A Frequency-Guided Polyp Segmentation Method Based on Adaptive Receptive Field

Abstract

The application discloses a kind of frequency guidance polyp segmentation methods based on adaptive receptive field, comprising: the original endoscope image is input into main network PVTv2 and extracts four-stage features;Lightweight DWSKAN middle layer semantic enhancement module enhances the second, third stage feature, Token-KAN deep layer semantic enhancement module enhances the fourth stage feature;Enhance the first stage feature;Adaptive receptive field module ARF processes the four-stage feature of enhancement, obtains adaptive receptive field enhanced polyp feature map, constructs cross-scale frequency guidance fusion module GFF and processes adaptive receptive field enhanced polyp feature map, obtains the final polyp segmentation result.The application is processed by lightweight DWSKAN middle layer semantic enhancement module, and the second, third stage feature output by main network PVTv2 is processed, the joint modeling of space and channel dimension is realized, and the nonlinear expression capability of polyp boundary feature is improved.

Description

Technical Field

[0001] This invention belongs to the field of endoscopic polyp segmentation technology, specifically relating to a frequency-guided polyp segmentation method based on adaptive receptive fields. Background Technology

[0002] Colonic polyps are a collective term for all raised lesions on the colonic mucosa and are widely recognized clinically as precancerous lesions of colorectal cancer (CRC). Colonoscopy, currently the most direct and effective diagnostic method, plays a crucial role in the early detection and removal of polyps, reducing colorectal cancer-related mortality. Besides the issue of missed diagnoses, colonoscopy also presents challenges related to the accuracy of morphological description. Doctors not only need to know whether polyps are present but also need to determine their benign or malignant nature based on their shape, thus deciding whether removal is necessary. Therefore, polyp segmentation plays a vital role in colonoscopy.

[0003] Traditional endoscopic polyp segmentation techniques primarily rely on manual extraction of low-level features such as color and texture, combined with traditional image processing algorithms (e.g., edge detection, region growing). However, these methods exhibit poor robustness in complex scenarios and struggle to adapt to variations in equipment or individual patient differences. Current deep learning-based endoscopic polyp segmentation techniques mainly include methods based on convolutional neural networks (e.g., U-Net, PraNet) and Transformer structures (e.g., Polyp-PVT, HSNet). These methods employ an encoder-decoder structure, achieving multi-scale information fusion through multi-layer feature extraction to improve the accuracy of polyp segmentation.

[0004] Although existing methods have improved polyp segmentation performance to some extent, the following technical bottlenecks and shortcomings still exist:

[0005] (1) Fixed receptive field size cannot adaptively adapt to boundary and region information: In polyp segmentation, expanding the receptive field allows the model to acquire a wider range of contextual information. However, existing polyp segmentation methods use a fixed dilation rate to expand the receptive field. This fixed-size receptive field cannot adaptively adjust to the feature differences of different regions in the polyp image. The polyp boundary belongs to a high-frequency detail region, which requires a smaller receptive field to capture fine edge textures; while the internal region of the polyp is mostly low-frequency, which requires a larger receptive field to cover complete region information. A fixed receptive field cannot simultaneously meet these two needs, often resulting in blurred boundaries or missing internal regions in the segmentation results, failing to accurately depict the true shape of the polyp. In other words, if the fixed dilation rate is too large, although the receptive field is expanded, the high-frequency detail information of the polyp boundary may be lost due to the reduced response frequency and effective bandwidth of the convolution kernel, making the segmentation results unable to accurately reflect the true boundary of the polyp; if the fixed dilation rate is too small, the advantage of a large receptive field cannot be fully utilized to acquire sufficient contextual information, which also affects the segmentation accuracy.

[0006] (2) Transformer models lack the ability to model local details and have weak edge perception: Transformer-based segmentation models are good at extracting global semantics and long-distance dependencies through self-attention mechanisms, but they often lack prior modeling of local details. In polyp segmentation tasks, this limitation manifests as insensitivity to small structures and edge changes, weakening the ability to perceive polyp edges, resulting in blurred and jagged segmentation boundaries. This weak edge modeling ability makes Transformer-type models prone to inaccurate segmentation when dealing with polyps with complex and variable boundaries.

[0007] (3) Insufficient suppression of high-frequency noise and significant background interference: In endoscopic polyp images, intestinal folds, light spot reflections, residues, and other background structures often form high-frequency texture noise. Existing polyp segmentation methods lack effective high-frequency noise suppression mechanisms, easily misclassifying these non-polyp edge textures as part of the polyp. High-frequency interference in the background significantly affects segmentation performance, manifested as an increase in false positives, i.e., incorrectly segmenting background noise as polyps. This high-frequency background interference not only reduces the accuracy and reliability of the segmentation results but may also lead to misjudgments by doctors. For example, when the contrast between the polyp and the surrounding tissue is low, the model is more easily misled by background noise, thus producing incorrect segmentation in non-lesion areas.

[0008] (4) Existing feature fusion strategies only process the spatial domain, making it difficult to fully exploit the complementarity of high and low frequency information: Most mainstream polyp segmentation models currently adopt spatial domain feature fusion strategies, such as skip connections, channel splicing, or simple weighted fusion. These fusion methods only integrate information in the spatial dimension, ignoring the structural features and texture details contained in the image itself in the frequency domain. In fact, key diagnostic elements such as the edge contour and protrusion morphology of polyps are high-frequency signals in the frequency domain, while their shape contour and internal structure mainly correspond to low-frequency components. In methods based solely on spatial domain operations, high-frequency and low-frequency features often interfere with each other, resulting in problems such as blurred edges, broken contours, or missegmentation of the background. Especially when processing polyp images with complex structures, low contrast, or similar textures, the lack of information separation and enhancement mechanisms in the frequency dimension makes it difficult to achieve differentiated modeling of edge and region information, ultimately affecting the accuracy and robustness of the segmentation results. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a frequency-guided polyp segmentation method based on adaptive receptive fields, which aims to solve the problems in the background technology.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a frequency-guided polyp segmentation method based on adaptive receptive fields, comprising:

[0011] Step S1: Input the original endoscopic image into the backbone network PVTv2 for processing to obtain the first-stage polyp feature map, the second-stage polyp feature map, the third-stage polyp feature map and the fourth-stage polyp feature map respectively;

[0012] Step S2: The second-stage polyp feature map and the third-stage polyp feature map are processed by the lightweight DWSKAN mid-layer semantic enhancement module to obtain the second-stage semantic enhancement feature map and the third-stage semantic enhancement feature map respectively.

[0013] Step S3: Process the fourth-stage polyp feature map using the Token-KAN deep semantic enhancement module to obtain the fourth-stage semantic enhancement feature map;

[0014] Step S4: Process the first-stage polyp feature map to obtain the first-stage semantic enhancement feature map; input the first-stage, second-stage, third-stage, and fourth-stage semantic enhancement feature maps into the adaptive receptive field module (ARF) for processing to obtain four-stage adaptive receptive field enhanced polyp feature maps; aggregate the four-stage adaptive receptive field enhanced polyp feature maps to obtain aggregated features; process the aggregated features through a lightweight convolutional prediction branch to obtain the initial boundary prediction map and the initial polyp prediction map, respectively.

[0015] Step S5: Input the adaptive receptive field enhanced polyp feature maps from the four stages into the cross-scale frequency-guided fusion module GFF for frequency modeling and cross-layer guided feature fusion processing to obtain high-frequency boundary enhancement features. Use the high-frequency boundary enhancement features as guiding features along with the initial boundary prediction map and the initial polyp prediction map. Figure 1 The data is then input into a dual-supervised decoder for both the boundary and foreground regions to obtain the final polyp segmentation result.

[0016] Furthermore, the process of inputting the raw endoscopic images into the backbone network PVTv2 for processing is as follows:

[0017] The original endoscopic images were input into the backbone network PVTv2 for shallow-to-deep extraction, resulting in polyp feature maps at four stages, denoted as follows: ;in This is a characteristic image of a polyp in the first stage; This is a characteristic image of a second-stage polyp. This is a characteristic diagram of a third-stage polyp. This is a characteristic diagram of a fourth-stage polyp.

[0018] Furthermore, the processing procedure of the lightweight DWSKAN mid-layer semantic enhancement module is as follows:

[0019] First, the second-stage polyp feature map is compressed using a 1×1 convolution to reduce the number of channels. and the characteristic diagram of stage III polyps Processing was performed to obtain the following results: and ,express:

[0020] ;

[0021] ;

[0022] In the formula, For the second stage of embedding feature maps; For the third stage of embedding feature maps; It is a 1×1 convolution;

[0023] The lightweight DWSKAN mid-layer semantic enhancement module introduces KAN linear layer pairs based on Kolmogorov-Arnold network theory. and Perform point-by-point nonlinear mapping to obtain the following results: and ,express:

[0024] ;

[0025] ;

[0026] In the formula, For the second stage of embedding feature maps, the first The spatial location of each sample Enhanced eigenvalues ​​at the location; This represents the total number of non-linear activation functions; Let be the nth learnable weighting coefficient; Let be the nth learnable unary nonlinear activation function; This indicates retrieving all channels; For the third stage of embedding feature maps The spatial location of each sample Enhanced eigenvalues ​​at the location;

[0027] Will According to the original spatial location Recombined to form a complete second-stage enhanced feature map ;Will According to the original spatial location Recombined to form a complete third-stage enhanced feature map ;

[0028] right and Process them separately to obtain the results. and ,express:

[0029] ;

[0030] ;

[0031] In the formula, This is the output feature map after spatial modeling in the second stage; This is the output feature map after spatial modeling in the third stage; For depthwise separable KAN convolution;

[0032] Will and Perform residual join and apply normalization to obtain ;Will and Perform residual join and apply normalization to obtain ,express:

[0033] ;

[0034] ;

[0035] In the formula, This is the second-stage semantic enhancement feature map; For batch normalization; This is the semantic enhancement feature map for the third stage.

[0036] Furthermore, the processing procedure of the Token-KAN deep semantic enhancement module is as follows:

[0037] First, the characteristic diagram of the fourth-stage polyp is shown. Through convolution, batch normalization, and The activation function processes the data, converting it into a deep embedding feature map, which is represented as follows:

[0038] ;

[0039] In the formula, Represents deep embedding feature maps; This represents a two-dimensional convolution operation;

[0040] Deeply embedding feature maps Divided into multiple Patch size; Indicates the height of the deep embedding feature map; Indicates the width of the deep embedding feature map; This represents the number of channels in the deep embedding feature map; Represents the set of real numbers; This represents the dimensions of each patch in the height and width directions of the deep embedding feature map; a patch is a small block in the deep embedding feature map.

[0041] Flatten each patch into a one-dimensional vector, forming a total of One token; Indicates the total number of generated tokens;

[0042] No. Each token is All tokens are mapped through a learnable linear mapping. Projecting onto a D-dimensional vector space forms an embedding sequence ,express:

[0043] ;

[0044] In the formula, This represents the embedded sequence formed by linearly mapping all tokens; This represents the vector obtained after the first token is processed by a learnable linear mapping E; This represents the vector obtained after the Nth token is processed by a learnable linear mapping E;

[0045] The Token-KAN deep semantic enhancement module uses a three-layer KAN structure to... The input is processed through three KAN layers, with residual connections and layer normalization performed after each layer to obtain the result. KAN layer output ,express:

[0046] ;

[0047] In the formula, For layer normalization; For the first The output of the KAN layer; ; For the first KAN layer output; For the first KAN layer output; For the first KAN layer output;

[0048] From the The output of the KAN layer to the th The transformation of the KAN layer is represented as:

[0049] ;

[0050] In the formula, For the first The KAN layer uses a set of nonlinear mappings, i.e., a set of unary activation functions;

[0051] Composed of multiple unary activation functions The resulting transformation matrix will change the input dimension. Mapped to output dimension , used in total An independent, learnable unary activation function ,express:

[0052] ;

[0053] In the formula, For input channel dimension index; Output channel dimension index; For the first KAN layer input dimension; For the first The output dimension of the KAN layer;

[0054] exist Then, the KAN attention mechanism was introduced to capture contextual information, representing:

[0055] ;

[0056] In the formula, The token sequence is semantically enhanced. This refers to the KAN attention mechanism.

[0057] semantically enhanced token sequence Rearranged into spatial feature maps The fourth stage polyp characteristic map After upsampling, it is compared with the spatial feature map. Perform residual connections to obtain the fourth-stage semantic enhancement feature map. .

[0058] Furthermore, the feature map of the first-stage polyp is enhanced, specifically through the following process:

[0059] First, let's look at the characteristic diagram of the first-stage polyps. Using 1×1 convolution, batch normalization and The activation operation performs semantic enhancement processing to obtain the first-stage semantic enhancement feature map. ;

[0060] in, ; For the first Semantic enhancement of polyp feature map at each stage.

[0061] Furthermore, the processing procedure of the Adaptive Receptive Field (ARF) module is as follows:

[0062] For the first Semantic enhancement polyp feature map of the stage use Discrete wavelet transform decomposes into four frequency components, represented as follows:

[0063] ;

[0064] In the formula, Indicates the first The low-frequency components of the polyp feature map are semantically enhanced in the stage; , and They represent the first The semantically enhanced polyp feature map of the stage includes the horizontal high-frequency components, vertical high-frequency components, and diagonal high-frequency components. Represents the discrete wavelet transform;

[0065] Define each local location pixel The local frequency energy is represented by:

[0066] ;

[0067] In the formula, Indicates the first Semantic enhancement of polyp feature maps at different stages in location Local frequency energy at that location; , and They represent the first Semantic enhancement of polyp feature maps at different stages in location The horizontal high-frequency components, vertical high-frequency components, and diagonal high-frequency components at the location;

[0068] right Use average pooling to obtain the position. The smoothed local average frequency energy is represented as:

[0069] ;

[0070] In the formula, Indicates the first Stage position The local average frequency energy after smoothing; This indicates an average pooling operation;

[0071] based on Calculate the first The maximum local frequency energy across the entire map of the semantically enhanced polyp feature map of the stage is represented by:

[0072] ;

[0073] In the formula, Indicates the first The maximum local frequency energy across the entire map of the polyp feature map in the semantic enhancement stage; This indicates the operation of retrieving the maximum value;

[0074] based on Define two threshold parameters, representing:

[0075] ;

[0076] ;

[0077] In the formula, Indicates the first High-frequency thresholds in the semantic enhancement polyp feature map of the stage; Indicates the first Low-frequency thresholds in the semantic enhancement polyp feature map of the stage;

[0078] Select the corresponding expansion rate ,express:

[0079] ;

[0080] In the formula, For the first Semantic enhancement of polyp feature map location in stages The corresponding expansion rate;

[0081] Indicates position It belongs to the high-frequency region, and the convolution dilation rate is... ; Indicates position Belongs to the low-frequency region, convolution dilation rate ;

[0082] Indicates position It belongs to the transition region, and the convolution dilation rate is... ;

[0083] By using adaptive dynamic receptive field convolution operations, based on For the first Semantic enhancement polyp feature map of the stage Convolution processing is performed to obtain the adaptive receptive field enhanced polyp feature map of stage n. ; This represents the feature map of the polyp enhanced by the adaptive receptive field in the first stage. This represents the feature map of the polyp, representing the adaptive receptive field enhancement in the second stage. This represents the feature map of the third-stage adaptive receptive field enhancement polyp. This represents the feature map of the fourth stage of adaptive receptive field enhancement polyp.

[0084] polymerization The aggregated features are obtained and then processed through a lightweight convolutional prediction branch to obtain the initial boundary prediction map and the initial polyp prediction map, respectively.

[0085] Furthermore, the processing procedure of the cross-scale frequency-guided fusion module GFF is as follows:

[0086] For the first Stage-adaptive receptive field enhancement polyp feature map use The discrete wavelet transform decomposes into four components, representing:

[0087] ;

[0088] In the formula, Indicates the first The adaptive receptive field of the stage enhances the low-frequency components of the polyp feature map; , and They represent the first The adaptive receptive field of the stage enhances the horizontal, vertical, and diagonal high-frequency components of the polyp feature map;

[0089] Will , and Add and merge into a single first High-frequency feature map of the stage ,express:

[0090] ;

[0091] Based on the High-frequency feature map of the stage The gated weight matrix is ​​generated by 1×1 convolution and the Sigmoid activation function, and is represented as follows:

[0092] ;

[0093] In the formula, Represents the gate weight matrix; This represents the Sigmoid activation function; The weight matrix represents the weights of the 1×1 convolution kernel; Indicates the bias term; This is an element-wise multiplication operation;

[0094] For the first High-frequency feature map of the stage Through 1×1 convolution and Nonlinear transformation, to obtain the first High-frequency features after stage enhancement ,express:

[0095] ;

[0096] In the formula, express Activation function; The weight matrix represents the weights of the 1×1 convolution kernel; Indicates the bias term;

[0097] The first High-frequency features after stage enhancement Spatial attention and channel attention Enhancement is performed to obtain the first High-frequency features after staged attention enhancement ,express:

[0098] ;

[0099] Combined with gated weight matrix , for the High-frequency features after staged attention enhancement Perform point-by-point modulation to obtain the first... Stage High Frequency Boundary Enhancement Feature Map ,express:

[0100] ;

[0101] The first Stage High Frequency Boundary Enhancement Feature Map As guiding features, initial boundary prediction map, and initial polyp prediction Figure 1 The data is then input into a dual-supervised decoder for both the boundary and foreground regions to obtain the final polyp segmentation result.

[0102] Compared with existing technologies, the present invention has the following advantages:

[0103] (1) This invention proposes a lightweight DWSKAN mid-layer semantic enhancement module, which realizes joint modeling of spatial and channel dimensions for the mid-layer features output by the backbone network PVTv2, improves the nonlinear expression ability of polyp boundary features, and enhances the polyp segmentation method proposed in this invention to model the local structure and boundary details of polyps.

[0104] (2) The present invention proposes an adaptive receptive field module (ARF), which adaptively adjusts the convolution dilation rate of each spatial location in the polyp feature map based on the local frequency energy of each pixel in the polyp feature map. This enables the polyp segmentation method proposed in this invention to break through the traditional fixed dilation rate limitation, solve the problem that boundary details and regional semantics cannot be taken into account at the same time, improve the accuracy of polyp edge segmentation and the coherence of the entire polyp region, and improve the robustness of the adaptive receptive field module (ARF).

[0105] (3) To address the problem that existing feature fusion strategies are limited to spatial domain operations and ignore frequency dimension features, this invention designs a cross-scale frequency-guided fusion module (GFF). This module performs discrete wavelet decomposition on each layer of enhanced polyp feature map to obtain high-frequency and low-frequency components. The high-frequency components adaptively suppress high-frequency background noise (such as intestinal folds and light spots) through an attention mechanism, while simultaneously enhancing real edge information, achieving the effect of high-frequency guided edge segmentation. This significantly reduces background interference, improves the accuracy of real edge localization, and solves the problem of false detection caused by complex background noise in endoscopic images. Attached Figure Description

[0106] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0107] like Figure 1As shown, the present invention provides a technical solution: a frequency-guided polyp segmentation method based on adaptive receptive fields, comprising:

[0108] Step S1: Input the original endoscopic image into the backbone network PVTv2 for processing to obtain the first-stage polyp feature map, the second-stage polyp feature map, the third-stage polyp feature map and the fourth-stage polyp feature map respectively;

[0109] Step S2: The second-stage polyp feature map and the third-stage polyp feature map are processed by the lightweight DWSKAN mid-layer semantic enhancement module to obtain the second-stage semantic enhancement feature map and the third-stage semantic enhancement feature map respectively.

[0110] Step S3: Process the fourth-stage polyp feature map using the Token-KAN deep semantic enhancement module to obtain the fourth-stage semantic enhancement feature map;

[0111] Step S4: Process the first-stage polyp feature map to obtain the first-stage semantic enhancement feature map; input the first-stage, second-stage, third-stage, and fourth-stage semantic enhancement feature maps into the adaptive receptive field module (ARF) for processing to obtain four-stage adaptive receptive field enhanced polyp feature maps; aggregate the four-stage adaptive receptive field enhanced polyp feature maps to obtain aggregated features; process the aggregated features through a lightweight convolutional prediction branch to obtain the initial boundary prediction map and the initial polyp prediction map, respectively.

[0112] Step S5: Input the adaptive receptive field enhanced polyp feature maps from the four stages into the cross-scale frequency-guided fusion module GFF for frequency modeling and cross-layer guided feature fusion processing to obtain high-frequency boundary enhancement features. Use the high-frequency boundary enhancement features as guiding features along with the initial boundary prediction map and the initial polyp prediction map. Figure 1 The result is then fed into a Boundary and Foreground Dual-Supervised Decoder (BFDS Decoder) for processing to obtain the final polyp segmentation result.

[0113] The process of inputting the raw endoscopic images into the backbone network PVTv2 for processing is as follows:

[0114] The original endoscopic images were input into the backbone network PVTv2 for shallow-to-deep extraction, resulting in polyp feature maps at four stages, denoted as follows: ;in This is a characteristic image of a polyp in the first stage; This is a characteristic image of a second-stage polyp. This is a characteristic diagram of a third-stage polyp. This is the feature map of the fourth stage polyp; the feature maps of the four stages of polyps contain global semantic structure and location context information.

[0115] The processing procedure of the lightweight DWSKAN mid-layer semantic enhancement module is as follows:

[0116] This invention is based on the Kolmogorov-Arnold Network (KAN) and targets the mid-layer features output by the backbone network PVTv2. , We propose a lightweight DWSKAN mid-layer semantic enhancement module (Lite Depthwise-Separable Kolmogorov–Arnold Network Block) to enhance the perception and modeling capabilities of polyp edges;

[0117] First, the second-stage polyp feature map is compressed using a 1×1 convolution to reduce the number of channels. and the characteristic diagram of stage III polyps Processing was performed to obtain the following results: and ,express:

[0118] ;

[0119] ;

[0120] In the formula, For the second stage of embedding feature maps; For the third stage of embedding feature maps; It is a 1×1 convolution;

[0121] The lightweight DWSKAN mid-layer semantic enhancement module introduces KANLinear pairs based on Kolmogorov-Arnold network theory. and Perform point-by-point nonlinear mapping to obtain the following results: and ,express:

[0122] ;

[0123] ;

[0124] In the formula, For the second stage of embedding feature maps, the first The spatial location of each sample Enhanced eigenvalues ​​at the location; This represents the total number of non-linear activation functions; Let be the nth learnable weighting coefficient; Let be the nth learnable unary nonlinear activation function; This indicates retrieving all channels; For the third stage of embedding feature maps The spatial location of each sample Enhanced eigenvalues ​​at the location;

[0125] and The second and third stages of embedding feature maps are respectively the first... The spatial location of each sample The enhanced eigenvalues ​​at point ; for the ... One sample, According to the original spatial coordinates Recombined to form a complete second-stage enhanced feature map For the first One sample, According to the original spatial coordinates Recombined to form a complete third-stage enhanced feature map The recombination operation ensures the consistency of the features in spatial structure after pointwise nonlinear mapping, providing a complete and structured input for subsequent depthwise separable KAN convolution (DWS-KANConv);

[0126] To further enhance the modeling capability of spatial dimensions, a depthwise separable KAN convolution (DWS-KANConv) is proposed to achieve [the following is unclear due to incomplete sentence fragment: "to ... Highly efficient sensing and nonlinear enhancement, for and Process them separately to obtain the results. and ,express:

[0127] ;

[0128] ;

[0129] In the formula, This is the output feature map after spatial modeling in the second stage; This is the output feature map after spatial modeling in the third stage; For depthwise separable KAN convolution;

[0130] Will and Perform residual join and apply normalization to obtain ;Will and Perform residual join and apply normalization to obtain ,express:

[0131] ;

[0132] ;

[0133] In the formula, This is the second-stage semantic enhancement feature map; For batch normalization; This is the semantic enhancement feature map for the third stage.

[0134] The processing procedure of the Token-KAN deep semantic enhancement module is as follows:

[0135] To compensate for the shortcomings of the backbone network PVTv2 in local detail modeling, this invention introduces the Token-KAN deep semantic enhancement module to enhance the polyp feature map of the fourth stage. Enhancement is achieved by combining the global modeling capabilities of the backbone network PVTv2 with the nonlinear expression capabilities of KAN (based on Kolmogorov-Arnold network), thereby enabling more accurate modeling of complex polyp boundaries and micro polyps.

[0136] First, the characteristic diagram of the fourth-stage polyp is shown. Through convolution, batch normalization, and The activation function processes the data, converting it into a deep embedding feature map, which is represented as follows:

[0137] ;

[0138] In the formula, Represents deep embedding feature maps; This represents a two-dimensional convolution operation;

[0139] Deeply embedding feature maps Divided into multiple Patch size; Indicates the height of the deep embedding feature map; Indicates the width of the deep embedding feature map; This represents the number of channels in the deep embedding feature map; Represents the set of real numbers; This represents the dimensions of each patch in the height and width directions of the deep embedding feature map; a patch is a small block in the deep embedding feature map.

[0140] Flatten each patch into a one-dimensional vector, forming a total of One token; Indicates the total number of generated tokens;

[0141] No. Each token is All tokens are mapped through a learnable linear mapping. Projecting onto a D-dimensional vector space forms an embedding sequence ,express:

[0142] ;

[0143] In the formula, This represents the embedded sequence formed by linearly mapping all tokens; This indicates that the first token has undergone a learnable linear mapping. The resulting vector; This indicates that the Nth token has undergone a learnable linear mapping. The resulting vector;

[0144] The Token-KAN deep semantic enhancement module uses a three-layer KAN structure to... The input is processed through three KAN layers, with residual connections and layer normalization performed after each layer to obtain the result. KAN layer output ,express:

[0145] ;

[0146] In the formula, For layer normalization; For the first The output of the KAN layer; ; For the first KAN layer output; For the first KAN layer output; For the first KAN layer output;

[0147] From the The output of the KAN layer to the th The transformation of the KAN layer is represented as:

[0148] ;

[0149] In the formula, For the first The KAN layer uses a set of nonlinear mappings, i.e., a set of unary activation functions;

[0150] Composed of multiple unary activation functions The resulting transformation matrix will change the input dimension. Mapped to output dimension , used in total An independent, learnable unary activation function ,express:

[0151] ;

[0152] In the formula, For input channel dimension index; Output channel dimension index; For the first KAN layer input dimension; For the first The output dimension of the KAN layer;

[0153] exist Then, the KAN attention mechanism was introduced to capture contextual information, representing:

[0154] ;

[0155] In the formula, The token sequence is semantically enhanced. This refers to the KAN attention mechanism.

[0156] semantically enhanced token sequence Rearranged into spatial feature maps The fourth stage polyp characteristic map After upsampling, it is compared with the spatial feature map. Perform residual connections to obtain the fourth-stage semantic enhancement feature map. .

[0157] The specific process for enhancing the polyp feature map in the first stage is as follows:

[0158] First, let's look at the characteristic diagram of the first-stage polyps. Using 1×1 convolution, batch normalization and The activation operation performs semantic enhancement processing to obtain the first-stage semantic enhancement feature map. ;

[0159] in, ; For the first Semantic enhancement of polyp feature map at each stage.

[0160] The processing procedure of the Adaptive Receptive Field (ARF) module is as follows:

[0161] right The adaptive receptive field module (ARF) proposed in this invention is used to dynamically and adaptively adjust the receptive field according to the frequency characteristics of local regions of the feature map, so that the convolution kernel can be adaptively adjusted in regions with rich details (high frequency) and smooth regions (low frequency), thereby improving the ability of the adaptive receptive field module (ARF) to adaptively capture polyp edges and regional features in the feature map.

[0162] For the first Semantic enhancement polyp feature map of the stage use The discrete wavelet transform decomposes into the following four frequency components, represented as follows:

[0163] ;

[0164] In the formula, Indicates the first The low-frequency components of the polyp feature map are semantically enhanced in the stage; , and They represent the first The semantically enhanced polyp feature map of the stage includes the horizontal high-frequency components, vertical high-frequency components, and diagonal high-frequency components. Represents the discrete wavelet transform;

[0165] Define each local location pixel The local frequency energy is represented by:

[0166] ;

[0167] In the formula, Indicates the first Semantic enhancement of polyp feature maps at different stages in location Local frequency energy at that location; , and They represent the first Semantic enhancement of polyp feature maps at different stages in location The horizontal high-frequency components, vertical high-frequency components, and diagonal high-frequency components at the location;

[0168] right Use average pooling to obtain the position. The smoothed local average frequency energy is represented as:

[0169] ;

[0170] In the formula, Indicates the first Stage position The local average frequency energy after smoothing; This indicates an average pooling operation;

[0171] based on Calculate the first The maximum local frequency energy across the entire map of the semantically enhanced polyp feature map of the stage is represented by:

[0172] ;

[0173] In the formula, Indicates the first The maximum local frequency energy across the entire map of the polyp feature map in the semantic enhancement stage; This indicates the operation of retrieving the maximum value;

[0174] based on Define two threshold parameters, representing:

[0175] ;

[0176] ;

[0177] In the formula, Indicates the first High-frequency thresholds in the semantic enhancement polyp feature map of the stage; Indicates the first Low-frequency thresholds in the semantic enhancement polyp feature map of the stage;

[0178] Select the corresponding expansion rate ,express:

[0179] ;

[0180] In the formula, For the first Semantic enhancement of polyp feature map location in stages The corresponding expansion rate;

[0181] Indicates position This is a high-frequency region, so a smaller convolution dilation rate is used. ; Indicates position This falls within the low-frequency region, so a larger convolution dilation rate is used. ;and Indicates position It belongs to the transition region and uses a moderate convolution dilation rate. ;

[0182] In obtaining Then, adaptive dynamic receptive field convolution is performed, representing:

[0183] ;

[0184] In the formula, Indicates the first Stage position Adaptive receptive fields enhance polyp features; This represents an adaptive dynamic receptive field convolution operation; This represents the dilation rate of the convolution operation;

[0185] By using adaptive dynamic receptive field convolution operations, based on For the first Semantic enhancement polyp feature map of the stage Convolution processing is performed to obtain the adaptive receptive field enhanced polyp feature map of stage n. ; This represents the feature map of the polyp enhanced by the adaptive receptive field in the first stage. This represents the feature map of the polyp, representing the adaptive receptive field enhancement in the second stage. This represents the feature map of the third-stage adaptive receptive field enhancement polyp. This represents the feature map of the fourth stage of adaptive receptive field enhancement polyp.

[0186] polymerization The aggregated features are obtained and then processed through a lightweight convolutional prediction branch to obtain the initial boundary prediction map and the initial polyp prediction map, respectively.

[0187] The processing procedure of the cross-scale frequency-guided fusion module GFF is as follows:

[0188] For the first Stage-adaptive receptive field enhancement polyp feature map use The discrete wavelet transform decomposes into four components, representing:

[0189] ;

[0190] In the formula, Indicates the first The adaptive receptive field of the stage enhances the low-frequency components of the polyp feature map; , and They represent the first The adaptive receptive field of the stage enhances the horizontal, vertical, and diagonal high-frequency components of the polyp feature map;

[0191] Will , and Add and merge into a single first High-frequency feature map of the stage ,express:

[0192] ;

[0193] Based on the High-frequency feature map of the stage The gated weight matrix is ​​generated by 1×1 convolution and the Sigmoid activation function, and is represented as follows:

[0194] ;

[0195] In the formula, Represents the gate weight matrix; This represents the Sigmoid activation function; The weight matrix represents the weights of the 1×1 convolution kernel; Indicates the bias term; This is an element-wise multiplication operation;

[0196] For the first High-frequency feature map of the stage Through 1×1 convolution and Nonlinear transformation, to obtain the first High-frequency features after stage enhancement ,express:

[0197] ;

[0198] In the formula, express Activation function; The weight matrix represents the weights of the 1×1 convolution kernel; Indicates the bias term;

[0199] The first High-frequency features after stage enhancement Spatial attention and channel attention Further enhancements were made to obtain the first... High-frequency features after staged attention enhancement ,express:

[0200] ;

[0201] Combined with gated weight matrix , for the High-frequency features after staged attention enhancement Perform point-by-point modulation to obtain the first... Stage High Frequency Boundary Enhancement Feature Map ,express:

[0202] ;

[0203] The first Stage High Frequency Boundary Enhancement Feature Map As guiding features, initial boundary prediction map, and initial polyp prediction Figure 1 The data is then input into a dual-supervised decoder for both the boundary and foreground regions to obtain the final polyp segmentation result.

[0204] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A frequency-guided polyp segmentation method based on adaptive receptive field, characterized in that, Comprise: Step S1: input the original endoscope image into the backbone network PVTv2 for processing to obtain a first-stage polyp feature map, a second-stage polyp feature map, a third-stage polyp feature map and a fourth-stage polyp feature map respectively; Step S2: process the second-stage polyp feature map and the third-stage polyp feature map respectively through the lightweight DWSKAN middle-layer semantic enhancement module to obtain a second-stage semantic enhancement feature map and a third-stage semantic enhancement feature map respectively; Step S3: process the fourth-stage polyp feature map through the Token-KAN deep-layer semantic enhancement module to obtain a fourth-stage semantic enhancement feature map; Step S4: process the first-stage polyp feature map to obtain a first-stage semantic enhancement feature map; input the first-stage semantic enhancement feature map, the second-stage semantic enhancement feature map, the third-stage semantic enhancement feature map and the fourth-stage semantic enhancement feature map into the adaptive receptive field module ARF for processing to obtain adaptive receptive field enhanced polyp feature maps of the four stages; The processing process of the adaptive receptive field module ARF is: To the first Semantic enhanced polyp feature map of the stage Using The discrete wavelet transform decomposes into four frequency components, denoted: ； wherein denotes the low frequency component of the semantic enhanced polyp feature map of the i-th stage; denotes the low frequency component of the semantic enhanced polyp feature map of the i-th stage; , and denote the horizontal, vertical and diagonal high frequency components of the semantic enhanced polyp feature map of the i-th stage, respectively; denote the horizontal, vertical and diagonal high frequency components of the semantic enhanced polyp feature map of the i-th stage, respectively; denotes the discrete wavelet transform; defining a local position pixel of each local position the local frequency energy of the local position pixel, is represented by: ； wherein represents the semantic enhanced polyp feature map at the first stage; represents the local frequency energy of the semantic enhanced polyp feature map at the first stage at position ; , and respectively represent the horizontal high frequency component, the vertical high frequency component and the diagonal high frequency component of the semantic enhanced polyp feature map at the first stage at position ; ​ For Using an average pooling operation, the smoothed local average frequency energy at position is obtained, denoted by: ； wherein denotes the phase position smoothed local average frequency energy at the denotes an average pooling operation; based on computing the maximum local frequency energy of the semantic enhanced polyp feature map across the full image range of the stage, denoted as: computing the maximum local frequency energy of the semantic enhanced polyp feature map across the full image range of the stage, denoted as: ； wherein represents the maximum local frequency energy within the full image range of the semantic enhanced polyp feature map of the stage represents the maximum local frequency energy within the full image range of the semantic enhanced polyp feature map of the stage represents a max operation based on two threshold parameters are defined, denoted: ； ； In the formula, represents the high frequency threshold in the semantic enhanced polyp feature map of the stage; represents the low frequency threshold in the semantic enhanced polyp feature map of the stage; Selecting a corresponding expansion rate , means: ； In the formula, is the first stage semantic enhanced polyp feature map position corresponding to the inflation rate; representative position belongs to the high frequency region, convolution expansion rate ; representative position belongs to the low frequency region, convolution expansion rate ; representative position belongs to in the transition region, the convolution expansion rate ; By using adaptive dynamic receptive field convolution operations, based on For the first Semantic enhancement polyp feature map of the stage Convolution processing is performed to obtain the adaptive receptive field enhanced polyp feature map of stage n. ; This represents the feature map of the polyp enhanced by the adaptive receptive field in the first stage. This represents the feature map of the polyp, representing the adaptive receptive field enhancement in the second stage. This represents the feature map of the third-stage adaptive receptive field enhancement polyp. This represents the feature map of the fourth stage of adaptive receptive field enhancement polyp. Step S5: input the adaptive receptive field enhanced polyp feature maps of the four stages into the cross-scale frequency guide fusion module GFF for frequency modeling and cross-layer guide feature fusion processing to obtain a final polyp segmentation result.

2. The method of claim 1, wherein: The process of inputting the original endoscope image into the backbone network PVTv2 for processing is: The original endoscope image is input into the backbone network PVTv2 for shallow-to-deep extraction to obtain four-stage polyp feature maps denoted as ; wherein is the first-stage polyp feature map; is the second-stage polyp feature map; is the third-stage polyp feature map; is the fourth-stage polyp feature map.

3. The method of claim 2, wherein: The processing process of the lightweight DWSKAN middle-layer semantic enhancement module is: First, the second-stage polyp feature map is compressed using a 1×1 convolution to reduce the number of channels. and the characteristic diagram of stage III polyps Processing was performed to obtain the following results: and ,express: ； ； wherein is a second stage embedding feature map; is a third stage embedding feature map; is a 1 x 1 convolution; The lightweight DWSKAN middle layer semantic enhancement module introduces KAN linear layer based on Kolmogorov-Arnold network theory and Point-by-point nonlinear mapping is performed, respectively and , indicating: ； ； In the formula, For the second stage of embedding feature maps, the first The spatial location of each sample Enhanced eigenvalues ​​at the location; This represents the total number of non-linear activation functions. Let be the nth learnable weighting coefficient; Let be the nth learnable unary nonlinear activation function; This indicates retrieving all channels; For the third stage of embedding feature map The spatial location of each sample Enhanced eigenvalues ​​at the location; will be combined according to the original spatial position recombined to form a complete second stage enhanced feature map will be combined according to the original spatial position recombined to form a complete third stage enhanced feature map ; To and were treated, respectively, to give and , respectively. ； ； In the formula, is the output feature map of the second stage after spatial modeling; is the output feature map of the third stage after spatial modeling; is a depth separable KAN convolution; Let and be the residual connection and apply the normalization operation to obtain ; Let and be the residual connection and apply the normalization operation to obtain , which means: ； ； In the formula, is a second-stage semantic enhancement feature map; is batch normalization; is a third-stage semantic enhancement feature map.

4. The method of claim 3, wherein: The processing process of the Token-KAN deep-layer semantic enhancement module is: First, the fourth stage polyp feature map is processed by convolution, batch normalization and activation function, and converted into a deep embedding feature map, which is represented as: ； In the formula, represents a deep embedding feature map; represents a two-dimensional convolution operation; deep embedding feature map divided into multiple patches of size denotes the height of the deep embedding feature map; denotes the width of the deep embedding feature map; denotes the number of channels of the deep embedding feature map; denotes the set of real numbers; denotes the size of each patch in the height and width directions of the deep embedding feature map; a patch is a small block in the deep embedding feature map; Each patch is flattened into a one-dimensional vector, constituting tokens tokens; denotes the total number of generated tokens tokens. No. Each token is All tokens are mapped through a learnable linear mapping. Projecting onto a D-dimensional vector space forms an embedding sequence ,express: ； In the formula, represents an embedding sequence composed of all token tokens after linear mapping; represents a vector obtained after the first token token passes through a learnable linear mapping E; represents a vector obtained after the Nth token token passes through a learnable linear mapping E; The Token-KAN deep semantic enhancement module adopts three KAN layers, and the residual connection and layer normalization are performed after each layer to obtain the first KAN layer output The input is processed in the three KAN layers, and the residual connection and layer normalization are performed after each layer to obtain the first layer KAN layer output , which represents: ； wherein is the layer normalization; is the layer output of the KAN layer; ; is the layer output of the KAN layer; is the layer output of the KAN layer; is the layer output of the KAN layer; From the first layer KAN layer output to the first layer KAN layer transformation, denoted: ； In the formula, is the KAN layer employs a set of nonlinear mappings, i.e., a unary activation function set; composed of a plurality of unary activation functions composed of a plurality of unary activation functions composed of a plurality of unary activation functions composed of a plurality of unary activation functions composed of a plurality of unary activation functions composed of a plurality of unary activation functions ； In the formula, For input channel dimension index; Output channel dimension index; For the first KAN layer input dimension; For the first The output dimension of the KAN layer; In After that, KAN attention mechanism is introduced to capture the context information, which is represented as: ； In the formula, is a semantic enhanced token sequence; is a KAN attention mechanism; the token sequence after semantic enhancement rearranged into a spatial feature map the fourth stage polyp feature map after upsampling and the spatial feature map a residual connection is performed to obtain the fourth stage semantic enhancement feature map .

5. The method of claim 4, wherein: Enhance the first-stage polyp feature map, and the specific process is: First, the first stage polyp feature map Adopt 1x1 convolution, batch normalization and Activation operation for semantic enhancement processing, obtain the first stage semantic enhancement feature map ; wherein, ; is the semantic enhanced polyp feature map of the first stage.

6. The method of claim 5, wherein: aggregation The aggregation feature is input to a lightweight convolution prediction branch for processing to obtain an initial boundary prediction map and an initial polyp prediction map, respectively.

7. The method of claim 6, wherein: The processing process of the cross-scale frequency guide fusion module GFF is: To the Adaptation of receptive field enhancement polyp feature map Using The discrete wavelet transform is decomposed into four components, denoted: ； In the formula, denotes the adaptive receptive field enhanced polyp feature map of the first stage; , and denote the horizontal high-frequency component, the vertical high-frequency component and the diagonal high-frequency component of the adaptive receptive field enhanced polyp feature map of the first stage, respectively. Adding , and together into a single first stage high-frequency feature map , which represents: ； Based on the first Stage high-frequency feature map The gating weight matrix is generated by 1x1 convolution and Sigmoid activation function, which is represented as: ； wherein, denotes a gating weight matrix; denotes a Sigmoid activation function; denotes a weight matrix of a 1 x 1 convolution kernel; denotes a bias term; is an element-wise multiplication operation; to the first stage high-frequency feature map of the second stage by 1x1 convolution and nonlinear transformation, to obtain the high-frequency feature enhanced by the first stage , represents: ； In the formula, represents activation function; represents a weight matrix of a 1x1 convolution kernel; represents a bias term; The high-frequency feature of the first stage is enhanced The high-frequency feature of the first stage is enhanced by spatial attention and channel attention to obtain the high-frequency feature of the second stage enhanced by attention , denoted as ​ ； Gating weight matrix , the first stage attention enhanced high-frequency feature , point-by-point modulation is performed to obtain the first stage high-frequency boundary enhanced feature map , which is represented as ； The first Stage high-frequency boundary enhancement feature map As the guide feature, the initial boundary prediction map and the initial polyp prediction map are input into the boundary and foreground region dual-supervised decoder for processing to obtain the final polyp segmentation result.

Images