Intestinal polyp segmentation method and system based on spiral scanning state space model

Abstract

The present application relates to the technical field of medical image processing, in particular to a kind of intestinal polyp segmentation method and system based on helical scanning state space model, comprising the following steps: S1, obtain and pre-process image;S2, multi-scale context feature coding;S3, adaptive frequency domain feature enhancement;S4, decoding and segmentation prediction;S5, thresholding processing generates binary segmentation mask image.The present application keeps the boundary integrity of irregular target through dynamic multi-focus helical scanning strategy, and introduces adaptive frequency domain filtering module to enhance the discriminability of key features, so as to comprehensively improve the segmentation accuracy and robustness in complex medical image tasks such as intestinal polyp.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method and system for segmenting intestinal polyps based on a spiral scan state space model. Background Technology

[0002] Medical image segmentation is a crucial step in modern clinical diagnosis, pathological classification, treatment planning, and surgical robotics. With the development of deep learning technology, methods based on Convolutional Neural Networks (CNNs), such as U-Net and its variants, have become the benchmark in this field due to their powerful feature extraction capabilities. However, CNNs are limited by their local receptive fields, making it difficult to capture long-range dependencies, resulting in insufficient segmentation accuracy for complex anatomical structures and variable lesion morphologies.

[0003] To address the problem of long-range dependencies, Transformer-based methods (such as TransUNet and Swin-UNet) have been introduced into medical image segmentation. Although Transformer performs well in global modeling, its computational complexity increases quadratically with sequence length and often ignores fine-grained local details, leading to a decrease in the spatial accuracy of the segmentation results.

[0004] Recently, the Mamba architecture based on the state-space model has attracted much attention due to its linear time complexity and efficient long-distance dependency modeling capabilities. However, existing Mamba-based methods face serious challenges when applied to medical images, especially for tasks such as intestinal polyp segmentation, where their limitations are more pronounced: First, their fixed unidirectional or row-column scanning strategies are severely mismatched with the irregular circular shape of medical targets (such as polyps), which can disrupt spatial adjacency and semantic continuity, leading to blurred and distorted boundaries; Second, the single-center scanning mechanism is difficult to effectively cover the multiple and scattered lesions commonly seen in colonoscopy images, easily resulting in missed detections; Third, most existing methods only perform feature modeling in the spatial domain, failing to effectively utilize the frequency domain information that is crucial for distinguishing tissue texture and edges, resulting in insufficient robustness of the model when intestinal images contain complex noise such as feces and bubbles.

[0005] Therefore, there is an urgent need for a new medical image segmentation architecture that can dynamically adapt to lesion morphology, maintain spatial continuity, effectively integrate local and global features, and make full use of frequency domain information. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a method for segmenting intestinal polyps based on a spiral scanning state-space model. This method maintains the boundary integrity of irregular targets through a dynamic multi-focus spiral scanning strategy and introduces an Adaptive Frequency Domain Filtering Module (AFFM) to enhance the discriminative power of key features. This comprehensively improves the segmentation accuracy and robustness in complex medical image tasks such as those involving intestinal polyps. The specific steps include:

[0007] S1. Image Acquisition and Preprocessing: Acquire the original image to be segmented and perform standardized preprocessing on the original image to obtain the input feature map. The preprocessing includes adjusting the size to a preset resolution and performing normalization processing to eliminate the differences caused by the imaging device and acquisition conditions, and to ensure that the input data meets the model requirements.

[0008] S2, Multi-scale contextual feature encoding: The input feature map is input into the encoder of the U-Net model to generate a multi-center attention map; the encoder contains multiple cascaded multi-center spiral scanning Mamba modules, each module dynamically generates multiple focal points based on its input feature map, and generates a spiral scanning sequence for each focal point to capture and fuse local details and global contextual information of the image, thereby outputting a multi-level feature map;

[0009] S3. Adaptive Frequency Domain Feature Enhancement: The multi-level feature map output by the encoder is input into the bottleneck layer of the U-Net model; the bottleneck layer contains an adaptive frequency domain filtering module, which is used to transform the features in the spatial domain to the frequency domain and perform adaptive filtering according to the energy distribution of the features to enhance the modeling of key organizational structures and obtain an enhanced context feature map.

[0010] S4. Decoding and Segmentation Prediction: The enhanced context feature map is input into the decoder of the U-Net model. The spatial resolution is gradually restored through upsampling and convolution operations, and skip connections are made with the features of the corresponding layer of the encoder to finally generate a segmentation probability map.

[0011] S5. Result generation: Threshold the segmentation probability map to output the binarized segmentation mask image corresponding to the original image.

[0012] Preferably, in step S2, inputting the input image into the encoder of the U-Net model to generate a multi-center attention map specifically includes the following steps: The multi-center attention map is generated through the following segmentation-guided attention mechanism:

[0013]

[0014] in Input feature map; Represents the linear rectification activation function; This represents a 3×3 convolution operation; is the sigmoid activation function; attn_map is a multi-center attention map.

[0015] Preferably, in step S2, each focal point generates a spiral scanning sequence, the spiral radius of which is... for: ,in This is a resolution-dependent reference radius constant; This is the scaling factor; This indicates variance operation; `weigthted_coords` represents the attention-weighted centroids. ,in, For the first The segmentation guides attention values ​​at each position. For the first Spatial coordinates of a location, A constant used to ensure numerical stability.

[0016] Preferably, the multi-center spiral scanning module in S2 further performs the following steps: dividing the feature map into multiple horizontal slices and vertical slices, performing bidirectional scanning on the horizontal slices and vertical slices, and generating a flipped spiral scanning sequence to enhance the robustness of the model.

[0017] Preferably, the adaptive frequency domain filtering module's processing steps are as follows: converting the input features to the frequency domain, dynamically calculating routing weights using a multilayer perceptron, adaptively guiding multiple frequency filters to filter the frequency domain features based on the routing weights, and converting the filtered frequency domain features back to the spatial domain. The specific formula is as follows:

[0018]

[0019] in For the input feature map, For adaptive frequency domain filtering, ⊙ represents the Hadamard product, i.e., element-wise multiplication. For two-dimensional fast Fourier transform, This refers to the composition of functions, indicating that preprocessing operations are first performed on the input feature map. Then perform a two-dimensional fast Fourier transform. , The preprocessing operations for the input feature map include pointwise convolution and identity mapping, used to optimize the input feature distribution to compensate for acquisition fluctuations in medical images. This is a two-dimensional inverse fast Fourier transform, used to convert the filtered frequency domain features back to the spatial domain. The output of the adaptive frequency domain filtering operation is the enhanced context feature map.

[0020] Preferably, the U-Net model in S4 includes skip connections, which integrate channel attention bridges and spatial attention bridges to achieve multi-scale feature fusion between the encoder and decoder.

[0021] Preferably, the encoder in S2 is built on the Mamba architecture, specifically by using a selective state-space model to model the spiral scan sequence and handling long-range dependencies in linear time complexity.

[0022] Preferably, bilinear interpolation is used for upsampling in S4, and multi-scale features are enhanced through channel attention bridge and spatial attention bridge modules to improve the model's sensitivity to lesions at different scales. The specific formula is as follows:

[0023]

[0024]

[0025] in The input feature map; To compute a one-dimensional channel attention map; To compute a two-dimensional spatial attention map, This indicates element-wise multiplication. F' is the feature map after channel dimension enhancement, and F'' is the final output feature map after spatial dimension enhancement.

[0026] Accordingly, the calculation process for the channel attention bridge and spatial attention bridge modules is as follows:

[0027]

[0028] Where F represents the input feature map. and These represent the global average pooling operation and the global max pooling operation, respectively. This represents a multilayer perceptron, containing two weighting matrices. and , This represents the sigmoid function. This represents the feature vector with the maximum channel dimension obtained after global average pooling. This represents the feature vector with the maximum channel dimension obtained after global max pooling. Indicates the kernel size as Convolution operation, This represents the feature vector with the maximum spatial dimension obtained after global average pooling. This represents the feature vector with the maximum spatial dimension obtained after global max pooling. This indicates a splicing operation.

[0029] A system for segmenting intestinal polyps based on a spiral scan state space model is characterized by segmenting using the intestinal polyp segmentation method described above.

[0030] By adopting the above solution, the present invention has the following advantages and beneficial effects:

[0031] (1) The multi-center spiral scanning strategy is adopted. The adaptively generated spiral curve replaces the traditional straight line scanning, which can better fit the common circular, elliptical or irregular lesion shapes in medical images, maintain the semantic continuity of the boundary area, and effectively reduce feature breakage and boundary distortion caused by the mismatch between the scanning path and the target shape.

[0032] (2) By dynamically selecting multiple scanning centers through the attention mechanism, the problem of insufficient modeling of multiple lesions and multiple organs by single-center scanning is overcome, enabling the model to capture multiple spatially dispersed but semantically related target regions at the same time, thereby improving the overall perception of complex anatomical structures.

[0033] (3) Adaptive frequency domain filtering module, which dynamically enhances the key frequency bands related to the organizational structure in the frequency domain, suppresses noise and irrelevant details, thereby strengthening the feature representation with diagnostic value in the global context, and further improving the segmentation accuracy and robustness. Attached Figure Description

[0034] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the specific embodiments of the present invention will be briefly introduced below. It should be understood that the following drawings only show some specific embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0035] Figure 1 This is a flowchart illustrating the method described in this invention;

[0036] Figure 2 This is a schematic diagram of the overall model architecture of the intestinal polyp segmentation method based on a spiral scanning state-space model and dynamic filtering according to the present invention.

[0037] Figure 3 The flowchart of the MSS Mamba module described in this invention includes attention map generation, NMS center selection, and spiral scan generation.

[0038] Figure 4 This is a schematic diagram comparing the horizontal scanning of natural images and the central spiral scanning of medical images according to the present invention;

[0039] Figure 5 This is a schematic diagram comparing the traditional scanning strategy of the present invention with the multi-center scanning strategy of the present invention in a multi-lesion scenario;

[0040] Figure 6 This is a flowchart illustrating the structure of the AFDFM module described in this invention. Detailed Implementation

[0041] 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 only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0043] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0044] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0045] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0046] Embodiments of the present invention disclose a method for segmenting intestinal polyps based on a helical scan state-space model and dynamic filtering, such as... Figure 1 As shown, this method includes 5 steps, and the complete model structure is as follows: Figure 2 As shown in the figure, MSS Mamba is a multi-center spiral scan Mamba module. The specific operations for each step are as follows:

[0047] S1. Acquire the original medical image to be segmented, and perform standardized preprocessing on the original image to obtain the input feature map. Preprocessing includes adjusting the size to a preset resolution and performing normalization to eliminate differences caused by imaging equipment and acquisition conditions, ensuring that the input data meets the model requirements.

[0048] S2, Input the input feature map into... Figure 3 The encoder of the U-Net model is shown. The encoder contains multiple cascaded MSS Mamba modules. Each module dynamically generates multiple focal points based on its input features and generates a spiral scan sequence for each focal point to capture and fuse local details and global contextual information of the image, outputting a multi-level feature map. Specifically, the MSS Mamba module first generates an attention map through a segmentation-guided attention mechanism:

[0049]

[0050] in For the input feature map, Represents the linear rectified activation function. This represents a 3×3 convolution operation. is the sigmoid activation function, and attn_map is the final output attention map.

[0051] The number of focus centers, k, is dynamically estimated based on the entropy of the normalized attention map. Given an attention map... First, normalize it:

[0052]

[0053] Where attn_map is a multi-center attention map, Softmax(·) represents the Softmax normalization function, and attn_norm is the attention map obtained after normalization.

[0054] And its entropy value is calculated as follows:

[0055]

[0056] Let be the value of the i-th element in attn_norm, where log(·) denotes the natural logarithm function, and ε is a minimal constant used for numerical stability. This is the final calculated entropy value.

[0057] The number of focal points is dynamically determined based on the entropy of this attention map. :

[0058]

[0059] in This represents the minimum number of focal points. The maximum number of focal points is determined by the feature map resolution. The entropy of the attention map after normalization. For logarithmic operations, and These represent the height and width of the feature map, respectively.

[0060] Attention-guided nonmaximum suppression is employed to select spatially dispersed focal points. For each focal point, a logarithmic spiral trajectory is generated for feature scanning, with the spiral radius adaptively determined based on the variance of the attention-weighted coordinates.

[0061]

[0062]

[0063] in This is a resolution-dependent reference radius constant. Scaling factor This indicates variance operation, and `weigthted_coords` represents the attention-weighted centroids. This is a square root operation. This represents the summation operation over all positions in the feature map. For the first The segmentation guides attention values ​​at each position. For the first Spatial coordinates of a location, To ensure numerical stability, a small constant is used. The spiral scan maintains the continuity of spatial adjacency relationships and combines bidirectional horizontal and vertical slicing to achieve effective extraction of multi-scale features. The scale of the output feature maps at each stage of the encoder decreases progressively, forming a multi-level feature representation.

[0064] The pixel features corresponding to each type of scanning trajectory are sequentially reassembled into a one-dimensional input sequence. A Selective State Space Model (S6) is selected to map it to a continuous system defined by linear differential equations, and the Zero-Order Hold (ZOH) method is used for discretization.

[0065]

[0066]

[0067] Let A be the learned step size parameter, B be the input projection matrix, exp(·) be the exponentiation of the natural constant, and I be the identity matrix. The discrete state transition matrix, It is a discrete input matrix.

[0068] To achieve selective focusing on different lesion regions, the model performs state evolution using the following discrete recursive formula:

[0069]

[0070]

[0071] x(t) is the input feature sequence, and h(t) is the hidden state. The discrete state transition matrix, For discrete input matrices, To output the projection matrix, This is a residual join. For example... Figure 4 , Figure 5 As shown, this process utilizes a selective scanning mechanism, enabling the model to dynamically adjust the attention weights for different anatomical structures based on the spiral trajectory. Figure 5 The MCSS in the image is a multicenter spiral scan image.

[0072] S3 inputs the multi-level feature map output by the encoder into the bottleneck layer of the U-Net; the bottleneck layer includes, for example, the multi-level feature map output by the encoder. Figure 6The AFDFM shown is used to transform features from the spatial domain to the frequency domain and perform adaptive filtering based on the energy distribution of the features to enhance the modeling of key tissue structures, resulting in an enhanced contextual feature map. AFDFM dynamically calculates routing weights through MLP to guide different frequency filters to selectively enhance features, thereby strengthening frequency components related to anatomical structures in the frequency domain, suppressing noise and irrelevant details, and improving the discriminative power of the feature representation.

[0073] The specific formula is as follows:

[0074]

[0075] Where X is the input feature map, which is the spatial feature output by the bottleneck layer of the model in this invention, carrying the initial representation information of the medical image; The adaptive frequency domain filter is dynamically generated by the MLP in this invention and can be adapted to the characteristics of the medical image corresponding to the input feature map; ⊙ is the Hadamard product, which is an element-wise multiplication operation used to adaptively weight and adjust the frequency domain features. It is a two-dimensional fast Fourier transform used to convert spatial domain features into frequency domain features; This refers to the composition of functions, indicating that preprocessing operations are first performed on the input feature map. Then perform a two-dimensional fast Fourier transform F; The preprocessing operations for the input feature map include pointwise convolution and identity mapping, which are used to optimize the distribution of input features to compensate for the acquisition fluctuations of medical images. It is a two-dimensional inverse fast Fourier transform used to convert the filtered frequency domain features back to the spatial domain; This is the output of the adaptive frequency domain filtering operation, i.e., the spatial feature map after frequency domain enhancement.

[0076] Adaptive frequency domain filter basis Defined as ,in C represents the complex field, H represents the feature map height, and W represents the feature map width. (for rounding up); filter ,in The number of feature channels, The generated adaptive frequency domain filter is weighted and correlated using a multilayer perceptron M, as specifically defined below:

[0077]

[0078]

[0079] s is the original score vector output by the MLP, whose elements correspond to the weight generation basis of different channels and different filter bases; e is the natural constant; c is the channel number of the feature map, used to locate the feature channel corresponding to the dedicated adaptive filter; N is the number of globally shared filter bases, which is the total number of basic components that constitute the adaptive filter; i is used to locate the i-th filter base under this channel, and the whole achieves accurate matching of "channel-filter base"; The summation operator is used to "accumulate and aggregate the N related items corresponding to the c-th channel"; To perform a linear combination operation on N basic filters according to their corresponding normalized weights, and to achieve adaptive optimization of the basic filters for the characteristics of the c-th channel; c refers to the dedicated adaptive frequency domain filter for the c-th channel. `c` is a tensor index wildcard, representing the selection of all elements in the height and width dimensions of the feature map for this channel. This filter can specifically capture the frequency domain features of the medical image corresponding to the c-th channel and is a core component of AFDFM. `h` represents the pixel index in the vertical direction of the input feature map; `w` represents the pixel index in the horizontal direction of the input feature map. This is used to sum the feature values ​​of all channels at all pixel positions in the input feature map, which is used to aggregate global image information; H is the height of the input feature map, and W is the width of the input feature map; MLP is used to map low-dimensional global mean features to high-dimensional original score vectors, realizing a non-linear transformation from global features to weights. The cross-channel feature vector of feature map X at height index h and width index w; Let be the output weight vector, where Let represent the i-th adaptive weight coefficient, N be the number of filter basis vectors, C' be the number of channels in the input feature map, and the superscript T denotes the matrix transpose operation.

[0080] S4. The enhanced context feature map is input into the decoder of the U-Net. Spatial resolution is gradually restored through upsampling and convolution operations, and skip connections are made with the features of the corresponding layer of the encoder to finally generate a segmentation probability map. During the decoding process, bilinear interpolation is used for upsampling, and multi-scale feature fusion is enhanced through channel attention bridge and spatial attention bridge modules to improve the model's sensitivity to lesions at different scales. The specific formula is as follows:

[0081]

[0082]

[0083] The input feature map; To compute a one-dimensional channel attention map; To compute a two-dimensional spatial attention map; where F' represents element-wise multiplication; F' is the feature map after channel dimension enhancement; F'' is the final refined output after spatial dimension enhancement.

[0084] The calculation process for channel attention and spatial attention is as follows:

[0085]

[0086]

[0087] Where F represents the input feature map; and These represent the global average pooling operation and the global max pooling operation, respectively. This represents a multilayer perceptron, containing two weighting matrices. and ; Represents the sigmoid function; This represents the feature vector with the maximum channel dimension obtained after global average pooling. This represents the feature vector with the maximum channel dimension obtained after global max pooling. Indicates the kernel size as Convolution operation; This represents the feature vector with the maximum spatial dimension obtained after global average pooling. This represents the feature vector with the maximum spatial dimension obtained after global max pooling. This indicates a stitching operation, merging the two 2D images along the channel dimension. The feature refinement mechanism introduced in the skip connection further optimizes the accuracy of boundary detail restoration.

[0088] S5. Threshold the segmentation probability map and output the binarized segmentation mask image corresponding to the original image. By setting an appropriate probability threshold, the probability map is transformed into a mask with clear distinction between lesion areas and background, thus completing the automatic segmentation of the medical image.

[0089] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0090] The present invention provides the preferred embodiments described above. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for segmenting intestinal polyps based on a spiral scan state-space model, characterized in that, Includes the following steps: S1. Image Acquisition and Preprocessing: Acquire the original image to be segmented and perform standardized preprocessing on the original image to obtain the input feature map. The preprocessing includes adjusting the size to a preset resolution and performing normalization processing to eliminate the differences caused by the imaging device and acquisition conditions, and to ensure that the input data meets the model requirements. S2. Multi-scale contextual feature encoding: The input feature map is input into the encoder of the U-Net model to generate a multi-center attention map; the encoder contains multiple cascaded multi-center spiral scanning Mamba modules, each module dynamically generates multiple focal points based on its input feature map, and generates a spiral scanning sequence for each focal point to capture and fuse local details and global contextual information of the image, thereby outputting a multi-level feature map; the specific steps of inputting the input feature map into the encoder of the U-Net model to generate the multi-center attention map are as follows: The multi-center attention map is generated through the following segmentation-guided attention mechanism: in Input feature map; Represents the linear rectification activation function; This represents a 3×3 convolution operation; It is the sigmoid activation function; attn_map is a multi-center attention map; S3. Adaptive Frequency Domain Feature Enhancement: The multi-level feature map output by the encoder is input into the bottleneck layer of the U-Net model; the bottleneck layer contains an adaptive frequency domain filtering module, which is used to transform the features in the spatial domain to the frequency domain and perform adaptive filtering according to the energy distribution of the features to enhance the modeling of key organizational structures and obtain an enhanced context feature map. S4. Decoding and Segmentation Prediction: The enhanced context feature map is input into the decoder of the U-Net model. The spatial resolution is gradually restored through upsampling and convolution operations, and skip connections are made with the features of the corresponding layer of the encoder to finally generate a segmentation probability map. S5. Result generation: Threshold the segmentation probability map to output the binarized segmentation mask image corresponding to the original image.

2. The intestinal polyp segmentation method based on a spiral scan state-space model according to claim 1, characterized in that, In step S2, each focal point generates a spiral scan sequence, the spiral radius of which is... for: ,in This is a resolution-dependent reference radius constant; This is the scaling factor; This indicates variance operation; `weigthted_coords` represents the attention-weighted centroids. ,in, For the first The segmentation guides attention values ​​at each position. For the first Spatial coordinates of a location, A constant used to ensure numerical stability.

3. The intestinal polyp segmentation method based on a spiral scan state-space model according to claim 1, characterized in that, The multi-center spiral scanning module in S2 also performs the following steps: dividing the feature map into multiple horizontal and vertical slices, performing bidirectional scanning on the horizontal and vertical slices, and generating a flipped spiral scanning sequence to enhance the robustness of the model.

4. The intestinal polyp segmentation method based on a spiral scan state-space model according to claim 1, characterized in that, The processing steps of the adaptive frequency domain filtering module are as follows: the input features are converted to the frequency domain, the routing weights are dynamically calculated using a multilayer perceptron, and based on the routing weights, multiple frequency filters are adaptively guided to filter the frequency domain features. The filtered frequency domain features are then converted back to the spatial domain. The specific formula is as follows: in For the input feature map, For adaptive frequency domain filtering, ⊙ represents the Hadamard product, i.e., element-wise multiplication. For two-dimensional fast Fourier transform, This refers to the composition of functions, indicating that preprocessing operations are first performed on the input feature map. Then perform a two-dimensional fast Fourier transform. , The preprocessing operations for the input feature map include pointwise convolution and identity mapping, used to optimize the input feature distribution to compensate for acquisition fluctuations in medical images. This is a two-dimensional inverse fast Fourier transform, used to convert the filtered frequency domain features back to the spatial domain. The output of the adaptive frequency domain filtering operation is the enhanced context feature map.

5. The intestinal polyp segmentation method based on a spiral scan state-space model according to claim 1, characterized in that, The U-Net model in S4 includes skip connections, which integrate channel attention bridges and spatial attention bridges to achieve multi-scale feature fusion between the encoder and decoder.

6. The intestinal polyp segmentation method based on a spiral scan state-space model according to claim 1, characterized in that, The encoder described in S2 is built on the Mamba architecture, specifically by using a selective state-space model to model the spiral scan sequence and handling long-range dependencies in linear time complexity.

7. The intestinal polyp segmentation method based on a spiral scan state-space model according to claim 1, characterized in that, In S4, bilinear interpolation is used for upsampling, and multi-scale features are enhanced through channel attention bridge and spatial attention bridge modules to improve the model's sensitivity to lesions at different scales. The specific formula is as follows: in Input feature map; To compute a one-dimensional channel attention map; To compute a two-dimensional spatial attention map, This indicates element-wise multiplication. F' is the feature map after channel dimension enhancement, and F'' is the final output feature map after spatial dimension enhancement.

8. The intestinal polyp segmentation method based on a spiral scan state-space model according to claim 7, characterized in that, Multi-scale feature fusion is achieved through channel attention bridge and spatial attention bridge modules. The calculation process of the channel attention bridge and spatial attention bridge modules is as follows: Where F represents the input feature map. and These represent global average pooling and global max pooling operations, respectively. This represents a multilayer perceptron, containing two weighting matrices. and , This represents the sigmoid function. This represents the feature vector with the maximum channel dimension obtained after global average pooling. This represents the feature vector with the maximum channel dimension obtained after global max pooling. Indicates the kernel size as Convolution operation, This represents the feature vector with the maximum spatial dimension obtained after global average pooling. This represents the feature vector with the maximum spatial dimension obtained after global max pooling. This indicates a splicing operation.

9. A system for segmenting intestinal polyps based on a spiral scan state-space model, characterized in that, The intestinal polyp segmentation method as described in any one of claims 1-8 is used for segmentation.

Images