Hybrid augmented medical image segmentation method and system based on uncertainty perception
By constructing a bidirectional hybrid enhancement framework based on uncertainty awareness, the problems of pseudo-label errors and data distribution differences in semi-supervised medical image segmentation are solved, achieving more efficient medical image segmentation and improving the model's generalization ability and segmentation accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANDONG UNIV
- Filing Date
- 2025-12-23
- Publication Date
- 2026-08-04
AI Technical Summary
Existing semi-supervised medical image segmentation methods are prone to generating false labels when dealing with complex medical images, leading to noise accumulation and insufficient model generalization ability. Furthermore, they fail to effectively utilize the distributional differences between labeled and unlabeled data and the complementary information of regions with high and low uncertainty.
A bidirectional hybrid enhancement framework based on uncertainty perception is constructed. The uncertainty of pseudo-labels is evaluated by KL divergence, high uncertainty regions are cropped and low uncertainty regions are removed. The student model and teacher model are combined to perform medical image segmentation, so as to achieve semantic information complementarity between labeled and unlabeled data.
It improves the model's attention and segmentation accuracy in complex regions, enhances the model's generalization performance and robustness, and enables it to better adapt to diverse clinical scenarios, thereby improving segmentation accuracy and model stability.
Smart Images
Figure CN121937713B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image segmentation technology, and particularly relates to a hybrid enhancement medical image segmentation method and system based on uncertainty perception. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] Medical image segmentation is a crucial task in computer-aided diagnosis and treatment planning, and its accurate identification and contour extraction of lesion areas have significant clinical implications. However, high-precision annotation of medical images heavily relies on manual drawing by professional physicians, which is costly and time-consuming, resulting in a scarcity of labeled data and the ineffective utilization of a large amount of unlabeled data. To alleviate the problem of insufficient annotation resources, semi-supervised learning methods are widely used. These methods aim to train the model using a small number of labeled samples and a large number of unlabeled samples to improve model performance.
[0004] Existing semi-supervised medical image segmentation methods mainly include pseudo-labeling methods, contrastive learning, and data augmentation strategies. Among these, pseudo-labeling techniques involve the model predicting unlabeled images and selecting high-confidence predictions as pseudo-labels to augment the supervision signal, thereby improving the model's generalization ability. However, this type of method suffers from the following significant problems: (1) The accuracy of pseudo-labels depends on the reliability of the model's current predictions. Due to the complex structure, blurred boundaries, and class imbalance in medical images, the model is prone to making incorrect predictions for uncertain regions. If these erroneous pseudo-labels are directly introduced into the training process, it will lead to noise accumulation and error propagation, thereby reducing the model's convergence stability and final segmentation accuracy.
[0005] (2) Most existing methods fail to adequately coordinate the distribution differences between labeled and unlabeled data. Labeled data typically has limited coverage and a single distribution, while unlabeled data contains richer anatomical variations and image features. If the training process relies too heavily on limited labeled data or fails to establish an effective consistency constraint mechanism, the model will struggle to fully exploit the potential semantic information in the unlabeled data, resulting in insufficient generalization ability and decreased performance in complex clinical scenarios.
[0006] (3) Existing methods for model tuning mostly focus on one-way information transmission and fail to effectively utilize complementary information in regions of high and low uncertainty. This approach can easily lead to insufficient model learning of samples and limit the model's ability to model the global data distribution. Summary of the Invention
[0007] To overcome the shortcomings of the prior art, this invention provides a hybrid enhanced medical image segmentation method and system based on uncertainty perception, which can perform accurate segmentation based on different medical image tasks and scenarios.
[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of the present invention provides a hybrid enhanced medical image segmentation method based on uncertainty perception.
[0009] Hybrid enhancement medical image segmentation methods based on uncertainty awareness include: Obtain a medical image dataset that includes both labeled and unlabeled images; A bidirectional hybrid enhancement framework incorporating both student and teacher models is constructed, and medical image segmentation is performed based on this framework. Specifically: Two sets of unlabeled images and two sets of labeled images are randomly selected from a medical image dataset. A pre-trained teacher model is used to predict the probability of the two sets of unlabeled images and solve the KL divergence to obtain the sub-regions with the highest and lowest uncertainty in the two sets of unlabeled images. The first mixed input, the second mixed input and the corresponding mixed label are generated through cropping and removal operations. A student model is used to learn the first and second mixed inputs; medical image segmentation is performed based on the learned bidirectional hybrid enhancement framework.
[0010] Furthermore, the KL divergence is used to assess uncertainty by importing the probability output by the teacher model into the KL divergence calculation model.
[0011] Furthermore, the implementation of the uncertainty assessment includes: first, arranging the predicted probabilities of each pixel output by the teacher model in descending order along the category dimension to obtain a sorted probability distribution map; and calculating the KL divergence of the corresponding pseudo-label based on the probability distribution map.
[0012] Furthermore, the generation of the first mixed input includes: cropping the region with the highest uncertainty in the first group of unlabeled images and pasting it into the first group of labeled images to generate the first mixed input.
[0013] Furthermore, the generation of the second mixed input includes: removing the region with the lowest uncertainty from the second set of unlabeled images and pasting it into the second set of labeled images to generate the second mixed input.
[0014] Furthermore, before performing medical image segmentation based on the bidirectional hybrid enhancement framework, the bidirectional hybrid enhancement framework is pre-trained. Specifically, two sets of labeled data are selected from the current batch and random binary masks are generated to construct a hybrid image and corresponding labels. The hybrid image is then input into the bidirectional hybrid enhancement framework for forward propagation, and supervised learning is performed using a dual loss function.
[0015] Furthermore, after pre-training is completed, the pre-trained model is used as the teacher model to generate pseudo-labels for unlabeled images. In each training iteration, the parameters of the student model are first optimized by stochastic gradient descent, and then the parameters of the teacher model are updated using its exponential moving average.
[0016] A second aspect of the present invention provides a hybrid enhanced medical image segmentation system based on uncertainty perception.
[0017] A hybrid augmented medical image segmentation system based on uncertainty perception includes: The data acquisition module is configured to acquire a medical image dataset that includes both labeled and unlabeled images. The medical image segmentation module is configured to: construct a bidirectional hybrid enhancement framework comprising a student model and a teacher model, and perform medical image segmentation based on the constructed bidirectional hybrid enhancement framework. Specifically: Two sets of unlabeled images and two sets of labeled images are randomly selected from a medical image dataset. The teacher model is used to perform probability prediction on the two sets of unlabeled images and solve the KL divergence to obtain the sub-regions with the highest and lowest uncertainty in the two sets of unlabeled images. The first mixed input, the second mixed input and the corresponding mixed label are generated through cropping and removal operations. A student model is used to learn the first and second mixed inputs; medical image segmentation is performed based on the learned bidirectional hybrid enhancement framework. A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the uncertainty-aware hybrid enhancement medical image segmentation method as described in the first aspect of the present invention.
[0018] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the uncertainty-aware hybrid enhancement medical image segmentation method as described in the first aspect of the present invention.
[0019] The above one or more technical solutions have the following beneficial effects: (1) This invention is based on a bidirectional hybrid enhancement framework that includes a student model and a teacher model for medical image segmentation. The uncertainty of the pseudo-labels generated by the teacher model is quantified region by region through the KL divergence evaluation mechanism. This enhances the model's attention to complex regions and improves model performance.
[0020] (2) This invention proposes a bidirectional hybrid enhancement mechanism. Through uncertainty-aware mask construction, high uncertainty regions in unlabeled images are fused with labeled images, while low uncertainty regions are removed from unlabeled images and then fused with labeled images. This operation achieves semantic information complementarity and distribution alignment between labeled and unlabeled data at the pixel level, enabling the model to make fuller use of the rich anatomical structural variations in unlabeled data, and significantly improving the model's generalization performance and robustness in diverse clinical scenarios.
[0021] (3) This invention employs a bidirectional hybrid strategy. On the one hand, it enhances the semantic consistency of the foreground target by introducing high-uncertainty regions; on the other hand, it enhances the learning of the background by eliminating low-uncertainty regions. This bidirectional information utilization mechanism enables the model to learn effective features in difficult samples more comprehensively, strengthens the representation ability of complex structures and regions with ambiguous boundaries, and thus comprehensively improves the modeling quality and segmentation accuracy of global data distribution.
[0022] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0023] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0024] Figure 1 This is a flowchart of the hybrid enhancement medical image segmentation method based on uncertainty perception in Embodiment 1 of the present invention.
[0025] Figure 2 This is a flowchart of mask construction based on KL divergence in Embodiment 1 of the present invention. Detailed Implementation
[0026] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0027] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.
[0028] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.
[0029] To facilitate understanding of the technical solution of this invention, the following terms are explained: 1) Semi-supervised learning: A learning method that combines a small amount of labeled data with a large amount of unlabeled data to train the model, aiming to improve model performance in scenarios where labeling costs are high. This method utilizes the latent structural information of unlabeled data to guide the model to improve its generalization ability while maintaining the accuracy of labeled samples.
[0030] 2) Pseudo-labels: In semi-supervised learning, the model predicts the class of unlabeled samples and selects the class with higher confidence as its "pseudo" label for subsequent training. This method can transform unlabeled data into a form that can be used for supervised learning, thereby expanding the scale of training data.
[0031] 3) Mean-Teacher Model: A typical semi-supervised learning framework consisting of a student model and a teacher model with identical structures. The student model updates its parameters through regular backpropagation, while the teacher model's parameters are updated through a moving average of the student model's parameters.
[0032] The overall concept of this invention is as follows: This invention provides a hybrid enhancement method for medical image segmentation based on uncertainty awareness, constructing a bidirectional hybrid enhancement framework based on uncertainty awareness. This invention uses a Mean-Teacher model as the framework. Specifically: First, two sets of unlabeled images and two sets of labeled images are randomly selected from the training set. The unlabeled images are input into the pre-trained teacher model to obtain pseudo-labels. The deterministic level of each region in the pseudo-labels is evaluated using the KL divergence formula, and the sub-regions with the highest and lowest uncertainty are extracted respectively. Based on this, a bidirectional hybridization operation is performed to obtain a hybrid image and label: For the first set of labeled and unlabeled images, the region with the highest uncertainty in the pseudo-labels corresponding to the unlabeled images is retained as the foreground and fused with the real label of the labeled images to obtain a hybrid label. The same operation is performed between the two sets of images to obtain a hybrid image. For the second set of labeled and unlabeled images, the region with the lowest uncertainty in the pseudo-labels corresponding to the unlabeled images is removed as the background and mixed with the real label to reduce the interference of noisy labels on model training. The same operation is performed between the two sets of images to obtain a hybrid image. Finally, these hybrid images are input into the student model, and training is performed using the hybrid labels as supervision signals.
[0033] The uncertainty-aware bidirectional hybrid enhancement framework, as a semi-supervised learning framework, fully leverages unlabeled information in medical image data. This approach will help address the high cost and difficulty of data labeling in the medical field. By utilizing unlabeled data, this invention can further improve the algorithm's performance and generalization ability, thus better adapting to different medical image tasks and scenarios. This will bring new breakthroughs to the field of medical image analysis, thereby advancing disease diagnosis and treatment.
[0034] Example 1 This embodiment discloses a hybrid enhancement medical image segmentation method based on uncertainty perception.
[0035] like Figure 1 As shown, the uncertainty-aware hybrid enhancement medical image segmentation method includes: Obtain a medical image dataset that includes both labeled and unlabeled images; A bidirectional hybrid enhancement framework incorporating both student and teacher models is constructed, and medical image segmentation is performed based on this framework. Specifically: Two sets of unlabeled images and two sets of labeled images are randomly selected from a medical image dataset. A pre-trained teacher model is used to perform probability prediction on the two sets of unlabeled images and calculate the KL divergence to obtain the sub-regions with the highest and lowest uncertainty in the two sets of unlabeled images. The first mixed input, the second mixed input, and the corresponding mixed labels are generated through cropping and removal operations. A student model is used to learn the first and second mixed inputs, and medical image segmentation is performed based on the learned bidirectional hybrid enhancement framework.
[0036] Based on the above process, this invention can perform accurate segmentation while adapting to different medical image tasks and scenarios. To facilitate understanding of the technical solution of this invention, the specific implementation methods of this invention will be further explained and described below.
[0037] In the specific implementation process, given a dataset Dataset By a containing N A labeled subset of samples and a containing M Unlabeled subset of samples Composition, and .in, and They represent the first A labeled medical image and its corresponding real segmentation label; Indicates the first A medical image of an unlabeled sample.
[0038] In the specific implementation process, firstly, two sets of unlabeled images are randomly selected from the dataset. ) and two sets of labeled images ( Two sets of unlabeled images are input into a pre-trained teacher model, and the output probabilities are... Import the KL divergence calculation formula and find the sub-regions with the highest and lowest uncertainty in the two sets of unlabeled images, respectively. Crop. Paste the area with the highest uncertainty into The first mixed input is obtained. Remove Paste the area with the lowest uncertainty into The second mixed input is obtained above. Performing the same operation on pseudo-labels and labels yields a mixed label used as a supervisory signal. and The student model is input and learned using mixed labels as supervision signals. Through these operations, unlabeled images can learn more comprehensive semantic information from labeled images.
[0039] Before performing medical image segmentation based on the bidirectional hybrid enhancement framework, the framework needs to be pre-trained. Influenced by previous work, data augmentation methods are used for labeled samples during the self-training phase. Specifically, two sets of labeled data are selected from the current batch (…). , )and( , ), and generate a random binary mask. M Construct the blended image and its corresponding label: ; ; Blended image The data is fed into the model for forward propagation, and supervised learning is performed using the Dice loss function and the cross-entropy loss function. ; in, This represents the sum of the Dice loss function and the cross-entropy loss function; where, the Dice loss function... With cross-entropy loss function They are represented as follows: ; ; in, This represents the probability predicted by the model. Indicates the true label, Total number of pixels; This is a smoothing factor used to avoid the denominator being zero.
[0040] Furthermore, the training framework proposed in this invention includes a teacher network. and a student network ;in, and Let represent the parameters of the teacher and student networks, respectively. The parameters of the student network are optimized using stochastic gradient descent (SGD), while the parameters of the teacher network are updated using an exponential moving average (EMA) of the student parameters.
[0041] The entire training process is divided into three stages: First, the model is pre-trained using labeled data; then, the pre-trained model is used as the teacher network to generate pseudo-labels for unlabeled images; in each training iteration, the parameters of the student network are first optimized using stochastic gradient descent. Then, its exponential moving average is used to update the parameters of the teacher network. .
[0042] Furthermore, medical image segmentation based on a bidirectional hybrid enhancement framework can be achieved through the following methods: 1) Pseudo-label generation and KL divergence calculation.
[0043] To generate pseudo-labels for unlabeled images and quantize the KL divergence of these pseudo-labels, and After applying a normalized softmax operation to the raw output of the input teacher model, its probability map is obtained. and .
[0044] Next, according to the probability diagram and Pseudo-labels are generated for subsequent learning. Specifically, for each pixel location, the category index corresponding to the highest probability is taken as the prediction, and a pseudo-label is generated. and .
[0045] To evaluate the uncertainty of pseudo-labels, the predicted probabilities of each pixel are sorted in descending order along the category dimension, thus obtaining a sorted probability distribution. Based on the sorted probability map, pseudo-labels are calculated using the following formula. and KL divergence plot, which can measure uncertainty , , : ; in, Denotes KL divergence, Indicates the total number of categories. Indicates the first Predicted probability plots for each category; This represents the class distribution with the highest predicted probability, i.e., the most reliable class at present. The formula uses KL divergence to measure the class distribution. The difference between the most probable class and the lower-ranked classes is significant, and overemphasizing these differences could obscure the model's uncertainty among the major classes. Therefore, this invention introduces a decay coefficient to reduce the impact of lower-ranked classes on the overall uncertainty assessment, highlighting the focus on the differences among the top few high-confidence classes.
[0046] 2) Mask construction based on confidence level.
[0047] To obtain more relevant regions that are difficult to segment, this invention introduces a KL divergence mask construction strategy, such as... Figure 2 As shown, , Divided into N The size is The sub-block, denoted as , ;in, and This is the sub-block index. For any sub-block, calculate its average KL divergence, expressed as: ; ; in, express The Middle The first sub-block pixel value, express The Middle The first sub-block Each pixel value.
[0048] This invention is from Select the sub-block with the highest KL divergence. and from Select the image patch with the lowest KL divergence. .
[0049] according to , Generate two binary masks and ;in, Middle and Index The corresponding region is set to 1, and the rest of the regions are set to 0; while Middle and Index The corresponding regions are set to 0, and the remaining regions are set to 1. These masks are then used for image and label blending operations.
[0050] 3) Image tag bidirectional mixing.
[0051] Using uncertainty-aware binary masks, a patch-based copy-paste (CP) operation is performed between unlabeled and labeled images. Specifically, the input images are mixed. , The definition is as follows: ; ; in, This indicates element-wise multiplication. This is a binary mask generated based on the image patch confidence scores. This operation preserves... The region with the highest uncertainty is used; the remaining regions are... Replace; Remove The region with the lowest uncertainty is used Replace it.
[0052] Similarly, to ensure consistency between the blended input and its supervisory signal, the same image patch-level blending operation is performed on the labels. Blending Labels , The definition is as follows: ; ; in, and These represent unlabeled images. and pseudo-tags and These represent labeled images. and The actual labels. Then the images are blended. , Input into the student model and using mixed labels , Conduct supervised learning.
[0053] Compared to the unidirectional mixing strategy in existing technologies, the bidirectional mixing method makes fuller use of the semantic information in pseudo-labels. Using only high-uncertainty regions may cause the model to favor learning the foreground and ignore the background structure, while bidirectional mixing can alleviate this problem.
[0054] This invention uses a loss function The overall loss function is defined as follows: ; ; ; in, This means that the Dice loss function is added to the cross-entropy loss function. and Mixed images and The output, and This is a scaling factor to prevent the loss function from being completely dominated by the uncertain region.
[0055] The total loss is used to optimize the parameters of the student network using stochastic gradient descent (SGD). Subsequently, in the... In this iteration, the parameters of the teacher network... Update using the following formula: ; Building upon the Mean-Teacher framework, this invention inputs unlabeled data into a teacher model to obtain pseudo-labels. It then calculates the uncertainty of these pseudo-labels using a formula based on KL divergence, and constructs a mask based on this uncertainty to mix unlabeled and labeled data and build hybrid labels. This uncertainty-based mixing strategy effectively enhances the model's discriminative ability in ambiguous or boundary regions, improving model stability and discriminative power. Mixing labeled and unlabeled data helps the model learn semantic information bidirectionally from both labeled and unlabeled data, effectively mitigating the negative impact of distribution differences between data sources and improving the model's generalization ability and training stability.
[0056] An uncertainty-aware bidirectional hybrid enhancement framework, as a semi-supervised learning framework, fully leverages unlabeled information in medical image data. The application of this method will help address the high cost and difficulty in obtaining data labeling in the medical field. By utilizing unlabeled data, this invention will be able to further improve the performance and generalization ability of the algorithm, thereby better adapting to different medical image tasks and scenarios. This will bring new breakthroughs to the field of medical image analysis, thereby promoting progress in early disease diagnosis and treatment.
[0057] To further highlight the advantages of the present invention, the following experiments were conducted in this embodiment: Table 1 shows the ACDC dataset based on 10% of the labeled data.
[0058] To ensure a fair comparison, this embodiment follows recent research and selects 5% and 10% of medical images from the ACDC dataset as labeled data for the experiment.
[0059] As shown in Table 1, using 10% labeled data, the proposed method achieves a Dice coefficient of 89.78%, a Jaccard index of 81.78%, a 95HD of 2.13, and an ASD of 0.60. Compared to the previous model BCP, Dice and Jaccard scores are improved by 0.94% and 1.35%, respectively, while 95HD and ASD scores are reduced by 1.85 and 0.57, respectively. This indicates that the proposed method can obtain higher quality segmentation results and achieve more accurate boundary characterization.
[0060] Table 2 is based on the ACDC dataset with 5% labeled data.
[0061] Similarly, as shown in Table 2, even with only 5% labeled data, this method still outperformed other comparative methods across all evaluation metrics in this study.
[0062] The improvements in Dice and Jaccard metrics fully demonstrate the superiority of this invention in segmentation quality, while the decreases in 95HD and ASD indicate a significant advantage in boundary accuracy. These results demonstrate that this invention, as a semi-supervised learning framework, fully utilizes unlabeled information in medical image data, providing more accurate guidance for model training.
[0063] Example 2 This embodiment discloses a hybrid enhanced medical image segmentation system based on uncertainty perception.
[0064] A hybrid augmented medical image segmentation system based on uncertainty perception includes: The data acquisition module is configured to acquire a medical image dataset that includes both labeled and unlabeled images. The medical image segmentation module is configured to: construct a bidirectional hybrid enhancement framework comprising a student model and a teacher model, and perform medical image segmentation based on the constructed bidirectional hybrid enhancement framework. Specifically: Two sets of unlabeled images and two sets of labeled images are randomly selected from a medical image dataset. A pre-trained teacher model is used to predict the probability of the two sets of unlabeled images and solve the KL divergence to obtain the sub-regions with the highest and lowest uncertainty in the two sets of unlabeled images. The first mixed input, the second mixed input and the corresponding mixed label are generated through cropping and removal operations. A student model is used to learn the first and second mixed inputs; medical image segmentation is performed based on the learned bidirectional hybrid enhancement framework. Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.
[0065] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the uncertainty-aware hybrid enhancement medical image segmentation method as described in Embodiment 1 of this disclosure.
[0066] Example 4 The purpose of this embodiment is to provide an electronic device.
[0067] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the uncertainty-aware hybrid augmented medical image segmentation method as described in Embodiment 1 of this disclosure.
[0068] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.
[0069] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.
[0070] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A hybrid enhancement medical image segmentation method based on uncertainty perception, characterized in that, include: Obtain a medical image dataset that includes both labeled and unlabeled images; A bidirectional hybrid enhancement framework comprising a student model and a teacher model is constructed, and the bidirectional hybrid enhancement framework is pre-trained: two sets of labeled data are selected from the current batch and random binary masks are generated to construct a hybrid image and corresponding labels; The hybrid image is input into the bidirectional hybrid enhancement framework for forward propagation, and supervised learning is performed using a dual loss function; Medical image segmentation is performed based on the established bidirectional hybrid enhancement framework, specifically: Two sets of unlabeled images and two sets of labeled images are randomly selected from a medical image dataset. A pre-trained teacher model is used to predict the probability of the two sets of unlabeled images and solve for the KL divergence to obtain the sub-regions with the highest and lowest uncertainty in the two sets of unlabeled images. A first mixed input, a second mixed input, and corresponding mixed labels are generated through cropping and removal operations. The KL divergence is used to evaluate uncertainty by importing the probability output by the teacher model into the KL divergence calculation model, i.e.: First, and After applying a normalized softmax operation to the raw output of the input teacher model, its probability map is obtained. and According to the probability diagram and To generate pseudo-labels, specifically, for each pixel location, the category index corresponding to the highest probability is used as the prediction to generate pseudo-labels. and ; Subsequently, the prediction probability of each pixel is arranged in descending order in the category dimension to obtain an ordered probability distribution; according to the ordered probability map, the pseudo label is calculated according to the following formula and the KL divergence map of , : ； in, Denotes KL divergence, Indicates the total number of categories. Indicates the first Predicted probability plots for each category; This represents the category distribution with the highest prediction probability, i.e., the most reliable category at present; Then, a KL divergence mask construction strategy is introduced to... , Divided into N The size is The sub-block, denoted as , ;in, and For each sub-block, calculate its average KL divergence, expressed as: ； ； in, express The Middle The first sub-block pixel value, express The Middle The first sub-block Each pixel value; from Select the sub-block with the highest KL divergence. and from Select the image patch with the lowest KL divergence. ;according to , Generate two binary masks and ;in, Middle and Index The corresponding area is set to 1, and the rest of the area is set to 0; Middle and Index The corresponding area is set to 0, and the rest of the area is set to 1; A student model is used to learn the first and second mixed inputs; medical image segmentation is performed based on the learned bidirectional hybrid enhancement framework.
2. The uncertainty perception based hybrid augmented medical image segmentation method of claim 1, wherein, The generation of the first mixed input includes: cropping the region with the highest uncertainty from the first group of unlabeled images and pasting it into the first group of labeled images to generate the first mixed input.
3. The uncertainty perception based hybrid augmented medical image segmentation method of claim 1, wherein, The generation of the second mixed input includes: removing the region with the lowest uncertainty from the second set of unlabeled images and pasting it into the second set of labeled images to generate the second mixed input.
4. The uncertainty perception based hybrid augmented medical image segmentation method of claim 1, wherein, After pre-training, the pre-trained model is used as the teacher model to generate pseudo-labels for unlabeled images. In each training iteration, the parameters of the student model are first optimized by stochastic gradient descent, and then the parameters of the teacher model are updated using its exponential moving average.
5. A hybrid augmented medical image segmentation system based on uncertainty perception employing the hybrid augmented medical image segmentation method as claimed in any one of claims 1-4, characterized in that, include: The data acquisition module is configured to acquire a medical image dataset that includes both labeled and unlabeled images. The medical image segmentation module is configured to: construct a bidirectional hybrid enhancement framework comprising a student model and a teacher model, and perform medical image segmentation based on the constructed bidirectional hybrid enhancement framework. Specifically: Two sets of unlabeled images and two sets of labeled images are randomly selected from a medical image dataset. A pre-trained teacher model is used to predict the probability of the two sets of unlabeled images and solve the KL divergence to obtain the sub-regions with the highest and lowest uncertainty in the two sets of unlabeled images. The first mixed input, the second mixed input and the corresponding mixed label are generated through cropping and removal operations. The student model is used to learn the obtained first and second mixed inputs; Medical image segmentation based on a learned bidirectional hybrid enhancement framework.
6. A computer-readable storage medium having stored thereon a program, characterized in that, When the program is executed by the processor, it implements the steps in the uncertainty-aware hybrid enhancement medical image segmentation method as described in any one of claims 1-4.
7. An electronic device comprising a memory, a processor, and a program stored in the memory and capable of running on the processor, characterized by When the processor executes the program, it implements the steps in the uncertainty-aware hybrid enhancement medical image segmentation method as described in any one of claims 1-4.