A semi-supervised medical image segmentation method based on causal uncertainty decomposition

Abstract

This invention discloses a semi-supervised medical image segmentation method based on causal uncertainty decomposition, belonging to the fields of medical image processing and artificial intelligence. The method constructs a segmentation model with a teacher-student architecture, integrating a causal uncertainty decomposition module, an adaptive consistency learning module, and a topologically aware consistency loss module. By decomposing total uncertainty into cognitive uncertainty and stochastic uncertainty, targeted processing is achieved; a dual-path weight fusion and differential modulation strategy are employed to achieve pixel-level adaptive learning; and the Betti number is introduced to calculate topological distance, preserving the integrity of anatomical structures. This invention improves the Dice coefficient by 3.2% to 4.8% on multiple medical image datasets using only 5% to 20% of the labeled data, significantly improving segmentation accuracy and boundary localization accuracy. It effectively solves the segmentation problem under conditions of scarce medical image annotation and has significant clinical application value.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and artificial intelligence, specifically relating to a semi-supervised medical image segmentation method based on causal uncertainty decomposition. Background Technology

[0002] In recent years, deep learning has made significant progress in the field of medical image processing, especially in medical image segmentation, where it has demonstrated tremendous potential. Medical image segmentation plays a crucial role in clinical diagnosis, treatment planning, and lesion region analysis. Deep learning methods, such as convolutional neural networks (CNNs), have achieved breakthrough results in several medical image tasks. However, deep learning models typically require large amounts of labeled data for training, while the labeling of medical image data often faces challenges such as high costs, strict privacy protection, and data scarcity, which limits the widespread application of deep learning in the field of medical imaging.

[0003] To overcome these challenges, semi-supervised learning methods have emerged, capable of learning on both limited labeled data and large amounts of unlabeled data. Current semi-supervised medical image segmentation methods largely rely on consistency regularization, aiming to improve model performance by maintaining prediction consistency under perturbations of the input image. However, existing methods suffer from several problems: First, they ignore the diversity of uncertainty, often treating uncertainty uniformly and failing to distinguish between cognitive uncertainty caused by the data itself (such as image noise, blurred boundaries, etc.) and random uncertainty caused by model parameters (such as insufficient training); second, traditional pixel-level consistency loss cannot effectively consider the topological characteristics of medical structures, leading to problems when segmenting complex structures in medical images.

[0004] To address these issues, recent methods have attempted to enhance the performance of medical image segmentation models by introducing causal inference and topological constraints. In particular, causal uncertainty decomposition, by decomposing uncertainties from different sources, can provide more accurate modeling and processing strategies, especially when dealing with medical images containing regions of high uncertainty. Topology-aware loss, on the other hand, ensures the structural integrity of the segmentation results by maintaining the anatomical connectivity of the medical image, which is crucial for handling complex anatomical structures and organ shapes.

[0005] However, existing semi-supervised learning methods still have some limitations. For example, many methods fail to effectively combine causal reasoning and topological awareness, and lack flexibility in handling different types of uncertainty. Therefore, there is an urgent need for an innovative method that can more effectively handle uncertainty in medical images, achieve high-precision medical image segmentation using limited labeled data, and maintain the anatomical consistency of medical images. Summary of the Invention

[0006] The technical problem to be solved by this invention is: under the condition of scarce medical image annotation, how to effectively distinguish and handle uncertainties from different sources (including cognitive uncertainties caused by data noise and random uncertainties caused by unstable model parameters), and to perform differentiated processing based on uncertainties in the consistency learning of unlabeled data, while maintaining the integrity of the anatomical topology of the segmentation results, so as to improve the accuracy and robustness of medical image segmentation.

[0007] To address the above problems, this invention provides a semi-supervised medical image segmentation method based on causal uncertainty decomposition, which effectively solves the uncertainty modeling and topology preservation problems in medical image segmentation. The key technical points are three main modules: First, a causal uncertainty decomposition module: based on a causal structure model, the total uncertainty in segmentation prediction is decomposed into cognitive uncertainty and stochastic uncertainty, achieving explicit modeling of different uncertainty sources; second, an adaptive consistency learning module: based on the uncertainty obtained from decomposition, adaptive weights and modulation factors are generated, and differentiated consistency constraints are applied to high-uncertainty regions of unlabeled images to improve the model's learning ability for difficult regions; third, a topology-aware consistency loss module: based on the Betti number, the topological distance predicted by the student model and the teacher model is calculated to maintain the anatomical connectivity and hole structure consistency of the segmentation results.

[0008] In a first aspect, the present invention provides a semi-supervised medical image segmentation method based on causal uncertainty decomposition, comprising the following steps:

[0009] Step 1: Obtain a medical image dataset and preprocess it. The dataset includes labeled data and unlabeled data.

[0010] Step 2: Construct a segmentation model for the teacher and student architectures. The segmentation model includes a teacher network and a student network with the same structure. Both the student network and the teacher network use a 3D U-Net as the backbone network and integrate a causal uncertainty decomposition module, an adaptive consistency learning module, and a topology-aware consistency loss module.

[0011] Step 3: Train the segmentation model using the labeled and unlabeled data, performing the following operations in each training iteration:

[0012] Step 3.1: Input the input image into the student network and the teacher network respectively for forward propagation to obtain the prediction results and feature maps of the student network, and the prediction results of the teacher network;

[0013] Step 3.2: By directly inputting the feature map output by the student network into the causal uncertainty decomposition module, the total uncertainty in medical image segmentation is decomposed into cognitive uncertainty and random uncertainty, wherein the cognitive uncertainty represents the noise inherent in the data, and the random uncertainty represents the uncertainty of the model parameters;

[0014] Step 3.3: Based on the cognitive uncertainty and stochastic uncertainty, generate adaptive weights and modulation factors through the adaptive consistency learning module;

[0015] Step 3.4: Calculate the topological distance between the student network prediction and the teacher network prediction using the topological awareness consistency loss module, and calculate the topological consistency loss based on the topological distance;

[0016] Step 3.5: Calculate the total loss function, which includes the supervision loss based on labeled data, the causal decomposition loss based on unlabeled data, the adaptive consistency loss, and the topological consistency loss;

[0017] Step 3.6: Update student network parameters via backpropagation;

[0018] Step 3.7: Update the parameters of the teacher network based on the exponential moving average of the student network parameters;

[0019] Step 4: Apply the trained segmentation model to segment the target medical image.

[0020] Optionally, in step 2, the causal uncertainty decomposition module, based on Pearl's hierarchy of causality theory, decomposes the total uncertainty into cognitive uncertainty and stochastic uncertainty by constructing a causal graph, and optimizes it through a decomposition loss function; the decomposition loss function The calculation method is as follows:

[0021]

[0022] in, for Reconstruction loss; Total uncertainty; It is a random uncertainty; To understand uncertainty; and These are adaptive learning parameters; These are the weighting coefficients; The regularization term is calculated as follows:

[0023]

[0024] in, The prior is a uniform distribution; Let KL divergence be denoted as KL divergence.

[0025] Optionally, in step 2, the adaptive consistency learning module generates an adaptive weight graph through a dual-path weight fusion network, and the calculation method is as follows:

[0026]

[0027] in, Student network feature graph; Capturing self-attention structures Spatial dependence; This is a feature-level attention function based on a self-attention mechanism, used to capture... Spatial dependencies; and These are respectively a random uncertainty weight generator and a cognitive uncertainty weight generator; This indicates element-wise multiplication; Dual-path adaptive weights;

[0028] Furthermore, a differentiated modulation factor is applied to enhance consistency constraints in regions with high cognitive uncertainty and to reduce consistency constraints in regions with high random uncertainty. The calculation method is as follows:

[0029]

[0030]

[0031] in, It is a random modulation factor; Cognitive modulation factor; and These represent the spatial standard deviations of random uncertainty and cognitive uncertainty, respectively. and These are learnable parameters;

[0032] The adaptive consistency loss is calculated by combining modulation factor, adaptive weights, and progressive scheduling to optimize unlabeled data; the adaptive consistency loss... The calculation method is as follows:

[0033]

[0034] in, Loss due to basic consistency; This indicates the unlabeled dataset. Take the expected value.

[0035] Optionally, in step 2, the topology-aware consistency loss module calculates Betti quantified topological features and defines the topological distance between the student model prediction and the teacher model prediction. The calculation formula is as follows:

[0036]

[0037] in, Indicates the number of connected components. Indicates the number of one-dimensional holes. Weighting for penalties related to holes. The segmentation results of the student model on the input medical image. The segmentation result of the teacher model on the input medical image;

[0038] The topological distance is combined with pixel-level loss to obtain topological awareness consistency loss. The calculation method is as follows:

[0039]

[0040] in, The topological loss weights for the current iteration. For maximum weight, This represents the current iteration number. For pixel-level consistency loss, The initial iteration is for topological loss. This is the weighting growth cycle.

[0041] Optionally, in step 3.5, the total loss function The calculation method is as follows:

[0042]

[0043] in, For the supervision loss function, The loss function is the causal decomposition function. For adaptive consistency loss function, Let the topology consistency loss function be... , , These are the corresponding weighting coefficients.

[0044] Optionally, in step 3.7, the teacher network parameters are updated using an exponential moving average method:

[0045]

[0046] in, For teacher network parameters, For student network parameters, This is the momentum factor, with a value ranging from 0.9 to 0.999.

[0047] Secondly, the present invention provides a semi-supervised medical image segmentation system based on causal uncertainty decomposition, comprising:

[0048] The data preprocessing module is used to acquire and preprocess medical image datasets, which include labeled and unlabeled data.

[0049] The segmentation model construction module is used to construct segmentation models of teacher and student architectures. The segmentation model includes teacher networks and student networks with the same structure. Both the student network and the teacher network use a 3D U-Net as the backbone network and integrate a causal uncertainty decomposition module, an adaptive consistency learning module, and a topology-aware consistency loss module.

[0050] A training module is used to train the segmentation model using the labeled and unlabeled data. This training module is configured to perform the following operations in each iteration: input the input image into a student network and a teacher network for forward propagation, respectively, to obtain the prediction results and feature maps of the student network and the prediction results of the teacher network; decompose the total uncertainty in medical image segmentation into cognitive uncertainty and random uncertainty using the causal uncertainty decomposition module; generate adaptive weights and modulation factors based on the cognitive uncertainty and random uncertainty using the adaptive consistency learning module; calculate the topological distance between the student network predictions and the teacher network predictions using the topological awareness consistency loss module, and calculate the topological consistency loss based on the topological distance; calculate the total loss function, which includes supervised loss based on labeled data, causal decomposition loss based on unlabeled data, adaptive consistency loss, and topological consistency loss; update the student network parameters through backpropagation; and update the teacher network parameters according to the exponential moving average of the student network parameters.

[0051] The segmentation execution module is used to segment the target medical image using a trained segmentation model.

[0052] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the semi-supervised medical image segmentation method based on causal uncertainty decomposition as described above.

[0053] Fourthly, the present invention provides an electronic device, comprising:

[0054] Memory, used to store computer programs;

[0055] A processor is configured to implement the semi-supervised medical image segmentation method based on causal uncertainty decomposition as described above when executing the computer program.

[0056] The present invention has the following beneficial effects:

[0057] The method of this invention significantly improves the accuracy of uncertainty estimation through a causal uncertainty decomposition module. This invention decomposes the total uncertainty in medical image segmentation into cognitive uncertainty and stochastic uncertainty, enabling the model to identify the source of uncertainty and address it accordingly. On the LA dataset, the expected calibration error (ECE) decreased from 0.12 in the traditional method to 0.05, indicating a high degree of consistency between the model's confidence and actual accuracy. Calibration curve analysis shows that the curve of this invention closely approximates the ideal diagonal, while the comparative methods generally suffer from overconfidence or overconservatism. More importantly, by visualizing the uncertainty map, it can be clearly seen that this invention can accurately identify areas that are difficult to segment, such as organ boundaries and lesion edges. The high uncertainty in these areas suggests that doctors need to pay extra attention, thereby improving the clinical interpretability and credibility of the model.

[0058] The method of this invention achieves differentiated learning for different uncertainty regions through an adaptive consistency learning module, effectively suppressing the interference of noise pseudo-labels in high-risk areas such as pancreatic boundaries and brain tumor edges. Furthermore, this invention generates adaptive weights based on cognitive and random uncertainties, avoiding the problem of noise pseudo-labels misleading all unlabeled pixels due to the indiscriminate weighting of existing methods. Ablation experiments show that adaptive weighting further improves the Dice coefficient by 2.15% to 4.55% on the BraTS2019 and Pancreas datasets, and reduces the 95% Hausdorff distance (HD95) by 1.8 mm, demonstrating its advantage in boundary localization accuracy.

[0059] The method of this invention solves the problem of false disconnections or false connections caused by traditional pixel-level loss through a topology-aware consistency loss module, effectively maintaining the topological correctness of structures in tasks such as left atrium of the heart and brain tumors. Furthermore, the topological distance constraint based on the Betti number in this invention enables the model to maintain the connectivity and internal hole structure of the target region, which also avoids the false disconnections and false connections caused by traditional pixel-level consistency loss. Experimental data show that topology-aware loss significantly improves the HD95 index, reducing it to 12.33 mm on the BraTS2019 dataset, while increasing the Dice coefficient by 1.49%, verifying its necessity in the segmentation of complex anatomical structures.

[0060] The method of this invention combines supervised loss, causal decomposition loss, adaptive consistency loss, and topological consistency loss to form a multi-task optimization objective, enabling the model to learn collaboratively on labeled and unlabeled data. The total loss function balances the contributions of each component through weight coefficients, avoiding overfitting or underfitting of a single loss term. Ablation experiments confirm that the complete loss function improves the Dice coefficient by 4.94% and reduces HD95 by 10.09 mm compared to the baseline (supervised loss only), demonstrating the effectiveness of multi-loss synergy.

[0061] The method of this invention employs exponential moving average (EMA) to update teacher network parameters, ensuring that the teacher model provides stable and reliable pseudo-labels. This update strategy smooths parameter changes, reduces the generation of noisy pseudo-labels, and improves the stability of semi-supervised learning. Experimental results show that the EMA strategy enables the model to maintain high performance even with a low proportion of labeled data (5%) (91.57% on the LA dataset Dice), and the training process converges faster.

[0062] In summary, the method of this invention achieves high-precision and robust medical image segmentation with a small amount of labeled data through the synergistic effect of causal uncertainty decomposition, adaptive consistency learning, and topology-aware constraints. It significantly reduces labeling costs, improves the performance of the model in segmenting complex anatomical structures, and provides a theoretically rigorous and practical solution for semi-supervised medical image segmentation. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is the overall model framework diagram provided by the present invention.

[0065] Figure 2 This is a flowchart of the causal uncertainty decomposition module provided by the present invention.

[0066] Figure 3 This is a flowchart of the topology-aware consistency loss module processing provided by the present invention.

[0067] Figure 4 This is a visual comparison image of the pancreas dataset provided by the present invention.

[0068] Figure 5 This is a performance comparison chart of the ISLES-2022 dataset provided by this invention.

[0069] Figure 6 This is a performance comparison chart of the BraTS2019 dataset provided by this invention.

[0070] Figure 7 This is a graph showing the performance comparison results of the LA dataset provided by this invention.

[0071] Figure 8 This is a graph showing the performance comparison results of the Pancreas-NIH dataset provided by this invention.

[0072] Figure 9 This is a diagram showing the ablation experiment results for different causal decomposition configurations provided by the present invention.

[0073] Figure 10 This is an ablation experiment result diagram of the adaptive consistency learning module provided by the present invention.

[0074] Figure 11 This is a graph showing the ablation experiment results on the BraTS2019 dataset with 10% of the labeled data provided by this invention. Detailed Implementation

[0075] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] Example 1

[0077] This embodiment provides a semi-supervised medical image segmentation method based on causal uncertainty decomposition (hereinafter referred to as the CaUCON method), which includes the following steps:

[0078] Step 1: Obtain and preprocess the medical image dataset, which includes labeled and unlabeled data;

[0079] In this embodiment, the experimental verification covers four representative medical image segmentation datasets. These datasets are diverse in terms of imaging modalities, anatomical structures, and segmentation difficulty, which can comprehensively evaluate the effectiveness and generalization ability of the proposed method.

[0080] The first dataset is the LA (Left Atrium Dataset), which contains 100 gadolinium-enhanced 3D MRI images of the left atrium. These images have a spatial resolution of 0.625 × 0.625 × 0.625 cubic millimeters, exhibiting high spatial accuracy. The task of this dataset is to accurately segment the structure of the left atrium, which is crucial for the diagnosis and treatment planning of cardiac diseases such as atrial fibrillation. The main challenges in left atrial segmentation lie in its thin-walled structure, complex morphological variations, and the delineation of its boundaries with surrounding cardiac tissue.

[0081] The second dataset is the Pancreas-NIH dataset, which collects 82 three-dimensional abdominal CT images. Pancreatic segmentation is recognized as one of the most challenging tasks in medical image segmentation. The main challenge stems from the often blurred boundaries between the pancreas and neighboring organs such as the stomach, duodenum, and spleen, as well as the significant variations in the shape, size, and location of the pancreas itself among individuals. Furthermore, the low contrast between pancreatic tissue and surrounding soft tissue in CT images further complicates the segmentation process.

[0082] The third dataset is the ISLES-2022 dataset, which contains 250 multimodal MRI scan images (including DWI, ADC, and FLAIR modal data) for ischemic stroke lesion segmentation. The task of this dataset is to perform accurate segmentation of stroke lesion regions, which is crucial for acute diagnosis and treatment assessment of stroke. The main challenge of this dataset lies in the significant diversity of lesion size, location, and sparse distribution.

[0083] The fourth dataset is the BraTS2019 dataset (Brain Tumor Segmentation Challenge 2019), which contains 335 multimodal brain MRI scans. Although the original dataset includes four modalities—T1, T1-enhanced, T2, and T2-FLAIR—this experiment primarily used T2-FLAIR sequences for whole-brain tumor segmentation. This dataset is divided into a training set of 250 cases, a validation set of 25 cases, and a test set of 60 cases. The difficulty in brain tumor segmentation lies in the high heterogeneity of the tumor region, its irregular shape, and the blurred boundaries with normal brain tissue.

[0084] To comprehensively evaluate the performance of semi-supervised learning methods under different data scarcity levels, this invention employed three different proportions of labeled data in experiments. The first scenario involved extremely few samples, using only 5% of the data as labeled data. This simulates a situation of extreme scarcity of labeled resources, placing extremely high demands on semi-supervised learning methods. The second scenario involved low samples, using 10% of the labeled data. This is a common situation in clinical practice, where only a small portion of cases received expert annotations. The third scenario involved moderate samples, using 20% ​​of the labeled data. This represents a situation where labeled resources are relatively abundant but still insufficient to support fully supervised learning. In all scenarios, the remaining unlabeled data participated in the training process, and semi-supervised learning techniques such as pseudo-labels and consistency regularization were used to improve model performance.

[0085] Data preprocessing is a crucial step in ensuring effective model training. For intensity normalization, this invention employs differentiated processing strategies for different imaging modalities. For CT images, the window level and width are first adjusted according to the characteristics of abdominal scans, limiting grayscale values ​​to a range relevant to soft tissues and organs. Then, Z-score normalization is performed to ensure comparability of intensity distributions across different scans. For MRI images, since intensity values ​​lack absolute physical meaning and intensity distributions vary significantly between different sequences, the original relative intensity distribution characteristics are preserved, and Z-score normalization is also used to standardize the intensity range of each scan. Regarding spatial normalization, because images generated by different scanning devices and protocols have different resolutions and field-of-view sizes, this invention first resamples all images to a uniform isotropic resolution, ensuring a consistent spatial scale for the model when processing data from different sources. Subsequently, images are cropped or padded to fixed-size 3D patches, which satisfies the network input size requirements while allowing for the processing of complete images of arbitrary sizes via a sliding window strategy.

[0086] Data augmentation strategies are applied only to support set samples (i.e., labeled samples) to increase the model's generalization ability and prevent overfitting. Specific augmentation operations include random rotation, with an angle range set to ±15 degrees, simulating slight positional changes in the patient during scanning. Elastic deformation augmentation simulates anatomical differences between individuals, increasing the diversity of training samples through smooth local deformation. Intensity perturbation simulates the effects of different scanning parameters by randomly adjusting image brightness and contrast. All augmentation operations are applied randomly with a probability of 0.5 to avoid unrealistic transformations caused by over-augmentation.

[0087] Step 2: Construct a partitioning model for the teacher and student architectures. The partitioning model includes structurally identical teacher and student networks. Both networks use a 3D U-Net as their backbone and integrate a causal uncertainty decomposition module (CUDM), an adaptive consistency learning module (ACLM), and a topology-aware consistency loss module (TACL). See [link to relevant documentation]. Figure 1 ;

[0088] The backbone network used in this invention is the classic 3D U-Net architecture, which is particularly suitable for medical image segmentation tasks because it can simultaneously capture multi-scale contextual information and fine spatial details through its encoder and decoder structure. The encoder consists of four consecutive downsampling layers. Each downsampling layer extracts features through convolution and reduces spatial resolution through pooling, thereby progressively expanding the receptive field to capture a wider range of contextual information. The decoder also consists of four upsampling layers, progressively restoring spatial resolution through transposed convolution or upsampling operations. Skip connections are used to concatenate the feature maps of corresponding encoder layers with the feature maps of the decoder. This design effectively preserves detailed information from the encoder and compensates for spatial details that may be lost during upsampling, which is especially important for segmentation tasks in medical images where high boundary sharpness is required.

[0089] The causal uncertainty decomposition module is one of the core innovations of this invention. Its design, based on Pearl's hierarchy of causality theory, aims to decompose the total uncertainty of the model output into two independent components: random uncertainty (originating from model parameters) and cognitive uncertainty (originating from inherent noise in the data). To achieve this goal, this module constructs a lightweight convolutional network with the number of parameters controlled to within the thousands. This design ensures sufficient expressive power without imposing excessive computational burden on the overall model.

[0090] This decomposer takes the feature maps from the segmentation network as input and is built upon the CaUCON backbone network based on 3D UNet. The backbone network contains two lightweight convolutional branches, one for estimating stochastic uncertainty and the other for estimating cognitive uncertainty. These two branches follow the design principle of "lightweight convolution and adaptive fusion," each outputting a single-channel uncertainty map, denoted as follows: (Cognitive uncertainty) and (Random uncertainty).

[0091] Through this explicit decomposition mechanism, the model can implement differentiated learning strategies for different types of uncertainty, thereby making more efficient use of unlabeled data and improving the robustness and generalization ability of the overall system.

[0092] Causal decomposition loss function for:

[0093]

[0094] in, for Reconstruction loss; Total uncertainty; It is a random uncertainty; To understand uncertainty; and These are adaptive learning parameters; These are the weighting coefficients; The regularization term is calculated as follows:

[0095]

[0096] in, Assuming a uniform prior, Let KL divergence be denoted as KL divergence.

[0097] The adaptive consistency learning module generates an adaptive weight graph through a dual-path weight fusion network, and the calculation method is as follows:

[0098]

[0099] in, Student network feature graph; For self-attention structures used to capture Spatial dependence; This is a feature-level attention function based on a self-attention mechanism, used to capture... Spatial dependencies; and These are a cognitive uncertainty weight generator and a stochastic uncertainty weight generator, both of which are lightweight convolutional networks (with an integrated spatial attention module) used to convert the corresponding uncertainty graph into a spatial weight graph. This indicates element-wise multiplication; Dual-path adaptive weights;

[0100] Each weight generator integrates a spatial attention module, enabling it to learn the importance of different spatial locations and output a weight map of the same spatial size as the input feature map. This pixel-level adaptive weighting mechanism allows the model to finely adjust the learning intensity spatially, strengthening the learning signal in high-confidence regions and reducing the impact of noise in low-confidence regions. The parameters of the weight network are jointly optimized with the main network through end-to-end backpropagation, thereby automatically learning the optimal weighting strategy without the need for manually setting fixed thresholds or weight decay strategies.

[0101] Then, a differential modulation factor is applied to enhance consistency constraints in regions with high cognitive uncertainty and to reduce consistency constraints in regions with high random uncertainty. The calculation method is as follows:

[0102]

[0103]

[0104] in, It is a random modulation factor; Cognitive modulation factor; It is the standard deviation of random uncertainty in spatial dimensions, corresponding to each location. The value reflects the intensity of spatial fluctuations in random uncertainty in the region; It is the standard deviation of cognitive uncertainty in the spatial dimension, which is essentially a spatial quantification of the degree of cognitive inadequacy of the model in different regions. and In the adaptive consistency learning mechanism, the learnable parameter is used to adjust the strength of the influence of uncertainty, where adjust Modulation factor The intensity of the impact, Then adjust Modulation factor The intensity of the impact.

[0105] The adaptive consistency loss is calculated by combining modulation factor, adaptive weights, and asymptotic scheduling to optimize unlabeled data; adaptive consistency loss The calculation method is as follows:

[0106]

[0107] in, Basic consistency loss is a variable loss function, including but not limited to the MSE loss function and the L2 loss function; This indicates the unlabeled dataset. Take the expected value.

[0108] The topology-aware consistency loss module calculates Betti quantified topological features and defines the topological distance between student model predictions and teacher model predictions. The calculation formula is as follows:

[0109]

[0110] in, The number of connected components is represented by the 26-connectivity labeling algorithm. Indicates the number of one-dimensional holes. Weighting of penalties for holes; The segmentation results of the student model on the input medical image. The segmentation result of the teacher model on the input medical image;

[0111] Topological distance is combined with pixel-level loss to obtain topological awareness consistency loss, the calculation method of which is as follows:

[0112]

[0113] in, The topological loss weights for the current iteration. For maximum weight, This represents the current iteration number. The pixel-level consistency loss, specifically in this embodiment, is the mean squared error (MSE) loss between the prediction results of the student network and the teacher network. The initial iteration is for topological loss. As the weight growth cycle occurs, these hyperparameters need to be adjusted based on the total number of training iterations and the characteristics of the dataset, and the optimal values ​​vary depending on the dataset.

[0114] Step 3: Train the segmentation model using labeled and unlabeled data, performing the following operations in each training iteration:

[0115] Step 3.1: Input the input image into the student network and the teacher network respectively for forward propagation to obtain the prediction results and feature maps of the student network, and the prediction results of the teacher network;

[0116] Step 3.2: See Figure 2 By directly inputting the feature map output by the student network into the causal uncertainty decomposition module, the total uncertainty in medical image segmentation is decomposed into cognitive uncertainty and random uncertainty, where cognitive uncertainty represents the inherent noise of the data and random uncertainty represents the uncertainty of the model parameters.

[0117] Step 3.3: Based on cognitive uncertainty and stochastic uncertainty, adaptive weights and modulation factors are generated through the adaptive consistency learning module;

[0118] Step 3.4: See Figure 3 The topological distance between student network predictions and teacher network predictions is calculated using the topological awareness consistency loss module, and the topological consistency loss is calculated based on the topological distance.

[0119] Step 3.5: Calculate the total loss function, which includes the supervision loss based on labeled data, the causal decomposition loss based on unlabeled data, the adaptive consistency loss, and the topological consistency loss;

[0120] Step 3.6: Update student network parameters via backpropagation;

[0121] Step 3.7: Update the parameters of the teacher network based on the exponential moving average of the student network parameters;

[0122] The total loss function is calculated as follows:

[0123]

[0124] in, To monitor losses, For causal decomposition loss, For adaptive consistency loss, For topology consistency loss, , , These are the corresponding weighting coefficients. , , The fixed value is selected through experimentation and hyperparameter search. Preferably, in this embodiment... The value is 0.1. The value is 1.0. The value is 0.05.

[0125] The method for calculating monitoring losses is as follows: = ,in, Let cross-entropy be the loss function. This is a loss for Dice.

[0126] The teacher network parameters are updated using an exponential moving average method:

[0127]

[0128] in, For teacher network parameters, For student network parameters, The momentum factor ranges from 0.9 to 0.999. This design makes the updates of the teacher model smoother and more stable, providing more reliable supervision signals for the student model.

[0129] The training parameters in this embodiment are set as follows: SDG optimizer is used, initial learning rate is 0.01, momentum is 0.9, batch size is 4 (limited by GPU memory), and number of iterations is 6000. To make full use of the limited labeled data, an online hard example mining strategy is adopted, that is, in each epoch, samples with large model prediction errors are sampled first. This strategy can accelerate model convergence and improve the segmentation performance of hard samples. At the same time, in order to prevent overfitting, in addition to data augmentation, a Dropout layer (dropout rate 0.1) and weight decay regularization (coefficient 0.0001) are added to the network. The training process selects the optimal model by monitoring the Dice coefficient on the validation set and adopts an early stopping strategy. If the validation performance does not improve within 20 consecutive epochs, the training is terminated early.

[0130] To comprehensively evaluate the performance of the proposed method, this invention establishes a multi-dimensional evaluation system covering aspects such as segmentation accuracy, uncertainty estimation quality, and computational efficiency. For segmentation accuracy evaluation, three of the most commonly used metrics in the field of medical image segmentation are employed. The first is the Dice coefficient (Dice Similarity Coefficient, DSC), which measures the degree of overlap between the predicted and actual segments. The calculation formula is twice the intersection divided by the sum of the two segments, with values ​​ranging from 0 to 1; a higher value indicates more accurate segmentation. The Dice coefficient is sensitive to the overall quality of segmentation and is considered the gold standard for evaluating segmentation performance.

[0131] Secondly, there's the Jaccard index (Intersection over Union, IoU), also known as the intersection-over-union ratio, which calculates the intersection of the predicted and actual segments divided by their union. Compared to the Dice coefficient, IoU is more stringent regarding segmentation errors because it only calculates the overlap region once in the denominator, resulting in a heavier penalty for missegmentation. The third metric is the 95% Hausdorff Distance (HD95), which measures the maximum distance between the predicted and actual boundaries, but excludes 5% of the maximum distance values ​​to avoid the influence of outliers. HD95 reflects the accuracy of the segmentation boundary and is particularly important for clinical applications requiring precise boundary localization (such as radiotherapy planning). This metric is measured in millimeters; a smaller value indicates more accurate boundary localization.

[0132] In assessing the quality of uncertainty estimation, two metrics were used: calibration curve and Expected Calibration Error (ECE). The calibration curve divides the predicted probabilities into several intervals, calculates the average predicted probability and the actual accuracy for each interval, and then plots the relationship between the two. Ideally, if the model's uncertainty estimation is accurate, the predicted probabilities should match the actual accuracy, and the calibration curve should be diagonal. ECE quantifies the deviation between the calibration curve and the ideal diagonal, calculating a weighted average of the differences between the predicted probabilities and the actual accuracy across all probability intervals. A smaller ECE indicates more accurate uncertainty estimation.

[0133] Furthermore, the quality of the uncertainty decomposition was evaluated. Due to the lack of true labels for random and cognitive uncertainties, an indirect evaluation method was employed. For cognitive uncertainty, the model's cognitive uncertainty was observed to change with different amounts of training data by gradually increasing the amount of training data. Theoretically, cognitive uncertainty should decrease as the model acquires more knowledge. If this trend is observed, the estimation of cognitive uncertainty is reasonable. For random uncertainty, its correlation with local image features was analyzed, with higher random uncertainty expected in regions with blurred boundaries and high noise levels.

[0134] For computational efficiency evaluation, the number of model parameters, floating-point operations (FLOPs), training time, and inference time were measured. The number of parameters reflects the model's storage overhead, and FLOPs reflect computational complexity; both metrics are particularly important for practical deployment. Training time assesses the time required for the model to converge under the same hardware conditions, while inference time measures the segmentation time of a single image, which directly affects the feasibility of clinical applications. For fair comparison, all methods were tested under the same hardware environment (NVIDIA Tesla V100 GPU) and software environment (PyTorch 1.10).

[0135] Figure 4 The visual comparison results of the pancreatic dataset are presented, from left to right: (a) original CT image, (b) segmentation result of the DyCON method, (c) segmentation result of the DTC method, (d) segmentation result of the MCF method, (e) segmentation result of the CML method, (f) segmentation result of the CaUCON method of this invention, and (g) the gold standard (GT). It can be clearly observed that, under the limited condition of using only 10% of the labeled data, other comparison methods (columns b to e) exhibit significant problems such as incomplete segmentation, boundary offset, or false positive annotations in areas with blurred pancreatic boundaries and areas bordering surrounding organs, especially showing a significant deficiency in capturing small structures such as the pancreatic tail. In stark contrast, the segmentation results of the CaUCON method of this invention (column f) are highly consistent with the gold standard (column g). CaUCON can accurately delineate the complete outline of the pancreas, effectively avoiding segmentation deviations caused by blurred boundaries, while accurately preserving the integrity of the pancreatic anatomical structure, clearly segmenting even small branch structures. This remarkable visual effect fully validates the synergistic effect of the adaptive consistency learning mechanism and topology-aware loss in CaUCON: the former performs differentiated learning for regions with high uncertainty, while the latter ensures the topological correctness of anatomical structures. Through these mechanisms, CaUCON ultimately achieves high-precision segmentation of complex pancreatic structures. Figure 4 The intuitive visual advantages provide a basis for subsequent quantitative performance evaluations on multiple datasets such as ISLES-2022, BraTS2019, and LA (e.g., Figure 5 , Figure 6 , Figure 7 , Figure 8 (as shown) and the effectiveness verification of each module (ablation experiments, such as...) Figure 9 , Figure 10 , Figure 11 (As shown) provided strong initial support.

[0136] Figure 5 This paper presents a performance comparison of the CaUCON method of this invention with mainstream semi-supervised learning (SSL) baseline methods such as MT, MCF, DTC, AC-MT, BCP, MagicNet, GALOSs, and DyCON on the ISLES-2022 ischemic stroke lesion segmentation dataset. This dataset exhibits significant variations in lesion size and location, and a sparse distribution, making it an ideal scenario for validating the effectiveness of uncertainty decomposition. With 10% labeled data, the CaUCON method achieves a Dice coefficient of 66.71%, significantly surpassing GALOSs' 60.05%. With 20% labeled data, the CaUCON method further improves its Dice coefficient to 70.58%, a 10.5 percentage point improvement over the baseline method BCP. This advantage primarily stems from the CaUCON method's causal uncertainty decomposition module—which accurately distinguishes between inherent data noise and model parameter uncertainty, thereby more efficiently handling the segmentation challenge of sparse lesion regions and avoiding the processing bias caused by traditional methods treating uncertainty as a single concept.

[0137] Figure 6 Experimental results of the CaUCON method and baseline methods on the BraTS2019 brain tumor segmentation dataset, which presents challenges such as class imbalance and tumor heterogeneity, are presented. With 10% labeled data, the CaUCON method achieves a Dice coefficient of 88.24%, higher than the DyCON method's 87.19%; with 20% labeled data, the CaUCON method achieves a Dice coefficient of 89.91%, consistently outperforming the DyCON method's 88.68%. Furthermore, the CaUCON method reduces the 95% Hausdorff distance by 1.8 mm compared to traditional methods. This is attributed to the topology-aware consistency loss—which uses Betti number calculations to capture the complex shape and anatomical connectivity of brain tumors, avoiding segmentation structure breaks or spurious connections caused by traditional pixel-level losses, thus ensuring the topological correctness of the tumor region.

[0138] Experimental results from the LA (left atrium) and Pancreas (pancreas) datasets ( Figure 7 , Figure 8This further validates the generalization ability of the CaUCON method. On the LA dataset, using only 5% labeled data, the Dice coefficient of the CaUCON method reaches 91.57%, and with 10% labeled data, the Dice coefficient increases to 92.83%, significantly surpassing UA-MT's 90.02%, demonstrating the adaptability of causal uncertainty decomposition to data-scarce scenarios in segmentation tasks. On the Pancreas dataset, which has blurred boundaries and is easily affected by surrounding organs, the Dice coefficient of the CaUCON method is 84.95% with 20% labeled data, an improvement of 2.17 percentage points compared to BCP's 82.78%. This is attributed to the adaptive consistency learning mechanism—which uses dual-path weight fusion and differential modulation to specifically learn the high-uncertainty region of the pancreatic boundary, improving the segmentation accuracy of blurred boundaries.

[0139] This embodiment verifies the effectiveness and necessity of each module of the present invention by setting up an ablation experiment. All experimental results have passed the statistical significance test (paired t test, p<0.05), ensuring the reliability of the conclusions.

[0140] The first set of ablation experiments verified the effect of causal uncertainty decomposition, setting up three comparative versions: (1) completely removing the uncertainty decomposition module, treating all pseudo-labels equally; (2) using traditional total uncertainty (i.e., prediction entropy) to weight pseudo-labels, without decomposing chance and cognition; (3) the complete method, using the decomposed dual uncertainty for adaptive weighting. By comparing the performance of these three versions, the improvement brought by uncertainty decomposition can be quantified. The experimental results are as follows: Figure 9 As shown.

[0141] from Figure 9 As can be seen in the figure ("No" represents no; "Yes" represents yes), when the decomposition module is completely removed (causal decomposition = No, policy = No), the Dice coefficients of the BraTS2019 and Pancreas datasets are only 81.55% and 71.58%, respectively, showing the worst performance. When only traditional total uncertainty weighting is used (causal decomposition = Yes, policy = No), the Dice coefficients of the two datasets are improved by 1.66% and 2.74%, respectively, proving that decomposing uncertainty alone can initially improve the accuracy of pseudo-label weighting. The complete version (causal decomposition = Yes, policy = Adaptive) achieves the highest Dice coefficient, reaching 85.76% on the BraTS2019 dataset and 78.43% on the Pancreas dataset, representing improvements of 4.21% and 6.85%, respectively, compared to the version with the decomposition module completely removed. This difference indicates that traditional total uncertainty cannot distinguish between inherent data noise and model parameter bias, while dual uncertainty decomposition can achieve accurate typological weighting, significantly improving segmentation performance.

[0142] The second set of ablation experiments focused on adaptive weight networks, verifying the advantages of spatial adaptive weighting compared to global uniform weighting. Comparative versions included: (1) a method using fixed weights, i.e., setting the same pseudo-label weights for all pixels; (2) a method using category-level weights, i.e., setting different weights for different categories but uniform weights within the same category; and (3) a complete pixel-level adaptive weighting method. This set of experiments revealed the importance of fine-grained weighting, with results as follows: Figure 10 As shown.

[0143] Figure 10 Data shows that the fixed-weight version of the Dice coefficient (Baseline plus causal decomposition) is 85.76% and 78.43% in the two datasets, respectively, resulting in coarse weighting due to neglecting spatial heterogeneity. The class-weighted version (with Dual PathWeighting) improves the Dice coefficient by 0.58% (BraTS2019) and 2.24% (Pancreas), demonstrating that class distinction can initially optimize weighting. The pixel-level adaptive version (Adaptive Consistency Learning) further improves the Dice coefficient to 87.91% and 82.98%, respectively, an improvement of 2.15% and 4.55% compared to the fixed-weighted version. This result verifies the necessity of fine-grained weighting—for example, high-uncertainty regions such as pancreatic boundaries and brain tumor edges need to have reduced weights to avoid noise interference, while low-uncertainty internal organ regions need to have increased weights to reinforce learning. Fixed / class-weighted methods cannot meet this spatial heterogeneity requirement.

[0144] The third set of ablation experiments analyzed the contributions of different loss function components. Using the supervised loss (cross-entropy and Dice loss) based solely on labeled data as a baseline, pseudo-label loss, consistency regularization loss, and uncertainty regularization loss were gradually added, and the incremental contribution of each loss term was observed. Specifically, for the pseudo-label loss, the difference between having and not having uncertainty weighting was further compared to verify the value of intelligent pseudo-label screening. Experimental results are as follows: Figure 11 As shown (using 10% of the BraTS2019 dataset labeled data as an example).

[0145] from Figure 11As can be seen, the baseline performance of supervised loss alone is the worst (Dice 81.55%, HD95 22.42mm). After adding unweighted pseudo-label loss (adding Causal Uncertainty Decomposition), Dice improves by 2.23%, but HD95 is still as high as 18.65mm, proving that pseudo-labels can expand the data but need to be filtered. Further adding consistency regularization loss (adding Adaptive Consistency Learning) improves Dice to 86.42% and HD95 decreases to 12.33mm, indicating that consistency constraints can stably predict. After adding uncertainty regularization loss (adding Topology-aware loss), Dice improves by another 1.49%, and the complete version of pseudo-label and uncertainty weighted collaboration (CompleteCaUCON) reduces HD95 by an additional 11.67mm compared to the unweighted pseudo-label loss version, verifying that intelligent pseudo-label filtering can effectively reduce noise label interference and improve segmentation accuracy.

[0146] Each ablation experiment was repeated on multiple datasets, including LA (left atrium), Pancreas (pancreas), and BraTS2019 (brain tumor), to ensure the generalizability of the results. The experimental results collectively demonstrate that CaUCON's core advantage stems from the synergistic effect of causal uncertainty decomposition, adaptive weight network, and topology-aware loss. The decomposition module provides accurate differentiation of uncertainty types, the weight network enables fine-grained learning scheduling, and the topology loss ensures the integrity of anatomical structures. Furthermore, teacher updates with EMA α=0.99, a combination of strong and weak enhancements, 10% annotation, and 6000 iterations of training configuration further enhance model stability and generalization. Ultimately, CaUCON achieves a 3.2% to 4.8% improvement in the Dice coefficient in low-annotation scenarios, providing an efficient solution for semi-supervised medical image segmentation.

[0147] Example 2

[0148] This embodiment provides a semi-supervised medical image segmentation system based on causal uncertainty decomposition. The system includes:

[0149] The data preprocessing module is used to acquire and preprocess medical image datasets, which include labeled and unlabeled data.

[0150] The segmentation model construction module is used to build segmentation models for teacher and student architectures. The segmentation model includes teacher and student networks with the same structure. Both the student and teacher networks use 3D U-Net as the backbone network and integrate a causal uncertainty decomposition module, an adaptive consistency learning module, and a topology-aware consistency loss module.

[0151] The training module trains the segmentation model using labeled and unlabeled data. In each iteration, the training module performs the following operations: Forward propagation of the input image into the student and teacher networks, respectively, to obtain the predictions and feature maps of the student network and the predictions of the teacher network; decomposing the total uncertainty in medical image segmentation into cognitive uncertainty and random uncertainty using a causal uncertainty decomposition module; generating adaptive weights and modulation factors based on cognitive and random uncertainties using an adaptive consistency learning module; calculating the topological distance between the student and teacher network predictions using a topology-aware consistency loss module, and calculating the topology consistency loss based on the topological distance; calculating the total loss function, which includes supervised loss based on labeled data, causal decomposition loss based on unlabeled data, adaptive consistency loss, and topology consistency loss; updating the student network parameters through backpropagation; and updating the teacher network parameters based on the exponential moving average of the student network parameters.

[0152] The segmentation execution module is used to segment the target medical image using a trained segmentation model.

[0153] Example 3

[0154] This embodiment provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the semi-supervised medical image segmentation method based on causal uncertainty decomposition as described in Embodiment 1.

[0155] Example 4

[0156] This embodiment provides an electronic device, including:

[0157] Memory, used to store computer programs;

[0158] A processor is used to execute a computer program to implement the semi-supervised medical image segmentation method based on causal uncertainty decomposition as described in Embodiment 1.

[0159] In summary, this invention achieves optimal or near-optimal performance in terms of segmentation accuracy across all test datasets. On the LA dataset, using only 5% labeled data, the invention achieves a Dice coefficient of 89.2%, representing a 12.5 percentage point improvement over fully supervised methods (using only 5% data), a 4.8 percentage point improvement over traditional semi-supervised methods (mean teacher), and a 2.3 percentage point improvement over the latest uncertainty-aware method, UA-MT. This significant improvement is particularly valuable in scenarios where labeled data is extremely scarce, as it greatly reduces reliance on expensive annotation work. On the Pancreas dataset, pancreas segmentation is a recognized challenging task. This invention achieves a Dice coefficient of 78.5% with 10% labeled data, a 3.4 percentage point improvement over UA-MT's 75.1%, and the 95% Hausdorff distance is reduced from 6.8 mm to 5.2 mm, indicating more accurate boundary localization. On the ACDC dataset, the average Dice coefficient of this invention reached 91.7% with a 20% annotation ratio, approaching the 92.5% of fully supervised methods using all labeled data, demonstrating its effectiveness as a replacement for fully supervised learning in moderate sample scenarios. On the BraTS2019 dataset, in the whole-brain tumor segmentation task, the overall Dice coefficient of this invention was 85.3% with a 10% annotation ratio, significantly higher than the comparison methods, especially in the difficult-to-segment tumor enhancement regions, where the performance improvement was even more significant.

[0160] In terms of uncertainty estimation quality, this invention demonstrates superior calibration performance, which is crucial for clinical decision support. Traditional methods often output uncertainties that do not match the actual error rate, making it difficult for physicians to accurately assess the reliability of model predictions. This invention significantly improves the accuracy of uncertainty estimation through causal uncertainty decomposition. On the LA dataset, the expected calibration error (ECE) decreased from 0.12 for traditional methods to 0.05, indicating a high degree of consistency between the model's confidence and actual accuracy. Calibration curve analysis shows that the curve of this invention closely approximates the ideal diagonal, while the comparative methods generally suffer from overconfidence or overconservatism. More importantly, visualization of the uncertainty plot clearly shows that this invention can accurately identify difficult-to-segment regions, such as organ boundaries and lesion edges. The high uncertainty in these regions prompts physicians to pay extra attention, thereby improving the model's clinical interpretability and reliability.

[0161] In terms of computational efficiency, this invention maintains reasonable computational overhead while improving performance. Despite the introduction of a causal uncertainty decomposer and an adaptive weight network, the number of model parameters only increases by less than 2% due to the lightweight design, from 14.3M in the original U-Net to 14.5M. The number of floating-point operations (FLOPs) increases by approximately 5%, still within an acceptable range. Regarding training time, compared to the basic semi-supervised method, the training time per epoch only increases by approximately 15%, mainly due to the additional overhead of uncertainty calculation and weight generation; however, considering the significant performance improvement, this time cost is entirely worthwhile. During the inference phase, the segmentation time for a single 512×512×64 3D image is approximately 0.8 seconds (on a V100 GPU), meeting the real-time requirements of clinical applications. Compared to Transformer-based semi-supervised methods, this invention achieves approximately 3 times faster inference and approximately 2 times faster training, while maintaining comparable or better segmentation performance, demonstrating an excellent balance between efficiency and effectiveness.

[0162] This invention demonstrates excellent robustness and generalization ability. Cross-dataset experiments show that when a model trained on one dataset is applied to another anatomical site, its performance degrades only slightly. For example, a model trained on the LA dataset, when directly applied to the ACDC dataset, still achieves a Dice coefficient of over 85%, while traditional methods often drop below 75%. This strong generalization ability is attributed to the causal uncertainty decomposition, which helps the model learn more essential feature representations rather than overfitting to the characteristics of a specific dataset. Furthermore, robustness tests against differences in image quality (such as different scanners and parameter settings) show that the performance fluctuations of this invention are significantly smaller than those of the comparative methods, maintaining relatively stable segmentation results even on low-quality images.

[0163] In terms of practical clinical application value, this invention solves a key bottleneck in the implementation of medical AI. Medical institutions typically possess a large amount of unlabeled image data, but expert annotation is costly and time-consuming. This invention can achieve near-fully supervised performance by utilizing unlabeled data with only a small number of labeled samples, significantly reducing the construction cost and time of AI systems. For rare diseases, due to the scarcity of cases, it is difficult to collect a large amount of labeled data. The small-sample learning capability of this invention makes it possible to develop auxiliary diagnostic systems for these diseases. In addition, the uncertainty estimation function provides doctors with a reliability reference for model predictions. In areas of high uncertainty, the system can prompt doctors to conduct manual review, forming a human-machine collaborative diagnostic mode that improves efficiency and ensures safety. Preliminary trial feedback from multiple hospitals indicates that this invention significantly reduces the workload of doctors in tasks such as liver tumor segmentation and lung nodule detection, while maintaining high diagnostic accuracy.

[0164] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A semi-supervised medical image segmentation method based on causal uncertainty decomposition, characterized in that, Includes the following steps: Step 1: Obtain a medical image dataset and preprocess it. The dataset includes labeled data and unlabeled data. Step 2: Construct a segmentation model for the teacher and student architectures. The segmentation model includes a teacher network and a student network with the same structure. Both the student network and the teacher network use a 3D U-Net as the backbone network and integrate a causal uncertainty decomposition module, an adaptive consistency learning module, and a topology-aware consistency loss module. Step 3: Train the segmentation model using the labeled and unlabeled data, performing the following operations in each training iteration: Step 3.1: Input the input image into the student network and the teacher network respectively for forward propagation to obtain the prediction results and feature maps of the student network, and the prediction results of the teacher network; Step 3.2: By directly inputting the feature map output by the student network into the causal uncertainty decomposition module, the total uncertainty in medical image segmentation is decomposed into cognitive uncertainty and random uncertainty, wherein the cognitive uncertainty represents the noise inherent in the data, and the random uncertainty represents the uncertainty of the model parameters; Step 3.3: Based on the cognitive uncertainty and stochastic uncertainty, generate adaptive weights and modulation factors through the adaptive consistency learning module; Step 3.4: Calculate the topological distance between the student network prediction and the teacher network prediction using the topological awareness consistency loss module, and calculate the topological consistency loss based on the topological distance; Step 3.5: Calculate the total loss function, which includes the supervision loss based on labeled data, the causal decomposition loss based on unlabeled data, the adaptive consistency loss, and the topological consistency loss; Step 3.6: Update student network parameters via backpropagation; Step 3.7: Update the parameters of the teacher network based on the exponential moving average of the student network parameters; Step 4: Apply the trained segmentation model to segment the target medical image.

2. The semi-supervised medical image segmentation method based on causal uncertainty decomposition according to claim 1, characterized in that, In step 2, the causal uncertainty decomposition module, based on Pearl's hierarchy of causality theory, decomposes total uncertainty into cognitive uncertainty and stochastic uncertainty by constructing a causal graph, and optimizes it through a decomposition loss function; the decomposition loss function... The calculation method is as follows: in, for Reconstruction loss; Total uncertainty; It is a random uncertainty; To understand uncertainty; and These are adaptive learning parameters; These are the weighting coefficients; The regularization term is calculated as follows: in, The prior is a uniform distribution; Let KL divergence be denoted as KL divergence.

3. The semi-supervised medical image segmentation method based on causal uncertainty decomposition according to claim 1, characterized in that, In step 2, the adaptive consistency learning module generates an adaptive weight graph through a dual-path weight fusion network, and the calculation method is as follows: in, Student network feature graph; Capturing self-attention structures Spatial dependence; This is a feature-level attention function based on a self-attention mechanism, used to capture... Spatial dependencies; and These are respectively a random uncertainty weight generator and a cognitive uncertainty weight generator; This indicates element-wise multiplication; Dual-path adaptive weights; Furthermore, a differentiated modulation factor is applied to enhance consistency constraints in regions with high cognitive uncertainty and to reduce consistency constraints in regions with high random uncertainty. The calculation method is as follows: in, It is a random modulation factor; Cognitive modulation factor; and These represent the spatial standard deviations of random uncertainty and cognitive uncertainty, respectively. and These are learnable parameters; The adaptive consistency loss is calculated by combining modulation factor, adaptive weights, and progressive scheduling to optimize unlabeled data; the adaptive consistency loss... The calculation method is as follows: in, Loss due to basic consistency; This indicates the unlabeled dataset. Take the expected value.

4. The semi-supervised medical image segmentation method based on causal uncertainty decomposition according to claim 1, characterized in that, In step 2, the topology-aware consistency loss module calculates Betti quantified topological features and defines the topological distance between the student model prediction and the teacher model prediction. The calculation formula is as follows: in, Indicates the number of connected components. Indicates the number of one-dimensional holes. Weighting for penalties related to holes. The segmentation results of the student model on the input medical image. The segmentation result of the teacher model on the input medical image; The topological distance is combined with pixel-level loss to obtain topological awareness consistency loss. The calculation method is as follows: in, The topological loss weights for the current iteration. For maximum weight, This represents the current iteration number. For pixel-level consistency loss, This is the initial iteration for topological loss. This is the weighting growth cycle.

5. The semi-supervised medical image segmentation method based on causal uncertainty decomposition according to claim 1, characterized in that, In step 3.5, the total loss function The calculation method is as follows: in, For the supervision loss function, The loss function is the causal decomposition function. For adaptive consistency loss function, Let the topology consistency loss function be... , , These are the corresponding weighting coefficients.

6. The semi-supervised medical image segmentation method based on causal uncertainty decomposition according to claim 1, characterized in that, In step 3.7, the teacher network parameters are updated using an exponential moving average method: in, For teacher network parameters, For student network parameters, This is the momentum factor, with a value ranging from 0.9 to 0.

999.

7. A semi-supervised medical image segmentation system based on causal uncertainty decomposition, characterized in that, include: The data preprocessing module is used to acquire and preprocess medical image datasets, which include labeled and unlabeled data. The segmentation model construction module is used to construct segmentation models of teacher and student architectures. The segmentation model includes teacher networks and student networks with the same structure. Both the student network and the teacher network use 3D U-Net as the backbone network and integrate a causal uncertainty decomposition module, an adaptive consistency learning module, and a topology-aware consistency loss module. A training module is used to train the segmentation model using the labeled and unlabeled data. This training module is configured to perform the following operations in each iteration: input the input image into a student network and a teacher network for forward propagation, respectively, to obtain the prediction results and feature maps of the student network and the prediction results of the teacher network; decompose the total uncertainty in medical image segmentation into cognitive uncertainty and random uncertainty using the causal uncertainty decomposition module; generate adaptive weights and modulation factors based on the cognitive uncertainty and random uncertainty using the adaptive consistency learning module; calculate the topological distance between the student network predictions and the teacher network predictions using the topological awareness consistency loss module, and calculate the topological consistency loss based on the topological distance; calculate the total loss function, which includes supervised loss based on labeled data, causal decomposition loss based on unlabeled data, adaptive consistency loss, and topological consistency loss; update the student network parameters through backpropagation; and update the teacher network parameters according to the exponential moving average of the student network parameters. The segmentation execution module is used to segment the target medical image using a trained segmentation model.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the semi-supervised medical image segmentation method based on causal uncertainty decomposition as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the semi-supervised medical image segmentation method based on causal uncertainty decomposition as described in any one of claims 1 to 6 when executing the computer program.

Images