Semi-supervised ultrasound image segmentation method based on SAM and double student-teacher framework

Abstract

The invention discloses a semi-supervised ultrasound image segmentation method based on SAM and a double student-teacher framework. The method comprises the following steps: S1, acquiring a plurality of birth time mother-fetal ultrasound videos; s2, extracting the maternal-fetal ultrasonic video obtained in the step S1 to obtain an extracted maternal-fetal ultrasonic image during delivery; and S3, inputting the extracted maternal and fetal ultrasonic images during delivery into the trained DSTCS network, performing automatic segmentation of an anatomical structure, and outputting a segmentation result. According to the method, good segmentation capability can be provided by using a small amount of labeled data, the precision, robustness and method capability of pubis symphysis fetal head segmentation in the intranatal ultrasound image are remarkably improved, and more reliable technical support is provided for clinical delivery decision making.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ultrasound images, and more particularly, to a semi-supervised ultrasound image segmentation method based on a SAM and a double student-teacher framework. BACKGROUND

[0002] Intrapartum ultrasound is a transformative tool to address maternal and neonatal health challenges, which can reduce fetal mortality by up to 20% through non-invasive real-time monitoring during the delivery process. The International Society of Ultrasound in Obstetrics and Gynecology (ISUOG), the World Association of Perinatal Medicine (WAPM), and the Perinatal Medicine Foundation (PMF) point out that, compared with prenatal ultrasound, intrapartum ultrasound needs to achieve rapid and robust segmentation in a dynamic and variable birth canal environment to replace vaginal examination and thus guide delivery decisions. Accurate segmentation of the pubic symphysis fetal head is crucial for quantifying key biological feature parameters (such as head circumference HC, biparietal diameter BPD, and occipitofrontal diameter OFD) and assessing fetal position, identifying labor abnormalities, and avoiding delivery risks, but still faces challenges in actual clinical applications: (1) Segmentation difficulty in complex environment: In real delivery scenarios, the model needs to segment the fetal head and pubic symphysis at the same time, and there are significant class imbalance and edge blur problems between the two types of structures. Secondly, the dynamic changes in the birth canal environment and individual differences pose extremely high requirements on the accuracy, robustness, and generalization ability of the segmentation algorithm.

[0003] (2) Lack of high-quality labeled data: Current segmentation methods based on convolutional neural networks (CNN) rely heavily on large-scale, high-quality labeled ultrasound image datasets. However, labeled data is not only time-consuming but also requires the participation of ultrasound doctors with professional knowledge, resulting in a limited number of labeled pubic symphysis fetal head ultrasound images, which severely limits the performance and generalization ability of supervised learning models.

[0004] Therefore, although two-dimensional ultrasound automatic segmentation technology has made some progress in the past two decades, the above two core problems have greatly limited its widespread application and effectiveness in actual clinical practice.

[0005] Early medical image segmentation mainly relies on traditional image processing algorithms and machine learning methods. In the field of ultrasound image segmentation, researchers have tried various technical routes: basic algorithms such as edge detection, probability filter, level set method, and hybrid methods combined with domain-specific knowledge. For fetal head contour extraction, Lu et al. proposed an iterative random Hough transform method, which achieved certain results in transabdominal ultrasound images, but its performance is heavily dependent on the quality of the pre-segmentation K-Means. Van et al. used a random forest combined with dynamic programming method, and Liu et al. developed a multi-scale feature asymmetry detection technology, both of which face the challenge of inherent noise and low contrast in ultrasound images. Li et al. integrated gestational age prior knowledge and phase symmetry detection to improve the automation level of head circumference measurement to some extent. However, these traditional methods generally have weak algorithm generalization ability, complex process, and other defects, making it difficult to meet the requirements of segmentation accuracy and robustness in pubic symphysis fetal head segmentation in real-time ultrasound images.

[0006] With the breakthrough of deep learning, fully supervised segmentation networks have shown significant advantages in medical image processing. From the pioneering end-to-end segmentation architecture of FCN to the encoder-decoder structure proposed by U-Net, CNN-based methods have made important progress in pubic symphysis fetal head segmentation tasks. Zhou et al. and Lu et al. improved the U-Net architecture to realize the multi-task joint segmentation of fetal head and pubic symphysis. Bai et al. designed a dual-branch segmentation network (DBSN) that introduced deformable convolution to enhance the model's adaptability to deformed structures. Chen et al. proposed FH-PSSNet, which significantly improved the boundary segmentation accuracy through dual attention mechanism and multi-scale feature selection. However, these supervised learning methods heavily rely on large-scale labeled data, and in the field of medical images with high labeling costs, this limitation directly affects their clinical application value.

[0007] To reduce the dependence on annotation, researchers have begun to explore the application of semi-supervised learning (SSL) in medical image segmentation. The mainstream SSL methods include basic paradigms such as pseudo-label consistency regularization, deep adversarial learning, and hybrid methods. The regularization strategy based on interpolation mixing by Basak et al., the enhanced mean teacher framework by Li et al., and the uncertainty-guided method proposed by Yu et al. have all improved model performance to varying degrees. It is particularly noteworthy that recent research has begun to explore the collaborative training mode of CNN and Transformer. Luo et al. first cross-teach CNN and Transformer networks, but lack an effective pseudo-label quality control mechanism. Xiao et al. improved this deficiency through dual-teacher uncertainty guidance, while Wang et al. directly use the average teacher as consistency guidance for the student network, and Li et al. use soft pseudo-labels and classifier certainty difference minimization to further make the model's predictions consistent and certain. These works lay an important foundation for the DSTCS framework proposed in this application, but existing methods still have obvious shortcomings in the noise sensitivity and boundary blur problems specific to ultrasound images.

[0008] Despite the significant progress made in the field of medical image segmentation, there are still three key limitations in the current study of pubic symphysis and fetal head segmentation. First, existing methods are mainly based on CNN architectures or Transformers, and have not explored the potential of the Segment Anything Model (SAM) large model in ultrasound image segmentation, particularly in the lack of generalization ability when dealing with complex anatomical structures. Second, to address the edge blur problem specific to ultrasound images, existing research has mostly adopted general segmentation strategies, lacking optimization methods specifically targeting the irregular edges of the pubic symphysis. Finally, in terms of model generalization, current semi-supervised methods have not effectively addressed the noise interference and class imbalance problems in ultrasound images, leading to a significant decline in segmentation performance. These limitations make it difficult for existing methods to achieve both high precision and the robustness required in clinical settings. SUMMARY

[0009] The present application aims to provide a semi-supervised ultrasound image segmentation method based on the SAM and dual student-teacher framework to overcome the deficiencies of the prior art.

[0010] To achieve the above-mentioned purposes, the technical solutions adopted by the present application are as follows: The semi-supervised ultrasound image segmentation method based on the SAM and dual student-teacher framework comprises the following steps: S1, obtaining a plurality of maternal-fetal ultrasound videos during childbirth; S2, extracting the maternal-fetal ultrasound videos obtained in step S1 to obtain extracted maternal-fetal ultrasound images during childbirth; S3, inputting the extracted perinatal ultrasound image into the trained DSTCS network to automatically segment the anatomical structure and output the segmentation result.

[0011] Further, the DSTCS network in the step S3 comprises a double student architecture and a single teacher architecture; the double student architecture comprises a CNN model UNet and a SAM model Aop-SAM, and the single teacher architecture adopts the CNN model UNet.

[0012] Further, the double student architecture uses the labeled data to train the model, and the supervision loss is composed of a cross-entropy loss and a Dice loss , and the formula is: ; In the formula, Y represents the label, and and respectively represent the prediction of the CNN model UNet and the SAM model Aop-SAM of the double student architecture on the input image. The double student architecture uses hard pseudo labels for cross supervision , which comprises using the argmax(.) function to generate hard pseudo labels for cross supervision between the two student networks, and then jointly training the network on unlabeled data, and the cross supervision loss of the unlabeled data is defined as: ; The double student architecture uses soft pseudo labels for consistency learning, and the final loss function is: ; In the formula, E is the mean square error loss. The formula for obtaining the soft pseudo label is: ; In the formula, tau is a hyperparameter for controlling the sharpening temperature. The double student architecture uses a classifier to determine the difference , and the loss definition formula is: ; In the formula, represents the element in the mth row and the nth column, and C represents the number of categories.

[0013] Further, the single teacher architecture uses consistency regularization teaching to the double student architecture to minimize the inconsistency of the student model, and the implementation process comprises: First, random noise is introduced in the unlabeled data input into the single teacher architecture network to realize data-level perturbation. Secondly, dropout layers are added to the CNN model UNet network of the student architecture to cause structural perturbation; the parameters of the teacher architecture network are updated from the network of the SAM model Aop-SAM through exponential moving average, denoted as where δ is a weight factor, θ is the network parameter of the SAM model Aop-SAM, and t represents different training rounds. The student-teacher framework is used to minimize the difference between the predictions of the student architecture and the teacher architecture under data and network perturbation, and the consistency loss of the output probability is as follows: where σ and σ' represent different data perturbations and random dropout operations of the network layer, and φ represents the network parameter of the CNN model UNet of the student architecture.

[0014] Further, the DSTCS network adopts an in-situ edge patch superposition strategy, divides the original image into a network of 16x16 pixels, extracts an edge local patch set by positioning the anatomical edge region with a label Y, and re-embeds the enhanced image , specifically denoted as: wherein, is the enhanced image, R represents an edge reorganization function, l is labeled data, and u is unlabeled data, is a model prediction function and its network parameter.

[0015] Further, the DSTCS network enhances the sensitivity of the model to edge geometric features through a neighborhood sensitive loss function; wherein the field difference weight formula is: wherein, is the weight of pixel (i,j), r is the maximum radius of the field, which is set to 5, II is an indicator function, and the difference is 1, otherwise 0, is a k-order field convolution operation, is the prediction probability of the model at (i,j), is the true label at (i,j); The weight c is introduced to adjust the category contribution, and the final loss function is:

[0016] wherein, C represents the number of categories, and c represents the weight of each category.

[0017] Further, the training step of the DSTCS network in step S3 comprises:​​​​ acquire a plurality of perinatal maternal-fetal ultrasound videos; extract the acquired plurality of perinatal maternal-fetal ultrasound videos, obtain extracted perinatal maternal-fetal ultrasound images, and divide them into a certain proportion of a training set, a validation set and a test set; each perinatal maternal-fetal ultrasound image in the training set is input into the DSTCS network for training, and each setting round is verified once with the test set, and the optimal one model is verified with the validation set as the final DSTCS network.

[0018] Further, the training scheme of the DSTCS network contains 30000 iterations, adopts a random gradient descent optimizer, configures a momentum of 0.9 and a weight decay of 0.0001, adopts a batch size of 16, includes equal division of 8 labeled images and 8 unlabeled images to support a semi-supervised learning paradigm, training starts with an initial learning rate of 0.01, uses a clustering-based learning rate strategy to dynamically adjust the initial learning rate, introduces data noise disturbance in the range of [-0.2, 0.2], in the training process, the image size is adjusted to 224*224 according to the nearest neighbor interpolation method, and the training data set is randomly rotated and flipped before training.

[0019] Compared with the prior art, the advantages of the present application are: The present application constructs a dual student heterogeneous collaborative learning mechanism combining CNN and SAM, effectively overcomes the problems that a single model is easily disturbed by ultrasound image noise and the quality of pseudo-label is unstable through hard / soft pseudo-label cross-validation and classifier certainty alignment, and improves the consensus ability and anti-interference ability of the model in the dynamic birth canal environment; The present application designs a boundary perception data enhancement strategy - in-situ edge patch superposition, performs data enhancement under the premise of maintaining the integrity of the pubic symphysis fetal head structure, avoids the destruction of traditional enhancement methods to the pubic symphysis fetal head edge information, and provides more discriminative training samples for the model; The present application proposes a field-sensitive loss function, which strengthens the geometric sensitivity of the model to the edge area through spatial weight mapping, directly optimizes the class imbalance and boundary ambiguity problem between the pubic symphysis and the fetal head, and improves the segmentation accuracy of the key area.

[0020] The present application can provide good segmentation ability with a small amount of labeled data, significantly improve the accuracy, robustness and method ability of pubic symphysis fetal head segmentation in perinatal ultrasound images, and provide more reliable technical support for clinical delivery decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative effort based on these drawings.

[0022] Figure 1 is a flow chart of the semi-supervised ultrasound image segmentation method based on SAM and double student-teacher framework of the present application; Figure 2 is an architecture diagram of the DSTCS network in the present application; Figure 3 is a principle diagram of the semi-supervised ultrasound image segmentation device based on SAM and double student-teacher framework of the present application; Figure 4 is a box plot with scatter points in the present application; Figure 5 is a visual comparison diagram of different methods tested using 20% labeled data in the present application; Figure 6 is a visual comparison diagram of different methods of generalization experiment in the present application; Figure 7 is a diagram of the relationship between segmentation performance and annotation ratio in the present application; Figure 8 is a visualization diagram obtained by different percentages of labeled data in the present application. DETAILED DESCRIPTION

[0023] The preferred embodiments of the present application will be described in detail below with reference to the accompanying drawings, so that the advantages and features of the present application can be more easily understood by those skilled in the art, and the protection scope of the present application can be more clearly and definitely defined.

[0024] Referring to Figure 1 , the present embodiment discloses a semi-supervised ultrasound image segmentation method based on SAM and double student-teacher framework, comprising the following steps: Step S1, obtaining a plurality of perinatal maternal-fetal ultrasound videos.

[0025] In the present embodiment, they can be obtained through professional medical databases and academic platforms, medical institutions and clinical practice, public video material libraries, etc., while strictly complying with ethical and legal norms.

[0026] Step S2, extracting the maternal-fetal ultrasound video obtained in step S1 to obtain the extracted perinatal maternal-fetal ultrasound images.

[0027] Step S3, input the extracted perinatal ultrasound images into the trained DSTCS network, automatically segment the anatomical structures, and output the segmentation results.

[0028] As Figure 2 shown, an overview of the DSTCS network, including the loss function of the single-teacher architecture and the dual-student architecture. L is the labeled data, U is the unlabeled data, U_edg is the data with only edge patches. L h and L sup both use L NW -DICE (neighborhood-sensitive loss Dice function L NW ) for calculation.

[0029] For general semi-supervised learning, the training set consists of two parts: a labeled data set containing N labeled images and an unlabeled data set containing M (original images ) unlabeled images , the complete training data set is denoted as . For an image , its true label Y is available, but for , its true label is not provided.

[0030] Figure 2 The DSTCS network in includes dual-student architecture and single-teacher architecture; the dual-student architecture consists of CNN model UNet (named student1) and SAM model Aop-SAM (named student2), and the single-teacher architecture uses CNN model UNet, DSTCS network solves semi-supervised image segmentation from four aspects: supervised learning , cross-supervision of hard pseudo-labels , consistency learning of soft pseudo-labels , minimum classifier certainty difference and consistency regularization constraint from teacher model , the minimum classifier certainty difference Therefore, the overall training loss function of student1 or student2 can be defined as: ; where α, β, γ, µ are the weighting weights, and in the proposed collaborative training process, α = 0.5, β = 1.0, γ = 3.0. In addition, µ is set to 0.1. It is worth noting that in addition to is the common part of the two students, other loss functions are similar between the two students.

[0031] In this embodiment, the dual student architecture uses the labeled data to train the model, and the supervision loss is composed of the cross-entropy loss and the Dice loss , whose formula is: ; wherein and represent the predictions of the CNN model UNet and the SAM model Aop-SAM of the dual student architecture on the input image, respectively, and Y represents the label.

[0032] In this embodiment, the dual student architecture uses hard pseudo-labels for cross-supervision . The difference between the predictions of the CNN and the SAM mainly comes from the different spatial sensitivity and semantic abstraction ability of the output features of the two: the CNN is biased towards local details and progressive induction, while the SAM focuses on global context and overall perception and zero-shot generalization. Based on these prediction results, this embodiment includes using the argmax(.) function to generate hard pseudo-labels for cross-supervision between the two student networks, and then jointly training the networks on unlabeled data. It is worth noting that there is no gradient backpropagation between them in each mini-batch training. The cross-supervision loss of the unlabeled data is defined as: ; In this embodiment, the dual student architecture uses soft pseudo-labels for consistency learning. The hard pseudo-labels generated by the maximum confidence inevitably have noise, which may cause confusion bias during segmentation training. In order to further reduce the noise of the hard pseudo-labels and focus on the challenging areas of the unlabeled data, a sharpening function is used to generate soft pseudo-labels, which can reduce the prediction uncertainty of the model. The way to obtain the soft pseudo-labels is as follows: ; wherein τ is a hyperparameter used to control the sharpening temperature, and is set to 0.1 in this experiment. Consistency learning is performed between the probability output of one model and the soft pseudo-label of the other model, which reduces the difference between the outputs of the models, and the predictions in these highly uncertain areas should be consistent. The purpose of this design can be summarized as follows: (1) implement consistency constraints by encouraging all models to produce constant outputs; (2) learn models to produce low-entropy results under the supervision of soft pseudo-labels, as an entropy minimization constraint. The final loss function of CLS (Consistency learning with soft pseudo-labels, i.e. Ls, consistency learning with soft pseudo-labels) is: ; wherein E is the mean square error (MSE) loss.

[0033] In this embodiment, the dual student architecture uses a classifier determinacy disparity The diversity caused by the structural difference between CNN and SAM can affect the performance of the model, and the diversity of the two classifiers is studied by The difference between the directly predicted distribution can ignore the inter-class correlation, which can potentially affect the confidence of the prediction. Therefore, the correlation matrix is predicted by the dual classifier to study the difference between the classifiers. The matrix A contains all the probabilities that the two classifiers disagree on, where the sum of the diagonal elements represents the consistency of the classifier prediction, and the off-diagonal elements reflect the uncertainty of the prediction. The goal of this strategy is to maximize the former and minimize the latter. CDD (Classifier Determinacy Disparity, that is The CDD loss function is defined as: ; In the formula, represents the element in the mth row and the nth column, and C represents the number of classes.

[0034] In this embodiment, the single teacher architecture uses consistency regularization teaching to the dual student architecture to minimize the inconsistency of the student model, and improves the generalization of the model by implementing two part perturbation operations. The implementation process includes: First, random noise is introduced into the unlabeled data input into the single teacher architecture network to realize data-level perturbation; second, a dropout layer is added to the CNN model UNet network of the dual student architecture to cause structural perturbation; the parameters of the single teacher architecture network are updated from the network of student2 by exponential moving average (EMA), denoted as The parameters of the teacher network are updated, where δ is a weight factor, θ is the network parameter of the SAM model Aop-SAM, and t represents different training rounds; the dual student-teacher framework is used to minimize the difference between the predictions of the dual student architecture and the single teacher architecture under data and network perturbation, and the consistency loss of the output probabilities of the two is as follows: ; In the formula, σ and σ' represent different data perturbations and random dropout operations of the network layer, and φ represents the network parameter of student1.

[0035] In this embodiment, the DSTCS network uses an in-situ edge patch superposition strategy to divide the original image into a network of 16x16 pixels, and extracts an edge local patch set by positioning the anatomical edge region with the label Y, and re-embeds the enhanced image , which is specifically represented as: ; In the formula, is an enhanced image, R represents an edge reorganization function, I is labeled data, and u is unlabeled data, is a model prediction function and its network parameters.

[0036] In the embodiment, the DSTCS network enhances the sensitivity of the model to edge geometric features through a neighborhood sensitive loss function; wherein the domain difference weight formula is: ; In the formula, is the weight of pixel (i, j), r is the maximum radius of the field, which is set to 5, II is an indicator function, and the difference is 1, otherwise 0, is a k-order field convolution operation, is the prediction probability of the model at (i, j), is the true label at (i, j); the convolution kernel size is dynamically changed, which depends on the parameter r; in the function, the loop is from 1-r: at each i, the size of the convolution kernel is calculated as: kernel_size = 2 * i + 1; For example: When i=1, the convolution kernel size = 2×1 + 1 = 3×3 When i=2, the convolution kernel size = 2×2 + 1 = 5×5 When i=3, the convolution kernel size = 2×3 + 1 = 7×7 ... When i=r, the convolution kernel size = 2×r + 1 The weight c is introduced to adjust the category contribution, and the final loss function is: ; In the formula, C represents the number of categories, and c represents the weight of each category.

[0037] In the embodiment, the training step of the DSTCS network in step S3 comprises: Step S31, obtaining a plurality of maternal-fetal ultrasound videos during delivery.

[0038] Step S32, extracting the obtained plurality of maternal-fetal ultrasound videos during delivery to obtain extracted maternal-fetal ultrasound images during delivery, and dividing them into a certain proportion of training set, validation set and test set.

[0039] Step S33, input each intrapartum maternal-fetal ultrasound image of the training set into the DSTCS network, train, verify once every set round with the test set, and then verify the optimal one model with the validation set as the final DSTCS network.

[0040] In this embodiment, the DSTCS network model training is implemented on an Ubuntu 20.04 operating system, using Python 3.8, PyTorch 1.10 and CUDA 11. The network training scheme contains 30000 iterations, using a stochastic gradient descent (SGD) optimizer, with a momentum of 0.9 and a weight decay of 0.0001. The batch size used in this paper is 16, including an equal partition of 8 labeled images and 8 unlabeled images to support the semi-supervised learning paradigm. The training starts with an initial learning rate of 0.01, and the initial learning rate is dynamically adjusted using a clustering-based learning rate strategy. In addition, this paper introduces data noise disturbance in the range of [-0.2, 0.2] to enhance the robustness of the model to input changes. During the training process, the image size is adjusted to 224x224 according to the nearest neighbor interpolation method to meet the requirements of the network. The training data set is randomly rotated and flipped before training. Specifically, first, a random integer k between 0 and 3 is generated, and then the image is rotated k times using the rot90 function of numpy. Next, a random integer axis between 0 and 1 is generated, and finally the image is flipped along the specified axis using the flip function of numpy.

[0041] As shown in Figure 3 To solve the above technical problems, the embodiments of the present application also provide a semi-supervised ultrasound image segmentation device based on SAM and double student-teacher framework, which adopts the technical scheme as follows: the semi-supervised ultrasound image segmentation device based on SAM and double student-teacher framework comprises: an acquisition module 1 configured to acquire a plurality of intrapartum maternal-fetal ultrasound videos; an extraction module 2 configured to extract the obtained maternal-fetal ultrasound videos to obtain extracted intrapartum maternal-fetal ultrasound images; and an automatic segmentation module 3 configured to input the extracted intrapartum maternal-fetal ultrasound images into a trained DSTCS network, perform automatic segmentation of anatomical structures, and output segmentation results.

[0042] The effects and advantages of the present application will be further introduced through the embodiments.

[0043] The DSTCS network of this embodiment is evaluated by participating in two MICCAI challenges.

[0044] Baseline experiment: The MICCAI 2023 challenge dataset is used, which includes 5101 labeled images. To ensure the scientificity of model training, the dataset is randomly divided into training set, validation set and test set, the proportion is 70%, 10% and 20%. All training images are subjected to uniform preprocessing operations, including denoising and normalization. In order to carry out the generalization experiment, the method uses the dataset from the MICCAI 2024 challenge, which includes 300 images.

[0045] The method of the embodiment is compared with several state-of-the-art SSL methods, including Mean Teacher (MT), Interpolation Consistency (ICT), Uncertainty-aware Mean Teacher (UAMT), Deep Adversarial Network (DAN), Deep Co-training (DCT), Cross-Loss Balancing (CLB), Cross Pseudo Supervision (CPS), Cross Teaching between CNN and Transformer (CTCT), Feature Similarity and Reliable Region Enhanced Network (FSRENet), and Self-Ensemble Method Based on Consistency-aware Pseudo Label (S4CVnet). As can be seen from Table 1, in the experiment with only 20% labeled data, the method of the embodiment shows a comprehensive leading performance advantage. Taking the key indicator DSC (the higher the value, the better) as an example, OURS achieves 0.871, 0.952 and 0.911 on the PS (pubic symphysis), FH (fetal head) and PSFH (pubic symphysis and fetal head) tasks, respectively, which is significantly better than the suboptimal methods CTCT (0.780), S4CVnet (0.905) and CTCT (0.842), with a relative improvement of 11.7%, 5.2% and 8.2%. In terms of ASD indicators, OURS comprehensively outperforms other methods with 0.514 (PS), 0.159 (FH) and 0.336 (PSFH). The HD95 indicator also shows the absolute advantage of OURS, and these data fully prove that the OURS method has significant superiority in different segmentation tasks and evaluation dimensions. From the perspective of the overall performance distribution of different methods, as shown in FIG. 1, the performance of the method of the embodiment is more stable and reliable, and the performance of the method of the embodiment is more stable and reliable. Figure 4 For the box plot with scattered points, the performance distribution of different methods tested using 20% labeled data is compared. Each scattered point represents a test case result. From the box plot, it can be further analyzed that the performance fluctuation range of the generalization is smaller, and the abnormal data is significantly less, which is crucial for decision support relying on continuous and reliable feedback in real-time clinical scenarios. In addition, from the perspective of the overall performance distribution of different methods, as shown in FIG. 1, the performance of the method of the embodiment is more stable and reliable, and the performance of the method of the embodiment is more stable and reliable. Figure 5 For visual comparison of different methods tested using 20% labeled data. Red represents the region labeled by doctors, green represents the prediction result of pubic symphysis, and yellow represents the prediction result of fetal head. From the visual output, it can be seen that the method of the embodiment is superior to other methods in both boundary positioning accuracy and region overlap consistency, which directly improves the ability to monitor the dynamic changes of key structures in real time, and helps to enhance the timeliness and accuracy of intervention decision-making.

[0046] Table 1

[0047] Generalization experiment: As can be seen from Table 2, the DSC, ASD and HD95 indexes of the pubic symphysis fetal head are only 0.025, 0.063 and 1.905 different from the basic experiment. Moreover, the DSC, ASD and HD95 indexes of the method of the embodiment are significantly higher than those of other semi-supervised methods. It is worth noting that the baseline model generally appears the phenomenon of pubic symphysis region missing segmentation, while the method of the embodiment completely and accurately realizes the segmentation of the two key structures of pubic symysis and fetal head in all test images, and there is no missing example. Figure 6 For visual comparison of different methods in generalization experiment. Red represents the region marked by the doctor, green represents the pubic symphysis prediction result, and yellow represents the fetal head prediction result. The visualization result further shows that the method of the embodiment is significantly better than the comparison methods in complex boundary preservation and structural integrity, and shows excellent generalization ability and clinical applicability.

[0048] Table 2

[0049] Among them, the bold represents the best result, and the underlined represents the second best result. PS is pubic symphysis, FH is fetal head, and PSFH is pubic symphysis fetal head. Lossing represents the number of images that the model cannot segment.

[0050] The embodiment also systematically evaluates the influence of different annotation data proportions (10%-80%) on the segmentation performance of the model. As can be seen from Table 3, Figure 7 It can be seen that as the proportion of labeled data in the training set increases, the model shows consistent performance improvement in all key indicators of pubic symphysis and fetal head segmentation. When the labeled data increases from 10% to 80%, the Dice coefficient (DSC) of the PS region continuously increases from 0.885 when the data is 10% to 0.92 when the data is 80%; the DSC of the fetal head region also increases from 0.95 to 0.96, in terms of segmentation accuracy, PSFH-HD95 decreases from 1.90 mm to 1.05 mm, and PSFH-ASD also decreases from 0.32 mm to 0.14 mm. Reflecting the importance of sufficient labeled samples for improving the generalization ability and robustness of the model.

[0051] Notably, the improvement of PS-DSC tends to flatten out after the annotation ratio reaches 40%, indicating that the model's recognition ability for the target subject has approached saturation; however, the boundary accuracy indicators (ASD, HD95) still show a significant decline from 40% to 80%, suggesting that larger-scale annotation data continues to help the model learn more refined boundary features. This finding has important practical significance: in practical applications, if the goal is to achieve high DSC, only about 40% of the labeled data is needed to achieve good performance; but if you pursue extremely high boundary segmentation quality, you still need to provide more sufficient labeled samples. Figure 8 The visualization chart for different percentages of labeled data is obtained from Figure 8 The visualization chart can intuitively feel that the segmentation result based on a small amount of labeled data has reached the clinical practical threshold and can meet the needs of most clinical scenarios. If you need to achieve edge adhesion to the entire structure and support more accurate measurement, it is necessary to further improve the model performance by increasing the labeled data.

[0052] This embodiment also explores the impact of different models on the framework. Table 3 experimental data shows that when Aop-SAM is used as the main segmenter (S1), its PS-DSC reaches 0.896, which is 5.3% and 9.3% higher than UNet (0.851) and SUNet (0.820); FH-DSC is improved to 0.943, which is better than UNet (0.900) and SUNet (0.920). More importantly, Aop-SAM shows outstanding advantages in boundary accuracy: PS-ASD is as low as 0.294mm, which is 62% lower than the baseline model (UNet: 0.781mm) and 70% lower than SUNet (0.966mm); FH-HD95 is only 1.624mm, which is 74.6% lower than UNet. The significant improvement of the joint indicators PSFH-ASD (0.284mm) and PSFH-HD95 (1.793mm) indicates that Aop-SAM not only can independently improve the segmentation quality of a single target, but also can effectively promote the collaborative learning among multiple structures, greatly improving the collective accuracy of contour drawing while maintaining high volume overlap accuracy, providing more reliable segmentation results for medical image analysis tasks.

[0053] Table 3 Performance comparison of different model configurations in experiments

[0054] To further investigate the influence of class_weight and r on the segmentation performance of the model, the embodiment systematically evaluates the performance of different hyperparameter configurations on the background, pubic symphysis and fetal head segmentation tasks. According to Table 4, class_weight = [1.0, 2.0, 1.0] and r = 5 perform best on multiple indicators, achieving the highest DICE (0.927), the lowest ASD (0.181), and lower PSFH-HD95 (1.439) in PSFH. This reflects the best balance between overall segmentation accuracy and boundary consistency for this configuration.

[0055] Specifically, the setting of class_weight is particularly crucial for the segmentation of pubic symphysis (PS). Compared with the baseline weight [1.0, 1.0, 1.0], moderately increasing the weight of PS to 1.5 or 2.0 can significantly improve DSC and reduce ASD and HD95, indicating that assigning a higher weight to PS helps the model focus on this difficult-to-segment structure and alleviate the class imbalance problem. However, when the weight is too high (such as 3.0), the performance of PS and overall PSFH declines, indicating that excessive weight bias may lead to an imbalance in the optimization direction of the model, compromising the discrimination ability of other classes. In terms of the local range hyperparameter r, r = 5 can achieve optimal or near-optimal performance under most experimental configurations, indicating that this scale can effectively capture the local contextual information that is beneficial for segmentation. When r increases to 7 or decreases to 3, most indicators show slight degradation, indicating that too large a context range may introduce irrelevant noise, while too small a range may not provide enough structural information.

[0056] In summary, the combination of class_weight = [1.0, 2.0, 1.0] and r = 5 is the optimal configuration, which significantly improves the segmentation accuracy of the model for multiple structures, especially the pubic symphysis. This provides empirical evidence for the selection of key hyperparameters in semi-supervised medical image segmentation.

[0057] Table 4 Hyperparameter analysis of class_weight and local range in NW-Dice loss

[0058] Although the embodiments of the present application are described in combination with the drawings, the patent owner can make various modifications or modifications within the scope of the appended claims, as long as they do not exceed the protection scope described in the claims of the present application, and should be within the protection scope of the present application.

Claims

1. A semi-supervised ultrasound image segmentation method based on SAM and a dual student-teacher framework, characterized in that, Includes the following steps: S1. Acquire multiple intrapartum maternal and fetal ultrasound videos; S2. Extract the maternal-fetal ultrasound video obtained in step S1 to obtain the extracted intrapartum maternal-fetal ultrasound images. S3. Input the extracted intrapartum maternal and fetal ultrasound images into the trained DSTCS network to perform automated segmentation of anatomical structures and output the segmentation results.

2. The semi-supervised ultrasound image segmentation method based on SAM and a dual student-teacher framework according to claim 1, characterized in that, The DSTCS network in step S3 includes a dual-student architecture and a single-teacher architecture; the dual-student architecture consists of a CNN model UNet and a SAM model Aop-SAM, and the single-teacher architecture uses a CNN model UNet.

3. The semi-supervised ultrasound image segmentation method based on SAM and a dual student-teacher framework according to claim 2, characterized in that, The dual-student architecture uses labeled data to train the model, with supervised loss. Cross-entropy loss and Dice loss The composition, its formula is: ； In the formula, and Y represents the predictions of the input image by the dual-student architecture CNN model UNet and the SAM model Aop-SAM, respectively, where Y represents the label. The dual-student architecture uses hard pseudo-labels for cross-supervision. This includes using the argmax(.) function to generate hard pseudo-labels for cross-supervision between two student networks, and then jointly training the networks on unlabeled data. The cross-supervision loss for unlabeled data is defined as: ； The dual-student architecture uses soft pseudo-labels for consistency learning, and its final loss function is: ； In the formula, E represents the mean square error loss; The formula for obtaining soft pseudo tags is: ； In the formula, τ is a hyperparameter used to control the sharpening temperature; The dual-student architecture uses a classifier for deterministic differential. The loss is defined by the following formula: ； In the formula, Let C represent the element in the m-th row and n-th column, and let C represent the number of categories.

4. The semi-supervised ultrasound image segmentation method based on SAM and a dual student-teacher framework according to claim 2, characterized in that, The single-teacher architecture employs consistent regularization teaching in the dual-student architecture to minimize inconsistencies in the student models. Its implementation process includes: First, random noise is introduced into the unlabeled data of the input single-teacher architecture network to achieve data-level perturbation; Secondly, dropout layers were added to the UNet CNN model with a dual-student architecture to induce structural perturbations; the parameters of the single-teacher architecture network... It is updated from the SAM model Aop-SAM network using an exponential moving average, expressed as: Update the parameters of the teacher network, where δ is a weight factor, θ is the network parameter of the SAM model Aop-SAM, and t represents the different training rounds; The described two-student-teacher framework is used to minimize the difference in predictions between the two-student architecture and the single-teacher architecture under data and network perturbations. The consistency loss of the output probabilities of the two is as follows: ； In the formula, σ and σ′ represent different data perturbations and random dropout operations in the network layer, and φ represents the network parameters of the dual-student architecture by the CNN model UNet.

5. The semi-supervised ultrasound image segmentation method based on SAM and a dual student-teacher framework according to claim 2, characterized in that, The DSTCS network employs an in-situ edge patch overlay strategy, dividing the original image into a 16×16 pixel network. It extracts local edge patch sets by locating dissected edge regions using the Y-label and then re-embedding them into the enhanced image. Specifically, it is expressed as: ； In the formula, For the enhanced image, R represents the edge reconstruction function, l is the labeled data, and u is the unlabeled data. This refers to the model prediction function and its network parameters.

6. The semi-supervised ultrasound image segmentation method based on SAM and a dual student-teacher framework according to claim 2, characterized in that, The DSTCS network enhances the model's sensitivity to edge geometric features through a neighborhood-sensitive loss function; the neighborhood difference weight formula is as follows: ； In the formula, Here, represents the weight of pixel (i,j), r is the maximum neighborhood radius (set to 5), and Ⅱ is the indicator function, with a difference of 1 otherwise 0. It is a k-order neighborhood convolution operation. Let be the predicted probability of the model at (i,j). Let (i,j) be the true label at position (i,j); Introducing weight c to adjust class contributions, the final loss function is: ； In the formula, C represents the number of categories, and c represents the weight of each category.

7. The semi-supervised ultrasound image segmentation method based on SAM and a dual student-teacher framework according to claim 2, characterized in that, The training steps of the DSTCS network in step S3 include: Acquire multiple intrapartum maternal and fetal ultrasound videos; Multiple intrapartum maternal and fetal ultrasound videos were extracted to obtain extracted intrapartum maternal and fetal ultrasound images, which were then divided into a certain proportion of training set, validation set, and test set. Each intrapartum ultrasound image of the mother and fetus in the training set is input into the DSTCS network for training. The model is validated once with the test set in each set round, and then the optimal model is validated with the validation set as the final DSTCS network.

8. The semi-supervised ultrasound image segmentation method based on SAM and a dual student-teacher framework according to claim 7, characterized in that, The training scheme of the DSTCS network includes 30,000 iterations, employs a stochastic gradient descent optimizer with a momentum of 0.9 and a weight decay of 0.0001, and uses a batch size of 16, comprising equal splits of 8 labeled images and 8 unlabeled images to support a semi-supervised learning paradigm. Training begins with an initial learning rate of 0.01, which is dynamically adjusted using a clustering-based learning rate strategy. Data noise perturbation is introduced within the range of [-0.2, 0.2]. During training, the image size is adjusted to 224 × 224 using nearest neighbor interpolation. The training dataset is randomly rotated and flipped before training.

Images