A polyp image segmentation system based on frequency-space dual-domain perception and feature interaction

Abstract

The application provides a polyp image segmentation system based on frequency-space dual-domain perception and feature interaction in the technical field of computer vision and medical image processing, which comprises a data division module for preprocessing and dividing a data set; a parallel feature extraction network, wherein local texture and edge space features are acquired by a CNN branch, and high-frequency details and low-frequency global semantic features are respectively extracted by a spatial window attention and a frequency domain complex filter through a Transformer branch; a dynamic feature interaction module enhances dual-domain features by using a cross-gating mechanism; a frequency-space gating fusion module generates adaptive weights, and weightedly fuses and smoothes CNN spatial features and enhanced features; a training module completes model training, verification and deployment; and a segmentation module outputs a target region result for an input image.The application has the advantages that the accuracy, boundary definition and robustness of polyp segmentation in a complex clinical scene are greatly improved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and medical image processing technology, and in particular to a polyp image segmentation system with frequency-space dual-domain perception and feature interaction. Background Technology

[0002] In recent years, deep learning-based segmentation models have been widely used in colonoscopy polyp detection and assisted diagnosis. Based on differences in core feature extraction architecture, existing technologies are mainly divided into two categories: methods based on convolutional neural networks (CNNs) and methods based on Transformers.

[0003] CNN-based methods, represented by U-Net and its variants (such as UNet++ and Attention U-Net), employ an encoder-decoder structure. They extract low-level spatial features like image texture and edges through the local receptive field of convolutional kernels and recover spatial details of the target using skip connections, achieving pixel-level segmentation. For example, Chinese invention patent CN117274226A proposes a parallel diagnostic method for colon polyps based on enhanced U-shaped networks, improving local feature utilization through multi-layer convolution and feature enhancement modules. However, the effective receptive field of CNNs is limited by the size of the convolutional kernels and the network depth, making it difficult to capture global contextual information far from the target region. When polyps and normal tissue have similar gray levels or complex backgrounds, the model lacks sufficient global discrimination criteria, easily leading to missed detections of small-scale or blurred-boundary polyps.

[0004] To overcome the inherent limitation of CNNs' receptive field, Transformer-based methods (such as VisionTransformer and Swin Transformer) have been introduced into the field of medical image segmentation. These methods establish inter-pixel contextual relationships globally through self-attention mechanisms, extracting feature representations rich in global semantic information, and exhibit significant segmentation advantages for polyps with varied morphologies and complex backgrounds. However, self-attention mechanisms are inherently weaker than CNNs in characterizing high-frequency spatial information such as local edges and textures, often resulting in overly smoothed segmentation results at boundaries, losing crucial structural details and reducing segmentation precision.

[0005] To combine the complementary advantages of the two methods mentioned above, hybrid architectures have become the mainstream research direction. Such systems are typically designed as dual-branch or multi-branch parallel networks: one branch uses a CNN to extract local detail features, while the other branch uses a Transformer to capture the global context. Finally, the two outputs are fused through concatenation or element-wise addition. However, this fusion method is too coarse, failing to address the fundamental gap between the two types of features in terms of semantic hierarchy and spatial scale. This results in the fused features failing to achieve true complementary advantages, often leading to feature redundancy or mutual inhibition, thus limiting the upper limit of model performance.

[0006] Furthermore, recent studies have begun to explore the potential of frequency domain information in medical image segmentation. For example, Chinese invention patent CN119832011A proposes a medical image segmentation method based on wavelet transform and attention mechanism. This method introduces wavelet transform during the encoding stage to obtain different frequency components of the image and uses an attention mechanism for feature enhancement. However, this type of method still treats frequency domain processing as an external preprocessing or auxiliary enhancement module, failing to achieve deep collaboration and interaction between frequency domain information (containing global structure and periodic patterns) and spatial domain information (containing local details and location information). Due to the lack of noise identification and edge enhancement mechanisms based on frequency domain information, spatial domain features are easily and severely interfered with when imaging noise is present, leading to blurred segmentation boundaries and decreased robustness.

[0007] In summary, existing polyp segmentation technologies generally suffer from the following interconnected bottlenecks: First, the challenge lies in representing local details and global semantics; neither a single architecture nor a simple fusion method can adequately balance segmentation precision and context awareness. Second, there is a lack of deep collaboration between spatial and frequency domain information; frequency domain information serves only as an auxiliary tool and fails to effectively guide spatial domain feature learning. Third, multi-branch feature interaction and fusion mechanisms are singular and static, unable to dynamically adjust the feature contribution of different branches or scales based on input content, and lack a bidirectional calibration mechanism between high-frequency details and low-frequency semantics, resulting in insufficient model localization and segmentation capabilities for complex boundaries and small-scale targets. These bottlenecks are intertwined, severely restricting the system's segmentation accuracy, boundary recognition ability, and robustness in complex clinical scenarios.

[0008] Therefore, how to provide a polyp image segmentation system with frequency and space dual-domain perception and feature interaction to achieve dynamic collaboration and deep fusion between local and global features and spatial and frequency domain features, thereby improving the accuracy, boundary clarity and robustness of polyp segmentation in complex clinical scenarios, has become an urgent technical problem to be solved. Summary of the Invention

[0009] The technical problem to be solved by this invention is to provide a polyp image segmentation system with frequency-space dual-domain perception and feature interaction, so as to realize dynamic collaboration and deep fusion between local-global features and spatial-frequency features, thereby improving the accuracy, boundary clarity and robustness of polyp segmentation in complex clinical scenarios.

[0010] This invention provides a polyp image segmentation system based on frequency-space dual-domain perception and feature interaction, comprising the following modules: The polyp medical image dataset partitioning module is used to acquire a polyp medical image dataset, and to preprocess and partition the polyp medical image dataset to obtain a training set, a validation set and a test set; The frequency-space dual-domain feature extraction module is used to construct a parallel feature extraction network that includes a convolutional neural network branch and a frequency-space dual-domain perceptual Transformer branch; The convolutional neural network branch is used to perform multi-level feature extraction on the input polyp medical image to obtain spatial features containing local texture information and edge structure information. The frequency-space dual-domain perception Transformer branch is used to extract high-frequency detail features through a spatial domain local window attention mechanism and extract low-frequency global semantic features through a frequency domain complex learnable filter to obtain frequency-space dual-domain features. The dynamic feature interaction module is used to construct a dynamic feature interaction network to perform information interaction and collaborative enhancement on the high-frequency detail features and low-frequency global semantic features in the frequency-space dual-domain features through a cross-gating mechanism, thereby obtaining enhanced dual-domain features. The frequency-space gated fusion module is used to construct a frequency-space gated fusion network to generate adaptive gate weights based on the enhanced dual-domain features, use the adaptive gate weights to weight the spatial features output by the convolutional neural network branches, fuse the weighted spatial features with the enhanced dual-domain features, and obtain the fused features through feature smoothing processing. The segmentation network model training module is used to train, validate, and test the segmentation network model created based on the parallel feature extraction network, dynamic feature interaction network, frequency-space gated fusion network, and decoder using the training set, validation set, and test set, respectively, and to deploy the segmentation network model that passes the test. The polyp medical image segmentation module is used to acquire the medical image of the polyp to be segmented, input it into the deployed segmentation network model, and output the corresponding target region segmentation result.

[0011] Furthermore, the polyp medical image dataset partitioning module is specifically used for: A polyp medical image dataset containing polyp regions and their annotation information is obtained. The polyp medical images in the polyp medical image dataset are preprocessed by size normalization, pixel value standardization and data augmentation. The preprocessed polyp medical image dataset is divided into training set, validation set and test set according to a preset ratio for subsequent model training, validation and performance testing.

[0012] Furthermore, in the frequency-space dual-domain feature extraction module, the frequency-space dual-domain perceptual Transformer branch is specifically used to perform channel division on the input polyp medical image to obtain high-frequency features and low-frequency features; to perform local feature modeling on the high-frequency features using a spatial domain local window attention mechanism to extract high-frequency detail features; to perform frequency domain transformation and global feature modeling on the low-frequency features using a frequency domain complex learnable filter to extract low-frequency global semantic features; and to fuse the high-frequency detail features and low-frequency global semantic features to obtain frequency-space dual-domain features.

[0013] Furthermore, the dynamic feature interaction module is specifically used for: Based on the high-frequency detail features and low-frequency global semantic features in the frequency-space dual-domain features, a first gating weight and a second gating weight are generated respectively. Through a cross-gating mechanism, the low-frequency global semantic features are weighted based on the first gating weight, and the high-frequency detail features are weighted based on the second gating weight, so as to perform information interaction and collaborative enhancement, thereby obtaining enhanced dual-domain features.

[0014] Furthermore, in the segmentation network model training module, the frequency-space gating fusion module is used to generate adaptive gating weights based on the enhanced dual-domain features, and to use the adaptive gating weights to weight the spatial features output by the convolutional neural network branches in order to suppress noise and enhance effective features; The weighted spatial features are fused with the enhanced dual-domain features, and the fused features are obtained through feature smoothing.

[0015] Furthermore, the frequency-space gating fusion module is set between feature layers of different scales, and multi-scale feature aggregation is achieved by fusing deep features, mid-level features and shallow features step by step.

[0016] Furthermore, in the segmentation network model training module, the segmentation network model adopts a network structure of encoder and decoder; The encoder consists of a convolutional neural network branch and a frequency-space dual-domain sensing Transformer branch; The encoder and decoder transmit features via skip connections; The decoder restores the spatial resolution of the fused features through step-by-step upsampling and outputs the target region segmentation result.

[0017] Furthermore, the decoder includes multiple decoding layers. Each decoding layer gradually recovers the spatial resolution of the fused features through convolution and upsampling operations, and combines the encoded features of the corresponding scale to perform feature reconstruction to obtain the target region segmentation result.

[0018] Furthermore, in the segmentation network model training module, the process of training, validating, and testing the segmentation network model specifically includes: The segmentation network model is trained using the training set, its performance is validated and parameters are adjusted using the validation set, and its performance is tested using the test set.

[0019] The advantages of this invention are: 1. By constructing a parallel dual-branch architecture of convolutional neural network and frequency-space dual-domain perception Transformer, the CNN branch fully preserves local texture and edge structure to maintain the boundary foundation, while the Transformer branch uses spatial domain local window attention to capture high-frequency details and uses frequency domain complex learnable filters to extract low-frequency global semantics, thus fundamentally taking into account both local fineness and global context awareness. On this basis, the dynamic feature interaction module realizes bidirectional calibration and collaborative enhancement of high-frequency details and low-frequency semantics through cross-gating mechanism, while the frequency-space gated fusion network generates adaptive weights based on the enhanced dual-domain features, performs targeted noise suppression and effective feature enhancement on the spatial features of the CNN branch, and deeply fuses the weighted spatial features with the enhanced dual-domain features. With the help of a multi-scale hierarchical aggregation strategy, it effectively copes with complex clinical interferences such as polyps and tissues having similar gray levels, complex backgrounds, specular reflections, and imaging noise. Finally, it systematically improves the global localization accuracy, boundary contour clarity, and model anti-interference robustness of segmentation.

[0020] 2. By constructing a parallel dual-branch feature extraction network—the convolutional neural network branch focuses on extracting spatial features containing local texture information and edge structure information, while the frequency-space dual-domain perceptual Transformer branch uses the spatial domain local window attention mechanism to extract high-frequency detail features, and at the same time uses the frequency domain complex learnable filter to extract low-frequency global semantic features—the system can simultaneously retain the local detail depiction capability of CNN and the global context modeling capability of Transformer in the same network, thus resolving the perceptual bias of a single architecture from the source, and providing a solid foundation for information complementarity for subsequent feature fusion, rather than simply splicing two types of features.

[0021] 3. Parallel processing of spatial and frequency domains in the Transformer branch: Local window attention in the spatial domain focuses on high-frequency details, while complex learnable filters in the frequency domain explicitly model low-frequency global semantics. The two are fused to form "frequency-space dual-domain features." This design does not simply superimpose frequency domain features, but uses a cross-gating mechanism in the dynamic feature interaction module to allow high-frequency detail features and low-frequency global semantic features to interact and calibrate bidirectionally—for example, using gating weights generated from high-frequency details to weight low-frequency semantics, and vice versa. This deep collaboration allows the global structure and periodic patterns inherent in the frequency domain to actively guide feature learning in the spatial domain, while the spatial domain's positional information can also refine the frequency domain representation, truly achieving synergistic enhancement of "using frequency to assist space and using space to regulate frequency," overcoming the shortcomings of isolated use of frequency domain information in existing technologies.

[0022] 4. Design a dual dynamic control mechanism: On the one hand, the dynamic feature interaction module generates first and second gating weights based on high-frequency and low-frequency features respectively. Through cross-weighting, dynamic information interaction between high-frequency and low-frequency features is achieved, allowing them to adaptively enhance each other according to the input content. On the other hand, the frequency-space gating fusion module generates adaptive gating weights based on the enhanced dual-domain features, weighting the spatial features output by the CNN branch. This effectively suppresses noise and enhances effective features, and then the weighted spatial features are fused with the enhanced dual-domain features. This process does not use fixed weights but dynamically changes with the input image content, enabling the fusion strategy to adaptively adjust according to the size, shape, boundary clarity, and background complexity of the polyps. This greatly improves the targeting and effectiveness of the fusion, avoiding the performance bottleneck caused by static fusion.

[0023] 5. At the encoding end, both high-frequency details (through local window attention and frequency domain complex filters) and low-frequency global semantics are preserved. The deep calibration of both is achieved through cross-gating and adaptive fusion, making the network particularly sensitive to high-frequency information such as edges and textures, while also providing a global basis for overall structure discrimination. When polyps and normal tissues have similar gray levels or complex backgrounds, the low-frequency global semantic features can provide discrimination clues, while the high-frequency detail features can be used to accurately locate the boundaries. Thus, when the decoder upsamples step by step to restore resolution, it can reconstruct target areas with clear contours and sharp boundaries, significantly improving the detection rate and segmentation precision of small polyps and polyps with blurred boundaries.

[0024] 6. Low-frequency global semantic features extracted through a frequency-domain complex learnable filter naturally possess the ability to suppress high-frequency noise (because noise is usually manifested as a high-frequency component), while high-frequency detail features extracted through spatial domain local window attention focus on meaningful edges and textures. After dynamic interaction and gating fusion, the two can effectively distinguish between noisy high frequencies and structural high frequencies. More importantly, the adaptive gating weights can dynamically adjust the contribution of different components according to the feature quality—when noise is severe in a certain area, the corresponding gating weights are automatically reduced, thereby suppressing noise propagation. This mechanism enables the system to maintain stable segmentation performance on noisy endoscopic images acquired in actual clinical practice, significantly improving the clinical applicability and generalization ability of the model.

[0025] 7. The internal processing logic of the frequency-space dual-domain perception Transformer branch is as follows: First, the input features are divided into channels to obtain high-frequency features and low-frequency features respectively. Then, they are fed into the spatial domain local window attention mechanism and the frequency domain complex learnable filter for independent modeling, and finally fused. This "separate then combine" design has significant engineering advantages compared to directly performing mixed frequency-space processing on all channels. From the perspective of representation learning, channel division forces the network to project different frequency components onto different feature subspaces, realizing the decoupled representation of frequency domain information. This avoids mutual interference or cancellation between high-frequency edge details and low-frequency global structure in the same feature channel, reducing the learning redundancy of the model. From the perspective of computational efficiency, the local window attention only operates on the high-frequency channel subset, and the complex filter only operates on the low-frequency channel subset. Compared to performing spatial and frequency domain transformations on all channels simultaneously, this greatly reduces the amount of computation and memory usage, enabling the model to maintain high performance while having better training and inference efficiency, and making it easier to deploy in real-world clinical environments with limited computing resources.

[0026] 8. First and second gating weights are generated based on high-frequency detail features and low-frequency global semantic features, respectively. Then, a cross-gating mechanism is used—that is, the weights derived from high frequencies are used to weight low-frequency semantics, and the weights derived from low frequencies are used to weight high-frequency details. This bidirectional cross-weighting mechanism is not merely a simple importance recalibration; its deeper advantage lies in achieving semantic alignment and mutual calibration between high and low frequencies. Specifically, high-frequency detail features are rich in edge location information but lack global context. The gating weights generated from them, when applied to low-frequency semantic features, can guide global semantic features to "focus" on boundary regions spatially, enhancing the spatial perception of key locations by low-frequency features. Conversely, low-frequency global semantic features are rich in category and structural priors. The gating weights generated from them, when applied to high-frequency detail features, can suppress false high-frequency responses caused solely by noise or texture, ensuring that high-frequency details are enhanced while maintaining semantic correctness. This bidirectional calibration effectively bridges the inherent gap between high-frequency and low-frequency features in semantic hierarchy and spatial distribution, truly integrating the two heterogeneous features into a synergistically consistent enhanced dual-domain feature.

[0027] 9. By placing the frequency-space gating fusion module between feature layers of different scales, multi-scale feature aggregation is achieved through the stepwise fusion of deep, mid, and shallow features. That is, the fusion strategy does not perform a one-time fusion at a single level, but deploys the gating fusion module in multiple downsampling stages of the encoder, forming a progressive fusion chain throughout the entire encoding path. The significant advantage of this design is that deep features are rich in semantics but have low spatial resolution, while shallow features are rich in detail but have weakened semantics. By introducing adaptive gating fusion at different scales, the spatial features at each level can receive guidance and control from the frequency-space dual-domain enhanced features at their most suitable scale—deep fusion focuses on global structural calibration, mid-level fusion balances semantics and detail, and shallow fusion emphasizes fine edge characterization. Finally, the fused features from each level are passed to the decoder via skip connections, enabling the decoder to obtain semantically matched and scale-adapted fused features at different upsampling stages. This effectively alleviates the scale mismatch and information loss problems in traditional single fusion, and is particularly beneficial for segmenting multiple polyps with significant size differences.

[0028] 10. The segmentation network model adopts an encoder-decoder structure, with feature transfer between the encoder and decoder via skip connections. The decoder includes multiple decoding layers, each of which gradually restores resolution through convolution and upsampling, and reconstructs the image by combining the encoded features of the corresponding scale. The deep advantage of this design is that although the enhanced dual-domain features fuse rich frequency-space information, the accuracy of spatial coordinates will inevitably be lost after multiple downsampling layers. Skip connections directly transfer the unfused original spatial features (especially shallow high-resolution features) from each stage of the encoder to the corresponding scale decoding layer, providing the decoder with "spatial anchor points" to ensure that pixel positioning during the upsampling process does not deviate from the real anatomical structure. At the same time, when restoring resolution, each decoding layer does not simply rely on bilinear interpolation, but combines the encoded features of the corresponding scale for convolutional reconstruction. This "feature-guided upsampling" can effectively suppress the checkerboard effect and jagged edges that are easily generated by traditional upsampling, so that the final output segmentation mask maintains global semantic consistency while presenting a fine boundary contour that highly matches the original image, greatly improving the credibility and clinical usability of the segmentation results. Attached Figure Description

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0030] Figure 1 This is a schematic diagram of the structure of a polyp image segmentation system based on frequency-space dual-domain perception and feature interaction according to the present invention.

[0031] Figure 2 This is a network architecture diagram of the segmentation network model of this invention.

[0032] Figure 3 This is a schematic diagram of the spatial domain local window attention in the frequency-space dual-domain perception Transformer branch of this invention.

[0033] Figure 4 This is a schematic diagram of the dynamic feature interaction network and frequency-space gating fusion network of the present invention.

[0034] Figure 5 This is a schematic diagram of the collaborative extraction of frequency domain complex learnable filters according to the present invention. Detailed Implementation

[0035] The overall approach of the technical solution in this application is as follows: A polyp image segmentation system based on frequency-space dual-domain perception and feature interaction is constructed. This system extracts local spatial features and frequency-space dual-domain features containing high-frequency details and low-frequency global semantics by designing parallel convolutional neural network branches and frequency-space dual-domain perception Transformer branches. A cross-gating mechanism in the dynamic feature interaction module is used to achieve bidirectional calibration and collaborative enhancement between high-frequency and low-frequency features. Subsequently, a frequency-space gating fusion module adaptively generates gating weights based on the enhanced dual-domain features to suppress noise and effectively enhance the spatial features of the CNN branch. The weighted spatial features are then deeply fused with the enhanced dual-domain features. Finally, an encoder-decoder structure with skip connections is used to progressively restore resolution and output a refined segmentation result. This systematically improves the global localization accuracy, boundary clarity, and model robustness of polyp segmentation in complex clinical scenarios.

[0036] Please refer to Figures 1 to 5 As shown, a preferred embodiment of the polyp image segmentation system based on frequency-space dual-domain perception and feature interaction of the present invention includes the following modules: The polyp medical image dataset partitioning module is used to acquire a polyp medical image dataset, and to preprocess and partition the polyp medical image dataset to obtain a training set, a validation set and a test set; In practice, the publicly available polyp datasets Kvasir-SEG and CVC-ClinicDB can be selected as experimental datasets. Data augmentation processing such as normalization, random flipping, and random rotation of the images can be performed, and the datasets can be divided into training set, validation set, and test set in a ratio of 8:1:1.

[0037] The frequency-space dual-domain feature extraction module is used to construct a parallel feature extraction network that includes a convolutional neural network branch and a frequency-space dual-domain perceptual Transformer branch; The convolutional neural network branch is used to perform multi-level feature extraction on the input polyp medical image to obtain spatial features of the lesion region containing local texture information and edge structure information; The frequency-space dual-domain perception Transformer branch is used to extract high-frequency detail features through a spatial domain local window attention mechanism and extract low-frequency global semantic features through a frequency domain complex learnable filter to obtain frequency-space dual-domain features. This involves performing collaborative feature modeling in the spatial and frequency domains on the input medical images of polyps. This includes: dividing the input features into channels to obtain spatial and frequency domain features; using a windowing mechanism to perform local attention modeling on the spatial domain features to obtain high-frequency structural information; mapping the frequency domain features to the frequency domain space using a two-dimensional fast Fourier transform and performing global semantic modulation using a complex learnable filter; restoring the processed frequency domain features to the spatial domain using an inverse Fourier transform to obtain enhanced frequency domain features; and achieving information interaction and adaptive fusion between the spatial and frequency domain features through a cross-gating mechanism to obtain dual-domain features (spatial and frequency domains).

[0038] The dynamic feature interaction module is used to construct a dynamic feature interaction network to perform information interaction and collaborative enhancement on the high-frequency detail features and low-frequency global semantic features in the frequency-space dual-domain features through a cross-gating mechanism, thereby obtaining enhanced dual-domain features. This is used to achieve bidirectional information modulation between spatial domain features and frequency domain features, including: generating frequency domain modulation weights based on spatial domain features; generating spatial domain modulation weights based on frequency domain features; performing weighted enhancement on spatial domain features and frequency domain features respectively; and performing channel-level splicing and linear mapping on the enhanced dual-domain features to obtain the final enhanced dual-domain features.

[0039] The frequency-space gated fusion module is used to construct a frequency-space gated fusion network to generate adaptive gate weights based on the enhanced dual-domain features, use the adaptive gate weights to weight the spatial features output by the convolutional neural network branches, fuse the weighted spatial features with the enhanced dual-domain features, and obtain the fused features through feature smoothing processing. The segmentation network model training module is used to train, validate, and test the segmentation network model created based on the parallel feature extraction network, dynamic feature interaction network, frequency-space gated fusion network, and decoder using the training set, validation set, and test set, respectively, and to deploy the segmentation network model that passes the test. The polyp medical image segmentation module is used to acquire the medical image of the polyp to be segmented, input it into the deployed segmentation network model, and output the corresponding target region segmentation result.

[0040] Specifically, the input medical image is represented as follows: Where: H and W represent the spatial resolution of the medical image, respectively; C represents the number of channels; and X represents the medical image tensor in the real number domain.

[0041] In this invention, to adapt to the Transformer-based sequence modeling approach, the two-dimensional feature map is expanded into a sequence form: Where: B represents the batch size; N = H × W represents the number of tokens; This represents the feature dimension. This transformation is used to achieve a unified representation between spatial features and sequence modeling, enabling subsequent spatial attention and frequency domain modeling to interact within the same feature space.

[0042] This invention proposes a frequency-space dual-domain sensing Transformer branch to address the problems of weak local details, insufficient global semantics, and severe noise interference in medical images. The feature decomposition mechanism formula is shown below: ; in: High-frequency spatial features mainly include edge, texture, and fine-grained structural information; Low-frequency semantic features mainly include region contours and global context information. This partitioning corresponds to a channel-dimensional splitting operation, used to achieve local-global decoupled modeling.

[0043] For high-frequency branches, this invention employs a local window self-attention mechanism for modeling, as shown below: .

[0044] WindowAttention means dividing the feature map into fixed-size segments. The design employs local windows; multi-head self-attention is computed independently within each window; and local structure modeling capabilities are enhanced by limiting the receptive field. The technical benefits of this design include: enhanced edge sensitivity; improved responsiveness to small targets (such as tiny polyp regions); and reduced oversmoothing issues caused by global attention.

[0045] Attention within the window is calculated as follows: ; Where Q, K, and V represent the query, key, and value matrices, respectively, used to characterize the similarity retrieval and information aggregation relationships between features; the feature dimension is denoted as d, used to control the scaling factor in attention calculation to avoid the inner product result being too large and causing gradient instability; the Softmax(⋅) function is used to normalize the attention weights to satisfy the probability distribution constraints, thereby highlighting key response regions and suppressing irrelevant responses.

[0046] Since pure window attention suffers from window boundary fragmentation, this invention further introduces depthwise separable convolution for cross-window information fusion: ; in: This represents depthwise separable convolution; BN represents batch normalization; and residual structures are used to stabilize training and preserve original features. This step is used to: eliminate spatial breaks between windows; enhance the continuity of local structures; and improve spatial consistency representation.

[0047] For the low-frequency branch, this invention employs frequency domain complex number modeling. First, a linear projection is performed on the low-frequency features: ; in: It is a learnable linear projection matrix; used to map the input to a low-frequency semantic subspace.

[0048] The purpose of this design is to reduce the impact of high-frequency noise and make frequency domain modeling more stable.

[0049] To achieve the modeling of global semantic information and high-frequency noise suppression in medical images, the input features are first subjected to frequency domain transformation.

[0050] Specifically, the low-frequency semantic features after channel mapping are represented as Then, a two-dimensional Fast Fourier Transform (FFT) is performed on it to map the spatial domain features to the frequency domain representation: ; Through the above transformation, the structural information in the original spatial domain is mapped to the frequency domain representation, thus transforming it into a spectral distribution form with explicit physical meaning. In this spectral space, information at different spatial scales is naturally decoupled into different frequency components, enabling the model to more clearly model multi-scale feature structures. The low-frequency components mainly correspond to the overall contour, regional distribution, and semantic hierarchical information of the image, effectively characterizing global consistency and structural stability; while the high-frequency components mainly reflect local detail changes, including edge contours, texture details, and potential noise perturbation information. This frequency domain decoupling mechanism allows the network to perform feature modeling and selective enhancement in different frequency subspaces, thereby effectively suppressing redundant high-frequency noise interference while preserving key semantic information, and improving the robustness and discriminative ability of feature representation.

[0051] After obtaining the frequency domain representation, this invention further introduces a learnable filtering mechanism in the complex domain to adaptively modulate the spectral features. Specifically, learnable weight parameters are applied to the real and imaginary parts of the frequency domain features, respectively. and This allows for independent control of responses at different frequencies, as expressed below: ; in: Used to adjust the intensity of the real part of the frequency domain response, thereby controlling the degree of enhancement of structural semantic information; Used to adjust the intensity of the imaginary part of the frequency domain response, in order to control phase information and fine-grained changes; j represents the imaginary unit, used to construct the complex frequency domain representation; ⊙ represents element-wise multiplication.

[0052] Through the above complex filtering operations, the model can adaptively recalibrate different frequency components in the frequency domain, thereby enhancing the stable frequency components related to lesions in medical images and effectively suppressing high-frequency disturbances caused by acquisition noise or imaging interference, thus improving the signal-to-noise ratio of feature expression.

[0053] After frequency domain modulation, the processed spectral features are mapped back to the spatial domain using an inverse two-dimensional fast Fourier transform (IFFT) to recover spatial structure information. The expression is as follows: ; IFFT is used to reproject the frequency-domain enhanced signal back into the spatial domain, enabling the model to maintain the global semantic enhancement effect while restoring spatial structure consistency.

[0054] The final result This refers to the low-frequency semantic features after frequency domain modeling and enhancement. These features significantly reduce noise interference while preserving global contextual information, providing a stable semantic foundation for subsequent space-frequency domain interactive fusion.

[0055] To achieve information complementarity and synergistic enhancement between spatial and frequency domain features, a bidirectional gating interaction mechanism is constructed to establish dynamic information flow paths between different feature domains, thereby enabling adaptive fusion of local structural information and global semantic information.

[0056] Specifically, in obtaining spatial domain enhanced features Frequency domain enhancement features Next, the frequency domain representation is guided by spatial domain features. This process generates gated weights by performing channel mapping and nonlinear transformation on the spatial features, and then uses these weights to adaptively modulate the frequency domain features. The expression is as follows: ; Subsequently, this gating weight is applied to the frequency domain feature FL to constrain and filter the semantic spectrum using spatial structure information, thereby suppressing frequency responses irrelevant to the target region. Its expression is as follows: ; Through the above process, frequency domain features, while maintaining their global semantic expressive power, are further calibrated and reweighted by spatial structural information, thereby achieving collaborative alignment between spatial and frequency domain information. Specifically, the structural priors inherent in the spatial domain can constrain the frequency domain response, enabling the model to adaptively suppress redundant frequency components irrelevant to the task in the spectral space and enhance discriminative frequency responses related to the target region (such as the lesion region). The spatially guided frequency domain recalibration mechanism transforms the frequency domain representation from a simple global transformation result into a task-related representation fused with structural awareness, thus more effectively focusing on key frequency components corresponding to the lesion region and improving the model's sensitivity and localization accuracy for fine-grained lesion regions.

[0057] Meanwhile, to achieve reverse guidance of local spatial features by frequency domain information, this invention further utilizes frequency domain features to generate spatial modulation weights. Specifically, by performing linear mapping and nonlinear activation on the frequency domain features, gate coefficients for spatial domain modulation are obtained: ; The gating weight then acts on the spatial domain features. This is used to enhance its ability to perceive the global semantic context, thereby improving the semantic consistency of local features. Its expression is as follows: ; After bidirectional modulation is completed, the spatial domain enhancement features and the frequency domain enhancement features are spliced ​​and fused in the same feature space to form a unified dual-domain representation: .

[0058] The polyp medical image dataset partitioning module is specifically used for: A polyp medical image dataset containing polyp regions and their annotation information is obtained. The polyp medical images in the polyp medical image dataset are preprocessed by size normalization, pixel value standardization and data augmentation. The preprocessed polyp medical image dataset is divided into training set, validation set and test set according to a preset ratio for subsequent model training, validation and performance testing.

[0059] In the frequency-space dual-domain feature extraction module, the frequency-space dual-domain perceptual Transformer branch is specifically used to divide the input polyp medical image into channels to obtain high-frequency features and low-frequency features; to model the high-frequency features locally using a spatial domain local window attention mechanism to extract high-frequency detail features; to perform frequency domain transformation and global feature modeling on the low-frequency features using a frequency domain complex learnable filter to extract low-frequency global semantic features; and to fuse the high-frequency detail features and low-frequency global semantic features to obtain frequency-space dual-domain features.

[0060] In specific implementation, the frequency-space dual-domain perception Transformer branch includes multiple frequency-space dual-domain perception units. Each frequency-space dual-domain perception unit first divides the input features along the channel dimension into high-frequency feature channels and low-frequency feature channels. The high-frequency feature channels use a spatial domain local window attention mechanism to model local features, thereby enhancing lesion edge and texture detail information. The low-frequency feature channels use a frequency domain complex learnable filter to perform frequency domain transformation and global feature modeling, thereby extracting long-distance dependencies and global semantic information. Finally, the high-frequency detail features and low-frequency global semantic features are fused to obtain frequency-space dual-domain features.

[0061] In the medical image segmentation task of polyps, accurately extracting lesion edge details and global semantic information is crucial for improving segmentation accuracy. However, traditional convolutional neural networks mainly rely on local receptive fields for feature extraction. While they can effectively acquire texture and edge information, they struggle to establish long-distance dependencies. Although the Transformer structure possesses strong global modeling capabilities, it is insufficient in depicting local details and boundary structures. Furthermore, polyp regions typically exhibit significant morphological variations, blurred boundaries, and severe interference from background tissue. A single feature extraction method cannot simultaneously capture both local detail information and global semantic information, thus limiting further improvements in segmentation performance. To address these issues, this invention constructs a frequency-space dual-domain perceptual Transformer branch. By dividing the input features into high-frequency and low-frequency feature channels, it achieves collaborative modeling of spatial and frequency domain information. Specifically, the high-frequency feature channel utilizes a spatial domain local window attention mechanism to extract lesion edge and texture details, while the low-frequency feature channel uses a frequency domain complex learnable filter for global feature modeling to obtain global semantic information of the target region. The two types of features are then fused to obtain frequency-space dual-domain features. Furthermore, to enhance the information synergy among different features, a cross-gating mechanism is used to achieve bidirectional information interaction between high-frequency detail features and low-frequency global semantic features, thereby obtaining enhanced dual-domain features. In addition, adaptive weighted fusion is performed on the spatial features output by the convolutional neural network branches and the enhanced dual-domain features. This suppresses noise while strengthening the features of the target region, improving the model's ability to represent complex backgrounds and polyps of different scales, thus enhancing the accuracy and robustness of polyp medical image segmentation.

[0062] The dynamic feature interaction module is specifically used for: Based on the high-frequency detail features and low-frequency global semantic features in the frequency-space dual-domain features, a first gating weight and a second gating weight are generated respectively. Through a cross-gating mechanism, the low-frequency global semantic features are weighted based on the first gating weight, and the high-frequency detail features are weighted based on the second gating weight, so as to perform information interaction and collaborative enhancement, thereby obtaining enhanced dual-domain features.

[0063] In the segmentation network model training module, the frequency-space gating fusion module is used to generate adaptive gating weights based on the enhanced dual-domain features, and to use the adaptive gating weights to weight the spatial features output by the convolutional neural network branches in order to suppress noise and enhance effective features. The weighted spatial features are fused with the enhanced dual-domain features, and the fused features are obtained through feature smoothing.

[0064] The frequency-space gating fusion module is set between feature layers of different scales. It achieves multi-scale feature aggregation by fusing deep features, mid-level features and shallow features step by step.

[0065] In the segmentation network model training module, the segmentation network model adopts a network structure of encoder and decoder; The encoder consists of a convolutional neural network branch and a frequency-space dual-domain sensing Transformer branch; The encoder and decoder transmit features via skip connections; The decoder restores the spatial resolution of the fused features through step-by-step upsampling and outputs the target region segmentation result.

[0066] The decoder includes multiple decoding layers. Each decoding layer gradually recovers the spatial resolution of the fused features through convolution and upsampling operations, and performs feature reconstruction by combining the encoded features at the corresponding scale to obtain the target region segmentation result.

[0067] In practice, the encoder adopts a layer-stacked Transformer structure, which includes multiple layers of attention computing units and feedforward network units. Each layer includes: a spatial frequency domain collaborative attention sublayer for extracting joint local and global features; a feedforward network sublayer for nonlinear feature mapping and dimensionality transformation; and residual connections and normalization layers for stabilizing the training process and enhancing gradient propagation capabilities.

[0068] The encoder generates multi-scale semantic representations through layer-by-layer feature extraction and fusion, and passes them to the decoder via skip connections; the decoder is used to restore the spatial resolution step by step and output pixel-level target region segmentation results.

[0069] In the segmentation network model training module, the process of training, validating, and testing the segmentation network model specifically includes: The segmentation network model is trained end-to-end using the training set, its performance is validated and parameters are adjusted using the validation set, and its performance is tested using the test set.

[0070] To verify the effectiveness of this invention, experiments were conducted on the publicly available polyp datasets Kvasir-SEG and CVC-ClinicDB, and comparisons were made with existing typical segmentation methods. To ensure experimental fairness, all models employed the same data augmentation strategy, training parameters, and evaluation metrics. Evaluation metrics included Dice coefficient, Intersection over Union (IoU), Recall, and Precision.

[0071] First, the performance of this invention was compared with several mainstream medical image segmentation models on the Kvasir-SEG polyp segmentation dataset. The experimental results are shown in Table 1. The results show that our invention achieved the best performance across all evaluation metrics, validating the effectiveness of the proposed frequency-space dual-domain perception and dynamic feature interaction mechanism. Specifically, our invention achieved a Dice coefficient of 91.84%, an IoU of 86.16%, a Recall of 94.05%, and a Precision of 91.41% on the Kvasir-SEG dataset. Compared to the best-performing comparison model, MISSFormer, our invention improved Dice from 89.76% to 91.84%, an increase of 2.08 percentage points; IoU from 84.21% to 86.16%, an increase of 1.95 percentage points; Recall from 92.48% to 94.05%, an increase of 1.57 percentage points; and Precision from 90.35% to 91.41%, an increase of 1.06 percentage points. The results show that this invention can more accurately locate polyp regions and obtain more complete target contours. Compared with TransUNet, a representative Transformer model, the advantages of this invention are also significant. Specifically, Dice is improved from 89.21% to 91.84%, an increase of 2.63 percentage points; IoU from 83.73% to 86.16%, an increase of 2.43 percentage points; and Recall from 90.85% to 94.05%, an increase of 3.20 percentage points. This indicates that the proposed frequency-space dual-domain collaborative feature extraction mechanism can effectively compensate for the shortcomings of traditional Transformers in local boundary modeling, thereby improving the model's ability to perceive fine-grained polyp structures and edge regions. Compared with Swin-Unet, this invention improves Dice, IoU, Recall, and Precision by 3.28, 5.22, 4.18, and 3.39 percentage points, respectively. Although Swin-Unet can obtain certain global modeling capabilities using a window self-attention mechanism, its ability to express complex boundary regions and small polyp structures remains limited. This invention enables the network to simultaneously focus on global semantic information and local edge details by combining frequency domain feature enhancement and spatial domain local window attention modeling, thereby obtaining more accurate segmentation results.

[0072] This invention also demonstrates significant advantages over traditional convolutional network models. Compared to U-Net, Dice is improved by 8.15 percentage points, and IoU by 10.17 percentage points; compared to UNet++, Dice is improved by 8.63 percentage points; and compared to AttentionU-Net, Dice is improved by 7.76 percentage points. This indicates that while traditional convolutional structures can effectively extract local texture features, they lack the ability to model long-distance dependencies. This invention, by introducing a frequency-space dual-domain perceptual Transformer branch and a dynamic feature interaction network, achieves deep fusion of local detail information and global contextual information, thereby significantly enhancing feature representation capabilities. Furthermore, this invention achieves a Recall metric of 94.05%, the highest among all models, indicating that it can more completely identify polyp regions and effectively reduce false negatives. Simultaneously, this invention achieves a Precision metric of 91.41%, indicating that the model can effectively suppress missegmentation of background regions and reduce the false positive rate. The simultaneous improvement in Recall and Precision demonstrates that this invention can not only accurately discover target regions but also maintain high boundary localization accuracy, exhibiting good segmentation robustness and generalization ability. In summary, this invention achieves state-of-the-art segmentation performance on the Kvasir-SEG dataset, fully validating the effectiveness of the proposed frequency-domain complex learnable filter module, spatial-domain local window attention module, feature interaction module, and frequency-space gating fusion module. Through collaborative modeling of frequency and spatial domain information and a cross-branch feature interaction mechanism, this invention can more fully mine the discriminative features of polyp regions, thereby obtaining more accurate and robust medical image segmentation results.

[0073] Table 1: Performance of different methods on the Kvasir-SEG dataset

[0074] To further verify the generalization ability of this invention in different polyp segmentation scenarios, a comparative experiment was conducted on the CVC-ClinicDB dataset with mainstream medical image segmentation models. The experimental results are shown in Table 2. The CVC-ClinicDB dataset contains a large number of complex intestinal polyp images, characterized by large variations in polyp size, complex boundary morphology, and diverse lighting conditions. It is an important publicly available dataset for evaluating the performance of polyp segmentation models. As can be seen from Table 2, our invention achieved the best results in all evaluation metrics. Specifically, the Dice coefficient reached 93.45%, IoU reached 88.60%, Recall reached 93.20%, and Precision reached 94.97%, with overall performance superior to all compared methods, fully demonstrating the effectiveness of the proposed frequency-space dual-domain perception and dynamic feature interaction mechanism in polyp segmentation tasks.

[0075] Specifically, compared with classic convolutional networks, this invention has significant advantages. Compared with U-Net, the Dice of this invention is improved from 86.20% to 93.45%, an improvement of 7.25 percentage points; the IoU is improved from 79.45% to 88.60%, an improvement of 9.15 percentage points. Compared with Unet++, Dice is improved by 6.73 percentage points; compared with AttentionU-Net, Dice is improved by 4.50 percentage points. The results show that although traditional CNN structures can effectively extract local texture information, they still have shortcomings in modeling complex polyp boundaries and long-range dependencies. This invention, through collaborative modeling in the frequency and spatial domains, can obtain richer and more robust feature representations. Compared with the Transformer series models, this invention also shows strong competitiveness. TransUNet achieved a Dice of 92.92% and an IoU of 87.15% on this dataset, demonstrating excellent segmentation performance. Swin-UNet achieved a Dice of 92.31% and an IoU of 85.96%, while MISSFormer achieved a Dice of 90.50% and an IoU of 84.21%. This invention further improves the Dice to 93.45% and the IoU to 88.60%, representing improvements of 0.53 and 1.45 percentage points respectively compared to TransUNet. Although the improvement may seem limited, within the high-performance segmentation range where Dice exceeds 90%, every 0.5 percentage point improvement signifies a further enhancement in the model's ability to locate polyp boundaries and improve region integrity, which has significant practical implications. In terms of Recall, this invention achieves 93.20%, essentially on par with Swin-UNet's 93.58%, indicating that the model can detect polyp regions relatively completely and effectively reduce missed segmentation. Meanwhile, the precision of this invention reaches 94.97%, significantly higher than TransUNet (91.82%), Swin-Unet (91.84%), and MISSFormer (90.35%), indicating that the model can effectively suppress false detections in background regions while maintaining high recall, thus improving the accuracy of prediction results. This demonstrates that the proposed frequency-domain complex learnable filter can effectively suppress noise interference, while the dynamic feature interaction module and the frequency-space gated fusion module further enhance the model's ability to discriminate real polyp regions. Combined with the visualized segmentation results, it can be found that in scenarios with blurred boundaries, small-sized polyps, and complex backgrounds, this invention can more accurately recover the polyp contours and obtain more complete and smooth segmentation results. In contrast, some comparative methods suffer from boundary loss, local oversegmentation, or undersegmentation, while the segmentation mask generated by this invention is closer to the real annotations.In summary, this invention achieves the best segmentation performance on the CVC-ClinicDB dataset, verifying that the synergistic effect of the frequency-space dual-domain feature extraction module, the dynamic feature interaction module, and the frequency-space gating fusion module can effectively improve the model's ability to locate polyp regions and characterize their boundaries, further demonstrating that this invention has good robustness and generalization ability.

[0076] Table 2: Performance of different methods on the ClinicDB dataset

[0077] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A polyp image segmentation system based on frequency-space dual-domain perception and feature interaction, characterized in that, Includes the following modules: The polyp medical image dataset partitioning module is used to acquire a polyp medical image dataset, and to preprocess and partition the polyp medical image dataset to obtain a training set, a validation set and a test set; The frequency-space dual-domain feature extraction module is used to construct a parallel feature extraction network that includes a convolutional neural network branch and a frequency-space dual-domain perceptual Transformer branch; The convolutional neural network branch is used to perform multi-level feature extraction on the input polyp medical image to obtain spatial features containing local texture information and edge structure information. The frequency-space dual-domain perception Transformer branch is used to extract high-frequency detail features through a spatial domain local window attention mechanism and extract low-frequency global semantic features through a frequency domain complex learnable filter to obtain frequency-space dual-domain features. The dynamic feature interaction module is used to construct a dynamic feature interaction network to perform information interaction and collaborative enhancement on the high-frequency detail features and low-frequency global semantic features in the frequency-space dual-domain features through a cross-gating mechanism, thereby obtaining enhanced dual-domain features. The frequency-space gated fusion module is used to construct a frequency-space gated fusion network to generate adaptive gate weights based on the enhanced dual-domain features, use the adaptive gate weights to weight the spatial features output by the convolutional neural network branches, fuse the weighted spatial features with the enhanced dual-domain features, and obtain the fused features through feature smoothing processing. The segmentation network model training module is used to train, validate, and test the segmentation network model created based on the parallel feature extraction network, dynamic feature interaction network, frequency-space gated fusion network, and decoder using the training set, validation set, and test set, respectively, and to deploy the segmentation network model that passes the test. The polyp medical image segmentation module is used to acquire the medical image of the polyp to be segmented, input it into the deployed segmentation network model, and output the corresponding target region segmentation result.

2. The polyp image segmentation system with frequency-space dual-domain perception and feature interaction as described in claim 1, characterized in that, The polyp medical image dataset partitioning module is specifically used for: A polyp medical image dataset containing polyp regions and their annotation information is obtained. The polyp medical images in the polyp medical image dataset are preprocessed by size normalization, pixel value standardization and data augmentation. The preprocessed polyp medical image dataset is divided into training set, validation set and test set according to a preset ratio for subsequent model training, validation and performance testing.

3. The polyp image segmentation system with frequency-space dual-domain perception and feature interaction as described in claim 1, characterized in that, In the frequency-space dual-domain feature extraction module, the frequency-space dual-domain perception Transformer branch is specifically used to divide the input polyp medical image into channels to obtain high-frequency features and low-frequency features. The high-frequency features are modeled locally using a spatial domain local window attention mechanism to extract high-frequency detail features; the low-frequency features are transformed in the frequency domain and modeled globally using a frequency domain complex learnable filter to extract low-frequency global semantic features. The high-frequency detail features and low-frequency global semantic features are fused to obtain frequency-space dual-domain features.

4. The polyp image segmentation system with frequency-space dual-domain perception and feature interaction as described in claim 1, characterized in that, The dynamic feature interaction module is specifically used for: Based on the high-frequency detail features and low-frequency global semantic features in the frequency-space dual-domain features, a first gating weight and a second gating weight are generated respectively. Through a cross-gating mechanism, the low-frequency global semantic features are weighted based on the first gating weight, and the high-frequency detail features are weighted based on the second gating weight, so as to perform information interaction and collaborative enhancement, thereby obtaining enhanced dual-domain features.

5. The polyp image segmentation system with frequency-space dual-domain perception and feature interaction as described in claim 1, characterized in that, In the segmentation network model training module, the frequency-space gating fusion module is used to generate adaptive gating weights based on the enhanced dual-domain features, and to use the adaptive gating weights to weight the spatial features output by the convolutional neural network branches in order to suppress noise and enhance effective features. The weighted spatial features are fused with the enhanced dual-domain features, and the fused features are obtained through feature smoothing.

6. The polyp image segmentation system with frequency-space dual-domain perception and feature interaction as described in claim 5, characterized in that, The frequency-space gating fusion module is set between feature layers of different scales. It achieves multi-scale feature aggregation by fusing deep features, mid-level features and shallow features step by step.

7. The polyp image segmentation system with frequency-space dual-domain perception and feature interaction as described in claim 1, characterized in that, In the segmentation network model training module, the segmentation network model adopts a network structure of encoder and decoder; The encoder consists of a convolutional neural network branch and a frequency-space dual-domain sensing Transformer branch; The encoder and decoder transmit features via skip connections; The decoder restores the spatial resolution of the fused features through step-by-step upsampling and outputs the target region segmentation result.

8. The polyp image segmentation system with frequency-space dual-domain perception and feature interaction as described in claim 7, characterized in that, The decoder includes multiple decoding layers. Each decoding layer gradually recovers the spatial resolution of the fused features through convolution and upsampling operations, and performs feature reconstruction by combining the encoded features at the corresponding scale to obtain the target region segmentation result.

9. The polyp image segmentation system with frequency-space dual-domain perception and feature interaction as described in claim 1, characterized in that, In the segmentation network model training module, the process of training, validating, and testing the segmentation network model specifically includes: The segmentation network model is trained using the training set, its performance is validated and parameters are adjusted using the validation set, and its performance is tested using the test set.

Images