Medical image segmentation model, method and equipment based on VRWKV

Abstract

The invention relates to the technical field of computer-aided diagnosis and treatment, and particularly discloses a VRWKV-based medical image segmentation model, method and equipment, the model adopts an encoder-bottleneck block-decoder U-shaped architecture, and a frequency sensing wavelet attention module is arranged on a jump connection path between an encoder and a decoder. The frequency sensing wavelet attention module decomposes feature representation in a frequency domain through wavelet transform, and performs adaptive modulation by using a VRWKV attention mechanism, so that high-frequency details and a low-frequency global structure can be better aligned; and a multi-scale channel fusion module is arranged between the decoder and the segmentation head, and the multi-scale channel fusion module performs cross-scale fusion on output features of all layers of the decoder by using a VRWKV channel mixing mechanism. According to the method, the optimal performance is obtained in various target segmentation tasks, and the method has higher robustness for complex scenes such as fuzzy boundaries, low contrast, noise interference and the like.

Description

Technical Field

[0001] This invention relates to the field of computer-aided diagnosis and treatment technology, and in particular to a medical image segmentation model, method and device based on VRWKV. Background Technology

[0002] In the field of computer-aided diagnosis and treatment, medical image segmentation is a fundamental and crucial task, its performance directly impacting the accuracy of clinical diagnosis and the efficiency of treatment planning. Currently, mainstream medical image segmentation techniques are primarily based on Convolutional Neural Networks (CNNs), Visual Transformers (ViTs), Mamba models, and hybrid architectures; however, these techniques all have significant limitations. CNN-based methods, represented by U-Net and its variants, possess strong local feature extraction capabilities. However, limited by the receptive field, they struggle to effectively model global contextual information and long-range dependencies, leading to issues such as blurred boundaries and lost details in segmentation tasks involving complex anatomical structures and lesions. ViT-based methods alleviate the challenge of global dependency modeling through self-attention mechanisms; however, the secondary computational complexity of these mechanisms results in significant computational costs and memory consumption when processing high-resolution medical images, severely restricting their deployment in clinical settings. Mamba-based methods employ state-space models to achieve long-range dependency modeling with linear computational complexity, improving computational efficiency. However, this typically comes at the cost of segmentation accuracy, making it difficult to achieve an ideal balance between efficiency and accuracy.

[0003] VRWKV, as an emerging and efficient modeling architecture, possesses linear computational complexity and strong long-distance dependency modeling capabilities, and has been initially applied in the field of medical image segmentation. For example, Med-URWKV: Pure RWKV With ImageNet Pre-training For Medical Image Segmentation discloses that pure RWKV models can be used for medical image segmentation. However, existing technologies have two major drawbacks: first, the use of a hybrid architecture of CNN and VRWKV cannot fully reuse the weight parameters of large-scale pre-trained VRWKV encoders; second, there is a lack of an effective fusion mechanism for frequency domain information and multi-scale features. VRWKV models are insufficient in terms of collaborative modeling of local details and global structure, and preservation of spatial continuity, making it difficult to meet the high requirements of boundary accuracy and structural integrity in medical image segmentation. Summary of the Invention

[0004] This invention aims to solve the aforementioned problems. To this end, this invention provides a VRWKV-based medical image segmentation model, method, and device. By reusing a pre-trained VRWKV encoder and integrating two core modules: a frequency-aware wavelet attention module (FAWA module) and a multi-scale channel fusion module (MSCF module), the former improves the accuracy of feature detail representation by introducing frequency domain information, while the latter enhances feature aggregation through multi-scale fusion. This achieves optimal performance in various target segmentation tasks and exhibits stronger robustness to complex scenes such as blurred boundaries, low contrast, and noise interference.

[0005] This invention provides a medical image segmentation model based on VRWKV, and the technical solution adopted is as follows: it includes a U-shaped architecture of encoder-bottleneck block-decoder, wherein the encoder is provided with multiple VRWKV blocks, the decoder is provided with multiple image block expansion and VRWKV blocks, and the encoder and decoder have skip connections between corresponding layers. A frequency-aware wavelet attention module is set on the path of the jump connection. The frequency-aware wavelet attention module decomposes the feature representation in the frequency domain through wavelet transform and uses VRWKV attention mechanism for adaptive modulation, so that the high-frequency details and the low-frequency global structure can be better aligned. A multi-scale channel fusion module is set between the decoder and the segmentation head. The multi-scale channel fusion module uses the channel mixing mechanism of VRWKV to perform cross-scale fusion of the output features of all layers of the decoder. The working process of the multi-scale channel fusion module is as follows: The decoded hierarchical features calculated at each layer of the decoder are scale-aligned; then, they are input into the channel mixing module of VRWKV for channel-dimensional interaction, and the interaction results are concatenated to obtain enhanced features; the scale-aligned decoded hierarchical features are concatenated along the channel dimension to obtain basic features; the basic features and enhanced features are added to obtain channel fusion features.

[0006] Furthermore, the working process of the frequency-aware wavelet attention module is as follows: The input features are subjected to wavelet transform and decomposed into low-frequency components, vertical high-frequency components, horizontal high-frequency components, and diagonal high-frequency components. Low-frequency components, vertical high-frequency components, horizontal high-frequency components, and diagonal high-frequency components are input into the spatial mixing module of VRWKV for frequency attention modulation. The output features of the spatial mixing module are used to generate a fused feature map through inverse wavelet transform; The fused feature map is added to the input feature to obtain the frequency-aware enhanced feature.

[0007] Furthermore, in the spatial mixing module, the low-frequency component, vertical high-frequency component, horizontal high-frequency component, and diagonal high-frequency component are used as receiving vectors, with the low-frequency component serving as both the key vector and the value vector.

[0008] Furthermore, in the channel mixing module, the decoded hierarchical features aligned to the minimum spatial resolution are used as the key vectors, and all the scale-aligned decoded hierarchical features are used as the receive vectors.

[0009] Furthermore, the decoding hierarchical features are scale-aligned through upsampling.

[0010] Furthermore, the encoder is a pre-trained VRWKV encoder.

[0011] Furthermore, the segmentation head includes image patch unpacking, upsampling, 3×3 convolutional blocks, and 1×1 convolutional layers.

[0012] This invention also provides a VRWKV-based medical image segmentation method, the technical solution of which is as follows: using the above-mentioned VRWKV-based medical image segmentation model, the method includes the following steps: S1: Acquire medical images; S2: Input the medical image into a multi-layer encoder to obtain features at multiple coding levels; Multiple encoded hierarchical features are passed through a frequency-aware wavelet attention module and then input into the corresponding layer of the decoder. The encoded hierarchical features output from the last layer of the encoder are abstracted through a bottleneck block and then input into the decoder. The decoder outputs multiple decoding layer features; S3: Multiple decoding layer features are input into the multi-scale channel fusion module to perform multi-scale channel interaction and obtain channel fusion features; S4: Input the channel fusion feature into the segmentation head to obtain the segmentation result.

[0013] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described VRWKV-based medical image segmentation method.

[0014] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: 1. Enhanced feature representation capability: The FAWA module effectively associates high-frequency details with low-frequency global structure through wavelet transform and frequency attention fusion, improving the accuracy of boundary segmentation; the MSCF module solves the representation problem caused by the scale variation of lesion morphology through adaptive multi-scale channel fusion. The synergistic effect of the two enables the model to achieve the best balance between detail preservation and structural integrity.

[0015] 2. Excellent architectural compatibility and scalability: This model adopts a U-shaped architecture design, which is fully compatible with existing VRWKV pre-trained weights and does not require structural modifications to the encoder, providing a flexible foundation for subsequent model upgrades and cross-modal transfer. At the same time, the linear computational complexity characteristics ensure that the model still has efficient computing power in high-resolution medical image processing, providing technical support for clinical applications.

[0016] 3. Significant transfer effect of pre-training: By reusing the large-scale pre-trained VRWKV encoder, this model significantly accelerates the model convergence speed and ultimately improves the best performance, fully verifying the transfer value of pre-trained weights in medical image segmentation tasks.

[0017] 4. Significantly improved segmentation accuracy: The model achieves an average Dice similarity coefficient (DSC) of 88.04% on five publicly available medical image segmentation datasets (ISIC2017, ISIC2018, GLAS, BUSI, and KvasirSEG), which is 1.84% higher than the current mainstream Swin-UMamba model. It achieves the best performance in various target segmentation tasks such as skin lesions, polyps, and glands, and has stronger robustness to complex scenes such as blurred boundaries, low contrast, and noise interference.

[0018] 5. Parameter efficiency optimization: This model achieves better segmentation performance using only 25.59M parameters, effectively resolving the contradiction between accuracy and parameter complexity in existing technologies, and is more suitable for clinical deployment.

[0019] Additional aspects and advantages 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

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a model structure diagram provided by the present invention.

[0022] Figure 2 This is a graph showing the qualitative results of the comparative experiment provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this 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 this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but should not be used to limit the scope of this invention.

[0024] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0025] The following is combined Figures 1 to 2 The present invention will be further described in detail below, including a VRWKV-based medical image segmentation model, method, and device: In this embodiment, as Figure 1 As shown, a medical image segmentation model based on VRWKV is provided, named Met-URWKV†, which includes: a U-shaped architecture of encoder-bottleneck block-decoder, in which multiple VRWKV blocks are set in the encoder, and multiple image block expansion and VRWKV blocks are set in the decoder, with skip connections between corresponding layers of the encoder and decoder. A frequency-aware wavelet attention module is set on the path of the jump connection. The frequency-aware wavelet attention module decomposes the feature representation in the frequency domain through wavelet transform and uses VRWKV attention mechanism for adaptive modulation, so that the high-frequency details and the low-frequency global structure can be better aligned. A multi-scale channel fusion module is provided between the decoder and the segmentation head. The multi-scale channel fusion module uses the channel mixing mechanism of VRWKV to perform cross-scale fusion of the output features of all layers of the decoder.

[0026] Overall workflow of the model: Medical Images After image patch embedding, the image is input into a pre-trained VRWKV encoder. The encoder has a multi-layer structure and extracts multi-scale features. The dimension is Where H is the height of the image, W is the width of the image, and 3 is the number of channels.

[0027] In this embodiment, the encoder adopts a four-layer structure, with each layer including three stacked VRWKV blocks. The multi-scale features output by the encoder include first-level coding features. Second-level coding features Third-level coding features and fourth-level coding features , These are the output features of the final stage of the encoder. With each layer of the encoder, the dimensionality of the encoded hierarchical features decreases, such as... The dimension is ,in, This represents the number of channels.

[0028] After further feature abstraction using the bottleneck block, the data is input into the decoder. In this embodiment, the bottleneck block includes stacked max pooling, two VRWKV blocks, and layer normalization.

[0029] The decoder also adopts a multi-layer structure. Corresponding to the encoder structure, in this embodiment, the decoder is also a four-layer structure, with each layer including image block unpacking and two VRWKV blocks.

[0030] In the skip connection path between the encoder and decoder, each encoding level feature is processed separately by the FAWA module to generate a feature representation with enhanced structural stability and richer detail information, which is then passed to the corresponding decoding block. The decoder output is represented in sequence and denoted as the first decoding level feature, the second decoding level feature, the third decoding level feature, and the fourth decoding level feature.

[0031] The four decoding-level features are input into the multi-scale channel fusion module for multi-scale channel interaction to obtain the channel fusion features. .

[0032] at last, It is passed to the segmentation head for further processing. In the segmentation head, After image patch expansion, upsampling is used to recover the original height and width of the feature map. Then, a 3×3 convolutional block, including convolutional layers, batch normalization (BN), and ReLU, is used to map the high-dimensional feature map back to its original dimensions. This stabilizes the feature distribution after channel concatenation and enhances feature fusion capabilities. Finally, a 1×1 convolutional layer generates the final segmentation result Y. The dimension of Y is... , where n is the number of categories.

[0033] The following section highlights the innovative features of this model: the frequency-aware wavelet attention module and the multi-scale channel fusion module.

[0034] (1) Frequency-Aware Wavelet Attention (FAWA) Module: Wavelet transform (WT) can decompose an image into multi-scale sub-bands: low-frequency components preserve global contours, while high-frequency components capture detailed information such as edges and textures. Compared with traditional methods such as Fourier transform, WT can generate low-frequency and high-frequency signals more efficiently, and it can locate image features in both the spatial and frequency domains. The inherent WKV attention mechanism of RWKV can also be applied to the frequency domain. Therefore, this embodiment proposes a frequency-aware wavelet attention module, the architecture of which is as follows: Figure 1 As shown in the diagram, this embodiment selects the low-frequency subband as the keys and values ​​in the WKV attention mechanism because it can preserve global structure and information with high stability, which is crucial for modeling long-range dependencies in medical images. In contrast, the high-frequency subband mainly captures local variations and edge cues, making it suitable for generating attention weights. This design allows high-frequency details to be selectively enhanced under low-frequency guidance, thereby improving boundary accuracy.

[0035] The working process of the FAWA module is as follows: First, the i-th encoding level features of the input... Perform Haar wavelet transform (WT) to decompose into low-frequency components. Vertical high frequency components Horizontal high-frequency components and diagonal high frequency components The spatial dimensions of each component are reduced to half of their original size. The decomposition process is as follows: in, This is a wavelet transform.

[0036] Then, frequency attention modulation is performed, using the four frequency components as the receiving vector, with the low-frequency component... The Spatial Mix module simultaneously serves as both key and value vector inputs to VRWKV. Figure 1 In the spatial blending module, Q-Shift is a four-way shift operation. This is the receive vector in spatial mixing. For value vectors in spatial mixing, Here, represents the key vector in spatial mixing, Bi-wkv is a bidirectional weighted key-value attention module, and Norm is normalization. The RWKV attention mechanism is used to fine-tune multi-frequency features. The received vector converts frequency information into a weight distribution, which is then used to filter low-frequency values. The modulation process is as follows: in, The output characteristics of the spatial mixing module, For spatial hybrid modules, It is the Sigmoid activation function. for The corresponding receive vector, To multiply point by point, for The corresponding weighted key-value attention matrix, To output the weight matrix, This is the index for the wavelet decomposition subband.

[0037] The output characteristics of the spatial mixing module include attention-adjusted low-frequency components. Vertical high-frequency components after attention adjustment Horizontal high-frequency components after attention adjustment and attention-adjusted diagonal high-frequency components The results obtained It matches the global structure and preserves high-frequency features, allowing high-frequency components to reference low-frequency global structure information and measure the importance of different frequency features to the overall spatial features.

[0038] Finally, the attention-adjusted feature map is processed using inverse wavelet transform (IWT). , , , Restore the original feature map size to generate a fused feature map. The frequency-aware enhancement features ultimately output by the FAWA module. To fuse feature maps With the original input The sum of, It preserves the basic information of the original features and enhances the stability of training. The relevant calculation formula is as follows: in, This is the inverse wavelet transform.

[0039] The FAWA module decomposes the feature representation in the frequency domain through wavelet transform and uses the VRWKV attention mechanism for adaptive modulation, which enables better alignment between high-frequency details and low-frequency global structure, especially in scenarios where the lesion size varies greatly and local features are prominent, resulting in better segmentation performance.

[0040] (2) Multi-Scale Channel Fusion (MSCF) Module: In medical imaging, the morphological scale of lesions exhibits significant intra-class variability, posing a challenge to comprehensive representation using single-scale features. Furthermore, medical images often contain substantial spatial redundancy, and due to the high localization of lesion regions, directly using global features may lead to significant information attenuation. To address these issues, this embodiment proposes a multi-scale channel fusion module, which adaptively learns scale weights for each instance of the input image. This mechanism enables the network to prioritize the most effective scale features for the current segmentation task while enhancing discriminative channel representation.

[0041] The MSCF module works as follows: Using the decoding layer features calculated at each layer of the decoder as input, in this embodiment, the input of the MSCF module includes the first decoding layer feature, the second decoding layer feature, the third decoding layer feature, and the fourth decoding layer feature.

[0042] The MSCF module performs scale alignment on the decoding hierarchy features to obtain scale-aligned decoding hierarchy features. In this embodiment, the scale-aligned decoding hierarchy features include... , , and Specifically, the features at the second, third, and fourth decoding levels are upsampled to a size equal to that of the features at the first decoding level. Similarly, the scale-aligned second-level decoding features are obtained. Third-level decoding features after scale alignment Fourth decoding level features after scale alignment This laid the foundation for subsequent cross-scale fusion.

[0043] Based on scale-aligned features, the MSCF module introduces the channel blending module from VRWKV to... As a reference anchor point, for the already aligned , , and Perform channel-level interaction, where, As a key vector , , and As the receiving vector. It has the smallest spatial resolution and carries the original global semantic information, then to Relying on their respective information The refined global content is weighted and selectively activated, which gives each layer of features a gating effect. The final output cross-scale interactive enhanced feature map can reflect the relative importance of each channel feature to the task at the current scale. Figure 1 In the channel mixing module, This is the receive vector in channel mixing. This represents the key vector in channel blending.

[0044] Finally, a residual fusion strategy is adopted to further optimize the feature representation. First, the scale-aligned decoded hierarchical features are concatenated along the channel dimension to form basic features that retain complete multi-scale information. Simultaneously, the cross-scale interactive enhancement feature maps are concatenated to obtain the enhanced features. Then, the basic features and enhanced features are combined by adding them element by element to prevent the loss of key information during feature fusion, thus obtaining the channel fusion features.

[0045] The working process can be represented by the following formula: in, The i-th decoded level feature after scale alignment For feature splicing operations, Basic features, To enhance features, For channel mixing modules, This is a channel fusion feature.

[0046] This embodiment combines the FAWA and MSCF modules. The two modules are complementary and indispensable to each other: the FAWA module mainly alleviates the mismatch between local details and global structural information, while the MSCF module enhances effective multi-scale feature fusion. Their combination ultimately brings about a gradual and consistent improvement in medical image segmentation performance.

[0047] This embodiment quantitatively compares the segmentation performance of our model with the current state-of-the-art methods on five publicly available medical image segmentation datasets (ISIC-2017, ISIC-2018, GLAS, BUSI, and Kvasir-SEG). The symbol "↓" indicates that a lower value indicates better performance. The evaluation metrics chosen are DSC (Dice similarity coefficient) and IoU (Intersection over Union). To ensure statistical reliability, all experiments were repeated three times, and the average results are reported. As shown in Table 1, Med-URWKV† achieves an average DSC of 88.04% on the five publicly available medical image segmentation datasets, a 1.84% improvement over the current mainstream Swin-Umamba model. It achieves state-of-the-art performance in various target segmentation tasks, including skin lesions, polyps, and glands, and demonstrates stronger robustness to complex scenes such as blurred boundaries, low contrast, and noise interference.

[0048] Table 1

[0049] In Table 1, Med-URWKV-Tiny and Med-URWKV-Small are sourced from the papers mentioned in the background section.

[0050] This embodiment also included ablation experiments on the pre-trained VRWKV encoder, with results shown in Table 2. Here, W / pre-training represents the VRWKV encoder initialized with pre-trained weights, and W / O pre-training represents the VRWKV encoder initialized without pre-training, with parameters randomly initialized and trained from scratch. Experiments were conducted on three datasets (BUSI, Kvasir-SEG, and GLAS), using DSC as the evaluation metric.

[0051] Table 2

[0052] This embodiment also included quantitative ablation analysis to evaluate the individual and combined efficacy of the FAWA and MSCF modules, as shown in Table 3. The symbol "-" indicates exclusion of the corresponding module, and the symbol "√" indicates inclusion of the corresponding module. Experiments were conducted on the BUSI and Kvasir-SEG datasets, using DSC and IoU as evaluation metrics.

[0053] Table 3

[0054] This embodiment also includes a qualitative analysis of the model and existing methods, the results of which are as follows: Figure 2 . Figure 2 (a) Figure 2(c) shows the analysis results on the datasets ISIC-2017, Kvasir-SEG, and GLAS, respectively.

[0055] This embodiment also provides a VRWKV-based medical image segmentation method, the technical solution of which is as follows: using the above-mentioned VRWKV-based medical image segmentation model, the method includes the following steps: S1: Acquire medical images; S2: Input the medical image into a multi-layer encoder to obtain features at multiple coding levels; Multiple encoded hierarchical features are passed through a frequency-aware wavelet attention module and then input into the corresponding layer of the decoder. The encoded hierarchical features output from the last layer of the encoder are abstracted through a bottleneck block and then input into the decoder. The decoder outputs multiple decoding layer features; S3: Multiple decoding layer features are input into the multi-scale channel fusion module to perform multi-scale channel interaction and obtain channel fusion features; S4: Input the channel fusion feature into the segmentation head to obtain the segmentation result.

[0056] This embodiment also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described VRWKV-based medical image segmentation method.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A medical image segmentation model based on VRWKV, comprising: A U-shaped architecture of encoder-bottleneck block-decoder is adopted. The encoder contains multiple VRWKV blocks, and the decoder contains multiple image block expansion and VRWKV blocks. The corresponding layers of the encoder and decoder are connected by skip connections. The feature is that a frequency-aware wavelet attention module is set on the path of the skip connection. The frequency-aware wavelet attention module decomposes the feature representation in the frequency domain through wavelet transform and uses VRWKV attention mechanism for adaptive modulation, so that the high-frequency details and the low-frequency global structure can be better aligned. A multi-scale channel fusion module is set between the decoder and the segmentation head. The multi-scale channel fusion module uses the channel mixing mechanism of VRWKV to perform cross-scale fusion of the output features of all layers of the decoder. The working process of the multi-scale channel fusion module is as follows: The decoded hierarchical features calculated at each layer of the decoder are scale-aligned; then, they are input into the channel mixing module of VRWKV for channel-dimensional interaction, and the interaction results are concatenated to obtain enhanced features; the scale-aligned decoded hierarchical features are concatenated along the channel dimension to obtain basic features; the basic features and enhanced features are added to obtain channel fusion features.

2. The medical image segmentation model based on VRWKV as described in claim 1, characterized in that, The working process of the frequency-aware wavelet attention module is as follows: The input features are subjected to wavelet transform and decomposed into low-frequency components, vertical high-frequency components, horizontal high-frequency components, and diagonal high-frequency components. Low-frequency components, vertical high-frequency components, horizontal high-frequency components, and diagonal high-frequency components are input into the spatial mixing module of VRWKV for frequency attention modulation. The output features of the spatial mixing module are used to generate a fused feature map through inverse wavelet transform; The fused feature map is added to the input feature to obtain the frequency-aware enhanced feature.

3. The medical image segmentation model based on VRWKV as described in claim 2, characterized in that, In the spatial mixing module, the low-frequency component, vertical high-frequency component, horizontal high-frequency component, and diagonal high-frequency component are used as receiving vectors, while the low-frequency component is used as both the key vector and the value vector.

4. The medical image segmentation model based on VRWKV as described in claim 1, characterized in that, In the channel mixing module, the decoded hierarchical features aligned to the minimum spatial resolution are used as the key vectors, and the decoded hierarchical features aligned to the full scale are used as the receive vectors.

5. A medical image segmentation model based on VRWKV as described in claim 1, characterized in that, Upsampling is used to scale-align the decoded hierarchical features.

6. A medical image segmentation model based on VRWKV as described in claim 1, characterized in that, The encoder is a pre-trained VRWKV encoder.

7. A medical image segmentation model based on VRWKV as described in claim 1, characterized in that, The segmentation head consists of image patch unpacking, upsampling, 3×3 convolutional blocks, and 1×1 convolutional layers.

8. A medical image segmentation method based on VRWKV, characterized in that, Using a VRWKV-based medical image segmentation model as described in any one of claims 1 to 7, the method includes the following steps: S1: Acquire medical images; S2: Input the medical image into a multi-layer encoder to obtain features at multiple coding levels; Multiple encoded hierarchical features are passed through a frequency-aware wavelet attention module and then input into the corresponding layer of the decoder. The encoded hierarchical features output from the last layer of the encoder are abstracted through a bottleneck block and then input into the decoder. The decoder outputs multiple decoding layer features; S3: Multiple decoding layer features are input into the multi-scale channel fusion module to perform multi-scale channel interaction and obtain channel fusion features; S4: Input the channel fusion feature into the segmentation head to obtain the segmentation result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the VRWKV-based medical image segmentation method of claim 8.

Images