Method for intima segmentation of carotid ultrasound image based on pseudo-label divergence active learning

Abstract

The application discloses a kind of based on pseudo-label divergence active learning's strong artery ultrasound image intima segmentation method, first by fitting student network parameter to image set, so that the performance of student network shows the trend of substantial rise, if the average intersection-over-union ratio change rate is less than early stop parameter, stop fitting to avoid overfitting in the later training period;And compare the maximum prediction value based on the updated teacher network for each frame pixel output of labeled image with label refinement parameter, update the labeled label based on the comparison result to obtain more accurate labeled label, based on the above two points, the method provided by the application can more accurately predict the intima of strong artery ultrasound image.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of image processing, and particularly relates to a pseudo-label divergence-based active learning carotid artery ultrasound image intima segmentation method. BACKGROUND

[0002] Carotid intima-media (CIM), i.e. the part between lumen-intima (LI) and media-adventitia (MA), is positively correlated with obesity and atherosclerotic diseases in morphology. One of the schemes for accurately identifying the CIM morphology is medical image segmentation based on deep learning. Compared with conventional ultrasound scanning, CIM segmentation can provide more accurate morphological basis for obesity and atherosclerotic diseases by measuring the thickness and roughness of the intima-media layer.

[0003] A carotid artery ultrasound image recognition method based on self-supervised learning is disclosed in Chinese Patent No. CN113159223A. The method comprises the following steps: (1) collecting carotid artery ultrasound images and preprocessing; (2) using the auxiliary task function of self-supervised learning to expand the data set and generate corresponding pseudo labels according to the obtained ultrasound image data set after preprocessing; (3) loading the new ultrasound image data set obtained from step (2) into a neural network for training, saving the learned optimal network weight parameters; (4) migrating the network weight parameters to the target neural network, learning the preprocessed ultrasound image data set obtained in step (1), obtaining the optimal network model of the target neural network, and then testing the test set to obtain the final test accuracy. This method applies the self-supervised learning method to the surface changes and internal characteristics of the carotid artery ultrasound image feature extraction, providing a quantitative analysis method for the prediction of the lesion area.

[0004] The document K. M. Meiburger, F. Marzola, G. Zahnd, F. Faita, C. P. Loizou, N. Laine', C. Carvalho, D. A. Steinman, L. Gibello, R. M. Bruno et al., "Carotid ultrasound boundary study (CUBS): Technical considerations on an open multi-center analysis of computerized measurement systems for intima media thickness measurement on common carotid artery longitudinal b-mode ultrasound scans," Computers in Biology and Medicine, vol. 144, p. 105333, 2022. proposes a region-based fully automatic segmentation method. The region of interest with a size of 128*512 is cropped along the y-axis in the ultrasound image with a pixel size of 618*512, and the regions have overlaps. The cropped region of interest image is used for fully supervised training of the dilated U-Net network. Finally, the trained model is used for prediction of each region of interest, and all predictions are spliced and combined into a complete prediction map, and the overlapping part is taken as the prediction.

[0005] The carotid ultrasound image recognition method disclosed in the above patent and document uses the CIM region on part of the ultrasound scan image, which is unclear and difficult to distinguish the accurate position of the intima. And the label in clinic is usually a discrete annotation point annotated at the boundary, which is not suitable for deep learning, and two adjacent annotation points need to be connected by interpolation to obtain a continuous boundary region annotation. Therefore, the data set will contain a large number of images with unclear CIM region and noisy labels, thereby affecting the training and generalization performance of the model, reducing the segmentation accuracy, and being not conducive to obtaining the morphological information of the intima from the image to assist in diagnosis. SUMMARY

[0006] The present application provides a carotid ultrasound image intima segmentation method based on pseudo-label divergence active learning, which can accurately predict the carotid ultrasound image intima and optimize the annotation label.

[0007] The carotid ultrasound image intima segmentation method based on pseudo-label divergence active learning provided by the embodiment of the present application has the characteristics that it comprises:

[0008] S1. Use the carotid artery ultrasound image set as the training set, divide the training set into an labeled image set and an unlabeled image set, and construct a label pool through the labeled image set and the label.

[0009] S2. Initialize the student and teacher network parameters, update the student and teacher network parameters through an inner loop, and optimize the labeling, including:

[0010] S21. The student network parameters are updated sequentially using the supervised loss function based on the labeled image set and the consistency loss function based on the unlabeled image set to obtain the optimized student network.

[0011] S22. Based on the updated student network parameters, the teacher network parameters are updated using the exponential moving average method. The predicted values ​​are obtained by using the optimized student network through the labeled image set, and the average intersection-union ratio of the predicted values ​​and the labeled labels is obtained.

[0012] S23. If the current cycle number of the inner loop reaches the first cycle number threshold, then fit the obtained average crossover ratio with the corresponding cycle number of the inner loop to obtain the fitting function; otherwise, iterate steps S21-S22.

[0013] S24. Differentiate the fitted function and use the differentiated function to obtain the average crossover ratio change rate based on the current and previous inner loop cycles. If the average crossover ratio change rate is less than the early stopping parameter, compare the maximum predicted value of each frame pixel output of the updated teacher network for the labeled image with the label refinement parameter. Update the label based on the comparison result to obtain the optimized label. Otherwise, iterate steps S21-S23 and input the carotid ultrasound image into the optimized teacher network to obtain the intima prediction value.

[0014] Furthermore, the optimized labeled pool and unlabeled images are obtained through training in the outer loop, resulting in the final teacher network, which includes:

[0015] S3. Labeling a portion of the unlabeled images in the unlabeled image set to optimize both the labeled and unlabeled image sets, including:

[0016] S31. Calculate the KL divergence of image samples in the unlabeled image set based on the prediction results of the optimized teacher network and student network.

[0017] S32. Sort the KL divergence of the image samples in the unlabeled image set in descending order, and select the unlabeled image samples corresponding to the top n KL divergence values.

[0018] S33. The unlabeled image samples selected are re-labeled to obtain new labels. The re-labeled image samples are added to the labeled image set and removed from the unlabeled image set to obtain the optimized labeled image set and the unlabeled image set, respectively. An optimized label pool is constructed using the optimized labeled images, the new labels, and the optimized labels.

[0019] S4. If the number of cycles in the outer loop reaches the second cycle threshold, then stop the outer loop to obtain the final teacher network; otherwise, iterate through steps S2-S3. The cycle of the outer loop is steps S2-S4.

[0020] Furthermore, based on the prediction results of the optimized teacher network and student network, the unlabeled image sample x is calculated. u KL divergence d(x) u )for:

[0021] d(x u ) = mean i，j Div KL (x u (i, j), θ S θ T )

[0022]

[0023] Where, mean i，j Div KL (·) represents the unlabeled image sample x u After calculating the KL divergence at each pixel location, the average value is taken, θ S θ represents the optimized student network parameters obtained through the inner loop. T The optimized teacher network parameters are obtained through the inner loop. For the optimized teacher network targeting unlabeled image samples x u The binary classification prediction probability of each pixel in each frame. To optimize the student network for unlabeled image samples x u The binary classification prediction probability of each pixel in each frame, where c = 0 represents an unlabeled image sample x. u For non-inner membranes, c=1 represents unlabeled image sample x. u It is the endometrium.

[0024] Furthermore, the obtained average crossover ratio is fitted to the corresponding number of cycles in the inner loop to obtain a fitting function. for:

[0025]

[0026] Where a, b, and c are fitting parameters, s is the index of the number of inner loop cycles, s = 1, 2, ..., k, and k is the current number of inner loop cycles.

[0027] Furthermore, the average crossover ratio change rate γ is obtained by using the differentiated function based on the current and previous inner loop cycle numbers. k for:

[0028]

[0029] in, To differentiate the fitted function, k is the number of cycles in the current inner loop.

[0030] Furthermore, based on the comparison results, the annotation labels are updated to obtain optimized annotation labels for the labeled image x. l The optimized label corresponding to the (i, j) pixel position. for:

[0031]

[0032]

[0033] Where, x l (i, j) represents the labeled image x. l The (i, j) pixel position in the text. F represents the teacher network parameters updated during the k-th inner loop cycle. T (·) represents the optimized teacher network. The optimized teacher network provides the binary classification prediction probability for each pixel in the frame, where λ is the set label refinement parameter, and y(i,j) is the labeled image x. l The original label corresponding to the (i, j) pixel position in the text.

[0034] Furthermore, the optimized student network is obtained by updating the student network parameters using a supervised loss function based on the labeled image set and a consistency loss function based on the unlabeled image set, where:

[0035] Supervision loss function L sup for:

[0036]

[0037] Consistency loss function L con for:

[0038]

[0039] in, X represents the student network parameters updated during the k-th inner loop cycle. L For the labeled image set, YL For the labeled set, x l For the labeled image, y represents the label, and F represents the image. T (·) represents the optimized teacher network, F S (·) represents the optimized student network, X U For the unlabeled image set, x u This is an unlabeled image.

[0040] Furthermore, based on the updated student network parameters of the k-th inner cycle, the teacher network parameters of the k-th inner cycle are updated using the exponential moving average method. for:

[0041]

[0042] in, The moving average coefficient is represented by `iter`, which represents the number of iterations in the training process within the current inner loop period `k`. These are the student network parameters updated during the k-th inner loop cycle. The teacher network parameters are updated during the k-th inner loop cycle.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] This invention uses labeled and unlabeled images to fit the parameters of the student network, resulting in a significant increase in the performance of the student network. In the later stages of training, the fitting performance tends to fluctuate slightly. The average crossover ratio (CWR) change rate is monitored, and if the average CWR change rate is less than the early stopping parameter, the fitting is stopped to avoid overfitting. The resulting optimized student network and teacher network can more accurately predict the intima of the carotid artery in ultrasound images. Furthermore, this invention compares the maximum predicted value of each pixel output of the updated teacher network for labeled images with the label refinement parameters, and updates the label based on the comparison results to obtain more accurate label.

[0045] This invention optimizes the teacher network and student network through an inner loop, calculates the KL divergence of the predicted values ​​of each pixel in the unlabeled image, and performs expert annotation on the unlabeled images corresponding to the top n KL divergence values. In other words, it performs expert annotation on the unlabeled images with large prediction deviations to expand the labeled image set, and further improves the model performance of the teacher network through the expanded labeled image set. Attached Figure Description

[0046] Figure 1 Flowchart of a method for intimal segmentation of carotid artery ultrasound images based on active learning of pseudo-label divergence, provided for a specific embodiment of the present invention;

[0047] Figure 2The inner loop flowchart provided for a specific embodiment of the present invention;

[0048] Figure 3 This is a flowchart for updating the labeled pool and the unlabeled image set, provided as a specific embodiment of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0050] This invention constructs a fitting function by fitting the predicted value output by the optimized student network with the average crossover ratio (CRO) calculated from the labeled values ​​and the number of cycles in the internal circulation. The change rate of the average CRO obtained from the fitting function is used to determine whether the internal circulation has ended, i.e., whether the student network training is complete. This allows for the acquisition of a student network and a teacher network that can accurately predict the intima of the carotid artery in ultrasound images.

[0051] The present invention also obtains the maximum value of the binary classification prediction probability output by the teacher network for each pixel position of the labeled image after inner loop optimization, compares the maximum value with the set label refinement parameter, and if it is greater than the set label refinement parameter, it indicates that the maximum value is reliable. Then, the label of the pixel position corresponding to the maximum value is modified to the prediction result corresponding to the maximum value, thereby optimizing the label.

[0052] The specific embodiments of this invention provide a method for intimal segmentation of carotid artery ultrasound images based on active learning of pseudo-label divergence, such as... Figure 1 As shown, it includes:

[0053] S1. Obtain the initial unlabeled image set and label pool, including: extracting a carotid ultrasound image of a region of interest of size I×J from each ultrasound scan. A training set was constructed using multiple carotid artery ultrasound images. A small number of carotid artery ultrasound images were randomly selected from the training set and sent to experts for annotation, resulting in an annotated image set X. L Its corresponding label is Y L D L =X L ×Y L This is the labeled pool. The remaining unlabeled data is denoted as X. U .

[0054] In one specific embodiment, the structure of the student network (such as the number of hidden layers and the number of nodes in each layer) is specified, while the teacher network has the same structure as the student network. Let F represent the carotid artery ultrasound image. Mapping to pixel-wise prediction probability neural networks, Let x represent the binary classification prediction probability at pixel position (i, j) (i∈{1, 2, ..., I}, j∈{1, 2, ..., J}).

[0055] In a specific embodiment of the invention, parameters are set, specifying two parameters, λ and τ, for label refinement and early stopping. The number of samples selected in each active learning loop is n.

[0056] S2. Initialize student network parameters θ S and teacher network parameters θ T Where θ S Initialized by ImageNet pre-training, θ T From θ S Initialization. (e.g.) Figure 2 As shown, the inner loop updates the parameters of the student and teacher networks and optimizes the labels. The inner loop trains the neural network model using the current label pool and refines the noisy labels using the teacher model. Assuming the current inner loop is in its k-th iteration, the specific steps are as follows:

[0057] S21. The optimized student network is obtained by sequentially applying the supervised loss function based on the labeled image set and the consistency loss function based on the unlabeled image set.

[0058] The supervised loss function L provided in the specific embodiments of the present invention sup for:

[0059]

[0060] The consistency loss function L provided in the specific embodiments of the present invention con for:

[0061]

[0062] in, X represents the student network parameters updated during the k-th inner loop cycle. L For the labeled image set, Y L For the labeled set, x l For the labeled image, y represents the label, and F represents the image. T (·) represents the optimized teacher network, F S (·) represents the optimized student network, X U For the unlabeled image set, x u This is an unlabeled image.

[0063] S22. Update teacher network parameters using the exponential moving average method based on the updated student network parameters. The predicted values ​​are obtained by using an optimized student network on a labeled image set, and the average intersection-union ratio (μ) between the predicted values ​​and the labeled labels is obtained. (k) .

[0064] In one specific embodiment, the teacher network parameters for the k-th inner loop cycle are updated using an exponential moving average method based on the updated student network parameters for the current inner loop cycle k. for:

[0065]

[0066] in, represents the moving average coefficient, and iter represents the number of iterations in the training process using the above loss function within the current inner loop period k. These are the student network parameters updated during the k-th inner loop cycle. The teacher network parameters are updated during the k-th inner loop cycle.

[0067] In one specific embodiment, D is calculated. L The mean intersection over union (mIoU) μ between the predicted and labeled outputs of the optimized student network (label pool) (k) for:

[0068]

[0069] Where IoU(A, B) = |A∩B| / |A∪B|, A is y, i.e., the label, and B is... That is, the prediction of the optimized student network output.

[0070] S23. If the current cycle number k of the inner loop reaches the first cycle number threshold, in one embodiment, the first cycle number threshold K1≥3, then the average crossover ratio obtained in the previous inner loop cycle is fitted with the corresponding cycle number of the inner loop to obtain the fitting function; otherwise, iterate steps S21-S22.

[0071] In one specific embodiment, the fitting function for:

[0072]

[0073] Where a, b, and c are fitting parameters, s is the index of the number of inner loop cycles, s = 1, 2, ..., k, and k is the current number of inner loop cycles.

[0074] S24. Differentiate the fitted function and use the differentiated function to obtain the average crossover ratio (CRO) rate of change based on the current and previous inner loop cycles. If the average CRO rate of change is less than the early stopping parameter, it indicates that continuing to fit will lead to overfitting. Therefore, the fitting should be stopped, and the labels should be optimized based on the optimized teacher network. This includes comparing the maximum predicted value of each frame pixel output of the updated teacher network for the labeled image with the label refinement parameters, and updating the labels based on the comparison results to obtain the optimized labels. Based on optimized labeling and the labeled image set is X L Construct an inner loop to optimize the annotation pool If the average crossover ratio changes at a rate γ k If the value is not less than the early stopping parameter τ, then iterative steps S21-S23 are performed. During the application process, the carotid artery ultrasound image is input into the optimized teacher network to obtain the intima prediction value.

[0075] In one specific embodiment, the average crossover ratio change rate γ is obtained using the differentiated function based on the current and previous inner loop cycle numbers. k for:

[0076]

[0077] in, To differentiate the fitted function, k is the number of cycles in the current inner loop.

[0078] In one specific embodiment, the labeled image x l The optimized label corresponding to the (i, j) pixel position. for:

[0079]

[0080]

[0081] Where, x l (i, j) represents the labeled image x. l The (i, j) pixel position in the text. F represents the teacher network parameters updated during the k-th inner loop cycle. T (·) represents the optimized teacher network. For the optimized teacher network, the binary classification prediction probability for each frame pixel is given by λ, which is a set label refinement parameter. This is a binary classification result. for The index of the maximum value is 0 or 1. If the result is 0, it indicates that it is not the endometrium, and if the result is 1, it indicates that it is the endometrium.

[0082] Existing carotid ultrasound image recognition methods require manual annotation by experts to obtain labels for deep learning. However, medical annotation is very expensive, and annotating all data sequentially requires significant time and manpower. Therefore, fully supervised learning using all labels is impractical. This invention optimizes both the labeled and unlabeled image sets by using an outer loop to select the most valuable labeled samples based on the divergence between the predicted values ​​from the student and teacher networks trained in the inner loop. This outer loop is an active learning loop that selects the most valuable labeled samples based on the divergence between the student and teacher network predictions. One cycle of this outer loop is steps S2-S4. The optimized labeled pool and unlabeled images are obtained through training in the outer loop, leading to the final teacher network, which includes:

[0083] S3. Label some unlabeled images in the unlabeled image set to optimize both the labeled and unlabeled image sets. Figure 3 As shown, it includes:

[0084] S31. Calculate the KL divergence of all unlabeled image samples in the unlabeled image set based on the prediction results of the optimized teacher network and student network.

[0085] In one specific embodiment, unlabeled image samples are input into the teacher network and student network after inner-loop optimization, respectively. Based on the prediction results of the teacher network and student network after inner-loop optimization, the unlabeled image sample x is calculated. u KL divergence d(x) u )for:

[0086] d(x u ) = mean i，j Div KL (x u (i, j), θ S θ T )

[0087]

[0088] Where, mean i，j Div KL (·) represents the unlabeled image sample x u After calculating the KL divergence at each pixel location, the average value is taken, θ S θ represents the optimized student network parameters obtained through the inner loop. T The optimized teacher network parameters are obtained through the inner loop. For the optimized teacher network targeting unlabeled image samples x u The binary classification prediction probability of each pixel in each frame. To optimize the student network for unlabeled image samples x u The binary classification prediction probability of each pixel in each frame, where c = 0 represents an unlabeled image sample x. u For the inner membrane, c=1 represents the unlabeled image sample x. u Not the endometrium.

[0089] S32. Sort the KL divergences of the image samples in the unlabeled image set in descending order, and select the unlabeled image samples corresponding to the top n KL divergences. This allows the unlabeled image samples with significantly different prediction results to be selected for expert labeling, denoted as set X. A The method involves expert annotation of poorly performing student and teacher network outputs for further training to improve the fitting ability of the student and teacher networks.

[0090] S33. For the selected unlabeled image samples X A The new label Y is obtained by re-annotating. A Thus, a new label pool D is obtained. A =X A ×Y A .

[0091] S34. The re-annotated image samples are added to the annotated image set and removed from the unannotated image set to obtain the optimized annotated image set and the unannotated image set, respectively. An optimized annotation pool is constructed using the optimized annotated images, the new annotation labels, and the optimized annotation labels.

[0092] Optimized label pool

[0093] Optimized unlabeled image set

[0094] S4. If the number of cycles in the outer loop reaches the second cycle threshold, then stop the outer loop to obtain the final teacher network; otherwise, iterate through steps S2-S3. The cycle of the outer loop is steps S2-S4.

[0095] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for intimal segmentation of carotid artery ultrasound images based on pseudo-label divergence active learning, characterized in that, include: S1. Use the carotid artery ultrasound image set as the training set, divide the training set into an labeled image set and an unlabeled image set, and construct a label pool through the labeled image set and the label. S2. Initialize the student and teacher network parameters, update the student and teacher network parameters through an inner loop, and optimize the labeling, including: S21. The student network parameters are updated sequentially using the supervised loss function based on the labeled image set and the consistency loss function based on the unlabeled image set to obtain the optimized student network. S22. Based on the updated student network parameters, the teacher network parameters are updated using the exponential moving average method. The predicted values ​​are obtained by using the optimized student network through the labeled image set, and the average intersection-union ratio of the predicted values ​​and the labeled labels is obtained. S23. If the current cycle number of the inner loop reaches the first cycle number threshold, then fit the obtained average crossover ratio with the corresponding cycle number of the inner loop to obtain the fitting function; otherwise, iterate steps S21-S22. S24. Differentiate the fitted function and use the differentiated function to obtain the average crossover ratio change rate based on the current and previous inner loop cycles. If the average crossover ratio change rate is less than the early stopping parameter, compare the maximum predicted value of each frame pixel output of the updated teacher network for the labeled image with the label refinement parameter. Update the label based on the comparison result to obtain the optimized label. Otherwise, iterate steps S21-S23 and input the carotid ultrasound image into the optimized teacher network to obtain the intima prediction value. The optimized labeled pool and unlabeled images are obtained through training via the outer loop, resulting in the final teacher network, which includes: S3. Labeling a portion of the unlabeled images in the unlabeled image set to optimize both the labeled and unlabeled image sets, including: S31. Calculate the KL divergence of image samples in the unlabeled image set based on the prediction results of the optimized teacher network and student network. S32. Sort the KL divergence of the image samples in the unlabeled image set in descending order, and select the unlabeled image samples corresponding to the top n KL divergence values. S33. The unlabeled image samples selected are re-labeled to obtain new labels. The re-labeled image samples are added to the labeled image set and removed from the unlabeled image set to obtain the optimized labeled image set and the unlabeled image set, respectively. An optimized label pool is constructed using the optimized labeled images, the new labels, and the optimized labels. S4. If the number of cycles in the outer loop reaches the second cycle threshold, then stop the outer loop to obtain the final teacher network; otherwise, iterate through steps S2-S3. The cycle of the outer loop is steps S2-S4.

2. The method for intimal segmentation of carotid artery ultrasound images based on pseudo-label divergence active learning according to claim 1, characterized in that, The unlabeled image sample x is calculated based on the prediction results of the optimized teacher network and student network. u KL divergence d(x) u )for: d(x u )=mean i,j Div KL (x u (i,j),θ S ,θ T ) Where, mean i,j Div KL (·) represents the unlabeled image sample x u After calculating the KL divergence at each pixel location, the average value is taken, θ S θ represents the optimized student network parameters obtained through the inner loop. T The optimized teacher network parameters are obtained through the inner loop. For the optimized teacher network targeting unlabeled image samples x u The binary classification prediction probability of each pixel in each frame. To optimize the student network for unlabeled image samples x u The binary classification prediction probability of each pixel in each frame, where c = 0 represents an unlabeled image sample x. u For non-inner membranes, c=1 represents unlabeled image sample x. u It is the endometrium.

3. The method for intimal segmentation of carotid artery ultrasound images based on pseudo-label divergence active learning according to claim 1, characterized in that, The obtained average crossover ratio is fitted to the corresponding number of cycles in the inner loop to obtain the fitting function. for: Where a, b, c are fitting parameters, s is the index of the number of inner loop cycles, s = 1, 2, ... k, and k is the current number of inner loop cycles.

4. The method for intimal segmentation of carotid artery ultrasound images based on pseudo-label divergence active learning according to claim 1, characterized in that, The average crossover ratio change rate γ is obtained by using the differentiated function based on the current and previous inner loop cycle numbers. k for: in, To differentiate the fitted function, k is the number of cycles in the current inner loop.

5. The method for intimal segmentation of carotid artery ultrasound images based on pseudo-label divergence active learning according to claim 1, characterized in that, Based on the comparison results, the annotation labels are updated to obtain optimized annotation labels for image x. l The optimized label corresponding to the (i,j) pixel position. for: Where, x l (i,j) represents the labeled image x. l The (i,j) pixel position in the text. F represents the teacher network parameters updated during the k-th inner loop cycle. T (·) represents the optimized teacher network. The optimized teacher network provides the binary classification prediction probability for each pixel in the frame, where λ is the set label refinement parameter, and y(i,j) is the labeled image x. l The original label corresponding to the (i,j) pixel position in the image.

6. The method for intimal segmentation of carotid artery ultrasound images based on pseudo-label divergence active learning according to claim 1, characterized in that, The optimized student network is obtained by updating the student network parameters using a supervised loss function based on the labeled image set and a consistency loss function based on the unlabeled image set, where: Supervision loss function L sup for: Consistency loss function L con for: in, X represents the student network parameters updated during the k-th inner loop cycle. L For the labeled image set, Y L For the labeled set, x l For the labeled image, y represents the label, and F represents the image. T (·) represents the optimized teacher network, F S (·) represents the optimized student network, X U For the unlabeled image set, x u This is an unlabeled image.

7. The method for intimal segmentation of carotid artery ultrasound images based on pseudo-label divergence active learning according to claim 1, characterized in that, The teacher network parameters for the k-th inner cycle are updated using the exponential moving average method based on the updated student network parameters for the k-th inner cycle. for: in, The moving average coefficient is represented by `iter`, which represents the number of iterations in the training process within the current inner loop period `k`. These are the student network parameters updated during the k-th inner loop cycle. The teacher network parameters are updated during the k-th inner loop cycle.

Images