A medical image segmentation method and system for true lumen, false lumen and false lumen thrombus of aortic dissection

Abstract

The application discloses a medical image segmentation method and system for a true lumen, a false lumen and a false lumen thrombus of a aortic dissection. Firstly, three-dimensional medical image data of the aortic dissection is acquired, pre-processing operation is performed on the image data to obtain a two-dimensional slice image, and the two-dimensional slice image is divided into local regions. Secondly, a hierarchical feature extraction module is used to extract features of the divided image. Then, a dynamic region adaptive module is used to adjust the extracted feature map, a region response map is generated through a guide path, and different spatial regions in the feature map are given explicit adjustment weights. Finally, the multi-scale feature map after the region adjustment is unified and aggregated, and a segmentation mask is output through a prediction module to generate a segmentation result with clear structure boundary and complete semantics. The application improves the recognition accuracy of complex anatomical structures, enhances the detection capability and boundary perception performance of small volume lesions, and has good universality and stability.

Description

Technical Field

[0001] This application relates to the field of medical image processing, particularly to the three-dimensional CT image segmentation of aortic dissection, specifically to a medical image segmentation method and system for aortic dissection of the true lumen, false lumen, and false lumen thrombus. Background Technology

[0002] Aortic dissection is a serious and life-threatening cardiovascular disease, primarily caused by a tear in the aortic intima, allowing blood to enter and create a true and false lumen structure. This can further lead to massive internal bleeding, shock, multiple organ failure, and even myocardial infarction or sudden death, posing an extremely high risk of immediate mortality and long-term disability. Therefore, rapid and accurate identification of the true and false lumens, as well as any potential false lumen thrombi, is crucial for developing individualized treatment plans and prognostic assessments.

[0003] Computed tomography (CTA) is an important tool for the diagnosis and preoperative assessment of aortic dissection, providing detailed information on the aortic anatomy, including the morphological characteristics of the true lumen, false lumen, and false lumen thrombus. Accurately quantifying the distribution of the true lumen, false lumen, and false lumen thrombus in CTA images is crucial for surgical planning and postoperative prognosis assessment.

[0004] Existing technologies have enabled automated segmentation of the aorta, true lumen, and false lumen. For example, methods based on multi-atlas registration and aortic centerline extraction can assess aortic diameter in non-ECG-gated, non-contrast CT scans, providing an effective tool for screening and clinical assessment of aortic dilatation. However, these methods still lack the ability to detect false lumen thrombosis, primarily due to the lack of datasets specifically designed for it.

[0005] Furthermore, while multi-task segmentation methods based on 3D convolutional neural networks can automatically segment the aorta, true lumen, and false lumen within a unified framework, they primarily focus on the segmentation of the entire aorta and the true and false lumens, with limited ability to identify false lumen thrombi. Moreover, these methods still fall short in their ability to precisely identify false lumen thrombi, struggling to accurately distinguish between true lumen, false lumen, and false lumen thrombi.

[0006] Chinese invention patent CN113674291B discloses an image segmentation method for aortic dissection false lumen, which can be used to identify the true and false lumen regions, but it does not involve specific modeling of thrombi in the false lumen and has limitations in expressing multi-scale structural details.

[0007] Some works have attempted to introduce Transformer structures to enhance image representation capabilities. For example, Chinese invention patent CN116797605A combines SCUNet with an attention module to identify pulmonary embolism, demonstrating certain advantages in structural modeling. However, this method is primarily geared towards pulmonary artery scenarios and lacks the ability to represent small volumes, multi-scale structures, and dynamically adjust regions of thrombus structures in aortic dissections.

[0008] In addition, Chinese invention patent CN115272389A proposes an aortic dissection segmentation method with an intimal valve attention module, which can help enhance certain spatial channel information. However, this mechanism has structural differences from the explicit regulation mechanism based on regional response maps proposed in this application, and it is still difficult to meet the needs of fine modeling of the blurred boundary region of false lumen thrombus.

[0009] Some methods are based on traditional 3D convolutional models for structural modeling. For example, CN108805134B discloses a method for segmenting and modeling aortic dissection, which mainly uses 3D CNN to extract spatial features, but does not integrate explicit attention or region adaptation mechanisms, and has limited ability to model the complex boundary between the true lumen, false lumen, and thrombus.

[0010] On the other hand, in recent years, 3D CTA datasets containing annotations of the true lumen, false lumen, and false lumen thrombi have emerged, making automatic segmentation of false lumen thrombi possible. However, existing false lumen thrombus segmentation models still have significant room for improvement in accuracy, mainly because false lumen thrombi have complex structures, large morphological variations, and uneven distribution in clinical data. Achieving accurate segmentation of the true lumen, false lumen, and false lumen thrombi within the complex aortic anatomy, especially efficient segmentation of false lumen thrombi, remains a major challenge in current research.

[0011] In summary, although existing technologies have made some progress in the segmentation of true and false lumens, they still have significant shortcomings in the accurate identification and boundary adjustment of low-contrast, small-volume targets such as false lumen thrombi. There is an urgent need for an intelligent segmentation method for aortic dissection with the capabilities of region perception, adaptive weight adjustment and multi-scale fusion to improve the segmentation accuracy and clinical usability of the model in complex structures. Summary of the Invention

[0012] This application provides a medical image segmentation method and system based on hierarchical feature extraction and dynamic region adaptation mechanism, which aims to achieve high-precision identification and segmentation of small-volume lesions in complex anatomical structures. It is especially suitable for the fine modeling of key structures such as the true lumen, false lumen, and false lumen thrombus in aortic dissection. By improving the model's ability to distinguish regions with blurred boundaries, it effectively reduces misclassification and missed detection of small-volume lesions and enhances the overall segmentation performance of the model.

[0013] In a first aspect, the present invention provides a medical image segmentation method for aortic dissection, comprising the true lumen, false lumen, and false lumen thrombus, including the following steps:

[0014] Image preprocessing: Acquire three-dimensional medical image data of aortic dissection, perform preprocessing operations on the image data, including window width and window level adjustment, grayscale normalization, image cropping and smoothing, and convert the three-dimensional image into two-dimensional slices to reduce computational complexity and improve spatial structure resolution.

[0015] Local structure segmentation: The two-dimensional slice image is segmented into local regions to preserve spatial structure information and enhance local detail expression, providing contextual support for subsequent feature extraction;

[0016] Feature extraction: Features are extracted from the segmented image through a hierarchical feature extraction module. The hierarchical feature extraction module integrates long-range modeling capabilities and local detail representation capabilities, enhances the perception of small volume structures through a multi-scale aggregation strategy, and introduces a channel attention mechanism to explicitly adjust and strengthen key channels in the feature map.

[0017] Region Adjustment Enhancement: The extracted feature map is adjusted by using a dynamic region adaptation module. A region response map is generated by guiding the path. Explicit adjustment weights are assigned to different spatial regions in the feature map to strengthen the expression of edge structures, enhance structural differences, suppress background interference and non-target responses, and improve the ability to distinguish small-volume structures.

[0018] Feature fusion and prediction: The multi-scale feature maps after region adjustment are uniformly aggregated and a segmentation mask is output through the prediction module. The prediction module adopts a pixel-by-pixel convolutional classification method to output the class probability of each pixel, generating a segmentation result with clear structural boundaries and complete semantics.

[0019] A second aspect of the present invention provides a medical image segmentation system for aortic dissection, including the true lumen, false lumen, and false lumen thrombus, comprising the following modules:

[0020] Image preprocessing module: used to acquire three-dimensional medical image data of aortic dissection and perform preprocessing operations on it, including window width and window level adjustment, grayscale normalization, image cropping and smoothing, and convert the three-dimensional image into two-dimensional slices to reduce computational complexity and improve spatial structure resolution.

[0021] Local structure segmentation module: used to segment the two-dimensional slice image into local regions, preserve spatial structure information and enhance the expression of local details, and provide contextual support for subsequent feature extraction;

[0022] Feature extraction module: used to extract features from the segmented image, integrates long-range modeling capabilities with local detail representation capabilities, enhances the perception of small volume structures through multi-scale aggregation strategy, and introduces a channel attention mechanism to explicitly adjust and enhance key channels in the feature map;

[0023] Region Adjustment Enhancement Module: This module is used to adjust the extracted feature map by generating a region response map through a guide path. It assigns explicit adjustment weights to different spatial regions in the feature map, strengthens the expression of edge structures, enhances structural differences, suppresses background interference and non-target responses, and improves the ability to distinguish small-volume structures.

[0024] Feature fusion and prediction module: This module is used to uniformly aggregate the multi-scale feature maps after region adjustment, and output a segmentation mask through the prediction module. The prediction module adopts a pixel-by-pixel classification method, outputs the class probability of each pixel, and generates a segmentation result with clear structural boundaries and complete semantics.

[0025] In summary, the medical image segmentation method and system proposed in this application can realize the entire process of modeling from 3D image preprocessing, structure preservation, feature extraction, region adjustment to result prediction. This invention improves the accuracy of complex anatomical structure recognition while enhancing the detection capability and boundary awareness performance of small lesions, possessing good versatility, stability, and clinical deployment prospects. Attached Figure Description

[0026] The accompanying drawings are used to illustrate the various modules and overall process in the specific implementation, and to help understand the technical solution of this application. The drawings include, but are not limited to, the following:

[0027] Figure 1 This is a schematic diagram of the overall process of the medical image segmentation method provided in the embodiments of this application;

[0028] Figure 2 This is a schematic diagram of the structure of the hierarchical feature extraction module provided in the embodiments of this application;

[0029] Figure 3 This is a schematic diagram of the structure of a single dynamic region adaptive module provided in an embodiment of this application;

[0030] Figure 4 This is a schematic diagram showing the connection relationship between the modules provided in the embodiments of this application;

[0031] Figure 5 This is a schematic diagram of the multi-scale feature fusion process provided in an embodiment of this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0033] This application provides a medical image segmentation method and system based on hierarchical feature extraction and dynamic region adaptation, which can effectively improve the segmentation effect and accuracy of small-volume lesions (such as false cavities and false-cavity thrombi) in complex anatomical structures. By combining a hierarchical feature extraction module and a dynamic region adaptation module, the sensitivity to small-volume structures is enhanced, and the error in the segmentation process of false-cavity thrombi is reduced. This application solves the problems of traditional methods being unable to effectively segment complex anatomical structures and having low segmentation accuracy of false-cavity thrombi.

[0034] The basic concept of this application:

[0035] This application first preprocesses the 3D CTA medical image data, including window width and level adjustment, grayscale normalization, image cropping and smoothing operations, to ensure data consistency and structural clarity, and converts the 3D image into 2D slices to reduce computational complexity and improve the model's ability to resolve spatial structures.

[0036] Subsequently, each 2D slice image is divided into patches to preserve spatial structure information and enhance local detail representation, providing rich contextual support for subsequent feature modeling. The patch division size and step size can be flexibly set according to task requirements to achieve a balance between local structure representation capability and computational efficiency.

[0037] In the feature extraction stage, a hierarchical feature extraction enhancement module is introduced. This module integrates the long-range modeling capabilities of the SwinTransformer with the local detail representation capabilities of the convolutional module. It captures long-distance dependencies through the Shifted Window mechanism and enhances the feature contrast between true and false cavities and thrombus regions by combining a channel-adaptive attention mechanism. Furthermore, a multi-scale aggregation strategy is used to perform feature fusion after each Transformer module layer, effectively reducing semantic dilution during information transmission and improving the model's ability to perceive small-volume structures.

[0038] Building upon the above, this application further introduces a dynamic region adaptive module. Addressing the issues of blurred boundaries and structural similarities between false lumen thrombi, this module generates region response maps via a guided path, assigning explicit adjustment weights to different spatial regions within the feature map. This module strengthens the representation of edge structures, enhances structural diversity, effectively suppresses background interference and non-target responses, and improves the model's ability and robustness in discriminating small-volume structures.

[0039] After region adjustment, the multi-scale feature maps are further uniformly aggregated along the channel dimension, and the prediction module outputs a segmentation mask. This module uses a pixel-wise convolutional classification method, outputting the class probability of each pixel through the Softmax activation function, ultimately generating a segmentation result with clear structural boundaries and complete semantics.

[0040] To achieve the above objectives, this application will, in conjunction with the accompanying drawings, describe in detail the specific implementation methods of the hierarchical extraction enhancement and dynamic region adaptation modules, demonstrating their application value in the segmentation of complex anatomical structures.

[0041] like Figure 1 As shown, the method for separating the true lumen, false lumen, and false lumen thrombus in aortic dissection provided in this application includes the following steps:

[0042] Example

[0043] S1: Image Preprocessing

[0044] S1.1 Acquire 3D CTA image data of aortic dissection. The original image data is in NII.GZ format, and the image size is denoted as W×H×S, where W and H represent the width and height of the image, respectively, and S is the number of image slices. To ensure data consistency in all directions and network input standardization, this embodiment uniformly resamples the images to a voxel spacing of 1mm. 3 This yields a three-dimensional tensor with dimensions of 512×512×S.

[0045] S1.2 To enhance the contrast between the aorta and surrounding soft tissues and improve model recognition capabilities, the original CTA image is cropped by window level and width. Specifically, based on the tissue density characteristics of the aorta and its surrounding structures, the window level (WL) is set to 90 and the window width (WW) to 500. At this point, the image's grayscale range is:

[0046]

[0047] The grayscale value x of each pixel in the image is cropped, and the cropped pixel value x' is expressed as follows:

[0048]

[0049] The above-mentioned cropping process can improve the contrast between the aortic lumen and the surrounding soft tissue, enhance the model's ability to identify the boundary between true and false lumens and small-volume structures, and provide a unified input basis for subsequent normalization and segmentation.

[0050] S1.3 performs grayscale normalization on the image, standardizing the cropped grayscale values ​​to the [0,1] range. The normalization formula is as follows:

[0051]

[0052] in These are the pixel values ​​after cropping by window width and window level. These are the normalized pixel values.

[0053] S1.4 The normalized image is sliced ​​into two dimensions. The three-dimensional data is divided into multiple two-dimensional slices along the axial direction, with a size of 512×512, which are used as the input images for subsequent models.

[0054] S2: Patch partitioning

[0055] S2.1 performs patch division processing on the two-dimensional slice image obtained in S1.4. Specifically, each 512×512 two-dimensional image is divided into multiple small-sized patches, and the size of each patch is set to P×P. The overlap step size S can be set during the division to control the coverage between patches.

[0056] Image segmentation in S2.2 employs a sliding window mechanism, translating and segmenting each 2D image spatially to generate several structure-preserving local patches. By appropriately setting the window size and stride, the model's ability to learn fine-grained local regions can be enhanced while maintaining the integrity of the spatial structure.

[0057] The patch image set after S2.3 partitioning serves as the input for the subsequent feature extraction module. This not only preserves the local details of the images but also improves the feature representation efficiency and memory utilization during model training through data block partitioning.

[0058] In this embodiment, the patch size is set to 128×128, the sliding step size is 64, and the insufficient edge areas are filled by symmetrical filling to ensure that all areas can be covered and learned by the network.

[0059] S3: Multi-level feature extraction

[0060] Based on the patches obtained in step S2.3, the process proceeds to the feature extraction stage. This application employs the following... Figure 2The hierarchical feature extraction module structure shown aims to extract hierarchical, highly semantic feature maps from the input image, enhancing the model's ability to express multi-scale, small objects, and complex structures. This module comprises three stages: a linear embedding module, a patch merging module, and a Swin Transformer Block, possessing excellent structural semantic preservation and feature modeling capabilities. Its overall process includes the following three steps:

[0061] In the initial stage of feature extraction (S3.1), the model first performs linear feature embedding on the input image patches. This process is completed by the Patch linear embedding module, which uses a 4×4 convolutional kernel with a stride of 4 to perform the embedding operation, mapping the pixel-level input of the two-dimensional image to a fixed-dimensional high-dimensional feature space, completing the transformation from image representation to preliminary feature expression, and providing a structured input foundation for subsequent contextual modeling.

[0062] The feature map output from the first stage (S3.2) serves as the input to the second stage. The second and third stages have the same structure, both containing: a patch merging module and two consecutive Swing Transformer Blocks.

[0063] The Patch merging module performs channel splicing on adjacent 2×2 regions in the first-stage output feature map, and achieves spatial downsampling (halving the size) and channel expansion (doubling the number of channels) through normalization and linear transformation.

[0064] Each Swin Transformer Block contains two substructures, which alternately stack multi-head self-attention mechanisms (W-MSA and SW-MSA) based on fixed windows and sliding windows to jointly model local and cross-window spatial dependencies. Each substructure sequentially executes the following flow: layer normalization → self-attention module (W-MSA or SW-MSA) → residual connection → layer normalization → feedforward network module → residual connection, to enhance nonlinear expressive power.

[0065] The output feature map from the second stage is then fed into the third stage. Through this hierarchical structure, the model extracts more abstract and compressed high-semantic features layer by layer, and gradually enhances its ability to perceive spatial structure and boundary features.

[0066] To enhance the ability of each channel in the feature map to distinguish key anatomical structures (such as true cavities, false cavities, and false cavity thrombi), S3.3 introduces a channel attention mechanism after each level of the Swing Transformer Block×2 module to explicitly adjust and strengthen the channel dimension. This mechanism is implemented using a Squeeze-and-Excitation structure. First, global average pooling is performed on each channel to extract channel-level semantic response information. Then, a channel attention weight vector is constructed using a two-layer fully connected network and a ReLU activation function, and normalized using a Sigmoid activation. These weights are used to adjust the input feature map channel-by-channel, highlighting the response intensity of key semantic channels and suppressing interference from non-target or redundant feature channels.

[0067] Channel attention module as Figure 2 The hierarchical feature extraction module shown is... Figure 3 The connection components between the dynamic region adaptive modules shown have undergone response optimization and structural enhancement in the channel dimension of their output feature maps, providing semantically clear and discriminative input representations for subsequent region-level modeling. This further supports fine segmentation and accurate recognition of complex structures. Their connection relationships are as follows: Figure 4 As shown.

[0068] S4: Enhanced Regional Regulation

[0069] To further enhance the model's ability to distinguish spatial boundaries between complex anatomical structures (such as true cavities, false cavities, and false cavity thrombi) and improve its robustness to ambiguous regions and preservation of structural details, this application proposes a dynamic region adaptive module. The module structure is as follows: Figure 3 As shown, the entire system consists of a main path for feature preservation and a guiding path for generating region response maps. These paths are used to preserve structural semantic information and generate spatial guiding weights, respectively. A weighted fusion operation is then used to output a region-enhanced feature map. The left side of the figure shows the overall structural flow of the main path and the guiding path, while the right side further refines the internal processing of the guiding path, including convolution, fully connected layers, and activation operations, used to construct the spatial response map.

[0070] The guiding path dynamically adjusts the response intensity of different spatial regions in the feature map by constructing a region response map, thereby improving the model's saliency representation of key anatomical structures, suppressing background interference, and enhancing region consistency. To achieve multi-scale modeling, this module is independently applied to feature maps at different scales, specifically including the following three steps:

[0071] S4.1 In step S3 of this application, the multi-level feature maps obtained correspond to semantic representations at different scales. To preserve the spatial structure and contextual information of each feature map layer, the three sets of feature maps at different resolutions are fed into three independent dynamic region adaptation modules for region-aware processing. This design allows the semantic information at each scale to independently complete boundary enhancement and region adjustment while maintaining its original spatial characteristics, avoiding semantic dilution or boundary blurring caused by information fusion between different scales. Each dynamic region adaptation module receives a set of feature maps at a specific scale as input, ensuring the independence of region modeling and the plasticity of multi-scale features.

[0072] S4.2 For each scale of input feature map, the dynamic region adaptation module first constructs a region response map to characterize the response intensity of each spatial location in the feature map to the target structure (such as a false lumen or a false lumen thrombus).

[0073] The response map extracts local contextual features through a set of lightweight convolutional operations, then compresses channel information using global average pooling, and finally generates normalized weight coefficients through two fully connected layers and a nonlinear activation function to represent the importance level of different regions. This response map can highlight anatomical boundary regions, suppress interference from non-target regions, and provide explicit semantic guidance for subsequent adjustment paths.

[0074] In this embodiment, the region response map construction path consists of two consecutive 3×3 convolutional operations. The number of output channels in each convolutional layer is set to half the number of input channels. Batch normalization and ReLU activation functions are applied after each convolutional layer to enhance nonlinear modeling capabilities and maintain feature stability. After global average pooling, the feature map generates a normalized region weight map through a set of 1×1 fully connected layers and a sigmoid activation function to achieve pixel-level saliency adjustment, enhancing salient regions and suppressing non-target regions, thereby providing spatial guidance for subsequent feature adjustment paths.

[0075] After the region response map described in S4.3 is generated, it is upsampled to the same spatial resolution as the original feature map and then weighted and fused with the input feature map along the channel-wise dimension. The fusion method adopts an element-wise multiplication operation, that is, the feature map of each channel and its corresponding response weight are weighted pixel by pixel, which enhances the semantic expression strength of the structural region and suppresses feature interference from the background or blurred regions. This process can be regarded as an explicit region attention mechanism, which guides the model to focus on regions with clear structural boundaries and significant semantics, significantly improving the discrimination performance and structural perception ability in the subsequent feature fusion and mask prediction stages. Finally, the output region-enhanced feature map retains the spatial structure consistent with the input resolution, providing optimized input for feature fusion and decoding in the S5 stage.

[0076] S5: Feature Fusion and Output Prediction

[0077] After completing multi-scale feature extraction and region modulation enhancement, this application further designs a feature fusion mechanism to uniformly aggregate region enhancement feature maps from different scales, generating the final segmentation prediction result. This fusion process aims to fully utilize the boundary detail representation capability of high-resolution features and the global semantic modeling capability of low-resolution features to improve the model's comprehensive recognition performance of structural regions (such as true cavities, false cavities, and false cavity thrombi) at different scales.

[0078] like Figure 5 As shown, specifically, the feature maps output by the dynamic region adaptive module from three scales are first spatially aligned. The low-resolution and medium-resolution feature maps are upsampled to restore the high-resolution scale, ensuring that the three sets of features remain consistent in spatial dimensions. Subsequently, they are cascaded and fused along the channel dimension, and a lightweight convolutional layer is introduced for channel compression and feature integration, thereby constructing a unified fused semantic representation.

[0079] The fused feature map is fed into the prediction module for class mapping. This module employs a pixel-wise classification architecture, consisting of a set of 1×1 convolutional operations, used to map the channel dimensions in the fused feature map to specific class spaces. The convolutional output is then fed into a pixel-wise Softmax activation function to obtain the probability distribution of each pixel across all classes.

[0080] In this embodiment, the number of output channels of the category mapping module is set to 3, corresponding to the three structural regions of the aortic dissection: the true lumen, the false lumen, and the false lumen thrombus. In the probability map output by Softmax, the final predicted category of each pixel is selected by the argmax operation, which selects the channel label corresponding to its highest probability, generating a final segmentation mask result with clear structural boundaries and complete semantic information.

[0081] Through the above fusion and prediction process, high-precision identification and regional segmentation of small-volume lesions in complex anatomical structures can be achieved, providing terminal decision-making basis for intelligent auxiliary diagnosis of aortic dissection.

[0082] By executing steps S1 to S5 sequentially, the medical image segmentation method proposed in this application can complete the entire modeling process from CTA image preprocessing, structure-preserving patch slicing, hierarchical feature extraction and enhancement, to the final high-precision prediction output. The synergistic effect among the key modules not only improves the model's ability to express multi-scale and complex anatomical structures, but also strengthens its ability to respond to and discriminate against regions with blurred boundaries and small lesions, ultimately achieving the generation of a segmentation mask with clear structure and accurate category discrimination.

[0083] Based on the same concept as the above method, this application also provides a medical image segmentation system for aortic dissection, including the true lumen, false lumen, and false lumen thrombus, comprising the following modules:

[0084] Image preprocessing module: used to acquire three-dimensional medical image data of aortic dissection and perform preprocessing operations on it, including window width and window level adjustment, grayscale normalization, image cropping and smoothing, and convert the three-dimensional image into two-dimensional slices to reduce computational complexity and improve spatial structure resolution.

[0085] Local structure segmentation module: used to segment the two-dimensional slice image into local regions, preserve spatial structure information and enhance the expression of local details, and provide contextual support for subsequent feature extraction;

[0086] Feature extraction module: used to extract features from the segmented image, integrates long-range modeling capabilities with local detail representation capabilities, enhances the perception of small volume structures through multi-scale aggregation strategy, and introduces a channel attention mechanism to explicitly adjust and enhance key channels in the feature map;

[0087] Region Adjustment Enhancement Module: This module is used to adjust the extracted feature map by generating a region response map through a guide path. It assigns explicit adjustment weights to different spatial regions in the feature map, strengthens the expression of edge structures, enhances structural differences, suppresses background interference and non-target responses, and improves the ability to distinguish small-volume structures.

[0088] Feature fusion and prediction module: This module is used to uniformly aggregate the multi-scale feature maps after region adjustment, and output a segmentation mask through the prediction module. The prediction module adopts a pixel-by-pixel classification method, outputs the class probability of each pixel, and generates a segmentation result with clear structural boundaries and complete semantics.

[0089] In summary, this application possesses structure preservation and hierarchical modeling capabilities, effectively enhancing the contextual semantic modeling of complex medical images; the channel and spatial joint enhancement mechanism finely enhances the responsiveness to key structures; the dynamic region adjustment mechanism strengthens the feature expression of boundary and lesion regions, significantly alleviating the problem of anatomical confusion; and the multi-scale aggregation and prediction structure are compact and efficient, adapting to the fine prediction needs of different granular structures. Thus, it can achieve highly robust and high-precision structural-level segmentation under complex CTA medical imaging conditions, possessing good versatility and clinical deployment prospects.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A medical image segmentation method for aortic dissection, including the true lumen, false lumen, and false lumen thrombus, characterized in that, Includes the following steps: Image preprocessing: Acquire three-dimensional medical image data of aortic dissection, perform preprocessing operations on the image data, including window width and level adjustment, grayscale normalization, image cropping and smoothing, and convert the three-dimensional image into two-dimensional slices; the three-dimensional medical image data is CT angiography CTA image data; Local structure segmentation: The two-dimensional slice image is segmented into local regions to preserve spatial structure information and enhance the expression of local details; Feature extraction: Features are extracted from the segmented image using a hierarchical feature extraction module; The hierarchical feature extraction module includes: The linear embedding module is used to map input image patches to a high-dimensional feature space; The Patch merging module is used to perform spatial downsampling and channel expansion on feature maps; The Swin Transformer Block contains a multi-head self-attention mechanism based on fixed windows and sliding windows, used to jointly model local and cross-window spatial dependencies; A channel attention mechanism is introduced after each Swin Transformer Block, which includes the following steps: Perform global average pooling on each channel to extract channel-level semantic response information; Channel attention weight vectors are constructed using a two-layer fully connected network and a non-linear activation function; The weight vector is normalized to generate weights for channel-by-channel weighted adjustment of the input feature map; Region Adjustment Enhancement: The extracted feature map is adjusted using a dynamic region adaptation module. A region response map is generated through a guide path, and explicit adjustment weights are assigned to different spatial regions in the feature map. The dynamic region adaptive module includes: A main path for feature preservation is used to retain structural semantic information; A guiding path is used to generate a region response map. The guiding path highlights the dissected boundary region and suppresses interference from non-target regions through lightweight convolution operations, global average pooling, fully connected networks and nonlinear activation functions. The region response map is upsampled to the same spatial resolution as the original feature map and then weighted and fused with the input feature map in a channel-wise dimension to enhance the semantic expression strength of the structural region. The fusion method adopts an element-wise multiplication operation, that is, the feature map of each channel and its corresponding response weight are weighted pixel by pixel to enhance the semantic expression strength of the structural region and suppress feature interference from background or blurred regions. Feature fusion and prediction, specifically: Spatial size alignment is performed on enhanced feature maps from regions of different scales; Cascaded fusion is performed along the channel dimension, and lightweight convolutional layers are introduced for channel compression and feature integration. By using pixel-by-pixel convolutional classification to output the class probability of each pixel, a segmentation result with clear structural boundaries and complete semantics is generated.

2. The medical image segmentation method for aortic dissection, including the true lumen, false lumen, and false lumen thrombus, according to claim 1, is characterized in that... The image preprocessing step includes the following preprocessing operations: The CTA image data was resampled to a voxel spacing of 1 mm. 3 ; The window width and window level of the image were adjusted to enhance the contrast between the aorta and the surrounding soft tissue; The adjusted image is then subjected to grayscale normalization, which normalizes the grayscale values ​​to the [0,1] range. The 3D image is divided into multiple 2D slices along the axial direction, which serve as the input images for subsequent models.

3. The medical image segmentation method for aortic dissection, including the true lumen, false lumen, and false lumen thrombus, according to claim 1, is characterized in that... In the local structure segmentation step, a sliding window mechanism is used to divide each two-dimensional slice image into multiple small patches to enhance the ability to express local details. Specifically, this includes: Set the size of each patch to 128×128; Set the sliding step size to 64 to control the coverage between patches.

4. The medical image segmentation method for aortic dissection, including the true lumen, false lumen, and false lumen thrombus, according to claim 3, is characterized in that... For areas with insufficient edge coverage, a symmetrical filling method is used to fill in the gaps, ensuring that all areas can be covered and learned by the network.

5. The medical image segmentation method for aortic dissection, including the true lumen, false lumen, and false lumen thrombus, according to claim 1, is characterized in that... The process of generating the regional response map in the guidance path includes: Two consecutive convolutional operations are performed, with a normalization and non-linear activation function applied after each convolutional layer. After global average pooling, a normalized region weight map is generated through a fully connected layer and a non-linear activation function to achieve pixel-level saliency adjustment.

6. A medical image segmentation system for aortic dissection, including the true lumen, false lumen, and false lumen thrombus, characterized in that, Includes the following modules: Image preprocessing module: used to acquire three-dimensional medical image data of aortic dissection and perform preprocessing operations on it, including window width and window level adjustment, grayscale normalization, image cropping and smoothing, and convert the three-dimensional image into two-dimensional slices. The three-dimensional medical image data is CT angiography CTA image data. Local structure segmentation module: used to segment the two-dimensional slice image into local regions, preserve spatial structure information and enhance the expression of local details; Feature extraction module: Used to extract features from the segmented image, including: The linear embedding module is used to map input image patches to a high-dimensional feature space; The Patch merging module is used to perform spatial downsampling and channel expansion on feature maps; The Swin Transformer Block contains a multi-head self-attention mechanism based on fixed windows and sliding windows, used to jointly model local and cross-window spatial dependencies; A channel attention mechanism is introduced after each Swin Transformer Block, which includes the following steps: Perform global average pooling on each channel to extract channel-level semantic response information; Channel attention weight vectors are constructed using a two-layer fully connected network and a non-linear activation function; The weight vector is normalized to generate weights for channel-by-channel weighted adjustment of the input feature map; The region adjustment and enhancement module is used to adjust the extracted feature maps by generating region response maps through a guide path and assigning explicit adjustment weights to different spatial regions in the feature maps. This includes: A main path for feature preservation is used to retain structural semantic information; A guiding path is used to generate a region response map. The guiding path highlights the dissected boundary region and suppresses interference from non-target regions through lightweight convolution operations, global average pooling, fully connected networks and nonlinear activation functions. The region response map is upsampled to the same spatial resolution as the original feature map and then weighted and fused with the input feature map in a channel-wise dimension to enhance the semantic expression strength of the structural region. The fusion method adopts an element-wise multiplication operation, that is, the feature map of each channel and its corresponding response weight are weighted pixel by pixel to enhance the semantic expression strength of the structural region and suppress feature interference from background or blurred regions. Feature fusion and prediction module: Used to perform the following operations: Spatial size alignment is performed on enhanced feature maps from regions of different scales; Cascaded fusion is performed along the channel dimension, and lightweight convolutional layers are introduced for channel compression and feature integration. By using pixel-by-pixel convolutional classification to output the class probability of each pixel, a segmentation result with clear structural boundaries and complete semantics is generated.

Images