample may also include a medical image sample obtained by adding random Gaussian noise η to the medical image containing the blood vessel structure.

[0093] The real images corresponding to each blood vessel segmentation training sample can be manually annotated or automatically annotated by a machine. The real labels automatically annotated by the machine can be manually reviewed. The automatic annotation of real labels uses other image recognition technologies other than the blood vessel segmentation model in the embodiment of the present invention.

[0094] Step S502: Using the blood vessel segmentation training samples and the true labels, a pre-constructed initial blood vessel segmentation model is trained to obtain any one of the blood vessel segmentation models in the above embodiments.

[0095] The specific training method may be a supervised training method in the related art, or may be a training method described below.

[0096] In some optional specific implementations, step S502, i.e., using the blood vessel segmentation training samples and the true labels to train the pre-constructed initial blood vessel segmentation model, includes:

[0097] Step S5021: during the training process, obtain the predicted blood vessel segmentation label M predicted by the initial blood vessel segmentation model P , and the corresponding true blood vessel segmentation label M GT .

[0098] Step S5022: from the predicted blood vessel segmentation label M P Extract the first vascular skeleton B of the target vascular branch P , from the true blood vessel segmentation label M GT Extract the second blood vessel skeleton B of the target blood vessel branch GT .

[0099] Specifically, a differentiable branch skeleton extraction method can be used to efficiently extract the skeleton of branch vessels. The specific calculation formula is:

[0100] B=ReLU(M-maxpool(minpool(M,s),s))

[0101] Among them, M is the predicted blood vessel segmentation label M P Or the true blood vessel segmentation label M GT , accordingly, B is the first vascular skeleton B P Or the second vascular skeleton B GT, s is the scale of the target blood vessel branch, minpool is the minimum pooling, and maxpool is the maximum pooling.

[0102] Step S5023: determining a global morphology preserving loss (GMPLoss) based on the first blood vessel skeleton, the second blood vessel skeleton, the predicted blood vessel segmentation label, and the true blood vessel segmentation label.

[0103] In some optional specific implementations, step S5023, i.e., determining the global morphology preservation loss based on the first blood vessel skeleton, the second blood vessel skeleton, the predicted blood vessel segmentation label, and the real blood vessel segmentation label, includes:

[0104] Step S50231: determining the topological accuracy of the target blood vessel branch based on the real blood vessel segmentation label and the first blood vessel skeleton.

[0105] Specifically, the topological accuracy T prec (B P ,M GT ) is the first vascular skeleton B P The real blood vessel segmentation label M GT The inner part occupies the first blood vessel skeleton B P The first vascular skeleton B P The real blood vessel segmentation label M GT The inner part, that is, the first vascular skeleton B P and the true blood vessel segmentation label M GT The overlapping part of the topological accuracy T prec (B P ,M GT ) is calculated as:

[0106]

[0107] Step S50232: based on the predicted blood vessel segmentation label M P and the second vascular skeleton B GT , determine the topological sensitivity T of the target vascular branch sens (B GT ,M P ).

[0108] Specifically, the topological sensitivity is the second vascular skeleton B GT The predicted vessel segmentation label M is located in P The proportion of the second vascular skeleton to the second vascular skeleton B GT The predicted vessel segmentation label M is located in P The inner part, that is, the second vascular skeleton B GTand predict the vessel segmentation label M P Overlapping part. Topological sensitivity T sens (B GT ,M P ) is calculated as:

[0109]

[0110] Step S50233: determining the global morphology preservation loss based on the topology accuracy and the topology sensitivity.

[0111] In some optional specific implementations, step S50233, determining the global morphology preservation loss based on the topology accuracy and the topology sensitivity, includes:

[0112] Step S502331, obtain the harmonic mean (F1) BranchDice of the topology accuracy and the topology sensitivity. The calculation formula of the harmonic mean BranchDice is:

[0113]

[0114] Step S502332: Determine the global shape preserving loss GMPLoss based on the harmonic mean. Specifically, GMPLoss=1-BranchDice.

[0115] The commonly used loss function for image segmentation is the Dice loss, which relies on the average voxel-level hit or miss prediction. In blood vessel segmentation, a continuous sequence of correct voxel predictions is more meaningful than some discontinuous correct prediction regions, especially for thin blood vessels, which are often ignored by the Dice loss. And it is obvious that the offset of a single voxel in the prediction will change the topological structure of the entire network. Therefore, the embodiment of the present invention proposes a global morphology preserving loss, which calculates the intersection (overlap) of the segmentation label and its morphological skeleton for branch blood vessels, i.e., thin blood vessels to maintain the connectivity of the blood vessels, effectively guiding the blood vessel segmentation model to learn tubular structures. That is, the embodiment of the present invention uses the global morphology preserving loss to focus on the thin blood vessels that are often ignored by the Dice loss, thereby ensuring the complete segmentation of the branch blood vessels.

[0116] Step S5024: obtaining the segmentation prediction loss of the initial blood vessel segmentation model.

[0117] Step S5025: determining a total loss of the initial blood vessel segmentation model based on the global morphology preservation loss and the segmentation prediction loss.

[0118] In an embodiment of the present invention, the target vascular branch may be one or more. In the case where there are multiple target vascular branches, the global morphology preservation loss needs to be calculated for each target vascular branch separately, and all global morphology preservation losses need to be accumulated, such as directly adding or weighted summing. The weight value in the weighted summing can be determined according to the level and / or importance of the target vascular branch.

[0119] The segmentation prediction loss can be Dice loss, cross-entropy loss, or a composite loss function that combines Dice loss and cross-entropy loss, namely DiceCE loss.

[0120] The total loss of the initial blood vessel segmentation model can be obtained by weighted summing the global morphology preservation loss and the segmentation prediction loss. Taking the segmentation prediction loss as the DiceCE loss as an example, the calculation formula for the total loss of the initial blood vessel segmentation model is:

[0121] l seg =α seg ×GMPLoss(P hq ,L hq )+(1-α seg )×DiceCELoss(P hq ,L hq )

[0122] Among them, l seg is the total loss of the initial vessel segmentation model, which is the high-quality supervised training loss; α seg is the weight value, GMPLoss is the global shape preservation loss, DiceCELoss is the DiceCE loss, P hq To predict the vessel segmentation label, L hq is the real blood vessel segmentation label.

[0123] In the embodiment of the present invention, the global morphology preservation loss and the segmentation prediction loss are combined to form the final high-quality supervised training loss l seg , l seg It supplements the topology preservation capability that is lacking in the Dice-based loss function, and emphasizes the importance of branched vessels in maintaining vascular morphology, thereby ensuring that the vascular segmentation model (HVMA-Net) proposed in the embodiment of the present invention can segment the complete vascular structure.

[0124] In some optional specific implementations, the real label includes a precise label and a noise label; the annotation quality of the precise label is higher than the annotation quality of the noise label. Specifically, the precise label can be a label obtained by manual and fine annotation by a professional, and the noise label can be, for example, a label automatically annotated by a machine, a label obtained by manual annotation by a non-professional, or a label roughly annotated by a professional. The precise label and the noise label can correspond to the same blood vessel segmentation training sample, that is, for the same blood vessel segmentation training sample, the precise label and the noise label can be annotated at the same time.

[0125] Step S502, i.e., using the blood vessel segmentation training samples and the true labels to train the pre-built initial blood vessel segmentation model, includes:

[0126] Step S502a, obtaining the initial blood vessel segmentation model established in advance based on the average teacher framework, wherein the initial blood vessel segmentation model includes a teacher model and a student model.

[0127] Step S502b, inputting the noise label and the corresponding blood vessel segmentation training sample into the teacher model and the student model respectively, obtaining a first blood vessel segmentation prediction label output by the teacher and a second blood vessel segmentation prediction label output by the student model.

[0128] Specifically, P noisy =F student (I noisy ), P perturb =F teacher (I noisy )

[0129] Among them, P noisy The second vessel segmentation prediction label output by the student model, F student For student models, I noisy is a blood vessel segmentation training sample, and the blood vessel segmentation true label corresponding to the blood vessel segmentation training sample is a noise label; P perturb is the first vessel segmentation prediction label output by the teacher, F teacher Model for teachers.

[0130] Step S502c: optimizing the noise label based on the first blood vessel segmentation prediction label and the second blood vessel segmentation prediction label to obtain an optimized label.

[0131] In some optional specific implementations, step S502c, i.e., optimizing the noise label based on the first blood vessel segmentation prediction label and the second blood vessel segmentation prediction label to obtain an optimized label, includes:

[0132] Step S502c1, as Figure 6 As shown, the pixel-by-pixel KL divergence (Kullback-Leibler Divergence) of the first blood vessel segmentation prediction label and the second blood vessel segmentation prediction label in two categories is obtained to obtain a segmented blood vessel uncertainty map and a background uncertainty map, and the two categories include segmented blood vessels and background.

[0133] Specifically, the segmented vessel uncertainty map U seg For: U seg =KLdiv(P noise [1,...],P perturb [1,...]); Background uncertainty map U bg For: U bg =KLdiv(P noise [0,...],P perturb [0,...]). Among them, P noise [1,...] is the prediction result of category 1 (blood vessel) output by the student model, P perturb [1,...] is the prediction result of category 1 (blood vessel) output by the teacher model, P noise [0,...] is the prediction result of category 0 (background) output by the student model, P perturb [0,...] is the prediction result of category 0 (background) output by the teacher model.

[0134] Step S502c2, obtaining the optimized label based on the segmented blood vessel uncertainty map and the background uncertainty map, and the noise label.

[0135] In some optional specific implementations, step S502c2, i.e., obtaining the optimized label based on the segmented blood vessel uncertainty map and the background uncertainty map, and the noise label, includes:

[0136] Step S502c21, obtaining a reliable label of each pixel based on the segmented blood vessel uncertainty map and the background uncertainty map.

[0137] In some optional specific implementations, step S502c21, i.e., obtaining a reliable label of each pixel based on the segmented blood vessel uncertainty map and the background uncertainty map, includes:

[0138] Step S502c211: using a negative exponential function, convert the segmented blood vessel uncertainty map and the background uncertainty map into corresponding score values.

[0139] Specifically, a negative exponential function may be used to map the segmented blood vessel uncertainty map and the background uncertainty map to score values ​​within the range of [0, 1].

[0140] Step S502c212, using the score value, determine the reliable label M reliable .

[0141] Specifically, the reliable label M of each pixel is determined as follows: reliable :

[0142]

[0143] in, is the score value corresponding to the segmented vessel uncertainty map, is the score value corresponding to the background uncertainty map (i.e., background uncertainty value), α seg and α bg These are the preset thresholds corresponding to segmenting blood vessels and background respectively.

[0144] Step S502c22: Based on the reliable label and the noise label, the optimized label is obtained. Specifically, the noise label L can be fused in the following manner: noisy and reliable label M reliable , get the optimized label L refined :

[0145]

[0146] Step S502d: calculating the segmentation prediction loss of the student model based on the optimized label and the second blood vessel segmentation prediction label.

[0147] In some optional specific implementations, step S502d, namely, calculating the segmentation prediction loss of the student model based on the optimized label and the second blood vessel segmentation prediction label, includes:

[0148] Step S502d1, obtaining the uncertainty weight value and / or regularization value of each pixel, wherein the uncertainty weight value and the regularization value are determined based on the segmented blood vessel uncertainty map.

[0149] Specifically, the uncertainty weight value of each pixel can be obtained using a negative exponential function, that is, a negative exponential function of the segmented vessel uncertainty map, that is, a score value in the range of [0,1] corresponding to the segmented vessel uncertainty map. The regularization value can be the absolute value of the segmented vessel uncertainty map.

[0150] Step S502d2: Determine the segmentation prediction loss of each pixel of the student model based on the uncertainty weight value and / or the regularization value.

[0151] The segmentation prediction loss can be Dice loss, cross-entropy loss, or a composite loss function that combines Dice loss and cross-entropy loss, namely DiceCE loss.

[0152] Taking the segmentation prediction loss as DiceCE loss as an example, the segmentation prediction loss of each pixel of the student model is:

[0153]

[0154] In the embodiment of the present invention, the segmentation prediction loss of the student model is determined by the second blood vessel segmentation prediction label P output by the student model. noisy And the optimized label is calculated, and based on the segmented vessel uncertainty map U seg An additional regularization term is added to ensure the consistency between the predictions of the teacher module and the student model. noisy and P perturb The difference between them further extracts the potential information in the image. In addition, for the segmentation prediction loss of each pixel, the weight value based on the uncertainty of the segmented blood vessels is multiplied during the loss calculation process, so that pixels with higher uncertainty are assigned smaller weight values, so that pixels with higher uncertainty contribute less to the loss, thereby reducing their impact on the final model training, and pixels with lower uncertainty are assigned larger weight values, thereby increasing their impact on the model training.

[0155] In addition, the final loss of the student model can also be recorded as above, integrating the global morphology preservation loss and the segmentation prediction loss.

[0156] The embodiment of the present invention utilizes the information of the noisy label data through the above self-distillation label optimization method, enhances the robustness of the blood vessel segmentation model to the noisy label data, and improves its ability to effectively learn from the noisy label data. This method is particularly beneficial for medical image segmentation tasks because the labeled data usually contains noise and uncertainty. By optimizing L noisy , the vascular segmentation model of the present application can more accurately learn the target structure in the medical image, thereby improving the accuracy and reliability of the segmentation.

[0157] Step S502e: Based on the segmentation prediction loss of the student model, adjust the weight parameters of the student model. In the embodiment of the present invention, the weight parameters of the student model are updated by back propagation.

[0158] Step S502f, based on the adjusted weight parameters of the student model, adjust the weight parameters of the teacher model. Specifically, during the training process, when iterating the training step t, the weight parameters of the student model are expressed as θ t , and the weight parameter of the teacher model It can be updated by the Exponential Moving Average (EMA) method. The update formula is: Among them, α ema is the decay factor of EMA, for example, it can be 0.99. A decay factor that is too small will cause the teacher model to quickly forget the knowledge in the past rounds, while a decay factor that is too large will slow down the update of the teacher model.

[0159] In the embodiment of the present invention, a vascular segmentation method of joint hierarchical vascular morphology learning and noise label refinement for vascular segmentation (JHN-Seg) is proposed, which is applied to the training process of the vascular segmentation model, and includes two hierarchical vascular morphology learning networks (student HVMA-Net and teacher HVMA-Net) and a self-distilled noise label optimization module for collaborative learning from a limited number of high-quality labels and a large number of low-quality labels. Specifically, HVMA-Net contains a multi-scale local morphology attention module, which captures complex multi-scale local blood vessels by integrating multi-mode convolution blocks. In addition, the embodiment of the present invention proposes a global morphology preservation loss, emphasizing the role of fine branch vessels in ensuring the integrity of the global vascular structure. The self-distilled noise label optimization module is intended to improve the quality and reliability of noise labels in the training stage. In addition, the self-distilled noise label optimization module utilizes the potential knowledge in the noise labels to enhance the learning process. By continuously optimizing the noise labels and utilizing the embedded knowledge, the self-distilled noise label optimization module significantly improves the model performance, ensuring more accurate and reliable recognition results in the vascular recognition task involving noisy labeled data.

[0160] The HN-Seg proposed in the embodiment of the present invention utilizes limited high-quality data D in different ways during the training process. hq and a large amount of noisy data D noisy Given that D noisyLow-quality labels in the image may hinder model training. The embodiment of the present invention adopts a semi-supervised learning method, which ignores these low-quality labels and focuses on the image data itself. Based on this strategy, the embodiment of the present invention selects the Mean-Teacher (MT) model as the self-distillation framework, which is a very effective framework in semi-supervised learning. The average teacher framework includes two key components: a student network updated by back-propagation and a teacher network updated by weighted average of the student network weights at each step. The embodiment of the present invention uses the independently proposed HVMA-Net designed for blood vessel segmentation as the student network and teacher network in the average teacher model. The network uses different methods to process D during training. hq and D noisy , as follows:

[0161] For high-quality data hq , HN-Seg uses the student HVMA-Net in a fully supervised manner. The training process is composed of a weighted segmentation loss l seg Guided, this loss combines the global shape preservation loss and Dice loss.

[0162] For the noisy data D noisy , the original image I noisy and an image I containing some random perturbations η noisy +η are input into the student and teacher HVMA-Net respectively, and the network is inferred to get P noisy and P perturb Subsequently, the predictions and noise labels are fed into the self-distillation noise label optimization module for knowledge distillation and pseudo-labeling (i.e., optimized labels) L pseudo Finally, use L pseudo and P noisy The segmentation loss is calculated to guide the training process of the student HVMA-Net model.

[0163] By using the average teacher framework, HN-Seg can be trained using two different training datasets separately. High-quality datasets can provide the main guidance for reliably training student models, and this high-quality knowledge can be transferred to images in noisy datasets. In an embodiment of the present invention, the self-distillation noisy label optimization method innovatively uses the teacher model and the student model to construct an uncertainty map, and combines this map with the noisy label to improve the noisy label. The uncertainty map is estimated using the KL divergence, which measures the difference between the predicted probability distribution and the given prior distribution.

[0164] In summary, the embodiments of the present invention provide a blood vessel segmentation method that combines multi-level blood vessel morphology learning and noise label optimization. Under the premise of ensuring accurate blood vessel segmentation, it reduces the dependence on high-quality labels and solves the problem of difficulty in learning blood vessel morphology caused by factors such as complex and changeable blood vessel structure and low label quality. Specifically, the vascular hierarchical morphology perception network (also known as the layered vascular morphology perception network) uses a multi-scale local morphology attention mechanism and a global morphology preservation loss function to ensure the integrity of the overall blood vessel segmentation morphology. The self-distillation noise label optimization method utilizes the uncertainty of the model output, and through uncertainty optimization and consistency regularization, it fully utilizes the knowledge in the image during the training process while optimizing the noisy labels (i.e., noise labels).

[0165] After the training is completed, the teacher model can be used as the final trained vascular segmentation model, or the student model can be used as the final trained vascular segmentation model. A large number of experiments conducted on the liver vascular dataset show that the method proposed in the embodiment of the present invention performs well in segmentation performance, outperforming the six most advanced algorithms described above, with the clDice index improving by 6.9% compared to the optimal method.

[0166] In this embodiment, a training device for a blood vessel segmentation model is also provided, and the device is used to implement the above-mentioned training method embodiment and preferred implementation of the blood vessel segmentation model, which have been described and will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.

[0167] This embodiment provides a training device for a blood vessel segmentation model. Figure 7 As shown, including:

[0168] A sample data acquisition module is used to obtain blood vessel segmentation training samples and corresponding true labels;

[0169] A training module is used to train a pre-built initial blood vessel segmentation model using the blood vessel segmentation training samples and the true labels to obtain any blood vessel segmentation model described in the above embodiments.

[0170] The further functional description of each of the above modules and units is the same as that of the above corresponding embodiments and will not be repeated here.

[0171] The training device of the blood vessel segmentation model in this embodiment is presented in the form of a functional unit, where the unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that executes one or more software or fixed programs, and / or other devices that can provide the above functions.

[0172] The embodiment of the present invention also provides a computer device having the above Figure 7 The training setup of the vessel segmentation model shown.

[0173] See also Figure 8 , Figure 8 is a schematic diagram of the structure of a computer device provided by an optional embodiment of the present invention, such as Figure 8 As shown, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components are connected to each other using different buses for communication, and can be installed on a common mainboard or installed in other ways as needed. The processor can process the instructions executed in the computer device, including instructions stored in or on the memory to display the graphical information of the GUI on an external input / output device (such as, a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple computer devices can be connected, and each device provides some necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 8 A processor 10 is taken as an example.

[0174] The processor 10 may be a central processing unit, a network processor or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be a dedicated integrated circuit, a programmable logic device or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic or any combination thereof.

[0175] The memory 20 stores instructions executable by at least one processor 10, so that the at least one processor 10 executes the method shown in the above embodiment.

[0176] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely arranged relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0177] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid state drive; the memory 20 may also include a combination of the above types of memory.

[0178] The computer device also includes an input device 30 and an output device 40. The processor 10, the memory 20, the input device 30 and the output device 40 may be connected via a bus or other means. Figure 8 The example of connecting through bus is taken in the following.

[0179] The input device 30 can receive input digital or character information, and generate key signal input related to the user settings and function control of the computer device, such as a touch screen, a keypad, a mouse, a track pad, a touch pad, an indicator bar, one or more mouse buttons, a trackball, a joystick, etc. The output device 40 may include a display device, an auxiliary lighting device (e.g., an LED) and a tactile feedback device (e.g., a vibration motor), etc. The above-mentioned display device includes but is not limited to a liquid crystal display, a light emitting diode, a display and a plasma display. In some optional embodiments, the display device can be a touch screen.

[0180] The computer device also includes a communication interface, which is used for the computer device to communicate with other devices or a communication network.

[0181] The embodiment of the present invention also provides a computer-readable storage medium. The method according to the embodiment of the present invention can be implemented in hardware, firmware, or can be implemented as a computer code that can be recorded in a storage medium, or can be implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and will be stored in a local storage medium through a network download, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state hard disk, etc.; further, the storage medium can also include a combination of the above types of memories. It can be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor, or hardware, the method shown in the above embodiment is implemented.

[0182] A part of the present invention may be applied as a computer program product, such as a computer program instruction, which, when executed by a computer, can call or provide the method and / or technical solution according to the present invention through the operation of the computer. Those skilled in the art should understand that the existence of the computer program instruction in a computer-readable medium includes, but is not limited to, a source file, an executable file, an installation package file, etc., and accordingly, the way in which the computer program instruction is executed by the computer includes, but is not limited to: the computer directly executes the instruction, or the computer compiles the instruction and then executes the corresponding compiled program, or the computer reads and executes the instruction, or the computer reads and installs the instruction and then executes the corresponding installed program. Here, the computer-readable medium may be any available computer-readable storage medium or communication medium accessible to the computer.

[0183] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A blood vessel segmentation model, characterized in that: include: Encoders and decoders; The encoder includes a multi-stage downsampling module, and the decoder correspondingly includes a multi-stage upsampling module; Each level of the downsampling module includes a multi-scale local morphology attention module; The multi-scale local morphological attention module includes a plurality of different convolution modules, and the convolution kernels of different convolution modules are different in size, so as to respectively extract the vascular morphological features at different scales; The multi-scale local morphological attention module further comprises a splicing module for splicing the vascular morphological features at different scales; The output layer in the decoder performs blood vessel segmentation based on the blood vessel morphological features at different scales.

2. The model according to claim 1, characterized in that The multi-scale local morphology attention module further includes a depthwise separable convolution module, and the output of the depthwise separable convolution module is used as the output of the multiple different convolution modules; and / or, The multi-scale local morphology attention module also includes a normalization layer and a multi-layer perceptron, and the splicing module, the normalization layer and the multi-layer perceptron are connected in sequence.

3. A method for training a blood vessel segmentation model, characterized in that: The method comprises: Obtain blood vessel segmentation training samples and corresponding true labels; The pre-constructed initial blood vessel segmentation model is trained using the blood vessel segmentation training samples and the true labels to obtain the blood vessel segmentation model according to claim 1 or 2.

4. The method according to claim 3, characterized in that The using the blood vessel segmentation training samples and the true labels to train the pre-built initial blood vessel segmentation model includes: During the training process, the predicted blood vessel segmentation labels predicted by the initial blood vessel segmentation model and the corresponding real blood vessel segmentation labels are obtained; Extracting a first blood vessel skeleton of a target blood vessel branch from the predicted blood vessel segmentation label, and extracting a second blood vessel skeleton of the target blood vessel branch from the real blood vessel segmentation label; Determining a global morphology preservation loss based on the first blood vessel skeleton, the second blood vessel skeleton, the predicted blood vessel segmentation label, and the true blood vessel segmentation label; Obtaining the segmentation prediction loss of the initial blood vessel segmentation model; Based on the global morphology preservation loss and the segmentation prediction loss, a total loss of the initial blood vessel segmentation model is determined.

5. The method according to claim 4, characterized in that The determining of the global morphology preservation loss based on the first blood vessel skeleton, the second blood vessel skeleton, the predicted blood vessel segmentation label, and the real blood vessel segmentation label includes: Determining the topological accuracy of the target blood vessel branch based on the real blood vessel segmentation label and the first blood vessel skeleton; Determining a topological sensitivity of the target blood vessel branch based on the predicted blood vessel segmentation label and the second blood vessel skeleton; The global morphology preserving loss is determined based on the topology accuracy and the topology sensitivity.

6. The method according to claim 5, characterized in that The topological accuracy is the proportion of the portion of the first blood vessel skeleton located within the real blood vessel segmentation label to the first blood vessel skeleton; And / or, the topological sensitivity is the proportion of the portion of the second blood vessel skeleton located within the predicted blood vessel segmentation label to the second blood vessel skeleton.

7. The method according to claim 5 or 6, characterized in that: The determining the global morphology preservation loss based on the topology accuracy and the topology sensitivity includes: Obtaining a harmonic mean of the topology accuracy and the topology sensitivity; The global shape preserving loss is determined based on the harmonic mean.

8. The method according to claim 3, characterized in that The real label includes a precise label and a noise label; the labeling quality of the precise label is higher than the labeling quality of the noise label; The using the blood vessel segmentation training samples and the true labels to train the pre-built initial blood vessel segmentation model includes: Acquire the initial blood vessel segmentation model established in advance based on the average teacher framework, wherein the initial blood vessel segmentation model includes a teacher model and a student model; Input the noise label and the corresponding blood vessel segmentation training sample into the teacher model and the student model respectively, and obtain a first blood vessel segmentation prediction label output by the teacher and a second blood vessel segmentation prediction label output by the student model; Based on the first blood vessel segmentation prediction label and the second blood vessel segmentation prediction label, optimizing the noise label to obtain an optimized label; Calculating the segmentation prediction loss of the student model based on the optimized label and the second blood vessel segmentation prediction label; Adjusting a weight parameter of the student model based on the segmentation prediction loss of the student model; Based on the adjusted weight parameters of the student model, the weight parameters of the teacher model are adjusted.

9. The method according to claim 8, characterized in that The step of optimizing the noise label based on the first blood vessel segmentation prediction label and the second blood vessel segmentation prediction label to obtain an optimized label includes: Obtaining pixel-by-pixel KL divergence of the first blood vessel segmentation prediction label and the second blood vessel segmentation prediction label on two categories to obtain a segmented blood vessel uncertainty map and a background uncertainty map, wherein the two categories include segmented blood vessels and background; The optimized label is obtained based on the segmented blood vessel uncertainty map and the background uncertainty map, as well as the noise label.

10. The method according to claim 9, characterized in that The step of obtaining the optimized label based on the segmented blood vessel uncertainty map and the background uncertainty map, and the noise label, comprises: Obtaining a reliable label for each pixel based on the segmented blood vessel uncertainty map and the background uncertainty map; The optimized label is obtained based on the reliable label and the noise label.

11. The method according to claim 10, characterized in that The obtaining of a reliable label of each pixel based on the segmented blood vessel uncertainty map and the background uncertainty map includes: Using a negative exponential function, converting the segmented blood vessel uncertainty map and the background uncertainty map into corresponding score values; The reliable label is determined using the score value.

12. The method according to claim 9, characterized in that The calculating the segmentation prediction loss of the student model based on the optimized label and the second blood vessel segmentation prediction label includes: Acquire an uncertainty weight value and / or a regularization value of each pixel, wherein the uncertainty weight value and the regularization value are determined based on the segmented blood vessel uncertainty map; Based on the uncertainty weight value and / or the regularization value, the segmentation prediction loss of each pixel of the student model is determined.

13. A training device for a blood vessel segmentation model, characterized in that: The device comprises: A sample data acquisition module is used to obtain blood vessel segmentation training samples and corresponding true labels; A training module is used to train a pre-built initial blood vessel segmentation model using the blood vessel segmentation training samples and the true labels to obtain the blood vessel segmentation model according to claim 1 or 2.

14. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the training method of the blood vessel segmentation model according to any one of claims 3 to 12 by executing the computer instructions.

15. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the blood vessel segmentation model training method according to any one of claims 3 to 12.

Images