A semi-supervised medical image segmentation method and system based on wavelet transform and multi-frequency auxiliary network

Abstract

A semi-supervised medical image segmentation method and system based on wavelet transform and multi-frequency auxiliary networks, belonging to the field of medical image processing technology, includes the following steps: First, an enhanced training sample is generated by mixing labeled and unlabeled images using a bidirectional copy-paste strategy. Then, wavelet transform is used to decompose the mixed sample into low-frequency and high-frequency components to explicitly separate global structural information and local detail features. Next, a multi-frequency auxiliary segmentation network is designed, comprising multiple targeted sub-networks to process different frequency components. The low-frequency sub-network uses a graph attention mechanism to capture semantic associations, while the high-frequency sub-network uses an edge attention mechanism to enhance boundary features. Finally, under the mean teacher framework, consistency constraint training is performed by weighted fusion of the losses of each sub-network, thereby achieving accurate and robust medical image segmentation with limited labeled data. This invention significantly improves the accuracy and boundary integrity of organ and lesion segmentation.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a semi-supervised medical image segmentation method and system based on wavelet transform and multi-frequency auxiliary network. Background Technology

[0002] Medical image segmentation, a key technology in medical image analysis, is of great significance for clinical diagnosis and treatment planning. In recent years, deep learning-based image segmentation methods, such as U-Net and its variants, have made significant progress in various tasks. However, these methods typically rely on large amounts of finely labeled data, while medical image annotation is costly and time-consuming, resulting in scarce labeled samples and limiting the application of these models in real-world scenarios.

[0003] To alleviate the problem of insufficient labeled data, semi-supervised learning methods have been introduced into the field of medical image segmentation, aiming to improve model performance by jointly utilizing a small amount of labeled data and a large amount of unlabeled data. Existing methods mainly include self-training strategies, consistency regularization, and teacher-student frameworks. Among them, consistency regularization methods force the model to maintain consistent output under different perturbations by applying image enhancement, noise injection, or model perturbation; while the teacher-student framework uses the teacher network to generate pseudo-labels, guiding the student network to learn effective features from unlabeled data.

[0004] However, existing methods still have significant limitations. On the one hand, the consistency perturbation methods used often disrupt the inherent spatial structure and statistical distribution of medical images, easily introducing distribution shifts and affecting the stability and reliability of segmentation results. On the other hand, most methods primarily perform overall modeling in the spatial domain, failing to fully explore the multi-scale structural and detailed information contained in the frequency domain of medical images. Medical images naturally possess multi-frequency characteristics: low-frequency components carry overall structural and semantic information, while high-frequency components contain local features such as edges and textures. Although some studies have attempted to introduce frequency domain analysis tools such as wavelet transform, they are mostly used as preprocessing or auxiliary feature enhancement methods, failing to deeply integrate with semi-supervised learning frameworks and lacking differentiated modeling mechanisms for different frequency features. Furthermore, while current attention-based methods can improve feature representation capabilities, they often employ a uniform attention modeling approach, failing to distinguish the differences in physical meaning and modeling objectives between low-frequency semantics and high-frequency edges, making it difficult to achieve coordinated optimization of global structure and local details. Summary of the Invention

[0005] One of the objectives of this invention is to provide a semi-supervised medical image segmentation method based on wavelet transform and multi-frequency auxiliary network, in order to solve the problems of insufficient image segmentation accuracy in existing semi-supervised medical image segmentation methods, such as insufficient utilization of frequency domain information, single feature modeling method, and poor stability of consistency learning.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a semi-supervised medical image segmentation method based on wavelet transform and multi-frequency assisted network, comprising the following steps:

[0007] S1. Mix the labeled and unlabeled images to generate mixed training samples;

[0008] S2. Perform wavelet transform on the mixed training samples to decompose them into multiple frequency components carrying different frequency domain information. The frequency components include at least low-frequency components and high-frequency components.

[0009] S3. Input the multiple frequency components into a multi-frequency auxiliary segmentation network for processing. The multi-frequency auxiliary segmentation network includes multiple sub-networks corresponding to the frequency components. Each sub-network is configured to extract and model the features of its corresponding frequency components and output the corresponding prediction results.

[0010] S4. Based on the teacher-student framework, the pseudo-labels generated by the teacher network are used to constrain the consistency of the prediction results of the student network. The parameters of the student network are optimized with a total loss function, which is the weighted sum of the losses corresponding to the prediction results output by the multiple sub-networks.

[0011] Preferably, in step S1, a two-way copy-paste strategy is adopted, including:

[0012] Randomly select one labeled image from the labeled dataset and one unlabeled image from the unlabeled dataset;

[0013] Using a randomly generated foreground mask, a first blending operation is performed to paste the foreground region of the labeled image onto the unlabeled image, generating a first blended image. ;

[0014] Using the foreground mask, a second blending operation is performed to paste the foreground region of the unlabeled image onto the labeled image, generating a second blended image. ;

[0015] Wherein, the first mixed image and the second mixed image This constitutes the mixed training samples.

[0016] More preferably, in step S2, the step of performing wavelet transform on the mixed training samples includes:

[0017] For the first mixed image and the second mixed image Each image is decomposed into a low-frequency component LF and at least one high-frequency component HF by applying discrete wavelet transform.

[0018] The set of frequency components obtained after decomposition is represented as follows: and The subscripts L, M, and H represent low frequency, original spatial, and high frequency components, respectively.

[0019] More preferably, the multi-frequency assisted segmentation network includes a low-frequency sub-network (LF-Net), an original image sub-network (M-Net), and a high-frequency sub-network (HF-Net), which are used to process the low-frequency component, the original spatial component, and the high-frequency component, respectively.

[0020] More preferably, the encoder of the low-frequency sub-network (LF-Net) integrates an adaptive graph attention mechanism to capture long-range dependencies and semantic consistency in low-frequency features.

[0021] More preferably, the encoder of the high-frequency sub-network (HF-Net) integrates an edge attention mechanism to accurately locate and enhance edge information in high-frequency features.

[0022] More preferably, in step S4, the step of performing consistency constraints based on the teacher-student framework includes:

[0023] The original spatial components and Input into the teacher network to generate corresponding pseudo-tags. and ;

[0024] The frequency components are set as follows and The input is fed into the student network, where the low-frequency sub-network (LF-Net), the original image sub-network (M-Net), and the high-frequency sub-network (HF-Net) output prediction results respectively. and ;

[0025] Calculate the segmentation loss between the prediction result of each sub-network and the corresponding pseudo-label; the segmentation loss of each sub-network is a combination of Dice loss and cross-entropy loss.

[0026] More preferably, the total loss function The weighted sum of the losses of the three subnetworks is expressed as:

[0027] ;

[0028] in, These represent the segmentation losses of the low-frequency subnetwork, the original image subnetwork, and the high-frequency subnetwork, respectively.

[0029] More preferably, the weighting coefficient Optimized settings: .

[0030] In addition, the present invention also provides a semi-supervised medical image segmentation system based on a wavelet transform-assisted multi-frequency network, which includes:

[0031] The mixed sample generation module is used to mix labeled and unlabeled images to generate mixed training samples;

[0032] The frequency domain decomposition module is used to perform wavelet transform on the mixed training samples, decomposing them into multiple frequency components carrying different frequency domain information. The frequency components include at least low-frequency components and high-frequency components.

[0033] The multi-frequency assisted segmentation network module includes multiple sub-networks corresponding to the frequency components. Each sub-network is configured to extract and model the features of its corresponding frequency component and output the corresponding prediction result.

[0034] The consistency training module, based on the teacher-student framework, is configured to use pseudo-labels generated by the teacher network to constrain the prediction results of the student network, and optimize the parameters of the student network with a total loss function, which is a weighted sum of the losses corresponding to the prediction results output by the multiple sub-networks.

[0035] The system operates using the method described above.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] 1. This invention proposes a novel frequency-aware representation learning module that uses wavelet transform to decompose an image into low-frequency and high-frequency components. This enables the model to simultaneously utilize both global anatomical background and fine-grained boundary details, overcoming the limitations of traditional SSL methods that primarily focus on a single-scale spatial domain.

[0038] 2. To maximize the utilization of frequency domain features, this invention introduces a dual attention design based on target multi-frequency attention: an edge attention mechanism specifically tailored for high-frequency boundary refinement and a graph attention mechanism for low-frequency semantic consistency. This specialized integration enhances the discriminative power of multi-frequency representations while maintaining spatial coherence.

[0039] 3. By using robust consistency regularization, we combine bidirectional copy-paste synthesis with LF-HF decomposition to establish a frequency-aware consistency objective. This strategy effectively mitigates learning bias caused by human perturbation and stabilizes the supervision signal of unlabeled data, ensuring higher fidelity in pseudo-label generation. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the model structure of the method of the present invention;

[0041] Figure 2 This is a schematic diagram of the MF-Net structure in the method of the present invention. Detailed Implementation

[0042] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0043] The overall architecture of this invention is as follows Figure 1 As shown, this method is based on the mean-teacher framework in semi-supervised learning, comprising a teacher network and a student network. The core of the method lies in performing frequency domain decomposition of the input image using wavelet transform, and designing a multi-frequency auxiliary network to specifically process features of different frequencies. Finally, through consistency regularization and bidirectional copy-paste enhancement strategies, efficient and accurate semi-supervised medical image segmentation is achieved.

[0044] To address the mismatch between labeled and unlabeled data distributions in semi-supervised scenarios, this invention first employs a bidirectional copy-paste strategy to generate mixed training samples. Specifically, two images are randomly selected from both the labeled and unlabeled sets, and a randomly generated foreground mask is used... Perform two blending operations: first, paste the foreground region of the labeled image onto the unlabeled image to generate... Another step involves pasting the foreground area of ​​the unlabeled image onto the labeled image to generate... This step aims to construct an intermediate training distribution to facilitate the bidirectional flow of knowledge between the two types of data.

[0045] Two blended images generated after bidirectional copy and paste and Discrete wavelet transform is applied to perform frequency domain decomposition. Each image is decomposed into a low-frequency component (LF) containing global semantic information and a high-frequency component (HF) containing edge detail information. Thus, each mixed sample is expanded into a set of frequency features with explicit physical meaning. and subscript This represents the original image. This step achieves an explicit mapping from image spatial features to multi-frequency features, laying the foundation for subsequent targeted processing.

[0046] further, Figure 2 The structure of the Multi-Frequency Assisted Segmentation Model (MF-Net) is demonstrated, comprising three main networks. M-Net consists of either UNet or VNet. HF-Net and LF-Net add input image quantity and an image fusion module to M-Net. Furthermore, to fully utilize the frequency domain spatial features within the original image, this embodiment employs an edge attention mechanism in the encoder of HF-Net to extract edge features, and an adaptive graph attention mechanism in the encoder of LF-Net to extract semantic features. Although low-frequency and high-frequency information is fused into complete information, the original image still contains unused information. Therefore, this embodiment also uses the original image as the network input. The three pathways output the prediction results corresponding to the frequency space. and This approach enables collaborative analysis and decision-making on the same target across different feature dimensions. The low-frequency fused image carries global structural and semantic information. Graph attention mechanisms are used to effectively capture the long-range dependencies and semantic relevance of LF features. The high-frequency fused image carries the most critical details and edge information of the image. However, this information is spatially sparse and easily affected by noise. Edge attention mechanisms are used to accurately locate edge regions and enhance high-frequency features.

[0047] Furthermore, this embodiment also utilizes a teacher network to... and Perform forward propagation to generate pseudo-labels and The student network then receives all the frequency features generated in step two as input and calculates the segmentation loss (combining Dice loss and cross-entropy loss) between each prediction and its corresponding pseudo-label. The total loss function... The weighted sum of the losses of the three subnetworks:

[0048] ;

[0049] Each of them Both contain information from and Supervisory signals, their weighting coefficients It is an adjustable hyperparameter. Verified through ablation experiments, the setting... The time-series model performs optimally on most medical image segmentation tasks. This setting effectively balances the contributions of low-frequency semantic features, raw spatial features, and high-frequency edge features to the training process. By optimizing this loss, the student network is forced to learn a robust representation consistent with mixed perturbations in the input spatial and frequency domains. The parameters of the student network are updated via gradient descent, while the parameters of the teacher network are updated via an exponential moving average of the student network parameters, forming a stable self-training loop.

[0050] As can be seen, this invention creatively integrates the multi-resolution analysis capabilities of wavelet transform with a semi-supervised learning framework through a coherent technical chain of "hybrid sample generation → frequency domain decoupling → multi-path collaborative processing → consistency optimization," achieving accurate and robust segmentation of global semantics and local boundaries in medical images under limited annotation. More specifically:

[0051] This invention explicitly separates images into frequency components carrying different physical meanings by performing wavelet transform decomposition on mixed training samples. This allows the model to utilize the inherent multi-scale information in medical images from the source, overcoming the limitations of existing methods that mainly perform overall modeling in the spatial domain and struggle to fully utilize frequency domain structure and details. Secondly, the multi-frequency assisted segmentation network uses parallel sub-networks to process different frequency components, achieving collaborative and deep modeling of global semantic information and local detailed features, effectively improving the discriminative ability of feature representation. Finally, within a teacher-student framework, a consistency constraint is applied to the student network based on the weighted loss function of the multi-sub-network output. This mechanism encourages the model to learn robust representations of the frequency and spatial perturbations of the input data, thus significantly improving the accuracy, boundary precision, and model training stability of the segmentation results even with limited labeled data. In summary, this invention achieves more accurate and robust segmentation of medical images in semi-supervised scenarios.

[0052] Experiment and effect verification

[0053] Overall Performance Validation of Semi-Supervised Medical Image Segmentation: This experiment verifies the segmentation performance of the wavelet transform-based semi-supervised multi-frequency auxiliary network (WT-MF-Net) for medical image segmentation proposed in the above embodiments under low annotation ratio conditions. The Automated Heart Diagnosis Challenge (ACDC) dataset was used in the experiment. The ratio of training set, test set, and validation set was 7:2:1. Only 10% of the samples in the training set were provided with pixel-level annotations, and the remaining samples were used as unlabeled data for semi-supervised training.

[0054] The methods compared in this experiment include the fully supervised U-Net model, the semi-supervised segmentation method based on MeanTeacher, the semi-supervised segmentation method based on consistency learning, and the WT-MF-Net method proposed in the above embodiments. The Dice coefficient, IoU, Hausdorff distance (HD95), and average surface distance (ASD) were used as evaluation metrics. The experimental results are shown in Table 1 below.

[0055] Table 1. Comparison of segmentation performance of different methods on the ACDC dataset under low annotation ratio conditions.

[0056]

[0057] Table 1 shows the average performance comparison of the proposed method with other excellent methods on the ACDC dataset across four evaluation metrics. On the ACDC dataset with a 10% annotation ratio, the proposed method outperforms other methods in all metrics. Due to the different attention mechanism modules used in MF-Net, the proposed method demonstrates stronger capabilities in extracting different frequency domain spatial features.

[0058] To facilitate understanding by those skilled in the art of the improvements of this invention over the prior art, some of the accompanying drawings and descriptions have been simplified. The above embodiments are preferred implementations of this invention. In addition, this invention can be implemented in other ways. Any obvious substitutions without departing from the concept of this technical solution are within the protection scope of this invention.

Claims

1. A semi-supervised medical image segmentation method based on wavelet transform and multi-frequency auxiliary network, characterized in that, Includes the following steps: S1. Mix the labeled and unlabeled images to generate mixed training samples; S2. Perform wavelet transform on the mixed training samples to decompose them into multiple frequency components carrying different frequency domain information. The frequency components include at least low-frequency components and high-frequency components. S3. Input the multiple frequency components into a multi-frequency auxiliary segmentation network for processing. The multi-frequency auxiliary segmentation network includes multiple sub-networks corresponding to the frequency components. Each sub-network is configured to extract and model the features of its corresponding frequency components and output the corresponding prediction results. S4. Based on the teacher-student framework, the pseudo-labels generated by the teacher network are used to constrain the consistency of the prediction results of the student network. The parameters of the student network are optimized with a total loss function, which is the weighted sum of the losses corresponding to the prediction results output by the multiple sub-networks.

2. The semi-supervised medical image segmentation method based on wavelet transform and multi-frequency auxiliary network according to claim 1, characterized in that: In step S1, a two-way copy-paste strategy is adopted, including: Randomly select one labeled image from the labeled dataset and one unlabeled image from the unlabeled dataset; Using a randomly generated foreground mask, a first blending operation is performed to paste the foreground region of the labeled image onto the unlabeled image, generating a first blended image. ; Using the foreground mask, a second blending operation is performed to paste the foreground region of the unlabeled image onto the labeled image, generating a second blended image. ; Wherein, the first mixed image and the second mixed image This constitutes the mixed training samples.

3. The semi-supervised medical image segmentation method based on wavelet transform and multi-frequency auxiliary network according to claim 2, characterized in that: Step S2, the step of performing wavelet transform on the mixed training samples, includes: For the first mixed image and the second mixed image Each image is decomposed into a low-frequency component LF and at least one high-frequency component HF by applying discrete wavelet transform. The set of frequency components obtained after decomposition is represented as follows: and The subscripts L, M, and H represent low frequency, original spatial, and high frequency components, respectively.

4. The semi-supervised medical image segmentation method based on wavelet transform and multi-frequency auxiliary network according to claim 3, characterized in that: The multi-frequency assisted segmentation network includes a low-frequency sub-network, an original image sub-network, and a high-frequency sub-network, which are used to process the low-frequency component, the original spatial component, and the high-frequency component, respectively.

5. The semi-supervised medical image segmentation method based on wavelet transform and multi-frequency auxiliary network according to claim 4, characterized in that: The encoder of the low-frequency sub-network integrates an adaptive graph attention mechanism to capture long-range dependencies and semantic consistency in low-frequency features.

6. The semi-supervised medical image segmentation method based on wavelet transform and multi-frequency auxiliary network according to claim 4, characterized in that: The encoder of the high-frequency sub-network integrates an edge attention mechanism to accurately locate and enhance edge information in high-frequency features.

7. The semi-supervised medical image segmentation method based on wavelet transform and multi-frequency auxiliary network according to claim 4, characterized in that: Step S4, the step of performing consistency constraints based on the teacher-student framework, includes: The original spatial components and Input into the teacher network to generate corresponding pseudo-tags. and ; The frequency components are set as follows and The input is fed into the student network, where the low-frequency subnetwork, the original image subnetwork, and the high-frequency subnetwork each output prediction results. and ; Calculate the segmentation loss between the prediction result of each sub-network and the corresponding pseudo-label; the segmentation loss of each sub-network is a combination of Dice loss and cross-entropy loss.

8. The semi-supervised medical image segmentation method based on wavelet transform and multi-frequency auxiliary network according to claim 7, characterized in that: The total loss function The weighted sum of the losses of the three subnetworks is expressed as: ； in, These represent the segmentation losses of the low-frequency subnetwork, the original image subnetwork, and the high-frequency subnetwork, respectively.

9. The semi-supervised medical image segmentation method based on wavelet transform and multi-frequency auxiliary network according to claim 8, characterized in that: The weighting coefficient Optimized settings: .

10. A semi-supervised medical image segmentation system based on a wavelet transform-assisted multi-frequency network, characterized in that, include: The mixed sample generation module is used to mix labeled and unlabeled images to generate mixed training samples; The frequency domain decomposition module is used to perform wavelet transform on the mixed training samples, decomposing them into multiple frequency components carrying different frequency domain information. The frequency components include at least low-frequency components and high-frequency components. The multi-frequency assisted segmentation network module includes multiple sub-networks corresponding to the frequency components. Each sub-network is configured to extract and model the features of its corresponding frequency component and output the corresponding prediction result. The consistency training module, based on the teacher-student framework, is configured to use pseudo-labels generated by the teacher network to constrain the prediction results of the student network, and optimize the parameters of the student network with a total loss function, which is a weighted sum of the losses corresponding to the prediction results output by the multiple sub-networks. The system operates using the method described in any one of claims 1-9.

Images