A 3D coronary artery image segmentation method based on semi-supervised consistency learning

Abstract

This invention discloses a 3D coronary artery image segmentation method based on semi-supervised consistency learning, comprising: acquiring raw 3D cardiac CT image data and preprocessing the acquired raw 3D cardiac CT image data by employing random rotation, contrast enhancement, and random cropping to augment the data; introducing consistency learning to construct a two-stage semi-supervised training 3D coronary artery segmentation model; inputting the 3D cardiac CT image to be segmented into the first-stage 3D coronary artery segmentation model for prediction to obtain a pseudo-label feature map, and inputting the pseudo-label feature map into the second-stage 3D coronary artery segmentation model to obtain the segmentation result. The semi-supervised consistency learning method can efficiently utilize a large amount of unlabeled data while achieving fully supervised performance. It inputs labeled and unlabeled data in a 1:4 ratio into the 3D coronary artery segmentation model, using labeled data to label the unlabeled data, thereby effectively increasing the training data.

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation technology, and in particular to a 3D coronary artery image segmentation method based on semi-supervised consistency learning. Background Technology

[0002] Coronary computed tomography (CT) angiography is a primary method for diagnosing coronary artery disease. Accurate and robust segmentation of coronary arteries from 3D cardiac CT images plays a crucial role in clinical diagnosis and treatment. However, precise segmentation of coronary arteries is challenging due to factors such as image quality, annotation accuracy, limited effective pixel count in 3D cardiac CT images of vessel distalities, and interference from venous structures. Commonly used vessel segmentation algorithms include statistically based methods, active contour model-based methods, region-based methods, boundary-tracking-based methods, and threshold-based methods. These methods have drawbacks such as poor adaptability to different cases, high computational cost, and a tendency to oversegment images. Therefore, problems exist such as incomplete segmentation due to coronary artery rupture or absence, and veins or other tissues being mistaken for coronary arteries.

[0003] With the emergence of Convolutional Neural Networks (CNNs), a series of supervised learning methods have sprung up. These methods perform end-to-end semantic segmentation of medical images based on pre-labeled text, greatly improving the overall performance of image segmentation. CNNs are now widely used in image segmentation tasks. Compared with traditional methods, CNNs have excellent feature representation capabilities and do not require manual extraction of image features. Therefore, they can effectively handle medical image segmentation tasks with complex tissue structures without requiring excessive preprocessing. CNNs generally have a five-layer structure: input layer, convolutional layer, activation layer, pooling layer, and output layer. Image preprocessing typically occurs at the input layer. After a series of preprocessing operations at the input layer, the image undergoes feature extraction and local perception through convolutional operations at the convolutional layer to obtain feature maps. The activation layer is designed to enhance the network's expressive power and is generally a non-linear mapping of the convolutional layer output. Commonly used activation functions include sigmoid, tanh, ReLU, and leaky ReLU. Pooling layers, also known as undersampling or subsampling layers, aim to reduce feature dimensionality, compress data volume, decrease overfitting, and improve the model's generalization ability. Commonly used pooling methods include max pooling and average pooling. The output layer, also known as a fully connected layer, connects classifiers, such as the softmax classifier.

[0004] However, coronary CT images are complex, and segmenting blood vessels in 3D medical images remains highly challenging. One challenge is the presence of false positives, meaning the segmented results cannot be guaranteed to be spatially connected and may contain holes or spurs. Furthermore, the complex long-distance topology of coronary arteries, coupled with imaging artifacts frequently present in CT images, makes CNN-based segmentation algorithms prone to missing certain segments of the vessels, resulting in discontinuous extraction of the vessel centerline. Supervised deep learning methods for medical imaging also rely on large amounts of labeled data; manually collecting coronary artery data is both demanding in terms of expertise and extremely time-consuming, especially in 3D images. Summary of the Invention

[0005] The purpose of this invention is to address the technical deficiencies in the existing technology by providing a 3D coronary artery image segmentation method based on semi-supervised consistency learning.

[0006] The technical solution adopted to achieve the purpose of this invention is:

[0007] A 3D coronary artery image segmentation method based on semi-supervised consistency learning, the method comprising:

[0008] Step 1: Acquire raw 3D cardiac CT image data and preprocess the acquired raw 3D cardiac CT image data by using random rotation, contrast enhancement and random cropping to augment and enhance the data.

[0009] Step 2: Introduce consistency learning to construct a 3D coronary artery segmentation model with a two-stage semi-supervised training method;

[0010] Step 3: Input the 3D cardiac CT image to be segmented into the first-stage 3D coronary artery segmentation model to predict and obtain a pseudo-label feature map, and input the pseudo-label feature map into the second-stage 3D coronary artery segmentation model to obtain the segmentation result.

[0011] In the above technical solution, the 3D cardiac CT image data includes a training dataset and a test dataset. The training dataset consists of N+M training samples, i.e., N labeled data and M unlabeled data. The labeled dataset is represented as follows: Unlabeled datasets are represented as in, Represents a 3D cardiac CT input image, y i ∈{0,1} H×W×D H represents the true label, and W and D represent the height, width and thickness of the 3D cardiac CT input image, respectively. 20% of the labeled data and 80% of the unlabeled data are used for training the 3D coronary artery segmentation model.

[0012] In the above technical solution, each stage of the 3D coronary artery segmentation model includes a set of teacher-student models.

[0013] In the above technical solution, the expression for the pseudo-label feature map is:

[0014] z i =f(x) i ;θ1,η)

[0015] In the formula, z i The pseudo-labels represent spatial features of the same size as the 3D cardiac CT input image, f(·) represents the neural network, and x i θ1 represents the weights of the first-stage 3DResUNet model, and η represents the regularization and perturbation applied to the 3D cardiac CT input image.

[0016] In the above technical solution, the expression for the segmentation result is:

[0017]

[0018] In the formula, The segmentation result represents spatial features of the same size as the 3D cardiac CT input image, f(·) represents the neural network, and z i θ represents the pseudo-label feature map, θ2 represents the weights of the second-stage model, and ξ represents the noise added to the pseudo-label.

[0019] In the above technical solution, the semi-supervised training process of the 3D coronary artery segmentation model in the first stage adopts a consistency learning method with one student model and two teacher models. The consistency of the model in the prediction process or intermediate feature extraction process is enhanced by adding perturbations. For the same unlabeled data, the image is randomly augmented during the perturbation process, so that the model obtains two different prediction results in the forward propagation, and consistency constraints are added between the predictions of the augmented image. The semi-supervised training process of the 3D coronary artery segmentation model in the second stage adopts a pseudo-label training method and introduces an attention mechanism.

[0020] In the above technical solution, the overall loss function expression for training the teacher-student model is as follows:

[0021] l total (.)=l sup (.)+λl con (.)

[0022] In the formula, l total (.) represents the total loss function for training the teacher-student model, l sup(.) represents the supervision loss between the predicted coronary artery segmentation map and the ground truth label. con (.) represents the unsupervised consistency loss, and λ represents the ramp weighting coefficient, used to control the trade-off between supervised and unsupervised losses, ensuring that the overall loss function is initially dominated by the supervised loss, thus preventing model degradation during training.

[0023] The specific expression for the slope weighting coefficient is as follows:

[0024]

[0025] In the formula, λ represents the slope weighting coefficient, t represents the current iteration number of the model, and t max This represents the maximum number of iterations the model can handle.

[0026] In the above technical solution, labeled data is input into the student model to obtain a coronary artery segmentation prediction map. The supervision loss between the coronary artery segmentation prediction map and the true labels is calculated based on cross-entropy (CE) and dice coefficient (DICE). The expression for the supervision loss is as follows:

[0027] l sup (.)=0.5×(l CE (y i ,p i )+l DICE (y i ,p i ))

[0028] In the formula, l sup (.) represents the supervision loss between the predicted coronary artery segmentation map and the ground truth label. CE (y i ,p i ) represent the cross-entropy loss between the predicted coronary artery segmentation map and the ground truth label, respectively. DICE (y i ,p i ) represent the dice coefficient loss between the predicted coronary artery segmentation map and the actual label, respectively.

[0029] In the above technical solution, for unlabeled data, a semi-supervised training strategy of mutual teaching is adopted. The same unlabeled data is input into the teacher model and the student model respectively, generating two prediction results, which are as follows:

[0030]

[0031]

[0032] In the formula, This represents the prediction result generated by inputting the same unlabeled data into the teacher model, where f(·) represents the neural network, and x u Represents unlabeled data, θ t Represents the teacher model weights, θ represents the prediction result generated by inputting the same unlabeled data into the student model. s Represents the student model weights;

[0033] Using the teacher model to provide pseudo-labels for the student model, and using the student model to provide pseudo-labels for the teacher model, the specific expressions for the pseudo-labels provided by the student model for the teacher model and by the teacher model for the student model are as follows:

[0034]

[0035]

[0036] In the formula, The pseudo-labels provided by the student model to the teacher model represent the function that evaluates the set of arguments of the function. This represents the prediction results generated by inputting the same unlabeled data into the student model. The pseudo-labels provided by the teacher model to the student model. This represents the prediction results generated by inputting the same unlabeled data into the teacher model.

[0037] In the above technical solution, the expression for the unsupervised consistency loss is as follows:

[0038]

[0039] In the formula, l con (.) represents the unsupervised consistency loss. The bi-cross-entropy loss represents the prediction results of the teacher model and the pseudo-labels. The prediction results of the student model are represented by the binary cross-entropy loss of the pseudo-label.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] 1. This invention utilizes the 3DResUNet network to extract rich global features and spatial information from 3D cardiac CT images. The extracted features are used as the stage training results. The segmentation results of the model are constrained by the self-attention mechanism in the second stage. The introduction of the self-attention mechanism makes the model pay more attention to the information of interest, overcoming some limitations of traditional neural networks, such as the performance of the system decreases with the increase of input size, low computational efficiency, and lack of feature extraction and enhancement.

[0042] 2. Semi-supervised consistency learning can efficiently utilize large amounts of unlabeled data while achieving fully supervised performance. It combines supervised and unsupervised learning, using a small subset of labeled examples and a large amount of unlabeled data from which the model learns and predicts new examples. The basic process involves using existing labeled data to label the remaining unlabeled data, effectively increasing the training data. Furthermore, it introduces more data and model perturbations on top of the teacher-student model to build consistency of the same input under different perturbations. This two-stage semi-supervised learning is more conducive to maintaining the connectivity of coronary artery segmentation results and satisfying shape constraints while learning a large amount of useful information from unlabeled data.

[0043] 3. This invention improves the accuracy of coronary artery segmentation to a certain extent, enabling multi-level analysis and processing of effective information in complex coronary artery images, thus better serving clinical medicine. It fully utilizes a large amount of unlabeled data during segmentation, significantly reducing the model's dependence on labeled data and decreasing the time and labor costs for doctors to annotate 3D data. In typical clinical scenarios, acquiring images of patients' disease sites using medical image acquisition equipment requires subjective analysis by specialist doctors. Accurate segmentation of coronary arteries is a crucial prerequisite for quantitative disease description and 3D vascular reconstruction. The method proposed in this invention plays a role in assisting doctors in clinical diagnosis and treatment, and the research on an intelligent image segmentation method is of great significance in clinical medical diagnosis. Attached Figure Description

[0044] Figure 1 The diagram shows the flowchart of the 3D coronary artery image segmentation method based on semi-supervised consistency learning of the present invention.

[0045] Figure 2 The diagram shown is a structural schematic of the 3DResUNet of the present invention. Detailed Implementation

[0046] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0047] Example 1

[0048] A 3D coronary artery image segmentation method based on semi-supervised consistency learning, see [link to relevant documentation]. Figure 1 The method includes:

[0049] Step 1: Acquire raw 3D cardiac CT image data and preprocess the acquired raw 3D cardiac CT image data by using random rotation, contrast enhancement and random cropping to augment and enhance the data.

[0050] The preprocessing includes: uniformly adjusting the size of the acquired original 3D cardiac CT images to 256*256*128, cropping all scan data with the heart region as the center and removing lung noise, limiting the Hu value of the 3D cardiac CT images to the range of [-90, 410], and normalizing each image to zero mean and unit variance.

[0051] The 3D cardiac CT image data includes a training dataset and a test dataset. The training dataset consists of N+M training samples, i.e., N labeled data and M unlabeled data. The labeled dataset is represented as follows: Unlabeled datasets are represented as in, Represents a 3D cardiac CT input image, y i ∈{0,1} H×W×D The labels represent the true values, and H, W, and D represent the height, width, and thickness of the 3D cardiac CT input image, respectively. 20% of the labeled data and 80% of the unlabeled data are selected for training the 3D coronary artery segmentation model.

[0052] Data augmentation methods include: during the training phase, data augmentation is performed using a combination of methods, including random rotation, contrast enhancement, and random cropping; based on the characteristics of consistency learning, the same unlabeled data is augmented using both strong augmentation (random masking, random grayscale, and random blurring) and weak augmentation (random noise); during the inference phase, data augmentation is performed using center cropping, cropping the 3D cardiac CT image to be segmented to 128*128*128 pixels.

[0053] Step 2: Introduce consistency learning to construct a 3D coronary artery segmentation model with a two-stage semi-supervised training method.

[0054] The 3D coronary artery segmentation model comprises two stages. Both the first and second stages of the 3D coronary artery segmentation model employ semi-supervised training, and each stage includes a teacher-student model. In the first stage, 3DResUNet is selected as the backbone network. 3DResUNet contains four encoder layers, four decoder layers, and a certain number of residual layers. Consistency learning is achieved by increasing model perturbation, data perturbation, and introducing additional teacher models. The network in the second stage includes convolutional layers and self-attention layers, with some noise added during the training phase.

[0055] Among them, see Figure 2Each encoder layer of the 3DresUNet contains a convolutional layer, a normalization layer, an activation layer, and a downsampling layer. The convolutional kernel size of the convolutional layer is 3*3*3 with a stride of 1. The activation layer uses the LeakyReLU activation function, and the normalization layer uses the BatchNorm normalization function. The convolutional kernel size of the downsampling layer is 3*3*3 with a stride of 2. After passing through the convolutional and downsampling layers, the number of channels in the feature map doubles, and the resolution is halved. The decoder layer replaces the downsampling layer in the encoder with an upsampling layer. In order to retain more low-level information, a certain number of skip connections are used.

[0056] Step 3: Input the 3D cardiac CT image to be segmented into the first-stage 3D coronary artery segmentation model to predict and obtain a pseudo-label feature map, and input the pseudo-label feature map into the second-stage 3D coronary artery segmentation model to obtain the segmentation result.

[0057] Specifically, the 3D cardiac CT image to be segmented is input into the 3DResUNet network for prediction, resulting in the first-stage pseudo-label feature map, which is as follows:

[0058] z i =f(x) i ;θ1,η)

[0059] In the formula, z i The pseudo-labels represent spatial features of the same size as the 3D cardiac CT input image, f(·) represents the neural network, and x i θ1 represents the weights of the first-stage 3DResUNet model, and η represents the regularization and perturbation applied to the 3D cardiac CT input image.

[0060] The input to the second stage is the pseudo-label feature map generated in the first stage. The pseudo-label feature map is then input into the network of this stage to obtain the segmentation result. The specific expression is as follows:

[0061]

[0062] In the formula, The segmentation result represents spatial features of the same size as the 3D cardiac CT input image, f(·) represents the neural network, and z i θ represents the pseudo-label feature map, θ2 represents the weights of the second-stage model, and ξ represents the noise added to the pseudo-label.

[0063] The first stage employs a consistency learning approach with one student model and two teacher models. Consistency learning enhances the model's consistency during the prediction process or intermediate feature extraction by adding perturbations. For the same unlabeled data, the image is randomly augmented during the perturbation process, causing the model to obtain two different prediction results during forward propagation. Consistency constraints are added between the predictions of the augmented image, making the model pay more attention to low-density regions to improve segmentation performance. To ensure the connectivity constraints of the coronary arteries, the second stage uses a pseudo-label training method and introduces an attention mechanism.

[0064] Specifically, in the semi-supervised training process, the parameters of the student model are updated through backpropagation, while the parameters of the teacher model are an exponentially weighted average of the student network parameters. To ensure consistent learning, both the teacher and student models can learn the features of the labeled data, supervisory constraints are established for the labeled data. The labeled data is input into the student model for training, resulting in a predicted coronary artery segmentation map. The supervisory loss between the predicted coronary artery segmentation map and the true labels is calculated based on cross-entropy (CE) and dice coefficient (DICE). The expression for the supervisory loss is as follows:

[0065] l sup (.)=0.5×(l CE (y i ,p i )+l DICE (y i ,p i ))

[0066] In the formula, l sup (.) represents the supervision loss between the predicted coronary artery segmentation map and the ground truth label. CE (y i ,p i ) represent the cross-entropy loss between the predicted coronary artery segmentation map and the ground truth label, respectively. DICE (y i ,p i ) represent the dice coefficient loss between the predicted coronary artery segmentation map and the true label, respectively. i Represents the true label, p i This represents a predicted segmentation diagram of the coronary arteries.

[0067] For unlabeled data, a semi-supervised training strategy of mutual teaching is adopted. The same unlabeled data is input into the teacher model and the student model respectively, generating two prediction results, which are as follows:

[0068]

[0069]

[0070] In the formula, This represents the prediction result generated by inputting the same unlabeled data into the teacher model, where f(·) represents the neural network, and x u Represents unlabeled data, θ t Represents the teacher model weights, θ represents the prediction result generated by inputting the same unlabeled data into the student model. s Represents the student model weights.

[0071] In semi-supervised training, pseudo-labels are a method that advocates using the model itself to generate artificial labels for unlabeled data. For unlabeled data, pseudo-labels are provided by the teacher model for the student model, and pseudo-labels are provided by the student model for the teacher model. The specific expressions for the pseudo-labels provided by the student model for the teacher model and the pseudo-labels provided by the teacher model for the student model are as follows:

[0072]

[0073]

[0074] In the formula, The pseudo-labels provided by the student model to the teacher model represent the function that evaluates the set of arguments of the function. This represents the prediction results generated by inputting the same unlabeled data into the student model. The pseudo-labels provided by the teacher model to the student model. This represents the prediction results generated by inputting the same unlabeled data into the teacher model.

[0075] In the semi-supervised training process, the confidence score output by the model after softmax represents its confidence in the corresponding pseudo-label. The higher the confidence score, the higher the potential accuracy of the pseudo-label. Specifically, the prediction results of the two teacher models on the same 3D cardiac CT image are averaged, and the final confidence score is used as a weight to penalize the output of the student model, discarding pixel labels of 3D cardiac CT images with low confidence scores. Two-way cross-entropy (BCE) is used to improve overall performance, thereby determining the unsupervised consistency loss. The expression for the unsupervised consistency loss is as follows:

[0076]

[0077] In the formula, l con (.) represents the unsupervised consistency loss. The bi-cross-entropy loss represents the prediction results of the teacher model and the pseudo-labels. The prediction results of the student model are represented by the binary cross-entropy loss of the pseudo-label.

[0078] Among them, the use of binary cross-entropy (BCE) improves overall performance and can reduce noise in the training process.

[0079] Therefore, the overall loss function expression for training the teacher-student model is as follows:

[0080] l total (.)=l sup (.)+λl con (.)

[0081] In the formula, l total (.) represents the total loss function for training the teacher-student model, l sup (.) represents the supervision loss between the predicted coronary artery segmentation map and the ground truth label. con (.) represents the unsupervised consistency loss, and λ represents the ramp weighting coefficient, used to control the trade-off between supervised and unsupervised losses, ensuring that the overall loss function is initially dominated by the supervised loss, thus preventing model degradation during training.

[0082] The specific expression for the slope weighting coefficient is as follows:

[0083]

[0084] In the formula, λ represents the slope weighting coefficient, t represents the current iteration number of the model, and t max This represents the maximum number of iterations the model can handle.

[0085] During the training process, unlabeled data is weakly augmented and then input into the teacher model. The results predicted by the two teacher models are averaged and weighted to form the first pseudo-label feature map. Unlabeled data is strongly augmented and then input into the student model, and the prediction result is recorded as the second pseudo-label feature map. The prediction result of labeled data by the student model is recorded as the third pseudo-label feature map. The specific expressions for each pseudo-label feature map are as follows:

[0086]

[0087] In the formula, This represents the first pseudo-label feature map. This represents the second pseudo-label feature map. and Used to calculate consistency loss. The third pseudo-label feature map is used to calculate the supervised loss. η′ represents unlabeled data after weak augmentation, η represents unlabeled data after strong augmentation, f(·) represents the neural network, and x u Represents unlabeled data. Representing the first teacher model, Representing the second teacher model, Representing the student model, x l Represents the label image.

[0088] In the first stage of the semi-supervised training process, the combination of supervised loss and unsupervised consistency loss is minimized based on the student model, as specifically expressed below:

[0089]

[0090] In the formula, l sup1 This represents a mixture of the first-stage cross-entropy (CE) loss and the dice coefficient (DICE) loss. con1 D represents the unsupervised consistency loss. L D represents the labeled dataset. U Let f(·) represent an unlabeled dataset, f(·) represent a neural network, and x represent an unlabeled dataset. i Represents a 3D cardiac CT input image. Representing the student model, y i Represents the true label, η′ represents the weighted average of the predictions from the two teacher models, η′ represents different levels of data augmentation, η represents the interference from unlabeled data, and λ represents the slope weighting coefficient.

[0091] In the first stage of the semi-supervised training process, the combination of supervised loss and unsupervised consistency loss is minimized based on the student model, as specifically expressed below:

[0092]

[0093] In the formula, l sup2 A mixture representing CE loss and DICE loss, l con2 D represents the unsupervised consistency loss. L D represents the labeled dataset. U Let f(·) represent an unlabeled dataset, f(·) represent a neural network, and z represent a neural network. i This represents the pseudo-label feature map generated in the first stage. y represents the student model weights. i Represents the true label, ξ represents the weights of the teacher model, ξ′ and ξ represent different noises added to the pseudo-labels, and λ represents the slope weighting coefficient.

[0094] In this semi-supervised training process, the two phases are performed sequentially. In the teacher-student model, the weights of the student model are updated through backpropagation. The exponential moving average (EMA) of the student model weights is used to update the weights of the teacher model to integrate information from different training steps. That is, in training step t, the weights of the teacher model are updated to: θ t '=αθt ' -1 +(1-α)θ t .

[0095] Where, θ t 'θ represents the weights of the teacher model updated after the semi-supervised training step t. t ' -1 θ represents the weights of the teacher model after the semi-supervised training step t-1 has been updated. t α represents the weights of the teacher model in the semi-supervised training step t, and α represents the ratio controlling EMA decay.

[0096] Furthermore, multiple coronary artery segmentation prediction maps are generated from the same 3D cardiac CT image. The generated coronary artery segmentation prediction maps are then reconstructed in 3D according to their original positions to obtain the reconstructed coronary artery segmentation prediction maps. Based on the reconstructed coronary artery segmentation prediction maps, the average precision, average recall, average dice similarity coefficient, and average topological skeleton similarity coefficient of the coronary arteries are calculated to evaluate the constructed 3D coronary artery segmentation model.

[0097] Example 2

[0098] Based on Example 1, a specific 3D cardiac CT image is used as an example for illustration.

[0099] Specifically, the original 3D cardiac CT images were first resized to 256*256*128, the Hu values ​​of the CT images were limited to the range of [-90, 410] and pulmonary vein noise was removed, and each image was normalized to zero mean and unit variance; labeled and unlabeled data were divided in a 1:4 ratio; during semi-supervised training, the CT images were cropped to 128*128*128 size, and the training images were enhanced to varying degrees; after the 3D coronary artery segmentation model was trained, the prediction results obtained from training with two teacher models were used as the final segmentation results.

[0100] The model was tested on publicly available datasets such as orCaScore and CASDQEF, and the segmentation results were evaluated using Precision, Recall, and Dice.

Claims

1. A 3D coronary artery image segmentation method based on semi-supervised consistency learning, characterized in that, The method comprises: Step 1, collecting original 3D heart CT image data, and preprocessing the collected original 3D heart CT image data, adopting random rotation, contrast enhancement and random cropping to expand and enhance the data; Step 2, introducing consistency learning, and constructing a 3D coronary artery segmentation model in a two-stage semi-supervised training mode; Step 3, inputting the 3D heart CT image to be segmented into the first-stage 3D coronary artery segmentation model to obtain a pseudo-label feature map, and inputting the pseudo-label feature map into the second-stage 3D coronary artery segmentation model to obtain a segmentation result; Each stage of the 3D coronary artery segmentation model comprises a group of teacher-student models, the semi-supervised training process of the first-stage 3D coronary artery segmentation model adopts a consistency learning mode of one student model and two teacher models, the consistency of the model in the prediction process and the intermediate feature extraction process is strengthened by adding disturbance, for the same unlabeled data, the image is randomly expanded in the disturbance process, so that the model obtains two different prediction results in the forward propagation, and a consistency constraint is added between the prediction of the expanded image; the semi-supervised training process of the second-stage 3D coronary artery segmentation model adopts a pseudo-label training mode and introduces an attention mechanism.

2. The method of claim 1, wherein, The 3D cardiac CT image data includes a training data set and a test data set, the training data set is composed of N+M training samples, that is, N labeled data and M unlabeled data, and the labeled data set is represented as The unlabeled data set is represented as Wherein, represents a 3D cardiac CT input image, y i ∈{0,1} H×W×D represents the true label, H, W and D represent the height, width and thickness of the 3D cardiac CT input image respectively; 20% of the labeled data and 80% of the unlabeled data are used for 3D coronary artery segmentation model training.

3. The method of claim 1, wherein, The expression of the pseudo-label feature map is: z i = f(x i ; θ1,η) wherein z i represents a spatial feature having the same size as the 3D cardiac CT input image, f(·) represents a neural network, x i represents a 3D cardiac CT input image, θ1 represents the weights of the first-stage 3D ResUNet model, and η represents the regularization and disturbance performed on the 3D cardiac CT input image.

4. The method of claim 1, wherein, The expression of the segmentation result is: In the formula, representing that the segmentation result has the same size of spatial features as the 3D heart CT input image, f(·) represents a neural network, z i representing a pseudo-label feature map, θ2 represents the weight of the second stage model, and ξ represents noise added to the pseudo-label.

5. The method of claim 1, wherein, The total loss function expression of the teacher-student model training is as follows: l total (.) = l sup (.) + λl con (.) wherein l total (.) represents the total loss function of the teacher-student model training, l sup (.) represents the supervised loss between the coronary artery segmentation prediction map and the real label, l con (.) represents the unsupervised consistency loss, λ represents the slope weighting coefficient, which is used to control the trade-off between the supervised loss and the unsupervised loss, to ensure that the total loss function is dominated by the supervised loss at the beginning, avoiding the degradation of the model during the training process, wherein The specific expression of the slope weighting coefficient is as follows: In the formula, λ represents a slope weighting coefficient, t represents the current iteration number of the model, and t max represents the maximum iteration number of the model.

6. The method of claim 5, wherein, Input the label data into the student model to obtain a coronary artery segmentation prediction map, and calculate the supervision loss between the coronary artery segmentation prediction map and the true label based on cross entropy (CE) and dice coefficient (DICE), the expression of the supervision loss is as follows: l sup (0 = 0.5 x (l CE (y i ,p i )+l DICE (y i ,p i )) where l sup (.) represents the supervised loss between the coronary artery segmentation prediction map and the true label, l CE (y i , p i ) respectively represent the cross-entropy loss between the coronary artery segmentation prediction map and the true label. l DICE (y i , p i ) respectively represent the dice coefficient loss between the coronary artery segmentation prediction map and the true label.

7. The method of claim 1, wherein, For unlabeled data, a mutual teaching semi-supervised training strategy is adopted, the same unlabeled data is input into the teacher model and the student model to generate two prediction results, the two prediction results are: In the formula, represents the prediction result generated by inputting the same label-free data into the teacher model, f(·) represents a neural network, x u represents the label-free data, θ t represents the teacher model weight, represents the prediction result generated by inputting the same label-free data into the student model, θ s represents the student model weight; The teacher model provides a pseudo-label for the student model, and the student model provides a pseudo-label for the teacher model, the specific expression of the pseudo-label provided by the student model for the teacher model and the pseudo-label provided by the teacher model for the student model is as follows: In the formula, represents the pseudo-label provided by the student model for the teacher model, argmax() represents a function of taking the parameter set of the function, represents the prediction results generated by inputting the same unlabeled data into the student model respectively, represents the pseudo-label provided by the teacher model for the student model, represents the prediction results generated by inputting the same unlabeled data into the teacher model respectively.

8. The method of claim 5, wherein, The unsupervised consistency loss expression is as follows: In the formula, l con (.) represents an unsupervised consistency loss, represents the binary cross-entropy loss of the prediction result of the teacher model and the pseudo label, represents the binary cross-entropy loss of the prediction result of the student model and the pseudo label.

Images