A real-time lung CT segmentation method and device based on a spatial frequency guide network

Abstract

The application discloses a real-time lung CT segmentation method and device based on a spatial frequency guide network, and relates to the technical field of image segmentation.The spatial frequency guide network comprises an encoder part, a decoder part and a segmentation head; the encoder part comprises stem blocks and a plurality of encoders connected in sequence, and sequentially extracts features of a lung CT image and outputs a plurality of features of different scales; the decoder part comprises a multi-scale feature interaction module and a plurality of decoders connected in sequence, integrates the plurality of features of different scales, and obtains final features; and the segmentation head generates a segmentation result based on the final features.The application realizes stable hierarchical coding by integrating a residual down-sampling block and a spatial frequency guide block, and realizes efficient feature fusion by combining a multi-scale feature interaction module; the spatial frequency representation is integrated into an encoder-decoder process, which significantly improves the ability to outline fine anatomical structures and the ability to process appearance variations driven by pathology, and does not affect efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image segmentation, in particular to a real-time lung CT segmentation method and device based on a spatial frequency guided network. BACKGROUND

[0002] Precise segmentation of lung nodules in CT images is crucial for lung cancer diagnosis, and the segmentation quality directly affects the subsequent analysis and diagnosis effect. However, lung nodule segmentation faces major challenges: the internal structure of the lung and the morphological characteristics of the nodule are often highly similar, especially when the nodule is located on the lung wall or attached to the end of the blood vessel. The solutions proposed in recent years can be roughly divided into traditional detection methods, machine learning algorithms, and convolutional neural network (CNN) based methods. In traditional detection methods, morphological operations (Pandey and Bhandari, 2024), threshold-based methods (Zheng et al., 2024; Mehta et al., 2024), clustering algorithms (Shi et al., 2021), and energy optimization algorithms are widely used to detect nodules in highly complex lung environments (Yan and Razmjooy, 2023). For example, Gupta et al. (2015) constructed a lung segmentation template by combining masks, flood fill algorithms, and morphological operations, and then used a multi-level threshold algorithm combined with multi-feature extraction techniques for tuberculosis detection. Rezaie et al. (2017) first screened the region of interest that may contain nodules by threshold method, and then applied edge detection algorithm to locate the nodules. However, due to the significant differences in position, shape, size and type of lung nodules, from adherent nodules adjacent to the pleura or blood vessels to ground glass opacity nodules, threshold and morphological based segmentation methods are less effective. With the continuous growth of lung CT images, developing a robust automatic detection model is particularly critical to reduce tedious manual annotation and detection result variation. SUMMARY

[0003] The present application aims to solve the problems in the prior art.

[0004] The technical solution adopted by the present application to solve its technical problems is to provide a real-time lung CT image segmentation method based on a spatial frequency guided network, comprising the following steps:

[0005] Constructing a spatial frequency guided network based on an encoder-decoder structure and training it;

[0006] Using the trained spatial frequency guided network for real-time lung CT image segmentation;

[0007] The spatial frequency guided network includes an encoder part, a decoder part and a segmentation head.

[0008] The encoder part comprises a stem block and a plurality of encoders connected in sequence, for sequentially extracting features of the lung CT image and outputting a plurality of features of different scales; the stem block adopts a stacked 3*3 convolution, for extracting preliminary features of the lung CT image; the decoder part comprises a plurality of decoders connected in sequence and a multi-scale feature interaction module; the multi-scale feature interaction module receives the output of the last-stage encoder and guides fine-grained spatial feature optimization through a coarse-grained semantic descriptor, to realize structured cross-scale interaction; the last-stage decoder receives the output of the multi-scale feature interaction module and the output of the last-stage encoder, and outputs a decoding feature; the other decoders receive the decoding feature output by the next-stage decoder and the encoding feature output by the same-stage encoder, and output a decoding feature; the decoding feature output by the first-stage decoder is taken as a final feature; and the segmentation head generates a segmentation result based on the final feature.

[0009] Preferably, each stage encoder comprises a spatial frequency guide module SFGM and a stage link fusion unit SLFU; the spatial frequency guide module SFGM comprises a spatial frequency guide block SFGB and a residual down-sampling block RDBlock, the residual down-sampling block RDBlock learns low-resolution features with refined information from the input features as a down-sampling tool; the spatial frequency guide block SFGB captures dense semantic context from the RDBlock output as a basic bottleneck structure; the stage link fusion unit SLFU fuses the features output by the spatial frequency guide module SFGM of the current stage and the features of the previous stage, to obtain the encoding features of the current-stage encoder and output the same; the features of the previous stage of the first-stage encoder are the preliminary features output by the stem block, and the features of the previous stage of the other encoders are the encoder features output by the encoder of the previous stage.

[0010] Preferably, the residual down-sampling block RDBlock has an internal processing procedure comprising the following steps:

[0011] initial down-sampling to obtain a down-sampled feature , denoted as:

[0012] ;

[0013] wherein, is a 3*3 convolution with a step s=2, denotes a convolution kernel of the convolution ; denotes the input feature of the residual down-sampling block RDBlock;

[0014] a 3*3 deep convolution is adopted to refine the down-sampled feature , and batch normalization is adopted to improve training stability, to obtain a feature , denoted as:

[0015] ;

[0016] wherein, denotes a convolution kernel of the convolution;

[0017] Feature concatenation along the channel dimension is denoted as:

[0018] ;

[0019] wherein, Concat denotes a concatenation operation;

[0020] The concatenated features are subjected to average pooling Āand channel redundancy reduction by 1x1 convolution to obtain features , which is denoted as:

[0021] ;

[0022] wherein, denotes a 1x1 convolution, denotes a convolution kernel of the operation ;

[0023] Subsequently, the residual connection is added to optimize the training process, which is denoted as:

[0024] ;

[0025] wherein, denotes the final output of the residual down-sampling block RDBlock.

[0026] Preferably, the spatial frequency guide block SFGB has an internal processing procedure including the following steps:

[0027] Local structure and multi-scale spatial context are extracted to obtain extracted features , which is denoted as:

[0028] ;

[0029] wherein, denotes the input feature map of the spatial frequency guide block SFGB; is a convolution for capturing fine-grained details, denotes a convolution kernel of the convolution, is a dilated depthwise convolution with dilation rate d, which can expand the receptive field without increasing the computational cost, denotes a convolution kernel of the convolution; denotes batch normalization;

[0030] The spatial branch of the spatial-frequency attention module captures broader local context with large kernel convolution, resulting in spatial tokens, denoted as:

[0031] ;

[0032] where S denotes spatial tokens; p is a ReLU activation, and Conv denotes a convolution operation.

[0033] The frequency branch of the spatial-frequency attention module processes the input X to obtain frequency tokens, denoted as: Each channel computes a two-dimensional fast Fourier transform, and computes a channel-dimension average amplitude, and finally projects to obtain frequency tokens, denoted as:

[0034] ;

[0035] ;

[0036] ;

[0037] where, denotes the result of computing a two-dimensional fast Fourier transform, denotes a two-dimensional fast Fourier transform, denotes the result of taking an absolute value of ; denotes the value of the c-th channel, denotes the total number of channels, denotes the result of computing a channel-dimension average amplitude of ; denotes that the frequency token a is a 1x1 convolution, denotes the convolution kernel of the convolution a;

[0038] Fusion and attention weighting are performed on the spatial tokens S and the frequency tokens to generate spatial-frequency attention :

[0039] ;

[0040] where || is a channel-level concatenation, is a sigmoid function; denotes a convolution, denotes the convolution kernel of the convolution ;

[0041] The final output is obtained by the following formula: :

[0042] ;

[0043] ;

[0044] wherein, ⊙ denotes element-wise multiplication, is a 1x1 convolution, is a convolution kernel of the convolution .

[0045] Preferably, the stage linking fusion unit SLFU has an internal processing procedure including the following steps:

[0046] receiving the current stage output and the previous stage fused feature , which are processed by two parallel branches respectively, denoted as:

[0047] ;

[0048] ;

[0049] wherein, and denote the output results of the two parallel branches respectively; and both denote separable convolution, and denote the convolution kernel of the convolution and the convolution respectively;

[0050] the processing results are optimized and fused to obtain the final output of the SLFU, denoted as:

[0051] ));

[0052] wherein, denotes batch normalization, denotes 1x1 convolution, denotes the convolution kernel of the convolution .

[0053] Preferably, the multi-scale feature interaction module has an internal processing procedure including the following steps:

[0054] performing successive double 3x3 convolution operations on the features output by the last SLFU to generate enhanced deep feature maps ;

[0055] performing 1x1 convolution on the encoded features output by the last encoder to generate transformed features , which are refined by a lightweight channel attention module to obtain coarse-grained features ; finally, the coarse-grained features are fused with the enhanced deep feature maps , denoted as:

[0056] ;

[0057] DK represents the output of the multi-scale feature interaction module.

[0058] Preferably, the internal processing flow of the decoder is represented as:

[0059] ;

[0060] wherein, is a 3x3 convolution optimization operator, represents channel concatenation, represents up-sampling; represents the decoder feature output by the k-th decoder; when k=K, the output of the multi-scale feature interaction module is adopted; when k≠K, represents the decoder feature output by the k+1-th decoder; represents the encoder feature output by the k-th encoder.

[0061] The application also provides a real-time lung CT image segmentation device based on a spatial frequency guided network, comprising:

[0062] a network training module for constructing and training a spatial frequency guided network based on an encoder-decoder structure;

[0063] a real-time segmentation module for performing real-time lung CT image segmentation using the trained spatial frequency guided network;

[0064] The spatial frequency guided network comprises an encoder part, a decoder part and a segmentation head; the encoder part comprises a stem block and a plurality of encoders connected in sequence, for sequentially extracting features from the lung CT image and outputting a plurality of features of different scales; the stem block adopts stacked 3x3 convolution to extract preliminary features of the lung CT image; the decoder part comprises a multi-scale feature interaction module and a plurality of decoders connected in sequence; the multi-scale feature interaction module receives the output of the last-stage encoder and guides fine-grained spatial feature optimization through coarse-grained semantic descriptors, realizing structured cross-scale interaction; the last-stage decoder receives the output of the multi-scale feature interaction module and the output of the last-stage encoder, and outputs decoding features; the other decoders receive the decoding features output by the next-stage decoder and the encoding features output by the same-stage encoder, and output decoding features; the decoding features output by the first-stage decoder are used as final features; and the segmentation head generates a segmentation result based on the final features.

[0065] The application has the following beneficial effects:

[0066] (1) The application realizes stable level coding by integrating the residual down-sampling block and the spatial frequency guide block, and completes efficient feature fusion in combination with a multi-scale feature interaction module, and constructs a standardized architecture of unified fine-grained detail preservation and global context modeling;

[0067] (2) The application integrates the space-frequency representation into the encoder-decoder process, significantly improves the fine anatomical structure sketching ability and the appearance variation processing ability driven by pathology, and does not affect the efficiency, so that it is suitable for various real-time medical segmentation scenes.

[0068] The application will be further described in detail in combination with the drawings and embodiments, but the application is not limited to the embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0069] Figure 1 A method step diagram of a real-time lung CT image segmentation method based on a spatial frequency guide network according to an embodiment of the application;

[0070] Figure 2 A structure schematic diagram of SFGNet of a real-time lung CT image segmentation method based on a spatial frequency guide network according to an embodiment of the application;

[0071] Figure 3 A main module schematic diagram of SFGNet of a real-time lung CT image segmentation method based on a spatial frequency guide network according to an embodiment of the application;

[0072] Figure 4 A structure schematic diagram of a spatial frequency guide block of a real-time lung CT image segmentation method based on a spatial frequency guide network according to an embodiment of the application;

[0073] Figure 5 A structure schematic diagram of a decoder of a real-time lung CT image segmentation method based on a spatial frequency guide network according to an embodiment of the application;

[0074] Figure 6 Resource utilization rates of different SFGNets of a real-time lung CT image segmentation method based on a spatial frequency guide network according to an embodiment of the application;

[0075] Figure 7 Segmentation performance comparison of different SFGNets of a real-time lung CT image segmentation method based on a spatial frequency guide network according to an embodiment of the application on a lung CT data set;

[0076] Figure 8 Segmentation performance comparison of a real-time lung CT image segmentation method based on a spatial frequency guide network according to an embodiment of the application and other segmentation algorithms;

[0077] Figure 9The method steps of the real-time lung CT image segmentation method based on the spatial frequency guide network are shown in the method step diagram of the embodiment of the present application, and the method steps include:

[0078] Figure 10 The segmentation performance comparison of different SFGNets of the real-time lung CT image segmentation method based on the spatial frequency guide network on the COVID19-Seg data set is shown in the embodiment of the present application.

[0079] Figure 11 The structure diagram of the real-time lung CT image segmentation device based on the spatial frequency guide network is shown in the embodiment of the present application. DETAILED DESCRIPTION

[0080] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0081] Referring to Figure 1 The method steps of the real-time lung CT image segmentation method based on the spatial frequency guide network are shown in the method step diagram of the embodiment of the present application, and the method steps include:

[0082] S101, constructing a spatial frequency guide network based on an encoder-decoder structure and training the spatial frequency guide network;

[0083] S102, using the trained spatial frequency guide network to perform real-time lung CT image segmentation;

[0084] The structure of the spatial frequency guide network SFGNet is shown in Figure 2 The encoder has a unique design, including a stem block composed of stacked 3x3 convolutions, as shown in Figure 2(a) in FIG. 1 for extracting preliminary features; the stem block as an initial feature extractor converts the original input CT image into a compact and informative feature map. By reducing the spatial resolution early, the network gains computational efficiency while preserving the essential structural cues required for subsequent processing. Subsequently, three consecutive spatial frequency guidance modules (SFGMs) are connected to extract contextual information at 1 / 4, 1 / 8, and 1 / 16 of the original spatial size, respectively. Each SFGM contains N SFGBs followed by a RDBlock. The proposed stage linking fusion unit is used to refine and bridge the features between different stages. The RDBlock serves as a downsampling tool to learn low-resolution features with refined information, while the SFGB, as a basic bottleneck structure, captures dense semantic context from the RDBlock output. The SLFU operates to fuse adjacent stage features, facilitating the reuse of current and previous spatial scale information and enhancing feature representation. Additionally, the dilation rate of the SFGB in the first SFGM is smaller than that in the second and third modules: the first module has a dilation rate of 2, and the subsequent modules use dilation rates of 4 and 8, respectively. Experimental verification shows that this setting effectively improves the performance of real-time lung tumor segmentation. Then a decoder based on a multi-scale feature interaction module is proposed, which integrates different spatial dimension information through a feature fusion block, such as Figure 2 (c) in FIG. 1, which enhances the performance of pixel-level prediction. Finally, a simple segmentation head is used to generate the final segmentation result.

[0085] Table 1 - SFGNet architecture details (input resolution is 256x256, "N" represents the number of SFGB blocks at each stage; "C" represents the number of prediction categories):

[0086]

[0087] Specifically, the spatial frequency guidance network uses a reparameterization technique to optimize the convolutional neural network architecture, achieving efficient inference while maintaining training performance. Inspired by RepVGG, diverse branch blocks, and ACNet, among other models for compression and deployment efficiency, this method eliminates redundant calculations in the inference process by integrating the batch normalization layer into the pre-convolution layer. The following details the reparameterization strategy for the core building blocks of SFGNet, namely the convolution-batch normalization fusion.

[0088] Given a convolutional layer followed by a batch normalization and activation function sequence, it is represented as:

[0089] ;

[0090] where, is the convolution kernel, and b is the bias;

[0091] BN is normalized by the following formula:

[0092] ;

[0093] where μ and σ2are the sliding mean and variance, and γ, β are learnable affine parameters.

[0094] At inference time, this combination is re-parameterized into a single convolution by pre-computing the equivalent kernel W' and bias b':

[0095] ;

[0096] This results in a functionally equivalent but structurally simplified model, denoted as:

[0097] ;

[0098] This transformation is applied to all convolutional blocks in SFGNet, including the attention modules and the convolutional layers inside the residual structure.

[0099] In particular, encoder-based segmentation architectures require down-sampling to expand the effective receptive field; however, standard operations such as strided convolution and pooling often discard fine structural cues that are critical for medical image segmentation. This limitation is particularly prominent in lung CTs, where lesions such as ground-glass opacities and nodules often exhibit weak boundaries and heterogeneous textures. While existing techniques have explored multi-branch down-sampling and residual feature preservation to mitigate information loss, these methods still mainly focus on the spatial domain and struggle to maintain boundary-aware details when reducing resolution. To address these challenges, we propose a residual down-sampling block, denoted as (b) in FIG. 1, which preserves high-fidelity semantic content while reducing resolution. Inspired by residual learning and efficient deep decomposition, the RDBlock introduces a hybrid down-sampling pathway to maintain structural integrity at coarse scales. Given an input feature tensor Figure 2 , the RDBlock computes as follows:

[0100] First, an initial down-sampling is performed, denoted as:

[0101] ;

[0102] where Φ is a 3x3 convolution with a stride s = 2.

[0103] Subsequently, a 3x3 depthwise convolution Θ is employed to refine the down-sampled features, and batch normalization is applied to enhance training stability: ;

[0104] Feature concatenation is performed along the channel dimension; to enhance the encoder's ability to model global context, an average pooling Ā is applied to the concatenated features, followed by a 1x1 convolution to reduce channel redundancy; the result is then added to the residual connection to optimize the training process:

[0105] ; ​

[0106] ;

[0107] ;

[0108] This residual connection helps to preserve fine-grained details and alleviate the feature degradation problem caused by deep network down-sampling. In the SFGNet framework, RDBlock serves as the core component of the encoder path, achieving efficient multi-scale feature extraction while supporting real-time performance. Its ability to preserve semantic details through the residual path directly improves segmentation accuracy, especially in challenging cases involving small or irregularly shaped lung tumors.

[0109] Specifically, traditional convolution blocks and attention mechanisms often struggle to model both high-frequency boundary details and low-frequency structural context, both of which are crucial for robust lung CT image segmentation. Typical attention modules mainly focus on channel reweighting or local saliency cues, but lack explicit modeling of the global spectral distribution of encoded lesion texture, sharp edges, and fine boundary transitions. Similarly, Transformer-based self-attention mechanisms can capture long-range context, but due to tokenization and block-wise self-aggregation properties, they tend to oversmooth fine structures. To address these limitations, we propose the spatial frequency guided block, which unifies spatial features with global frequency descriptors. Its design inspiration comes from the observation that pathological structures in medical images often appear as high-frequency abnormalities embedded in a low-frequency anatomical background. By explicitly modeling dual-domain features, SFGB provides a stronger inductive bias than pure attention architectures and improves robustness to contrast fluctuations and morphological heterogeneity, which are common challenges in lung tumor and COVID-19 lesion segmentation. See Figure 4 The architecture of SFGB and its core component, the spatial frequency attention module, is shown in FIG. 1.

[0110] Given input feature maps , SFGB first extracts local structures and multi-scale spatial context:

[0111] ;

[0112] where Ψ is a convolution that captures fine-grained details, δ is a dilated depth convolution with dilation rate d, which can expand the receptive field without increasing computational cost. η represents batch normalization. The residual connection stabilizes the optimization process.

[0113] Dilated convolution has been shown to enhance the structural continuity of biomedical images, but it is not sufficient to encode sufficient long-range global information, so we introduce frequency domain feature integration and propose the spatial frequency attention module, which refines and enhances features Xc by jointly computing spatial and frequency domain attention.

[0114] The spatial branch captures broader local context through large kernel convolutions, denoted as:

[0115] ;

[0116] Large kernel convolutions have been shown to have significant effect in modeling geometric continuity and long-range anatomical consistency, especially in organ segmentation tasks.

[0117] The frequency branch captures global periodicity and boundary frequency characteristics by computing 2D fast Fourier transform for each channel to extract global spectral cues, denoted as:

[0118] ;

[0119] The channel-wise average magnitude is then computed, denoted as:

[0120] ;

[0121] This spectrum can capture global periodicity and boundary frequency characteristics, which is analogous to the frequency enhanced segmentation method and provides complementary global cues missing from spatial attention. The frequency descriptor is projected to channel:

[0122] ;

[0123] Fusion and attention weighting: The spatial and frequency tokens are concatenated and fused:

[0124] ;

[0125] The final output is obtained by:

[0126] ;

[0127] where denotes element-wise multiplication.

[0128] Output projection and residual refinement: The final projection layer aligns the channel dimension:

[0129] ;

[0130] Compared with the classic channel / space attention mechanism (such as SE, CBAM), the SFGB proposed in the embodiment of the present application explicitly fuses the frequency domain descriptors, which can capture the global structural regularity and high-frequency boundary transition characteristics specific to the lung. This characteristic of the lesion is particularly important in CT imaging, because the pathological region often appears as a subtle frequency disturbance embedded in a smooth anatomical background. In addition, although the existing multi-scale context module based on the cavity convolution can expand the receptive field, it is still essentially limited to the spatial filtering category. The SFGB effectively connects local geometric features and overall frequency characteristics by coupling multi-scale spatial responses with global frequency spectrum statistics, thereby expanding these designs. This coupling mechanism alleviates the common failure mode of pure spatial pyramid methods, which perform poorly under common lesion texture variations, contrast drift, or noise interference in CT acquisition.

[0131] Specifically, to promote progressive feature optimization between encoder stages, a stage linking fusion unit is proposed. This unit establishes hierarchical connections between adjacent stages to preserve multi-scale context information while achieving feature fusion. As shown in (a) of FIG. 1, Figure 3 The SLFU serializes the current stage output and the previous stage fusion feature (the first stage is the stem block output) to generate optimized multi-scale representations and propagate them forward. This connection mechanism ensures effective integration of early local details and deep global semantics, alleviating the information loss problem in traditional down-sampling processes. Figure 3 The detailed structure of the SLFU shown in (a) of FIG. 1 includes two parallel branches, which process and respectively. Each branch uses depthwise separable convolution υ, κ for local pattern capture and channel dimension adjustment:

[0132] ;

[0133] ;

[0134] Finally, a 1x1 convolution with batch normalization η optimizes the fused features, and its operation formula is represented as:

[0135] .

[0136] Specifically, the multi-scale feature interaction module processes the encoder features output by the last stage encoder through a spatial frequency guiding unit to inject global frequency spectrum priors, and then outputs the processing results to the decoder. Referring to FIG. 2, Figure 3As shown in (b), a multi-scale feature interaction decoder with a multi-scale feature interaction module at its core is designed specifically for achieving structured cross-scale fusion, enabling coarse-grained semantic representation to guide high-resolution spatial feature optimization. This is particularly important in lung CT image segmentation: when low-resolution or high-resolution features are used independently, small nodules and irregular boundaries are easily missed. Let E1, E2, and E3 represent progressively coarsened encoder features. MFIM first performs continuous double 3×3 convolution operations through staged link fusion units to generate enhanced deep feature maps. The enhanced representation is then coupled with the coarse-grained encoder feature E3: transform features are generated through 1×1 convolution. Then, the coarse-grained feature M is obtained by refining through a lightweight channel attention module. E This allows semantically rich activations to modulate fine-grained spatial cues; this selective interaction avoids the indiscriminate fusion common in traditional skip connections, enabling the decoder to preserve key boundary information without overloading high-resolution paths. This process forms the first-layer decoded output:

[0137] ;

[0138] See Figure 5 As shown, the decoder integrates MFIM into a top-down refining process, and the decoding process is described as follows: ;

[0139] Here, φ is a 3×3 convolution optimization operator, and || denotes channel concatenation. This formula enables controlled propagation of semantic information from deep to shallow layers, generating representations that combine spatial accuracy and global consistency. Compared to UNet++'s nested dense aggregation, MFIM avoids parameter inflation by focusing on semantically effective paths. Unlike Transformer-based decoders, MFIM maintains strong long-range context modeling capabilities through a frequency-aware conditional mechanism without incurring secondary computational costs associated with self-attention. Furthermore, by primarily applying attention at coarse scales, MFIM alleviates the spatial over-suppression problem common in dense attention gating. This selective guidance is crucial for CT-based lesion analysis—lesion structures span multiple spatial frequencies, requiring simultaneous assurance of reconstruction accuracy and robustness. Experiments on the LungCT and COVID19-Seg datasets demonstrate that MFIM significantly improves the Dice coefficient and cross-union ratio, particularly in cases with irregular boundaries and diffuse small lesions. These findings confirm the importance of structured cross-scale interactions, establishing MFIM as a standardized and efficient decoding mechanism for frequency-guided architectures such as SFGNet.

[0140] From an architectural perspective, MFIM can better balance representation richness and computational feasibility than the nested aggregation of UNet++ or the multi-head self-attention mechanism of Transformer. By mainly focusing on coarse-scale attention, the decoder avoids the spatial over-suppression phenomenon in dense attention gating and the parameter growth problem based on Transformer fusion. This selective guidance is crucial for lung CT image segmentation, as the lesion morphology may span multiple spatial scales, but the computation must remain lightweight. Empirical evaluation in LungCT and COVID19-Seg benchmark tests shows that MFIM significantly improves model performance, which not only confirms the importance of structured cross-scale information interaction, but also verifies the effectiveness of MFIM as a standardized fusion mechanism for frequency-guided architectures such as SFGNet.

[0141] A real-time lung CT image segmentation method based on a spatial frequency-guided network was verified by an embodiment of the application. Extensive experiments were conducted on two computed tomography segmentation benchmarks to evaluate the effectiveness and generalization ability of the proposed SFGNet. Both datasets contain heterogeneous lung lesions and are suitable for benchmark testing of different methods.

[0142] For lung tumor and nodule segmentation tasks, the public CT slice dataset provided by the Kaggle platform was used. This dataset integrates multiple sources of scan images, with all 2534 images having an original plane resolution of 512x512 pixels. To ensure input consistency and support efficient real-time inference, all images were resampled to 256x256 resolution using bilinear interpolation. To avoid patient-level information leakage and ensure evaluation fairness, 70% of the training set and 30% of the test set were divided by subject level. By integrating multiple public databases and uniformly sampling, a COVID-19 lung infection segmentation dataset containing 1277 high-quality chest CT images was constructed, with 894 training images and 383 test images to support robust model evaluation. All experiments were based on the PyTorch2.2.1 framework and CUDA12.4 environment, and were completed on an NVIDIA RTX2080 GPU. A unified training scheme was used: the Adam optimizer with a weight decay of 1x10 -4 -4 ; the calculation formula is ; all models were trained for 100 cycles with a batch size of 8. The loss function used a combination of cross-entropy and Dice loss to provide dual supervision for region classification and boundary outlining. Data augmentation included random horizontal / vertical flipping and size adjustment operations, and all training images were uniformly adjusted to 256x256 resolution.

[0143] ​First, ablation experiments were conducted. To explore the balance between accuracy and efficiency at different design scales, three model variants were evaluated: SFGNet-Base (basic version, abbreviated as SFGNet-B), SFGNet-Small (small version, abbreviated as SFGNet-S), and SFGNet-Tiny (micro version, abbreviated as SFGNet-T). Each variant was built by scaling the encoder-decoder architecture depth proportionally while retaining core modules (including residual downsampling blocks and stage-linked fusion units); SFGNet-B had a different configuration of the number of space-frequency guiding blocks in stages 2 and 3. Figure 7 and Figure 6 The comparison results are summarized as follows: Figure 7 As shown, SFGNet-B achieves the highest segmentation accuracy, with an F1 score of 96.98%, an intersection-over-union (IoU) ratio of 97.41%, precision of 97.33%, and recall of 97.81%. This demonstrates that this variant effectively captures fine-grained spatial structure and high-level semantic context. The medium-sized SFGNet-S follows closely behind, with an F1 score of 96.50% and an IoU ratio of 96.77%, and significantly reduced parameter and computational costs compared to the basic version. The lightest SFGNet-T, using only 0.314M parameters and 0.876 GFLOPs, still maintains a competitive F1 score of 96.01%, highlighting the efficiency advantages of its modular architecture. Figure 6 As shown, SFGNet-T leads with an inference speed of 294.67 FPS, while SFGNet-S and SFGNet-B achieve 254.57 FPS and 183.11 FPS, respectively. The negative correlation between network size and throughput confirms SFGNet's scalability under different resource constraints. Notably, the miniature variant maintains approximately 96% of the F1 score while reducing the number of parameters by about 20 times and improving computational efficiency by about 12 times compared to the basic version. Experimental results demonstrate SFGNet's excellent scalability on the accuracy-efficiency curve. The high performance of SFGNet-T validates the effectiveness of its core architectural innovations (especially the space-frequency bootstrapping block and multi-scale feature interaction decoder) even under strong compression conditions. SFGNet-S provides an ideal balance for real-time inference, combining high performance and high throughput. SFGNet-B is best suited for scenarios prioritizing extreme accuracy (such as offline tumor progression analysis).

[0144] To quantify the contribution of each architectural component, we conducted a series of ablation experiments by incrementally adding each component to the baseline model. The baseline model adopted the same encoder depth and down-sampling strategy as SFGNet, but replaced the spatial-frequency guidance block with a standard 3x3 convolution and used a normal convolution decoder without integrated multi-scale interaction. This design separates the performance gains from frequency-aware feature modeling and multi-scale fusion. The experimental results, as shown in Table 3, show that the SFGB with basic convolution fusion significantly improves the baseline model (F1 score +2.87%), demonstrating that designing an efficient feature extractor is superior to simply stacking features; the SFGB with spectral feature aggregation module further optimizes performance (F1 of 94.32%), verifying that joint processing of spatial and frequency cues can enhance lesion discrimination and boundary sensitivity. The final complete model integrating the multi-scale feature interaction module achieves the optimal segmentation accuracy (F1: 97.58%, IoU: 96.70%). These results demonstrate that gradual multi-scale guidance can continuously optimize spatial structures while preserving lesion spectral features, making MFIM a key component for tumor boundary reconstruction.

[0145] To evaluate the contribution of the proposed residual down-sampling module, we compared it with two widely used down-sampling strategies: max-pooling and strided convolution. As shown in Table 2, replacing RDBlock with max-pooling results in a significant drop in all indicators (especially F1 score and IoU), indicating that simple pooling operations lose boundary-sensitive information that is crucial for locating small nodules and irregular lesion contours. Strided convolution achieves modest improvements by learning a spatially adaptive down-sampling kernel, but its performance still lags far behind RDBlock. In contrast, RDBlock achieves the best results in all indicators (e.g., F1 score of 97.58% ± 0.31 and IoU of 96.70% ± 1.00), verifying the effectiveness of combining deep convolution optimization and residual preservation of fine-grained spatial cues. This design effectively alleviates the semantic information decay caused by resolution reduction, enabling the encoder to retain high-frequency structural information necessary for accurate lung CT image segmentation.

[0146] Table 2 - Ablation experiments of residual down-sampling block based on SFGNet-base on lung CT scan dataset:

[0147]

[0148] This ablation study reveals three key conclusions: (1) frequency guidance is crucial for lesion separation in low-contrast CT slices; (2) spectral feature aggregation outperforms simple fusion techniques in modeling texture and boundary cues; (3) multi-scale interaction brings the greatest performance gain by aligning coarse-grained semantics and fine-grained anatomical details. These findings collectively verify the rationality of SFGNet's design and highlight the practical value of spatial-frequency modeling in medical image segmentation.

[0149] Table 3 - SFGNet core architecture module ablation experiments:

[0150]

[0151] To explore the influence of receptive field scaling on spatial-frequency aggregation, ablation experiments were conducted on the dilated configuration of spatial-frequency guidance block. Since the dilated convolution can adjust the effective receptive field without increasing the parameters, choosing a suitable dilation factor is crucial to capture the lesions of different sizes in lung CT, especially the lesions with subtle morphological boundaries. Table 4 records the performance of five dilated configurations in three SFGB stages. Under the same training and evaluation protocol, it is found that when using a small dilation value (1, 1, 1), the receptive field is not enough to expand, resulting in a decline in the ability of context semantic encoding, and the performance is the weakest (F1: 90.83% ± 1.48). After increasing the dilation rate to (2, 2, 2), the segmentation quality is significantly improved (F1: 94.39% ± 2.61), which proves that moderately expanding the receptive field is advantageous to analyze medium-sized structures. However, uniform high dilation rate (3, 3, 3) will lead to performance degradation, which is due to grid artifacts and loss of fine boundary details. Although the non-uniform high dilation configuration (3, 7, 9) partially restores the performance, it shows that the heterogeneous receptive field pattern can improve the robustness to anatomical variations, but the stability is still insufficient. The (2, 4, 8) configuration proposed in the embodiment of the present invention achieves the optimal F1 score (97.58% ± 0.31) and intersection over union (96.70% ± 1.00), which verifies that gradually increasing the dilation rate between SFGB stages can form a coarse-to-fine aggregation hierarchy. This design makes the shallow layer retain local texture and the deep layer capture global spectral clues, ultimately enhancing the model's segmentation ability for irregular and low-contrast boundary tumors.

[0152] Table 4 - Ablation study of dilated configuration in SFGB:

[0153]

[0154] Secondly, comparative experiments. To strictly verify the performance and robustness of SFGNet, a variety of advanced medical image segmentation architectures are compared on the Lung CT dataset. The comparison methods include classic encoder-decoder CNN and attention enhanced methods (UNet, UNet++, Dense-UNet), specially designed architectures (R2UNet, BCDU-Net, Attention-UNet) and hybrid systems based on Transformer (UNETR, TransUNet, SCUNet++, FusionLungNet, PYAttUNet). All methods use the same data preprocessing, enhancement strategy and optimization settings, and are trained and evaluated at a resolution of 256x256. To ensure statistical reliability, each model is independently trained three times and the mean ± standard deviation is reported. The final results (mean% ± standard deviation%) of each model are shown in Table 5 and Figure 8 Figure 8 Each column of images in Table 5 is shown from left to right as the input image, the true label, the R2U-Net segmentation result, the Att-UNet segmentation result, the Dense-UNet segmentation result, the LadderNet segmentation result, the UNet segmentation result, the UNet++ segmentation result, the UNeXt segmentation result, the UNETR segmentation result, the FusionLungNet segmentation result and the SFGNet segmentation result.

[0155] Table 5 - Comparison of SFGNet variants and existing methods on the Lung CT scan dataset:

[0156]

[0157] ​Comprehensive benchmarking shows that SFGNet-B achieves the best performance with an F1 score of 97.58% ± 0.31 and an IoU of 96.70% ± 1.00. This result not only surpasses the classic CNN baseline, but also outperforms the parameter-heavy Transformer architecture (e.g., UNETR with 94.92M parameters and SCUNet++ with 77.23M parameters). The superiority of SFGNet-B highlights the effectiveness of introducing frequency domain priors and multi-scale spectral attention, which have recently shown potential in medical vision. Unlike Transformer systems that mainly rely on global self-attention, SFGNet employs spatial-spectral fusion to enhance the ability to distinguish subtle lesion boundaries and high-frequency anatomical cues. This is consistent with recent research findings that convolutional networks enhanced by FFT and DCT can improve boundary localization and enhance local noise robustness, which is crucial for representing lung CT lesions. The SFGNet-S variant demonstrates the excellent scalability of the architecture, achieving an F1 score of 96.50% ± 0.91 with an ultra-high throughput of 254.57 FPS. This performance significantly outperforms lightweight baseline models such as LadderNet and UNeXt, both of which exhibit lower accuracy and operational stability. The ultra-compact SFGNet-T (0.314M parameters) achieves an F1 score of 96.15% ± 1.07 FPS at an astonishing speed of 294.67 FPS, outperforming similar-sized models such as MalUNet, demonstrating that combining architecture compression and normalized spectral guidance can avoid the performance collapse problem commonly seen in aggressive pruning and miniature UNet variants.

[0158] To further evaluate the generalization ability of the SFGNet series, three variants were benchmarked on the COVID19-Seg dataset. Compared to the LungCT dataset, this dataset has higher intra-class variability, severe low-contrast boundaries, and irregular lesion morphology, making it difficult for encoder-decoder architectures to maintain fine structural cues. All comparative models were evaluated under the same settings at an input resolution of 256x256, with F1 score, IoU, precision, and recall used to measure segmentation quality, and parameter quantity, computational load, and frame rate used to evaluate real-time performance.

[0159] As Figure 9 , Figure 10and Table 6, SFGNet reconstructed clearer and more coherent tumor boundaries than competing models. UNet and Dense-UNet suffered from over-smoothing small nodules, while the Transformer baseline tended to produce contour discontinuities in low-contrast regions, a known drawback of purely attention-driven architectures. In contrast, the spectral modulation mechanism of SFGB ensured that high-frequency lesion features were preserved and enhanced throughout the network. The compact variant SFGNet-S further validated the efficiency advantage of the design. This model achieved an F1 score of 72.35% ± 1.51 with an ultra-high throughput of 254.57 FPS using only 1.35M parameters, outperforming UNeXt and LadderNet lightweight baselines by +1.90% and +5.31%, respectively, while maintaining comparable or higher running speeds, demonstrating that the space-frequency coupling mechanism provides significant performance gains for small models, which are typically limited by representation bottlenecks. The ultra-light configuration SFGNet-T achieved an F1 score of 70.64% ± 1.25 with 0.314M parameters and 0.876 GFLOPs, becoming the highest throughput among all models at 294.67 FPS. Despite its compact size, its F1 score still led MalUNet by 4.93%, indicating that introducing spectral domain priors is more valuable than simply pursuing extreme architecture compression.

[0160] Table 6 - Comparison of proposed SFGNet variants and existing methods on COVID19-Seg scan dataset

[0161]

[0162] From the above comparative experiments, it can be seen that in the LungCT and COVID19-Seg two benchmark tests, the empirical results reveal consistent and interpretable rules: integrating spatial-frequency representation into the encoder-decoder process can significantly improve the ability to outline fine anatomical structures and handle pathological-driven appearance variations. On the LungCT dataset with clear boundaries and relatively uniform intensity distribution, the SFGNet variant systematically outperforms the Transformer model (such as UNETR, SCUNet++) with 10-15 times the number of parameters, indicating that frequency-guided modulation can provide more efficient inductive bias. It is worth noting that the improvement in recall rate and intersection over union shows that SFGNet is superior to traditional residual or attention-enhanced backbone networks in capturing global lung morphology and subtle boundary transitions. On the more challenging COVID19-Seg dataset (lesions present diffuse, low-contrast and irregular morphology), the advantage of frequency-aware modeling is more prominent. Traditional CNN architectures tend to over-smooth high-frequency structures, and lightweight baseline models struggle to maintain sufficient sensitivity under aggressive parameter compression. In contrast, SFGNet variants (including compact SFGNet-T) still maintain excellent boundary localization and lesion integrity, demonstrating that the coupling of spatial features and Fourier domain cues can effectively improve the robustness to noise, scale variation and boundary ambiguity. The results of the two datasets can be summarized as follows: spatial-frequency interaction provides a standardized mechanism to improve representation expression, and does not affect efficiency, making it suitable for various real-time medical segmentation scenarios.

[0163] Referring to Figure 11 As shown in FIG. 1, it is a structural schematic diagram of a real-time lung CT image segmentation device based on a spatial frequency guide network according to an embodiment of the present application, which comprises:

[0164] The network training module 1101 constructs and trains the spatial frequency guide network based on the encoder-decoder structure;

[0165] The real-time segmentation module 1102 uses the trained spatial frequency guide network to perform real-time lung CT image segmentation.

[0166] The application proposes an SFGNet segmentation framework guided by spatial frequency, aiming to solve the core challenge of medical image analysis: achieving high-fidelity delineation of anatomical structures under strict efficiency constraints. By integrating residual down-sampling blocks with spatial frequency guiding blocks to achieve stable hierarchical coding, and combining multi-scale feature interaction modules to complete efficient feature fusion, SFGNet builds a standardized architecture that unifies fine-grained detail preservation and global context modeling. Comprehensive experiments on two complementary CT benchmarks, Lung CT and COVID19-Seg, show that SFGNet consistently outperforms classic CNN baselines, attention-enhanced designs, and Transformer-based architectures in key indicators such as F1 score, intersection over union, and recall rate. Importantly, these advantages are achieved with significantly reduced computational overhead, with even the most compact configuration (SFGNet-T) maintaining a real-time inference capability of 294.67 FPS. These achievements highlight the significant advantages of frequency-guided representation learning in improving boundary sensitivity and lesion integrity while maintaining excellent parameter efficiency.

[0167] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit it; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A real-time lung CT image segmentation method based on a spatial frequency-guided network, characterized in that, Includes the following steps: A spatial frequency guidance network was constructed and trained based on an encoder-decoder structure. Real-time lung CT image segmentation is performed using a trained spatial frequency-guided network. The spatial frequency steering network includes an encoder section, a decoder section, and a segmentation head. The encoder section includes a stem block and several encoders connected in sequence, used to extract features from lung CT images sequentially and output features at several different scales. The stem block uses stacked 3×3 convolutions to extract preliminary features from the lung CT images. The decoder section includes a multi-scale feature interaction module and several decoders connected in sequence. The multi-scale feature interaction module receives the output of the last-level encoder and guides fine-grained spatial feature optimization through coarse-grained semantic descriptors to achieve structured cross-scale interaction. The last-level decoder receives the output of the multi-scale feature interaction module and the output of the last-level encoder, and outputs decoded features. Other decoders receive the decoded features output by the next-level decoder and the encoded features output by the encoder at the same level, and output decoded features. The decoded features output by the first-level decoder are used as the final features. The segmentation head generates segmentation results based on the final features. Each level encoder includes a Spatial Frequency Guiding Module (SFGM) and a Stage Link Fusion Unit (SLFU). The SFGM consists of a Spatial Frequency Guiding Block (SFGB) and a Residual Downsampling Block (RDBlock). The RDBlock acts as a downsampling tool, learning low-resolution features with refined information from the input features. The SFGB acts as a basic bottleneck structure, capturing dense semantic context from the output of the RDBlock. The SLFU fuses the features output by the SFGM with the features from the previous stage to obtain the encoded features of the current level encoder and outputs them. The previous stage features of the first level encoder are the preliminary features output by the stem block, while the previous stage features of other encoders are the encoder features output by the previous level encoder. The spatial frequency guide block (SFGB) includes the following internal processing steps: Extracting local structure and multi-scale spatial context yields extracted features. , represented as: ； in, The input feature map represents the spatial frequency guide block (SFGB). To capture fine-grained details convolution, Represents convolution convolution kernel, It is a depthwise convolution with an inflation rate of d, which can expand the receptive field without increasing computational cost. Represents convolution The convolution kernel; Indicates batch normalization; The spatial branch of the spatial frequency attention module uses large kernel convolutions to capture a wider-area local context, resulting in a spatial token, represented as: ； Where S represents the spatial token; ρ is the ReLU activation; and Conv represents the convolution operation. Frequency domain branch pairs of spatial frequency attention modules For each channel, a two-dimensional Fast Fourier Transform is calculated, and the average amplitude across the channel dimensions is also calculated. Finally, the frequency token is projected to obtain the token, which is represented as: ； ； ； in, X represents c Calculate the result of the two-dimensional fast Fourier transform. This represents a two-dimensional Fast Fourier Transform. Indicates to The result of finding the absolute value; express The value of the c-th channel, Indicates the total number of channels. Indicates to The result of calculating the average amplitude of the channel dimension; This indicates that the frequency token α is a 1x1 convolution. Describes the kernel of convolution α; Fusion and attention weighting, combining spatial token S and frequency token Connect them to generate spatial frequency attention : ； Where || represents channel cascading. It is the sigmoid function; Represents convolution. Represents convolution The convolution kernel; The final output of the spatial frequency guide block SFGB is obtained by the following formula. : ； ； Here, ⊙ represents element-wise multiplication. It is a 1x1 convolution. It is convolution The convolution kernel; The multi-scale feature interaction module includes the following internal processing steps: The features output by the final stage link fusion unit SLFU are subjected to consecutive double 3×3 convolution operations to generate enhanced deep feature maps. ; The encoded features output by the last encoder stage are convolved with 1×1 to generate transform features. Then, coarse-grained features are obtained through refinement by a lightweight channel attention module. Finally, it is combined with enhanced deep feature maps. The fusion is represented as: ； Wherein, DK represents the output of the multi-scale feature interaction module.

2. The real-time lung CT image segmentation method based on a spatial frequency-guided network according to claim 1, characterized in that, The internal processing flow of the residual downsampling block RDBlock includes the following steps: Initial downsampling yields downsampling features. , represented as: ； in, For a 3×3 convolution with stride s=2, Represents convolution The convolution kernel; This represents the input features of the residual downsampling block RDBlock; Using 3×3 depthwise convolution downsampling features The features were refined and batch normalization was used to improve training stability, resulting in the desired features. , represented as: ； in, Represents convolution The convolution kernel; Feature concatenation along the channel dimension is represented as follows: ； Wherein, Concat represents the concatenation operation; splicing features Average pooling (k) is applied, and channel redundancy is reduced through 1×1 convolution to obtain features. , represented as: ； in, Represents a 1×1 convolution. Indicates operation The convolution kernel; The subsequent optimization training process, combined with residual connections, is represented as follows: ； in, This represents the final output of the residual downsampling block RDBlock.

3. The real-time lung CT image segmentation method based on a spatial frequency-guided network according to claim 1, characterized in that, The Stage Link Fusion Unit (SLFU) internal processing flow includes the following steps: Receive the output of the current stage Features of integration with preceding stages The process is handled using two parallel branches, as follows: ； ； in, and These represent the output results of the two parallel branches respectively; and Both represent separable convolutions. and Representing convolution and convolution The convolution kernel; The processing results are optimized and fused to obtain the final output of the Stage Linked Fusion Unit (SLFU), which is represented as: ))； in, Indicates batch normalization. Represents a 1×1 convolution. Represents convolution The convolution kernel.

4. The real-time lung CT image segmentation method based on a spatial frequency-guided network according to claim 1, characterized in that, The internal processing flow of the decoder is as follows: ； Where 𝜙(⋅) is a 3×3 convolution optimization operator, and ‖ represents channel concatenation. Indicates upsampling; This represents the decoder characteristics output by the k-level decoder; when k=K, The output of the multi-scale feature interaction module is used; when k≠K, This represents the decoder feature output by the k+1 level decoder; This represents the encoder characteristics output by the k-level encoder.

5. A real-time lung CT image segmentation device based on a spatial frequency-guided network, characterized in that, include: The network training module constructs and trains a spatial frequency guided network based on an encoder-decoder structure. The real-time segmentation module uses a trained spatial frequency-guided network to segment lung CT images in real time. The spatial frequency guided network comprises an encoder, a decoder, and a segmentation head. The encoder consists of a stem block and several encoders connected in sequence, used to extract features from lung CT images sequentially and output features at several different scales. The stem block uses stacked 3×3 convolutions to extract preliminary features from the lung CT images. The decoder consists of a multi-scale feature interaction module and several decoders connected in sequence. The multi-scale feature interaction module receives the output of the last-level encoder and guides fine-grained spatial feature optimization through a coarse-grained semantic descriptor, achieving structured cross-scale interaction. The last-level decoder receives the output of the multi-scale feature interaction module and the output of the last-level encoder, and outputs decoded features. Other decoders receive the decoded features output by the next-level decoder and the encoded features output by the same-level encoder, and output decoded features. The decoded features output by the first-level decoder serve as the final features. The segmentation head generates segmentation results based on these final features. Each level encoder includes a Spatial Frequency Guiding Module (SFGM) and a Stage Link Fusion Unit (SLFU). The SFGM consists of a Spatial Frequency Guiding Block (SFGB) and a Residual Downsampling Block (RDBlock). The RDBlock acts as a downsampling tool, learning low-resolution features with refined information from the input features. The SFGB acts as a basic bottleneck structure, capturing dense semantic context from the output of the RDBlock. The SLFU fuses the features output by the SFGM with the features from the previous stage to obtain the encoded features of the current level encoder and outputs them. The previous stage features of the first level encoder are the preliminary features output by the stem block, while the previous stage features of other encoders are the encoder features output by the previous level encoder. The spatial frequency guide block (SFGB) includes the following internal processing steps: Extracting local structure and multi-scale spatial context yields extracted features. , represented as: ； in, The input feature map represents the spatial frequency guide block (SFGB). To capture fine-grained details convolution, Represents convolution convolution kernel, It is a depthwise convolution with an inflation rate of d, which can expand the receptive field without increasing computational cost. Represents convolution The convolution kernel; Indicates batch normalization; The spatial branch of the spatial frequency attention module uses large kernel convolutions to capture a wider-area local context, resulting in a spatial token, represented as: ； Where S represents the spatial token; ρ is the ReLU activation; and Conv represents the convolution operation. Frequency domain branch pairs of spatial frequency attention modules For each channel, a two-dimensional Fast Fourier Transform is calculated, and the average amplitude across the channel dimensions is also calculated. Finally, the frequency token is projected to obtain the token, which is represented as: ； ； ； in, X represents c Calculate the result of the two-dimensional fast Fourier transform. This represents a two-dimensional Fast Fourier Transform. Indicates to The result of finding the absolute value; express The value of the c-th channel, Indicates the total number of channels. Indicates to The result of calculating the average amplitude of the channel dimension; This indicates that the frequency token α is a 1x1 convolution. Describes the kernel of convolution α; Fusion and attention weighting, combining spatial token S and frequency token Connect them to generate spatial frequency attention : ； Where || represents channel cascading. It is the sigmoid function; Represents convolution. Represents convolution The convolution kernel; The final output of the spatial frequency guide block SFGB is obtained by the following formula. : ； ； Here, ⊙ represents element-wise multiplication. It is a 1x1 convolution. It is convolution The convolution kernel; The multi-scale feature interaction module includes the following internal processing steps: The features output by the final stage link fusion unit SLFU are subjected to consecutive double 3×3 convolution operations to generate enhanced deep feature maps. ; The encoded features output by the last encoder stage are convolved with 1×1 to generate transform features. Then, coarse-grained features are obtained through refinement by a lightweight channel attention module. Finally, it is combined with enhanced deep feature maps. The fusion is represented as: ； Wherein, DK represents the output of the multi-scale feature interaction module.

Images