Polyp segmentation method based on adaptive multi-path fusion and hierarchical perception cooperation

Abstract

The present application relates to the technical field of image segmentation, and discloses a polyp segmentation method based on adaptive multi-path fusion and hierarchical perception cooperation, which comprises the following steps: inputting a polyp image to be segmented into a colorectal polyp segmentation model; an adaptive multi-path polyp feature fusion module is used for adaptive channel redistribution preprocessing, parallel feature extraction, adaptive adjustment of path weight according to polyp features, and feature calibration through double attention and double residual connection; the final output feature map of the adaptive multi-path polyp feature fusion module is input into a hierarchical perception bidirectional feature cooperation module, feature fusion is performed through an encoder shallow layer feature processing path and a decoder deep layer feature processing path, a double-branch adaptive fusion strategy is adopted in the fusion process, and a final polyp segmentation result corresponding to the fused features is output; the present application efficiently deals with polyp heterogeneity, makes the segmentation result more consistent with the actual morphology of the polyp, and significantly improves the segmentation accuracy and robustness.

Description

Polyp segmentation method based on adaptive multi-path fusion and hierarchical perception collaboration Technical Field

[0001] This invention relates to the field of image segmentation technology, specifically to a polyp segmentation method based on adaptive multipath fusion and hierarchical perception collaboration. Background Technology

[0002] As primary precancerous lesions, early detection and resection of colorectal polyps are of great significance in reducing the incidence of colorectal cancer. Currently, computer-aided polyp segmentation technology based on deep learning is gaining popularity as an effective means of early detection of colorectal polyps. Existing colorectal polyp image segmentation techniques mainly rely on CNN and Transformer architectures, but both have limitations. On the one hand, polyps in endoscopic images are highly heterogeneous, ranging in size from a few millimeters to several centimeters, with diverse shapes and boundaries that are either clear or blurred. Small polyps have weak features and are easily missed, while large polyps exceed the local receptive field and are difficult to grasp as a whole. Irregular shapes and edges are easily confused with normal tissue. Existing feature extraction modules have obvious limitations in dealing with this diversity. Fixed convolutional kernels are difficult to capture multi-scale targets, simple concatenation ignores semantic differences, and a single enhancement path cannot take into account both local and global aspects. On the other hand, in polyp segmentation networks based on the U-Net architecture, skip connections are used to connect the encoder and decoder, passing shallow high-resolution detailed features to deep layers to achieve accurate boundary localization. However, existing methods for handling cross-level feature fusion have a one-way limitation. They focus more on the "guidance" of deep information to shallow information and ignore the "calibration" and "supplement" of shallow features to deep information. This results in insufficient respect for shallow features, difficulty in bridging semantic gaps, rough boundaries of deep features, and segmentation results that are prone to distortion at the edges of polyps. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing a polyp segmentation method based on adaptive multipath fusion and hierarchical perception collaboration.

[0004] The first aspect of this invention provides a polyp segmentation method based on adaptive multi-path fusion and hierarchical perception collaboration, the method comprising the following steps:

[0005] The polyp image to be segmented is input into a colorectal polyp segmentation model that includes an adaptive multi-path polyp feature fusion module and a hierarchical perception bidirectional feature collaboration module.

[0006] The adaptive multi-path polyp feature fusion module first performs adaptive channel redistribution preprocessing, and then uses local detail-aware branch, multi-scale gradient aggregation branch, edge orientation enhancement branch and global context modeling branch for parallel feature extraction. The path weights are adaptively adjusted according to the polyp features, and feature calibration is achieved through dual attention and dual residual connections.

[0007] The final output feature map of the adaptive multi-path polyp feature fusion module is input into the hierarchical perception bidirectional feature collaboration module. Feature fusion is performed through the shallow feature processing path of the encoder and the deep feature processing path of the decoder. A dual-branch adaptive fusion strategy is adopted in the fusion process, and the final polyp segmentation result map corresponding to the fused features is output.

[0008] Optionally, in a first implementation of the first aspect of the present invention, the adaptive channel reallocation preprocessing includes:

[0009]

[0010]

[0011]

[0012] in, Represents the input feature map, where These represent the height and width of the feature map, respectively. For the number of channels, express Convolution operation, Indicates The initial feature map after convolution processing. Indicates global average pooling. This represents the learnable weight matrix of the first fully connected layer, enabling the number of channels to increase from... Compress to , This represents the learnable weight matrix of the second fully connected layer, enabling the number of channels to increase from... Restore to , Represents the ReLU activation function. This represents the Sigmoid activation function. Represents the channel attention weight vector. This indicates element-wise multiplication, which multiplies the weight vector. With the initial feature map Each channel is matched and weighted. This represents the output feature map after channel attention calibration.

[0013] Optionally, in a second implementation of the first aspect of the present invention, the local detail-aware branch retains high-frequency texture information through lightweight convolution; the multi-scale gradient aggregation branch employs parallel multi-kernel convolution and introduces a gradient-weighted aggregation mechanism to perform scale adaptation and enhance boundary awareness processing; the edge orientation enhancement branch processes edge features through a cross-direction attention mechanism; and the global context modeling branch uses a self-attention mechanism to capture long-range dependencies.

[0014] Optionally, in a third implementation of the first aspect of the present invention, the feature calibration achieved through dual attention and dual residual connections includes:

[0015] Channel attention:

[0016]

[0017] in, Represents the channel attention weight vector. This represents the sigmoid activation function. This represents the weight matrix of the first-level channel attention fully connected layer. This represents the weight matrix of the second-level channel attention fully connected layer. Represents the ReLU activation function. Indicates the fusion feature map Perform global average pooling. Indicates the fusion feature map Perform global max pooling. This indicates a splicing operation, which splices the outputs of GAP and GMP along the channel dimension;

[0018] Spatial attention:

[0019]

[0020] in Represents the spatial attention weight map. express Convolution operation, This indicates the fusion feature map along the channel dimension. Perform average pooling. This indicates the fusion feature map along the channel dimension. Perform max pooling. This indicates a concatenation operation, which concatenates the outputs of channel average pooling and channel max pooling along the channel dimension.

[0021] First residual connection:

[0022]

[0023] in This represents the recalibrated feature map. This represents the feature map after joint attention using channel attention and spatial attention. Indicates the feature map conduct convolution, The feature map representing the output of the adaptive channel redistribution preprocessing. This represents element-wise addition, enabling residual joins.

[0024] Double residual connection:

[0025]

[0026] in, This represents the final output feature map of the adaptive multipath polyp feature fusion module. Represents the recalibration feature map conduct convolution.

[0027] Optionally, in the fourth implementation of the first aspect of the present invention, the deep feature processing path of the decoder captures the overall semantic and detail correspondence of large polyps and locates the local edge features of small polyps by simultaneously calculating the cross attention between deep semantic features and shallow detail features at three different scales: 7×7, 5×5, and 3×3. The scale combination is automatically selected through scale-aware weight learning.

[0028] Optionally, in a fifth implementation of the first aspect of the present invention, the dual-branch adaptive fusion strategy includes feature difference measurement, difference-aware fusion weight generation, and progressive feature fusion, wherein the feature difference measurement includes both feature distribution distance and gradient similarity.

[0029] Optionally, in a sixth implementation of the first aspect of the present invention, the distribution distance is:

[0030] ,

[0031] ,

[0032]

[0033] in, Representing shallow feature maps The channel mean vector, Representing shallow feature maps The channel variance vector, Representing deep feature maps The channel mean vector, Representing deep feature maps The channel variance vector, Indicates global average pooling. This represents the variance calculation function. This represents the output feature map of the encoder's shallow feature processing path. This represents the output feature map of the decoder's deep feature processing path. This represents the approximate KL divergence. This represents the ratio of the variance of shallow to deep channels. This represents the ratio of the square of the mean difference to the depth variance. The logarithm of the variance ratio;

[0034] The gradient similarity:

[0035]

[0036]

[0037] in, Representing shallow feature maps Horizontal gradient plot, Representing deep feature maps Horizontal gradient plot, express Along the width dimension Extract a sub-image from column 2 to the last column. express Extract a sub-image along the width dimension from column 1 to the second-to-last column;

[0038] Comprehensive difference indicators:

[0039]

[0040]

[0041] in, This represents the gradient cosine similarity, with values ​​ranging from [−1, 1]. Represents the vector dot product. and They represent and L2 norm, Represents the smoothing factor. Indicators representing comprehensive differences Represents the balance coefficient. This indicates that the Sigmoid activation function acts on... The distribution distance is normalized to the range [0,1]. The complementary term representing gradient similarity will Convert to a difference measure.

[0042] Optionally, in a seventh implementation of the first aspect of the present invention, the generation of the difference-aware fusion weights includes:

[0043] Basic fusion weights:

[0044]

[0045] in, Represents the basic fusion weight vector. This represents the Sigmoid activation function. Represents the learnable weight matrix. Indicates feature concatenation operation;

[0046] Spatial adaptive fusion weights:

[0047]

[0048]

[0049]

[0050]

[0051] in, Spatial weighted feature map representing shallow features. Spatial weighted feature map representing deep features express Convolution operation, A spatial adaptive weighted graph representing shallow features. A spatial adaptive weighted graph representing deep features. Represents an exponential function;

[0052] Global-local fusion weight combination:

[0053]

[0054]

[0055] in, The final fusion weight map representing shallow features. The final fusion weight map of deep features. This represents the first element of the basic fusion weight vector. This represents the second element of the basic fusion weight vector.

[0056] Optionally, in the eighth implementation of the first aspect of the present invention, the progressive feature fusion performs weighted fusion in the first stage, cross-calibration in the second stage, semantic and detail decoupling enhancement in the third stage, and three-level residual connection and output in the fourth stage, wherein the third stage includes separating low-frequency and high-frequency components, enhancing high-frequency details, and recombining features.

[0057] Optionally, in a ninth implementation of the first aspect of the invention, the weighted fusion is:

[0058]

[0059] in, This represents the first-stage weighted fusion feature map. The final fusion weight map representing shallow features. The final fusion weight map representing deep features. This represents the output feature map of the encoder's shallow feature processing path. This represents the output feature map of the decoder's deep feature processing path;

[0060] The cross-calibration:

[0061]

[0062]

[0063]

[0064] in, This represents the cross-calibration feature between shallow features and the first-stage fusion features. The cross-calibration feature represents the fusion feature between deep features and the first-stage features. This represents the feature map after the second stage of cross-calibration. This indicates a splicing operation. express Convolution operation, Indicates the cross-calibration intensity coefficient;

[0065] The separation of low-frequency and high-frequency components:

[0066]

[0067]

[0068] in, Represents low-frequency characteristic components. Represents high-frequency characteristic components, express Average pooling, This represents the feature map after the second-stage cross-calibration.

[0069] The enhanced high-frequency details:

[0070]

[0071] in, This indicates the enhanced high-frequency features. Represents the absolute value of the high-frequency component. express Convolution operation, This represents the Sigmoid activation function. Indicates the high-frequency enhancement factor. Indicates the enhancing factor;

[0072] The recombination feature:

[0073]

[0074] in, This represents the final feature map after recombination. Indicates low-frequency semantic components. This indicates the enhanced high-frequency detail components;

[0075] The third-level residual connection:

[0076]

[0077] in, This indicates that the feature map is output to the next higher level. Indicates two levels convolution, This represents the final feature map after recombination. This indicates a downsampling operation.

[0078] The polyp segmentation method based on adaptive multipath fusion and hierarchical perception collaboration in this invention has the following advantages:

[0079] 1. This invention efficiently addresses polyp heterogeneity through an adaptive multi-path polyp feature fusion module, effectively preserving the weak features of small polyps to avoid omissions, fully capturing the global information of large polyps to avoid local biases, and enhancing the edge and gradient features of polyps with blurred boundaries to avoid confusion with normal tissue, thus significantly improving segmentation accuracy and robustness.

[0080] 2. This invention constructs a bidirectional collaborative enhancement mechanism through a hierarchical perceptual bidirectional feature collaboration module. The encoder's shallow feature processing path adaptively adjusts the semantic guidance intensity through a hierarchical difference-perceptual modulation factor, avoiding excessive suppression of shallow features and fully respecting the characteristics of shallow details. The decoder's deep feature processing path achieves reverse calibration of shallow to deep layers through a multi-scale spatial alignment mechanism, supplementing the missing details in deep layers. Simultaneously, the dual-branch adaptive fusion strategy and the four-stage progressive fusion process can dynamically adjust the fusion strategy according to the differences in feature distribution and edge consistency, prioritizing the enhancement of detailed branch contributions in the polyp edge region, effectively bridging the cross-hierarchical semantic gap, reducing edge segmentation distortion, and making the segmentation results more consistent with the actual morphology of the polyp.

[0081] 3. The local detail-aware branch of this invention uses lightweight convolution, and the global context modeling branch uses a simplified self-attention mechanism, which avoids redundant computation caused by complex architecture. It not only solves the problem that existing deep learning segmentation methods may be difficult to deploy on medical endoscopy equipment due to model complexity, but also meets the segmentation speed requirements in clinical scenarios, realizing efficient detection and segmentation of polyps. It provides a feasible technical solution for computer-aided early screening of colorectal polyps and has significant clinical application and engineering transformation value. Attached Figure Description

[0082] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0083] Figure 1 is a flowchart of the polyp segmentation method based on adaptive multipath fusion and hierarchical perception collaboration provided in an embodiment of the present invention;

[0084] Figure 2 is a flowchart of the adaptive multipath polyp feature fusion module provided in an embodiment of the present invention;

[0085] Figure 3 is a flowchart of the hierarchical perception bidirectional feature collaboration module provided in an embodiment of the present invention. Detailed Implementation

[0086] 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 some embodiments of the present invention, and not all embodiments. 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.

[0087] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to Figure 1 for a flowchart of the polyp segmentation method based on adaptive multi-path fusion and hierarchical perception collaboration provided by the embodiments of the present invention. The method specifically includes the following steps:

[0088] Step 101. Input the polyp image to be segmented into a colorectal polyp segmentation model that includes an adaptive multi-path polyp feature fusion module and a hierarchical perception bidirectional feature collaboration module;

[0089] Step 102. The adaptive multi-path polyp feature fusion module first performs adaptive channel redistribution preprocessing, and then uses local detail-aware branch, multi-scale gradient aggregation branch, edge orientation enhancement branch and global context modeling branch to perform parallel feature extraction. The path weights are adaptively adjusted according to the polyp features, and feature calibration is achieved through dual attention and dual residual connections.

[0090] Step 103. Input the final output feature map of the adaptive multi-path polyp feature fusion module into the hierarchical perception bidirectional feature collaboration module, perform feature fusion through the shallow feature processing path of the encoder and the deep feature processing path of the decoder, and adopt a dual-branch adaptive fusion strategy in the fusion process to output the final polyp segmentation result map corresponding to the fused features.

[0091] In this embodiment, as shown in Figure 2, the adaptive multi-path polyp feature fusion module includes adaptive channel redistribution preprocessing, four-branch parallel feature extraction, and an adaptive multi-branch fusion strategy. The four-branch parallel feature extraction includes a local detail-aware branch, a multi-scale gradient aggregation branch, an edge-oriented enhancement branch, and a global context modeling branch. The adaptive multi-branch fusion strategy includes dynamic evaluation of branch importance, weighted feature fusion, a dual attention refinement module, and feature recalibration and output.

[0092] The adaptive multi-path polyp feature fusion module utilizes multi-path parallel processing and intelligent fusion to allow different feature extraction paths to focus on different segmentation subtasks: the local detail-aware branch preserves high-frequency texture information through lightweight convolution; the multi-scale gradient aggregation branch employs parallel multi-kernel convolution and introduces explicit gradient modeling, achieving both scale adaptation and enhanced boundary awareness; the edge orientation enhancement branch specifically handles anisotropic edge features through a cross-directional attention mechanism; and the global context modeling branch captures long-range dependencies using a simplified self-attention mechanism. Abandoning the traditional fixed-weight concatenation strategy, a task-driven dynamic fusion mechanism is designed, enabling the module to automatically adjust the contribution of each path based on the characteristics of the input features, and achieving fine feature calibration through dual attention and dual residual connections.

[0093] The adaptive multi-path polyp feature fusion module demonstrates strong adaptability when faced with complex and varied polyp features: for small polyps, the module automatically enhances the weights of detail branches and context branches; for large polyps, multi-scale branches play a dominant role; and for polyps with blurred boundaries, gradient enhancement and directional attention are activated preferentially. Through this intelligent feature collaboration mechanism, the adaptive multi-path polyp feature fusion module can significantly improve the accuracy and robustness of polyp segmentation while maintaining lightweight design.

[0094] In this embodiment, a channel importance-aware mechanism is introduced during the adaptive channel reallocation preprocessing:

[0095]

[0096]

[0097]

[0098] in, This represents the intermediate feature map of the input, where The spatial dimensions (height, width) of the feature map. For the number of channels, express Convolution operations are used to adjust the number of channels in a feature map (by default, the spatial size is preserved). constant), Indicates The initial feature map after convolution processing, the number of channels and Keep as , This represents global average pooling, which performs average pooling on each channel of the feature map along its spatial dimension. Feature map conversion vector (i.e.) ), This represents the learnable weight matrix of the first fully connected layer, enabling the number of channels to increase from... Compress to ( The compression ratio is usually an integer such as 16. This represents the learnable weight matrix of the second fully connected layer, enabling the number of channels to increase from... Restore to , Represents the ReLU activation function ( ), Introducing nonlinear transformation, This represents the Sigmoid activation function ( The output is mapped to the [0, 1] interval to generate attention weights. This represents the channel attention weight vector, where each element corresponds to the importance weight of a channel in the feature map. This indicates element-wise multiplication, which multiplies the weight vector. With feature map Each channel is matched and weighted. This represents the output feature map after channel attention calibration.

[0099] In this embodiment, branch 1: the local detail-aware branch adopts a progressive channel separation strategy:

[0100]

[0101]

[0102]

[0103] in, Indicates the input feature map, For spatial dimensions (height, width) For the number of channels, express Convolution operation, superscript This indicates that the number of output channels is (Achieving channel dimensionality reduction), convolution kernel size is , This represents the feature map after depthwise convolution. This represents a learnable scalar parameter used to dynamically adjust the weights of deep convolutional features. Represents the adaptive weighting factor, expressed by the Sigmoid function ( ,in This generates and controls the contribution ratio of deep convolutional features in residual fusion. This represents the output feature map after residual fusion, obtained by converting the original dimensionality-reduced features. Weighted depthwise convolutional features This is achieved by adding elements one by one, preserving the original details while incorporating spatial features extracted by depthwise convolution. Indicates the adaptive weight parameters;

[0104] Branch 2: Multi-scale gradient aggregation branch adopts a parallel multi-scale + gradient weighted aggregation mechanism:

[0105]

[0106] The introduced scale-aware weight generator is:

[0107]

[0108]

[0109] The introduced gradient enhancement mechanism is as follows:

[0110]

[0111]

[0112]

[0113]

[0114] in, Indicates the input feature map, For spatial dimensions (height, width) For the number of channels, Indicates depthwise separable convolution, subscript Indicates the kernel size (corresponding to respectively) , , ), consisting of depthwise convolution (channel-wise convolution) and pointwise convolution ( Composed of convolutions, the number of output channels is... Keep it consistent (all) ), Indicates the first Depth-separable convolutional output feature maps at various scales ( (corresponding to different convolution kernel sizes) This represents global average pooling, which performs average pooling on each channel of the feature map along its spatial dimension. Feature map conversion vector (i.e.) ), Indicates the first Learnable sensitive parameters at each scale are used to adjust the contribution of features at the corresponding scale in the weight calculation. Indicates the first The normalized weights for each scale (calculated using the Softmax function) satisfy... The aggregation ratio used to dynamically balance multi-scale features. This represents the aggregated feature map after multi-scale features have been weighted. This represents the horizontal gradient map, calculated by aggregating the differences between adjacent pixels in the width dimension of the feature map. This represents the vertical gradient map, calculated by aggregating the differences between adjacent pixels in the height dimension of the feature map. This represents a gradient magnitude map, with dimensions consistent with the gradient map (horizontal gradient corresponds to...). Vertical gradient corresponds ), This represents the local constant, used to prevent numerical instability when the sum of squared gradients is zero. express Convolution operations are used to adjust the number of channels in the gradient magnitude map. This represents the Sigmoid activation function ( The gradient magnitude map after convolution is mapped to the [0, 1] interval to generate gradient attention weights. This represents the learnable gradient enhancement strength coefficient, which controls the magnitude of the enhancement effect of gradient information on features. This represents the output feature map after gradient enhancement;

[0115] Branch 3: Edge-oriented enhancement branch introduces a bidirectional anisotropic convolution + cross-attention mechanism:

[0116]

[0117]

[0118] Among them, the horizontal and vertical convolution kernel sizes satisfy ;

[0119] Cross-directional attention mechanism:

[0120]

[0121]

[0122]

[0123] in, Indicates the input feature map, For spatial dimensions (height, width) For the number of channels, This represents a horizontal convolution operation with a kernel size of [size missing]. Extract features along the width direction, and output the number of channels. Keep it consistent (all) ), This represents the feature map extracted by horizontal convolution. This represents a vertical convolution operation with a kernel size of [size missing]. Features are extracted along the height direction, and the number of output channels is equal to the number of output channels. Maintain consistency. This represents the feature map extracted by vertical convolution. This represents global average pooling along the width dimension, applying it to each channel of the feature map in the width direction ( Average the dimensions, and Feature map conversion ,Right now , This represents global average pooling along the height dimension, applied to each channel of the feature map in the height direction ( Average the dimensions, and Feature map conversion ,Right now , This indicates a channel-dimensional concatenation operation, which merges two feature maps along the channel dimension (the spatial size needs to be unified first through a broadcast mechanism). express Convolution operations are used to compress the number of channels in the concatenated feature map to a smaller number. (Same as the number of channels in the input feature map), preparing for the generation of attention weights. This represents the Sigmoid activation function ( The convolution output is mapped to the [0, 1] interval to generate attention weights. This represents the horizontal attention weight map, which is broadcast from... Extend to Same size, used to enhance key regions of the horizontal feature map. Represents the vertical attention weight map, broadcast from... Extend to Same size, used to enhance key regions of the vertical feature map. This indicates element-wise multiplication, where the attention weights are weighted by the feature maps in the corresponding directions. This indicates channel-dimensional concatenation, which involves weighting the horizontal feature map. With vertical feature map Merging at the channel level, the number of output channels is... , This represents the output feature map after attention fusion in the cross directions, which integrates key information in the horizontal and vertical directions, and has twice the number of channels as the input feature map;

[0124] Branch 4: The global context modeling branch introduces a simplified version of non-local operations and a context recalibration mechanism.

[0125]

[0126]

[0127]

[0128] The remodeled features are in sequence form:

[0129]

[0130]

[0131]

[0132] Context gating:

[0133]

[0134]

[0135] in, Indicates the input feature map, For spatial dimensions (height, width) For the number of channels, express Convolution operation, Indicates the number of output channels. This represents the query feature graph, after... Convolution dimensionality reduction to aisle, This represents the key feature map, which has the same dimension as the query features. Representational feature map, after Convolution dimensionality reduction to aisle, This represents a sequence of queries and keys. , , Represents a sequence of values. Characterizing the correlation strength at different spatial locations, Representation context feature map, This indicates channel-dimensional concatenation, which combines the input features. ( Channel and context features ( (Channel) spliced ​​into Channel characteristics, express The convolution operation will concatenate the results. Channel features compressed to Channels, matched with contextual feature dimensions. This represents the Sigmoid activation function, which maps the convolution output to the [0, 1] interval to generate gate weights. Context-local balance gating controls the fusion ratio of contextual features and local features. Represents the local information mapping of input features, after Convolution dimensionality reduction to aisle, This represents the final output feature map, which dynamically fuses contextual features and local features through a gating mechanism to balance global correlations and local details.

[0136] In this embodiment, the adaptive multi-branch fusion strategy is as follows:

[0137] 1. Dynamic evaluation of branch importance introduces a task-driven branch weight learning mechanism:

[0138] ,

[0139]

[0140]

[0141] in, Indicates the first The global features of each branch, i.e. , This represents the branch index, with values ​​{1, 2, 3, 4}. This represents the height of the output feature map of a single branch. This represents the width of the output feature map of a single branch. This represents the total number of pixels in the output feature map of a single branch, i.e., the spatial size of the feature map. Indicates the first The spatial location of each branch's output feature map The feature vector at that location, , Its dimensions are C , The four branches represent the global features, and the concatenation of these features results in the total feature vector. This indicates a splicing operation. Let represent the global feature representations for branches 1 through 4, with each vector having a dimension of . (Right now ), express The dimension identifier, due to the 4 branches Each for The total dimension after splicing is , This represents the branch weight vector of the final output, i.e. in Indicates the first The importance weight of each branch satisfies , Indicates the first The weight coefficients of each branch, This represents the softmax activation function. This represents the weight matrix of the first-level fully connected layer. This represents the weight matrix of the second-level fully connected layer. This represents the activation function (here, the ReLU function).

[0142] 2. Weighted Feature Fusion

[0143]

[0144] in, This represents the feature map after multi-branch weighted fusion, with dimensions consistent with the feature maps of each branch. This represents the branch index, with values ​​{1, 2, 3, 4}, corresponding to the 4 branches to be merged. Indicates the first The weight coefficients of each branch, Indicates the first The original feature maps output by each branch, This indicates that the weighted features of the four branches are summed to obtain the fused features;

[0145] 3. Dual Attention Refinement Module

[0146] Channel attention:

[0147] in, Represents the channel attention weight vector. This represents the sigmoid activation function, with an output range of [0,1]. This represents the weight matrix of the first-level channel attention fully connected layer. This represents the weight matrix of the second-level channel attention fully connected layer. This represents the activation function (usually the ReLU function), which introduces a non-linear transformation to enhance the model's expressive power. Indicates the fusion feature map Perform global average pooling. Indicates the fusion feature map Perform global max pooling. This indicates a splicing operation, which splices the outputs of GAP and GMP along the channel dimension;

[0148] Spatial attention:

[0149] in, Represents the spatial attention weight map. This indicates the sigmoid activation function, which normalizes the spatial weights to [0,1]. express Convolution operation, This indicates the fusion feature map along the channel dimension. Perform average pooling. This indicates the fusion feature map along the channel dimension. Perform max pooling. This indicates a concatenation operation, which concatenates the outputs of channel average pooling and channel max pooling along the channel dimension.

[0150] Joint attention:

[0151] in, This represents the feature map after channel and spatial attention refinement. This represents the original feature map after weighted fusion. Represents the channel attention weight vector. Represents the spatial attention weight map. This indicates element-wise multiplication, with channel attention and spatial attention applied sequentially to refine the fused features.

[0152] 4. Feature Recalibration and Output

[0153]

[0154] in, This represents the recalibrated feature map. Represents the recalibration feature map conduct convolution, This represents the initial input feature map. This represents element-wise addition, achieving residual connection (first residual);

[0155]

[0156] The last item is a double residual connection. This represents the final output feature map of the module. This indicates the sigmoid activation function, which will... The output of the convolution is normalized to [0,1]. Represents the recalibration feature map conduct convolution, This represents the feature map after recalibration.

[0157] In this embodiment, a hierarchical perceptual bidirectional feature collaboration module is designed based on the skip connection optimization method. Its flowchart is shown in Figure 3. The hierarchical perceptual bidirectional feature collaboration module includes shallow feature processing of the encoder, deep feature processing of the decoder, and a dual-branch adaptive fusion strategy. The shallow feature processing of the encoder includes spatial-channel decoupling attention, cross-level semantic transmission mechanism, dual-path channel attention generation, and residual enhancement and output. The deep feature processing of the decoder includes multi-scale spatial alignment, scale-aware cross-attention, and channel-space joint modulation. The dual-branch adaptive fusion strategy includes feature difference measurement, difference-aware fusion weight generation, and progressive feature fusion.

[0158] In this embodiment, the hierarchical perception bidirectional feature collaboration module breaks away from the traditional unidirectional guidance paradigm and constructs a bidirectional collaborative enhancement mechanism. It achieves bidirectional information flow through two parallel and interconnected processing paths. In the shallow feature processing path of the encoder, not only is deep semantics transmitted downwards, but a hierarchical difference perception modulation factor is also innovatively introduced. The intensity of semantic guidance is adaptively adjusted according to the actual distance between the shallow and deep feature distributions. When the difference between the two is large, the module will rely more on the feature judgment of the shallow layer itself to avoid excessive semantic suppression. When the difference is small, the proportion of deep semantics is increased to achieve effective semantic enhancement. At the same time, a dual-path channel attention mechanism is designed. It captures fine-grained dependencies between adjacent channels through local interaction paths, models long-range channel relationships through global interaction paths, and achieves dynamic balance between the two paths through learnable weights, thereby generating more accurate channel importance weights.

[0159] In the deep feature processing path of the decoder, the traditional single-scale spatial attention approach is abandoned, and a multi-scale spatial alignment mechanism is proposed: by simultaneously calculating the cross attention between deep semantic features and shallow detail features at three different scales of 7×7, 5×5, and 3×3, the module can capture the overall semantic-detail correspondence of large polyps at the coarse-grained level, accurately locate the local edge features of small polyps at the fine-grained level, and automatically select the most effective scale combination for the current sample through scale-aware weight learning; this design enables deep features to selectively extract and enhance missing detail information from shallow features, achieving reverse calibration;

[0160] In the dual-branch feature fusion stage, a difference-aware adaptive fusion strategy is proposed. First, the difference in the distribution of the two branches is measured by KL divergence, and the consistency of edge details is evaluated by gradient similarity. The feature difference index is formed by combining the two. The global-level fusion weight is generated based on the feature difference index module, and local fusion weights are generated independently for each spatial location. The two are multiplied to form a global-local joint modulation mechanism. This means that different spatial regions can have completely different fusion strategies: the polyp center may rely more on the semantic branch, while the edge region relies more on the detail branch.

[0161] A four-stage progressive fusion process was designed: the first stage performs weighted fusion; the second stage allows the two branches to complement each other through cross-calibration; the third stage explicitly decouples low-frequency semantics and high-frequency details and enhances them respectively; and the fourth stage ensures the complete preservation of the original information through three-level residual connections. This progressive design avoids the information loss that may be caused by one-step fusion and makes the fusion process more robust and controllable.

[0162] In this embodiment, the encoder's shallow layer features are enhanced with hierarchical perception:

[0163] 1. Spatial-channel decoupling attention

[0164] Spatial saliency extraction:

[0165] in, Represents the spatial saliency weighting plot. This indicates the sigmoid activation function, which will... The output of the convolution is normalized to [0,1]. express Convolution operation, This represents the input feature map along the channel dimension. Perform max pooling. This represents the input feature map along the channel dimension. Perform average pooling. This indicates a concatenation operation, which concatenates the outputs of channel max pooling and channel average pooling along the channel dimension. Indicates the input feature map;

[0166]

[0167] in, This represents the feature map after spatial saliency weighting. Represents the original input feature map. Represents the spatial saliency weighting plot. This represents element-wise multiplication;

[0168] Preliminary assessment of the importance of the passage:

[0169]

[0170]

[0171]

[0172] in, This represents the channel global feature vector obtained through global average pooling. Indicates global average pooling. This represents a feature map refined by spatial saliency. This represents the channel global feature vector obtained through global max pooling. Indicates global max pooling. This represents the initial channel weight vector. This indicates the sigmoid activation function, which normalizes the weights to [0,1]. This represents the learnable weight matrix of the fully connected layer. This indicates a splicing operation, which will... and By concatenating along the channel dimension, we obtain a dimension of eigenvectors.

[0173] 2. Cross-level semantic transmission mechanism, introducing semantic modulation factors and hierarchical difference perception:

[0174] Deep semantic embedding vector generation (from the previous stage of the decoder):

[0175]

[0176] Hierarchical difference modeling:

[0177]

[0178]

[0179] in, Represents a deep semantic embedding vector. The channel dimension represents the deep semantic vector, and the feature dimension represents the size of the vector. express The dimension identifier is the length. A one-dimensional vector, This represents the hierarchical difference metric. This represents the L2 norm (Euclidean distance). The channel global average vector representing shallow features. Indicates a linear projection layer. This represents the hierarchical difference modulation factor, with a value range of [0,1]. This represents the sigmoid activation function, which maps the input to the range [0,1], thereby normalizing the modulation factor. Indicates a learnable scalar parameter;

[0180] Adaptive semantic fusion:

[0181]

[0182]

[0183] in, This represents the feature vector after adaptive fusion. This represents the shallow feature weight coefficient. The hierarchical difference modulation factor ([0,1]) is used as a weighting coefficient for deep semantic information. A concatenated vector representing channel statistics of shallow features. The projection concatenation vector representing deep semantics is composed of the deep semantic vectors after two projections. It is pieced together along the channel dimension. express Dimension identifier;

[0184] 3. Dual-path channel attention generation, introducing parallel dual-path processing to fuse features:

[0185] Path A: Local interaction path

[0186]

[0187]

[0188] in, This indicates the size of the one-dimensional convolution kernel, which is an odd number. This represents the adaptive convolution kernel size function, with the number of channels as input. The output is the corresponding convolution kernel size. This enables dynamic matching of kernel size and feature dimension. This indicates the number of channels in the current feature map (as mentioned above). , (The channel dimensions are consistent). This represents the base-2 logarithm of the number of channels. Denotes hyperparameters, where Scaling factor These offsets collectively control the range of kernel size calculations. This indicates that the nearest odd number is selected, ensuring the symmetry of the convolution operation. Indicates the local channel attention weights. This indicates the sigmoid activation function, which normalizes the weights to the range [0,1]. This represents a one-dimensional convolution operation with a kernel size of . (in (for adaptive computation of kernel size), the number of convolution kernels is The step size is 1. This represents the feature vector after adaptive semantic fusion;

[0189] Path B: Global Interaction Path

[0190]

[0191] in, Represents the global channel attention weights, with dimensions of By modeling the long-distance dependencies between all channels using a fully connected layer, This indicates the sigmoid activation function, which normalizes the weights to the range [0,1]. This represents the weight matrix of the first-level fully connected layer. This represents the weight matrix of the second-level fully connected layer. This represents the activation function (usually ReLU), which introduces a nonlinear transformation between two fully connected layers. This represents the feature vector after adaptive semantic fusion;

[0192] Adaptive path fusion:

[0193]

[0194]

[0195] in, This represents the path fusion weight, with a value range of [0,1], used to dynamically balance the contributions of local and global paths. This indicates the fusion feature vector Perform global max pooling (take the maximum value along the channel dimension). This indicates the fusion feature vector Perform global average pooling (average value along the channel dimension). This represents the final channel attention weight, with dimensions of [dimension number missing]. It integrates information from local and global channel interactions. The path fusion weights ([0,1]) are used as local weights. coefficient, Represents the global weight coefficient, and Complementary, as global weight coefficient, Indicates the local channel attention weights. Indicates the global channel attention weight;

[0196] 4. Residual Enhancement and Output

[0197]

[0198]

[0199] in, This represents the feature map after channel attention enhancement, with dimension 1. Highlighting the characteristics of important passages, This represents a feature map refined by spatial saliency. Represents the final channel attention weight It interacts with each spatial location of the feature map through a broadcast mechanism. Indicates the residual enhancement factor. The final output feature map of the module has dimensions of . It combines enhanced features with original features. The feature map after channel enhancement ( ), The original input feature map of the module;

[0200] Attention vectors passed up to the next level (for use in the next layer):

[0201] in This represents the attention vector passed to the upper layer, with dimension . This allows the channel importance information at the current level to be passed to the next level, enabling cross-level attention collaboration. This represents the final channel attention weight at the current level. It is directly used as the transfer vector.

[0202] In this embodiment, the decoder performs detailed reverse calibration of deep features:

[0203] 1. Multi-scale spatial alignment

[0204] Upsampling deep features:

[0205] in, This represents the upsampled deep feature map, with dimensions equal to or greater than the subsequent shallow feature map to be matched (e.g., ...). If they are consistent, it is denoted as , This indicates an upsampling operation (common methods include bilinear interpolation, transposed convolution, etc.). This represents the original deep feature map before upsampling.

[0206] Multi-scale pooling generates query-key-value pairs:

[0207] For scale :

[0208]

[0209]

[0210]

[0211] in, Represents the target pooling scale, with a value of , representing the spatial dimensions (height × width) of the feature map after pooling. Indicates the first The query feature map at the scale has dimensions of From deep upsampling features Generate, used to initiate attention queries. Indicates the first Key feature map at scale, dimension is , by shallow features Generate, used to calculate similarity with query features, Indicates the first Value feature map at scale, dimension is , by shallow features Generate, used to output response features based on attention weights, This indicates an adaptive average pooling operation, which pools the input feature maps uniformly to... Space dimensions, maintaining the number of channels Without changing the model, multi-scale feature alignment is achieved. This represents the upsampled deep feature map, which serves as the source for generating query features. The shallow feature map to be interacted with serves as the source for generating key and value features;

[0212] 2. Scale-aware cross-attention

[0213] Calculate the attention response for each scale:

[0214]

[0215]

[0216] in, Indicates the first Attention matrix at scale, dimension is Each element represents the similarity between the query and the key feature element. Indicates the first Attention response feature map at scale, dimension 1 It is obtained by weighted features of the attention matrix. Indicates a flattening operation, The 3D feature map is reconstructed into a 1D column vector with dimension 1. , This represents the transpose of the flattened key feature vector, with dimension . Used to perform matrix multiplication with the query vector. This represents a scaling factor to prevent the inner product of the query and the key from becoming too large, which could lead to errors. After gradient vanishing, where Let be the dimension of the flattened vector. This represents the Softmax function, calculated along the row dimension of the matrix to ensure that the sum of elements in each row is 1, thus converting similarity into an attention weight distribution. This represents a reshaping operation, which transforms the weighted one-dimensional vector... Restore to 3D feature map;

[0217] Upsample back to original size:

[0218] in, Indicates the first The feature map after upsampling in the scale attention response. This indicates an upsampling operation of a specified size. of Zoom in , Indicates the first Attention response feature map at different scales.

[0219] Scale-weighted adaptive learning:

[0220]

[0221]

[0222] in, Indicates the first The adaptive weighting coefficients for scaling, with values ​​ranging from [0,1]. This represents a multi-scale attention fusion feature map. This represents an exponential function, amplifying the differences between responses at different scales and enhancing the weighting's discriminative power. Represents the response feature map after oversampling Perform global average pooling. Represents all scales of Summation is performed to normalize the weights. This indicates that the weighted response feature maps at different scales are summed element-wise to obtain the multi-scale fusion result;

[0223] 3. Channel-spatial joint modulation

[0224] Channel dimension compression:

[0225] in, Represents the global compressed feature vector in the channel dimension, with dimension 1. Each element corresponds to The global average response of a channel. Indicates global average pooling, along Spatial Dimensions Pooling, will Feature map compression After being flattened The vector, This represents the attention feature map after multi-scale fusion.

[0226] Spatial dimension compression:

[0227] in, Represents a compressed feature vector with spatial dimensions of . Each element corresponds to A global channel response for a spatial location. This represents average pooling along the channel dimension. Indicates a flattening operation, Feature maps reshaped into A one-dimensional vector (of length ) );

[0228] Joint modulation weight generation:

[0229]

[0230]

[0231] in, This represents the channel-space joint modulation weight vector. This indicates the sigmoid activation function, which normalizes the weights to the range [0,1]. Represents the learnable weight matrix. Represents the compression vector of space Perform global max pooling, retrieve the maximum value of all elements, and output a scalar. Key responses characterizing spatial dimensions Representation space refined weight graph, express The convolution operation uses a kernel of 1 and a stride of 1 (keeping the spatial dimensions constant). This indicates a reshaping operation, which will... The spatial compression vector is restored to The feature map is used as the input to the convolution. Represents spatial dimension compressed feature vector ;

[0232] Dual modulation:

[0233] in, This represents the deep feature map after channel-spatial joint modulation. This represents the original deep feature map after upsampling. This represents the channel-space joint modulation weight vector, which is used by the broadcaster for... Each spatial location, Representation space refined weight graph, produced by broadcaster for use Each channel;

[0234] Residual connection:

[0235] in, This represents the deep feature map after modulation and addition of residuals. This represents a deep feature map after joint modulation. This represents the original deep feature map after upsampling.

[0236] In this embodiment, the dual-branch adaptive fusion strategy is as follows:

[0237] 1. Measurement of Feature Difference

[0238] Introducing a dual metric: feature distribution distance and gradient similarity.

[0239] Distribution distance:

[0240] ,

[0241] ,

[0242]

[0243] in, Representing shallow feature maps The channel mean vector, Representing shallow feature maps The channel variance vector, Representing deep feature maps The channel mean vector, Representing deep feature maps The channel variance vector, Indicates global average pooling. This represents the variance calculation function. This represents the output feature map of the shallow branch. The output feature map represents the deep branch. This represents the approximate KL divergence. This represents the ratio of the variance of shallow to deep channels. This represents the ratio of the square of the mean difference to the depth variance. The logarithm of the variance ratio further describes the differences in distribution patterns;

[0244] Gradient similarity:

[0245]

[0246]

[0247] in Representing shallow feature maps Horizontal gradient plot, Representing deep feature maps Horizontal gradient plot, express Along the width dimension Extract a sub-image from column 2 to the last column. express Extract a sub-image along the width dimension from column 1 to the second-to-last column;

[0248]

[0249] in This represents the gradient cosine similarity, with values ​​ranging from [−1, 1]. Represents the vector dot product. and They represent and The L2 norm, Represents the smoothing factor, with values ​​ranging from 1 to 2. ;

[0250] Comprehensive difference indicators:

[0251] in, Indicators representing comprehensive differences These represent balance coefficients, 0.6 and 0.4 respectively, controlling the weights of distribution distance and gradient similarity in the overall index. This indicates that the sigmoid activation function acts on... The distribution distance is normalized to the range [0,1]. The complementary term representing gradient similarity will Convert to a difference measure;

[0252] 2. Difference-aware fusion weight generation

[0253] Basic fusion weights:

[0254] in, This represents the basic fusion weight vector, with dimension . , in the form of ,in These are the global basic weights for shallow features. These are the global basic weights for deep features. Both belong to [0,1] and jointly reflect the feature importance distribution at the global level. This indicates that the sigmoid activation function maps the output to the range [0,1]. Represents the learnable weight matrix. This indicates a feature concatenation operation. Representing shallow feature maps The channel mean vector, Representing deep feature maps The channel mean vector, This indicates a comprehensive index of differences.

[0255] Spatial adaptive fusion weights:

[0256]

[0257]

[0258]

[0259]

[0260] in, Spatial weighted feature map representing shallow features. Spatial weighted feature map representing deep features express Convolution operation, This represents the output feature map of the shallow branch. The output feature map represents the deep branch. A spatial adaptive weighted graph representing shallow features. A spatial adaptive weighted graph representing deep features. Represents an exponential function, denominator This represents an exponential summation of the spatial weight characteristics of shallow and deep layers;

[0261] Global-local fusion weight combination:

[0262]

[0263]

[0264] in The final fusion weight map representing shallow features. The final fusion weight map of deep features. This represents the first element of the basic fusion weight vector, i.e., the global basic weight of the shallow features. , The second element represents the global base weights of the deep features. , A spatial adaptive weighted graph representing shallow features. A spatial adaptive weighted graph representing deep features;

[0265] In this embodiment, unlike direct splicing, a progressive fusion + self-calibration mechanism is adopted:

[0266] Phase 1: Weighted Fusion

[0267]

[0268] in, This represents the first-stage weighted fusion feature map. The final fusion weight map representing shallow features ( (from global-local weight combination). The final fusion weight map representing deep features. This represents the output feature map of the shallow branch. The output feature map represents the deep branch;

[0269] Phase Two: Cross-calibration

[0270]

[0271]

[0272]

[0273] in This represents the cross-calibration feature between shallow features and the first-stage fusion features. The cross-calibration feature represents the fusion feature between deep features and the first-stage features. This represents the feature map after the second stage of cross-calibration. This indicates a splicing operation. express Convolution operation, This represents the cross-calibration strength coefficient (with a value of 0.1).

[0274] Phase 3: Semantic-Detail Decoupling Enhancement

[0275] Separate low-frequency and high-frequency components:

[0276]

[0277]

[0278] in Represents low-frequency characteristic components. Represents high-frequency characteristic components, express Average pooling, This represents the feature map after the second stage of cross-calibration. );

[0279] Enhanced high-frequency details:

[0280]

[0281] in, This indicates the enhanced high-frequency features. Represents the absolute value of the high-frequency component. express Convolution operation, This represents the sigmoid activation function, which normalizes the convolution output to [0,1]. This represents the high-frequency enhancement factor (with a value of 0.3). Indicates the enhancing factor;

[0282] Recombination characteristics:

[0283]

[0284] in, This represents the final feature map after recombination. Represents low-frequency semantic components ( ), This indicates the enhanced high-frequency detail components ( );

[0285] Phase 4: Dual Residuals and Output

[0286]

[0287] in, This indicates that the output feature map is fed into the next higher level, and its dimensions match the target level. Indicates two levels convolution, The final feature map after recombination ( ), This indicates a downsampling operation. Residual fusion representing shallow and deep features ( ).

[0288] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A polyp segmentation method based on adaptive multi-path fusion and hierarchical perception collaboration, characterized in that, The method includes the following steps: The polyp image to be segmented is input into a colorectal polyp segmentation model comprising an adaptive multi-path polyp feature fusion module and a hierarchical perceptual bidirectional feature collaboration module; the adaptive multi-path polyp feature fusion module first performs adaptive channel redistribution preprocessing, then uses a local detail-aware branch, a multi-scale gradient aggregation branch, an edge-oriented enhancement branch, and a global context modeling branch for parallel feature extraction, adaptively adjusting path weights according to polyp features, and achieving feature calibration through dual attention and dual residual connections; the local detail-aware branch retains high-frequency texture information through lightweight convolution; the multi-scale gradient aggregation branch uses parallel multi-kernel convolution and introduces a gradient-weighted aggregation mechanism for scale adaptation and enhanced boundary perception processing; the edge-oriented enhancement branch processes edge features through a cross-directional attention mechanism; the global context modeling branch utilizes… Long-range dependencies are captured using a self-attention mechanism. The final output feature map of the adaptive multi-path polyp feature fusion module is input to the hierarchical perceptual bidirectional feature collaboration module. Feature fusion is performed through the shallow feature processing path of the encoder and the deep feature processing path of the decoder. A dual-branch adaptive fusion strategy is adopted during the fusion process to output the final polyp segmentation result map corresponding to the fused features. The dual-branch adaptive fusion strategy includes feature difference measurement, difference-aware fusion weight generation, and progressive feature fusion. The feature difference measurement includes both feature distribution distance and gradient similarity. The progressive feature fusion performs weighted fusion in the first stage, cross-calibration in the second stage, semantic and detail decoupling enhancement in the third stage, and three-level residual connection and output in the fourth stage. The third stage includes separating low-frequency and high-frequency components, enhancing high-frequency details, and recombining features.

2. The polyp segmentation method based on adaptive multi-path fusion and hierarchical perception collaboration as described in claim 1, characterized in that, The adaptive channel reallocation preprocessing includes: in, Represents the input feature map, where These represent the height and width of the feature map, respectively. For the number of channels, express Convolution operation, Indicates The initial feature map after convolution processing. Indicates global average pooling. This represents the learnable weight matrix of the first fully connected layer, enabling the number of channels to increase from... Compress to ， This represents the learnable weight matrix of the second fully connected layer, enabling the number of channels to increase from... Restore to ， Represents the ReLU activation function. This represents the Sigmoid activation function. Represents the channel attention weight vector. This indicates element-wise multiplication, which multiplies the weight vector. With the initial feature map Each channel is matched and weighted. This represents the output feature map after channel attention calibration.

3. The polyp segmentation method based on adaptive multi-path fusion and hierarchical perception collaboration as described in claim 1, characterized in that, The feature calibration achieved through dual attention and dual residual connections includes: channel attention: in, Represents the channel attention weight vector. This represents the sigmoid activation function. This represents the weight matrix of the first-level channel attention fully connected layer. This represents the weight matrix of the second-level channel attention fully connected layer. Represents the ReLU activation function. Indicates the fusion feature map Perform global average pooling. Indicates the fusion feature map Perform global max pooling. This indicates a splicing operation, concatenating the outputs of GAP and GMP along the channel dimension; spatial attention: in Represents the spatial attention weight map. express Convolution operation, This indicates the fusion feature map along the channel dimension. Perform average pooling. This indicates the fusion feature map along the channel dimension. Perform max pooling. This indicates a concatenation operation, which concatenates the outputs of channel average pooling and channel max pooling along the channel dimension; first residual connection: in This represents the recalibrated feature map. This represents the feature map after joint attention using channel attention and spatial attention. Indicates the feature map conduct convolution, The feature map representing the output of the adaptive channel redistribution preprocessing. This represents element-wise addition, implementing residual join; double residual join: in, This represents the final output feature map of the adaptive multipath polyp feature fusion module. Represents the recalibration feature map conduct convolution.

4. The polyp segmentation method based on adaptive multi-path fusion and hierarchical perception collaboration as described in claim 1, characterized in that, The decoder's deep feature processing path captures the overall semantic and detail correspondence of large polyps by simultaneously calculating the cross-attention between deep semantic features and shallow detail features at three different scales: 7×7, 5×5, and 3×3. It also locates the local edge features of small polyps and automatically selects scale combinations through scale-aware weight learning.

5. The polyp segmentation method based on adaptive multi-path fusion and hierarchical perception collaboration as described in claim 1, characterized in that, The distribution distance: ， ， in, Representing shallow feature maps The channel mean vector, Representing shallow feature maps The channel variance vector, Representing deep feature maps The channel mean vector, Representing deep feature maps The channel variance vector, Indicates global average pooling. This represents the variance calculation function. This represents the output feature map of the encoder's shallow feature processing path. This represents the output feature map of the decoder's deep feature processing path. This represents the approximate KL divergence. This represents the ratio of the variance of shallow to deep channels. This represents the ratio of the square of the mean difference to the depth variance. The logarithm of the variance ratio; the gradient similarity: in, Representing shallow feature maps Horizontal gradient plot, Representing deep feature maps Horizontal gradient plot, express Along the width dimension Extract a sub-image from column 2 to the last column. express Extract a subplot along the width dimension from column 1 to the second-to-last column; Combine the dissimilarity index: in, This represents the gradient cosine similarity, with values ​​ranging from [−1, 1]. Represents the vector dot product. and They represent and L2 norm, Represents the smoothing factor. Indicators representing comprehensive differences Represents the balance coefficient. This indicates that the Sigmoid activation function acts on... The distribution distance is normalized to the range [0,1]. The complementary term representing gradient similarity will Convert to a difference measure.

6. The polyp segmentation method based on adaptive multi-path fusion and hierarchical perception collaboration as described in claim 5, characterized in that, The generation of the difference-aware fusion weights includes: basic fusion weights: in, Represents the basic fusion weight vector. This represents the Sigmoid activation function. Represents the learnable weight matrix. Indicates feature concatenation operation; spatial adaptive fusion weights: in, Spatial weighted feature map representing shallow features. Spatial weighted feature map representing deep features express Convolution operation, A spatial adaptive weighted graph representing shallow features. A spatial adaptive weighted graph representing deep features. Represents an exponential function; global-local fusion weight combination: in, The final fusion weight map representing shallow features. The final fusion weight map of deep features. This represents the first element of the basic fusion weight vector. This represents the second element of the basic fusion weight vector.

7. The polyp segmentation method based on adaptive multi-path fusion and hierarchical perception collaboration as described in claim 1, characterized in that, The weighted fusion: in, This represents the first-stage weighted fusion feature map. The final fusion weight map representing shallow features. The final fusion weight map representing deep features. This represents the output feature map of the encoder's shallow feature processing path. The output feature map represents the deep feature processing path of the decoder; the cross-calibration: in, This represents the cross-calibration feature between shallow features and the first-stage fusion features. The cross-calibration feature represents the fusion feature between deep features and the first-stage features. This represents the feature map after the second stage of cross-calibration. This indicates a splicing operation. express Convolution operation, Indicates the cross-calibration intensity coefficient; the separation of low-frequency and high-frequency components: in, Represents low-frequency characteristic components. Represents high-frequency characteristic components, express Average pooling, This represents the feature map after the second-stage cross-calibration; the enhanced high-frequency details: in, This indicates the enhanced high-frequency features. Represents the absolute value of the high-frequency component. express Convolution operation, This represents the Sigmoid activation function. Indicates the high-frequency enhancement factor. Indicates the enhancing factor; the recombination feature: in, This represents the final feature map after recombination. Indicates low-frequency semantic components. This represents the enhanced high-frequency detail components; the third-level residual connection: in, This indicates that the feature map is output to the next higher level. Indicates two levels convolution, This represents the final feature map after recombination. This indicates a downsampling operation.

Images