A medical image segmentation method based on boundary enhancement and double decoder
The medical image segmentation method using boundary enhancement and a dual-decoder architecture solves the problems of inaccurate contour localization and boundary leakage in existing technologies, achieving more accurate structural segmentation and feature representation, and improving the overall performance of medical image segmentation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANDONG AGRI & ENG UNIV
- Filing Date
- 2026-04-29
- Publication Date
- 2026-05-29
AI Technical Summary
Existing deep learning-based medical image segmentation methods are prone to problems such as inaccurate contour localization, boundary leakage, and structural discontinuity when dealing with adjacent structures with huge anatomical variations and blurred, low-contrast boundaries. Furthermore, they lack explicit boundary modeling and effective integration of semantic and boundary features.
We adopt a network architecture based on boundary enhancement and dual decoders. We extract multi-scale features through a shared encoder, capture contextual dependencies by combining an attention convolution module, generate a boundary probability map through the boundary decoder and introduce it into the segmentation decoder through skip connections, and use a hybrid loss function for supervision to achieve collaborative modeling of boundary and region features.
It significantly improves the structural consistency and contour localization accuracy of medical image segmentation, effectively suppresses boundary blurring and leakage, enhances multi-scale feature representation capabilities, and achieves segmentation accuracy superior to mainstream algorithms.
Smart Images
Figure CN122115484A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision and medical image processing technology, specifically to a medical image segmentation method based on boundary enhancement and dual decoders. Background Technology
[0002] Accurate segmentation of anatomical structures in medical images plays a crucial role in computer-aided diagnosis, treatment planning, and quantitative clinical assessment. In recent years, the rapid development of deep learning has significantly advanced research in medical image segmentation. For example, convolutional neural networks (CNNs), represented by 3D U-Net, have demonstrated powerful local representation capabilities through hierarchical multi-scale feature extraction; while Transformer-based models utilize self-attention mechanisms to capture long-range contextual dependencies, thereby enhancing global feature representation. To further improve structural delineation, some studies have introduced boundary-aware learning strategies (such as auxiliary boundary branches and boundary constraint loss functions) to provide explicit structural supervision.
[0003] However, due to significant differences in anatomical structures and the often blurred or low-contrast boundaries between adjacent structures, accurate contour delineation remains a major challenge. Inaccurate boundary localization not only leads to structural discontinuities in the segmentation results but may also affect the reliability of subsequent clinical analysis. Existing mainstream techniques mainly suffer from the following drawbacks:
[0004] 1. Lack of explicit boundary modeling: Most existing deep learning methods emphasize region semantic learning, but lack explicit boundary modeling in complex regions, resulting in inaccurate contour localization.
[0005] 2. Weak coupling between semantic and boundary features: Existing segmentation models have limitations in effectively integrating regional semantic information and guiding boundary structural information. Regional semantic features and boundary structural features are usually weakly coupled, making it difficult to produce segmentation results that are structurally consistent and detailed.
[0006] 3. Prone to boundary leakage: Under conditions of blurred or low-contrast boundaries, the above limitations become more obvious, and the model often experiences boundary leakage and structural discontinuities.
[0007] 4. Insufficient utilization of boundary information: Many existing boundary-aware methods treat boundary learning as an auxiliary supervision signal and fail to integrate boundary modeling depth into the feature refinement process, resulting in limited effectiveness in improving structural consistency and contour localization accuracy.
[0008] Therefore, a boundary enhancement network (BENet) based on a dual-decoder architecture is needed for medical image segmentation. This aims to address the technical problems in existing deep learning-based medical image segmentation methods that emphasize region semantic learning but lack explicit boundary modeling. These problems lead to inaccurate contour localization, boundary leakage, and structural discontinuities when dealing with adjacent structures exhibiting significant anatomical variations and blurred, low-contrast boundaries. Summary of the Invention
[0009] To address the problems existing in the prior art, the present invention aims to provide a medical image segmentation method based on boundary enhancement and dual decoders. By effectively fusing boundary structural information with regional semantic features, the method improves the contour localization accuracy and structural consistency of complex regions in medical images.
[0010] The technical solution adopted by this invention to solve its technical problem is: a medical image segmentation method based on boundary enhancement and dual decoders, comprising the following steps:
[0011] S1. Construct and initialize the boundary augmentation network model: The network model adopts a dual-decoder architecture, including a shared encoder, a segmentation decoder, and a boundary decoder.
[0012] S2. Define a hybrid loss function for model training: In order to jointly optimize segmentation and boundary prediction, a hybrid loss function is used to supervise the two decoding branches;
[0013] S3. Model Training and Parameter Configuration: Acquiring 3D Medical Images;
[0014] S4. Medical Image Prediction and Segmentation Output: Input the 3D medical image to be segmented into the trained boundary enhancement network model, and output the final 3D medical image segmentation result.
[0015] Specifically, the construction of the shared encoder in step S1 is used to extract hierarchical feature representations from the input image and provide contextual information for the two decoding branches. The shared encoder consists of convolutional layers and downsampling operations.
[0016] The shared encoder embeds an attention convolution module and adopts a two-branch feature extraction structure: given the input feature map, one branch is processed by the attention convolution module to capture long-range contextual dependencies; the other branch applies a 3×3×3 three-dimensional convolutional neural network to preserve local structural information; the outputs of the two branches are fused by element-wise summation.
[0017] Specifically, in step S1, the boundary decoder is constructed by receiving multi-scale features from the shared encoder and gradually refining the boundary representation through upsampling and convolution operations to generate a boundary probability map.
[0018] Specifically, the segmentation decoder in step S1 is constructed for feature-guided fusion. The segmentation decoder receives multi-scale features from the shared encoder to gradually restore spatial resolution. At the same time, the boundary representation generated by the boundary decoder is introduced into the segmentation decoder through feature guidance and skip connections.
[0019] Specifically, the steps for training the model using the hybrid loss function in step S2 include:
[0020] S21. Basic Loss Combination: For each prediction scale, the loss function is composed of Dice loss and cross-entropy loss, providing complementary supervision signals for region consistency and voxel-level classification.
[0021] S22, Multi-scale depth supervision: Both the boundary decoder and the segmentation decoder are supervised using the same form of loss function;
[0022] S23. Total Loss Calculation: The overall loss function includes segmentation loss and boundary detection loss, and assigns greater weight to the high-resolution output through weighting coefficients.
[0023] Specifically, the segmentation loss is defined as:
[0024] ;
[0025] In the formula: L Dice Indicates Dice loss; L CE p represents the cross-entropy loss; i This represents the predicted probability map at the i-th scale; y i Represents the corresponding true label for the i-th scale;
[0026] The boundary decoder uses the same loss formula for supervision, and employs a deep supervision strategy across multiple scales. The overall loss function is defined as:
[0027] ;
[0028] In the formula: w is the weighting coefficient; Let represent the segmentation loss at the i-th scale; Let represent the boundary detection loss at the i-th scale.
[0029] Specifically, the model training and parameter configuration steps in step S3 include:
[0030] S31. Data preparation: Obtain a medical CT image dataset, divide it into training and test sets in a 9:1 ratio, and use a five-fold cross-validation mechanism;
[0031] S32. Environment and Hyperparameter Settings: A boundary augmentation network model was implemented on an NVIDIA GPU based on the PyTorch deep learning framework. The Adam optimizer was used to update the network parameters, and the initial learning rate was set to 0.0003.
[0032] S33. Sliding window sampling: During the training and testing phases, a sliding window strategy is used to crop the 3D volume data, and the size of the 3D medical image block is set to 48×128×128.
[0033] Specifically, in step S4, the three-dimensional medical image is sequentially processed by a shared encoder to extract multi-scale features, a boundary decoder to generate a boundary probability map, and a segmentation decoder, guided by the boundary probability map and combined with multi-scale features, to output the final three-dimensional medical image segmentation result map.
[0034] The present invention has the following beneficial effects:
[0035] The medical image segmentation method based on boundary enhancement and dual decoders designed in this invention significantly improves structural consistency and contour localization accuracy: by proposing a boundary-guided dual decoder architecture, this invention incorporates boundary modeling as part of the feature refinement process, rather than merely as an auxiliary supervision signal, effectively strengthening the collaborative modeling between boundaries and segmentation features, and achieving more accurate contour localization in highly challenging regions.
[0036] The medical image segmentation method based on boundary enhancement and dual decoders designed in this invention effectively suppresses boundary blurring and boundary leakage: the boundary representation is integrated into the segmentation decoder through skip connections, enabling the network to better distinguish adjacent anatomical structures and significantly reducing boundary leakage in areas with blurred organ boundaries or low contrast.
[0037] The medical image segmentation method based on boundary enhancement and dual decoders designed in this invention enhances the comprehensive expressive power of contextual features and local features: by innovatively introducing an attention-based convolution module in the encoder, joint modeling of contextual relationships and local structural patterns is achieved, thereby improving the quality of multi-scale feature representation.
[0038] The medical image segmentation method based on boundary enhancement and dual decoders designed in this invention has a segmentation accuracy superior to mainstream algorithms: On publicly available medical image datasets (such as FLARE21 and LiTS17), the method of this invention (BENet) achieves the best overall performance on multiple evaluation metrics such as the Dice coefficient, demonstrating its outstanding effectiveness in improving segmentation accuracy and boundary delineation. Attached Figure Description
[0039] Figure 1 This is a network architecture diagram of a medical image segmentation method based on boundary enhancement and dual decoders.
[0040] Figure 2 This is a flowchart of the attention convolution module. Detailed Implementation
[0041] The technical solutions of the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0042] like Figures 1-2 As shown, a boundary enhancement network (BENet) based on a dual-decoder architecture is used for medical image segmentation, mainly for processing three-dimensional medical volume data (such as CT images). The specific implementation steps of this method are as follows:
[0043] Step 1: Construct and initialize the Boundary Reinforcement Network (BENet) model
[0044] The network adopts a dual-decoder architecture, which includes a shared encoder, a segment decoder and a boundary decoder.
[0045] 1. Construct a shared encoder: This encoder extracts hierarchical feature representations from the input image and provides multi-scale contextual information for the two decoding branches. The encoder consists of a series of convolutional layers and downsampling operations to progressively expand the receptive field.
[0046] 2. Embedded Attention Convolutional Module (ACM): An ACM module is embedded in the shared encoder to enhance feature representation capabilities. Specifically, a two-branch feature extraction structure is adopted: given the input feature map, one branch processes it through ACM to capture long-range contextual dependencies; the other branch applies a standard 3×3×3 three-dimensional convolution to preserve local structural information. The outputs of the two branches are fused by element-wise summation. This joint modeling of contextual relationships and local structural patterns effectively enhances the shared encoder's ability to capture boundary-sensitive and semantically discriminative features.
[0047] 3. Constructing a boundary decoder: The boundary decoder receives multi-scale features from the shared encoder and gradually refines the boundary representation through a series of upsampling and convolution operations to generate boundary probability maps. These generated boundary maps not only enhance contour perception but also provide structural guidance for subsequent segmentation processes.
[0048] To further enhance the interaction between boundary features and segmentation features, the generated boundary representations are integrated into the segmentation decoder through feature guidance and skip connections. This boundary guidance mechanism enables the network to better distinguish adjacent anatomical structures, suppress boundary ambiguity, and reduce missed detections in blurred or low-contrast regions. Unlike methods that use boundary learning as an auxiliary supervision signal, this design integrates boundary modeling into the feature optimization process, effectively improving the structural consistency and contour localization accuracy of the segmentation results.
[0049] 4. Construct a segmentation decoder and implement feature-guided fusion: The segmentation decoder receives multi-scale features from the shared encoder to progressively restore spatial resolution. Simultaneously, the boundary representation generated by the boundary decoder is introduced into the segmentation decoder through feature guidance and skip connections to achieve boundary-aware feature refinement and accurate contour localization.
[0050] The role of the segmentation decoder is to progressively recover the spatial resolution from the encoded features and generate the final segmentation map. The decoder receives multi-scale features from the shared encoder and fuses boundary-aware information provided by the boundary decoder to optimize the segmentation result.
[0051] Step 2: Define the mixture loss function for model training
[0052] To jointly optimize segmentation and boundary prediction, a hybrid loss function is used to supervise the two decoding branches.
[0053] 1. Basic Loss Combination: For each prediction scale, the loss function is composed of a combination of Dice loss and cross-entropy (CE) loss, providing complementary supervision signals for region consistency and voxel-level classification.
[0054] 2. Multi-scale deep supervision: Both the boundary decoder and the segmentation decoder are supervised using the same form of loss function. Deep supervision strategies are applied at multiple scales to improve feature learning and stabilize the training process.
[0055] 3. Total Loss Calculation: The overall loss function includes segmentation loss and boundary detection loss, and assigns greater weight to the high-resolution output through weighting coefficients.
[0056] To jointly optimize segmentation and boundary prediction, this paper employs a hybrid loss function to supervise the two decoding branches. At each prediction scale, the loss function is a combination of Dice loss and cross-entropy (CE) loss, providing complementary supervision for region consistency and voxel-level classification. Let p be the predicted probability map at the i-th scale. i The true label corresponding to the i-th scale is y. i The segmentation loss is then defined as:
[0057] ;
[0058] In the formula: L Dice Indicates Dice loss; L CE This represents the cross-entropy loss.
[0059] The boundary decoder uses the same loss formula for supervision to ensure the accuracy of contour prediction. To improve feature learning performance and stabilize the training process, a deep supervision strategy is employed at multiple scales. The overall loss function is defined as:
[0060] ;
[0061] In the formula: w is the weighting coefficient; Let represent the segmentation loss at the i-th scale; Let represent the boundary detection loss at the i-th scale.
[0062] Step 3: Model Training and Parameter Configuration
[0063] 1. Data preparation: Obtain a medical CT image dataset, divide it into training and test sets in a 9:1 ratio, and use a five-fold cross-validation mechanism to improve the robustness of the model.
[0064] 2. Environment and Hyperparameter Settings: The model of this invention was implemented on an NVIDIA GPU based on the PyTorch deep learning framework. The Adam optimizer was used to update network parameters, and the initial learning rate was set to 0.0003.
[0065] 3. Sliding window sampling: During the training and inference (testing) phases, a sliding window strategy is used to crop the 3D volume data, and the size of the 3D image patch is set to 48×128×128.
[0066] Step 4: Medical Image Prediction and Segmentation Output
[0067] 1. Input the 3D medical image to be segmented into the trained BENet model.
[0068] 2. The image is sequentially processed by a shared encoder to extract multi-scale features, and a boundary decoder generates a boundary probability map.
[0069] 3. Guided by the boundary probability map and combined with multi-scale features, the segmentation decoder outputs the final high-precision three-dimensional medical image segmentation result.
[0070] Implementation plan effectiveness verification:
[0071] To verify the effectiveness of this solution, the BENet method of this invention was compared with typical three-dimensional convolutional neural networks (such as SegNet, ResUNet, U-Net++, and Attention U-Net).
[0072] Experiments were conducted on two public CT datasets, FLARE21 and LiTS17. The FLARE2 dataset focuses on abdominal multi-organ segmentation and contains 361 training samples, 50 validation samples, and 100 test samples. The LiTS17 dataset is a liver and tumor segmentation dataset collected from multiple medical centers and contains 131 training samples and 70 test samples. To ensure fairness in the comparison, each dataset was split into training and test sets in a 9:1 ratio, and five-fold cross-validation was used to improve the robustness of the experiments.
[0073] All experiments were implemented on an NVIDIA GPU using the PyTorch framework, with the Adam optimizer and an initial learning rate of 0.0003. A sliding window strategy was used for both training and inference, with a window size of 48×128×128.
[0074] To verify the effectiveness of the proposed method, BENet was compared with several representative segmentation models, including SegNet, ResUNet, U-Net++, and Attention U-Net, on two datasets. The quantitative experimental results are shown in Table 1. BENet achieved the best overall performance on most evaluation metrics, demonstrating its effectiveness in improving segmentation accuracy and boundary rendering.
[0075] Table 1 Performance evaluation of various segmentation methods
[0076]
[0077] As shown in Table 1, although traditional convolutional base models such as SegNet and ResUNet can achieve certain segmentation performance, they have limitations in capturing complex anatomical structures; methods such as U-Net++ and Attention U-Net, which have enhanced feature fusion capabilities, have significantly improved segmentation performance due to their stronger multi-scale representation capabilities.
[0078] Compared to the methods described above, the proposed BENet achieves optimal overall performance on most evaluation metrics across both datasets. This performance improvement is attributed to the integration of boundary-aware learning and region feature modeling within a unified framework: the explicit contour supervision provided by the boundary decoder enables the network to better distinguish adjacent anatomical structures and reduce segmentation errors in blurred boundary regions; simultaneously, the attention-based encoder and deep supervision strategy further enhance the quality of feature representation and stabilize the model's optimization process.
[0079] Experimental results show that, by integrating boundary-aware learning and region feature modeling within a unified framework, this invention achieves optimal overall performance in core evaluation metrics such as the Dice coefficient when dealing with complex anatomical structures. In particular, it effectively reduces boundary leakage errors and significantly improves structural consistency in regions with complex organ morphology and blurred boundaries.
[0080] This invention is not limited to the above-described embodiments. Anyone should know that any structural changes made under the guidance of this invention, and any technical solutions that are the same as or similar to this invention, fall within the protection scope of this invention.
[0081] The technologies, shapes, and structures not described in detail in this invention are all known technologies.
Claims
1. A medical image segmentation method based on boundary enhancement and dual decoders, characterized in that, Includes the following steps: S1. Construct and initialize the boundary augmentation network model: The network model adopts a dual-decoder architecture, including a shared encoder, a segmentation decoder, and a boundary decoder. S2. Define a hybrid loss function for model training: In order to jointly optimize segmentation and boundary prediction, a hybrid loss function is used to supervise the two decoding branches; S3. Model Training and Parameter Configuration: Acquiring 3D Medical Images; S4. Medical Image Prediction and Segmentation Output: Input the 3D medical image to be segmented into the trained boundary enhancement network model, and output the final 3D medical image segmentation result.
2. The medical image segmentation method based on boundary enhancement and dual decoders according to claim 1, characterized in that, The construction of the shared encoder in step S1 is used to extract hierarchical feature representations from the input image and provide context information for the two decoding branches. The shared encoder consists of convolutional layers and downsampling operations. The shared encoder embeds an attention convolution module and adopts a two-branch feature extraction structure: given the input feature map, one branch is processed by the attention convolution module to capture long-range contextual dependencies; the other branch applies a 3×3×3 three-dimensional convolutional neural network to preserve local structural information; the outputs of the two branches are fused by element-wise summation.
3. The medical image segmentation method based on boundary enhancement and dual decoders according to claim 1, characterized in that, The boundary decoder in step S1 receives multi-scale features from the shared encoder, and gradually refines the boundary representation through upsampling and convolution operations to generate a boundary probability map.
4. The medical image segmentation method based on boundary enhancement and dual decoders according to claim 1, characterized in that, The segment decoder in step S1 is constructed for feature-guided fusion. The segment decoder receives multi-scale features from the shared encoder to gradually restore spatial resolution. At the same time, the boundary representation generated by the boundary decoder is introduced into the segment decoder through feature guidance and skip connections.
5. The medical image segmentation method based on boundary enhancement and dual decoders according to claim 1, characterized in that, The specific steps for training the model using the hybrid loss function in step S2 include: S21. Basic Loss Combination: For each prediction scale, the loss function is composed of Dice loss and cross-entropy loss, providing complementary supervision signals for region consistency and voxel-level classification. S22, Multi-scale depth supervision: Both the boundary decoder and the segmentation decoder are supervised using the same form of loss function; S23. Total Loss Calculation: The overall loss function includes segmentation loss and boundary detection loss, and assigns greater weight to the high-resolution output through weighting coefficients.
6. The medical image segmentation method based on boundary enhancement and dual decoders according to claim 5, characterized in that, The segmentation loss is defined as: ； In the formula: L Dice Indicates Dice loss; L CE p represents the cross-entropy loss; i This represents the predicted probability map at the i-th scale; y i Represents the corresponding true label for the i-th scale; The boundary decoder uses the same loss formula for supervision, and employs a deep supervision strategy across multiple scales. The overall loss function is defined as: ； In the formula: w is the weighting coefficient; Let represent the segmentation loss at the i-th scale; Let represent the boundary detection loss at the i-th scale.
7. The medical image segmentation method based on boundary enhancement and dual decoders according to claim 1, characterized in that, The specific steps for model training and parameter configuration in step S3 include: S31. Data preparation: Obtain a medical CT image dataset, divide it into training and test sets in a 9:1 ratio, and use a five-fold cross-validation mechanism; S32. Environment and Hyperparameter Settings: A boundary augmentation network model was implemented on an NVIDIA GPU based on the PyTorch deep learning framework. The Adam optimizer was used to update the network parameters, and the initial learning rate was set to 0.0003. S33. Sliding window sampling: During the training and testing phases, a sliding window strategy is used to crop the 3D volume data, and the size of the 3D medical image block is set to 48×128×128.
8. The medical image segmentation method based on boundary enhancement and dual decoders according to claim 1, characterized in that, In step S4, the three-dimensional medical image is sequentially processed by a shared encoder to extract multi-scale features, a boundary decoder to generate a boundary probability map, and a segmentation decoder, guided by the boundary probability map and combined with multi-scale features, to output the final three-dimensional medical image segmentation result map.