A cross-center trustworthy semi-supervised echocardiogram image segmentation method

Abstract

This invention discloses a cross-center, reliable, semi-supervised method for cardiac ultrasound image segmentation, belonging to the fields of medical image processing and computer vision. It aims to solve problems such as high annotation costs, poor cross-center generalization ability, and unreliable predictions in cardiac ultrasound image segmentation. The method includes: extracting multi-scale features using the image encoder of the segmentation base model SAM; designing a lightweight decoder to output a segmentation probability map and uncertainty estimation; employing semi-supervised training based on the mean teacher framework at the source centers, combined with a pseudo-label filtering strategy of consistency regularization and uncertainty weighting; and introducing an online adaptive mechanism at the target centers that does not require source data. Through multiple enhanced consistency constraints and pseudo-label selection and confidence weighting that fuses heterogeneity and cognitive uncertainty, robust cross-center segmentation is achieved. This invention effectively reduces annotation costs, improves segmentation accuracy and cross-center generalization ability, and enhances clinical reliability.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and computer vision technology, and in particular relates to a cross-center reliable semi-supervised cardiac ultrasound image segmentation method. Background Technology

[0002] Cardiovascular disease is a leading cause of death and disability worldwide. Echocardiography, due to its non-invasive, low-cost, repeatable, and real-time bedside imaging capabilities, has become a first-line tool for assessing cardiac structure and function. Accurate segmentation of the cardiac chambers, including the left ventricle, right ventricle, left atrium, and right atrium, is fundamental for calculating key quantitative indicators such as ventricular volume and ejection fraction, and is crucial for disease diagnosis, treatment planning, and prognostic assessment. Currently, deep learning, particularly encoder-decoder architectures based on convolutional neural networks, has achieved significant success in medical image segmentation tasks. However, the performance of these models heavily relies on large-scale, high-quality manually labeled data. In the field of cardiac ultrasound, frame-by-frame accurate annotation is not only time-consuming and labor-intensive but also requires experienced experts, making large-scale labeled datasets difficult to obtain. The high cost of annotation in practical clinical applications has become a major bottleneck.

[0003] To reduce annotation costs, semi-supervised learning methods have been introduced, utilizing a large amount of unlabeled data combined with a small amount of labeled data for model training. Existing semi-supervised segmentation methods mainly follow paradigms such as consistency regularization, pseudo-label generation, and adversarial learning. However, while these methods perform well on single-center datasets, in actual clinical deployments, differences in ultrasound equipment models, acquisition protocols, operators, and patient groups across different medical centers lead to significant domain shifts in image data. This causes a sharp decline in model performance and insufficient generalization ability when applied across centers (i.e., target centers). This domain shift problem stems from the heterogeneity of the clinical environment; for example, different manufacturers' equipment produces images with varying contrast and noise levels, or demographic differences in patient groups lead to variations in anatomical structures, making it difficult for the model to adapt to the data distribution of new centers.

[0004] In recent years, the emergence of basic models has provided new opportunities for medical image analysis. For example, segmentation models, pre-trained on large-scale general data, have learned powerful transferable visual representations. However, relying solely on basic models is insufficient to address the credibility challenges in clinical deployment. In cases of domain shift or poor image quality, model predictions may become unreliable, and the lack of quantitative assessment of prediction reliability makes it difficult to establish trust in clinical practice. Simultaneously, existing semi-supervised methods suffer from high pseudo-label noise in cross-center settings, impacting model performance and stability. This is because domain shift leads to a decline in the quality of pseudo-labels generated by the teacher model, and traditional methods cannot effectively filter out reliable pseudo-labels. Addressing these issues has been challenging, including how to achieve cross-center adaptation without accessing source center data and how to integrate uncertainty estimation to improve pseudo-label quality and model robustness. These challenges limit the widespread application of semi-supervised learning in real-world multi-center scenarios. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention proposes a cross-center, reliable, semi-supervised cardiac ultrasound image segmentation method, which effectively reduces annotation costs, improves segmentation accuracy and cross-center generalization ability, and enhances clinical reliability.

[0006] To achieve the above objectives, the present invention provides a cross-center reliable semi-supervised cardiac ultrasound image segmentation method, comprising:

[0007] Multiscale features of input cardiac ultrasound images are extracted using an image encoder based on a segmentation model.

[0008] A lightweight decoder is designed to fuse and upsample the multi-scale features, and output a segmentation probability map and a random uncertainty map.

[0009] At the source center, a mean teacher framework is used for semi-supervised training, combined with supervised loss, consistency regularization and uncertainty-weighted pseudo-label filtering strategy.

[0010] At the target center, an online passive adaptive mechanism is introduced to achieve robust cross-center segmentation through multiple enhanced consistency constraints and pseudo-label selection and confidence weighting that integrates heterogeneity and cognitive uncertainty.

[0011] Optionally, the image encoder utilizing the segmentation-based model can extract multi-scale features from the input echocardiogram image, including:

[0012] The image encoder of the Segment Anything Model, pre-trained on a large-scale general dataset, is used as the feature extractor;

[0013] The image encoder is based on a visual Transformer architecture;

[0014] The image encoder extracts multi-scale feature maps from the input image, including feature maps from multiple layers;

[0015] The parameters of the image encoder are kept frozen during training to preserve general visual knowledge.

[0016] Optionally, designing a lightweight decoder to fuse and upsample multi-scale features includes:

[0017] Design a lightweight UNet-like decoder;

[0018] Convolutional projection and upsampling to a common spatial resolution are performed on each encoder feature map;

[0019] By fusing all the upsampled feature maps, a decoded feature map is obtained;

[0020] Fine-grained spatial details are gradually recovered through a series of upsampling and skip connections to output a segmentation probability map and a random uncertainty map.

[0021] Optionally, the output segmentation probability map and random uncertainty map include:

[0022] The segmentation probability map is output through the segmentation branch, which includes a convolutional layer and a softmax activation function, to obtain the class probability distribution of each pixel;

[0023] The uncertainty branch outputs a random uncertainty map, which includes convolutional layers and ReLU activation functions, and predicts a pixel-level variance map to measure the inherent noise of the data.

[0024] Optionally, the method further includes a cognitive uncertainty estimation process, which includes:

[0025] During inference, multiple random forward propagations are performed using Monte Carlo Dropout to obtain multiple prediction probability maps;

[0026] The average of the multiple predicted probability maps is calculated as the final prediction;

[0027] The variance of the multiple predicted probability maps is calculated as cognitive uncertainty, which is used to quantify the uncertainty of the model due to insufficient knowledge.

[0028] Optionally, semi-supervised training using the mean-teacher framework at the source center includes:

[0029] Using a student network and a teacher network, wherein the parameters of the teacher network are an exponential moving average of the parameters of the student network;

[0030] For labeled samples, a supervised loss is calculated, which includes a combination of Dice loss and cross-entropy loss;

[0031] For unlabeled samples, two different augmented versions are generated and input into the student network and teacher network respectively. Consistency regularization is applied by minimizing the KL divergence of the two predictions.

[0032] Optional, uncertainty-weighted pseudo-label filtering strategies include:

[0033] Generating pseudo-labels from teacher network predictions using unlabeled samples;

[0034] Calculate the overall uncertainty for each pixel, where the overall uncertainty is the sum of accidental uncertainty and cognitive uncertainty;

[0035] Select pixels with uncertainty below a threshold and assign them exponentially decaying weights;

[0036] The pseudo-label loss is calculated based on weighted cross-entropy loss and used to train the student network.

[0037] Optionally, introducing an online passive adaptive mechanism at the target center includes:

[0038] For each sample at the target center, multiple random augmented versions are generated, including speckle noise, geometric transformation, and intensity perturbation;

[0039] Calculate the average forecast for all enhanced versions;

[0040] Minimize the KL divergence between each enhanced version prediction and the average prediction as a multi-enhancement consistency constraint.

[0041] Optional, pseudo-label selection and confidence weighting that integrates heterogeneity and cognitive uncertainty include:

[0042] Calculate the overall uncertainty for each pixel, where the overall uncertainty is the sum of accidental uncertainty and cognitive uncertainty;

[0043] Pixel-level weights are calculated based on uncertainty, and these weights are negatively correlated with uncertainty.

[0044] An uncertainty-weighted multi-enhancement consistency loss is used for adaptive model updates.

[0045] Optionally, the online passive adaptive mechanism also includes a stability control process, which includes:

[0046] Use a small learning rate for model updates;

[0047] Limit the number of update steps for each sample;

[0048] A teacher model is used to provide stable guidance, the parameters of which are exponential moving averages of the student model parameters, to prevent catastrophic forgetting or overfitting of the model.

[0049] Technical Advantages of this Invention: This invention discloses a cross-center, reliable, semi-supervised cardiac ultrasound image segmentation method. By integrating a base model and uncertainty estimation, and combining a two-stage learning strategy, it effectively reduces the dependence of cardiac ultrasound image segmentation on large amounts of labeled data. Semi-supervised learning fully utilizes unlabeled data, significantly reducing labeling costs. This invention improves the model's generalization ability in cross-center scenarios. Utilizing the powerful feature representation capabilities of the base model and an online passive adaptive mechanism, it effectively alleviates the domain shift problem caused by differences in equipment, protocols, and patient groups. Simultaneously, by simultaneously estimating accidental and cognitive uncertainties, this invention provides pixel-level confidence information for the segmentation results, enhancing the credibility of the results and assisting clinical decision-making. Furthermore, this uncertainty is used to guide pseudo-label selection and loss weighting during training and adaptation, improving the model's robustness. Moreover, the target center adaptation process does not require access to source center data, complying with clinical data privacy and security requirements. Attached Figure Description

[0050] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0051] Figure 1 This is a schematic diagram of the network architecture of a cross-center trusted semi-supervised cardiac ultrasound image segmentation method according to an embodiment of the present invention;

[0052] Figure 2 This is a visualization of the test results of the source center dataset CardiacUDA when the percentage of labeled data is 5%, 10%, and 20%, respectively, according to an embodiment of the present invention.

[0053] Figure 3 This is a visualization of the test results of the model on the target center private dataset when the percentage of labeled data in the CardiacUDA training set is 5%, 10%, and 20%, respectively, according to an embodiment of the present invention.

[0054] Figure 4 This is a flowchart illustrating a cross-center trusted semi-supervised cardiac ultrasound image segmentation method according to an embodiment of the present invention, wherein (a) is a flowchart of semi-supervised training of the model at the source center, and (b) is a flowchart of online passive adaptation of the model at the target center. Detailed Implementation

[0055] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0056] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0057] like Figure 4 As shown, this embodiment provides a cross-center reliable semi-supervised cardiac ultrasound image segmentation method, including:

[0058] Multiscale features of input cardiac ultrasound images are extracted using an image encoder based on a segmentation model.

[0059] A lightweight decoder is designed to fuse and upsample the multi-scale features, and output a segmentation probability map and a random uncertainty map.

[0060] At the source center, a mean teacher framework is used for semi-supervised training, combined with supervised loss, consistency regularization and uncertainty-weighted pseudo-label filtering strategy.

[0061] At the target center, an online passive adaptive mechanism is introduced to achieve robust cross-center segmentation through multiple enhanced consistency constraints and pseudo-label selection and confidence weighting that integrates heterogeneity and cognitive uncertainty.

[0062] Furthermore, such as Figure 1 As shown, the image encoder using the segmentation base model extracts multi-scale features from the input echocardiogram image, including:

[0063] An image encoder of the Segment Anything Model (SAM) pre-trained on a large-scale general dataset is used as a feature extractor;

[0064] The image encoder is based on a visual Transformer architecture;

[0065] The image encoder extracts multi-scale feature maps from the input image, including feature maps from multiple layers;

[0066] The parameters of the image encoder are kept frozen during training to preserve general visual knowledge.

[0067] Specifically, semi-supervised pre-training based on source centers:

[0068] Utilizing a small number of annotated echocardiogram images from source centers and a large number of unannotated images, semi-supervised learning is performed using a mean-teacher framework, and uncertainty-guided pseudo-label generation and filtering are introduced. The encoder employs an image encoder pre-trained on a large-scale general dataset, the Segment Anything Model, as the feature extractor. This encoder, based on a visual Transformer architecture, is used to extract features from the input echocardiogram images. Multi-scale feature maps:

[0069] ;

[0070] in, Indicates the first l Feature map of the layer This represents the encoder of the SegmentAnything Model. The encoder parameters are typically frozen during training to preserve general visual knowledge.

[0071] Furthermore, such as Figure 1 As shown, the design of a lightweight decoder for multi-scale feature fusion and upsampling includes:

[0072] Design a lightweight UNet-like decoder;

[0073] Convolutional projection and upsampling to a common spatial resolution are performed on each encoder feature map;

[0074] By fusing all the upsampled feature maps, a decoded feature map is obtained;

[0075] Fine-grained spatial details are gradually recovered through a series of upsampling and skip connections to output a segmentation probability map and a random uncertainty map.

[0076] Specifically, a lightweight UNet-like decoder is designed for multi-scale feature fusion and upsampling. For each encoder feature map... conduct Convolutional projection and upsampling to a common spatial resolution are then performed, followed by fusion:

[0077] ;

[0078] in express convolution, Indicates upsampling, Indicates a fusion operation. This represents the feature map output by the decoder.

[0079] Fine-grained spatial details are then gradually recovered through a series of upsampling and skip connections.

[0080] Furthermore, the output segmentation probability map and random uncertainty map include:

[0081] The segmentation probability map is output through the segmentation branch, which includes a convolutional layer and a softmax activation function, to obtain the class probability distribution of each pixel;

[0082] The uncertainty branch outputs a random uncertainty map, which includes convolutional layers and ReLU activation functions, and predicts a pixel-level variance map to measure the inherent noise of the data.

[0083] Specifically, the decoder produces two outputs:

[0084] Segmentation probability graph: through convolutional layers and Softmax activation function Obtain the class probability distribution for each pixel:

[0085] ;

[0086] in, , C Number of categories;

[0087] Random uncertainty graph: through another convolutional layer The pixel-level variance map predicted by the ReLU activation function is used to measure the inherent noise of the data.

[0088] ;

[0089] in, This represents a graph depicting randomness and uncertainty.

[0090] Furthermore, the method also includes a cognitive uncertainty estimation process, which includes:

[0091] During inference, multiple random forward propagations are performed using Monte Carlo Dropout to obtain multiple prediction probability maps;

[0092] The average of the multiple predicted probability maps is calculated as the final prediction;

[0093] The variance of the multiple predicted probability maps is calculated as cognitive uncertainty, which is used to quantify the uncertainty of the model due to insufficient knowledge.

[0094] Specifically, during reasoning, Monte Carlo Dropout is used. T After one random forward propagation, the prediction set is obtained. Cognitive uncertainty is quantified by the variance of predicted values:

[0095] , ;

[0096] in, Represents the sum of all predictions. This indicates cognitive uncertainty. Given the prediction result of the t-th forward pass, the final output of the network is: .

[0097] Furthermore, such as Figure 4 As shown in (a), semi-supervised training using the mean teacher framework at the source center includes:

[0098] Using a student network and a teacher network, wherein the parameters of the teacher network are an exponential moving average of the parameters of the student network;

[0099] For labeled samples, a supervised loss is calculated, which includes a combination of Dice loss and cross-entropy loss;

[0100] For unlabeled samples, two different augmented versions are generated and input into the student network and teacher network respectively. Consistency regularization is applied by minimizing the Kullback-Leibler (KL) divergence of the predictions of the two networks.

[0101] Specifically, there are supervisory losses. For labeled samples ,in The image is a true segmentation map, and Dice loss is used. With cross-entropy loss Combinations:

[0102] .

[0103] Consistent regularization For unlabeled samples Two different enhanced versions were generated. and The student network and the teacher network (whose parameters are exponential moving averages of the student network parameters) each generate predictions. and By minimizing the KL divergence predicted by both. To impose consistency constraints:

[0104] .

[0105] Furthermore, uncertainty-weighted pseudo-label filtering strategies include:

[0106] Generating pseudo-labels from teacher network predictions using unlabeled samples;

[0107] Calculate the overall uncertainty for each pixel, where the overall uncertainty is the sum of accidental uncertainty and cognitive uncertainty;

[0108] Select pixels with uncertainty below a threshold and assign them exponentially decaying weights;

[0109] The pseudo-label loss is calculated based on weighted cross-entropy loss and used to train the student network.

[0110] Specifically, the pseudo-label of uncertainty guidance Teacher network predictions without labeled samples Generating pseudo tags:

[0111] ;

[0112] in This represents the input that makes the function reach its maximum value, and c represents the category index.

[0113] Calculate the overall uncertainty for each pixel. Choose options with uncertainty below a threshold. τ The pixels are assigned exponentially decaying weights. ,in False label loss Weighted cross-entropy loss:

[0114] .

[0115] Overall Loss Function: Combining the above loss functions, the overall training objective is:

[0116] ,

[0117] in, and It is over time t The weight coefficients (increasing gradually with each training iteration) are used to stabilize the initial training phase.

[0118] Furthermore, such as Figure 4 As shown in (b), the online passive adaptive mechanism introduced at the target center includes:

[0119] For each sample at the target center, multiple random augmented versions are generated, including speckle noise, geometric transformation, and intensity perturbation;

[0120] Calculate the average forecast for all enhanced versions;

[0121] Minimize the KL divergence between each enhanced version prediction and the average prediction as a multi-enhancement consistency constraint.

[0122] Specifically, the model trained in the first stage is deployed to the target center. Without accessing the source center data, the unlabeled images of the target center are used for online adaptation to mitigate the impact of domain offset.

[0123] Enhance consistency For each sample at the target center ,generate K A random enhanced version (For example, speckle noise, geometric transformations, intensity perturbations). The model produces predictions for each enhanced image. Calculate the average forecast And minimize the KL divergence between each prediction and the average prediction:

[0124] .

[0125] Furthermore, the pseudo-label selection and confidence weighting that integrates heterogeneity and cognitive uncertainty includes:

[0126] Calculate the overall uncertainty for each pixel, where the overall uncertainty is the sum of accidental uncertainty and cognitive uncertainty;

[0127] Pixel-level weights are calculated based on uncertainty, and these weights are negatively correlated with uncertainty.

[0128] An uncertainty-weighted multi-enhancement consistency loss is used for adaptive model updates.

[0129] Specifically, updates to uncertainty perception:

[0130] Calculate the overall uncertainty for each pixel. .

[0131] Calculate pixel-level weights based on uncertainty. .

[0132] The uncertainty-weighted multi-reinforcement consistency loss is:

[0133] .

[0134] Stability control: Small learning rates, limiting the number of update steps per sample, and using a teacher model (with parameters that are exponential moving averages of the student model parameters) are employed to provide stability guidance and prevent catastrophic forgetting or overfitting of the model.

[0135] Overall adaptive loss: The final adaptive objective function is:

[0136] ,

[0137] in, This is a slight weighted regularization term. and This is the balance coefficient.

[0138] Furthermore, the online passive adaptive mechanism also includes a stability control process, which includes:

[0139] Use a small learning rate for model updates;

[0140] Limit the number of update steps for each sample;

[0141] A teacher model is used to provide stable guidance, the parameters of which are exponential moving averages of the student model parameters, to prevent catastrophic forgetting or overfitting of the model.

[0142] The verification process in this embodiment includes:

[0143] All methods were trained on the source-centric dataset CardiacUDA and evaluated on its internal test set. The proposed method was compared with classic and state-of-the-art semi-supervised segmentation methods, including UA-MT, URPC, MC-Net, CNN_ViT, BCP, SDCL, CCA, and UniMatchV2. This embodiment evaluated the proposed method using four commonly used metrics: Dice similarity coefficient (DSC), Jaccard similarity coefficient (JSC), 95% Hausdorff distance (95HD), and average surface distance (ASD). DSC and JSC quantify the overlap between the predicted and ground truth segmentation results. 95HD measures boundary consistency by calculating the maximum of all minimum distances between two point sets. ASD calculates the average distance from each point on the predicted surface to the corresponding point on the ground truth surface. All experiments were implemented in PyTorch. The model was optimized using stochastic gradient descent (SGD). The experiment was conducted on a system equipped with an Intel Core i7-14700KF CPU, 32GB of RAM, and an NVIDIA GeForce RTX 4090 GPU (24GB of VRAM). Training ran for 30,000 iterations with an initial learning rate of 0.01, which was reduced by 0.1 every 12,000 iterations.

[0144] Based on the experimental results shown in Tables 1-3, the method proposed in this embodiment consistently demonstrates excellent performance on the source-center dataset CardiacUDA under various annotation ratios. Figure 2The table visually compares the segmentation performance of different methods on the source center dataset (CardiacUDA) with different proportions of labeled data (5%, 10%, and 20%). These results demonstrate that the method in this embodiment can effectively utilize limited labeled data and a large amount of unlabeled data, and consistently outperforms existing methods under varying degrees of supervision. Table 1 shows the quantitative comparison results based on the source center dataset CardiacUDA, which was trained using a training set containing 20% ​​labeled data and 80% unlabeled data; bold values ​​indicate best performance. Table 2 shows the quantitative comparison results based on the source center dataset CardiacUDA, which was trained using a training set containing 10% labeled data and 90% unlabeled data; bold values ​​indicate best performance. Table 3 shows the quantitative comparison results based on the source center dataset CardiacUDA, which was trained using a training set containing 5% labeled data and 95% unlabeled data; bold values ​​indicate best performance.

[0145] Table 1

[0146]

[0147] Table 2

[0148]

[0149] Table 3

[0150]

[0151] To evaluate the cross-center generalization ability of our proposed method, experiments were conducted using a private dataset collected from a hospital (as the target center). This external validation aimed to evaluate the performance of the model trained on the publicly available CardiacUDA dataset under domain shifts caused by differences in scanners, protocols, and patient populations. As shown in Tables 4-6, our proposed method consistently outperforms existing semi-supervised segmentation methods across all annotation ratios (20%, 10%, and 5% of the labeled data from the source center). Figure 3The segmentation results of different methods on the target center private dataset are intuitively compared under different proportions of labeled data (5%, 10%, and 20%). This shows that even with limited labeling, appearance-based methods can reliably generalize across centers by effectively utilizing base model features and uncertainty-driven adaptability, demonstrating superior capabilities. Table 4 shows the quantitative comparison results of the models on the target center private dataset. The models are trained using the training set from the source center dataset CardiacUDA, which contains 20% labeled data and 80% unlabeled data. Bold values ​​indicate best performance. Table 5 shows the quantitative comparison results of the models on the target center private dataset. The models are trained using the training set from the source center dataset CardiacUDA, which contains 10% labeled data and 90% unlabeled data. Bold values ​​indicate best performance. Table 6 shows the quantitative comparison results of the models on the target center private dataset. The models are trained using the training set from the source center dataset CardiacUDA, which contains 5% labeled data and 95% unlabeled data. Bold values ​​indicate best performance.

[0152] Table 4

[0153]

[0154] Table 5

[0155]

[0156] Table 6

[0157]

[0158] To quantify the contributions of stochastic and cognitive uncertainty estimations, a comprehensive ablation study was conducted using a semi-supervised learning approach and 20% of the labeled source data. Results for source centers (CardiacUDA) and target centers (private dataset) are summarized in Tables 7 and 8, respectively. It can be seen that the complete model, incorporating both stochastic and cognitive uncertainty, performs best on all evaluation metrics. Removing either stochastic or cognitive uncertainty leads to performance degradation, indicating that both uncertainties have benefits. In particular, as shown in Table 8, in the cross-center scenario (target center), the absence of cognitive uncertainty (No-Epistemic) results in a more significant performance degradation, suggesting that cognitive uncertainty plays a crucial role in handling domain shifts. On the other hand, stochastic uncertainty (No-Aleatoric) also helps improve the model's robustness, especially in handling data noise and image quality variations. The performance degradation is most significant when both uncertainties are removed (no uncertainty), indicating that uncertainty estimation plays a crucial role in semi-supervised learning and cross-center adaptive pseudo-label selection and weight allocation. These results confirm that the combination of stochastic uncertainty and cognitive uncertainty provides complementary information, enhancing the robustness and generalization ability of this method, especially in cross-center scenarios. Table 7 shows the ablation results of the uncertainty components based on the source-center dataset CardiacUDA (20% labeled data), with the best results shown in bold. Table 8 shows the ablation results of the uncertainty components on the target-center private dataset (20% labeled source data), with the best results shown in bold.

[0159] Table 7

[0160]

[0161] Table 8

[0162]

[0163] An application embodiment of the present invention:

[0164] First, data preparation was performed. The CardiacUDA public dataset, containing 150 echocardiogram video sequences, was obtained from the source center. 20% of these were randomly selected as labeled data (left ventricle, right ventricle, and myocardial region labeled frame-by-frame by experienced physicians), and the remaining 80% were left as unlabeled data. The target center data consisted of 80 echocardiogram video sequences from a hospital, without any annotations. All images underwent uniform preprocessing. Resolution is normalized and standardized, and random cropping and rotation are applied. ), speckle noise ( Data augmentation such as intensity scaling (0.8-1.2x) and intensity scaling (0.8-1.2x) are also available.

[0165] The network architecture employs a three-layer structure: the feature extractor uses SAM's ViT-H image encoder (parameter frozen), outputting feature maps at four scales. , , and The lightweight decoder contains four feature fusion modules, each employing... The convolution projects the features onto 256 channels, then upsamples them using bilinear methods. Resolution is fused; the segmentation branch contains three upsampling layers (with 128, 64, and 32 channels respectively) and one... The convolution outputs a 4-channel probability map (3 heart structures + background); the uncertainty branch contains two... Convolutional layers (64 and 32 channels) and one Convolution outputs a single-channel variance plot. Cognitive uncertainty is estimated by applying a dropout rate of 0.2 during inference using 5 Monte Carlo sampling iterations.

[0166] The source-center semi-supervised training employs a mean-teacher framework, with student and teacher models having the same initial parameters. Teacher model parameters are updated using an exponential moving average (EMA decay rate of 0.99) of student model parameters. Training parameters are set as follows: batch size 16 (8 labeled samples + 8 unlabeled samples), initial learning rate 0.01, decay rate of 10 times every 12,000 iterations, totaling 30,000 iterations. The optimizer is SGD (momentum 0.9, weight decay 1e-4). Loss weights... and The pseudo-label screening threshold increases linearly from 0 to 5.0. Set to 0.2, uncertainty weight ,in For pixels i The overall uncertainty.

[0167] The online adaptive target centering phase requires no source data; it generates a target center for each input image. An enhanced version (including random affine transformation, gamma correction, Gaussian noise, and elastic deformation) was implemented. The adaptive learning rate was set to 1e-3, with no more than 3 updates per sample. The teacher model's EMA decay rate was increased to 0.999 to ensure stability. The weight α for the multi-enhancement consistency loss was set to 2.0, and the uncertainty weight β was set to 1.0, comprehensively considering the uncertainty... On the target center's private dataset, an adaptive process is triggered after every 100 new images are received, processing mini-batches of 20 images each time. The validation set performance is continuously monitored, and the adaptive process stops when the DSC metric decreases by more than 0.5% for three consecutive iterations. The system is implemented on an NVIDIA RTX 4090 GPU, with an average inference time of 35ms per image, plus an additional 15ms for uncertainty estimation.

[0168] This invention discloses a cross-center, reliable, semi-supervised cardiac ultrasound image segmentation method. By integrating a base model and uncertainty estimation, and combining a two-stage learning strategy, it effectively reduces the dependence of cardiac ultrasound image segmentation on large amounts of labeled data. Semi-supervised learning fully utilizes unlabeled data, significantly reducing annotation costs. This invention improves the model's generalization ability in cross-center scenarios. Utilizing the powerful feature representation capabilities of the base model and an online passive adaptive mechanism, it effectively alleviates the domain shift problem caused by differences in equipment, protocols, and patient groups. Simultaneously, by simultaneously estimating accidental and cognitive uncertainties, this invention provides pixel-level confidence information for the segmentation results, enhancing the credibility of the results and assisting clinical decision-making. Furthermore, this uncertainty is used to guide pseudo-label screening and loss weighting during training and adaptation, improving the model's robustness. Moreover, the target center adaptation process does not require access to source center data, complying with clinical data privacy and security requirements.

[0169] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A cross-center, reliable, semi-supervised cardiac ultrasound image segmentation method, characterized in that, include: Multiscale features of input cardiac ultrasound images are extracted using an image encoder based on a segmentation model. A lightweight decoder is designed to fuse and upsample the multi-scale features, and output a segmentation probability map and a random uncertainty map. At the source center, a mean teacher framework is used for semi-supervised training, and a loss function is constructed by combining supervised loss, consistency regularization, and uncertainty-weighted pseudo-label filtering strategies. At the target center, an online passive adaptive mechanism is introduced. Through multiple reinforcement consistency constraints, combined with pseudo-label selection and confidence weighting based on the fusion of heterogeneity and cognitive uncertainty, robust cross-center segmentation is achieved. Uncertainty-weighted pseudo-label filtering strategies include: Generating pseudo-labels from teacher network predictions using unlabeled samples; Calculate the overall uncertainty for each pixel, where the overall uncertainty is the sum of accidental uncertainty and cognitive uncertainty; Select pixels with uncertainty below a threshold and assign them exponentially decaying weights; The pseudo-label loss is calculated based on weighted cross-entropy loss and used to train the student network. Introducing an online passive adaptive mechanism at the target center includes: For each sample at the target center, multiple random augmented versions are generated, including speckle noise, geometric transformation, and intensity perturbation; Calculate the average forecast for all enhanced versions; Minimize the KL divergence between each enhanced version prediction and the average prediction as a multi-enhancement consistency constraint; Pseudo-label selection and confidence weighting that integrates heterogeneity and cognitive uncertainty include: Calculate the overall uncertainty for each pixel, where the overall uncertainty is the sum of accidental uncertainty and cognitive uncertainty; Pixel-level weights are calculated based on uncertainty, and these weights are negatively correlated with uncertainty. The aforementioned weights are used to apply uncertainty-weighted summaries to the multi-enhancement consistency loss for adaptive model updates. The online passive adaptive mechanism also includes a stability control process, which includes: Use a small learning rate for model updates; Limit the number of update steps for each sample; A teacher model is used to provide stable guidance, the parameters of which are exponential moving averages of the student model parameters, to prevent catastrophic forgetting or overfitting of the model.

2. The cross-center reliable semi-supervised cardiac ultrasound image segmentation method as described in claim 1, characterized in that, The image encoder, utilizing a segmentation-based model, extracts multi-scale features from the input echocardiogram image, including: The image encoder of the Segment Anything Model, pre-trained on a large-scale general dataset, is used as the feature extractor; The image encoder is based on a visual Transformer architecture; The image encoder extracts multi-scale feature maps from the input image, including feature maps from multiple layers; The parameters of the image encoder are kept frozen during training to preserve general visual knowledge.

3. The cross-center reliable semi-supervised cardiac ultrasound image segmentation method as described in claim 1, characterized in that, The design of a lightweight decoder for multi-scale feature fusion and upsampling includes: Design a lightweight UNet-like decoder; Each encoder feature map is convolved, projected, and upsampled to a common spatial resolution; By fusing all the upsampled feature maps, a decoded feature map is obtained; Fine-grained spatial details are gradually recovered through a series of upsampling and skip connections to output a segmentation probability map and a random uncertainty map.

4. The cross-center reliable semi-supervised cardiac ultrasound image segmentation method as described in claim 1, characterized in that, The output segmentation probability graph and random uncertainty graph include: The segmentation probability map is output through the segmentation branch, which includes a convolutional layer and a softmax activation function, to obtain the class probability distribution of each pixel; The uncertainty branch outputs a random uncertainty map, which includes convolutional layers and ReLU activation functions, and predicts a pixel-level variance map to measure the inherent noise of the data.

5. The cross-center reliable semi-supervised cardiac ultrasound image segmentation method as described in claim 1, characterized in that, The method further includes a cognitive uncertainty estimation process, which includes: During inference, multiple random forward propagations are performed using Monte Carlo Dropout to obtain multiple prediction probability maps; The average of the multiple predicted probability maps is calculated as the final prediction; The variance of the multiple predicted probability maps is calculated as cognitive uncertainty, which is used to quantify the uncertainty of the model due to insufficient knowledge.

6. The cross-center reliable semi-supervised cardiac ultrasound image segmentation method as described in claim 1, characterized in that, Semi-supervised training using the mean-teacher framework at the source center includes: Using a student network and a teacher network, wherein the parameters of the teacher network are an exponential moving average of the parameters of the student network; For labeled samples, a supervised loss is calculated, which includes a combination of Dice loss and cross-entropy loss; For unlabeled samples, two different augmented versions are generated and input into the student network and teacher network respectively. Consistency regularization is applied by minimizing the KL divergence of the two predictions.

Images