A biomedical image segmentation method, system, device, and medium

Abstract

The application discloses a biomedical image segmentation method, system, device and medium, and relates to the technical field of medical images, and comprises the following steps: collecting a biomedical image, and decomposing the biomedical image into a low-frequency LF component and a high-frequency HF component; the low-frequency LF component and the high-frequency HF component are respectively output to two parallel branches, a cross-branch pseudo-label guiding strategy is introduced, the pseudo-label generated by the low-frequency LF branch guides the high-frequency HF branch, the semantic context is captured, the pseudo-label generated by the high-frequency branch supervises the low-frequency branch, and the perception of fine-grained texture is enhanced; the spatial quadrant of a feature map is subjected to dynamic weight adjustment, a better frequency branch is selected as a final reasoning path, another branch is discarded, and a corresponding segmentation prediction result is generated. The application realizes collaborative modeling of low-frequency semantics and high-frequency structures of an image, and further enhances the segmentation accuracy and generalization ability of a model under full supervision and semi-supervised conditions.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, and in particular to a biomedical image segmentation method, system, device and medium. Background Technology

[0002] Automatic segmentation of biomedical images is of paramount importance in numerous clinical diagnostic and biomedical research applications, widely used in tumor detection, organ boundary delineation, and cell morphology analysis. With the rapid development of deep learning technology, U-shaped network structures based on U-Net and its variants have achieved significant results in biomedical image segmentation tasks. However, most existing research focuses on improving the model's feature extraction capabilities through architectural design or attention mechanisms, paying less attention to the intrinsic properties of biomedical images in the frequency domain.

[0003] Existing research has shown that neural networks exhibit spectral bias during training, meaning they are more likely to learn low-frequency (LF) information in images but struggle to capture high-frequency (HF) details. This bias limits the model's performance in recognizing fine anatomical structures, complex texture boundaries, and weak-contrast regions. High-frequency components, however, carry a wealth of crucial information about tissue edges and structural textures, making them invaluable in biomedical image segmentation tasks.

[0004] Artificially decoupling images in the frequency domain and extracting their low-frequency and high-frequency components separately not only enhances the model's ability to represent structures at different scales but also provides a new approach to constructing more discriminative learning signals. Against this backdrop, an increasing number of studies are exploring the multi-frequency information representation of images using frequency analysis tools such as wavelet transform.

[0005] While fully supervised learning excels in segmentation performance, its heavy reliance on acquiring large-scale, high-quality labeled data presents a significant challenge in the field of medical imaging. The labeling process for medical images is not only time-consuming and labor-intensive but also requires experts with specialized medical knowledge, resulting in extremely high data labeling costs and limited scalability.

[0006] To overcome this bottleneck, semi-supervised learning (SSL) has attracted increasing attention. This method combines a small amount of labeled data with a large amount of unlabeled data, significantly reducing the dependence on labels while maintaining segmentation performance. Common semi-supervised strategies include self-training, consistency regularization, and contrastive learning. Among them, consistency regularization has achieved good results in multiple tasks due to its strong practicality and ease of implementation. However, current methods generally use artificially designed image perturbations (such as rotation, blurring, color jitter, etc.) to construct consistency signals. These perturbations may introduce unnatural morphological changes or even artifacts in medical images, destroying the original anatomical structure, thereby affecting model performance and introducing negative learning bias.

[0007] When performing biomedical image segmentation, most current methods focus on solving a specific problem in fully supervised or semi-supervised scenarios. Some existing methods, such as XNet, attempt to fuse information from different frequencies to improve performance and achieve optimal segmentation results. However, the feature-level fusion method it employs weakens the differences in the distribution of low-frequency and high-frequency features, but also weakens the consistency signal used for regularization. This limits the adaptability and accuracy of the method when handling biomedical image segmentation in this semi-supervised task. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of the prior art by providing a biomedical image segmentation method, system, device, and medium to solve the problems in the prior art.

[0009] This invention specifically provides the following technical solution: a biomedical image segmentation method, comprising:

[0010] Acquire biomedical images and decompose them into low-frequency (LF) and high-frequency (HF) components;

[0011] Based on the cross-branch pseudo-label guidance strategy, the low-frequency LF component and the high-frequency HF component are processed in parallel to generate the first pseudo-label of the low-frequency LF component and the second pseudo-label of the high-frequency HF component. The first pseudo-label is used to semantically guide the high-frequency HF component processing, and the second pseudo-label is used to supervise the low-frequency LF component processing, thereby generating the first frequency feature map and the second frequency feature map respectively.

[0012] The first and second frequency feature maps are subjected to quadrant frequency sensing. The spatial quadrants of all frequency feature maps are dynamically weighted and recalibrated in a region-specific manner. Based on the frequency sensing recalibration results, the frequency branch that performs best in the current task is selected as the final inference path, and the corresponding segmentation prediction results are generated with the inference path.

[0013] Preferably, a pre-trained dual-branch segmentation network model is used to segment the low-frequency (LF) component and the high-frequency (HF) component, generating corresponding segmentation prediction results. The pre-trained dual-branch segmentation network model includes:

[0014] Two parallel-connected frequency processing branches; each frequency processing branch is composed of a deep neural network, and the skip connections of the deep neural network are connected to quadrant frequency sensing units.

[0015] Preferably, the dynamic weight adjustment of the spatial quadrants of all frequency feature maps and the frequency-aware recalibration in a region-specific manner specifically involve:

[0016] Within each quadrant of the quadrant frequency sensing unit, two complementary mechanisms are integrated to perform frequency sensing recalibration on all frequency feature maps; the complementary mechanisms include: channel-frequency modulation (CFM) and spatial-frequency enhancement (SFE).

[0017] Channel-frequency modulation (CFM) recalibrates the characteristic responses on different channels; the specific expression is:

[0018] ;

[0019] in, and These represent adaptive max pooling and average pooling operations, respectively. Indicates bottleneck transformation. It is the Sigmoid function. As a variable;

[0020] Among them, Spatial-Frequency Enhancement (SFE) is based on frequency-saliency enhancement of pixel-level feature responses; the specific expression is:

[0021] ;

[0022] in, and These are the maximum and average values ​​in the channel dimension, respectively. It is a 3×3 spatial convolution.

[0023] Preferably, the specific expression for generating the first frequency feature map and the second frequency feature map is as follows:

[0024] ;

[0025] ;

[0026] in, This is the first frequency feature map. This is the second frequency feature map. and These are learnable scalar parameters used to adjust the contribution of complementary frequency components. cA This represents the approximate coefficient corresponding to the low-frequency LF component. cH , cV and cD These represent the horizontal, vertical, and diagonal detail coefficients, respectively.

[0027] Preferably, before segmenting the low-frequency LF component and the high-frequency HF component using the pre-trained dual-branch segmentation network model, the process further includes pre-training the dual-branch segmentation network model. The pre-training process specifically involves:

[0028] The bi-branch splitting network model is jointly optimized using supervised loss and consistency loss; where, under fully supervised or semi-supervised settings, the total loss function is... All are defined as follows:

[0029] ;

[0030] in, L sup Indicates monitoring losses, L cons This indicates a loss of consistency.

[0031] The supervision loss is defined as the sum of the Dice losses of each branch:

[0032] ;

[0033] in, and Let represent the segmentation prediction results of the low-frequency LF and high-frequency HF branches for the i-th sample, respectively. It is the corresponding target truth value;

[0034] Consistency loss is defined as cross-pseudo-supervision (CPS) loss:

[0035] ;

[0036] in, and They are from or Pseudo-tags generated in the middle.

[0037] This invention provides a biomedical image segmentation system, comprising:

[0038] The acquisition module is used to acquire biomedical images and decompose them into low-frequency (LF) and high-frequency (HF) components.

[0039] The model processing module is used to process the low-frequency LF component and the high-frequency HF component in parallel based on the cross-branch pseudo-label guidance strategy, generate the first pseudo-label of the low-frequency LF component and the second pseudo-label of the high-frequency HF component, and use the first pseudo-label to semantically guide the high-frequency HF component processing process, and use the second pseudo-label to supervise the low-frequency LF component processing process, generating the first frequency feature map and the second frequency feature map respectively.

[0040] The segmentation module is used to perform quadrant frequency sensing on the first frequency feature map and the second frequency feature map, dynamically adjust the spatial quadrants of all frequency feature maps, and perform frequency sensing recalibration in a region-specific manner. Based on the frequency sensing recalibration results, the frequency branch that performs best in the current task is selected as the final inference path, and the corresponding segmentation prediction results are generated with the inference path.

[0041] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a program, and when the program is executed by the processor, the processor performs the steps of the above-described biomedical image segmentation method.

[0042] The present invention provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described biomedical image segmentation method.

[0043] Compared with the prior art, the present invention has the following significant advantages:

[0044] This invention decomposes biomedical images into low-frequency (LF) and high-frequency (HF) components. Through parallel branching and cross-branch pseudo-label guidance strategies, the LF and HF components are enhanced to obtain frequency feature maps, promoting collaborative learning between the two branches at different frequency feature levels. This achieves the construction of a natural frequency consistency signal, effectively balancing the information distribution between the low-frequency and high-frequency branches while maintaining their respective characteristics. Simultaneously, quadrant frequency sensing is performed, dynamically adjusting the spatial quadrants of all frequency feature maps and recalibrating the frequency sensing in a region-specific manner. Based on the frequency sensing recalibration results, the frequency branch performing best in the current task is selected as the final inference path, generating corresponding segmentation prediction results. This invention achieves enhanced local frequency learning capabilities through a natural frequency consistency signal combined with a quadrant frequency sensing mechanism (QFAU), enabling collaborative modeling of low-frequency semantics and high-frequency structure in images, thereby enhancing the model's segmentation accuracy and generalization ability under fully supervised and semi-supervised conditions. Attached Figure Description

[0045] Figure 1 This is the overall architecture diagram provided by the present invention; wherein, Figure 1 (a) is an overall architecture diagram of a biomedical image segmentation method. Figure 1 (b) is a specific module of the quadrant frequency sensing unit (QFAU);

[0046] Figure 2 This is a flowchart of the Discrete Wavelet Transform (DWT) processing in this invention;

[0047] Figure 3 This is an overall flowchart of a biomedical image segmentation method according to the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below 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 should fall within the scope of protection of the present invention.

[0049] Based on the research, the following core challenges were found in the current biomedical image segmentation: 1. Insufficient modeling ability for high-frequency details; 2. Difficulty in obtaining a large amount of high-quality labeled data; 3. Existing consistency regularization strategies are not friendly to biomedical images; 4. Lack of a unified and efficient framework applicable to both fully supervised and semi-supervised learning.

[0050] Based on the aforementioned problems in the existing solutions Figure 1This paper presents the overall framework of the proposed method, outlining a wavelet-based equilibrium decoupling strategy that decomposes biomedical images into low-frequency (LF) and high-frequency (HF) components. The LF component primarily preserves the global structure and semantic contours of the image, while the HF component emphasizes local texture and edge details. Considering the complementarity of these two frequency components in perceptual representation, a bi-branch segmentation network model is designed, where the LF and HF images are respectively input into two parallel branches to generate corresponding segmentation predictions. To promote collaborative learning between the two branches at different frequency feature levels, a cross-branch pseudo-label guidance strategy is introduced to achieve consistency constraints: pseudo-labels generated by the LF branch guide the HF branch, helping it better capture semantic context while preserving detailed features; conversely, pseudo-labels generated by the HF branch supervise the LF branch, enabling it to enhance its perception of fine-grained texture while maintaining global consistency. This bi-directional interaction promotes effective cross-frequency collaboration, thereby constructing a more stable and efficient consistency framework. Furthermore, to enhance the model's frequency awareness in key spatial regions, a quadrant frequency awareness unit (QFAC) is introduced at skip connections. This unit dynamically adjusts the weights of the spatial quadrants of the feature map, guiding the network to focus on salient semantic or texture regions.

[0051] Biomedical images are segmented using a pre-trained two-branch segmentation network model to generate corresponding segmentation prediction results. The pre-trained two-branch segmentation network model includes:

[0052] The discrete wavelet transform unit is connected in sequence and two parallel-connected frequency processing branches are connected in parallel. Each frequency processing branch is composed of a deep neural network, and the skip connections of the deep neural network are connected to quadrant frequency sensing units.

[0053] Based on the above framework, such as Figure 3 As shown, this embodiment of a biomedical image segmentation method includes the following steps:

[0054] Step S1: Acquire biomedical images and decompose them into low-frequency (LF) components and high-frequency (HF) components.

[0055] Raw biomedical image data is acquired; the images are decomposed into frequency domain information using discrete wavelet transform (DWT).

[0056] Wavelet transform is a powerful signal processing technique that can losslessly decompose image signals into multiple frequency subbands. Given an input image I, the Discrete Wavelet Transform (DWT) decomposes it into four distinct subbands:

[0057] A cross-branch pseudo-tag guidance strategy was introduced, specifically:

[0058] (1);

[0059] Here, cA represents the approximation coefficient corresponding to the low-frequency (global semantic structure LF) component, while cH, cV, and cD represent the horizontal, vertical, and diagonal detail coefficients, respectively. These three components together constitute the high-frequency (fine-grained texture feature HF) component, achieving information decoupling. Since both the low-frequency and high-frequency components are decoupled from the original image, each may lack crucial information. Furthermore, the low-frequency component contains most of the information from the original image, while the high-frequency component often lacks sufficient semantic content, potentially making it difficult for the model to effectively learn meaningful high-frequency features.

[0060] Step S2: Based on the cross-branch pseudo-label guidance strategy, the low-frequency LF component and the high-frequency HF component are processed in parallel to generate the first pseudo-label of the low-frequency LF component and the second pseudo-label of the high-frequency HF component. The first pseudo-label is used to provide semantic guidance for the high-frequency HF component processing, and the second pseudo-label is used to supervise the low-frequency LF component processing, thereby generating the first frequency feature map and the second frequency feature map respectively.

[0061] A cross-branch pseudo-label guidance strategy is proposed: a bidirectional pseudo-label constraint mechanism is formed to enhance semantic and detail co-modeling. To address this information imbalance problem, a learnable compensation mechanism is proposed to achieve mutual enhancement of low-frequency and high-frequency components. The enhanced low-frequency and high-frequency images, i.e., the generated first and second frequency feature maps, are defined as follows:

[0062] (2);

[0063] (3);

[0064] in, This is the first frequency feature map. This is the second frequency feature map. and These are learnable scalar parameters used to adjust the contribution of complementary frequency components. cA This represents the approximate coefficient corresponding to the low-frequency LF component. cH , cV and cD These represent the horizontal, vertical, and diagonal detail coefficients, respectively.

[0065] This strategy effectively balances the information distribution between low-frequency and high-frequency branches while maintaining their respective characteristics. Furthermore, since these learnable parameters are continuously updated during training, this mechanism also acts as an implicit data augmentation technique in the frequency domain. Figure 2 The complete DWE processing flow is demonstrated.

[0066] IL focuses on global structure and semantic information, while IH emphasizes local edges and texture details. By leveraging the inherent differences between IL and IH, semi-supervised learning can be carried out based on consistency regularization, which helps the model to understand image features more comprehensively.

[0067] LF branch (low-frequency channel):

[0068] deal with cA The generated images focus on modeling the global semantic contours and structural features of the images; the generated pseudo-labels feed back into high-frequency branches, providing semantic guidance.

[0069] HF branch (high-frequency channel):

[0070] deal with cA The generated image focuses on modeling the local edge and texture details of the image; the generated pseudo-labels feed back into the low-frequency branches, enhancing the perception of fine-grained textures.

[0071] Step S3: Perform quadrant frequency sensing on the first and second frequency feature maps, dynamically adjust the spatial quadrants of all frequency feature maps, and perform frequency sensing recalibration in a region-specific manner. Based on the frequency sensing recalibration results, select the frequency branch that performs best in the current task as the final inference path, and generate the corresponding segmentation prediction results with the inference path.

[0072] To reduce computational redundancy and model complexity during inference, at the end of training, the higher-performing frequency branch (low-frequency branch LF or high-frequency branch HF) is selected as the final inference path, and the other branch is discarded.

[0073] Inspired by the frequency decomposition behavior of Discrete Wavelet Transform (DWT), a Quadrant Frequency Aware Unit (QFAU) is introduced to enhance frequency adaptive learning capabilities in the low-frequency (LF) and high-frequency (HF) branches. Biomedical images typically exhibit spatially varying frequency distributions, with different regions containing distinct semantic and textural features. To effectively capture local frequency cues and improve the model's sensitivity to frequency features in spatially local regions, the QFAU spatially divides each feature map into four quadrants, similar to the four sub-bands of wavelet decomposition, and performs frequency enhancement processing in each quadrant, enabling the model to perform frequency-sensitive modulation in a region-specific manner. Within each quadrant of the Quadrant Frequency Aware Unit, two complementary mechanisms are integrated for frequency-aware recalibration of all frequency feature maps; these complementary mechanisms include Channel-Frequency Modulation (CFM) and Spatial-Frequency Enhancement (SFE). These two mechanisms jointly facilitate frequency-aware recalibration along both the channel and spatial dimensions. Furthermore, a residual global modulation path is added to enhance the model's ability to perceive multi-scale frequency information. The structure of the QFAU is as follows: Figure 1 As shown in (b).

[0074] Channel frequency modulation (CFM) focuses on the channel dimension, enabling recalibration of characteristic responses across different channels. It selectively highlights informative frequency features in different channels while suppressing redundant or noise components. Max pooling is used. and average pooling Combined with bottleneck transformation and the Sigmoid function Frequency calibration is completed to improve the model's response to key channel frequencies. CFM can be expressed as:

[0075] (4);

[0076] in, and These represent adaptive max pooling and average pooling operations, respectively. Indicates bottleneck transformation. It is the Sigmoid function. For variables.

[0077] Spatial Frequency Enhancement (SFE) operates in the spatial dimension, enhancing pixel-level feature responses based on frequency saliency. It utilizes the maximum value of each channel. Compared with the average By using 3×3 convolution, local texture response is enhanced; this improves the model's ability to recognize spatial frequency features such as edges and details, thereby enhancing its ability to extract high-frequency information. SFE can be represented as:

[0078] (5);

[0079] in, and These are the maximum and average values ​​in the channel dimension, respectively. It is a 3×3 spatial convolution.

[0080] The two-branch segmentation network model is pre-trained. The pre-training process is as follows:

[0081] The proposed method is compatible with both fully supervised and semi-supervised scenarios. During training, the bi-branch segmentation network model is jointly optimized using supervised loss and consistency loss. In the semi-supervised setting, consistency loss effectively mines latent information in unlabeled data, thereby improving the model's generalization ability. In the fully supervised setting, consistency loss ensures that the model's prediction results remain stable across different frequency components, thus enhancing the model's robustness. Therefore, regardless of whether the setting is fully supervised or semi-supervised, the total loss function... All are defined as follows:

[0082] (6);

[0083] in, L sup Indicates monitoring losses, L cons This indicates a loss of consistency.

[0084] The supervision loss is defined as the sum of the Dice losses of each branch:

[0085] (7);

[0086] in, and Let represent the segmentation prediction results of the low-frequency LF and high-frequency HF branches for the i-th sample, respectively. It is the corresponding target truth value.

[0087] Consistency loss is defined as cross-pseudo-supervision (CPS) loss:

[0088] (8);

[0089] in, and They are from or Pseudo-tags generated in the middle.

[0090] Extensive experiments were conducted on colorectal gland segmentation datasets, neuronal structure segmentation datasets in electron microscopy stacks, and ultrasound neural segmentation datasets, demonstrating that the invention achieves state-of-the-art performance in both fully supervised and semi-supervised training.

[0091] This invention proposes a biomedical image segmentation system, comprising: an acquisition module, a model processing module, and a segmentation module.

[0092] The acquisition module acquires biomedical images and decomposes them into low-frequency (LF) and high-frequency (HF) components. The model processing module processes the LF and HF components in parallel based on a cross-branch pseudo-label guidance strategy, generating a first pseudo-label for the LF component and a second pseudo-label for the HF component. The first pseudo-label provides semantic guidance for the HF component processing, while the second pseudo-label supervises the LF component processing, generating a first frequency feature map and a second frequency feature map respectively. The segmentation module performs quadrant frequency sensing on the first and second frequency feature maps, dynamically adjusts the spatial quadrants of all frequency feature maps, and performs frequency sensing recalibration in a region-specific manner. Based on the frequency sensing recalibration results, the frequency branch that performs best in the current task is selected as the final inference path, and the corresponding segmentation prediction results are generated using the inference path.

[0093] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a program, and when the program is executed by the processor, the processor performs the steps of a biomedical image segmentation method.

[0094] According to the disclosed embodiments, the computer device can communicate with one or more external devices (e.g., keyboard, pointing device, Bluetooth communication, etc.) or with any device that enables the computing device to communicate with one or more other computing devices (e.g., router, demodulator, etc.).

[0095] The present invention also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of a biomedical image segmentation method.

[0096] According to the disclosed embodiments, the storage medium can be a non-volatile computer-readable storage medium, such as, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, the storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0097] The above description, in conjunction with specific preferred embodiments, provides a more detailed explanation of the present invention. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered to fall within the scope of protection of the present invention.

Claims

1. A biomedical image segmentation method, characterized in that, include: Acquire biomedical images and decompose them into low-frequency (LF) and high-frequency (HF) components; Based on the cross-branch pseudo-label guidance strategy, the low-frequency LF component and the high-frequency HF component are processed in parallel to generate the first pseudo-label of the low-frequency LF component and the second pseudo-label of the high-frequency HF component. The first pseudo-label is used to semantically guide the high-frequency HF component processing, and the second pseudo-label is used to supervise the low-frequency LF component processing, thereby generating the first frequency feature map and the second frequency feature map respectively. The first and second frequency feature maps are subjected to quadrant frequency sensing. The spatial quadrants of all frequency feature maps are dynamically weighted and recalibrated in a region-specific manner. Based on the frequency sensing recalibration results, the frequency branch that performs best in the current task is selected as the final inference path. The corresponding segmentation prediction results are generated using the final inference path in the deep neural network. Biomedical images are segmented using a pre-trained two-branch segmentation network model to generate corresponding segmentation prediction results. The pre-trained two-branch segmentation network model includes: The discrete wavelet transform unit is connected in sequence and two parallel-connected frequency processing branches are connected in parallel. Each frequency processing branch is composed of a deep neural network, and the jump connection of the deep neural network is connected to a quadrant frequency sensing unit. The dynamic weight adjustment of the spatial quadrants of all frequency feature maps and the frequency-aware recalibration in a region-specific manner are specifically as follows: Within each quadrant of the quadrant frequency sensing unit, two complementary mechanisms are integrated to perform frequency sensing recalibration on all frequency feature maps; the complementary mechanisms include: channel-frequency modulation (CFM) and spatial-frequency enhancement (SFE). Channel-frequency modulation (CFM) recalibrates the characteristic responses on different channels; the specific expression is: ； in, and These represent adaptive max pooling and average pooling operations, respectively. Indicates bottleneck transformation. It is the Sigmoid function. As a variable; Among them, Spatial-Frequency Enhancement (SFE) is based on frequency-saliency enhancement of pixel-level feature responses; the specific expression is: ； in, and These are the maximum and average values ​​in the channel dimension, respectively. It is a 3×3 spatial convolution.

2. The biomedical image segmentation method as described in claim 1, characterized in that, The specific expression for generating the first frequency feature map and the second frequency feature map is as follows: ； ； in, This is the first frequency feature map. This is the second frequency feature map. and These are learnable scalar parameters used to adjust the contribution of complementary frequency components. cA This represents the approximate coefficient corresponding to the low-frequency LF component. cH , cV and cD These represent the horizontal, vertical, and diagonal detail coefficients, respectively.

3. The biomedical image segmentation method as described in claim 1, characterized in that, Before segmenting the low-frequency (LF) and high-frequency (HF) components using the pre-trained dual-branch segmentation network model, the process includes pre-training the dual-branch segmentation network model. The pre-training process specifically involves: The bi-branch splitting network model is jointly optimized using supervised loss and consistency loss; where, under fully supervised or semi-supervised settings, the total loss function is... All are defined as follows: ； in, L sup Indicates monitoring losses, L cons This indicates a loss of consistency. The supervision loss is defined as the sum of the Dice losses of each branch: ； in, and Let represent the segmentation prediction results of the low-frequency LF and high-frequency HF branches for the i-th sample, respectively. It is the corresponding target truth value; Consistency loss is defined as cross-pseudo-supervision (CPS) loss: ； in, and They are from or Pseudo-tags generated in the middle.

4. A biomedical image segmentation system, characterized in that, include: The acquisition module is used to acquire biomedical images and decompose them into low-frequency (LF) and high-frequency (HF) components. The model processing module is used to process the low-frequency LF component and the high-frequency HF component in parallel based on the cross-branch pseudo-label guidance strategy, generate the first pseudo-label of the low-frequency LF component and the second pseudo-label of the high-frequency HF component, and use the first pseudo-label to semantically guide the high-frequency HF component processing process, and use the second pseudo-label to supervise the low-frequency LF component processing process, generating the first frequency feature map and the second frequency feature map respectively. The segmentation module is used to perform quadrant frequency sensing on the first frequency feature map and the second frequency feature map, dynamically adjust the spatial quadrants of all frequency feature maps, and perform frequency sensing recalibration in a region-specific manner. Based on the frequency sensing recalibration results, the frequency branch that performs best in the current task is selected as the final inference path, and the corresponding segmentation prediction results are generated with the inference path. Biomedical images are segmented using a pre-trained two-branch segmentation network model to generate corresponding segmentation prediction results. The pre-trained two-branch segmentation network model includes: The discrete wavelet transform unit is connected in sequence and two parallel-connected frequency processing branches are connected in parallel. Each frequency processing branch is composed of a deep neural network, and the jump connection of the deep neural network is connected to a quadrant frequency sensing unit. The dynamic weight adjustment of the spatial quadrants of all frequency feature maps and the frequency-aware recalibration in a region-specific manner are specifically as follows: Within each quadrant of the quadrant frequency sensing unit, two complementary mechanisms are integrated to perform frequency sensing recalibration on all frequency feature maps; the complementary mechanisms include: channel-frequency modulation (CFM) and spatial-frequency enhancement (SFE). Channel-frequency modulation (CFM) recalibrates the characteristic responses on different channels; the specific expression is: ； in, and These represent adaptive max pooling and average pooling operations, respectively. Indicates bottleneck transformation. It is the Sigmoid function. As a variable; Among them, Spatial-Frequency Enhancement (SFE) is based on frequency-saliency enhancement of pixel-level feature responses; the specific expression is: ； in, and These are the maximum and average values ​​in the channel dimension, respectively. It is a 3×3 spatial convolution.

5. A computer device, characterized in that, The device includes a memory and a processor, wherein the memory stores a program that, when executed by the processor, causes the processor to perform the steps of a biomedical image segmentation method as described in any one of claims 1 to 3.

6. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the biomedical image segmentation method according to any one of claims 1 to 3.

Images