A medical image segmentation method and system based on a channel-level frequency separation mechanism

Abstract

The application discloses a medical image segmentation method and system based on a channel-level frequency separation mechanism, and the method comprises the following steps: collecting a two-dimensional medical image containing a lesion area and performing data preprocessing to obtain a preprocessed medical image; introducing a local-global KAN-Mamba module, a double-scale fusion bottleneck layer and a KAN split channel attention bridging layer, adopting an encoder-decoder symmetric topological structure, and constructing a frequency decoupling type KAN-Mamba U-shaped segmentation network; and based on the frequency decoupling type KAN-Mamba U-shaped segmentation network, performing feature extraction and segmentation decoding processing on the preprocessed medical image to obtain a pixel-level medical image segmentation result. The application can accurately extract key lesion edge information to improve the medical image segmentation precision. The application can be widely applied to the technical field of medical image segmentation as a medical image segmentation method and system based on a channel-level frequency separation mechanism.

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation technology, and in particular to a medical image segmentation method and system based on a channel-level frequency separation mechanism. Background Technology

[0002] Rapid and accurate identification of lesion regions is crucial in medical image segmentation. For many years, the U-Net architecture, based on fully convolutional networks, has been the "gold standard" in this field due to its encoder-decoder structure and skip connection design. CNNs extract features layer by layer through local convolutional operations, effectively capturing high-frequency information such as texture and edges in images. However, CNNs are limited by their local receptive field, making it difficult to establish long-distance pixel dependencies, leading to discontinuous organs or misclassifications in complex anatomical environments. To overcome the limitations of CNNs, Vision Transformers (ViT) based on self-attention mechanisms have been introduced into medical image segmentation, such as TransUNet and Swin-UNet. ViT significantly improves the model's ability to model long-distance dependencies by calculating global pixel correlations. However, the computational complexity of self-attention mechanisms increases quadratically with image resolution. This places a huge computational burden and memory pressure on medical images with high resolution characteristics, limiting the expansion of model depth and deployment on clinical edge devices.

[0003] Against this backdrop, structured state-space models (SSMs), especially the Mamba architecture, have rapidly become a research hotspot as a sequence modeling paradigm with linear complexity (O(N)). Mamba, through its selective scanning mechanism, can achieve dual optimization of inference speed and memory efficiency while maintaining the global receptive field. Furthermore, the emergence of works such as U-Mamba and VM-UNet has demonstrated Mamba's enormous potential in processing three-dimensional (3D) volumetric data and long sequence slices.

[0004] While Mamba excels in global context modeling, recent in-depth theoretical analysis has revealed its inherent limitations in the frequency domain: Mamba exhibits a significant low-frequency bias. Due to the recursive properties of SSMs, which resemble low-pass filters, it tends to preferentially fit smooth trends in the data (i.e., low-frequency components), while being less adept at capturing high-frequency information such as abrupt edges and subtle textures. In medical images, the differentiation between benign and malignant lesions often relies on extremely minute edge burrs or internal texture heterogeneity. This characteristic of Mamba can lead to "overly smoothed" segmentation results, thus missing crucial diagnostic information.

[0005] Parallel to the development of Mamba, Kolmogorov-Arnold networks (KANs) have provided a new perspective on neural network design. Based on the Kolmogorov-Arnold representation theorem, KANs transfer the activation functions fixed on neurons in traditional multilayer perceptrons (MLPs) to the connection edges, employing learnable B-splines as activation functions. This design allows KANs to far surpass MLPs with the same number of parameters in function approximation capabilities and exhibits faster neural scaling. More importantly, research shows that KANs demonstrate stronger robustness than MLPs when fitting high-frequency functions and are less susceptible to low-frequency biases.

[0006] However, KAN is not a perfect solution. Since each edge requires the computation of a B-spline basis function, the computational cost and parameter count of standard KAN increase dramatically when processing high-dimensional image data, resulting in slow training speeds. Although works such as U-KAN have attempted to apply it to medical segmentation, simply using KAN as the backbone network often requires compromises in network width and depth, thus limiting the full realization of its potential. Summary of the Invention

[0007] To address the aforementioned technical problems, the present invention aims to provide a medical image segmentation method and system based on a channel-level frequency separation mechanism, which can accurately extract key lesion edge information, thereby improving the accuracy of medical image segmentation.

[0008] The first technical solution adopted in this invention is: a medical image segmentation method based on a channel-level frequency separation mechanism, comprising the following steps: Two-dimensional medical images containing lesion areas are acquired and data preprocessed to obtain preprocessed medical images; By introducing a local-global KAN-Mamba module, a dual-scale fusion bottleneck layer, and a KAN split channel attention bridging layer, and adopting an encoder-decoder symmetrical topology, a frequency-decoupled KAN-Mamba U-shaped segmentation network is constructed. Based on the frequency-decoupled KAN-Mamba U-shaped segmentation network, feature extraction and segmentation decoding are performed on the preprocessed medical images to obtain pixel-level medical image segmentation results.

[0009] Furthermore, the step of acquiring a two-dimensional medical image containing the lesion area and performing data preprocessing to obtain a preprocessed medical image specifically includes: Acquire two-dimensional medical images containing the lesion area and corresponding gold standard masks annotated by experts; All two-dimensional medical images and gold standard masks are uniformly adjusted to the preset spatial resolution to obtain the adjusted two-dimensional medical images and the adjusted gold standard mask; The pixel values ​​of the adjusted two-dimensional medical image are normalized and mapped to a preset value range to obtain a normalized two-dimensional medical image. Data augmentation operations involving random rotation and random flipping are performed on the normalized two-dimensional medical image and the adjusted gold standard mask to obtain the augmented medical image. The enhanced medical image is converted into a standard three-dimensional feature tensor to obtain a preprocessed medical image.

[0010] Furthermore, the frequency-decoupled KAN-Mamba U-shaped segmentation network specifically includes a feature encoding path, cross-scale skip connections, a dual-scale fusion bottleneck layer, and a decoding recovery path, wherein: The feature encoding path adopts a multi-level hierarchical processing design. Each level consists of a block merging downsampling layer and a local-global KAN-Mamba module cascaded together, which is used to extract multi-scale high and low frequency features step by step. The decoding recovery path is symmetrical to the encoder structure and consists of a block extended upsampling layer and a local-global KAN-Mamba module, which is used to recover spatial resolution step by step and fuse shallow features. The dual-scale fusion bottleneck layer receives encoder end features and performs cross-spectrum attention fusion to achieve deep fusion of high and low frequency information. The KAN split-channel attention bridging layer is embedded at each cross-scale jump connection location to perform dynamic channel recalibration on shallow features on the encoder side.

[0011] Furthermore, the step of performing feature extraction and segmentation decoding on the preprocessed medical image based on the frequency-decoupled KAN-Mamba U-shaped segmentation network to obtain pixel-level medical image segmentation results specifically includes: Based on the frequency-decoupled KAN-Mamba U-shaped segmentation network, the preprocessed medical image is input into the starting block embedding layer and mapped to the multi-scale feature space to obtain the mapped medical image. The mapped medical image is then passed through a four-level coding path for downsampling and a local-global KAN-Mamba module to extract high- and low-frequency features at different scales. High and low frequency features at different scales are input into the dual-scale fusion bottleneck layer to perform cross-spectral deep fusion and obtain the final fused features of high and low frequency information. The final high and low frequency information fusion features are sequentially processed by upsampling and local-global KAN-Mamba modules in a four-level decoding path. At each level, the recalibrated features of the corresponding encoder level are fused through the KAN split channel attention bridging layer to obtain the recalibrated features. The recalibrated features are restored to the original image resolution through a final block expansion upsampling layer, and the feature channels are mapped to the number of segmentation categories through a pointwise convolutional layer to obtain pixel-level medical image segmentation results.

[0012] Furthermore, the data processing flow of the local-global KAN-Mamba module specifically includes: Based on the preset channel splitting ratio, the preprocessed medical image is divided into a high-frequency feature subset and a low-frequency feature subset; High-frequency enhanced features are obtained by performing edge detail enhancement processing on a subset of high-frequency features based on high-frequency streams. Based on the low-frequency flow, a global structure modeling process is performed on the low-frequency feature subset to obtain low-frequency global features; High-frequency enhanced features and low-frequency global features are concatenated along the channel dimension, and inter-channel information interaction and mixing are performed through convolutional layers to obtain fused features; The fused features are added to the global residual of the preprocessed medical image to obtain medical image features with high and low frequency information.

[0013] Furthermore, the step of performing edge detail enhancement processing on the high-frequency feature subset based on the high-frequency stream to obtain high-frequency enhanced features specifically includes: The high-frequency feature subset is input into the average pooling layer for smooth downsampling, and then bilinear interpolation is used to restore it to the original resolution to extract the slowly changing background smoothing component. The high-frequency feature subset is subtracted element-wise from the background smoothing component to obtain the high-frequency detail residual field; The high-frequency detail residual field is enhanced by nonlinear mapping based on the KAN two-dimensional convolution operator to obtain the enhanced high-frequency detail residual field. The KAN two-dimensional convolution is composed of the superposition of linear basis paths and high-order spline paths. The enhanced high-frequency detail residual field is processed sequentially through a batch normalization layer and then added to the residual of the high-frequency feature subset to obtain the high-frequency enhanced features.

[0014] Furthermore, the step of performing global structural modeling on the low-frequency feature subset based on the low-frequency flow to obtain low-frequency global features specifically includes: The low-frequency feature subset is spatially downsampled through an average pooling layer to obtain the compressed low-frequency feature tensor; The compressed low-frequency feature tensor is input into the visual state space block, and a two-dimensional selective scanning operation is performed to obtain the scanned low-frequency feature tensor. The scanned low-frequency feature tensor is restored to its original physical resolution through bilinear interpolation, resulting in the low-frequency feature tensor with restored resolution. The low-frequency feature tensor after resolution restoration is processed sequentially through pointwise convolutional layers and batch normalization layers to obtain low-frequency global features.

[0015] Furthermore, the data processing flow of the dual-scale fusion bottleneck layer specifically includes: Medical image features with high and low frequency information are used as input features and input to the spatial flow branch and the feature flow branch respectively. The visual state spatial block based on spatial flow branch performs a global scan of the input features and, combined with layer normalization, extracts the global structure and long-range dependencies of the lesions, and outputs spatial query basis vectors. The KAN linear operator based on feature flow branching, combined with layer normalization, maps the input features and extracts fine semantic attributes, outputting a feature dictionary vector. The spatial query basis vector is passed through a linear layer to generate a query vector, and the feature dictionary vector is passed through the corresponding linear layers to generate key vectors and value vectors, respectively. The query vector, key vector, and value vector are flattened into a two-dimensional sequence matrix in the spatial dimension, and cross-attention weighting operation is performed to obtain the attention-weighted features; The attention-weighted features are reconstructed in spatial dimension and then regularized using dropout to obtain hybrid fusion features; The hybrid fusion features are added to the query vector by residual addition, and then subjected to layer normalization to obtain the final fusion features of high and low frequency information.

[0016] Furthermore, the dynamic channel recalibration step of the KAN split channel attention bridging layer specifically includes: The shallow features output from the corresponding layer of the encoder are received as input to the bridging layer; The input features are divided into multiple channel groups through a channel splitting operation; Attention weights are predicted for each channel group based on KAN layers or fully connected layers, generating dynamic weighting coefficients for each channel. The shallow features of the output are recalibrated channel by channel based on the dynamic weighting coefficients, and adaptive alignment of high-frequency texture and low-frequency semantics is performed to obtain the recalibrated features.

[0017] The second technical solution adopted in this invention is: a medical image segmentation system based on a channel-level frequency separation mechanism, comprising: The first module is used to acquire two-dimensional medical images containing lesion areas and perform data preprocessing to obtain preprocessed medical images. The second module is used to introduce a local-global KAN-Mamba module, a dual-scale fusion bottleneck layer and a KAN split channel attention bridging layer. It adopts an encoder-decoder symmetrical topology to construct a frequency-decoupled KAN-Mamba U-shaped segmentation network. The third module is used to perform feature extraction and segmentation decoding on the preprocessed medical images based on the frequency-decoupled KAN-Mamba U-shaped segmentation network to obtain pixel-level medical image segmentation results.

[0018] The beneficial effects of the method and system of this invention are as follows: This invention acquires two-dimensional medical images containing lesion regions and performs data preprocessing to obtain preprocessed medical images; further, it introduces a local-global KAN-Mamba module, a dual-scale fusion bottleneck layer, and a KAN split channel attention bridging layer, adopts an encoder-decoder symmetrical topology, and constructs a frequency-decoupled KAN-Mamba U-shaped segmentation network. Through the channel separation mechanism, it guides the decoupling of feature flows. The low-frequency flow is processed using downsampling Mamba, which significantly reduces the amount of computation while capturing global anatomical structures; the high-frequency flow is processed by optimized and efficient KAN convolution, focusing on nonlinear enhancement of edge details; finally, based on the frequency-decoupled KAN-Mamba U-shaped segmentation network, feature extraction and segmentation decoding are performed on the preprocessed medical images to obtain pixel-level medical image segmentation results, which can accurately extract key lesion edge information and thus improve the accuracy of medical image segmentation. Attached Figure Description

[0019] Figure 1 This is a flowchart of the steps of a medical image segmentation method based on a channel-level frequency separation mechanism according to the present invention; Figure 2 This is a structural block diagram of a medical image segmentation system based on a channel-level frequency separation mechanism according to the present invention. Figure 3 This is a schematic diagram of the overall architecture of the FKM-UNet model provided in a specific embodiment of the present invention; Figure 4 This is a detailed structural diagram of the LKMBlock provided in a specific embodiment of the present invention; Figure 5 This is a schematic diagram of the specific calculation process of the DSFusion bottleneck layer provided in a specific embodiment of the present invention; Figure 6 These are schematic diagrams illustrating the segmentation results of the CVC-ClinicDB dataset using different methods provided in specific embodiments of the present invention. Figure 7 This is a schematic diagram illustrating the segmentation results of the ISIC-2018 dataset using different methods provided in specific embodiments of the present invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0021] First, it's important to note that medical image segmentation is crucial for identifying lesion regions and aiding clinical diagnosis. Convolutional Neural Networks (CNNs) and Transformer models have achieved significant success in this field. In particular, recent Structured State Space Models (SSMs) (such as Mamba) have addressed the efficiency issues of long sequence modeling. However, research indicates they suffer from significant low-frequency spectral bias, often exhibiting a smoothing effect when processing high-frequency boundary details, thus losing crucial lesion edge information. To address this issue, this embodiment combines the Kolmogorov-Arnold Network (KAN), which excels at high-frequency nonlinear fitting, with Mamba, which excels at low-frequency global modeling, to propose a novel medical image segmentation model based on a channel-level frequency separation mechanism—FKM-UNet (Fusion KAN-Mamba U-Net). This model achieves independent processing of structure and details through an explicit frequency decoupling strategy. First, a Local-Global KAN-Mamba Block (LKM-Block) is proposed, which decouples features into two independent feature streams along the channel dimension using a dynamic channel splitting mechanism. One stream is processed by a downsampling Mamba module to capture low-frequency global anatomical skeletons, while the other stream is processed by efficient KAN convolution to enhance high-frequency nonlinear representations of edge textures. Furthermore, a Cross-Spectrum Attention Fusion (CSAF) mechanism is designed, using the spatial location prior extracted by Mamba as a query to dynamically weight and filter high-dimensional semantic features extracted by KAN. Finally, the model is evaluated on three publicly available datasets. Experimental results show that FKM-UNet outperforms other models in segmentation tasks and is more accurate in edge detail processing.

[0022] This embodiment designs a local-global dual-stream architecture (LKM Block) based on dynamic channel splitting. This module guides the decoupling of feature streams through a channel separation mechanism. The low-frequency stream utilizes downsampling Mamba processing, significantly reducing computational cost while capturing global anatomical structures; the high-frequency stream, on the other hand, focuses on nonlinear enhancement of edge details through optimized and efficient KAN convolution processing. This design implicitly achieves the separation and enhancement of frequency domain features by leveraging the architectural inductive bias of the two operators.

[0023] In the bottleneck layer of the U-shaped network, this embodiment proposes a fusion strategy based on cross-attention. This strategy utilizes the global spatial prior extracted by Mamba as a query to retrieve and weight the high-dimensional semantic features (key / value) extracted by KAN. Finally, this embodiment compares the proposed model with multiple models on three publicly available datasets, demonstrating its superior performance both quantitatively and qualitatively.

[0024] Furthermore, it's worth noting that since the introduction of U-Net, medical image segmentation architectures have evolved from pure CNNs to hybrid CNN-Transformer architectures. TransUNet and Swin-UNet addressed long-distance dependencies by introducing self-attention mechanisms, but their high computational cost limited their application on high-resolution 3D data. In recent years, to balance efficiency and performance, researchers have begun exploring more lightweight global modeling schemes. Mamba, as a prominent representative of SSM, has been rapidly adapted to the medical imaging field due to its linear complexity and hardware affinity. U-Mamba was the first to integrate Mamba modules into the encoder of U-Ne; VM-UNet constructed a U-shaped structure based on pure Mamba. However, most of these works directly use Mamba as a general feature extractor, neglecting its low-pass filtering characteristics in the frequency domain, resulting in poor performance when processing high-frequency details such as retinal vessels or irregular tumor edges.

[0025] The introduction of KAN disrupted the dominance of MLP. In the field of medical imaging, U-KAN demonstrated the feasibility of using KAN for segmentation, proving its advantage in parameter efficiency. However, standard KAN suffers from computational bottlenecks and struggles to handle complex anatomical backgrounds without the aid of a global receptive field. Wav-KAN enhances the time-frequency localization capability of KAN by introducing wavelet transform, but its structure is relatively complex and difficult to directly embed into existing deep networks. The network proposed in this embodiment does not attempt to replace all layers with KAN, but rather uses it as a dedicated processor for high-frequency residuals. This design approach maximizes the nonlinear fitting advantage of KAN while ensuring efficiency.

[0026] Frequency learning has been widely applied in computer vision. Octave convolution reduces spatial redundancy by decomposing feature maps into high- and low-frequency groups. In the SSM (Single-Speed ​​Modeling) domain, FreqMamba and ZigMa attempted to apply state-space models in the frequency domain to enhance rain removal or restoration tasks. In medical imaging, LKM-UNet proposed the concept of large-kernel Mamba, but it focuses on expanding the receptive field rather than explicit frequency separation. Unlike the above works, the model in this embodiment utilizes the Laplacian pyramid principle to construct an end-to-end frequency decoupling-fusion mechanism, specifically designed to address the low-frequency bias problem of Mamba.

[0027] Reference Figure 1 This invention provides a medical image segmentation method based on a channel-level frequency separation mechanism, which includes the following steps: S100: Acquire two-dimensional medical images containing lesion areas and perform data preprocessing to obtain preprocessed medical images; Specifically, a two-dimensional medical image containing the lesion region and a corresponding gold standard mask annotated by experts are acquired; all two-dimensional medical images and gold standard masks are uniformly adjusted to a preset spatial resolution to obtain adjusted two-dimensional medical images and adjusted gold standard masks; the pixel values ​​of the adjusted two-dimensional medical images are normalized and mapped to a preset numerical range to obtain normalized two-dimensional medical images; random rotation and random flipping data augmentation operations are performed on the normalized two-dimensional medical images and the adjusted gold standard masks to obtain enhanced medical images; the enhanced medical images are converted into standard three-dimensional feature tensors to obtain preprocessed medical images.

[0028] In this embodiment, rigorous data acquisition and preprocessing are required first. This embodiment acquires the original two-dimensional medical image containing the lesion and its corresponding pixel-level expert-annotated mask (Ground Truth, GT). To enable the network to extract features stably and efficiently, the original image must be standardized. Specifically, all input medical images and masks are first uniformly adjusted to a specific spatial resolution (e.g., 256×256 pixels) to meet the physical requirements of deep networks for fixed-size input tensors. Subsequently, the pixel values ​​of the image are normalized and mapped to the [0, 1] interval. This not only accelerates the gradient descent process of the network but also mitigates the impact of uneven lighting in medical device imaging. Furthermore, to enhance the model's generalization ability and prevent overfitting, data augmentation operations such as random rotation and flipping are introduced. After preprocessing, the image is converted into a standard three-dimensional feature tensor. (in , and (representing height, width, and number of channels respectively), which are then formally input into FKM-UNet for feature extraction, segmentation, and decoding.

[0029] In this embodiment, as Figure 3 As shown, FKM-UNet follows a symmetrical U-shaped encoder-decoder topology, mainly composed of four parts: feature encoding path, cross-scale feature skip connection, dual-scale fusion bottleneck layer (DSFusion Bottleneck), and decoding recovery path. The internal structure and interaction relationships of each module are as follows: The encoder employs a multi-level hierarchical processing design, with each level consisting of a Patch Merging downsampling layer cascaded with the core LKMBlock. The LKMBlock decouples features through a Dynamic Channel Split mechanism, processing them in parallel via a KAN high-frequency stream (enhancing edge nonlinear fitting) and a Mamba low-frequency stream (performing two-dimensional selective scanning to capture the global anatomical skeleton). Finally, residual mixing is achieved through 1×1 convolution for output. The decoder, symmetrical to the encoder structure, consists of a patch expanding upsampling layer and an LKMBlock. Each stage of the decoder receives high-dimensional semantic features from the upper layers and fuses the shallow spatial features of the corresponding encoder layers through skip connections. To alleviate the semantic gap between cross-scale features in skip connections, a KAN-SCA Bridge (KAN-Split Channel Attention) is embedded at the connection point. This submodule uses KAN or fully connected layers to dynamically recalibrate the features, ensuring effective alignment between high-frequency textures and low-frequency semantics.

[0030] Dual-Scale Fusion Bottleneck: This module is located at the deepest part of the network and is the core for achieving deep fusion of high- and low-frequency information. It receives features from the end of the encoder and uses the Cross-Spectrum Attention Fusion (CSAF) mechanism to use the low-frequency global features output by the Mamba path as a query to perform weighted filtering on the high-frequency detail features of the KAN path, thereby achieving adaptive dynamic fusion of semantics and details.

[0031] Information flow logic: After image input, it is mapped to a multi-scale feature space via initial Patch Embedding; then features are extracted through four-level encoding, and semantic refinement is performed by DSFusion at the bottleneck; finally, the decoder upsamples step by step and the residuals are fused through KAN-SCA Bridge, and finally pixel-level segmentation results are output through Final Patch Expanding and Final Conv.

[0032] S200 introduces a local-global KAN-Mamba module, a dual-scale fusion bottleneck layer and a KAN split channel attention bridging layer, and adopts an encoder-decoder symmetrical topology to construct a frequency-decoupled KAN-Mamba U-shaped segmentation network. Specifically, the frequency-decoupled KAN-Mamba U-shaped segmentation network includes a feature encoding path, cross-scale skip connections, a dual-scale fusion bottleneck layer, and a decoding recovery path. The feature encoding path employs a multi-level hierarchical design, with each level consisting of a block merging downsampling layer cascaded with a local-global KAN-Mamba module, used to extract multi-scale high- and low-frequency features step-by-step. The decoding recovery path is symmetrical to the encoder structure, consisting of a block expanding upsampling layer and a local-global KAN-Mamba module, used to restore spatial resolution step-by-step and fuse shallow features. The dual-scale fusion bottleneck layer receives features from the encoder end and performs cross-spectral attention fusion to achieve deep fusion of high- and low-frequency information. The KAN split channel attention bridging layer is embedded at each cross-scale skip connection location to dynamically recalibrate the shallow features on the encoder side.

[0033] The S300, based on the frequency-decoupled KAN-Mamba U-shaped segmentation network, performs feature extraction and segmentation decoding on preprocessed medical images to obtain pixel-level medical image segmentation results.

[0034] S310, based on the frequency-decoupled KAN-Mamba U-shaped segmentation network, inputs the preprocessed medical image into the starting block embedding layer, maps it to the multi-scale feature space, and obtains the mapped medical image.

[0035] S320. The mapped medical image is sequentially processed through a four-level coding path downsampling and a local-global KAN-Mamba module to extract high- and low-frequency features at different scales. Specifically, based on a preset channel splitting ratio, the preprocessed medical image is divided into a high-frequency feature subset and a low-frequency feature subset. Edge detail enhancement processing is performed on the high-frequency feature subset based on the high-frequency stream to obtain high-frequency enhanced features. Global structure modeling processing is performed on the low-frequency feature subset based on the low-frequency stream to obtain low-frequency global features. The high-frequency enhanced features and low-frequency global features are concatenated along the channel dimension, and inter-channel information interaction and mixing are performed through a convolutional layer to obtain fused features. The fused features are then added to the global residual of the preprocessed medical image to obtain medical image features with both high and low-frequency information.

[0036] For high-frequency enhancement features, the high-frequency feature subset is first input into an average pooling layer for smooth downsampling, and then bilinear interpolation is used to restore it to the original resolution to extract the slowly changing background smoothing component. Then, the high-frequency feature subset is subtracted from the background smoothing component element by element to obtain the high-frequency detail residual field. Further, the high-frequency detail residual field is enhanced by nonlinear mapping based on the KAN two-dimensional convolution operator to obtain the enhanced high-frequency detail residual field. The KAN two-dimensional convolution is composed of the superposition of linear basis paths and high-order spline paths. Finally, the enhanced high-frequency detail residual field is processed sequentially through a batch normalization layer and added to the high-frequency feature subset for residual processing to obtain the high-frequency enhancement features.

[0037] For low-frequency global features, the low-frequency feature subset is first spatially downsampled through an average pooling layer to obtain a compressed low-frequency feature tensor. Then, the compressed low-frequency feature tensor is input into a visual state space block, and a two-dimensional selective scanning operation is performed to obtain a scanned low-frequency feature tensor. Further, the scanned low-frequency feature tensor is restored to its original physical resolution through bilinear interpolation to obtain a low-frequency feature tensor with restored resolution. Finally, the low-frequency feature tensor with restored resolution is processed sequentially through a pointwise convolutional layer and a batch normalization layer to obtain the low-frequency global features.

[0038] In this embodiment, as Figure 4 As shown, to address the challenge of feature extraction from macroscopic anatomical displacement (low frequency) and microscopic infiltration margins (high frequency) of lesion tissue, this embodiment designs a local-global KAN-Mamba module (LKMBlock). This module abandons the traditional single feature extraction paradigm and explicitly adopts a parallel two-stream architecture, including a channel splitting mechanism, a high-frequency stream, a low-frequency stream, and a feature fusion module.

[0039] First, for the input feature tensor LKMBlock uses a channel splitting mechanism (ChannelSplit) to achieve a preset ratio. It is divided into two parts: high-frequency feature subset and low-frequency feature subset These two subsets are then processed in separate feature streams.

[0040] In high-frequency streaming, the core objective is to extract and enhance microscopic edge details. To obtain pure high-frequency residuals, First, the slowly changing background smoothing component is extracted using average pooling with a kernel size of 4 × 4 and a stride of 4, followed by bilinear interpolation. Then, the original... Subtracting from the smoothing component reveals the high-frequency detail residual field. Subsequently, efficient KAN 2D convolution (KANConv2d) was used to... Nonlinear enhancement is performed. The feature map is formed by superimposing linear basis paths and higher-order spline paths: ; In the formula, This represents a standard linear convolution mapping with a kernel size of 1 × 1. The feature tensor elements of a higher-order spline path are represented as: ; The enhanced features are then subjected to two-dimensional batch normalization (BatchNorm2d) and internal residual connections (with the original features). (Add together) to finally output the high-frequency characteristic state. .

[0041] In low-frequency streams, the core objective is to capture the global anatomy and long-range dependencies. To reduce computational complexity, First, spatial downsampling is performed using an average pooling layer with a stride of 2. The compressed tensor is then fed into the Visual State Space Block (VSSBlock) to perform a 2D Selective Scan (SS2D), modeling the global spatial context in linear complexity. After the scan, the features are restored to full physical resolution through bilinear interpolation and then passed through a 1×1 convolution and batch normalization layer to form a low-frequency stream output. .

[0042] In the Fusion & Residual module, the network will and The channels are concatenated (Concat), followed by inter-channel information exchange and feature fusion through 1×1 convolution, batch normalization (BN), and SiLU activation. Finally, the fused features are compared with the initial input of the module. The global residuals are summed to output a final feature that combines high- and low-frequency information.

[0043] S330. Input high- and low-frequency features of different scales into the dual-scale fusion bottleneck layer, perform cross-spectrum deep fusion, and obtain the final fused features of high- and low-frequency information. Specifically, medical image features with high and low frequency information are input features to the spatial flow branch and the feature flow branch, respectively. The visual state spatial block based on the spatial flow branch performs a global scan of the input features, and combined with layer normalization, extracts the global structure and long-range dependencies of lesions, outputting a spatial query basis vector. The KAN linear operator based on the feature flow branch, combined with layer normalization, maps the input features and extracts fine semantic attributes, outputting a feature dictionary vector. The spatial query basis vector is passed through a linear layer to generate a query vector, and the feature dictionary vector is passed through corresponding linear layers to generate key and value vectors, respectively. The query vector, key vector, and value vector are flattened into a two-dimensional sequence matrix in the spatial dimension, and cross-attention weighting is performed to obtain attention-weighted features. The attention-weighted features are reconstructed in the spatial dimension and subjected to dropout regularization to obtain hybrid fusion features. The hybrid fusion features are added to the query vector by residual addition, and then subjected to layer normalization to obtain the final fusion features of high and low frequency information.

[0044] In this embodiment, as Figure 5 As shown, at the deepest point of the U-shaped network (i.e., the bottleneck layer), the spatial resolution of the feature tensor reaches its minimum, and the microscopic image topology is transformed into a highly abstract pathological semantic representation. To avoid semantic mixing and mutual cancellation between features of different frequency bands, this embodiment pioneers a dual-scale fusion bottleneck layer (DSFusion). This module comprises three main sub-components: spatial flow, feature flow, and cross-modal attention fusion. The tensor input to the bottleneck layer... Implicitly split into two branches, the Spatial Stream is fed into a bidirectional independent modeling branch. The Spatial Stream performs a global scan of the input features using VSSBlock, combined with LayerNorm, aiming to extract the global structure and long-range dependencies of lesions, and outputs a spatial query basis vector as a "localization reference." The Feature Stream employs a continuous KAN linear operator combined with LayerNorm for mapping. Its aim is to extract fine-grained semantic attributes within a highly compressed abstract space using the non-linear capabilities of KAN, outputting a feature dictionary. .

[0045] Subsequently, the module enters the Cross-Modal Attention Fusion stage. The algorithm constructs an asymmetric attention mechanism, generating a query vector Q from the spatial flow through a linear layer, and generating a key vector K and a value vector V from the feature flow through corresponding linear layers. Its expression is as follows: ; ; ; Before calculating cross attention, the three-dimensional feature tensor is first flattened into a two-dimensional sequence matrix in spatial dimension. This allows for strict matrix dot product and transpose operations: ; After weighted operations using the attention matrix and followed by spatial reshaping and Dropout regularization, hybrid features are obtained. Ultimately, this is compared with the original spatial flow prior. Apply identity residual jumpers and perform layer normalization (LayerNorm) to output the final fused features with high-order semantic alignment. ;

[0046] S340. The final high and low frequency information fusion features are sequentially processed by upsampling and local-global KAN-Mamba modules in the four-level decoding path. At each level, the recalibrated features of the corresponding encoder level are fused through the KAN split channel attention bridging layer to obtain the recalibrated features. Specifically, the shallow features output from the corresponding layer of the encoder are received as the input to the bridging layer; the input features are divided into multiple channel groups through a channel splitting operation; attention weights are predicted for each channel group based on the KAN layer or a fully connected layer to generate dynamic weighting coefficients for each channel; the output shallow features are recalibrated channel by channel according to the dynamic weighting coefficients, and adaptive alignment of high-frequency textures and low-frequency semantics is performed to obtain the recalibrated features.

[0047] S350. The recalibrated features are restored to the original image resolution through the final block expansion upsampling layer, and the feature channels are mapped to the number of segmentation categories through the pointwise convolutional layer to obtain pixel-level medical image segmentation results.

[0048] Finally, the present invention is evaluated and analyzed experimentally in conjunction with the accompanying drawings and specific embodiments: First, to verify the effectiveness and generalization ability of FKM-UNet, this embodiment uses three publicly available medical image datasets for evaluation: CVC-ClinicDB, Kvasir-SEG, and ISIC 2018. CVC-ClinicDB contains 612 images from colonoscopy video sequences. Kvasir-SEG contains 1000 high-quality images of gastrointestinal polyps and their corresponding pixel-level segmentation masks. ISIC 2018 contains 2594 training images and 1000 test images, specifically designed for challenging skin lesion boundary segmentation. In this embodiment, all datasets are divided into a 70% training and 30% testing ratio, and all images are uniformly resized to 256 × 256 before being input into the network to maintain consistency in the evaluation criteria.

[0049] Furthermore, to verify the effectiveness of FKM-UNet in this embodiment, it was compared with a series of representative advanced models, including UNet, UNet++, Att-UNet, U-Next, UTNet V2, SANet, TransFuse, MALUNet, VM-UNet, U-Mamba, and U-KAN.

[0050] This embodiment's experiments were conducted independently on an NVIDIA GeForce RTX 3090 GPU equipped with 24 GB of video memory. On the CVC-ClinicDB and Kvasir-SEG datasets, the proposed model used the AdamW optimizer with an initial learning rate of 0.001 and a batch size of 16. On the ISIC 2018 dataset, the model used the Adam optimizer in conjunction with a cosine annealing learning rate scheduler, with an initial learning rate of 1e... -4 The batch size is 8. The loss function combines binary cross-entropy (BCE) and Dice loss in a 1:1 ratio. All models are trained for 400 epochs to ensure full convergence.

[0051] Furthermore, Tables 1 and 2 present the quantitative comparison results of different comparison methods on the ISIC 2018, CVC-ClinicDB, and Kvasir-SEG datasets, respectively. Evaluation metrics include mean Intersection over Union (mIoU), Dice Similarity Coefficient, Accuracy, Specificity, Sensitivity, and Precision. In the tables, the best results are shown in bold.

[0052] As shown in Table 1, on the ISIC 2018 dataset with highly irregular morphology and severe background interference, FKM-UNet achieved state-of-the-art performance across almost all key metrics. Specifically, the model in this embodiment achieved 82.58% mIoU, 91.46% Dice score, 96.52% Acc, and 98.35% Spe, comprehensively outperforming traditional UNet++ and TransFuse and SANet based on the Transformer architecture. This demonstrates that the proposed frequency separation mechanism can effectively capture complex skin lesion boundaries and significantly reduce the misdiagnosis rate.

[0053] Table 2 further summarizes the model's performance on endoscopic datasets. On the CVC-ClinicDB dataset, FKM-UNet achieved the highest Dice score (0.9405), IoU (0.8877), and precision Pre (0.9518), significantly outperforming U-KAN (0.9439) and U-Mamba (0.9170) in precision. On the more complex Kvasir-SEG dataset, the model in this embodiment maintained strong competitiveness, achieving the best Dice (0.9120), IoU (0.8450), Sensitivity Sen (0.9250), and precision Pre (0.9310). These quantitative data demonstrate the theoretical advantage of FKM-UNet, combining the high-frequency local fitting of KAN with the low-frequency global decoding of Mamba, enabling it to maintain robust performance across different medical imaging modalities.

[0054] Table 1. Comparison of key metrics for different methods on different datasets.

[0055] Table 2. Quantitative comparison results of different methods on the CVC-ClinicDB and Kvasir-SEG datasets.

[0056] Furthermore, to further explore the model's generalization and robustness in real-world scenarios, Figure 6 and Figure 7 Qualitative visual comparisons of various methods on the CVC-ClinicDB and ISIC 2018 datasets are presented. In these mask comparison images, red areas clearly indicate false positives (oversegmented regions) that were incorrectly predicted by the model, while green areas mark missed false negatives (undersegmented regions).

[0057] like Figure 6 The segmentation results of colorectal polyps are shown, where red indicates false positives and green indicates false negatives. Due to the extremely low contrast between the lesion tissue and the surrounding healthy mucosa, traditional UNet++, Att-UNet, and Mamba-based VM-UNet are easily interfered with by the reflection of intestinal folds, producing a large number of obvious green missed diagnosis bands and red misdiagnosis blocks at the outer edge of the lesion. In stark contrast to these models that are limited by local receptive fields or suffer from excessive smoothing interference, the example FKM-UNet achieves the sharpest definition of the lesion area, with very few red and green pixel residues in its prediction results, demonstrating a near-perfect topological fit to the ground truth (GT).

[0058] In processing Figure 7 When dealing with ISIC 2018 skin lesion images exhibiting extremely irregular morphology and accompanied by noise interference such as hair, the shortcomings of baseline models are further amplified. Mainstream Transformer-type models (such as TransFuse, SANet, and UTNet V2) often output highly blurred and harsh contours when handling complex jagged edges, incorrectly classifying many irregular pathological edges into normal skin (resulting in large areas of green pixels). In contrast, FKM-UNet not only accurately depicts the complex, spiky morphology of the outermost layer but also successfully suppresses background noise interference, outputting the highest fidelity segmentation mask.

[0059] Furthermore, while ensuring excellent segmentation accuracy, the FKM-UNet in this embodiment, thanks to its carefully designed dual-algorithm weight reduction mechanism, keeps the number of parameters to approximately 30M. This not only significantly lowers the computational barrier for clinical deployment but also demonstrates a computational efficiency advantage far exceeding other heavyweight Mamba or Transformer models.

[0060] To thoroughly verify the specific contributions of each core innovative module in FKM-UNet (i.e., the local-global KAN-Mamba module LKM-Block and the dual-scale fusion bottleneck layer DSFusion) to the overall segmentation performance, this embodiment conducted detailed ablation experiments on the Kvasir-SEG dataset. Based on the FKM-UNet framework, this embodiment constructed different network variants by successively removing or replacing core modules, and used the Dice similarity coefficient and mean intersection-union ratio (mIoU) as quantitative evaluation metrics. The experimental results are shown in Table 3.

[0061] For w / o DSFusion (removing the dual-scale fusion bottleneck layer): When the deepest DSFusion module in this embodiment is replaced with regular feature concatenation, the model's Dice score drops to 0.8952, and mIoU drops to 0.8315. This significant performance decrease demonstrates that in a highly compressed semantic space, direct concatenation leads to semantic mixing of spatial and feature flows. The DSFusion module, through a cross-modal cross-attention mechanism, effectively utilizes the global spatial prior extracted by Mamba to guide the fusion of high-frequency semantics in KAN, playing a crucial role in improving the final segmentation accuracy.

[0062] For w / o LKMBlock (removing the local-global KAN-Mamba module): When the LKMBlock in the encoder and decoder of this embodiment is replaced with a conventional feature extraction module that does not include high- and low-frequency dual-stream separation, the model's Dice score drops to 0.9103, and the mIoU also drops to 0.8338. This indicates that without Laplacian-based frequency domain decoupling for dual-stream processing, the model cannot simultaneously consider smooth anatomical structures and complex high-frequency edges, resulting in a significant decrease in overall boundary fitting accuracy and region overlap. LKMBlock successfully decouples these two conflicting optimization objectives, thereby improving the overall segmentation performance.

[0063] For FKM-UNet (the complete model): When both LKMBlock and DSFusion are enabled, the model achieves optimal overall performance, obtaining the highest Dice score (0.9120) and mIoU (0.8450). This fully demonstrates the high complementarity and necessity of frequency separation feature extraction and cross-modal attention fusion in the architecture.

[0064] Table 3. Ablation Study Results of Different Module Combinations on the Kvasir-SEG Dataset

[0065] The symbol × indicates that the module is removed, and √ indicates that the module is included.

[0066] Reference Figure 2 A medical image segmentation system based on a channel-level frequency separation mechanism includes: The first module 201 is used to acquire two-dimensional medical images containing lesion areas and perform data preprocessing to obtain preprocessed medical images. The second module 202 is used to introduce a local-global KAN-Mamba module, a dual-scale fusion bottleneck layer and a KAN split channel attention bridging layer, and to construct a frequency-decoupled KAN-Mamba U-shaped segmentation network using an encoder-decoder symmetrical topology. The third module 203 is used to perform feature extraction and segmentation decoding on the preprocessed medical image based on the frequency-decoupled KAN-Mamba U-shaped segmentation network to obtain pixel-level medical image segmentation results.

[0067] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0068] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this is not intended to limit the scope of the embodiments of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the embodiments of the present application.

Claims

1. A medical image segmentation method based on a channel-level frequency separation mechanism, characterized in that, Includes the following steps: Two-dimensional medical images containing lesion areas are acquired and data preprocessed to obtain preprocessed medical images; By introducing a local-global KAN-Mamba module, a dual-scale fusion bottleneck layer, and a KAN split channel attention bridging layer, and adopting an encoder-decoder symmetrical topology, a frequency-decoupled KAN-Mamba U-shaped segmentation network is constructed. Based on the frequency-decoupled KAN-Mamba U-shaped segmentation network, feature extraction and segmentation decoding are performed on the preprocessed medical images to obtain pixel-level medical image segmentation results.

2. The medical image segmentation method based on channel-level frequency separation mechanism according to claim 1, characterized in that, The step of acquiring a two-dimensional medical image containing the lesion area and performing data preprocessing to obtain a preprocessed medical image specifically includes: Acquire two-dimensional medical images containing the lesion area and corresponding gold standard masks annotated by experts; All two-dimensional medical images and gold standard masks are uniformly adjusted to the preset spatial resolution to obtain the adjusted two-dimensional medical images and the adjusted gold standard mask; The pixel values ​​of the adjusted two-dimensional medical image are normalized and mapped to a preset value range to obtain a normalized two-dimensional medical image. Data augmentation operations involving random rotation and random flipping are performed on the normalized two-dimensional medical image and the adjusted gold standard mask to obtain the augmented medical image. The enhanced medical image is converted into a standard three-dimensional feature tensor to obtain a preprocessed medical image.

3. The medical image segmentation method based on channel-level frequency separation mechanism according to claim 2, characterized in that, The frequency-decoupled KAN-Mamba U-shaped segmentation network specifically includes a feature encoding path, cross-scale skip connections, a dual-scale fusion bottleneck layer, and a decoding recovery path, wherein: The feature encoding path adopts a multi-level hierarchical processing design. Each level consists of a block merging downsampling layer and a local-global KAN-Mamba module cascaded together, which is used to extract multi-scale high and low frequency features step by step. The decoding recovery path is symmetrical to the encoder structure and consists of a block extended upsampling layer and a local-global KAN-Mamba module, which is used to recover spatial resolution step by step and fuse shallow features. The dual-scale fusion bottleneck layer receives encoder end features and performs cross-spectrum attention fusion to achieve deep fusion of high and low frequency information. The KAN split-channel attention bridging layer is embedded at each cross-scale jump connection location to perform dynamic channel recalibration on shallow features on the encoder side.

4. The medical image segmentation method based on channel-level frequency separation mechanism according to claim 3, characterized in that, The step of performing feature extraction and segmentation decoding on the preprocessed medical image based on the frequency-decoupled KAN-Mamba U-shaped segmentation network to obtain pixel-level medical image segmentation results specifically includes: Based on the frequency-decoupled KAN-Mamba U-shaped segmentation network, the preprocessed medical image is input into the starting block embedding layer and mapped to the multi-scale feature space to obtain the mapped medical image. The mapped medical image is then passed through a four-level coding path for downsampling and a local-global KAN-Mamba module to extract high- and low-frequency features at different scales. High and low frequency features at different scales are input into the dual-scale fusion bottleneck layer to perform cross-spectral deep fusion and obtain the final fused features of high and low frequency information. The final high and low frequency information fusion features are sequentially processed by upsampling and local-global KAN-Mamba modules in a four-level decoding path. At each level, the recalibrated features of the corresponding encoder level are fused through the KAN split channel attention bridging layer to obtain the recalibrated features. The recalibrated features are restored to the original image resolution through a final block expansion upsampling layer, and the feature channels are mapped to the number of segmentation categories through a pointwise convolutional layer to obtain pixel-level medical image segmentation results.

5. The medical image segmentation method based on channel-level frequency separation mechanism according to claim 4, characterized in that, The data processing flow of the local-global KAN-Mamba module specifically includes: Based on the preset channel splitting ratio, the preprocessed medical image is divided into a high-frequency feature subset and a low-frequency feature subset; High-frequency enhanced features are obtained by performing edge detail enhancement processing on a subset of high-frequency features based on high-frequency streams. Based on the low-frequency flow, a global structure modeling process is performed on the low-frequency feature subset to obtain low-frequency global features; High-frequency enhanced features and low-frequency global features are concatenated along the channel dimension, and inter-channel information interaction and mixing are performed through convolutional layers to obtain fused features; The fused features are added to the global residual of the preprocessed medical image to obtain medical image features with high and low frequency information.

6. The medical image segmentation method based on channel-level frequency separation mechanism according to claim 5, characterized in that, The step of performing edge detail enhancement processing on a subset of high-frequency features based on high-frequency streams to obtain high-frequency enhanced features specifically includes: The high-frequency feature subset is input into the average pooling layer for smooth downsampling, and then bilinear interpolation is used to restore it to the original resolution to extract the slowly changing background smoothing component. The high-frequency feature subset is subtracted element-wise from the background smoothing component to obtain the high-frequency detail residual field; The high-frequency detail residual field is enhanced by nonlinear mapping based on the KAN two-dimensional convolution operator to obtain the enhanced high-frequency detail residual field. The KAN two-dimensional convolution is composed of the superposition of linear basis paths and high-order spline paths. The enhanced high-frequency detail residual field is processed sequentially through a batch normalization layer and then added to the residual of the high-frequency feature subset to obtain the high-frequency enhanced features.

7. The medical image segmentation method based on channel-level frequency separation mechanism according to claim 6, characterized in that, The step of performing global structural modeling on a subset of low-frequency features based on low-frequency flow to obtain low-frequency global features specifically includes: The low-frequency feature subset is spatially downsampled through an average pooling layer to obtain the compressed low-frequency feature tensor; The compressed low-frequency feature tensor is input into the visual state space block, and a two-dimensional selective scanning operation is performed to obtain the scanned low-frequency feature tensor. The scanned low-frequency feature tensor is restored to its original physical resolution through bilinear interpolation, resulting in the low-frequency feature tensor with restored resolution. The low-frequency feature tensor after resolution restoration is processed sequentially through pointwise convolutional layers and batch normalization layers to obtain low-frequency global features.

8. The medical image segmentation method based on channel-level frequency separation mechanism according to claim 7, characterized in that, The data processing flow of the dual-scale fusion bottleneck layer specifically includes: Medical image features with high and low frequency information are used as input features and input to the spatial flow branch and the feature flow branch respectively. The visual state spatial block based on spatial flow branch performs a global scan of the input features and, combined with layer normalization, extracts the global structure and long-range dependencies of the lesions, and outputs spatial query basis vectors. The KAN linear operator based on feature flow branching, combined with layer normalization, maps the input features and extracts fine semantic attributes, outputting a feature dictionary vector. The spatial query basis vector is passed through a linear layer to generate a query vector, and the feature dictionary vector is passed through the corresponding linear layers to generate key vectors and value vectors, respectively. The query vector, key vector, and value vector are flattened into a two-dimensional sequence matrix in the spatial dimension, and cross-attention weighting operation is performed to obtain the attention-weighted features; The attention-weighted features are reconstructed in spatial dimension and then regularized using dropout to obtain hybrid fusion features; The hybrid fusion features are added to the query vector by residual addition, and then subjected to layer normalization to obtain the final fusion features of high and low frequency information.

9. The medical image segmentation method based on channel-level frequency separation mechanism according to claim 8, characterized in that, The dynamic channel recalibration step of the KAN split channel attention bridging layer specifically includes: The shallow features output from the corresponding layer of the encoder are received as input to the bridging layer; The input features are divided into multiple channel groups through a channel splitting operation; Attention weights are predicted for each channel group based on KAN layers or fully connected layers, generating dynamic weighting coefficients for each channel. The shallow features of the output are recalibrated channel by channel based on the dynamic weighting coefficients, and adaptive alignment of high-frequency texture and low-frequency semantics is performed to obtain the recalibrated features.

10. A medical image segmentation system based on a channel-level frequency separation mechanism, characterized in that, Includes the following modules: The first module is used to acquire two-dimensional medical images containing lesion areas and perform data preprocessing to obtain preprocessed medical images. The second module is used to introduce a local-global KAN-Mamba module, a dual-scale fusion bottleneck layer and a KAN split channel attention bridging layer. It adopts an encoder-decoder symmetrical topology to construct a frequency-decoupled KAN-Mamba U-shaped segmentation network. The third module is used to perform feature extraction and segmentation decoding on the preprocessed medical images based on the frequency-decoupled KAN-Mamba U-shaped segmentation network to obtain pixel-level medical image segmentation results.

Images