A boundary enhancement model for semi-supervised medical image segmentation

Abstract

The application discloses a boundary enhancement model for semi-supervised medical image segmentation and relates to the technical field of medical image segmentation.The SBEM-UNet model is provided, which significantly improves the boundary precision and detail recovery effect of semi-supervised medical image segmentation and provides an efficient solution for complex scenes; a semantic boundary enhancement module (SBEM) is designed, the module significantly improves the boundary sensitivity and semantic expression capability through structure modeling and multi-scale feature fusion, and provides high-quality features for medical image segmentation; and a contour enhancement decoder is introduced, the decoder combines a dynamic contour attention mechanism and a multi-scale feature fusion technology, can accurately focus on the boundary area and recover fine details, and thus enhances the segmentation effect and improves the prediction accuracy in complex scenes.

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation technology, specifically to a boundary enhancement model for semi-supervised medical image segmentation. Background Technology

[0002] Medical image segmentation provides indispensable support for clinical diagnosis, treatment planning, and disease monitoring by accurately distinguishing anatomical structures from pathological regions. However, complex anatomical structures (such as the heart and prostate) significantly increase the difficulty of segmentation tasks due to their diverse shapes and textures, blurred boundaries caused by low contrast or imaging defects, and individual patient differences. Traditional algorithms based on manual features struggle to effectively capture complex boundary information, while early deep learning models (such as U-Net) are insufficient to meet the clinical needs for high-precision boundary and detail restoration in semi-supervised scenarios where labeled data is scarce.

[0003] Recent developments, such as semi-supervised frameworks based on contrastive learning (e.g., methods inspired by SimCLR) and segmentation frameworks based on diffusion models (e.g., Denoising Diffusion Probabilistic Models), have improved segmentation performance by enhancing feature representations or generating pseudo-labels, but their robustness in scenarios with blurred boundaries and their adaptability to complex anatomical structures remain limited.

[0004] Therefore, a new solution is needed to address the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a boundary enhancement model for semi-supervised medical image segmentation to solve the technical problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a boundary enhancement model for semi-supervised medical image segmentation, wherein the boundary enhancement model is an SBEM-UNet model, including a contour enhancement decoder, a semantic boundary enhancement module, and a loss function;

[0007] The semantic boundary enhancement module, also known as the SBEM module, aims to extract and enhance boundary-related information from the deepest feature E5 of the encoder, so as to significantly improve the boundary accuracy of semi-supervised medical image segmentation.

[0008] By integrating a contour enhancement module, the prediction accuracy of the target boundary is significantly improved. The contour enhancement decoder receives features extracted by the encoder, fuses multi-scale features through skip connections, and progressively upsamples to generate the final prediction result.

[0009] The loss function employs a combination of supervised and semi-supervised loss functions.

[0010] Furthermore, the SBEM module is divided into two parallel data streams: a structural modeling stream and a multi-scale context stream, which are responsible for local boundary detail extraction and global semantic information modeling, respectively. The structural modeling stream is the top of the graph, and the multi-scale context stream is the bottom of the graph.

[0011] The operation process of the SBEM module includes at least the following steps:

[0012] The structural modeling flow (the upper part of the graph, from the input (x) to the projected block) is responsible for extracting boundary and contextual features, providing basic information for boundary enhancement.

[0013] Data stream from input It begins by extracting local boundary details through edge convolutional blocks. The computation process is described as follows:

[0014] f e =ReLU(BN(Conv) 3×3,s＝2 (E5))) (1)

[0015] In this process, a standard 3×3 convolution (with a stride of 2) is applied to the input feature map E5, followed by batch normalization (BN) and ReLU activation to obtain the final output feature map. It focuses on local gradient changes and emphasizes edge details, making it particularly suitable for detecting low-contrast boundaries in medical images;

[0016] Meanwhile, the contextual convolutional block extracts global semantic information from the same input x and computes it as follows:

[0017] f g =ReLU(BN(Conv) 3×3,d＝2,s＝2 (E5)))(2)

[0018] The context convolutional block applies a 3×3 dilated convolution (dilation rate set to 2, stride of 2) to the E5, followed by Batch Normalization (BN) and ReLU processing, finally outputting f. g By expanding the receptive field, global semantic information can be captured, and the understanding of the context can be enhanced;

[0019] Two branch features f e and f g The data will be stitched together along the channel dimension, and the calculation result is as follows:

[0020] f concat =Concat(f c ,f g ,dim=1) (3)

[0021] Here, this step combines local boundary information with global semantic information, providing diverse feature representations for subsequent fusion and enhancing the model's understanding of boundary regions;

[0022] The fusion block integrates and concatenates features through a 1×1 convolution, and the calculation result is as follows:

[0023] E′5=ReLU(BN(Conv 1×1 (f fused (4)

[0024] In this case, the 1x1 convolution will convert f, which is used to generate the target dimension (e.g., the number of output channels), into the desired dimension. fused The number of channels is adjusted to the target value, and key features are extracted at the same time; then, batch normalization (BN) standardizes the distribution of features to enhance linearity, thereby generating more refined features E5.

[0025] Next, multi-scale feature extraction will be performed, and the formula is as follows:

[0026] A1 = Conv 1×1 (E′5) (5)

[0027] Then, the receptive field features of E'5 are extracted using the expanded spatial pyramid pooling technique with convolutional kernels of different sizes:

[0028] A2=ASPPConv1(E′5) (6)

[0029] A3=ASPPConv3(E′5) (7)

[0030] A4=ASPPConv4(E′5) (8)

[0031] Among them, ASPPConv k This represents a dilated convolution with an expansion rate of k, which can capture contextual information at different scales;

[0032] Furthermore, by performing pooling operations in the expanded spatial pyramid pooling, global context information can be extracted:

[0033] A5 = ASPPool(E5') (9)

[0034] Finally, by concatenating along the channel dimension and integrating contextual information at different scales, the final multi-scale feature representations, A1, A2, A3, A4, and A5, are obtained, with the following formula:

[0035] A concat =Concat(A1,A2,A3,A4,A5) (10)

[0036] After SEBlock processing, an enhanced feature representation E5 is obtained, and its formula is as follows:

[0037] E5” = SEBlock(A concat (11)

[0038] Then, the feature is weighted and processed by the channel attention module and the spatial attention module respectively;

[0039] The channel attention module adaptively adjusts the channel weights, while the spatial attention module highlights important spatial regions, thereby generating a weighted feature E. 5,CA and E 5,SA This provides a more discriminative representation for fusion, and its formula is as follows:

[0040]

[0041] Among them, channel attention first applies to the input feature map E5 ” Global average pooling (GAP) is performed to obtain a global description for each channel. Then, two convolutional layers (W1 for dimensionality reduction and W2 for dimensionality increase) and the ReLU activation function are used to learn the non-linear relationships between channels. Finally, channel attention weights are generated using the sigmoid function and passed through element-wise multiplication. This is applied to the original features to achieve weighted importance for each channel.

[0042]

[0043] Here, MaxPool and AvgPool are max pooling and average pooling, respectively, used to extract spatial features. Their outputs are concatenated ([.,.]) and then processed by a convolutional layer (Conv), followed by the application of the Sigmoid function (σ) to generate an attention map. This attention map is then processed by element-wise multiplication (·) with E5. ” Multiply to obtain the output E 5,SA This highlights the key features.

[0044] Finally, channel attention E 5,CA Spatial attention E 5,SA By fusing attention combinations, a weighted feature representation is obtained:

[0045]

[0046] Among them, the feature E weighted by channel attention 5,CA and features E weighted by spatial attention 5,SA By element-wise multiplication ( The multiplication of corresponding elements is used to fuse the data, thereby simultaneously emphasizing important information in both the channel and spatial domains, and obtaining the final weighted feature representation E. 5,final .

[0047] Furthermore, the contour enhancement decoder introduces a dynamic contour extraction and feature weighting mechanism, effectively improving the segmentation effect at blurred boundaries, through the segmentation head S i Generate prediction probability map The formula is as follows:

[0048]

[0049] Where f d It comes from the previous layer decoder D. i+1 Features;

[0050] Contour-enhanced decoders can extract contour features more accurately by simulating the erosion and dilation processes in morphological operations. c The formula for its use is as follows:

[0051]

[0052] I min =sof tmax(-I) (17)

[0053] I max =max pool(I min (18)

[0054] Here, I represents the prediction probability map obtained through the softmax function. The normalization process is performed to generate a smooth probability distribution; I min Local minimum features are extracted by simulating the erosion process in morphology by applying max pooling to the negative probability distribution -I; max to I min Max pooling is performed again to simulate dilation, thereby enhancing the expressive power of boundary features;

[0055] The formula for calculating the profile is as follows:

[0056]

[0057] Here, f c Describes the contour features, where D is the number of categories. and Let each represent the maximum and minimum eigenvalues ​​of the i-th category, and ReLU is a modified linear unit function that ensures the result is non-negative;

[0058] Based on the original contour enhancement decoder, an attention mechanism is introduced to optimize the original contour features f. c To enhance its boundary representation ability, the specific method is as follows: f c The input is fed into the enhancement module, which then undergoes channel expansion convolution, LeakyReLU activation, and channel recovery convolution to achieve feature weighting and refinement. The formula for the enhancement process is as follows:

[0059] f c '=Conv(LeakyReLU(convenience(f c ))) (20)

[0060] Here, Conv represents the convolution operation (kernel size is 3, padding is 1), and LeakyReLU represents the activation function;

[0061] To further optimize and enhance the features, a residual structure was adopted, and the final enhanced feature fo was obtained through the HardSigmoid function, as shown in the following formula:

[0062] f0 = HardSigmoid((f e Mf d )+(f e Mf d )×f c ) (twenty one)

[0063] Here, M represents the merge operation, and HardSigmoid is an activation function.

[0064] Furthermore, the loss function includes a supervised loss L for labeled data. sup and semi-supervised loss L for unlabeled data semi ;

[0065] It can fully utilize both labeled and unlabeled data, thereby improving the model's ability to perceive boundaries and its segmentation accuracy. The loss function is defined as follows:

[0066] L = L sup +λL semi (twenty two)

[0067] in, Represents the dynamic balance weights, where t and t total These represent the current iteration number and the total number of iterations, respectively.

[0068] Compared with the prior art, the beneficial effects of the present invention are:

[0069] 1. This invention proposes the SBEM-UNet model, which significantly improves the boundary accuracy and detail restoration effect of semi-supervised medical image segmentation, providing an efficient solution for complex scenarios;

[0070] 2. This invention designs a semantic boundary enhancement module (SBEM), which significantly improves boundary sensitivity and semantic expressiveness through structural modeling and multi-scale feature fusion, providing high-quality features for medical image segmentation;

[0071] 3. This invention introduces a contour enhancement decoder that combines a dynamic contour attention mechanism and multi-scale feature fusion technology. This decoder can accurately focus on the boundary region and restore fine details, thereby enhancing the segmentation effect and improving the prediction accuracy in complex scenes. Attached Figure Description

[0072] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0073] Figure 1 This is a schematic diagram of the SBEM module of the present invention (first page of the attached diagram);

[0074] Figure 2 This is a schematic diagram of the universal decoder and contour enhancement decoder of the present invention (first page of the attached figures);

[0075] Figure 3 This is a schematic diagram of the contour decoder structure of the present invention (see page 2 of the attached diagram);

[0076] Figure 4 This is a schematic diagram of Table 1 of the present invention (see page 2 of the attached diagram);

[0077] Figure 5 This is a schematic diagram of Table 2 of the present invention (see page 3 of the attached diagram);

[0078] Figure 6 This is a schematic diagram of Table 3 of the present invention (see page 3 of the attached diagram);

[0079] Figure 7 This is a schematic diagram of Table 4 of the present invention (see page 3 of the attached diagram);

[0080] Figure 8 This is a schematic diagram of Table 5 of the present invention (see page 4 of the attached diagram);

[0081] Figure 9 A schematic diagram comparing the Dice index of this invention and ABD (see page 4 of the attached diagram);

[0082] Figure 10 This is a schematic diagram comparing the 95HD index of the present invention and ABD (see page 4 of the attached diagram). Detailed Implementation

[0083] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0084] Medical image segmentation plays a crucial role in clinical diagnosis, treatment planning, and disease monitoring, providing key support for medical decision-making by accurately dividing anatomical structures and pathological regions. However, complex anatomical morphology, low-contrast boundary blurring, noise interference, and inter-individual variability significantly increase the difficulty of segmentation tasks, especially in semi-supervised scenarios where labeled data is scarce. To address these challenges, this invention proposes an innovative model, SBEM-UNet, aimed at improving the boundary accuracy and robustness in semi-supervised medical image segmentation. This model integrates a self-developed Semantic Boundary Enhancement Module (SBEM) with an introduced contour enhancement decoder, constructing an efficient boundary enhancement architecture. SBEM effectively captures global semantic information and local boundary details through an innovative strategy of structural modeling and multi-scale feature fusion, significantly enhancing the sensitivity and semantic expressiveness of boundary regions. Complementing this, the contour enhancement decoder employs a dynamic contour attention mechanism, adaptively focusing on boundary regions and combining multi-scale feature fusion to accurately recover the fine details of complex anatomical structures. Experimental results demonstrate that SBEM-UNet significantly improves segmentation performance in complex scenes across various medical imaging modalities (such as CT and MRI), exhibiting particularly outstanding robustness and accuracy in low-contrast and blurred boundary regions. Its efficient semi-supervised learning capabilities make it a promising candidate for application in clinical environments with limited annotation resources, providing an innovative solution for high-precision medical image segmentation tasks.

[0085] See Figures 1-3 A boundary enhancement model for semi-supervised medical image segmentation, the boundary enhancement model is the SBEM-UNet model, which includes a contour enhancement decoder, a semantic boundary enhancement module and a loss function;

[0086] The Semantic Boundary Enhancement Module (SBEM) aims to extract and enhance boundary-related information from the deepest feature E5 of the encoder, thereby significantly improving the boundary accuracy of semi-supervised medical image segmentation.

[0087] By integrating a contour enhancement module, the prediction accuracy of target boundaries is significantly improved. The contour enhancement decoder receives features extracted by the encoder, fuses multi-scale features through skip connections, and progressively upsamples to generate the final prediction result.

[0088] The loss function employs a combination of supervised and semi-supervised loss functions.

[0089] The SBEM module consists of two parallel data streams: the structural modeling stream and the multi-scale context stream, which are responsible for local boundary detail extraction and global semantic information modeling, respectively. The structural modeling stream is the top of the graph, and the multi-scale context stream is the bottom of the graph.

[0090] The SBEM module's operation process includes at least the following steps:

[0091] The structural modeling flow (the upper part of the graph, from the input (x) to the projected block) is responsible for extracting boundary and contextual features, providing basic information for boundary enhancement.

[0092] Data stream from input It begins by extracting local boundary details through edge convolutional blocks. The computation process is described as follows:

[0093] f e =ReLU(BN(Conv) 3×3,s＝2 (E5))) (1)

[0094] In this process, a standard 3×3 convolution (with a stride of 2) is applied to the input feature map E5, followed by batch normalization (BN) and ReLU activation to obtain the final output feature map. It focuses on local gradient changes and emphasizes edge details, making it particularly suitable for detecting low-contrast boundaries in medical images;

[0095] Meanwhile, the contextual convolutional block extracts global semantic information from the same input x and computes it as follows:

[0096] f g =ReLU(BN(Conv) 3×3,d＝2,s＝2 (E5))) (2)

[0097] The context convolutional block applies a 3×3 dilated convolution (dilation rate set to 2, stride of 2) to the E5, followed by Batch Normalization (BN) and ReLU processing, finally outputting f. g By expanding the receptive field, global semantic information can be captured, and the understanding of the context can be enhanced;

[0098] Two branch features f e and f g The data will be stitched together along the channel dimension, and the calculation result is as follows:

[0099] f concat =Concat(f c ,f g ,dim=1) (3)

[0100] Here, this step combines local boundary information with global semantic information, providing diverse feature representations for subsequent fusion and enhancing the model's understanding of boundary regions;

[0101] The fusion block integrates and concatenates features through a 1×1 convolution, and the calculation result is as follows:

[0102] E′5=ReLU(BN(Conv 1×1 (f fused (4)

[0103] In this case, the 1x1 convolution will convert f, which is used to generate the target dimension (e.g., the number of output channels), into the desired dimension. fused The number of channels is adjusted to the target value, and key features are extracted at the same time; then, batch normalization (BN) standardizes the distribution of features to enhance linearity, thereby generating more refined features E5.

[0104] Next, multi-scale feature extraction will be performed, and the formula is as follows:

[0105] A1 = Conv 1×1 (E′5) (5)

[0106] Then, the receptive field features of E'5 are extracted using the expanded spatial pyramid pooling technique with convolutional kernels of different sizes:

[0107] A2=ASPPConv1(E′5) (6)

[0108] A3=ASPPConv3(E′5) (7)

[0109] A4=ASPPConv4(E′5) (8)

[0110] Among them, ASPPConv k This represents a dilated convolution with an expansion rate of k, which can capture contextual information at different scales;

[0111] Furthermore, by performing pooling operations in the expanded spatial pyramid pooling, global context information can be extracted:

[0112] A5 = ASPPool(E5) ' (9)

[0113] Finally, by concatenating along the channel dimension and integrating contextual information at different scales, the final multi-scale feature representations, A1, A2, A3, A4, and A5, are obtained, with the following formula:

[0114] A concat=Concat(A1,A2,A3,A4,A5) (10)

[0115] After SEBlock processing, an enhanced feature representation E5 is obtained. ” The formula is as follows:

[0116] E5 ” =SEBlock(A concat (11)

[0117] Then, the feature is weighted and processed by the channel attention module and the spatial attention module respectively;

[0118] The channel attention module adaptively adjusts the channel weights, while the spatial attention module highlights important spatial regions, thereby generating a weighted feature E. 5,CA and E 5,SA This provides a more discriminative representation for fusion, and its formula is as follows:

[0119]

[0120] Among them, channel attention first applies to the input feature map E5 ” Global average pooling (GAP) is performed to obtain a global description for each channel. Then, two convolutional layers (W1 for dimensionality reduction and W2 for dimensionality increase) and the ReLU activation function are used to learn the non-linear relationships between channels. Finally, channel attention weights are generated using the sigmoid function and passed through element-wise multiplication. This is applied to the original features to achieve weighted importance for each channel.

[0121]

[0122] Here, MaxPool and AvgPool are max pooling and average pooling, respectively, used to extract spatial features. Their outputs are concatenated ([.,.]) and then processed by a convolutional layer (Conv), followed by the application of the Sigmoid function (σ) to generate an attention map. This attention map is then processed by element-wise multiplication (·) with E5. ” Multiply to obtain the output E 5,SA This highlights the key features.

[0123] Finally, channel attention E 5,CA Spatial attention E 5,SA By fusing attention combinations, a weighted feature representation is obtained:

[0124]

[0125] Among them, the feature E weighted by channel attention 5,CAand features E weighted by spatial attention 5,SA By element-wise multiplication ( The multiplication of corresponding elements is used to fuse the data, thereby simultaneously emphasizing important information in both the channel and spatial domains, and obtaining the final weighted feature representation E. 5,final .

[0126] The contour enhancement decoder introduces dynamic contour extraction and feature weighting mechanisms, effectively improving segmentation performance at blurred boundaries. This is achieved through the segmentation head S... i Generate prediction probability map The formula is as follows:

[0127]

[0128] Where f d It comes from the previous layer decoder D. i+1 Its characteristics.

[0129] Contour-enhanced decoders can extract contour features more accurately by simulating the erosion and dilation processes in morphological operations. c The formula for its use is as follows:

[0130]

[0131] I min =softmax(-I)(17)

[0132] I max =maxpool(I min (18)

[0133] Here, I represents the prediction probability map obtained through the softmax function. The normalization process is performed to generate a smooth probability distribution; I min Local minimum features are extracted by simulating the erosion process in morphology by applying max pooling to the negative probability distribution -I; max to I min Max pooling is performed again to simulate dilation, thereby enhancing the expressive power of boundary features;

[0134] The formula for calculating the profile is as follows:

[0135]

[0136] Here, f c Describes the contour features, where D is the number of categories. and Let each represent the maximum and minimum eigenvalues ​​of the i-th category, and ReLU is a modified linear unit function that ensures the result is non-negative;

[0137] Based on the original contour enhancement decoder, an attention mechanism is introduced to optimize the original contour features f. c To enhance its boundary representation ability, the specific method is as follows: f c The input is fed into the enhancement module, which then undergoes channel expansion convolution, LeakyReLU activation, and channel recovery convolution to achieve feature weighting and refinement. The formula for the enhancement process is as follows:

[0138] f′ c =Conv(LeakyReLU(convenience(f c ))) (20)

[0139] Here, Conv represents the convolution operation (kernel size is 3, padding is 1), and LeakyReLU represents the activation function;

[0140] To further optimize and enhance the features, a residual structure was adopted, and the final enhanced feature fo was obtained through the HardSigmoid function, as shown in the following formula:

[0141] f0 = HardSigmoid((f e Mf d )+(f e Mf d )×f c ) (twenty one)

[0142] Here, M represents the merge operation, and HardSigmoid is an activation function.

[0143] The loss function includes the supervised loss L for labeled data. sup and semi-supervised loss L for unlabeled data semi ;

[0144] It can fully utilize both labeled and unlabeled data, thereby improving the model's ability to perceive boundaries and its segmentation accuracy. The loss function is defined as follows:

[0145] L = L sup +λL semi (twenty two)

[0146] in, Represents the dynamic balance weights, where t and t total These represent the current iteration number and the total number of iterations, respectively.

[0147] The workflow is as follows:

[0148] Input data and preprocessing;

[0149] Input data is medical images Where C represents the number of channels, H and W represent the height and width of the image, respectively, and the input image size is defined as [224, 224] by parameters. Data augmentation is performed during the data preprocessing stage to enhance data diversity and improve model robustness. The encoder adopts a five-layer downsampling structure, namely E1, E2, E3, E4 and E5, and extracts multi-scale features through layer-by-layer max pooling, where each feature map The number of channels is [16, 32, 64, 128, 256], and the spatial resolution decreases to H / 2 with the number of layers. i ×W / 2 i (i = 0, 1, 2, 3, 4); The deepest feature E5 (256 channels, spatial resolution reduced to H / 32×W / 32) is input into the SBEM module to enhance boundary-related information, thereby improving its representation ability. The enhanced multi-scale features are then passed to the contour enhancement decoder to achieve fine recovery of boundary information.

[0150] Semantic Boundary Enhancement Module: The Semantic Boundary Enhancement Module (SBEM) receives features from the deepest layer of the encoder. (256 channels, spatial resolution H / 32×W / 32), aiming to improve the boundary accuracy of semi-supervised medical image segmentation by enhancing boundary-related information. The module's processing flow first extracts edge features f through two independent convolution operations. e and context features f g The number of channels is reduced to 64, and then after batch normalization (BN) and ReLU activation, the channels are concatenated along the channel dimension to generate the fused feature f. concat This feature is further optimized to E′5 through convolutional operations. To capture multi-scale semantic information, the module employs Extended Spatial Pyramid Pooling (ASPP) (Chenetal., 2018, CVPR) to expand the receptive field of E′5 and extract contextual features at different scales. Then, these multi-scale features are fused along the channel dimension to generate a comprehensive multi-scale contextual representation A. concat Finally, A concat The data are weighted by the channel attention module and the spatial attention module respectively, and the final output E is generated by feature concatenation. 5,final .

[0151] Contour Enhancement Decoder: This decoder receives multi-scale features extracted by the encoder, fuses shallow details with deep semantic information through skip connections, and progressively upsamples to generate accurate prediction results. To improve the accuracy of fuzzy boundary segmentation, this module introduces a dynamic contour extraction mechanism, generates a probability distribution I, and extracts Imin and Imax through max pooling to enhance boundary features, and calculates contour features f. c Subsequently, fc f′ is generated through refined module processing. c The features are then fused using the residual structure and the merging operation M. Combined with the Hard Sigmod activation function, the final output is an enhanced feature f0, which significantly improves the boundary robustness of low-contrast regions.

[0152] In summary:

[0153] This invention proposes SBEM-UNet, an innovative semi-supervised medical image segmentation model designed to significantly enhance segmentation accuracy by improving boundary awareness and detail recovery capabilities. The model integrates a self-designed Semantic Boundary Enhancement Module (SBEM) with an introduced contour enhancement decoder, constructing a highly efficient boundary enhancement architecture. SBEM employs parallel structural modeling and multi-scale contextual flows to extract boundary information from the deepest features of the encoder: the structural modeling flow focuses on capturing local boundary details, while the multi-scale contextual flow expands the receptive field to obtain global semantic information. These two flows, combined with feature concatenation and fusion and an attention mechanism, significantly improve boundary sensitivity and semantic expressiveness. The contour enhancement decoder utilizes a dynamic contour attention mechanism, combined with learnable enhancement factors, to adaptively focus on boundary regions. Through multi-scale feature fusion and residual connections, it accurately recovers the fine details of complex anatomical structures, thereby ensuring high segmentation accuracy. SBEM-UNet, through its optimized feature processing flow, effectively improves the performance of semi-supervised medical image segmentation, providing an efficient and practical solution for clinical scenarios with limited labeled data.

[0154] See Figures 4-10 Based on the above embodiments, the following experimental scheme is proposed:

[0155] Dataset

[0156] The ACDC dataset contains 200 labeled short-axis cardiac cine MRI images from 100 patients, categorized into four classes. While two-dimensional segmentation is more common than direct three-dimensional segmentation, both methods have been explored in the literature. The dataset was divided into 70 patient scans for training, 10 for validation, and 20 for testing.

[0157] The PROMISE12 dataset was provided for the MICCAI 2012 Prostate Segmentation Challenge. This dataset contains magnetic resonance imaging (MRI) images of 50 patients with various diseases from different locations. All 3D scans were converted to 2D slices. Following previous methods, DSC and ASD were used as evaluation metrics.

[0158] Evaluation indicators

[0159] In the experiments, to comprehensively evaluate the performance of SBEM-UNet in semi-supervised medical image segmentation, this invention employed four standard evaluation metrics: Dice similarity coefficient (DSC), Jaccard index, 95% Hausdorff distance (95HD), and average surface distance (ASD). These metrics cover region overlap accuracy and boundary accuracy. DSC measures the degree of overlap between the predicted segmentation region and the ground truth annotation; the Jaccard index calculates the ratio of the intersection area to the union area of ​​the predicted and ground truth regions, both ranging from 0 to 1, with higher values ​​indicating better segmentation accuracy and consistency. 95HD evaluates the 95th percentile of the maximum distance between the predicted and ground truth boundaries; ASD calculates the average distance between the predicted and ground truth surfaces, with lower values ​​reflecting higher boundary overlap and detail recovery capabilities, respectively. These metrics collectively ensure comprehensive validation of SBEM-UNet's segmentation performance in complex scenes.

[0160] Implementation details

[0161] This invention evaluates the performance of SBEM-UNet on two baseline methods—Cross-Teaching and BCP (Based on Average Teacher Framework). All experiments were run on an NVIDIA GeForce RTX 3090 using the ACDC dataset (containing 68 slices from 3 labeled patients). To ensure fair comparison, this invention strictly adheres to the default settings of the original Cross-Teaching and BCP methods, including a random seed (1337) and partially cached data loading.

[0162] Cross-teaching baseline: Cross-teaching is a semi-supervised cross-supervised learning framework that utilizes two different architectures, UNet and Swin-UNet, to implement network perturbations. During training, input images are uniformly cropped to a resolution of 224×224 and divided into K=16 image patches (each patch is 56×56, divided by a 4×4 grid). Each batch contains 16 samples (8 labeled patches and 8 unlabeled patches). Training uses the SGD optimizer (initial learning rate 0.01, momentum 0.9, weight decay 0.0001), combined with cross-entropy loss and Dice loss as the comprehensive loss function, with a maximum of 30,000 iterations. Consistency weights are gradually increased using an sigmoid ramp function (maximum weight 0.1, ramp period 200 iterations) to enhance model robustness.

[0163] BCP Baseline: BCP is based on the average teacher framework, and training is divided into two phases: pre-training and self-training. The pre-training phase uses a single UNet model for 10,000 iterations; the self-training phase introduces two student networks (both based on UNet) with different initializations and an exponential moving average (EMA) teacher model (decay rate 0.95) to achieve network perturbation. Input images are cropped to 256×256 resolution and divided into K=16 image patches (each patch is 64×64, divided by a 4×4 grid). Each batch contains 24 samples (12 labeled slices and 12 unlabeled slices). Training uses the SGD optimizer (initial learning rate 0.01, momentum 0.9, weight decay 0.0001) combined with cosine annealing learning rate scheduling (minimum learning rate 1×10). -5 The maximum number of iterations is 30,000. The consistency weight is gradually increased through an S-shaped ramp function (maximum weight 1.0, ramp period 300 iterations), and optimized by combining cross-entropy loss and Dice loss.

[0164] Comparison with existing technologies

[0165] ACDC dataset: such as Figure 4 As shown, the BoundaryNet framework outperforms existing methods across various key evaluation metrics, particularly in direct comparisons with ABD (CVPR 2023). BCP, a newly released high-performance method, demonstrates significant potential on the ACDC dataset. Seamless integration into the BoundaryNet framework, forming BoundaryNet(BCP), further enhances segmentation performance. Specifically, under experimental conditions with a 10% annotation rate, BoundaryNet(BCP) outperforms ABD-BCP and SCP-Net in metrics such as Dice similarity coefficient (DSC), Hausdorff distance (HD), and average surface distance (ASD), showcasing its significant advancements in segmentation accuracy and boundary conditions.

[0166] PROMISE12 dataset: Based on the SS-Net method specifications, this invention only uses 20% of the labeled data. Because...

[0167] The SBEM-UNet model is compared with state-of-the-art methods such as ABD, CCT, URPC, SS-Net, SLC-Net, and SCP-Net. Figure 5 As shown, the SBEM-UNet model of this invention outperforms all other methods in terms of performance. In particular, compared with the recently proposed ABD method, the framework of this invention achieves a significant improvement in the ASD metric, with an improvement of up to 30.82%.

[0168] ablation experiment

[0169] This invention employs a cross-disciplinary teaching method as a benchmark for experiments. Unless otherwise stated, all ablation experiments were conducted based on the ACDC dataset, which includes 10% labeled data.

[0170] SBEM and CAD: To evaluate the contributions of each component of SBEM-UNet in semi-supervised medical image segmentation, this invention conducted ablation experiments to analyze the impact of the self-designed semantic boundary enhancement module (SBEM) and the introduced contour enhancement decoder (CAD) on model performance. Experiments were conducted on the ACDC dataset, validating the roles of SBEM and CAD in improving boundary awareness and detail recovery capabilities.

[0171] Experimental results show that the basic model (without SBEM and CAD) has a DSC of 88.52%, a Jaccard index of 79.97%, a 95HD of 5.06, and an ASD of 1.43. After adding SBEM, the DSC increases to 89.33%, the Jaccard index increases to 81.27%, the 95HD decreases to 4.96, and the ASD decreases to 1.04, indicating that SBEM significantly enhances boundary sensitivity and semantic expressiveness through structural modeling and multi-scale feature fusion. Further introducing CAD slightly reduces the DSC to 89.15%, the Jaccard index to 81.00%, and the 95HD to 4.85, but the ASD increases to 1.33, indicating that CAD may have a complex impact on detail recovery in certain scenarios. When SBEM and CAD are integrated into the basic model simultaneously, the DSC reaches 89.55%, the Jaccard index reaches 81.59%, the 95HD decreases to 4.25, and the ASD is 1.07, demonstrating the maximum improvement in boundary awareness and detail recovery achieved through their synergistic effect. These results validate the effectiveness and complementarity of SBEM and CAD in SBEM-UNet, providing a high-precision boundary enhancement solution for semi-supervised medical image segmentation tasks.

[0172] ABD and SBEM-UNet: To evaluate the performance of SBEM-UNet in semi-supervised medical image segmentation, this invention was compared with the ABD method in comparative experiments. Figure 9The study demonstrates the trend of DSC (Distributed Target Concentration) over training iterations: at 2000 iterations, the ABD method achieved a DSC of 80.1%, while SBEM-UNet achieved 78.3%; by 4000 iterations, the ABD method rose to 81.5%, while SBEM-UNet significantly improved to 83.3%; at 6000 iterations, the ABD method achieved 85.6%, and SBEM-UNet achieved 84.2%; at 8000 iterations, the ABD method achieved 87.8%, and SBEM-UNet was close to 87.7%; at 10000 and 12000 iterations, the ABD method stabilized at 88.2% and 88.5%, respectively, while SBEM-UNet reached 89.3% and 89.5%, respectively, indicating that it surpassed the ABD method in the later stages of training and demonstrated stronger target region capture capabilities.

[0173] Figure 10 The trend of 95HD is shown: at 2000 iterations, the ABD method has a 95HD of 8.4, while SBEM-UNet has 8.9; at 4000 iterations, the ABD method drops to 8.2, and SBEM-UNet drops to 7.6; at 6000 iterations, the ABD method has 6.3, and SBEM-UNet has 6.8; at 8000 iterations, the ABD method has 6.0, and SBEM-UNet has 6.4; after 10000 iterations, the ABD method has 5.7, and SBEM-UNet further drops to 5.5; by 12000 iterations, the ABD method stabilizes at 5.1, while SBEM-UNet reaches 4.2. This indicates that SBEM-UNet gradually surpasses the ABD method in boundary localization accuracy, especially showing a significant advantage in the later stages of training. Overall, although the ABD method has a slight advantage in DSC in the early stages of training, SBEM-UNet reverses this trend in subsequent iterations and continues to lead in the 95HD metric. Its superior performance stems from the self-designed SBEM module, which enhances semantic boundary perception through structural modeling and multi-scale feature fusion, and the introduced CAD module, which improves contour detail recovery through dynamic contour attention mechanism. The synergistic effect of the two significantly improves the boundary accuracy and robustness of SBEM-UNet in semi-supervised medical image segmentation.

[0174] Enhancement Factor: To evaluate the impact of the enhancement factor on the performance of the contour enhancement decoder in semi-supervised medical image segmentation. Experimental results show that when the enhancement factor is set to 0.5, the DSC is 89.69%, the Jaccard index is 81.47%, the 95HD is 5.31, and the ASD is 1.32. At this value, the model's sensitivity to target boundaries is improved, but its detail recovery ability is slightly insufficient. When the enhancement factor is increased to 1.0, the DSC reaches 89.55%, the Jaccard index reaches 81.59%, the 95HD decreases to 4.25, and the ASD is 1.07, indicating that the model achieves the best balance between boundary perception and detail recovery, resulting in the optimal overall segmentation performance. Further increasing the enhancement factor to 1.5 slightly reduces the DSC to 89.33%, the Jaccard index to 81.27%, the 95HD to 4.96, and the ASD to 1.04, showing that an excessively high enhancement factor may cause the model to over-focus on boundary information, affecting the segmentation accuracy of internal regions. In summary, when the enhancement factor is 1.0, SBEM-UNet enhances semantic boundary awareness through its self-designed SBEM module and optimizes contour detail recovery by combining it with the introduced CAD module, achieving the best performance in semi-supervised medical image segmentation tasks and demonstrating its robustness and efficiency in complex scenarios.

[0175] In summary:

[0176] This invention evaluates the performance of SBEM-UNet using two authoritative datasets: ACDC (heart segmentation) and PROMISE12 (prostate segmentation). Experimental results show that SBEM-UNet outperforms existing methods, such as U-Net and DeepLab, in both segmentation accuracy and boundary quality. On the ACDC dataset, SBEM-UNet improves the Dice similarity coefficient by approximately 1.2%, the Jaccard coefficient by approximately 2%, and reduces the boundary distance metrics 95% Hausdorff distance (95HD) and average surface distance (ASD) by 16% and 25%, respectively. These significant improvements validate the robustness and generalization ability of SBEM-UNet in handling complex anatomical structures and scenes with blurred boundaries. Furthermore, SBEM-UNet's structural modeling and boundary optimization methods exhibit good scalability and can be applied to other high-precision tasks (such as small object detection), laying the foundation for its application prospects in a wider range of clinical scenarios.

[0177] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A boundary enhancement system for semi-supervised medical image segmentation, wherein the boundary enhancement model is the SBEM-UNet model, characterized in that: This includes a contour enhancement decoder, a semantic boundary enhancement module, and a loss function; The semantic boundary enhancement module, also known as the SBEM module, aims to enhance the semantic boundary from the deepest features of the encoder. Extracting and enhancing boundary-related information to significantly improve the boundary accuracy of semi-supervised medical image segmentation; By integrating a contour enhancement module, the prediction accuracy of the target boundary is significantly improved. The contour enhancement decoder receives features extracted by the encoder, fuses multi-scale features through skip connections, and progressively upsamples to generate the final prediction result. ; The loss function adopts a combination of supervised and semi-supervised loss functions; The SBEM module is divided into two parallel data streams: a structural modeling stream and a multi-scale context stream, which are responsible for local boundary detail extraction and global semantic information modeling, respectively. The structural modeling stream is the top of the graph, and the multi-scale context stream is the bottom of the graph. The operation process of the SBEM module includes at least the following steps: The structural modeling flow is responsible for extracting boundary and contextual features, providing basic information for boundary enhancement; Data stream from input It begins by extracting local boundary details through edge convolutional blocks. The computation process is described as follows: (1) In this process, a standard 3×3 convolution is applied to the input feature map E5, followed by batch normalization (BN) and ReLU activation to obtain the final output feature map. It focuses on local gradient changes and emphasizes edge details, making it particularly suitable for detecting low-contrast boundaries in medical images; Meanwhile, the contextual convolutional block extracts global semantic information from the same input x and computes it as follows: (2) Among them, the context convolution block will... Apply one The dilated convolution is then processed by BN and ReLU to finally output the result. By expanding the receptive field, global semantic information can be captured, and the understanding of the context can be enhanced; Two branch features and The data will be stitched together along the channel dimension, and the calculation result is as follows: (3) Here, this step combines local boundary information with global semantic information, providing diverse feature representations for subsequent fusion and enhancing the model's understanding of boundary regions; Merge block through Convolution integrates and concatenates features, and the calculation result is: (4) Among them, the 1x1 convolution will be used to generate the target dimension. The number of channels is adjusted to the target value, and key features are extracted simultaneously. Then, batch normalization (BN) standardizes the distribution of features to enhance linearity, thereby generating more refined features. ; Next, multi-scale feature extraction will be performed, and the formula is as follows: （5） Then, using the expanded spatial pyramid pooling technique with convolutional kernels of different sizes, the following was performed: Extracting receptive field features: (6) (7) (8) in, Indicates a kind of expansion rate Dilated convolutions can capture contextual information at different scales; Furthermore, by performing pooling operations in the expanded spatial pyramid pooling, global context information can be extracted: （9） Finally, by concatenating along the channel dimension and integrating contextual information at different scales, the final multi-scale feature representation was obtained. , , , and The formula is: (10) After SEBlock processing, an enhanced feature representation is obtained. The formula is as follows: (11) Then, the feature is weighted and processed by the channel attention module and the spatial attention module respectively; The channel attention module adaptively adjusts the channel weights, while the spatial attention module highlights important spatial regions, thereby generating weighted features. and This provides a more discriminative representation for fusion, and its formula is as follows: (12) Among them, channel attention first applies to the input feature map Global average pooling is performed to obtain a global description of each channel. Then, two layers of convolution and the ReLU activation function are used to learn the non-linear relationship between channels. Finally, the sigmoid function is used to generate channel attention weights, which are then applied to the original features through element-wise multiplication to achieve the weighted importance of each channel weight. (13) Here, MaxPool and AvgPool are max pooling and average pooling, respectively, used to extract spatial features. Their outputs are concatenated and processed by a convolutional layer, then an attention map is generated using the Sigmoid function. This attention map is then processed element-wise by multiplication. Multiply to get the output. This highlights the key features; Finally, channel attention and spatial attention By fusing attention combinations, a weighted feature representation is obtained: (14) Among them, the features are weighted by channel attention. and features weighted by spatial attention By fusing elements-wise, important information in both the channel and spatial domains is emphasized simultaneously, resulting in a final weighted feature representation. .

2. The boundary enhancement system for semi-supervised medical image segmentation according to claim 1, characterized in that: The contour enhancement decoder introduces a dynamic contour extraction and feature weighting mechanism, effectively improving the segmentation effect at blurred boundaries, through the segmentation head. Generate prediction probability map The formula is as follows: (15) in It comes from the previous layer decoder. Features; Contour-enhanced decoders can extract contour features more accurately by simulating the erosion and dilation processes in morphological operations. The formula for its use is as follows: (16) (17) (18) Here, Indicates passage Function Pair Prediction Probability Map The process of normalization is performed to generate a smooth probability distribution; By analyzing the negative probability distribution Max pooling is applied to simulate the erosion process in morphology, thereby extracting local minimum features; right Max pooling is performed again to simulate dilation, thereby enhancing the expressive power of boundary features; The formula for calculating the profile is as follows: （19） Here, Indicates contour features, For the number of categories, and Let each represent the maximum and minimum eigenvalues ​​of the i-th category, and ReLU is a modified linear unit function that ensures the result is non-negative; Based on the original contour enhancement decoder, an attention mechanism is introduced to optimize the original contour features. To enhance its boundary representation ability, the specific methods are as follows: The input is fed into the enhancement module, which then undergoes channel expansion convolution, LeakyReLU activation, and channel recovery convolution to achieve feature weighting and refinement. The formula for the enhancement process is as follows: (20) Here, Conv represents the convolution operation, and LeakyReLU represents the activation function; To further optimize and enhance the features, a residual structure was adopted, and the final enhanced feature fo was obtained through the HardSigmoid function, as shown in the following formula: (21) Here, M represents the merge operation, and HardSigmoid is an activation function.

3. The boundary enhancement system for semi-supervised medical image segmentation according to claim 2, characterized in that: The loss function includes a supervised loss L for labeled data. sup and semi-supervised loss L for unlabeled data semi ; It can fully utilize both labeled and unlabeled data, thereby improving the model's ability to perceive boundaries and its segmentation accuracy. The loss function is defined as follows: (22) in, Represents the dynamic balancing weights, where and These represent the current iteration number and the total number of iterations, respectively.

Images