A colonoscopy polyp segmentation method and system based on an improved KAN network

Abstract

The application provides a colonoscopy polyp segmentation method and system based on an improved KAN network, and relates to the technical field of polyp segmentation. The method comprises the following steps: acquiring a colonoscopy image to be segmented and inputting the image into an encoder; the encoder comprises a plurality of double-attention enhanced deep separable Tok-KAN modules, which are used for feature extraction and output encoding features; the encoding features are transmitted to a decoder through a jump connection; a hollow space pyramid pooling module is arranged in the jump connection, the encoding features are subjected to multi-scale context modeling, and multi-scale fusion features are obtained; the multi-scale fusion features are input into the decoder, and the enhanced features of an edge-guided learnable residual attention fusion module are fused with the output of the decoder in the decoding process to obtain segmentation prediction results; and the segmentation prediction results are output. The application can effectively solve the problems of low segmentation precision and insufficient generalization caused by background noise, scale difference and boundary blur of the colonoscopy image.

Description

Technical Field

[0001] This invention relates to the field of polyp segmentation technology, specifically to a colonoscopy polyp segmentation method and system based on an improved KAN network. Background Technology

[0002] Automatic segmentation of polyps in colonoscopy images is of significant clinical value for the early screening and diagnosis of colorectal cancer. In recent years, deep learning technology has been widely applied to colonoscopy polyp segmentation, significantly improving the automation level of segmentation. U-Net and its variants (e.g., Attention U-Net, TransUNet, etc.) are currently the mainstream polyp segmentation network architectures. Building upon this, the U-KAN model combines the nonlinear expressive power of the Kolmogorov-Arnold network (KAN) with the hierarchical features of the U-shaped structure, providing a new approach to medical image segmentation and achieving high segmentation accuracy while reducing computational costs.

[0003] However, real-world colonoscopy images are often accompanied by complex interferences such as mucus, reflections, bubbles, folds, and differences in intestinal cleanliness. Furthermore, polyps often have blurred boundaries, irregular shapes, and large scales. Existing models still face the following shortcomings when dealing with these complex scenarios:

[0004] First, the feature extraction unit at the encoding end has limited ability to screen effective features in complex colonoscopy images, making it difficult to fully suppress background noise and enhance lesion-related characterization. Second, the skip connection adopts a direct splicing method, lacking sufficient multi-scale context modeling capability, making it difficult to take into account both the details of small polyps and the overall outline of large lesions. Third, the feature fusion module in the decoding process has insufficient adaptability to scenarios such as blurred boundaries and low contrast, making it difficult to achieve accurate boundary characterization in complex noisy regions. Summary of the Invention

[0005] To address the problems in related technologies, this invention provides a colonoscopy polyp segmentation method and system based on an improved KAN network.

[0006] To achieve the above objectives, the technical solution adopted by the present invention includes:

[0007] According to a first aspect of the present invention, a method for colonoscopic polyp segmentation based on an improved KAN network is provided, comprising the following steps:

[0008] Step S1: Obtain the colonoscopy image to be segmented;

[0009] Step S2: Input the colonoscopy image into the encoder, which contains multiple dual-attention enhanced deep separable Tok-KAN modules. The dual-attention enhanced deep separable Tok-KAN modules are used to extract dual-attention enhanced deep separable convolutional features from the input features and output encoded features.

[0010] Step S3: The encoded features are passed to the decoder through skip connections. A hollow spatial pyramid pooling module is set in the skip connections to perform multi-scale context modeling on the encoded features and obtain multi-scale fused features.

[0011] Step S4: Input the multi-scale fusion features into the decoder, and during the decoding process, fuse the enhanced features of the edge-guided learnable residual attention fusion module with the output of the decoder to obtain the segmentation prediction result;

[0012] Step S5: Output the segmentation prediction result.

[0013] Optionally, in step S2, the dual-attention enhanced deep separable Tok-KAN module performs dual-attention enhanced deep separable convolution feature extraction on the input features, specifically including:

[0014] Step S2-1: Reshape the input sequence into a two-dimensional feature map;

[0015] Step S2-2: Perform 7×7 depthwise convolution, pointwise convolution, activation function processing, pointwise convolution and batch normalization on the two-dimensional feature map in sequence to obtain the first convolutional feature;

[0016] Step S2-3: Apply channel attention weighting and spatial attention weighting to the first convolutional feature sequentially to obtain dual attention enhanced features;

[0017] Step S2-4: Add the residuals of the two-dimensional feature map and the dual attention enhancement feature and reshape it into a sequence form, which is used as the output feature of the dual attention enhancement depth separable Tok-KAN module.

[0018] Optionally, in step S2-2, the first convolutional feature is obtained in the following manner:

[0019]

[0020] In the formula, Represents the first convolutional feature. This represents the two-dimensional feature map. This represents a 7×7 depthwise convolution. This represents pointwise convolution. This represents the GELU activation function. This indicates batch normalization.

[0021] Optionally, in steps S2-4, the output features of the dual-attention enhanced depth-separable Tok-KAN module are obtained as follows:

[0022]

[0023] In the formula, This represents the output features of the dual-attention enhanced depthwise separable Tok-KAN module. This represents the operation of reshaping a two-dimensional feature map into a sequence. This represents the two-dimensional feature map. This represents spatial attention operations. This indicates channel attention operations. This represents the first convolutional feature.

[0024] Optionally, in step S4, the processing procedure of the edge-guided learnable residual attention fusion module specifically includes:

[0025] Step S4-1: Receive encoder backbone features, external edge guiding features, and preliminary segmentation prediction map as input;

[0026] Step S4-2: Perform background suppression, high-frequency edge information extraction, and contour guidance processing on the input of step S4-1 through parallel branches to obtain multi-path processing features;

[0027] Step S4-3: After concatenating the multi-channel processed features, perform convolutional fusion and spatial attention weighting in sequence to obtain enhanced features focused on the polyp region;

[0028] Step S4-4: Perform weighted residual fusion of the enhanced features obtained in step S4-3 and the encoder backbone features using trainable parameters to obtain the output features of the edge-guided learnable residual attention fusion module.

[0029] Optionally, in step S4-4, the output features of the edge-guided learnable residual attention fusion module are obtained in the following manner:

[0030]

[0031] In the formula, This indicates that the edge-guided learnable residual attention fusion module can output features. Indicates trainable parameters, This represents the enhanced features obtained in step S4-3. This represents the encoder backbone features.

[0032] Optionally, in step S4-4, before performing weighted residual fusion of the enhanced features and the encoder backbone features, the enhanced features are further subjected to layer normalization processing:

[0033]

[0034] In the formula, The enhanced features are represented by the normalized layer. This represents the enhanced features obtained in step S4-3. Presentation layer normalization operation;

[0035] And Weighted residual fusion is performed with the encoder backbone features.

[0036] Optionally, in step S3, the hollow spatial pyramid pooling module performs multi-scale context modeling on the encoded features, specifically including:

[0037] Step S3-1: Input the encoded features into four parallel branches respectively. The four parallel branches are a 1×1 convolution branch, a 3×3 dilated convolution branch with a dilation rate of 2, a 3×3 dilated convolution branch with a dilation rate of 4, and a 3×3 dilated convolution branch with a dilation rate of 6, to obtain four branch features respectively.

[0038] Step S3-2: After concatenating the four branch features along the channel dimension, perform batch normalization and ReLU activation to obtain multi-scale fusion features.

[0039] Optionally, in step S3-1, the four branch features are obtained respectively in the following ways:

[0040]

[0041]

[0042]

[0043]

[0044] In the formula, Represents coding features, Represents a 1×1 convolution. This represents a 3×3 dilated convolution with a dilation rate of 2. This represents a 3×3 dilated convolution with an expansion rate of 4. This represents a 3×3 dilated convolution with a dilation rate of 6. , , and These represent the four branch features respectively;

[0045] In step S3-2, the multi-scale fusion features are obtained in the following manner:

[0046]

[0047] In the formula, This represents the multi-scale fusion feature. Represents the ReLU activation function. Indicates batch normalization. This indicates a channel dimension splicing operation.

[0048] According to a second aspect of the present invention, a colonoscopy polyp segmentation system based on an improved KAN network is also provided, for performing the colonoscopy polyp segmentation method based on an improved KAN network as described in any of the technical solutions of the first aspect of the present invention, the system comprising:

[0049] The encoder contains multiple dual-attention enhanced deep separable Tok-KAN modules for extracting dual-attention enhanced deep separable convolutional features from colonoscopy images to be segmented, and outputs encoded features.

[0050] The skip connection module, located between the encoder and the decoder, includes a hole spatial pyramid pooling module for multi-scale context modeling of the encoded features to obtain multi-scale fused features;

[0051] A decoder is used to decode the multi-scale fused features;

[0052] An edge-guided learnable residual attention fusion module is set in the decoding path of the decoder to fuse enhanced features with the output of the decoder;

[0053] The output module is used to output the segmentation prediction results.

[0054] Beneficial effects:

[0055] 1. First, this invention can effectively improve the segmentation accuracy in complex colonoscopy scenarios. Specifically, at the encoding end, this invention enhances the depthwise separable Tok-KAN module through dual attention. While retaining the local modeling efficiency of depthwise separable convolution, it adaptively strengthens polyp-related features and suppresses background noise (such as mucus, reflection, bubbles, etc.) from both channel and spatial dimensions, providing more discriminative encoded features for subsequent processing. In the skip connections, a multi-receptive-field parallel sampling structure is constructed through a hollow spatial pyramid pooling module, so that the features passed to the decoder simultaneously contain local details and global context, effectively balancing the preservation of details of small polyp targets with the integrity of the overall contour of large lesions.

[0056] Secondly, this invention effectively improves segmentation integrity in scenes with weak boundaries and small targets. Specifically, during the decoding process, this invention introduces an edge-guided learnable residual attention fusion module. This module dynamically adjusts the fusion ratio of edge enhancement information and backbone features through trainable parameters. In areas with blurred boundaries or discontinuous contours, it emphasizes edge enhancement information, while in noisy areas such as highlights and slime, it retains more robust backbone features, balancing boundary characterization and background robustness. Simultaneously, the ASPP module effectively preserves small target features through parallel sampling across multiple receptive fields, enabling the model to adapt well to polyps with extremely large size ranges.

[0057] Third, this invention demonstrates excellent cross-dataset generalization ability and real-world deployment value. Specifically, in external validation with 405 real-world colonoscopy images, the invention achieved accuracy, Dice, IoU, and Recall of 0.9761, 0.8146, 0.6871, and 0.7995, respectively, all superior to the baseline U-KAN; the precision was 0.8302, comparable to the control model overall, indicating that the invention maintains an acceptable level of false positives while improving the completeness of lesion detection. These results demonstrate that the invention not only performs excellently on standard laboratory datasets but also possesses good generalization ability and practical deployment value in complex real-world clinical scenarios.

[0058] Fourth, the modular architecture design of this invention facilitates engineering deployment and clinical integration. Specifically, this invention designs the encoder, skip connection module (including ASPP), decoder, and edge-guided learnable residual attention fusion module as independently coupled modular components. Data is transferred between modules through standardized feature interfaces, facilitating targeted adjustments and optimizations for different clinical application scenarios without requiring a complete redesign of the system architecture. Furthermore, the system inference phase only performs forward propagation computation, enabling efficient real-time inference while ensuring segmentation accuracy, meeting the timeliness requirements for real-time lesion identification in clinical colonoscopy.

[0059] Fifth, this invention provides reliable methodological support for computer-aided diagnosis in colonoscopy. Specifically, through synergistic improvements in feature extraction, multi-scale perception, and boundary optimization, this invention effectively addresses core technical challenges in real-world colonoscopy images, such as complex background noise, significant differences in lesion scale, and blurred boundaries. It provides a high-precision, highly generalizable, and deployable automated solution for computer-aided colonoscopy segmentation, serving as a reference for subsequent computer-aided diagnostic research.

[0060] 2. Other beneficial effects or advantages of the present invention will be described in detail in the specific embodiments. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] in:

[0063] Figure 1 This is a flowchart illustrating the steps of a colonoscopy polyp segmentation method based on an improved KAN network, provided by an exemplary embodiment of the present invention.

[0064] Figure 2 This is a schematic diagram of the structure of a DA TOK-KAN module provided in an exemplary embodiment of the present invention, wherein, Figure 2 a is a schematic diagram of the DA-Tok-KAN structure. Figure 2 b is a schematic diagram of the KAN Layer structure;

[0065] Figure 3 This is a schematic diagram of the structure of DA DwConv provided in an exemplary embodiment of the present invention;

[0066] Figure 4 This is a schematic diagram of the structure of an ELRA module provided in an exemplary embodiment of the present invention;

[0067] Figure 5 This is a schematic diagram of the structure of an ASPP module provided in an exemplary embodiment of the present invention;

[0068] Figure 6 This is a comparison diagram of the image processing process of the method of this invention and the processing methods in existing related technologies on a public dataset;

[0069] Figure 7 This is a comparison diagram of the image processing process of the method of the present invention and the processing methods in existing related technologies on a private dataset. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0071] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0072] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices. It should also be noted that in embodiments of this invention, the words "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in embodiments of this invention should not be construed as preferred or advantageous over other embodiments or designs. Specifically, the use of the words "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0073] To facilitate a clearer and more accurate understanding of the technical solution of this invention by those skilled in the art, the technical concept of this invention will be explained below.

[0074] The technical concept of this invention is as follows: First, at the encoding end, a depthwise separable convolutional module with dual attention enhancement is used to strengthen polyp-related features and suppress background interference while preserving local modeling efficiency, providing high-quality input for subsequent modeling. Second, a hollow spatial pyramid pooling module is introduced into the skip connections to achieve contextual awareness of polyps at different scales through parallel sampling of multiple receptive fields, thus taking into account both small target details and the overall structure of large lesions. Finally, during the decoding process, an edge-guided learnable residual attention fusion module is used to dynamically adjust the fusion ratio of edge enhancement information and backbone features through trainable parameters, enabling the model to focus more on contour characterization in weak boundary regions and retain more robust backbone information in noisy regions such as highlights and mucus. Furthermore, the above three improvement directions synergize in feature extraction, multi-scale modeling, and boundary optimization to jointly improve the model's segmentation accuracy and cross-dataset generalization ability in complex colonoscopy scenarios.

[0075] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0076] like Figure 1 As shown, according to a first aspect of the present invention, a colonoscopy polyp segmentation method based on an improved KAN network is provided, comprising the following steps:

[0077] Step S1: Obtain the colonoscopy image to be segmented;

[0078] Step S2: Input the colonoscopy image into the encoder. The encoder contains multiple dual-attention enhanced deep separable Tok-KAN modules. The dual-attention enhanced deep separable Tok-KAN modules are used to extract dual-attention enhanced deep separable convolutional features from the input features and output encoded features.

[0079] Step S3: Pass the encoded features to the decoder through skip connections. Set up a hollow spatial pyramid pooling module in the skip connections to perform multi-scale context modeling on the encoded features and obtain multi-scale fused features.

[0080] Step S4: Input the multi-scale fused features into the decoder, and during the decoding process, fuse the enhanced features of the edge-guided learnable residual attention fusion module with the output of the decoder to obtain the segmentation prediction result;

[0081] Step S5: Output the segmentation prediction results.

[0082] Through the above technical solution, firstly, the method of the present invention can improve the discriminative power and anti-interference ability of the encoding end features. Specifically, in the prior art, the basic feature extraction unit of the encoding end of U-KAN mainly relies on conventional deep convolution to complete spatial modeling. Its receptive field and feature selection method are relatively fixed, making it difficult to effectively distinguish polyp lesions from noise regions in the context of strong interference such as mucus, reflection, and bubbles, resulting in insufficient feature discriminative power, and subsequent KAN layers are unable to obtain high-quality input representations.

[0083] In this invention, the encoder includes a dual-attention enhanced depthwise separable Tok-KAN module. This module introduces a dual-attention mechanism (channel attention and spatial attention) to further enhance high-response features related to polyp texture and morphology from the channel dimension, building upon the local feature extraction by depthwise separable convolution, and focusing the model on the lesion region from the spatial dimension. This allows the encoder to retain the local modeling efficiency of depthwise separable convolution while adaptively adjusting feature weights based on the local distribution of the colonoscopy image. This effectively filters polyp-related features and suppresses background noise interference, providing more discriminative encoded features for subsequent processing and thus effectively improving segmentation accuracy.

[0084] Second, the method of this invention can enhance the adaptability and segmentation integrity of polyps of different sizes. Specifically, in the prior art, the skip connections of U-KAN use a direct splicing method to pass the encoded features to the decoder, lacking multi-scale receptive field modeling of the features. Because the size range of polyps in colonoscopy images is extremely wide, small polyps may only occupy tens of pixels in the image, while large lesions may cover hundreds of pixels. Single-scale feature fusion methods are difficult to take into account the semantic information of polyps of different sizes, resulting in small targets being easily ignored and large lesions having insufficient contour integrity.

[0085] In this invention, a dilated spatial pyramid pooling module is incorporated into the skip connections. This module constructs a multi-receptive-field feature extractor through parallel dilated convolutions with different dilation rates. This allows for multi-scale sampling of encoded features without increasing the number of parameters, acquiring contextual information from different receptive fields. Consequently, the features passed to the decoder from the skip connections contain both local details and global contextual information, effectively balancing the preservation of details in small polyps with the integrity of the overall contour of large lesions. This enhances the model's adaptability to polyps of different scales and improves the completeness and consistency of the segmentation results.

[0086] Third, the method of this invention can improve the segmentation continuity and positioning accuracy in weak boundary scenes. Specifically, in the prior art, the feature fusion in the decoding stage adopts residual connection with fixed coefficients, which cannot dynamically adjust the fusion weights of enhanced features and original features according to the image content. This makes it difficult to handle complex scenes in colonoscopy images with blurred boundaries, low contrast, and local noise such as highlights and mucus, resulting in insufficient segmentation continuity in weak boundary areas and a tendency to produce false detections in noisy areas.

[0087] In this invention, the enhanced features of the edge-guided learnable residual attention fusion module are fused with the decoder output during the decoding process. This module is based on encoder backbone features, external edge-guided features, and preliminary segmentation predictions. Figure 3 The input is processed through parallel branches to perform background suppression, high-frequency edge information extraction, and contour guidance, ultimately obtaining enhanced features focused on the polyp region. By integrating these boundary-rich enhanced features into the decoder output, this invention effectively strengthens the contour representation of the lesion region, improves the model's boundary delineation ability in weak-boundary, low-contrast scenes, and enhances the spatial continuity and localization accuracy of the segmentation results in the boundary region.

[0088] Fourth, the three improved modules (the dual-attention enhanced depth-separable Tok-KAN module, the hollow spatial pyramid pooling module, and the edge-guided learnable residual attention fusion module) can effectively form a synergistic effect. Specifically, the dual-attention enhancement at the encoding end can provide higher-quality input features for multi-scale modeling at skip connections, enabling the ASPP module (hollow spatial pyramid pooling module) to perform multi-receptive field sampling based on better features. At the same time, the multi-scale fusion features output by the ASPP module can provide decoder outputs with rich contextual information for edge-guided fusion at the decoding end, making its boundary enhancement effect more accurate and stable. The three modules work together to improve the segmentation accuracy, boundary characterization quality, and cross-sample generalization ability of the model in complex colonoscopy scenarios, thereby effectively solving the problems of complex background noise, significant differences in lesion scale, and blurred boundaries faced by existing technologies in real-world colonoscopy image segmentation.

[0089] In one embodiment of the present invention, in step S2, the dual-attention enhanced deep separable Tok-KAN module performs dual-attention enhanced deep separable convolution feature extraction on the input features, specifically including:

[0090] Step S2-1: Reshape the input sequence into a two-dimensional feature map;

[0091] Step S2-2: Perform 7×7 depthwise convolution, pointwise convolution, activation function processing, pointwise convolution and batch normalization on the two-dimensional feature map in sequence to obtain the first convolution feature;

[0092] Step S2-3: Apply channel attention weighting and spatial attention weighting to the first convolutional feature in sequence to obtain dual attention enhanced features;

[0093] Step S2-4: After adding the residuals of the two-dimensional feature map and the dual attention enhancement feature, reshape it into a sequence form, which serves as the output feature of the dual attention enhancement depth separable Tok-KAN module.

[0094] In this implementation, the specific processing flow of the dual-attention enhanced deep separable Tok-KAN module includes, in sequence: reshaping, depthwise convolution, pointwise convolution, activation, pointwise convolution, batch normalization, channel attention, spatial attention, residual addition, and sequence reshaping. This process embeds the dual-attention mechanism into the backbone structure of deep separable convolution, allowing attention-weighted features to be residually fused with the original features. This strengthens the representation of lesion regions while preserving the original information of the input features, preventing the degradation of effective features in deep networks, and providing high-quality sequential input for nonlinear modeling of subsequent KAN layers.

[0095] In one embodiment of the present invention, in step S2-2, the first convolutional feature can be obtained in the following manner:

[0096]

[0097] In the formula, Represents the first convolutional feature. Represents a two-dimensional feature map. This represents a 7×7 depthwise convolution. This represents pointwise convolution. This represents the GELU activation function. This indicates batch normalization.

[0098] In this embodiment, a large-size (7×7) depth convolution can fully capture the local texture details and spatial structure of polyp lesions. Combined with the introduction of nonlinear mapping by the GELU activation function, it can effectively improve the ability of convolution features to express nonlinear distribution features in complex colonoscopy images.

[0099] In one embodiment of the present invention, in steps S2-4, the output features of the dual-attention enhanced depth-separable Tok-KAN module can be obtained in the following manner:

[0100]

[0101] In the formula, This represents the output features of the dual-attention enhanced depthwise separable Tok-KAN module. This represents the operation of reshaping a two-dimensional feature map into a sequence. Represents a two-dimensional feature map. This represents spatial attention operations. This indicates channel attention operations. This represents the first convolutional feature.

[0102] In this embodiment, for the output features of the dual attention-enhanced depth-separable Tok-KAN module, response features related to polyps are first screened along the channel dimension, and then the lesion area is located along the spatial dimension, thereby effectively suppressing noise. At the same time, the basic information in the original features can be preserved through the residual addition operation, so that the output features are both discriminative and maintain feature stability.

[0103] In one embodiment of the present invention, step S4 of the present invention specifically includes the following processing steps of the edge-guided learnable residual attention fusion module:

[0104] Step S4-1: Receive encoder backbone features, external edge guiding features, and preliminary segmentation prediction map as input;

[0105] Step S4-2: Perform background suppression, high-frequency edge information extraction, and contour guidance processing on the input of step S4-1 through parallel branches to obtain multi-path processing features;

[0106] Step S4-3: After concatenating the multi-channel processed features, perform convolutional fusion and spatial attention weighting in sequence to obtain enhanced features focused on the polyp region;

[0107] Step S4-4: Perform weighted residual fusion of the enhanced features obtained in step S4-3 and the encoder backbone features using trainable parameters to obtain the output features of the edge-guided learnable residual attention fusion module.

[0108] In this embodiment, the specific processing flow of the edge-guided learnable residual attention fusion module includes, in sequence: receiving three types of inputs, parallel branch processing, stitching and fusion, spatial attention weighting, and learnable residual fusion. This process completes background suppression, high-frequency edge information extraction, and contour guidance through parallel branches, and finally focuses the model on the polyp boundary region through spatial attention; finally, weighted residual fusion is performed through trainable parameters, so that the feature fusion ratio can be adaptively adjusted according to the image content, enhancing the adaptability of the module in different scenarios.

[0109] In one embodiment of the present invention, in step S4-4, the output features of the edge-guided learnable residual attention fusion module can be obtained in the following manner:

[0110]

[0111] In the formula, This indicates the output features of the edge-guided learnable residual attention fusion module. Indicates trainable parameters, This represents the enhanced features obtained in step S4-3. This represents the encoder backbone features.

[0112] In this implementation, trainable parameters are introduced. It can achieve enhanced features With main characteristics The dynamic balance between these factors can enhance the weight of boundary information in areas with blurred boundaries, while preserving the robustness of the main features in noisy areas. This can improve the lack of flexibility of traditional fixed-coefficient fusion strategies in complex colonoscopy scenarios.

[0113] In this implementation, it can be understood that for trainable parameters In this regard, the determination method can be automatic updating through backpropagation and optimizers (such as Adam), or it can be designed as a scheduling parameter that changes dynamically with training rounds, for example, relying more on backbone features in the early stages of training ( (Smaller), edge guidance gradually enhanced later ( Increase), or adaptively increase when the loss decrease stagnates. To enhance edge details, this invention does not specifically limit the value; for example, the range of values ​​can be: ...

[0114] In one embodiment of the present invention, in step S4-4, before performing weighted residual fusion of the enhanced features and the encoder backbone features, the enhanced features may be subjected to layer normalization processing:

[0115]

[0116] In the formula, The enhanced features are represented by the normalized layer. This represents the enhanced features obtained in step S4-3. Presentation layer normalization operation;

[0117] And Weighted residual fusion is performed with the encoder backbone features.

[0118] In this implementation, the layer normalization operation can calibrate the distribution of enhanced features to eliminate feature magnitude fluctuations caused by differences in image quality between different samples. This makes the subsequent weighted residual fusion process more stable and helps the model maintain stable performance under different imaging conditions and intestinal cleanliness.

[0119] In one embodiment of the present invention, in step S3, the hollow spatial pyramid pooling module performs multi-scale context modeling on the encoded features, which may specifically include:

[0120] Step S3-1: Input the encoded features into four parallel branches respectively. The four parallel branches are a 1×1 convolution branch, a 3×3 dilated convolution branch with a dilation rate of 2, a 3×3 dilated convolution branch with a dilation rate of 4, and a 3×3 dilated convolution branch with a dilation rate of 6, and obtain four branch features respectively.

[0121] Step S3-2: After concatenating the four branch features along the channel dimension, perform batch normalization and ReLU activation to obtain multi-scale fused features.

[0122] In this implementation, the specific processing flow of the dilated spatial pyramid pooling module can be as follows: four parallel branches, concatenation, batch normalization, and ReLU activation. In this way, through parallel 1×1 convolution and dilated convolution with three different dilation rates, this processing flow can achieve feature sampling of multiple receptive fields in the same layer, so that multi-scale fused features can simultaneously contain local details (small dilation rate) and global context (large dilation rate), which can effectively improve the problem that single-scale features are difficult to take into account polyps of different sizes.

[0123] In one embodiment of the present invention, in step S3-1, the four branch features can be obtained respectively in the following ways:

[0124]

[0125]

[0126]

[0127]

[0128] In the formula, Represents coding features, Represents a 1×1 convolution. This represents a 3×3 dilated convolution with a dilation rate of 2. This represents a 3×3 dilated convolution with an expansion rate of 4. This represents a 3×3 dilated convolution with a dilation rate of 6. , , and These represent the four branch features respectively;

[0129] In step S3-2, the multi-scale fusion features can be obtained in the following way:

[0130]

[0131] In the formula, Indicates multi-scale fusion features, Represents the ReLU activation function. Indicates batch normalization. This indicates a channel dimension splicing operation.

[0132] In this implementation, the specific dilation rates (2, 4, 6) are defined. Dilated convolutions with different dilation rates can exponentially expand the receptive field without increasing the number of parameters, enabling the model to perceive contextual information at different scales. In the fusion formula, channel concatenation can preserve all multi-scale information of each branch. Combined with ReLU activation to introduce sparsity, the final output can contain fusion features with rich multi-scale semantic information, improving the model's adaptability to small targets and complex scenes.

[0133] The technical solution of the present invention will be further described below with reference to an exemplary embodiment.

[0134] I. Description of this exemplary implementation.

[0135] Colorectal cancer is one of the most common malignant tumors worldwide, and its disease burden continues to increase. Most colorectal cancers develop gradually from an adenoma to a carcinoma. Therefore, timely detection and accurate identification of polyps during colonoscopy are crucial for reducing missed diagnoses and optimizing clinical decision-making. In recent years, the application of deep learning in colonoscopy image analysis has been expanding, showing great potential, especially in polyp detection, classification, and segmentation. However, real-world colonoscopy images are often accompanied by complex interferences such as mucus, reflections, bubbles, folds, and differences in cleanliness. Furthermore, polyps often have blurred boundaries, irregular shapes, and large scale ranges, posing challenges to the robustness and generalization of existing models.

[0136] U-KAN combines the nonlinear expressive power of KAN with the hierarchical features of the U-shaped structure, providing a new approach for medical image segmentation. However, its ability to perform feature selection, boundary representation, and multi-scale modeling in complex colonoscopy scenarios still has room for improvement.

[0137] Based on this, the present invention proposes the DAR-UKAN segmentation network, which introduces a deep separable feature extraction module with dual attention enhancement at the encoding end, designs an edge-guided learnable residual attention fusion module in the boundary optimization stage, and adds an ASPP multi-scale context modeling unit in the skip connections.

[0138] Furthermore, in this exemplary embodiment, the performance and generalization ability of the model of the present invention were verified by combining four public datasets with real-world data from two hospitals (the First Affiliated Hospital of Shihezi University and the General Hospital of the Third Division of Xinjiang Production and Construction Corps).

[0139] II. Research Subjects and Methods

[0140] 1. Data sources and inclusion criteria.

[0141] This exemplary implementation incorporates four publicly available colorectal polyp segmentation datasets (Kvasir-SEG, CVC-ClinicDB, ETIS-LaribPolypDB, and CVC-300), with all images accompanied by expert-annotated pixel-level segmentation masks. Kvasir-SEG contains 1000 images, offering rich image resolution and wide applicability; CVC-ClinicDB contains 612 images derived from colonoscopy video frames; ETIS-LaribPolypDB contains 196 images, including numerous small polyps and complex scenes; and CVC-300 contains 60 images, commonly used to test the model's cross-dataset generalization ability.

[0142] In addition, 405 images of colonic polyps were collected from two hospitals, all annotated by senior physicians. This exemplary implementation excluded patients with inflammatory bowel disease, active colitis, coagulation disorders, familial adenomatous polyposis, and those undergoing emergency colonoscopy. Colonoscopy procedures involved multiple brands of endoscopy systems to enhance the heterogeneity of the data sources.

[0143] All images were anonymized. To improve the model's generalization ability and clinical applicability, images covering different intestinal cleanliness levels and imaging conditions were included.

[0144] 2. DAR-UKAN network structure design.

[0145] 2.1 Dual attention enhanced depth-separable ConvNeXt module.

[0146] In the Tok-KAN basic structure of U-KAN, the original feature extraction unit mainly relies on depthwise convolution to complete spatial modeling, which has limited ability to effectively filter features in complex colonoscopy images. To address the problems of large scale differences in polyps, complex textures, and strong background interference, this invention improves the basic feature extraction unit, such as... Figure 2 As shown, a dual-attention enhanced deep separable Tok-KAN module (DA-Tok-KAN) is proposed. Its core is the dual-attention deep separable ConvNeXt unit, which introduces channel attention and spatial attention while preserving the efficiency of local feature modeling, in order to enhance polyp-related representations and suppress background noise.

[0147] exist Figure 2 middle, Figure 2 a is the structural diagram of DA-Tok-KAN. Figure 2In diagram 'a', the modules from left to right represent: input feature map (i.e., the two-dimensional feature representation output by the previous network layer), tokenization (i.e., dividing and flattening the two-dimensional feature map into a series of feature blocks and mapping them to a token sequence), KAN Layer (i.e., the Kolmogorov-Arnold Network layer, which achieves non-linear feature mapping through learnable unary functions set on the network connection edges), DA DwConv (i.e., deep convolution with dual attention mechanism, which adaptively enhances local features through channel attention and spatial attention), and LayerNorm (i.e., layer normalization, used to normalize features and improve the stability of network training). Among these, the core role of tokenization is to convert convolutional features into token representations suitable for further processing by KAN. Figure 2 b is a structural diagram of the KAN Layer. Figure 2 In b, the meanings of each character are as follows: Z represents the input feature (input token feature) of the KAN layer. Represents the first-layer KAN nonlinear mapping, This represents the second-layer KAN nonlinear mapping. KAN(Z) represents the third-layer KAN nonlinear mapping, and KAN(Z) represents the output feature obtained after the input feature Z has passed through multiple layers of KAN mapping; where each... Each is composed of multiple learnable unary activation functions located on the node connection edges. Instead of the fixed-node activation function in traditional MLPs, it is constructed using a different method.

[0148] The module first reshapes the input sequence into a two-dimensional feature map, extracts local spatial and channel information through multi-layer deep convolution and pointwise convolution, and completes nonlinear mapping and distribution calibration by combining activation functions and normalization.

[0149]

[0150] Based on this, the model further enhances the high-response features related to polyp texture and morphology from the channel dimension, and focuses on the lesion area from the spatial dimension, and finally obtains the optimized output features through residual fusion.

[0151]

[0152] Compared with the original single convolutional structure, this module can adaptively adjust the feature weights according to the local distribution of colonoscopy images, improve feature discriminativeness and anti-interference ability, and provide higher quality input for the nonlinear modeling of subsequent KAN layers.

[0153] The structure of DA DwConv is as follows: Figure 3 As shown, in Figure 3 The diagram illustrates the processing flow of the main stem convolutional network and the dual attention enhancement mechanism. Here, BatchNorm2d represents 2D batch normalization, Channel Attention represents channel attention, Spatial Attention represents spatial attention, and Flatten represents the flatten layer.

[0154] 2.2 Edge-guided learnable residual attention fusion module (ELRA).

[0155] To address the issues of blurred boundaries, low contrast, and susceptibility to interference from mucus, reflections, and air bubbles in colon polyp segmentation, this invention designs an Edge-Guided Learnable Residual Attention Fusion Module (ELRA) based on the edge-guided attention module. This module retains the basic framework of multi-branch edge extraction and spatial attention, and improves the adaptive fusion capability under different image scenarios by introducing learnable residual weights.

[0156] Please see Figure 4 , Figure 4 A structural diagram of the ELRA module is shown. Figure 4In this text, the meanings of each character are as follows: EdgeFeature represents edge features, Input Feature represents input features, Prediction represents prediction results, BilinearInterpolation represents bilinear interpolation, 1×H×W represents the height × width of a single channel, C×H×W represents the height × width of the C channel, Edge_input represents edge input, Edge_guided Feature represents edge guided features, Background_x represents background feature input, 1-pred Background Interpolation represents 1-predicted background interpolation (backward predicted background interpolation), Background Feature represents background features, Laplacian Pyramid (make_laplace) represents the Laplacian pyramid (generating a Laplacian graph), Boundary Feature represents boundary features, Pred-feature represents predicted features, Concat (Along channel) represents concatenation (along the channel direction), 3C×H×W represents 3 times the number of channels × height × width, Fusion Conv Block represents the fusion convolutional module, and Conv 2d(3c→c) represents the fusion convolutional module. 3×3 represents two-dimensional convolution (3c channels reduced to c channels, 3×3 kernel), BatchNorm represents batch normalization, ReLU represents modified linear unit activation function, Spatial Attention represents spatial attention, Sigmoid represents Sigmoid activation function, Learnable ResidualWeight represents learnable residual weight, a Learnable represents learnable parameter a, 1-a represents 1 minus parameter a, Element-wise Multiplication represents element-wise multiplication, and Element-wise Addition represents element-wise addition.

[0157] The ELRA module receives three types of input: encoder backbone features x, external edge guidance features E, and preliminary segmentation prediction map P. The module performs background suppression, high-frequency edge information extraction, and contour guidance through parallel branches, and then obtains enhanced features focused on the polyp region through channel concatenation, convolutional fusion, and spatial attention weighting.

[0158] In the output stage, the original EGA module uses fixed coefficients to complete the residual connections, which is difficult to adapt to different image content. This invention introduces a trainable parameter α to construct a learnable residual structure, so as to achieve a dynamic balance between enhanced features and original features.

[0159]

[0160] In the feature fusion output stage, the original EGA module uses fixed coefficients to complete the residual connections, and the weights cannot dynamically change with the training process and image scene. This invention improves upon this by introducing a trainable parameter α to construct a learnable residual structure. The final output of the module is calculated as follows:

[0161]

[0162] During inference, the learnable residual mechanism can adjust the feature fusion ratio according to the local image content: in areas with blurred boundaries or discontinuous contours, it emphasizes edge enhancement information; in noisy areas such as highlights and slime, it retains more backbone features, thus balancing boundary characterization and background robustness.

[0163] 2.3. Hollow Space Pyramid Pooling Module (ASPP).

[0164] The original U-KAN skip connections use a direct splicing method, which lacks sufficient multi-scale context modeling capabilities. Considering the large size range of polyps and the complexity of local noise in colonoscopy images, single-scale feature fusion is difficult to take into account both the details of small polyps and the overall outline of large lesions. Therefore, this invention introduces a voided spatial pyramid pooling (ASPP) module at the skip connections.

[0165] Please see Figure 5 , Figure 5 This is a structural diagram of the ASPP module. Figure 5 In this context, Dilation represents the dilation rate, BatchNorm represents batch normalization, and ReLU represents the modified linear unit activation function.

[0166] The ASPP module constructs a multi-receptive-field feature extractor by using parallel 1×1 convolutions and 3×3 dilated convolutions with different dilation rates (2, 4, 6) to simultaneously model local details and global context.

[0167]

[0168]

[0169]

[0170]

[0171] After the features of each branch are concatenated along the channel dimension, they are fused and output after batch normalization and ReLU activation.

[0172]

[0173] This module helps improve the model's adaptability to polyps of different sizes, especially in scenarios with small targets and weak boundaries, where it can improve segmentation integrity and robustness.

[0174] 3. Experimental parameter settings.

[0175] All models were trained on an NVIDIA RTX 4090 GPU using the PyTorch framework. The training batch size was set to 32, the number of training epochs to 200, and the initial learning rate was [missing value]. The optimizer used was Adam, with 4 data loading threads. To standardize the input scale, all images were scaled to 256×256, and data augmentation strategies such as random flipping and rotation were employed.

[0176] Dice, IoU, Precision, Recall, and Accuracy are used as performance metrics for model segmentation. Among them, Dice and IoU reflect the degree of overlap between the predicted region and the ground truth labeled region, and are the core metrics for the segmentation task; Precision reflects the reliability of the predicted lesion pixels, Recall reflects the completeness of the lesion region capture, and Accuracy evaluates the overall pixel classification accuracy.

[0177] The calculation formulas for each indicator are as follows:

[0178]

[0179]

[0180]

[0181]

[0182]

[0183] In the formula, This represents the number of true instance pixels, that is, the number of pixels in both the predicted result and the true label that are in the target region. This represents the number of false positives, that is, the number of pixels that were mistakenly identified as background or target areas. This represents the number of false negative pixels, that is, the number of pixels in the target area that were mistakenly identified as background. The number of true negative pixels represents the number of pixels where both the predicted result and the true label are background. Among them, the Dice coefficient is used to measure the degree of overlap between the predicted segmented region and the true labeled region, and IoU reflects the ratio of their intersection to their union. These two are the most core evaluation metrics in image segmentation tasks. Precision measures the reliability of the model's prediction of the target region, Recall reflects the model's ability to completely capture the target region, and Accuracy is the overall prediction accuracy of all pixels. The values ​​of the above five metrics are all in the range of [0,1]. The closer the value is to 1, the better the model's segmentation effect.

[0184] III. Results.

[0185] The training and test sets for the four public datasets are shown in Table 1 below.

[0186] Table 1. Training and Test Set Allocation Table

[0187]

[0188] To evaluate the overall segmentation performance of DAR-UKAN, U-Net, Attention U-Net, TransUNet, and baseline U-KAN were selected as control models, and the results are shown in Table 2 below.

[0189] Table 2 Comparison of experimental results

[0190]

[0191] To further evaluate the model's generalization ability on real-world data, this invention included 405 colonoscopy images from two hospitals for external validation. These images were divided into a training set of 324 images and a test set of 81 images in an 8:2 ratio, and the same preprocessing and testing procedures as the publicly available dataset were used. The results are shown in Table 3 below.

[0192] Table 3. Statistics of Dataset Generalization Validation Experiment

[0193]

[0194] In Table 3, Private_Dataset represents data from two hospitals (the First Affiliated Hospital of Shihezi University and the General Hospital of the Third Division of Xinjiang Production and Construction Corps).

[0195] Experimental results combining publicly available datasets and real-world data from two hospitals demonstrate that DAR-UKAN possesses certain advantages in boundary characterization, robustness to interference, and multi-scale modeling. Specifically,

[0196] First, for the CVC-300 dataset, the polyp boundaries are relatively blurred and the contour transitions are gentle, requiring a high level of boundary representation capability. DAR-UKAN outperforms the control model in both Dice and IoU on this dataset, suggesting that the ELRA module can help improve the segmentation performance in weakly boundary regions.

[0197] Second, for Kvasir-SEG, the dataset features diverse scenes and significant differences in polyp scales. DAR-UKAN achieved superior overall performance on this dataset, demonstrating that dual-attention feature extraction and ASPP multi-scale modeling can enhance the model's adaptability to complex scenes.

[0198] Third, for CVC-ClinicDB, this dataset contains a lot of mucus, highlight, and wrinkle interference. The performance improvement of DAR-UKAN on this dataset suggests that the dual attention mechanism has a positive effect on suppressing background noise and enhancing lesion-related features.

[0199] Fourth, for ETIS-LaribPolypDB, this dataset mainly consists of small polyps and high-difficulty samples. The improvements of DAR-UKAN in Recall, Dice, and IoU demonstrate that the ASPP module has practical value in preserving features of small targets and modeling multi-scale receptive fields.

[0200] Overall, DAR-UKAN outperforms the control model in core metrics on public datasets, indicating that the three proposed improvement modules have a synergistic effect in feature extraction, boundary optimization, and multi-scale fusion.

[0201] In real-world data from both hospitals, DAR-UKAN achieved optimal accuracy, dice, IoU, and recall, while its precision was slightly lower than some control models. This suggests that while improving the completeness of lesion detection, the model may introduce a small number of additional false positives, but overall it achieves a good balance between sensitivity and specificity.

[0202] From a clinical application perspective, a higher recall helps reduce missed segmentation of lesions, especially suitable for scenarios with weak boundaries and small polyps; at the same time, the precision did not show a significant decrease, indicating that the model maintained an acceptable level of false detection while enhancing its detection capabilities.

[0203] From the perspective of process data processing, the image processing process data in public datasets can be found in [reference needed]. Figure 6 For details on image processing of private datasets, please refer to [link / reference]. Figure 7 .exist Figure 6 and Figure 7 In this text, each character has the following meaning: OriginalImage represents the original image, UNet represents the U-Net model, Att-UNet represents the Attention U-Net, TransUNet represents the TransUNet model, UKAN represents the UKAN model, Ours represents the method of this invention, and GT represents the ground truth annotation.

[0204] Overall, the proposed DAR-UKAN colon polyp segmentation method was validated on four public datasets and two real-world datasets from hospitals. The results show that the method demonstrates good overall performance in key metrics such as Dice, IoU, and Recall, particularly showing advantages in scenarios with weak boundaries, complex noise, and small targets. DAR-UKAN can serve as a candidate method for automatic colonoscopy image segmentation, providing a reference for subsequent computer-aided diagnosis and treatment research.

[0205] According to a second aspect of the present invention, a colonoscopy polyp segmentation system based on an improved KAN network is also provided, for executing the colonoscopy polyp segmentation method based on an improved KAN network according to any of the technical solutions in the first aspect of the present invention. The system includes an encoder, a skip connection module, a decoder, an edge-guided learnable residual attention fusion module, and an output module.

[0206] The encoder contains multiple dual-attention enhanced deep separable Tok-KAN modules for extracting dual-attention enhanced deep separable convolutional features from colonoscopy images to be segmented, and outputting encoded features;

[0207] The skip connection module is located between the encoder and the decoder and includes a hole spatial pyramid pooling module, which is used to perform multi-scale context modeling of encoded features to obtain multi-scale fused features.

[0208] The decoder is used to decode multi-scale fused features;

[0209] The edge-guided learnable residual attention fusion module is set in the decoder's decoding path to fuse the enhanced features with the decoder's output;

[0210] The output module is used to output the segmentation prediction results.

[0211] Through this implementation, firstly, the system of the present invention enables end-to-end automated polyp segmentation, improving the operability of clinical applications. Specifically, the encoder receives the colonoscopy image to be segmented and sequentially transmits it to the decoder via a skip connection module, and finally the output module directly outputs the segmentation prediction result. The entire system forms an end-to-end processing link from image input to segmentation result output, eliminating the need for manual design of intermediate features or additional post-processing interventions. This effectively lowers the technical threshold for clinical operation and improves the deployability and ease of use of the system in actual colonoscopy examination scenarios.

[0212] Second, the components of the system of this invention are connected sequentially according to the data flow order: the encoded features output by the encoder are transmitted to the decoder via a skip connection module. During the transmission process, the skip connection module utilizes the ASPP module to complete multi-scale fusion, while the edge-guided learnable residual attention fusion module performs boundary enhancement in the decoding path. The high-quality features output by the encoder provide a better input basis for the multi-scale sampling of the ASPP module, and the multi-scale fused features output by the ASPP module can further enhance the accuracy of the boundary enhancement of the ELRA module. The three components form a tightly coupled processing link with the data flow as the link, so that the feature selection, multi-scale perception and boundary optimization stages generate synergistic gains, jointly improving the segmentation accuracy and robustness of the system of this invention in complex colonoscopy scenarios.

[0213] Third, the encoder, skip connection module, decoder, edge-guided learnable residual attention fusion module, and output module are all independent modular components, and data is transferred between modules through standardized feature interfaces. This allows for targeted adjustments when a specific step needs optimization (such as replacing ASPP modules with different parameter settings or adjusting the number of modules stacked in the encoder), while keeping other modules unchanged. This eliminates the need to redesign the entire system architecture, effectively improving the maintainability and scalability of the system and providing an architectural foundation for subsequent system customization and iterative upgrades for different clinical applications.

[0214] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A colonoscopy polyp segmentation method based on an improved KAN network, characterized in that, Includes the following steps: Step S1: Obtain the colonoscopy image to be segmented; Step S2: Input the colonoscopy image into the encoder, which contains multiple dual-attention enhanced deep separable Tok-KAN modules. The dual-attention enhanced deep separable Tok-KAN modules are used to extract dual-attention enhanced deep separable convolutional features from the input features and output encoded features. Step S3: The encoded features are passed to the decoder through skip connections. A hollow spatial pyramid pooling module is set in the skip connections to perform multi-scale context modeling on the encoded features and obtain multi-scale fused features. Step S4: Input the multi-scale fusion features into the decoder, and during the decoding process, fuse the enhanced features of the edge-guided learnable residual attention fusion module with the output of the decoder to obtain the segmentation prediction result; Step S5: Output the segmentation prediction result.

2. The colonoscopy polyp segmentation method based on an improved KAN network according to claim 1, characterized in that, In step S2, the dual-attention enhanced deep separable Tok-KAN module performs dual-attention enhanced deep separable convolution feature extraction on the input features, specifically including: Step S2-1: Reshape the input sequence into a two-dimensional feature map; Step S2-2: Perform 7×7 depthwise convolution, pointwise convolution, activation function processing, pointwise convolution and batch normalization on the two-dimensional feature map in sequence to obtain the first convolutional feature; Step S2-3: Apply channel attention weighting and spatial attention weighting to the first convolutional feature sequentially to obtain dual attention enhanced features; Step S2-4: Add the residuals of the two-dimensional feature map and the dual attention enhancement feature and reshape it into a sequence form, which is used as the output feature of the dual attention enhancement depth separable Tok-KAN module.

3. The colonoscopy polyp segmentation method based on an improved KAN network according to claim 2, characterized in that, In step S2-2, the first convolutional feature is obtained in the following manner: ； In the formula, Represents the first convolutional feature. This represents the two-dimensional feature map. This represents a 7×7 depthwise convolution. This represents pointwise convolution. This represents the GELU activation function. This indicates batch normalization.

4. The colonoscopy polyp segmentation method based on an improved KAN network according to claim 2, characterized in that, In steps S2-4, the output features of the dual-attention enhanced depth-separable Tok-KAN module are obtained as follows: ； In the formula, This represents the output features of the dual-attention enhanced depthwise separable Tok-KAN module. This represents the operation of reshaping a two-dimensional feature map into a sequence. This represents the two-dimensional feature map. This represents spatial attention operations. This indicates channel attention operations. This represents the first convolutional feature.

5. The colonoscopy polyp segmentation method based on an improved KAN network according to claim 1, characterized in that, In step S4, the processing procedure of the edge-guided learnable residual attention fusion module specifically includes: Step S4-1: Receive encoder backbone features, external edge guiding features, and preliminary segmentation prediction map as input; Step S4-2: Perform background suppression, high-frequency edge information extraction, and contour guidance processing on the input of step S4-1 through parallel branches to obtain multi-path processing features; Step S4-3: After concatenating the multi-channel processed features, perform convolutional fusion and spatial attention weighting in sequence to obtain enhanced features focused on the polyp region; Step S4-4: Perform weighted residual fusion of the enhanced features obtained in step S4-3 and the encoder backbone features using trainable parameters to obtain the output features of the edge-guided learnable residual attention fusion module.

6. The colonoscopy polyp segmentation method based on an improved KAN network according to claim 5, characterized in that, In step S4-4, the output features of the edge-guided learnable residual attention fusion module are obtained in the following manner: ； In the formula, This indicates that the edge-guided learnable residual attention fusion module can output features. Indicates trainable parameters, This represents the enhanced features obtained in step S4-3. This represents the encoder backbone features.

7. The colonoscopy polyp segmentation method based on an improved KAN network according to claim 5, characterized in that, In step S4-4, before performing weighted residual fusion of the enhanced features and the encoder backbone features, the enhanced features are further subjected to layer normalization processing: ； In the formula, The enhanced features are represented by the normalized layer. This represents the enhanced features obtained in step S4-3. Presentation layer normalization operation; And Weighted residual fusion is performed with the encoder backbone features.

8. The colonoscopy polyp segmentation method based on an improved KAN network according to claim 1, characterized in that, In step S3, the hollow spatial pyramid pooling module performs multi-scale context modeling on the encoded features, specifically including: Step S3-1: Input the encoded features into four parallel branches respectively. The four parallel branches are a 1×1 convolution branch, a 3×3 dilated convolution branch with a dilation rate of 2, a 3×3 dilated convolution branch with a dilation rate of 4, and a 3×3 dilated convolution branch with a dilation rate of 6, to obtain four branch features respectively. Step S3-2: After concatenating the four branch features along the channel dimension, perform batch normalization and ReLU activation to obtain multi-scale fusion features.

9. The colonoscopy polyp segmentation method based on an improved KAN network according to claim 8, characterized in that, In step S3-1, the four branch features are obtained in the following ways: ； ； ； ； In the formula, Represents coding features, Represents a 1×1 convolution. This represents a 3×3 dilated convolution with a dilation rate of 2. This represents a 3×3 dilated convolution with an expansion rate of 4. This represents a 3×3 dilated convolution with a dilation rate of 6. , , and These represent the four branch features respectively; In step S3-2, the multi-scale fusion features are obtained in the following manner: ； In the formula, This represents the multi-scale fusion feature. Represents the ReLU activation function. Indicates batch normalization. This indicates a channel dimension splicing operation.

10. A colonoscopy polyp segmentation system based on an improved KAN network, characterized in that, The system is used to perform the colonoscopy polyp segmentation method based on the improved KAN network according to any one of claims 1-9, the system comprising: The encoder contains multiple dual-attention enhanced deep separable Tok-KAN modules for extracting dual-attention enhanced deep separable convolutional features from colonoscopy images to be segmented, and outputs encoded features. The skip connection module, located between the encoder and the decoder, includes a hole spatial pyramid pooling module for multi-scale context modeling of the encoded features to obtain multi-scale fused features; A decoder is used to decode the multi-scale fused features; An edge-guided learnable residual attention fusion module is set in the decoding path of the decoder to fuse enhanced features with the output of the decoder; The output module is used to output the segmentation prediction results.

Images