A gated CT image segmentation method and system based on prior guidance and morphology perception
By introducing a priori guided attention module and a morphological feature extraction module, combined with an instance-level gating module, the problem of segmenting small targets and irregular boundaries in CT image segmentation is solved, achieving more accurate and robust segmentation results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU UNIV OF SCI & TECH
- Filing Date
- 2026-03-16
- Publication Date
- 2026-06-19
Smart Images
Figure CN122244065A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image processing and relates to CT image segmentation technology, specifically to a gated CT image segmentation method and system based on prior guidance and morphological perception. Background Technology
[0002] Accurate segmentation of the core target area is an important prerequisite for image analysis and subsequent processing. However, target CT images are often accompanied by complex background interference and similar gray levels in the regions, resulting in strong background noise and low contrast. At the same time, the targets are often small in size and have extremely irregular shapes, which poses a severe challenge to the segmentation model in terms of feature extraction and boundary localization.
[0003] In recent years, deep learning segmentation methods have continued to develop and made progress in specific images. However, they still have significant limitations in scenarios involving core targets, such as "small targets + complex backgrounds + irregular boundaries." On the one hand, although the classic U-Net structure achieves multi-scale fusion through encoder-decoder and skip connections, its skip connections lack an effective filtering mechanism, easily transmitting a large amount of background noise from the encoder end directly to the decoder end, resulting in missegmentation and spurious responses. On the other hand, attention-based improved structures (such as Attention U-Net) tend to respond to large areas during weight generation, leading to insufficient attention to low-contrast and small targets, resulting in missed detections. Meanwhile, residual structures (such as ResU-Net) rely heavily on square convolution kernels with fixed geometric shapes for convolution modeling, making it difficult to fully characterize the highly irregular edge shapes of targets, easily leading to problems such as rough boundaries, undersegmentation, or oversegmentation.
[0004] Therefore, there is an urgent need for an image target segmentation technology that can enhance the representation of small targets in complex backgrounds, suppress cross-scale noise transmission, and improve the ability to depict irregular boundaries. Summary of the Invention
[0005] Purpose of the invention: To address the challenges of complex tissue background noise and the difficulties in feature extraction, skip connection noise propagation, and poor boundary characterization caused by small and highly irregular targets in CT images, this invention provides a gated CT image segmentation method and system based on prior guidance and morphology awareness. It proposes the PMG-Unet segmentation framework, introduces a prior-guided attention module to enhance the representation of small targets, constructs a morphological feature extraction module to enhance the characterization of irregular edges, and designs an instance-level gating module to suppress the propagation of background noise in skip connections, thereby improving the segmentation accuracy and robustness of core target regions.
[0006] Technical Solution: To achieve the above objectives, this invention provides a gated CT image segmentation method based on prior guidance and morphology awareness, comprising the following steps:
[0007] S1: Acquire the target CT image to be segmented;
[0008] S2: Preprocess the target CT image;
[0009] S3: Construct and train a CT image segmentation network. The CT image segmentation network is a U-shaped encoder-decoder structure. In the encoding and decoding stages, the convolutional unit PG-Conv Block with embedded prior guided attention module PGA is used for feature extraction. The morphological feature extraction module MFE is integrated in the bottleneck layer of the network, and the instance-level gating module IGM is deployed on the skip connection path to achieve enhancement of small targets, characterization of irregular boundaries and suppression of background noise in the entire process of feature extraction, deep convergence and cross-scale transmission.
[0010] S4: Input the preprocessed image data into the trained CT image segmentation network. After inference by the CT image segmentation network, the segmentation result of the target CT image is output.
[0011] Furthermore, the operation of the embedded prior-guided attention module PGA in step S3 includes:
[0012] Input features As input, a small target prior map is generated through convolution operations. And satisfy the following formula:
[0013]
[0014] in, For convolution operations, Use the Sigmoid activation function;
[0015] Based on prior diagrams The probability distribution in the spatial dimension Importance sampling is performed on the input features to obtain a global context descriptor through weighted aggregation. And satisfy the following formula:
[0016]
[0017] in, Representational space prior graph In coordinates scalar value at that point Used to maintain numerical stability;
[0018] The PGA module utilizes a multilayer perceptron. For global context descriptors Generate channel weights and apply them to the original features. Perform dynamic recalibration to enhance output features And satisfy the following formula:
[0019]
[0020] in, This indicates channel-wise multiplication, and introduces residual connections to enhance the output features. .
[0021] Furthermore, in step S3, the instance-level gating module (IGM) is deployed at the jump connection to control the input features. The IGM module performs cascaded optimization processing to suppress noise layer by layer from three dimensions: channel, space, and sample. The IGM module includes a semantic attention module, a positional attention module, and an instance-level gating unit.
[0022] Furthermore, the operation of the semantic attention module includes:
[0023] The semantic attention module computes input features in parallel. The global average pooling (Avg) and max pooling (Max) statistics are obtained by passing them through a multilayer perceptron. After fusion, a channel weight vector is generated, and the input features are recalibrated according to the channel dimensions to obtain the features. And satisfy the following formula:
[0024] .
[0025] Furthermore, the operation of the position attention module includes:
[0026] The location attention module generates a spatial attention map using the mean and extrema information of the channel dimension, and performs convolution operations. Focus on the target region and output spatially enhanced features. And satisfy the following formula:
[0027]
[0028] in, These represent the average pooling and max pooling operations performed along the channel dimension, respectively. This indicates splicing along the channel dimension.
[0029] Furthermore, the operation of the instance-level gating unit includes:
[0030] Instance-level gating units utilize global average pooling (GAP) to enhance the spatial features. It is compressed into a one-dimensional descriptor and then utilized with a scoring network. Dynamic calculation of scalar gating factor And satisfy the following formula:
[0031]
[0032] Instance-level gating units utilize scalar gating factors The feature stream is weighted overall to output the gated features. And satisfy the following formula:
[0033]
[0034] in, It is used to reflect the confidence level that the current sample contains valid target information, thereby effectively preserving high-value sample information in skip connections and blocking noise propagation in pure background slices.
[0035] Furthermore, in step S3, the morphological feature extraction module (MFE) is deployed in the network bottleneck layer to process the input features. Perform structured processing using "split-transform-recombination": First, split the input features along the channel dimension into... Subset Each subset sequentially enters an asymmetric multi-scale convolutional AM-Conv unit, and the AM-Conv unit uses a cascaded residual structure to superimpose the output of the previous scale onto the current input to generate multi-scale features. ;
[0036] Standard convolution, asymmetric convolution, and dilated convolution are applied at different scales to generate multi-scale features with rich receptive fields. And satisfy the following formula:
[0037]
[0038] in, Indicates the first Scale-based convolution operations, This indicates the number of multi-scale groups.
[0039] Furthermore, the morphological feature extraction module (MFE) includes a structural recalibration unit (SR-Block), which performs multi-scale feature extraction. Morphological enhancement is performed by parallel computation of global average pooling, local structural pooling (Avg), and global max pooling (Max), which are then fused to generate morphological weights. And satisfy the following formula:
[0040]
[0041] SR-Block utilizes shape weights Multiscale features Dynamic recalibration is performed to obtain And satisfy the following formula:
[0042]
[0043] All recalibrated features are compared with the finest-grained detail features generated by AM-Conv. The pieces are then joined together via a fusion layer. After dimensionality reduction, the final output is obtained by adding the residual to the original input. And satisfy the following formula:
[0044] .
[0045] This invention also provides a gated CT image segmentation system based on prior guidance and morphological awareness, comprising:
[0046] Acquisition unit, used to acquire CT images;
[0047] The preprocessing unit is used to preprocess CT images to form network input;
[0048] Segmentation network building blocks are used to construct CT image segmentation networks.
[0049] The segmentation unit is used to call the CT image segmentation network to perform inference on the network input and output the segmentation result.
[0050] Furthermore, the CT image segmentation network includes a PG-Conv Block convolutional unit with an embedded prior-guided attention module PGA in the encoding and decoding stages, a morphological feature extraction module MFE integrated in the network bottleneck layer, and an instance-level gating module IGM deployed on the skip connection path.
[0051] Beneficial Effects: Compared with existing technologies, this invention, based on a U-shaped encoder-decoder segmentation network, makes targeted improvements to address three core issues: easy loss of small targets, irregular target boundaries, and noise propagation through skip connections, forming the overall framework of PMG-Unet. The input raw CT slices are first truncated and normalized in terms of window width and level, and adjacent slices are stacked to supplement 3D contextual information. After unifying the size, they are fed into the network to complete end-to-end segmentation. The network backbone downsamples at each level along the encoding end to extract multi-scale semantic features, while the decoding end upsamples at each level to restore spatial details, and high- and low-level features are fused through skip connections. To enhance the representation of small targets, a priori-guided attention mechanism is introduced into the convolutional units of each encoding and decoding stage: a priori response to the target is generated through lightweight branches, and features are weighted and aggregated based on this prior to obtain a global context description. Then, the channel features are dynamically recalibrated, enabling the model to focus more on suspected target regions in complex backgrounds. To suppress background interference caused by skip connections, an instance-level gating module is set up on the cross-scale transmission path. First, channel semantic attention is used to highlight high-level semantics related to the target, then spatial location attention is used to strengthen the response of key regions, and finally, sample-level gating weights are calculated to adaptively filter the entire skip connection feature stream, thereby reducing the misleading effect of pure background slices or noisy regions on the decoding end. To improve the ability to characterize irregular boundaries, a morphological feature extraction module is integrated into the network bottleneck layer. A strategy of "splitting-multi-scale transformation-recombination and fusion" is adopted. Asymmetric multi-scale convolution is used to capture morphological cues of different directions and scales, and a structural recalibration mechanism is combined to enhance the recognition of boundaries and core regions. Through the synergistic effect of the above three types of modules in feature extraction, deep convergence, and cross-scale fusion, more accurate and robust segmentation output of core and target regions is achieved. Attached Figure Description
[0052] Figure 1 This is a flowchart of the method of the present invention.
[0053] Figure 2 This is a diagram of the overall network framework of PMG-Unet;
[0054] Figure 3 This is a schematic diagram of the PGA prior guidance module structure;
[0055] Figure 4 This is a schematic diagram of the IGM instance-level gating module;
[0056] Figure 5 This is a schematic diagram of the MFE morphological feature extraction module;
[0057] Figure 6 This is a flowchart of the model's training and inference process;
[0058] Figure 7 This is a schematic diagram of the segmentation effect. Detailed Implementation
[0059] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.
[0060] Example 1:
[0061] like Figure 1 As shown, this embodiment provides a gated CT image segmentation method based on prior guidance and morphology awareness, including the following steps:
[0062] S1: Acquire the target CT image to be segmented;
[0063] S2: Preprocess the target CT image;
[0064] S3: Construct and train a CT image segmentation network, such as... Figure 2 As shown, the CT image segmentation network is a U-shaped encoder-decoder structure. In the encoding and decoding stages, the convolutional unit PG-ConvBlock with embedded prior guided attention module PGA is used for feature extraction. The morphological feature extraction module MFE is integrated in the bottleneck layer of the network, and the instance-level gating module IGM is deployed on the skip connection path to achieve enhancement of small targets, characterization of irregular boundaries and suppression of background noise in the entire process of feature extraction, deep convergence and cross-scale transmission.
[0065] like Figure 3 As shown, the operation of the embedded prior-guided attention module PGA includes: to enhance the expressive ability of small targets, a prior-guided attention mechanism is introduced into the convolutional units of each encoding and decoding stage: a prior response of the target is generated through a lightweight branch, and the features are weighted and aggregated according to the prior. After obtaining the global context description, the channel features are dynamically recalibrated, so that the model can focus more on the suspected target region in complex backgrounds.
[0066] Input features As input, a small target prior map is generated through convolution operations. And satisfy the following formula:
[0067]
[0068] in, For convolution operations, Use the Sigmoid activation function;
[0069] Based on prior diagrams The probability distribution in the spatial dimension Importance sampling is performed on the input features to obtain a global context descriptor through weighted aggregation. And satisfy the following formula:
[0070]
[0071] in, Representational space prior graph In coordinates scalar value at that point Used to maintain numerical stability;
[0072] The PGA module utilizes a multilayer perceptron. For global context descriptors Generate channel weights and apply them to the original features. Perform dynamic recalibration to enhance output features And satisfy the following formula:
[0073]
[0074] in, This indicates channel-wise multiplication, and introduces residual connections to enhance the output features. .
[0075] like Figure 4 As shown, in order to suppress background interference caused by skip connections, an instance-level gating module (IGM) is set on the cross-scale transmission path. The operation mode of the instance-level gating module (IGM) includes: first, highlighting the high-level semantics related to the target through channel semantic attention, then strengthening the response of key regions through spatial location attention, and finally calculating the sample-level gating weights to adaptively filter the entire skip connection feature stream, thereby reducing the misleading effect of pure background slices or noisy regions on the decoding end.
[0076] Instance-level gating (IGM) modules are deployed at skip connections to control input features. The IGM module performs cascaded optimization processing to suppress noise layer by layer from three dimensions: channel, space, and sample. The IGM module includes a semantic attention module, a positional attention module, and an instance-level gating unit.
[0077] The operation of the semantic attention module includes:
[0078] The semantic attention module computes input features in parallel. The global average pooling (Avg) and max pooling (Max) statistics are obtained by passing them through a multilayer perceptron. After fusion, a channel weight vector is generated, and the input features are recalibrated according to the channel dimensions to obtain the features. And satisfy the following formula:
[0079] .
[0080] The operation of the location attention module includes:
[0081] The location attention module generates a spatial attention map using the mean and extrema information of the channel dimension, and performs convolution operations. Focus on the target region and output spatially enhanced features. And satisfy the following formula:
[0082]
[0083] in, These represent the average pooling and max pooling operations performed along the channel dimension, respectively. This indicates splicing along the channel dimension.
[0084] The operation of an instance-level gating unit includes:
[0085] Instance-level gating units utilize global average pooling (GAP) to enhance the spatial features. It is compressed into a one-dimensional descriptor and then utilized with a scoring network. Dynamic calculation of scalar gating factor And satisfy the following formula:
[0086]
[0087] Instance-level gating units utilize scalar gating factors The feature stream is weighted overall to output the gated features. And satisfy the following formula:
[0088]
[0089] in, It is used to reflect the confidence level that the current sample contains valid target information, thereby effectively preserving high-value sample information in skip connections and blocking noise propagation in pure background slices.
[0090] like Figure 5 As shown, in order to improve the ability to characterize irregular boundaries, a morphological feature extraction module (MFE) is integrated into the bottleneck layer of the network. The strategy of "splitting-multi-scale transformation-recombination and fusion" is adopted. Asymmetric multi-scale convolution is used to capture morphological cues of different directions and scales, and the recognition of boundaries and core regions is enhanced by combining a structural recalibration mechanism.
[0091] The morphological feature extraction module (MFE) is deployed in the network bottleneck layer to process input features. Perform structured processing using "split-transform-recombination": First, split the input features along the channel dimension into... Subset Each subset sequentially enters an asymmetric multi-scale convolutional AM-Conv unit, and the AM-Conv unit uses a cascaded residual structure to superimpose the output of the previous scale onto the current input to generate multi-scale features. ;
[0092] Standard convolution, asymmetric convolution, and dilated convolution are applied at different scales to generate multi-scale features with rich receptive fields. And satisfy the following formula:
[0093]
[0094] in, Indicates the first Scale-based convolution operations, This indicates the number of multi-scale groups.
[0095] The morphological feature extraction module (MFE) includes a structure recalibration unit (SR-Block), which performs multi-scale feature extraction. Morphological enhancement is performed by parallel computation of global average pooling, local structural pooling (Avg), and global max pooling (Max), which are then fused to generate morphological weights. And satisfy the following formula:
[0096]
[0097] SR-Block utilizes shape weights Multiscale features Dynamic recalibration is performed to obtain And satisfy the following formula:
[0098]
[0099] All recalibrated features are compared with the finest-grained detail features generated by AM-Conv. The pieces are then joined together via a fusion layer. After dimensionality reduction, the final output is obtained by adding the residual to the original input. And satisfy the following formula:
[0100] .
[0101] S4: Input the preprocessed image data into the trained CT image segmentation network. After inference by the CT image segmentation network, the segmentation result of the target CT image is output.
[0102] Example 2:
[0103] This embodiment provides a gated CT image segmentation system based on prior guidance and morphological awareness, including:
[0104] Acquisition unit, used to acquire CT images;
[0105] The preprocessing unit is used to preprocess CT images to form network input;
[0106] Segmentation network building blocks are used to construct CT image segmentation networks.
[0107] The segmentation unit is used to call the CT image segmentation network to perform inference on the network input and output the segmentation result.
[0108] The CT image segmentation network includes a PG-Conv Block convolutional unit with an embedded prior-guided attention module (PGA) in the encoding and decoding stages, a morphological feature extraction module (MFE) integrated in the network bottleneck layer, and an instance-level gating module (IGM) deployed on the skip connection path.
[0109] Example 3:
[0110] To verify the effectiveness and impact of this invention, this embodiment focuses on target CT core and target segmentation, and utilizes a public dataset to complete the entire implementation process from data preparation, model training, to model inference. For the dataset, 131 labeled core and target samples from the LiTS public dataset were selected for training and evaluation.
[0111] Reference Figure 6 The specific process is as follows:
[0112] I. Data Processing and Division
[0113] The original CT volume data was sliced layer by layer, and the CT values were truncated to [-200, 200] according to a preset window width and level range and linearly normalized. The labels were remapped into three categories: background, core, and target, to adapt to multi-class segmentation output. To supplement 3D context information, adjacent three-layer slices were stacked sequentially as a three-channel input, and the input size was unified to 512×512. Following the principle of independent partitioning at the sample level, data leakage was avoided by having the same sample slice appear in both training and testing: 20% of the samples were randomly selected from LiTS as the test set, and the remaining samples were further divided into training and validation sets in an 8:2 ratio. The validation set was used to tune and select the optimal model. To improve the robustness of the model to different imaging conditions and deformations, the input slices and corresponding labels were simultaneously subjected to random rotation, flipping, and elastic deformation during the training phase, and enhancement strategies such as contrast perturbation and noise perturbation were superimposed to expand sample diversity and alleviate overfitting.
[0114] II. Model Training Platform and Parameter Settings
[0115] The network was built and trained using the PyTorch framework, and accelerated training was performed in a GPU environment. The key training hyperparameters were set as follows: batch size of 8, training epoch of 100, initial learning rate of 0.0001, AdamW optimizer with weight decay of 0.001, and CosineAnnealingLR learning rate scheduling to improve convergence stability and fine-tuning capability in the later stage.
[0116] III. Loss Function and Optimization Objective
[0117] To address the class imbalance problem caused by the target region being much smaller than the background and core, a combination of Focal loss and Dice loss is used during training to simultaneously enhance attention to hard-to-classify samples and minority classes and directly optimize segmentation overlap. The weight coefficients of the two losses are set to 0.5 and 1.0, respectively, to balance the numerical magnitude and ensure the stability of training convergence.
[0118] IV. Training Process
[0119] First, the network parameters and optimizer state are initialized. Then, iterative sampling is performed from the training set in mini-batches, with each input being a three-channel slice. The network performs forward computation to obtain the segmentation probabilities for the three classes. The output is then combined with the corresponding pixel-level labels to calculate the combined loss and backpropagate to update the network parameters. After each epoch, inference evaluation is performed on the validation set. Based on the validation set metrics, the model weights with the best performance are selected and saved for final testing and deployment. The evaluation metrics use Dice and IoU to measure the consistency between the predicted region and the ground truth annotation, facilitating a quantitative comparison of the core and target segmentation performance.
[0120] V. Reasoning Process
[0121] In practical applications or testing, the samples to be segmented undergo the same preprocessing as during training (window width and level truncation, normalization, three-layer slice stacking, and uniform size). Then, the optimal model weights obtained from training are loaded, and each slice is input into the network to obtain segmentation probability maps of the core and the target. Finally, based on the maximum probability or threshold strategy, three-class segmentation masks are generated for target localization and quantitative analysis. To ensure output reliability, the inference results can be visually verified, and statistical evaluation can be performed on an independent test set to validate the method's ability to detect small targets in complex backgrounds and its ability to characterize irregular boundaries.
[0122] In this embodiment, compared with existing U-shaped segmentation networks, the CT image segmentation network provided by this invention, through a collaborative design of "prior-guided injection + skip connection gating and screening + bottleneck layer morphological reconstruction," significantly improves the detection and segmentation accuracy of image targets, especially low-contrast small targets, while ensuring the stability of the core overall contour, and effectively improves the characterization quality of irregular boundaries. Ablation results show that the prior-guided attention mechanism can directionally enhance the features of small targets, improving target Dice and IoU by 5.12 and 4.94 percentage points, respectively; the instance-level gating mechanism can suppress the accumulation of skip connection noise, bringing a Dice gain of 2.34 percentage points; and the morphological feature extraction mechanism can enhance the expression of complex boundaries, improving target Dice by 3.27 percentage points. After combining the three, the target Dice reaches 71.87% and the IoU reaches 58.08% on the LiTS dataset, which is an improvement of 8.87 and 9.75 percentage points, respectively, compared with the U-Net baseline, demonstrating the complementarity of modules and the effectiveness of the overall scheme.
[0123] Segmentation effect as follows Figure 7 As shown, this invention has fewer missed detections and spurious responses in target segmentation, clear boundaries, and is close to the gold standard, making it suitable for fine segmentation needs in complex image scenes.
Claims
1. A gated CT image segmentation method based on prior guidance and morphological perception, characterized in that, Includes the following steps: S1: Acquire the target CT image to be segmented; S2: Preprocess the target CT image; S3: Construct and train a CT image segmentation network. The CT image segmentation network is a U-shaped encoder-decoder structure. In the encoding and decoding stages, the convolutional unit PG-Conv Block with embedded prior guided attention module PGA is used for feature extraction. The morphological feature extraction module MFE is integrated in the bottleneck layer of the network, and the instance-level gating module IGM is deployed on the skip connection path to achieve enhancement of small targets, characterization of irregular boundaries and suppression of background noise in the entire process of feature extraction, deep convergence and cross-scale transmission. S4: Input the preprocessed image data into the trained CT image segmentation network. After inference by the CT image segmentation network, the segmentation result of the target CT image is output.
2. The gated CT image segmentation method based on prior guidance and morphological perception according to claim 1, characterized in that, The operation of the embedded prior-guided attention module PGA in step S3 includes: Input features As input, a small target prior map is generated through convolution operations. And satisfy the following formula: ； in, For convolution operations, Use the Sigmoid activation function; Based on prior diagrams The probability distribution in the spatial dimension Importance sampling is performed on the input features to obtain a global context descriptor through weighted aggregation. And satisfy the following formula: ； in, Representational space prior graph In coordinates scalar value at that point Used to maintain numerical stability; The PGA module utilizes a multilayer perceptron. For global context descriptors Generate channel weights and apply them to the original features. Perform dynamic recalibration to enhance output features And satisfy the following formula: ； in, This indicates channel-wise multiplication, and introduces residual connections to enhance the output features. .
3. The gated CT image segmentation method based on prior guidance and morphological perception according to claim 2, characterized in that, In step S3, the instance-level gating module (IGM) is deployed at the skip connection to process the input features. The IGM module performs cascaded optimization processing to suppress noise layer by layer from three dimensions: channel, space, and sample. The IGM module includes a semantic attention module, a positional attention module, and an instance-level gating unit.
4. The gated CT image segmentation method based on prior guidance and morphological perception according to claim 3, characterized in that, The operation of the semantic attention module includes: The semantic attention module computes input features in parallel. The global average pooling (Avg) and max pooling (Max) statistics are obtained by passing them through a multilayer perceptron. After fusion, a channel weight vector is generated, and the input features are recalibrated by channel dimension to obtain the features. And satisfy the following formula: 。 5. The gated CT image segmentation method based on prior guidance and morphological perception according to claim 4, characterized in that, The operation of the location attention module includes: The location attention module generates a spatial attention map using the mean and extreme values of the channel dimension, and then performs convolution operations. Focus on the target region and output spatially enhanced features. And satisfy the following formula: ； in, These represent the average pooling and max pooling operations performed along the channel dimension, respectively. This indicates splicing along the channel dimension.
6. The gated CT image segmentation method based on prior guidance and morphological perception according to claim 5, characterized in that, The operation of the instance-level gating unit includes: Instance-level gating units utilize global average pooling (GAP) to enhance the spatial features. It is compressed into a one-dimensional descriptor and then utilized with a scoring network. Dynamic calculation of scalar gating factor And satisfy the following formula: ； Instance-level gating units utilize scalar gating factors The feature stream is weighted overall to output the gated features. And satisfy the following formula: ； in, It is used to reflect the confidence level that the current sample contains valid target information, thereby effectively preserving high-value sample information in skip connections and blocking noise propagation in pure background slices.
7. The gated CT image segmentation method based on prior guidance and morphological perception according to claim 6, characterized in that, In step S3, the morphological feature extraction module (MFE) is deployed in the network bottleneck layer to process the input features. Perform structured processing using "split-transform-recombination": First, split the input features along the channel dimension into... Subset Each subset sequentially enters an asymmetric multi-scale convolutional AM-Conv unit, and the AM-Conv unit uses a cascaded residual structure to superimpose the output of the previous scale onto the current input to generate multi-scale features. ; Standard convolution, asymmetric convolution, and dilated convolution are applied at different scales to generate multi-scale features with rich receptive fields. And satisfy the following formula: ； in, Indicates the first Scale-based convolution operations, This indicates the number of multi-scale groups.
8. The gated CT image segmentation method based on prior guidance and morphological perception according to claim 7, characterized in that, The morphological feature extraction module (MFE) includes a structural recalibration unit (SR-Block), which performs multi-scale feature extraction. Morphological enhancement is performed by parallel computation of global average pooling, local structural pooling (Avg), and global max pooling (Max), which are then fused to generate morphological weights. And satisfy the following formula: ； SR-Block utilizes shape weights Multiscale features Dynamic recalibration is performed to obtain And satisfy the following formula: ； All recalibrated features are compared with the finest-grained detail features generated by AM-Conv. The pieces are then joined together via a fusion layer. After dimensionality reduction, the final output is obtained by adding the residual to the original input. And satisfy the following formula: 。 9. A gated CT image segmentation system based on prior guidance and morphological perception, characterized in that, For implementing the method of claim 1, the system comprises: Acquisition unit, used to acquire CT images; The preprocessing unit is used to preprocess CT images to form network input; Segmentation network building blocks are used to construct CT image segmentation networks. The segmentation unit is used to call the CT image segmentation network to perform inference on the network input and output the segmentation result.
10. A gated CT image segmentation system based on prior guidance and morphological perception according to claim 9, characterized in that, The CT image segmentation network includes a PG-Conv Block convolutional unit with an embedded prior guided attention module (PGA) in the encoding and decoding stages, a morphological feature extraction module (MFE) integrated in the network bottleneck layer, and an instance-level gating module (IGM) deployed on the skip connection path.