A medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning

Abstract

This application discloses a medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning. First, the input medical image is preprocessed to generate pixel-level semantic cues. Multi-level features are extracted using an encoder and decomposed into three factors: noise, style, and content, through a decoupling unit. Channel correlation weights are calculated based on foreground-background differences, and the channels are dynamically divided into three branches—local boundaries, non-local weak targets, and global semantics—using an optimal partitioning operator. These branches are then enhanced using gated depthwise convolution, sparse non-local attention, and global channel attention, respectively. Sparse cues for weak boundaries are generated using a cues expert database, and contrast alignment constraints are applied. Finally, multi-scale features are fused through cues-guided deformable convolution, resulting in a segmentation mask. This application effectively suppresses background interference, sharpens soft boundaries, and improves segmentation accuracy in low-contrast and complex scenes.

Description

Technical Field

[0001] This application relates to the fields of medical image processing, deep learning and computer-aided diagnosis technology, and specifically to a medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning. Background Technology

[0002] Medical image segmentation aims to accurately delineate the pixel-level boundaries of lesions, organs, or tissues from various medical images, and is a key technology for computer-aided diagnosis, treatment planning, and disease tracking. Accurate segmentation results can significantly improve diagnostic consistency, reduce the risk of misdiagnosis, and provide reliable support for clinical decision-making. In recent years, with the development of deep learning technology, methods based on convolutional neural networks (CNNs) and Transformers have made significant progress in medical image segmentation tasks, resulting in a series of representative works such as U-Net, TransUNet, and Swin-UNet. With the development of attention mechanisms, researchers have used inverse attention networks (such as PraNet) to guide models to focus on background regions to capture boundary details, or utilized feature decoupling strategies (such as UACANet and ConDSeg) to explicitly decompose encoded features into foreground, background, and uncertain regions to mitigate soft boundary interference.

[0003] However, the inherent imaging characteristics of medical images mean that segmentation tasks still face two core challenges:

[0004] (1) The problem of soft boundaries in the blurred transition between foreground and background. Due to physical limitations of imaging equipment, low contrast between tissues, and uneven illumination, there are often continuous gray-scale transition areas between lesion areas and normal tissues in medical images, forming "soft boundaries". These boundaries are neither clear edges nor pure background, making it difficult for the model to determine the accurate contour of the lesion. Existing edge detection methods mainly focus on the extraction of strong edges and lack effective modeling of soft boundary areas; although the reverse attention mechanism indirectly captures the boundary by focusing on the background area, it fails to actively suppress background noise interference. Feature decoupling methods mainly rely on heuristic learning of hard thresholds or independent branches during the decoupling process. The model parameter accuracy is insufficient and the effect of blurring region suppression is poor, resulting in blurred segmentation boundaries at soft boundaries.

[0005] (2) The problem of misleading coexistence between salient and insignificant targets. In complex medical images, lesions (salient targets) often appear simultaneously with normal tissue structures, artifacts, or background noise (insignificant targets), forming a "coexistence" phenomenon. In low-contrast environments, the features of the two are highly similar, which can easily cause confusion in the model. Existing methods mostly focus on the prediction of specific modal images, simply learning the mapping relationship between input and output, and cannot achieve accurate segmentation of complex scenes such as low contrast and coexistence of multiple targets. Although multi-scale convolution methods can extract features from different receptive fields, they lack a mechanism for distinguishing between salient and insignificant targets; although self-attention-based methods can capture global dependencies, they do not specifically suppress interference from insignificant regions. To achieve effective decoupling of foreground, background, and uncertain regions, it is necessary not only to consider the inherent contrast relationship at the feature numerical level, but also to mine additional supervisory information to actively suppress blurred regions.

[0006] In addition, existing multi-task methods often simply align shallow textures with deep semantics, ignoring the nonlinear differences between the two and the complementary information interaction during the decoupling process. This leads to information redundancy and alignment difficulties during feature fusion, affecting the model's discriminative ability.

[0007] In summary, sharpening soft boundaries in blurred transition regions and accurately distinguishing salient from non-salient objects in co-occurring target scenes remain core challenges that urgently need to be addressed in the field of medical image segmentation. To this end, this application proposes a medical image segmentation framework based on adaptive pixel-level correlation modulation optimal partitioning, combining frequency domain prior-guided feature decoupling, contrast-driven feature fusion, and a weak boundary adaptive sparse cueing mechanism to systematically address the aforementioned challenges. Summary of the Invention

[0008] To address the aforementioned issues, this application proposes a medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning. By constructing a feature decoupling and group attention collaborative optimization framework, it achieves fine modeling of soft boundary regions and adaptive fusion of multi-scale features, thereby significantly improving segmentation accuracy and robustness.

[0009] The technical solution of this application is:

[0010] A medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning includes the following steps:

[0011] Step 1: Preprocess the medical image data to be segmented, map the preprocessed medical image to the initial feature space, and generate pixel-level semantic prompts based on the initial image features;

[0012] Step 2: Construct a multi-level encoder network to extract hierarchical features from the initial image features, thereby obtaining multi-level encoded features;

[0013] Step 3: Through feature decoupling, the multi-level coding features obtained in Step 2 are decomposed into noise factors, style factors, and content factors;

[0014] Step 4: Adaptive pixel-level correlation modulation optimal partitioning. Based on the pixel-level semantic cues, the content factors, and the multi-level coding features, calculate the channel correlation weights, and use the adaptive pixel-level correlation modulation optimal partitioning operator to divide the feature channels into local boundary branches, non-local weak target branches, and global semantic branches.

[0015] Step 5: Branch-specific feature enhancement. Specialized feature enhancement is performed on the local boundary branch, non-local weak target branch, and global semantic branch respectively to obtain local boundary enhancement features, non-local weak target enhancement features, and global semantic enhancement features; after three-branch adaptive fusion, grouped attention output features are output.

[0016] Step 6: Weak boundary contrast cue alignment. Based on the grouped attention output features from Step 5, weak boundary sparse cue is generated using a pre-built cue expert library, and contrast alignment and weak boundary contrast cue constraints are applied to output contrast-enhanced multi-scale features. Step 7: Weak boundary adaptive fusion. The contrast-enhanced multi-scale features output from Step 6 are subjected to cue-guided deformable convolutional fusion to output multi-scale fused features. Step 8: Segmentation result output. Based on the multi-scale fusion features output in Step 7, fuse the features of each branch and output the final segmentation result.

[0017] In the above technical solution, the synergistic effect of the feature decoupling unit, the adaptive pixel-level correlation modulation optimal partitioning and grouping attention module, and the weak boundary contrast prompt alignment module achieves full-process optimization from feature separation and channel allocation to weak boundary enhancement.

[0018] The beneficial effects of this application are as follows:

[0019] This application enhances the modeling ability of blurred transition regions through soft boundary discrimination and pixel-level semantic prompting mechanisms, achieves effective separation of noise, style and content information through feature decoupling units, realizes on-demand allocation of channel-level computing resources through adaptive pixel-level correlation modulation optimal partitioning and grouping attention modules, and provides targeted enhancement to complex boundary regions through weak boundary contrast prompting alignment modules and weak boundary adaptive fusion modules. As a result, it significantly improves segmentation accuracy and stability under conditions of low contrast, multiple coexistence of targets and complex backgrounds. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall architecture of this application.

[0021] Figure 2This is the logic diagram of the feature decoupling unit in this application.

[0022] Figure 3 This is a logic diagram of the adaptive pixel-level correlation modulation optimal partitioning and grouping attention module in this application.

[0023] Figure 4 This is a visualization comparison of the segmentation results of the method in this application with several other comparative methods on five datasets: ISIC2018, Kvasir, MonuSeg, COVID-19, and BUSI. Detailed Implementation

[0024] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, and the terms "inner" and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0025] like Figure 1 , Figure 2 and Figure 3 As shown, this embodiment provides a medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning. First, the input medical image is preprocessed to generate pixel-level semantic cues such as foreground confidence and boundary ambiguity. Then, the hierarchical features of the image are extracted through a multi-level encoder, and the encoded features are decomposed into noise factors, style factors, and content factors using a feature decoupling unit. Further, channel correlation weights are calculated based on pixel-level semantic cues and content factors. The feature channels are divided into local boundary branches, non-local weak target branches, and global semantic branches through an adaptive pixel-level correlation modulation optimal partitioning operator. Subsequently, specialized feature enhancement is performed on the three branches respectively, and multi-scale features are fused through weak boundary contrast cues alignment and cue-guided deformable convolution. Finally, a pixel-level segmentation mask of the medical image is output.

[0026] Specifically, the steps include the following:

[0027] Step 1: Preprocessing and pixel-level semantic cue generation;

[0028] Step 11: Represent the input medical image as ,satisfy ,in, Indicates the image height. Indicates the image width. This represents the number of image channels. The medical image to be segmented undergoes preprocessing, including size normalization, pixel intensity normalization, and format unification. The preprocessed medical image is then mapped to the initial feature space to obtain the initial image features.

[0029] Pixel-level semantic cues are generated based on initial image features, including foreground confidence. and ambiguity ;in, , is used to represent the probability that the corresponding pixel belongs to the foreground lesion region or the target region; This indicates the degree to which a corresponding pixel lies within the blurred transition region between the foreground and background regions; boundary blur. Through local foreground features With local background features The difference between them was calculated as follows:

[0030] ;

[0031] in, Foreground features, This represents local background features.

[0032] The smaller the difference between foreground and background features, the higher the ambiguity, and the more difficult it is to distinguish the region, thus providing an effective pixel-level supervision signal for subsequent soft boundary modeling.

[0033] Step 2: Multi-level encoder hierarchical feature extraction;

[0034] An encoder network consisting of four cascaded coding modules is constructed to perform hierarchical feature extraction on the initial image features. Convolutional layers are then used to process the preprocessed medical image. Mapping to a high-dimensional feature space yields For the first Level coding module, in which =1,2,3,4, its input features are counted as , defined as: when When =1, ;when >1 hour ,in The intra-layer progressive features are the outputs of the previous level encoding module; each encoding module processes the input features. The process sequentially performs frequency domain feature mining, attention weighting, spatial feature enhancement, and nonlinear compensation. Its formalized flow is as follows:

[0035] ,

[0036] The operation of a multilayer perceptron is defined as follows:

[0037] ;

[0038] in Indicates the first The intra-layer progressive features output by the level coding module; Indicates upsampling convolution; Indicates downsampling convolution. Indicates normalization, represents the linear rectified activation function; MLP represents the multilayer perceptron;

[0039] Each level of the encoding module outputs two types of core hierarchical features. The first type is the intra-layer progressive feature. The first category serves as input to the next level of encoding modules; the second category consists of cross-layer interaction features, including high-frequency hierarchical features. and low-frequency hierarchical features The high-frequency and low-frequency hierarchical features are used for hierarchical retention and transmission to the corresponding level decoding modules in the decoder, forming a complete encoding and decoding interaction.

[0040] After being processed layer by layer by a four-level encoding module, the output is a vertically layered feature sequence. and cross-layer interaction feature set Together, they constitute multi-level coding features, enabling hierarchical mining and representation from pixel-level basic features to semantic-level advanced features.

[0041] Step 3: Decoupling through features;

[0042] Step 31: Decompose the multi-level coding features obtained in Step 2 into noise factors. Style factors and content factors The latent variables satisfy independent prior distributions. ,in It follows a standard normal distribution. It is the identity matrix;

[0043] Step 32: Generate via reparameterization:

[0044] ;

[0045] ;

[0046] ;

[0047] in, The mean of the noise factor. The mean of the style factor. The mean of the content factors; The standard deviation of the noise factor; The standard deviation of the style factor; The standard deviation of the content factor; For global Gaussian noise variables, It is the identity matrix. For element-wise multiplication; Let the Gaussian random noise variable in the reparameterized sampling satisfy the following conditions: Furthermore, its dimensions are consistent with the mean and standard deviation of the corresponding latent variables.

[0048] Style factors obtained from decoupling pass The formula implements style augmentation, and the augmented image reverse optimization decouples the processes, forming a collaborative closed loop:

[0049]

[0050] in For the original image, The mixing coefficient is a uniformly distributed coefficient. , To control the augmentation intensity using hyperparameters.

[0051] The total loss function for feature decoupling is derived from variational loss. Reconstruction loss Segmentation loss and decoupling loss Composition; specifically as follows:

[0052] (1) Variational loss This is used to constrain the distribution of latent variables to closely approximate the prior, achieving independent distribution optimization of the three factors:

[0053] ;

[0054] in, for divergence, , , These are the variational distributions of the noise factor, style factor, and content factor, respectively. The prior distribution of the noise factor is a standard normal distribution. , The prior distribution of style factors is a standard normal distribution. , The prior distribution of the content factors is a standard normal distribution. , Indicates whether the data has labels or not;

[0055] (2) Reconstruction loss The method for constraining the linear combination of noise factor, style factor, and content factor to reconstruct the input image is represented as:

[0056] ;

[0057] in, Mean square error, The image-level noise factor at the decoder output. The image-level style factor is the output of the decoder. The image-level content factor is the output of the decoder. The input image after style augmentation;

[0058] (3) Segmentation loss Used to constrain content factors to capture segmentation semantic features, denoted as

[0059] ;

[0060] in, The similarity coefficient is Dice. Content latent variables Decoded segmentation image, Can be a real label or a pseudo label?

[0061] (4) Decoupling loss The feature separation used to constrain noise factor, style factor, and content factor is represented as:

[0062] ;

[0063] in, The cosine similarity coefficient is... This represents the average cosine similarity between any two pairs of the three vectors. Style latent variables Decoded pseudo-segmentation image, For the latent variable characteristics of noise factors, For style factor latent variable characteristics, Features of latent variables for content factors;

[0064] There is loss of labeled data. and unlabeled data loss The sum of the above four losses is expressed as:

[0065] ;

[0066] ;

[0067] superscript This indicates that there is labeled data, superscript. This indicates unlabeled data.

[0068] Step 4: Adaptive pixel-level correlation modulation optimal partitioning;

[0069] Based on pixel-level semantic cues, content factors, and multi-level coding features, the representational contribution of each feature channel to the foreground target is calculated, and channel resources are allocated on demand according to relevance.

[0070] First, regarding the first Level encoder output features, calculate channel correlation weights :

[0071] ;

[0072] in, For the first Features of the level encoder output; where For encoder level index; Indicates the first Channel correlation weights for each feature channel; Indicates the channel index; The larger the value, the more significant the number. The higher the contribution of each feature channel to the characterization of the foreground lesion region or target region; It is a very small positive number;

[0073] according to and preset threshold The total channels are divided by the pixel-level correlation modulation optimal partitioning operator PRMOP. The adaptive division is divided into three dedicated branches:

[0074] ;

[0075] in, For local boundary branches, For non-local weak target branches, It is a global semantic branch, and ; This is the local boundary correlation threshold. The threshold for correlation between non-local weak targets;

[0076] Subsequently, the first Features of Level Encoder Output Based on the channel segmentation results, the features are divided into three branches:

[0077] ;

[0078] ;

[0079] ;

[0080] in, For local features, It is a non-local feature. For global features;

[0081] Simultaneously, a sparse gating function is introduced for feature enhancement:

[0082]

[0083] in, It is the sigmoid activation function. for Convolution operation, Value It is used to suppress sparsity in the background region and enhance features in the foreground and weak boundary regions.

[0084] Step 5: Enhance branch specialization features;

[0085] Specialized feature enhancements are performed on the three branches respectively, and lightweight operators are used to strengthen weak boundary, weak target context and global semantic features, thereby eliminating spatial and channel redundancy.

[0086] Local boundary branch: Gated depthwise separable convolutions are used to enhance fine-grained features of weak boundaries, and a gating mechanism is used to suppress local noise in the background region.

[0087] = ;

[0088] = ;

[0089] in, Representing local features Gated depth-separable intermediate features after convolution enhancement; It is a 3×3 depthwise convolution; This represents a 1×1 point convolution; Element-wise multiplication; This represents the weak boundary fine-grained enhancement feature output by the local boundary branch; Indicates basis The condition is to perform spatial sparse sampling on the branch features of nonlocal weak targets; This represents the normalization exponential function, used to normalize the nonlocal attention weights; This indicates that the feature interpolation is restored to... Spatial resolution operations.

[0090] In the nonlocal weak target branch, spatial sparse sampling is performed using a pixel-level correlation modulation gating map, only for targets that satisfy... Non-local attention is performed on the region, and the enhanced sparse features are restored to their original spatial resolution and then residually fused with the input features:

[0091] , ;

[0092] = ;

[0093] ;

[0094] in, This represents the input features of a nonlocal weak target branch; Represents a pixel-level correlation modulation gating map; Indicates the sparse sampling threshold; This represents the nonlocal weak target features after sparse sampling; This represents the features after non-local attention enhancement; This represents the output features of the nonlocal weak target branch; and These represent the height and width of the feature after sparse sampling, respectively. This represents the number of channels in the non-local weak target branch. and These represent the height and width of the original input features, respectively.

[0095] In the global semantic branch: spatial pyramid pooling and multilayer perceptron are used to obtain global channel attention weights. ;

[0096] = ;

[0097] ;

[0098] = ;

[0099] in, This represents the input features of the global semantic branch; Represents the multi-scale global context features after spatial pyramid pooling; The number of channels representing the global semantic branch; Indicates the global channel attention weight; This represents the enhanced features output by the global semantic branch; This demonstrates the average pooling operation. This represents the max pooling operation. This represents the operation of expanding multidimensional features into a one-dimensional vector;

[0100] Three-branch adaptive fusion: combines the enhanced features from the three branches, namely the weak-boundary fine-grained enhanced features output by the local boundary branch. Non-local weak target branch output features Enhanced features of global semantic branch output Stitching along the channel dimension, introducing adaptive fusion gating The concatenated features are weighted at the channel level, refined by lightweight convolution, and then the fused candidate features, adaptive fusion gating weights, and grouping attention features are calculated sequentially.

[0101] ;

[0102] = ;

[0103] = ;

[0104] in, This represents the fused candidate feature obtained by concatenating the three-branch enhanced features; Indicates adaptive fusion gating weights; Indicates the first Grouped attention output features corresponding to level-coded features; express Point convolution operation.

[0105] Step 6: Weak boundary comparison prompts for alignment;

[0106] Step 61: Pre-build a suggestion expert library ; To encode grayscale transition features, To enhance features for encoding high-frequency details, Encode shape-adaptive features;

[0107] For the group attention output features of step 5 Calculate the sparse activation weights:

[0108] ;

[0109] in, Indicates global average pooling. Indicates the first Sparse activation weights corresponding to each expert's hints: Only the maximum weight is retained, and the remaining weights are reset to 0. Weak boundary hints are generated as follows:

[0110] ;

[0111] Then, a comparison and alignment suggestion is performed, and the loss compared to the suggestion alignment is:

[0112] ;

[0113] Simultaneously, a weak boundary contrastive cueing loss is defined, which includes positive and negative sample constraints:

[0114] ;

[0115] ;

[0116] ;

[0117] in, For the pre-trained feature extraction network, and These are the features of positive and negative samples, respectively. For real labels, and These are constraints for positive and negative samples, respectively. These are the loss weighting coefficients used to balance the negative sample constraint term. This indicates a weak boundary comparison to suggest a loss;

[0118] The final output is contrast-enhanced multi-scale features. .

[0119] in, This indicates a depthwise separable convolution operation, which performs lightweight convolution enhancement on positive sample features.

[0120] Step 7: Weak Boundary Adaptive Fusion

[0121] For various weak boundary types such as blurred transitions, low contrast, and irregular edges, multi-scale feature adaptive fusion is achieved through deformable convolution guided by sparse cueing.

[0122] Step 71: First, process the contrast-enhanced multi-scale features output from Step 6. Enhance cross-level cascading with helpful prompts:

[0123] ;

[0124] in, Indicates the first Multi-scale features after contrast enhancement at each scale Indicates the relationship with the first Weak boundary sparsity hints corresponding to each scale express Convolution operation, This indicates element-wise multiplication. Indicates the enhanced version of the first... Individual scale features;

[0125] , ,

[0126] in, Indicates the first Each scale of cues enhances cross-layer cascaded features. This indicates cue enhancement features at an adjacent higher-level scale. Indicates scale index;

[0127] Step 72: Cue-guided deformable convolution:

[0128] ;

[0129] ;

[0130] ;

[0131] in, Indicates the offset prediction network, Represents a deformable convolutional network; where, Indicates the first The dilation rate at each scale in deformable convolution Deformable convolution offset under the given conditions This represents the offset prediction network. Indicates the relationship with the first Weak boundary sparsity hints corresponding to each scale This represents the dilation rate of deformable convolution; Indicates the first Deformable convolution output features enhanced with level contrast; Indicates the first Deformation enhancement features resulting from multi-scale deformable convolutional fusion with multiple dilation rates This represents a deformable convolution operation. This indicates that the convolution results under different dilation rates are summed and merged. This represents the set of dilation rates for deformable convolution; Indicates the first Each scale of prompts guides and enhances the fusion features. This indicates a batch normalization operation. This represents a non-linear activation function.

[0132] Step 73: Multi-scale fusion output yields multi-scale fusion features:

[0133] ; ∈{1,2,3},

[0134] in This indicates the corresponding convolution operation. This indicates a channel-level concatenation operation. These represent cue-guided enhancement fusion features at three different scales. For scale indexing.

[0135] Step 8: Output the segmentation results;

[0136] Based on the multi-scale fusion features output in step 7, the segmentation head... and learnable weights Generate segmentation probability map :

[0137] ,

[0138] Among them, for After thresholding, the final binary segmentation mask is obtained: 1 represents the foreground and 0 represents the background.

[0139] Training and Implementation Details

[0140] The entire network is trained end-to-end, and the total loss function is a weighted sum of segmentation loss and weak boundary contrast cueing loss.

[0141] Total loss from fusion

[0142] To lose weight, This refers to the segmentation loss of the content factors in step 3; The comparison of weak boundaries indicates the loss.

[0143] In practice, the entire network is implemented in the PyTorch framework, using the AdamW optimizer, and the learning rate scheduling adopts the CosineAnnealingLR strategy. The initial learning rate is set to... The training run consisted of 200 epochs with a batch size of 12. Input images were uniformly resized to 256×256 pixels, and data augmentation operations such as horizontal flipping, vertical flipping, and random rotation were employed to improve the model's generalization and robustness.

[0144] Experimental verification

[0145] To verify the effectiveness of the proposed medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning, five publicly available medical image segmentation datasets were selected for training, validation, and testing:

[0146] The ISIC2018 dataset is used for skin cancer lesion segmentation, containing 2594 skin lesion images, of which 2076 were used for training and 518 for testing. The Kvasir dataset is used for pixel-level segmentation of colorectal polyps, containing 1000 endoscopic images, of which 800 were used for training and 200 for testing. The BUSI dataset is a breast ultrasound image dataset containing normal, benign, and malignant breast ultrasound images, totaling 780 images, of which 624 were used for training and 156 for testing. The MonuSeg dataset is used for cell nucleus segmentation, containing 74 medical microscopic images, of which 59 were used for training and 15 for testing. The COVID-19 dataset is used for lung CT infection region segmentation, containing 894 images, of which 716 were used for training and 178 for testing. These datasets cover various medical imaging modalities, including dermatoscopy, endoscopy, ultrasound, microscopy, and CT images, and can be divided into training and testing sets in an 8:2 ratio.

[0147] The experiment used commonly used metrics in medical image segmentation, including mean Intersection over Union (mIoU), Dice similarity coefficient (DSC), pixel accuracy (Acc), specificity (Spe), and sensitivity (Sen). Specifically, mIoU measures the average overlap between the predicted segmented region and the ground truth labeled region; DSC evaluates the region similarity between the predicted result and the ground truth label; Acc evaluates the overall pixel-level classification accuracy; Spe evaluates the model's ability to identify background or negative regions; and Sen evaluates the model's ability to identify foreground lesion or positive regions. Higher values ​​for these metrics indicate better model segmentation performance.

[0148] The method described in this application is compared with several representative medical image segmentation methods, including MALU-Net (BIBM2022), SAM-Med2D, VPTTA, and this application. The segmentation performance of each method on five datasets is shown in Table 1, with the best performance indicated in bold.

[0149] Table 1 shows the segmentation performance of each medical image segmentation method on the ISIC2018, Kvasir, MonuSeg, COVID-19, and BUSI datasets, with the best performance indicated in bold.

[0150] Table 1 shows the performance comparison between the method of this application and several representative medical image segmentation methods on multiple datasets.

[0151]

[0152] As can be seen from the experimental results in Table 1, the medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning proposed in this application achieves good segmentation results on multiple publicly available medical image segmentation datasets. On the ISIC2018, Kvasir, COVID-19, and BUSI datasets, the mIoU, DSC, and Acc of the proposed method are generally superior to the comparative methods; on the MonuSeg dataset, the proposed method achieves the best mIoU index, indicating that it has a good modeling ability for the overall overlap segmentation of the cell nucleus region.

[0153] like Figure 4 As shown, the segmentation results of this application were visually compared with those of various contrast methods on five datasets. The results show that when there are blurred edges or soft boundaries in medical images, other contrast methods perform poorly in detecting significant target edges and are easily affected by soft boundaries, while this application can more accurately identify target segmentation boundaries and still has stable segmentation capabilities under low contrast and complex background conditions.

[0154] The comprehensive quantitative results show that the method proposed in this application is applicable to various medical image segmentation tasks, including dermoscopic images, endoscopic images, lung CT images, breast ultrasound images, and cell nuclear microscopy images. Experimental results demonstrate that this application, through pixel-level semantic prompts, feature decoupling, dynamic PRMOP channel partitioning, three-branch specialization enhancement, and weak boundary adaptive fusion, can enhance the model's segmentation ability for blurred boundaries, low-contrast targets, and complex background interference, thereby improving the accuracy and robustness of medical image segmentation.

Claims

1. A medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning, characterized in that, Includes the following steps: Step 1: Preprocess the medical image data to be segmented, map the preprocessed medical image to the initial feature space, and generate pixel-level semantic prompts based on the initial image features; Step 2: Construct a multi-level encoder network to extract hierarchical features from the initial image features, thereby obtaining multi-level encoded features; Step 3: Through feature decoupling, the multi-level coding features obtained in Step 2 are decomposed into noise factors, style factors, and content factors; Step 4: Adaptive pixel-level correlation modulation optimal partitioning. Based on the pixel-level semantic cues, the content factors, and the multi-level coding features, calculate the channel correlation weights, and use the adaptive pixel-level correlation modulation optimal partitioning operator to divide the feature channels into local boundary branches, non-local weak target branches, and global semantic branches. Step 5: Branch specialization feature enhancement. Specialization feature enhancement is performed on the local boundary branch, non-local weak target branch and global semantic branch respectively to obtain local boundary enhancement features, non-local weak target enhancement features and global semantic enhancement features; The output features are grouped by attention after three-branch adaptive fusion. Step 6: Weak boundary contrast cue alignment. Based on the grouped attention output features from Step 5, weak boundary sparse cue is generated using a pre-built cue expert library, and contrast alignment and weak boundary contrast cue constraints are applied to output contrast-enhanced multi-scale features. Step 7: Weak boundary adaptive fusion. The contrast-enhanced multi-scale features output from Step 6 are subjected to cue-guided deformable convolutional fusion to output multi-scale fused features. Step 8: Segmentation result output. Based on the multi-scale fusion features output in Step 7, fuse the features of each branch and output the final segmentation result.

2. The medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning according to claim 1, characterized in that, Step 1 specifically includes: Step 11: Represent the input medical image as ,satisfy ,in, Representing medical images, Indicates the image height. Indicates the image width. Indicates the number of image channels; Step 12: Perform size normalization, pixel intensity normalization and format unification processing on the medical image to be segmented, and map the preprocessed medical image to the initial feature space to obtain the initial image features; Step 13: Generate pixel-level semantic cues from the preprocessed medical image features, wherein the pixel-level semantic cues include foreground confidence. and boundary ambiguity ;in, , is used to represent the probability that the corresponding pixel belongs to the foreground lesion region or the target region; This is used to indicate the degree to which the corresponding pixel is located in the blurred transition area between the foreground and background regions; Boundary ambiguity Through local foreground features With local background features The difference between them was calculated as follows: ； in, Foreground features, This represents local background features.

3. The medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning according to claim 2, characterized in that, Step 2 specifically includes: Step 21: Use convolutional layers to process medical images Mapping to a high-dimensional feature space yields initial high-dimensional features. ; Step 22: The encoder consists of four cascaded encoding modules. For the first... Level coding module, in which =1,2,3,4, its input features are counted as ; when When =1, = ; when When >1, = ,in The intra-layer progressive features are the outputs of the previous level encoding module; each encoding module processes the input features. The process sequentially performs frequency domain feature mining, attention weighting, spatial feature enhancement, and nonlinear compensation. Its formalized flow is as follows: ； in, ; in Indicates the first The intra-layer progressive features output by the level coding module; Indicates upsampling convolution; Downsampling convolution, Indicates normalization. Represents the linear rectification activation function; This represents a multilayer perceptron; Each level of the encoding module outputs two types of core hierarchical features. The first type is the intra-layer progressive feature. The first category serves as input to the next level of encoding modules; the second category consists of cross-layer interaction features, including high-frequency hierarchical features. and low-frequency hierarchical features High-frequency and low-frequency hierarchical features are used for hierarchical retention and transmission to the corresponding level decoding modules in the decoder; Step 23: After processing layer by layer through the four-level encoding module, output the vertically layered feature sequence. and cross-layer interaction feature set Together, they constitute multi-level coding features.

4. The medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning according to claim 3, characterized in that, Step 3 specifically includes: Step 31: Feature decoupling decomposes multi-level encoded features into noise factors. Style factors and content factors The latent variables satisfy independent prior distributions. ,in It follows a standard normal distribution. It is the identity matrix; Step 32: Generate via reparameterization: ； ； ； in, The mean of the noise factor. The mean of the style factor. This represents the mean of the content factors; The standard deviation of the noise factor; The standard deviation of the style factor; The standard deviation of the content factor; Let the Gaussian random noise variable in the reparameterized sampling satisfy the following conditions: Furthermore, its dimensions are consistent with the mean and standard deviation of the corresponding latent variables. It is the identity matrix. For element-wise multiplication; Step 33: Construct the total loss function for feature decoupling, derived from variational loss. Reconstruction loss Segmentation loss and decoupling loss constitute.

5. The medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning according to claim 4, characterized in that, The loss functions in step 33 specifically include: Step 331: Variational Loss Represented as: ； in, for divergence, , , These are the variational distributions of the noise factor, style factor, and content factor, respectively. The prior distribution of the noise factor is a standard normal distribution. , The prior distribution of style factors is a standard normal distribution. , The prior distribution of the content factors is a standard normal distribution. , Indicates whether the data has labels or not; Step 332: Reconstruct the loss Represented as: ； in, Mean square error, The image-level noise factor at the decoder output. The image-level style factor is the output of the decoder. The image-level content factor is the output of the decoder. The input image after style augmentation; Step 333: Segmentation Loss Represented as: ； in, The similarity coefficient is Dice. Content latent variables Decoded segmentation image, Can it be a real label or a pseudo label? Step 334: Decoupling Loss Represented as: ； in, The cosine similarity coefficient is... Style latent variables Decoded pseudo-segmentation image, For the latent variable characteristics of noise factors, For style factor latent variable characteristics, Features of latent variables for content factors; There is loss of labeled data. and unlabeled data loss The sum of the above four losses is expressed as: ； ； Among them, superscript This indicates that there is labeled data, superscript. This indicates unlabeled data.

6. The medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning according to claim 5, characterized in that, Step 4 specifically includes: Step 41: Construct the adaptive pixel-level correlation modulation optimal partitioning operator PRMOP, which is used to dynamically partition the encoded feature channels according to the channel correlation weights; Step 42: Based on local foreground features Local background features and the Level encoder output features, calculate channel correlation weights : ； ； in, For the first Features of the encoder output, It is a very small positive number; among which For encoder level index; Indicates the first Channel correlation weights for each feature channel; Indicates the channel index; The larger the value, the more significant the number. The higher the contribution of each feature channel to the characterization of the foreground lesion region or target region; Step 43: According to and preset threshold The total channels are adaptively divided into three dedicated branches using the pixel-level correlation modulation optimal partitioning operator PRMOP: ； in, For local boundary branches, For non-local weak target branches, It is a global semantic branch, and ; This is the local boundary correlation threshold. The threshold for correlation between non-local weak targets; Step 44: Place the first Features of Level Encoder Output Based on the channel segmentation results, the features are divided into three branches: ； ； ； in, For local features, It is a non-local feature. For global features; Step 45: Feature enhancement using sparse gating functions: ； in, It is the sigmoid activation function. for Convolution operation, Value It is used to suppress sparsity in the background region and enhance features in the foreground and weak boundary regions.

7. The medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning according to claim 6, characterized in that, Step 5 specifically includes: Step 51: In local boundary branches, use gated depthwise separable convolutions to enhance fine-grained features of weak boundaries: = ； = ； in, Local features Gated depth-separable intermediate features after convolution enhancement; It is a 3×3 depthwise convolution. This represents a 1×1 point convolution. For element-wise multiplication, Indicates feature fusion operation; This represents the weak boundary fine-grained enhancement feature output by the local boundary branch; Step 52: In the non-local weak target branch, spatial sparse sampling is performed using a pixel-level correlation modulation gating map, only sampling that satisfies... Non-local attention is performed on the region, and the enhanced sparse features are restored to their original spatial resolution and then residually fused with the input features: ， ； = ； ； in, This represents the input features of a nonlocal weak target branch; Represents a pixel-level correlation modulation gating map; Indicates the sparse sampling threshold; This represents the nonlocal weak target features after sparse sampling; This represents the features after non-local attention enhancement; This represents the output features of the nonlocal weak target branch; and These represent the height and width of the feature after sparse sampling, respectively. This represents the number of channels in the non-local weak target branch. and These represent the height and width of the original input features, respectively. Indicates basis The condition is to perform spatial sparse sampling on the branch features of nonlocal weak targets; This represents the normalization exponential function, used to normalize the nonlocal attention weights; This indicates that the feature interpolation is restored to... Spatial resolution operations; Step 53: In the global semantic branch: Spatial pyramid pooling and multilayer perceptron are used to obtain global channel attention weights. ; = ； ； = ； in, This represents the input features of the global semantic branch; Represents the multi-scale global context features after spatial pyramid pooling; The number of channels representing the global semantic branch; Indicates the global channel attention weight; This represents the enhanced features output by the global semantic branch; This demonstrates the average pooling operation. This represents the max pooling operation. This represents the operation of expanding multidimensional features into a one-dimensional vector; Step 54: Fine-grained enhancement features of weak boundaries output from the local boundary branches of the three branches. Non-local weak target branch output features Enhanced features of global semantic branch output Stitching along the channel dimension, introducing adaptive fusion gating The concatenated features are weighted at the channel level, refined by lightweight convolution, and then output as grouped attention output features. : ； = ； = ； in, This represents the fused candidate feature obtained by concatenating the enhanced features from the three branches; Indicates adaptive fusion gating weights; Indicates the first The group attention output features corresponding to the level coding features, express Point convolution operation.

8. The medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning according to claim 7, characterized in that, Step 6 specifically includes: Step 61: Pre-build a suggestion expert library ; To encode grayscale transition features, To enhance features for encoding high-frequency details, Encode shape adaptation features; Step 62 For the input features output in step 54 Calculate the sparse activation weights: ; in, Indicates global average pooling. Indicates the first Sparse activation weights are assigned to each expert, with only the maximum activation weight retained and the remaining weights reset to 0. Weak boundary hints are generated as follows: ； Step 63: Perform a comparison alignment. The comparison indicates that the alignment loss is: ； The weak boundary comparison suggests the following loss: ； ； ； in, For the pre-trained feature extraction network, and These are the features of positive and negative samples, respectively. For real labels, and These are constraints for positive and negative samples, respectively. These are the loss weighting coefficients used to balance the negative sample constraint term. This indicates a weak boundary comparison to suggest a loss; Step 64: Output the contrast-enhanced multi-scale features. ; in, This indicates a depthwise separable convolution operation, which performs lightweight convolution enhancement on positive sample features.

9. A medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning according to claim 8, characterized in that, Step 7 specifically includes: Step 71: Enhance the contrast-enhanced multi-scale features output from Step 64 Enhance cross-level cascading with helpful prompts: ； in, Indicates the first Multi-scale features after contrast enhancement at each scale Indicates the relationship with the first Weak boundary sparsity hints corresponding to each scale express Convolution operation, This indicates element-wise multiplication. Indicates the enhanced version of the first... Individual scale features; ， ， in, Indicates the first Each scale of cues enhances cross-layer cascaded features. This indicates cue enhancement features at an adjacent higher-level scale. Indicates scale index; Step 72: Cue-guided deformable convolution: ； in, Indicates the first The dilation rate at each scale in deformable convolution Deformable convolution offset under the given conditions This represents the offset prediction network. Indicates the relationship with the first Weak boundary sparsity hints corresponding to each scale This represents the dilation rate of deformable convolution; ； in, Indicates the first Deformation enhancement features resulting from multi-scale deformable convolutional fusion with multiple dilation rates This represents a deformable convolution operation. This indicates that the convolution results under different dilation rates are summed and merged. This represents the set of dilation rates for deformable convolution; ； in, Indicates the first Each scale of prompts guides and enhances the fusion features. This indicates a batch normalization operation. Represents a nonlinear activation function; Step 73: Multi-scale fusion output yields multi-scale fusion features: ； ∈{1,2,3}， in, Indicates the first Deformable convolution output features enhanced with level contrast; This indicates the corresponding convolution operation. This indicates a channel-level concatenation operation. These represent cue-guided enhancement fusion features at three different scales. For scale indexing.

10. A medical image segmentation method based on adaptive pixel-level correlation modulation optimal partitioning according to claim 9, characterized in that, In step 8, the multi-scale fusion features output in step 73 are processed by the segmentation head. and learnable weights Generate segmentation probability map : ， Among them, for After thresholding, the final binary segmentation mask is obtained: 1 represents the foreground and 0 represents the background.

Images