A Medical Image Segmentation Method and System Based on Frequency Context Feature Hybridization

Abstract

This invention relates to the field of region segmentation technology in medical image analysis, specifically a medical image segmentation method and system based on frequency context feature fusion. First, the medical image to be processed undergoes image block-based graph embedding, feature rearrangement, and normalization to obtain a preprocessed medical image. Then, the preprocessed medical image undergoes multiple frequency-domain feature extractions and downsampling to obtain medical features of different depths. The channels of the mid-level and deep features are concatenated, and global context extraction and fusion processing are performed. Simultaneously, other depth features are processed using target features combined with auxiliary features for channel and spatial awareness, resulting in global enhancement features at each layer. Next, the global semantic enhancement features undergo multiple rounds of frequency-domain feature extraction and upsampling, and are fused with global enhancement features of different depths to obtain a global modality supplement image, which is used for segmentation processing, thus improving the accuracy of medical image segmentation results.

Description

Technical Field

[0001] This invention relates to the field of region segmentation technology in medical image analysis, specifically to a medical image segmentation method and system based on frequency context feature mixing. Background Technology

[0002] Medical image segmentation is a key problem at the intersection of medical analysis and computer vision. With the rapid development of artificial intelligence, smart healthcare based on computer-aided diagnosis is emerging. Medical image segmentation, as a crucial branch of smart healthcare, has become a current research hotspot. Recently, neural network models (such as the Mamba model) have demonstrated excellent modeling capabilities in various fields, exhibiting superior long-sequence processing capabilities while maintaining linear complexity. However, due to the multimodal nature, image heterogeneity, blurred boundaries, and disease complexity of medical images, existing methods often achieve low accuracy in medical image segmentation. Summary of the Invention

[0003] The purpose of this invention is to provide a medical image segmentation method and system based on frequency context feature mixing.

[0004] The technical solution of this invention is as follows:

[0005] A medical image segmentation method based on frequency context feature fusion includes the following operations:

[0006] S1. After the medical image to be processed is processed by graph embedding based on image block, feature rearrangement and normalization are performed to obtain the preprocessed medical image.

[0007] S2. The preprocessed medical image undergoes multiple feature extraction and downsampling processes based on frequency domain to obtain shallow medical features, mid-level medical features, mid-deep medical features, and deep medical features. The mid-level and deep medical features are concatenated by channel number to obtain structural semantic features. The structural semantic features are then extracted using global context to obtain context features. The context features are then mixed using global context to obtain global semantic enhancement features. Using any one of the shallow, mid-level, and mid-deep medical features as the target feature and the others as auxiliary features, channel-aware processing and spatial-aware processing are used to obtain shallow global enhancement features, mid-level global enhancement features, and mid-deep global enhancement features.

[0008] S3. The global semantic enhancement features are extracted using frequency-domain-based feature extraction to obtain the first global modality reconstruction image. The first global modality reconstruction image is upsampled and fused with the mid-to-deep global enhancement features. After frequency-domain-based feature extraction, the second global modality reconstruction image is obtained. The second global modality reconstruction image is upsampled and fused with the mid-level global enhancement features. After frequency-domain-based feature extraction, the third global modality reconstruction image is obtained. The third global modality reconstruction image is upsampled and fused with the shallow global enhancement features. After frequency-domain-based feature extraction, the global modality supplementary image is obtained. The global modality supplementary image is linearly projected and then segmented to obtain the medical image segmentation result.

[0009] In S2, deep medical features are obtained by downsampling and frequency-domain-based feature extraction from mid-deep medical features; mid-deep medical features are obtained by downsampling and frequency-domain-based feature extraction from mid-level medical features; and shallow medical features are obtained by frequency-domain-based feature extraction from preprocessed medical images.

[0010] The method for obtaining superficial medical features is as follows: the preprocessed medical image is mapped into four sub-frequency bands to obtain low-frequency band features, horizontal high-frequency band features, vertical high-frequency band features, and diagonal high-frequency band features; the low-frequency band features are processed by convolution, feature enhancement, and convolution to obtain low-frequency band detail features; the horizontal high-frequency band features, vertical high-frequency band features, and diagonal high-frequency band features are processed by convolution respectively, and then combined with the low-frequency band detail features by inverse wave transformation to obtain initial superficial enhancement features; the initial superficial enhancement features are processed by linear processing, depth convolution, and bidirectional detail scanning, and then subjected to layer normalization and linear processing to obtain superficial medical features.

[0011] In S2, the global context extraction operation is as follows: obtain the convolutional and pooling features of the structural semantic features, divide them equally by the number of channels, obtain the affinity matrix information, and obtain the context features.

[0012] In S2, the global context hybrid processing operation is as follows: after linear aggregation of context features, normalization and feature reshaping are performed to obtain context aggregated features; the context aggregated features and deep medical features are processed by channel attention mechanism to obtain global semantic enhancement features.

[0013] In S2, the method for obtaining shallow global enhancement features is as follows: shallow medical features are used as target features, and mid-level and mid-to-deep medical features are used as auxiliary features; the dimensions of the auxiliary features are mapped to the dimensions of the target features, and then concatenated with the target features to obtain multi-dimensional fusion features; the multi-dimensional fusion features are subjected to average pooling, max pooling, and convolution respectively, and then concatenated and nonlinearly processed to obtain multi-dimensional fusion weights; based on the multi-dimensional fusion weights, the multi-dimensional fusion features are weighted to obtain channel-aware features; the multi-dimensional fusion features are fused after convolution at different scales to obtain multi-scale convolutional fusion features; the multi-scale convolutional fusion features are subjected to average pooling and max pooling to obtain spatial-aware features; the spatial-aware features and channel-aware features are superimposed to obtain shallow global enhancement features.

[0014] The linear projection operation in S3 is achieved by linearly interpolating the global modality supplement image to reshape it to a dimension scale four times that of the global modality supplement image, and then normalizing it.

[0015] A medical image segmentation system based on frequency context feature fusion, used to implement the above-mentioned medical image segmentation method based on frequency context feature fusion, includes:

[0016] The preprocessing medical image generation module is used to perform feature rearrangement and normalization on the medical image to be processed after it has undergone image block-based graph embedding processing to obtain the preprocessed medical image.

[0017] The enhanced feature generation module is used to preprocess medical images through multiple frequency-domain-based feature extraction and downsampling processes to obtain shallow, medium, medium-deep, and deep medical features. The medium and deep medical features are then concatenated by channel number to obtain structural semantic features. These structural semantic features are then extracted using global context to obtain context features. Finally, the context features are processed through global context mixing to obtain global semantic enhancement features. Using any one of the shallow, medium, or medium-deep medical features as the target feature, and the others as auxiliary features, channel-aware and spatial-aware processing is applied to obtain shallow, medium, and medium-deep global enhancement features.

[0018] The medical image segmentation result generation module is used to obtain a first global modality reconstructed image by extracting global semantic enhancement features based on frequency division domain; after upsampling, the first global modality reconstructed image is fused with mid-to-deep global enhancement features and then extracted based on frequency division domain to obtain a second global modality reconstructed image; after upsampling, the second global modality reconstructed image is fused with mid-level global enhancement features and then extracted based on frequency division domain to obtain a third global modality reconstructed image; after upsampling, the third global modality reconstructed image is fused with shallow global enhancement features and then extracted based on frequency division domain to obtain a global modality supplementary image; after linear projection, the global modality supplementary image is segmented to obtain the medical image segmentation result.

[0019] A medical image segmentation device based on frequency context feature mixing includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the above-described medical image segmentation method based on frequency context feature mixing.

[0020] A computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described medical image segmentation method based on frequency context feature mixing.

[0021] The beneficial effects of this invention are as follows:

[0022] This invention provides a medical image segmentation method based on frequency context feature fusion. First, the medical image to be processed undergoes image block-based graph embedding, feature rearrangement, and normalization to obtain a preprocessed medical image, providing a regular and high-quality input for subsequent analysis. Then, the preprocessed medical image is subjected to multiple frequency domain feature extractions and downsampling to obtain medical features of different depths. The channels of the mid-level and deep features are concatenated, and global context extraction and fusion processing are performed to further explore the correlations between features and strengthen the coherence and dominance of global semantic information. Simultaneously, other deep features are processed using target features combined with auxiliary features for channel and spatial awareness to obtain global context information at each layer. Enhanced features; then, starting from global semantic enhancement features, through multiple rounds of frequency domain feature extraction, upsampling, and fusion of global enhancement features at different depths, the gradual refinement and precise integration of multi-scale global information is achieved. This not only preserves the overall coherence of global semantics, but also continuously injects more detailed local details and structural information through layer-by-layer fusion with mid-deep, mid-level, and shallow features, so that the final global modality supplement image can take into account both global context and local features at each level, improving the richness and completeness of feature expression; finally, after linear projection, the global modality supplement image is segmented to obtain medical image segmentation results, improving the accuracy of medical image segmentation. Attached Figure Description

[0023] The advantages and solutions of this application will become clear to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0024] In the attached diagram:

[0025] Figure 1 This is a flowchart illustrating the method of this embodiment.

[0026] Figure 2 The image shown here is a medical image segmentation result of the method described in this embodiment. Detailed Implementation

[0027] Exemplary embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings.

[0028] This embodiment provides a medical image segmentation method based on frequency context feature fusion, see [link to relevant documentation]. Figure 1 This includes the following operations:

[0029] S1. After the medical image to be processed is processed by graph embedding based on image block, feature rearrangement and normalization are performed to obtain the preprocessed medical image.

[0030] S2. The preprocessed medical image undergoes multiple feature extraction and downsampling processes based on frequency domain to obtain shallow medical features, mid-level medical features, mid-deep medical features, and deep medical features. The mid-level and deep medical features are concatenated by channel number to obtain structural semantic features. The structural semantic features are then extracted using global context to obtain context features. The context features are then mixed using global context to obtain global semantic enhancement features. Using any one of the shallow, mid-level, and mid-deep medical features as the target feature and the others as auxiliary features, channel-aware processing and spatial-aware processing are used to obtain shallow global enhancement features, mid-level global enhancement features, and mid-deep global enhancement features.

[0031] S3. The global semantic enhancement features are extracted using frequency-domain-based feature extraction to obtain the first global modality reconstruction image. The first global modality reconstruction image is upsampled and fused with the mid-to-deep global enhancement features. After frequency-domain-based feature extraction, the second global modality reconstruction image is obtained. The second global modality reconstruction image is upsampled and fused with the mid-level global enhancement features. After frequency-domain-based feature extraction, the third global modality reconstruction image is obtained. The third global modality reconstruction image is upsampled and fused with the shallow global enhancement features. After frequency-domain-based feature extraction, the global modality supplementary image is obtained. The global modality supplementary image is linearly projected and then segmented to obtain the medical image segmentation result.

[0032] The specific operational details are as follows.

[0033] S1. After image block-based graph embedding processing, the medical image to be processed undergoes feature rearrangement and normalization to obtain a preprocessed medical image.

[0034] The medical images to be processed undergo image block-based graph embedding, feature rearrangement, and normalization to transform the images into structured features and optimize the feature distribution, resulting in preprocessed medical images that provide regular and high-quality input for subsequent analysis.

[0035] First, the medical image to be processed undergoes image block-based graph embedding processing to obtain multi-block graph embedding features. Specifically, the medical image to be processed is subjected to image block-based graph embedding processing to obtain multi-block graph embedding features. For input, where H For height, W For width, In C Given the number of channels, graph embedding processing implements projection operations through 2D convolution (Conv2D), which can be represented as:

[0036] ,

[0037] Embed features for multi-block graphs, x For the medical image to be processed, The convolution kernel weights include embedded dimensionality information, input channels, and block size. This is a bias term.

[0038] Then, the multi-block image embedding features, which exist in multiple blocks, are rearranged according to their position order in the medical image to be processed, and normalized to obtain the preprocessed medical image. This process can be represented as:

[0039] ,

[0040] in For preprocessing medical images, B For batch size, C For the number of channels, For feature rearrangement, For normalization.

[0041] S2. The preprocessed medical image undergoes multiple feature extraction and downsampling processes based on frequency domain to obtain shallow medical features, intermediate medical features, mid-deep medical features, and deep medical features. The mid-deep medical features and deep medical features are concatenated by channel number to obtain structural semantic features. The structural semantic features are then extracted through global context to obtain context features. The context features are then mixed through global context to obtain global semantic enhancement features. Using any one of the shallow, intermediate, or mid-deep medical features as the target feature and the others as auxiliary features, channel-aware processing and spatial-aware processing are used to obtain shallow global enhancement features, mid-deep global enhancement features, and mid-deep global enhancement features.

[0042] By extracting and downsampling features in multiple frequency domains, medical features at different depths were obtained, providing multi-scale and multi-level basic information for subsequent analysis. Channel concatenation of mid-level and deep-level features yielded structural semantic features, achieving the fusion of abstract and structural information and enhancing the structural expression of features. Global context extraction and hybrid processing further explored the correlation between features, strengthening the coherence and dominance of global semantic information. At the same time, channel and spatial perception processing was performed using target features combined with auxiliary features to obtain global enhanced features at each layer. This not only highlighted the core position of the target features but also supplemented details and context with auxiliary features, enabling each layer of features to retain its own characteristics while possessing richer semantic and spatial information, providing comprehensive and accurate feature support for subsequent tasks such as global modality reconstruction.

[0043] First, the preprocessed medical image undergoes multiple feature extraction and downsampling processes based on frequency domain to obtain detailed information at different depths in the preprocessed medical image. This results in shallow medical features that provide basic texture, contour, and color; mid-level medical features that provide relatively clear structural information of the lesion area; mid-to-deep medical features that delve into the semantic-related features of the lesion and contain "qualitative" feature clues of the lesion; and deep medical features that provide highly abstract semantic features.

[0044] Among them, deep medical features are obtained by downsampling and frequency-domain-based feature extraction of mid-deep medical features; mid-deep medical features are obtained by downsampling and frequency-domain-based feature extraction of mid-level medical features; mid-level medical features are obtained by downsampling and frequency-domain-based feature extraction of shallow medical features; and shallow medical features are obtained by frequency-domain-based feature extraction of preprocessed medical images.

[0045] Taking the acquisition method of superficial medical features as an example, the operation steps of feature extraction based on frequency division domain are as follows.

[0046] Step 1: Map the preprocessed medical image into four sub-bands to obtain low-frequency band features that provide global context for subsequent analysis, horizontal high-frequency band features that help focus on lesion morphology details in the vertical direction, vertical high-frequency band features that facilitate the analysis of lesion structural information in the horizontal direction, and diagonal high-frequency band features that improve the lesion edge and internal details.

[0047] Step 2: Low-frequency features are processed by convolution (preferably 3×3 convolution), feature enhancement, and convolution to obtain low-frequency detailed features.

[0048] The feature enhancement operation is as follows: the low-frequency convolutional features are processed by layer normalization and linearization to obtain low-frequency linear features; the low-frequency linear features are processed by depthwise convolution to obtain low-frequency deep convolutional features; the low-frequency deep convolutional features are processed by semantic enhancement to evolve the low-frequency linear features from basic contour materials into semantic features for precise discrimination, resulting in low-frequency semantically enhanced features; the low-frequency semantically enhanced features are normalized by layer and multiplied element-wise with the low-frequency linear features to obtain initial enhanced features; the initial low-frequency enhanced features are processed by linearization and added element-wise with the low-frequency convolutional features to obtain low-frequency enhanced features, which are then used to perform convolution operations.

[0049] The aforementioned low-frequency convolutional features are obtained by convolutional processing of low-frequency features.

[0050] The semantic enhancement process described above is as follows: Low-frequency deep convolutional features are spatially decoupled (achieved through adaptive or deformable convolution) and channel decoupled (achieved through attention weighting of low-frequency deep convolutional features) to capture global contour deformation information in low-frequency features, such as the approximate shape distortion of skin lesions, while simultaneously highlighting low-frequency features such as pigment distribution and edge gradient corresponding channels, resulting in low-frequency global contour features and low-frequency channel feature maps; the low-frequency global contour features and low-frequency channel feature maps are fused to obtain low-frequency intermediate features; the low-frequency intermediate features are enhanced by multi-scale dilated convolution to enhance low-frequency features at different scales, resulting in multi-scale low-frequency enhanced features; after fusing the multi-scale low-frequency enhanced features with the low-frequency deep convolutional features, channel attention processing is performed to strengthen the key semantic information of low-frequency features for lesion diagnosis while retaining the basic contour information of the initial low-frequency features, resulting in low-frequency semantically enhanced features.

[0051] Step 3: The horizontal high-frequency band features, vertical high-frequency band features, and diagonal high-frequency band features are convolved (preferably 3×3 convolution) and then combined with the low-frequency band detail features through inverse wave transformation to obtain the initial shallow enhancement features.

[0052] Step 4: After linear processing, depth convolution and bidirectional detail scanning, the initial shallow enhancement features are normalized and linearly processed to obtain shallow medical features.

[0053] The aforementioned bidirectional detail scanning includes forward scanning and reverse scanning. Forward scanning is achieved through semantic enhancement processing, while reverse scanning is achieved through bidirectional deformable convolution. This allows bidirectional scanning to preserve the continuous spatial correlation of medical features and is suitable for analyzing the gradual details of lesion boundaries (such as the bidirectional transition zone from normal skin to lesions).

[0054] The frequency-domain based feature extraction methods used in acquiring mid-level, mid-deep, and deep medical features are similar to those used for shallow medical features, and will not be repeated here to save space.

[0055] Then, the mid-level medical features and deep medical features are concatenated by channel number to combine structural morphology and lesion abstract attribute information to obtain structural semantic features. The structural semantic features are extracted by global context to accurately capture the semantic space association of the same type and clearly depict the lesion structure to obtain context features. The context features are then mixed by global context to enhance semantic information and global perception capabilities to obtain global semantic enhancement features.

[0056] The above global context extraction operation is as follows: obtain the convolutional features and pooling features of structural semantic features, divide them equally by the number of channels, and let different channel branches focus on specific semantics (such as contours, pigments, etc.) to avoid mixed semantic interference, obtain affinity matrix information, and obtain context features.

[0057] The specific calculation formula is as follows:

[0058] ,

[0059] ,

[0060] ,

[0061] For structural semantic features, , These are structural semantic convolutional features and structural semantic pooling features, respectively. , These are convolution weights and pooling weights, respectively. For shape reshaping, For pooling operations, G is the total number of channels for even distribution. For the first g Transpose of the structural semantic convolutional features of each channel For the first g Structural semantic pooling features of each channel For the first g Affinity matrix information for each channel.

[0062] The above global context blending process is as follows: after linear aggregation of context features, normalization (achieved through the softmax function) and feature reshaping are performed to obtain aggregated context features; these aggregated context features are then combined with deep medical features through a channel attention mechanism to obtain globally semantically enhanced features. Linear aggregation is achieved by weighting the context features.

[0063] Finally, using any one of the superficial medical features, intermediate medical features, and intermediate-deep medical features as the target feature, and the others as auxiliary features, the superficial global enhancement features, intermediate global enhancement features, and intermediate-deep global enhancement features are obtained through channel perception processing and spatial perception processing.

[0064] The method for obtaining shallow global enhancement features is as follows: shallow medical features are used as target features, and mid-level and mid-to-deep medical features are used as auxiliary features; the dimensions of the auxiliary features are mapped to the dimensions of the target features, and then concatenated with the target features to obtain multi-dimensional fusion features; the multi-dimensional fusion features are subjected to average pooling, max pooling, and convolution, and then concatenated and nonlinearly processed to obtain multi-dimensional fusion weights; based on the multi-dimensional fusion weights, the multi-dimensional fusion features are weighted to obtain channel-aware features; the multi-dimensional fusion features are fused after convolution at different scales, with 3×3, 5×5, and 7×7 being the preferred scales for convolution, to obtain multi-scale convolutional fusion features (achieving channel-aware processing); the multi-scale convolutional fusion features are subjected to average pooling and max pooling to capture global information in the feature map and retain the saliency of local features, to obtain spatial-aware features (achieving spatial-aware processing); the spatial-aware features and channel-aware features are superimposed to obtain shallow global enhancement features.

[0065] The acquisition methods for mid-layer global enhancement features and mid-deep global enhancement features are similar to those for shallow global enhancement features. The acquisition method for mid-layer global enhancement features uses mid-layer medical features as the target features, while the acquisition method for mid-deep global enhancement features uses mid-deep medical features as the target features.

[0066] S3. The global semantic enhancement features are extracted using frequency-domain-based feature extraction to obtain the first global modality reconstruction image. The first global modality reconstruction image is upsampled and fused with the mid-to-deep global enhancement features. After frequency-domain-based feature extraction, the second global modality reconstruction image is obtained. The second global modality reconstruction image is upsampled and fused with the mid-level global enhancement features. After frequency-domain-based feature extraction, the third global modality reconstruction image is obtained. The third global modality reconstruction image is upsampled and fused with the shallow global enhancement features. After frequency-domain-based feature extraction, the global modality supplementary image is obtained. The global modality supplementary image is linearly projected and then segmented to obtain the medical image segmentation result.

[0067] By starting with global semantic enhancement features, and through multiple rounds of frequency-domain feature extraction, upsampling, and fusion with global enhancement features of different depths (shallow, mid-level, and mid-to-deep), a gradual refinement and precise integration of multi-scale global information is achieved. This preserves the overall coherence of global semantics while continuously injecting more detailed local details and structural information through layer-by-layer fusion with mid-to-deep, mid-level, and shallow features. This allows the final global modality supplement image to take into account both global context and local features at various levels, improving the richness and completeness of feature representation and restoring image resolution. It enhances the ability to process image context information while maintaining computational efficiency, providing more comprehensive and accurate modal information support for subsequent medical image segmentation and improving the accuracy of image segmentation. (Segmentation results are shown in [reference]). Figure 2 The segmentation effect of skin lesion images is shown, where the original image is a three-channel image, and the label image and the experimental result image (segmentation image) are binary images.

[0068] The above linear projection operation can be achieved by linearly interpolating the global modality supplement image to reshape it to a dimension scale four times that of the global modality supplement image, and then normalizing it.

[0069] This embodiment also provides a medical image segmentation system based on frequency context feature fusion, used to implement the above-mentioned medical image segmentation method based on frequency context feature fusion, including:

[0070] The preprocessing medical image generation module is used to perform feature rearrangement and normalization on the medical image to be processed after it has undergone image block-based graph embedding processing to obtain the preprocessed medical image.

[0071] The enhanced feature generation module is used to preprocess medical images through multiple frequency-domain-based feature extraction and downsampling processes to obtain shallow, medium, medium-deep, and deep medical features. The medium and deep medical features are then concatenated by channel number to obtain structural semantic features. These structural semantic features are then extracted using global context to obtain context features. Finally, the context features are processed through global context mixing to obtain global semantic enhancement features. Using any one of the shallow, medium, or medium-deep medical features as the target feature, and the others as auxiliary features, channel-aware and spatial-aware processing is applied to obtain shallow, medium, and medium-deep global enhancement features.

[0072] The medical image segmentation result generation module is used to obtain a first global modality reconstructed image by extracting global semantic enhancement features based on frequency division domain; after upsampling, the first global modality reconstructed image is fused with mid-to-deep global enhancement features and then extracted based on frequency division domain to obtain a second global modality reconstructed image; after upsampling, the second global modality reconstructed image is fused with mid-level global enhancement features and then extracted based on frequency division domain to obtain a third global modality reconstructed image; after upsampling, the third global modality reconstructed image is fused with shallow global enhancement features and then extracted based on frequency division domain to obtain a global modality supplementary image; after linear projection, the global modality supplementary image is segmented to obtain the medical image segmentation result.

[0073] This embodiment also provides a medical image segmentation device based on frequency context feature mixing, including a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the above-described medical image segmentation method based on frequency context feature mixing.

[0074] This embodiment also provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described medical image segmentation method based on frequency context feature mixing.

[0075] This embodiment provides a medical image segmentation method based on frequency context feature fusion. First, the medical image to be processed undergoes image block-based graph embedding, feature rearrangement, and normalization to obtain a preprocessed medical image, providing a regular and high-quality input for subsequent analysis. Then, the preprocessed medical image is subjected to multiple frequency domain feature extractions and downsampling to obtain medical features of different depths. After concatenating the channels of the mid-level and deep features, global context extraction and fusion processing are performed to further explore the correlation between features and strengthen the coherence and dominance of global semantic information. At the same time, other deep features are processed with target features combined with auxiliary features for channel and spatial awareness to obtain the full context of each layer. The process begins with global semantic enhancement features. Then, through multiple rounds of frequency-domain feature extraction, upsampling, and fusion of global enhancement features at different depths, a gradual refinement and precise integration of multi-scale global information is achieved. This preserves the overall coherence of the global semantics while continuously injecting more detailed local details and structural information through layer-by-layer fusion with mid-deep, mid-level, and shallow features. This ensures that the final global modality supplement image can consider both the global context and local features at each level, enhancing the richness and completeness of feature expression. Finally, the global modality supplement image is linearly projected and then segmented to obtain medical image segmentation results, improving the accuracy of medical image segmentation.

Claims

1. A medical image segmentation method based on frequency context feature mixing, characterized in that, The method comprises the following steps: S1, after the to-be-processed medical image is processed based on image block-based graph embedding, feature rearrangement and normalization processing are performed to obtain a preprocessed medical image; S2, the preprocessed medical image is processed by multiple times of feature extraction and downsampling based on frequency domain to obtain shallow medical features, middle medical features, middle-deep medical features and deep medical features; The deep medical features are obtained by downsampling and feature extraction based on frequency domain on the middle-deep medical features; The middle-deep medical features are obtained by downsampling and feature extraction based on frequency domain on the middle medical features; The shallow medical features are obtained by feature extraction based on frequency domain on the preprocessed medical image; the method for obtaining the shallow medical features comprises the following steps: the preprocessed medical image is mapped to four sub-frequency bands to obtain low-frequency band features, horizontal high-frequency band features, vertical high-frequency band features and diagonal high-frequency band features; the low-frequency band features are processed by convolution, feature enhancement and convolution to obtain low-frequency band detail features; the horizontal high-frequency band features, the vertical high-frequency band features and the diagonal high-frequency band features are processed by convolution respectively, and then are inversely changed with the low-frequency band detail features to obtain initial shallow enhanced features; The initial shallow enhanced features are processed by linear processing, deep convolution and bidirectional detail scanning, and then are subjected to layer normalization and linear processing to obtain the shallow medical features; The middle medical features and the deep medical features are subjected to channel number splicing processing to obtain structure semantic features; The structure semantic features are subjected to global context extraction to obtain context features; The context features are subjected to global context mixing processing to obtain global semantic enhanced features; Any one of the shallow medical features, the middle medical features and the middle-deep medical features is taken as a target feature, and others are taken as auxiliary features, and then the target feature and the auxiliary features are subjected to channel perception processing and spatial perception processing to obtain shallow global enhanced features, middle global enhanced features and middle-deep global enhanced features; The method for obtaining the shallow global enhanced features comprises the following steps: the shallow medical features are taken as target features, and the middle medical features and the middle-deep medical features are taken as auxiliary features; after the dimensions of the auxiliary features are mapped to the dimensions of the target features, the target features and the auxiliary features are spliced to obtain multi-dimensional fusion features; the multi-dimensional fusion features are subjected to average pooling, maximum pooling and convolution respectively, and then are subjected to splicing and nonlinear processing to obtain multi-dimensional fusion weights; based on the multi-dimensional fusion weights, the multi-dimensional fusion features are subjected to weighted processing to obtain channel perception features; the multi-dimensional fusion features are subjected to convolution processing of different scales and then are fused to obtain multi-scale convolution fusion features; the multi-scale convolution fusion features are subjected to average pooling and maximum pooling to obtain spatial perception features; the spatial perception features and the channel perception features are subjected to superposition processing to obtain the shallow global enhanced features; S3, the global semantic enhancement feature is obtained by frequency domain-based feature extraction on the first global modal reconstruction image; the first global modal reconstruction image is up-sampled and fused with the middle-deep global enhancement feature, and the second global modal reconstruction image is obtained by frequency domain-based feature extraction; the second global modal reconstruction image is up-sampled and fused with the middle global enhancement feature, and the third global modal reconstruction image is obtained by frequency domain-based feature extraction; the third global modal reconstruction image is up-sampled and fused with the shallow global enhancement feature, and the global modal supplementary image is obtained by frequency domain-based feature extraction; After linear projection of the global modal supplementary image, segmentation processing is performed to obtain a medical image segmentation result.

2. The medical image segmentation method based on frequency context feature mixing according to claim 1, characterized in that, In S2, the operation of global context extraction is: The convolutional features and the pooling features of the structural semantic features are obtained, and the affinity matrix information is obtained after the channel number is evenly divided, to obtain the context features. 3.The medical image segmentation method based on frequency context feature mixing according to claim 1, characterized in that, In S2, the operation of global context mixing processing is: After linear aggregation of the context features, normalization and feature remodeling processing are performed to obtain the context aggregation features; The context aggregation features and the deep medical features are processed by the channel attention mechanism to obtain the global semantic enhancement features. 4.The method of claim 1, wherein, In S3, the operation of linear projection is to perform linear interpolation on the global modal supplementary image, reshape the global modal supplementary image to four times the dimension scale, and perform normalization processing.

5. A medical image segmentation system based on frequency context feature blending for implementing the frequency context feature blending medical image segmentation method of claim 1, characterized in that, It includes: A preprocessed medical image generation module is configured to perform feature rearrangement and normalization processing on the processed medical image after image block-based graph embedding processing to obtain a preprocessed medical image; An enhancement feature generation module is configured to obtain shallow medical features, middle medical features, middle-deep medical features, and deep medical features by multiple frequency domain-based feature extraction and down-sampling processing on the preprocessed medical image; and perform channel number splicing processing on the middle medical features and the deep medical features to obtain structural semantic features; The structural semantic features are subjected to global context extraction to obtain context features; The context features are subjected to global context mixing processing to obtain global semantic enhancement features; Any one of the shallow medical features, the middle medical features, and the middle-deep medical features is taken as a target feature, and the others are taken as auxiliary features, and channel perception processing and spatial perception processing are performed to obtain shallow global enhancement features, middle global enhancement features, and middle-deep global enhancement features; A medical image segmentation result generation module is configured to obtain a first global modal reconstruction image by frequency domain-based feature extraction on the global semantic enhancement features; the first global modal reconstruction image is up-sampled and fused with the middle-deep global enhancement feature, and a second global modal reconstruction image is obtained by frequency domain-based feature extraction; the second global modal reconstruction image is up-sampled and fused with the middle global enhancement feature, and a third global modal reconstruction image is obtained by frequency domain-based feature extraction; The third global modality reconstructed image is up-sampled and fused with the shallow layer global enhancement feature, and a global modality supplementary image is obtained through frequency domain-based feature extraction; the global modality supplementary image is linearly projected and segmented to obtain a medical image segmentation result.

6. A medical image segmentation apparatus based on frequency context feature mixing, characterized by, A processor and a memory are included, wherein the processor implements the medical image segmentation method based on frequency context feature mixing according to any one of claims 1-4 when executing a computer program stored in the memory.

7. A computer readable storage medium characterized in that, A computer program for storing is included, wherein the computer program is executed by a processor to implement the medical image segmentation method based on frequency context feature mixing according to any one of claims 1-4.

Images