An automatic segmentation method and system for aortic dissection images

Abstract

This invention relates to an automatic segmentation method and system for aortic dissection images. Utilizing the temporal continuity of two-dimensional CT image sequences, and combining label combination processing, dynamic anchor point selection, and a bidirectional propagation mechanism, it achieves cross-frame matching to supplement global information in long sequences, solving the problem of incomplete single-frame labels. Simultaneously, this application employs an adaptive label propagation strategy formed by label combination processing, dynamic anchor point selection, and the bidirectional propagation mechanism to correct propagation errors and focus on feature learning of similar combinations, balancing efficiency and accuracy. Furthermore, by combining spatial attention and channel attention mechanisms to enhance contextual awareness of true and false cavity boundaries in single-frame labels, it improves high-quality local features, achieving high-precision segmentation of three-dimensional CT images.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and artificial intelligence technology, and more specifically, to an automatic segmentation method and system for aortic dissection images. Background Technology

[0002] In recent years, deep learning technology has made significant progress in the field of medical image segmentation, but it still faces unique challenges in the segmentation of sparse tubular structures such as type B aortic dissection (TBAD), mainly in the following aspects: (1) The morphology of the true and false lumens of aortic dissection is varied, resulting in inconsistent label distribution in different CT slices. Existing methods (such as U-Net, nnU-Net, etc.) rely on local feature extraction and are insufficient in terms of global context modeling. In addition, traditional attention mechanisms (such as channel attention) lack the synergistic optimization of spatial and channel features and are not sensitive enough to boundaries. (2) The tubular structure of the aorta in CT sequences has significant spatial continuity, but existing methods are mostly based on independent segmentation of single-frame images, ignoring the temporal correlation between sequences. Two-dimensional models (such as CS-Net) cannot model axial dynamic changes, while three-dimensional models (such as 3D U-Net) are difficult to adapt to long sequence processing due to high computational complexity, resulting in segmentation results with discontinuities or label jumps in adjacent slices, which seriously affects the reliability of clinical diagnosis. (3) As an acute condition, TBAD requires rapid segmentation in clinical practice to assist in emergency decision-making. However, existing methods (such as Swin-Unet) suffer from large parameter counts and computational complexity, making it difficult to meet real-time inference requirements; lightweight models (such as MobileNet) suffer from insufficient feature extraction capabilities, leading to a significant decrease in accuracy. How to improve efficiency while maintaining high accuracy has become a key challenge for clinical deployment. Summary of the Invention

[0003] The technical problem to be solved by this invention is how to achieve high-precision and high-efficiency image segmentation.

[0004] This invention provides an automatic segmentation method for aortic dissection images, comprising:

[0005] Step 1: Acquire labeled 3D CT images and process them to obtain a 2D CT image sequence composed of labeled 2D CT images. The labels include three categories: aortic wall labels, true lumen labels, and false lumen labels.

[0006] Step 2: Morphological modeling is used to identify and repair the breakpoints and discontinuous regions in the labels corresponding to each 2D CT image, and 2D CT images with the same label category combination are grouped into a subset.

[0007] Step 3: Randomly select a frame of 2D CT image from a subset containing all label categories as the anchor frame;

[0008] Step 4: Based on the anchor frame, the image frames adjacent to the anchor frame are matched as query frames using the bidirectional propagation mechanism, and the anchor frame features, query frame features, and anchor frame segmentation label features are extracted using a feature extractor.

[0009] Step 5: Calculate the similarity matrix between the query frame features and the anchor frame features;

[0010] Step 6: Based on the similarity matrix and anchor frame segmentation label features, a weighted propagation mechanism is used to generate high-dimensional features of the query frame;

[0011] Step 7: Based on the high-dimensional features and query frame features, a spatial attention weight map is generated using a spatial attention mechanism; then, a channel attention feature map is generated using a channel attention mechanism.

[0012] Step 8: Input the channel attention feature map into the decoder to obtain the segmentation label of the query frame, iterate the new anchor frame according to the segmentation label, and return to step 4.

[0013] Compared with the prior art, this application has the following advantages: This application utilizes the temporal continuity of two-dimensional CT image sequences, combined with label combination processing, dynamic anchor point selection, and bidirectional propagation mechanism, to achieve cross-frame matching to supplement the global information of long sequences, solving the problem of incomplete single-frame labels. At the same time, this application corrects propagation errors and focuses on feature learning of similar combinations through an adaptive label propagation strategy formed by label combination processing, dynamic anchor point selection, and bidirectional propagation mechanism, balancing efficiency and accuracy. In addition, by combining spatial attention mechanism and channel attention mechanism to enhance the context awareness of true and false cavity boundaries in single-frame labels, high-quality local features are improved, achieving high-precision segmentation of three-dimensional CT images.

[0014] In one possible implementation, step 1 specifically includes:

[0015] Step 101: Acquire labeled 3D CT images;

[0016] Step 102: Resample, standardize, and denoise the labeled 3D CT images;

[0017] Step 103: Extract continuous slices along the Z-axis of the 3D CT image to obtain a 2D CT image sequence composed of multiple labeled 2D CT images.

[0018] Compared with existing technologies, converting 3D CT images into 2D CT image sequences reduces hardware memory requirements while improving the speed and accuracy of target segmentation. At the same time, the spatiotemporal continuity of the segmented targets in the 2D CT image sequence eliminates the problems of poor model generalization and robustness caused by data heterogeneity due to different equipment and scanning parameters.

[0019] In one possible implementation, step 2 specifically includes:

[0020] Step 201: Obtain the binary mask of each label category in each frame of the 2D CT image;

[0021] Step 202: Morphological modeling is used to identify and repair the breakpoints and discontinuous regions of the labels in the binary mask;

[0022] Step 203: Based on the repaired binary mask, identify the label category combinations on each frame of the 2D CT image. The combination of label categories for a frame of 2D CT image is represented as follows:

[0023] ;

[0024] In the formula, Indicates the first Label categories in 2D CT images The binary mask, Represents spatial pixel coordinates; 0 represents the aortic wall label, 1 represents the true lumen label, and 2 represents the false lumen label.

[0025] Step 204: Group those with the same label category The two-dimensional CT images were divided into subsets .

[0026] Compared with existing technologies, this method uses morphological repair to fix label breakpoints and discontinuous regions, improving sample quality. At the same time, it uses the method of dividing 2D CT images with the same label combination into subsets to ensure semantic consistency of samples within the subsets.

[0027] In one possible implementation, step 3 specifically includes:

[0028] Step 301: Define the subset containing all tag categories as the candidate anchor set. The expression is:

[0029] ;

[0030] In the formula, Indicates from the set of candidate anchor points Anchor frames randomly selected from the data. Indicates a combination of label categories;

[0031] Step 302, initialize the tag propagation process, represented as:

[0032] ;

[0033] In the formula, This indicates the network propagation mechanism, starting from the anchor frame. To query frame Propagate the segmentation results. Indicates anchor frame The segmentation mask.

[0034] Compared with existing technologies, using random anchor points to select anchor frames helps reduce dependence on specific locations and ensures stable learning across different sequence positions.

[0035] In one possible implementation, the feature extractor in step 4 includes an initial convolutional layer and four sequentially connected residual layers, the stride of the initial convolutional layer being 1; and the channel configuration parameters of the feature extractor are modified according to the grayscale characteristics of the two-dimensional CT image.

[0036] In one possible implementation, the bidirectional propagation mechanism in step 4 includes:

[0037] Forward propagation: from anchor frame To subsequent frames Transmit features and optimize the current frame using the propagated results from the previous frame;

[0038] Backpropagation, from anchor frame To previous frame Transmit features and correct potential deviations in the current frame using subsequent frames;

[0039] Then, a complete prediction sequence is generated through weighted fusion.

[0040] Compared with existing technologies, this method transforms the 3D sequence segmentation problem into a local matching problem between anchor frames and query frames by propagating forward and backward from the anchor frame as the center. It also utilizes the spatiotemporal continuity of adjacent frames to extend the information of the anchor point in a single frame to the entire sequence, thus solving the problem of information loss in the target region in long sequences and avoiding the high computational complexity of directly processing the entire sequence.

[0041] In one possible implementation, the expression for calculating the similarity matrix between the query frame features and the anchor frame features in step 5 is:

[0042] ;

[0043] In the formula, The query frame characteristics are the location of the query frame. The feature vector at that location, The anchor frame feature is the position of the anchor frame. eigenvectors; This represents the cosine similarity between the query frame features and the anchor frame features.

[0044] Compared with existing technologies, the similarity matrix is ​​used to intuitively reflect the matching process between the query frame and the anchor frame, providing clear trust weights for subsequent weighted propagation. It also filters irrelevant areas through the threshold of the similarity matrix to avoid noise features interfering with the propagation of key information in the anchor frame.

[0045] In one possible implementation, the calculation formula for the high-dimensional features of the query frame generated using the weighted propagation mechanism in step 6 is as follows:

[0046] ;

[0047] In the formula, The segmentation label representing the anchor frame. Indicates the location of the query frame. Predicted high-dimensional features.

[0048] In one possible implementation, step 7 specifically includes:

[0049] Step 701, based on high-dimensional features and query frame features A spatial attention weight map is generated using a spatial attention mechanism, expressed as follows:

[0050] ;

[0051] In the formula, To query frame features , High-dimensional features , Indicates generation 1×1 convolution of channel features, This represents the Sigmoid activation function. express Activation function;

[0052] Step 702, spatial attention weight map and query frame features Multiplication yields the attention map of filtered low-level features. ;

[0053] Step 703: Use a channel attention mechanism to filter low-level feature attention maps. The channel attention weights are obtained through processing. The calculation formula is:

[0054] ;

[0055] In the formula, This indicates a global average pooling operation. This represents the global max pooling operation. Indicates that the core is One-dimensional convolution;

[0056] Step 704: Filter the low-level feature attention map. and channel attention weights Multiplication yields the channel attention feature map. .

[0057] Compared with existing technologies, anchor frame instance features serve as prior knowledge to compensate for the lack of information in the query frame due to noise or occlusion. Similarity weighting amplifies the feature contribution of high-matching regions and reduces the weight of low-matching regions, thereby improving the signal-to-noise ratio of the query frame features.

[0058] The present invention also provides an automatic segmentation system, comprising:

[0059] The data acquisition and processing unit is used to convert the acquired labeled 3D CT images into a labeled 2D CT image sequence.

[0060] The anchor frame selection unit, connected to the data acquisition and processing unit, is used to perform morphological recognition and label repair on the two-dimensional CT image sequence, group two-dimensional CT images with the same label category combination into a subset, and randomly select anchor frames from the subset containing all category label combinations.

[0061] The memory bank, connected to the anchor frame selection unit, is used to store the 2D CT image of the current anchor frame and its segmentation labels;

[0062] The bidirectional propagation unit, connected to the anchor frame selection unit and the data acquisition and processing unit, is used to match the query frames adjacent to the anchor frame using a bidirectional propagation mechanism.

[0063] The feature extraction unit is connected to the bidirectional propagation unit and is configured with a feature extractor to extract anchor frame features, query frame features, and anchor frame segmentation label features.

[0064] The query frame high-dimensional feature generation unit is connected to the feature extraction unit and the memory bank. It is used to calculate the similarity matrix and generate high-dimensional features of the query frame through a weighted propagation mechanism.

[0065] The joint attention unit, connected to the feature extraction unit and the query frame high-dimensional feature generation unit, includes cascaded spatial attention units and channel attention units, and generates channel attention feature maps based on spatial attention mechanisms and channel attention mechanisms.

[0066] The decoder, connected to the joint attention module, is used to decode the channel attention feature map to generate segmentation labels for the query frame. Its output is connected to the memory bank for iterating new anchor frames. Attached Figure Description

[0067] Figure 1This is a flowchart of a specific embodiment 1 of the present invention. Detailed Implementation

[0068] First, those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of the embodiments of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0069] In the description of the embodiments of this application, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of this application based on the specific circumstances.

[0070] In the embodiments of this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0071] The present application will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0072] Example 1:

[0073] See Figure 1 This application discloses an automatic segmentation method for aortic dissection images, comprising:

[0074] Step 1: Acquire labeled 3D CT images and process them to obtain labeled 2D CT image sequences; specifically including:

[0075] Step 101: Acquire tagged 3D CT images. 3D CT images are typically stored in DICOM format and contain metadata such as slice thickness (e.g., 1-2 mm), slice interval (usually equal to slice thickness to ensure continuity), and pixel spacing (e.g., 0.35 × 0.35 mm). These images need to be read and parsed using a medical image processing library (e.g., SimpleITK, ITK-SNAP). Tagged 3D CT images mean that each voxel has been actively labeled with tags for the aortic wall, true lumen of the dissection, false lumen of the dissection, and surrounding tissues. In this embodiment, the tags mainly include three categories: aortic wall tags, true lumen tags, and false lumen tags.

[0076] Step 102 involves resampling, standardizing, and denoising the labeled 3D CT images. In this embodiment, the resampling process primarily addresses the issue of uneven slice thickness, using linear interpolation to unify the slice thickness to 1mm, ensuring consistent spatial resolution of the subsequent 2D sections. The standardization process includes truncating the CT values ​​(HU values) to [-1000, 2000] (the range of lung tissue and bone) and normalizing them to the [0,1] interval to reduce the impact of device differences on the model. The denoising process employs nonlocal mean filtering (NLM) or bilateral filtering to suppress Gaussian noise (such as metal artifacts) in the CT images while preserving vascular edge details.

[0077] Step 103: Extract continuous slices along the Z-axis of the 3D CT image to obtain a 2D CT image sequence composed of multiple labeled 2D CT images.

[0078] Step 2 involves using morphological modeling to identify and repair breakpoints and discontinuous regions in the labels corresponding to each 2D CT image, grouping 2D CT images with the same label category combination into a subset; specifically including:

[0079] Step 201: Obtain the binary mask of each label category in each frame of the 2D CT image. The label combination of each frame of the 2D CT image is the combination of label categories that exist in that frame.

[0080] Step 202 involves using morphological modeling to identify and repair breakpoints and discontinuous regions in the labels of the binary mask; specifically including:

[0081] Fracture repair: For fractured areas of the aortic wall or intima (label discontinuity), a closed operation (expansion followed by corrosion) is used, and the core size is adjusted according to the fracture scale (e.g., 3×3 or 5×5).

[0082] Hole filling: For small holes (missing labels) in the true / false lumen, use a filling algorithm (such as based on connected component analysis, only fill small holes with an area smaller than the threshold to avoid accidentally filling large blood vessel branches).

[0083] Boundary smoothing: For rough label edges (such as the irregular contour of the inner membrane), morphological gradient (dilation and erosion reduction) is used to extract the edges, and then Gaussian filtering is used for smoothing.

[0084] In the embodiments of this application, the closing operation, filling algorithm, morphological gradient and Gaussian filtering are all existing technologies, and will not be described in detail here;

[0085] Step 203: Based on the repaired binary mask, identify the label category combinations on each frame of the 2D CT image. The combination of label categories for a frame of 2D CT image is represented as follows:

[0086] ;

[0087] In the formula, Indicates the first Categories in 2D CT Images The binary mask, Represents spatial pixel coordinates; 0 represents the aortic wall label, 1 represents the true lumen label, and 2 represents the false lumen label. Indicates label category The binary mask in spatial pixel coordinates If there is a non-zero region in the label, then the label category... Included .

[0088] Step 204: Group those with the same label category The two-dimensional CT images were divided into subsets This application embodiment uses a hash table to count all possible tag category combinations, groups frames with the same tag category combination together, and obtains the following three tag category combinations. , , .

[0089] Step 3, from a subset containing all tag categories A frame of a two-dimensional CT image is randomly selected as the anchor frame; specifically, it includes:

[0090] Step 301: Define the subset containing all tag categories as the candidate anchor set. Soon The subset of the set of candidate anchor points is expressed as:

[0091] ;

[0092] In the formula, Indicates from the set of candidate anchor points Anchor frames randomly selected from the data. Indicates a combination of label categories;

[0093] Step 302, initialize the tag propagation process, represented as:

[0094] ;

[0095] In the formula, This indicates the network propagation mechanism, starting from the anchor frame. To query frame Propagate the segmentation results. Indicates anchor point The segmentation mask.

[0096] Step 4: Based on the anchor frame, the query frame is matched in the two-dimensional CT image sequence using the bidirectional propagation mechanism. The anchor frame features, query frame features, and anchor frame segmentation label features are extracted using a feature extractor.

[0097] The bidirectional propagation mechanism includes:

[0098] Forward propagation: from anchor frame To subsequent frames The features are passed along, and the current frame is optimized using the propagated results from the previous frame. The expression is as follows:

[0099] ;

[0100] Backpropagation, from anchor frame To previous frame The feature is transmitted, and potential deviations in the current frame are corrected using subsequent frames. The expression is:

[0101] ;

[0102] After matching the query frame, a feature extractor is used to extract features from the anchor frame and the query frame respectively, obtaining anchor frame features, query frame features, and anchor frame segmentation label features. In this embodiment, the feature extractor uses an improved ResNet-50 as the feature extractor, whose structure includes an initial convolutional layer (layer 0) and four consecutive residual blocks (layers 1-4). To adapt to the characteristics of medical images, the original ResNet-50 is modified as follows: the stride of the initial convolutional layer is adjusted from 2 to 1 to retain more detailed information; and the channel configuration parameters of the feature extractor are modified according to the grayscale characteristics of the two-dimensional CT image.

[0103] The mechanism of dynamic selection of random anchor points reduces keyframe dependence, the bidirectional propagation mechanism corrects accumulated errors, and the combined label processing focuses on feature learning under the same combination to reduce cross-class interference.

[0104] Step 5, calculate the similarity matrix between the query frame features and the anchor frame features; the expression is:

[0105] ;

[0106] In the formula, The query frame characteristics are the location of the query frame. The feature vector at that location, The anchor frame feature is the position of the anchor frame. eigenvectors; It represents the cosine similarity between the query frame features and the anchor frame features; by using the similarity matrix calculation method, it can still accurately establish cross-frame pixel-level correspondences and achieve matching of corresponding anatomical structures even in the presence of morphological changes.

[0107] Step 6: Based on the similarity matrix and anchor frame segmentation labels, a weighted propagation mechanism is used to generate high-dimensional features of the query frame; the calculation formula is:

[0108] ;

[0109] In the formula, The segmentation label representing the anchor frame. Indicates the location of the query frame. Predicted high-dimensional features; the embodiments of this application employ a weighted propagation mechanism to integrate global information, which is particularly suitable for processing structures with continuous features such as the aorta.

[0110] Step 7: Based on high-dimensional features and query frame features, a spatial attention weight map is generated using a spatial attention mechanism; then, a channel attention feature map is generated using a channel attention mechanism; specifically including:

[0111] Step 701, based on high-dimensional features and query frame features A spatial attention weight map is generated using a spatial attention mechanism, expressed as follows:

[0112] ;

[0113] In the formula, To query frame features , High-dimensional features , Indicates generation 1×1 convolution of channel features, This represents the Sigmoid activation function. express The activation function guides the network to focus on spatially significant regions at the pixel level, thereby suppressing irrelevant background interference and improving the recognition accuracy of blurred boundary structures such as false cavities. This application performs explicit attention calibration before high- and low-level feature fusion, which effectively improves the spatial representation purity of the feature map and lays a structural foundation for subsequent channel modeling.

[0114] Step 702: Multiply the spatial attention weight map and the query frame features to obtain the low-level feature attention map. The calculation formula is:

[0115] ;

[0116] Step 703: Use a channel attention mechanism to filter low-level feature attention maps. The channel attention weights are obtained through processing. The calculation formula is:

[0117] ;

[0118] In the formula, This indicates a global average pooling operation. This represents the global max pooling operation. Indicates that the core is One-dimensional convolution;

[0119] Step 704: Filter the low-level feature attention map. and channel attention weights Multiplication yields the channel attention feature map. The calculation formula is:

[0120] ;

[0121] In this step, this application processes the input features using a sum of global max pooling and global average pooling to generate a more informative global feature map. Furthermore, since reducing the dimensionality of the global feature map would degrade model performance, a one-dimensional convolution with a kernel size of k is used to achieve cross-channel interaction of the global feature map without reducing its dimensionality, thus achieving this goal without increasing computational cost. The main objective of the channel attention mechanism employed in this application is to further enhance the structure-related semantic representation and strengthen the model's responsiveness to key channels by modeling the dependencies between channels.

[0122] To improve segmentation performance in complex anatomical structures, this embodiment employs a decoder input layer that cascades spatial attention (SA) and channel attention (CA) to enhance the fusion of high- and low-level features. First, the SA module guides the model to focus on key regions, reducing background interference. Then, the CA module further filters significant channels, improving the ability to express semantic elements. This approach not only fully leverages the complementarity of the two attention mechanisms in spatial localization and semantic enhancement but also avoids information conflict and redundant weighting, providing more accurate feature representations for subsequent segmentation decisions. This bottom-up attention nesting strategy significantly improves model performance in false cavity region boundary recognition, exhibiting stronger robustness, especially in structurally complex or adherent regions.

[0123] Step 8: Input the channel attention feature map into the decoder to obtain the segmentation label of the query frame, and iterate the new anchor frame according to the segmentation label, then return to step 4 until all 2D CT images have been iterated, and output the segmentation label; then generate the complete prediction sequence through weighted fusion, the expression is:

[0124] .

[0125] Example 2:

[0126] An automatic segmentation system is provided to implement the automatic segmentation method for aortic dissection images in this embodiment. The automatic segmentation system for aortic dissection images in this application uses a Space-Time Correspondence Network (STCN) as the segmentation model and integrates spatial attention and channel attention at the input of the decoder to optimize feature fusion. It also introduces an anchor point random selection mechanism to store anchor point frames in a memory pool and combines a bidirectional propagation mechanism to match query frames, thereby establishing a dense pixel-level correspondence between video frames and achieving efficient propagation of label information. The automatic segmentation system for aortic dissection images includes:

[0127] The data acquisition and processing unit is used to acquire labeled 3D CT images and convert the labeled 3D CT images into a series of labeled 2D CT images.

[0128] The anchor frame selection unit is connected to the data acquisition and processing unit. It is used to perform morphological restoration on the two-dimensional CT image sequence to bridge the label break points and discontinuous areas, group two-dimensional CT images with the same category of label combination into a subset, and randomly select a frame from the subset containing all category of label combinations as the anchor frame.

[0129] The memory bank, connected to the anchor frame selection unit, is used to store the 2D CT image of the current anchor frame and its corresponding segmentation label;

[0130] The bidirectional propagation unit, connected to the anchor frame selection unit and the data acquisition and processing unit, is used to match query frames from a two-dimensional CT image sequence using a bidirectional propagation mechanism.

[0131] A feature extraction unit is configured with a feature extractor. The feature extraction unit is connected to a bidirectional propagation unit. In this embodiment, the feature extractor includes a query encoder, an anchor encoder, and a value encoder. The query encoder is used to extract features from a query frame to obtain query frame features. The anchor encoder is used to extract features from anchor frames in the memory to obtain anchor frame features. The value encoder is used to calculate value features from anchor segmentation labels.

[0132] The query frame high-dimensional feature generation unit is connected to the feature extraction unit and the memory bank. It adopts the key affinity calculation method, calculates the similarity matrix based on the query frame features and anchor frame features based on the skip connection, and generates the high-dimensional features of the query frame through a weighted propagation mechanism based on the similarity matrix and the value features corresponding to the segmentation labels of the anchor frames in the memory bank.

[0133] A joint attention unit is connected to a feature extraction unit and a query frame high-dimensional feature generation unit. The joint attention unit includes a cascaded spatial attention unit and a channel attention unit. It is used to generate a spatial attention weight map based on high-dimensional features and query frame features using a spatial attention mechanism; then, it generates a channel attention feature map using a channel attention mechanism.

[0134] The decoder, connected to the joint attention module, is used to decode the channel attention feature map to generate segmentation labels for the query frame. The output of the decoder is connected to the memory bank to iterate new anchor frames based on the segmentation labels of the query frame.

[0135] After constructing the automatic segmentation system for aortic dissection images according to this application, a loss function is constructed, and the automatic segmentation system for aortic dissection images is trained by acquiring three-dimensional CT images as a dataset. After training, in the testing phase, a frame of a two-dimensional CT image sequence is selected as an anchor frame and added to the memory bank. The preceding and following frames adjacent to the anchor frame are input into the automatic segmentation system for aortic dissection images to perform video target segmentation. In this embodiment, the cross-entropy loss function is used to measure the difference between the predicted segmentation label and the true label. Therefore, the loss function L of the method of this invention uses the cross-entropy loss function.

[0136] In the description of the embodiments of this application, it should be noted that the terms "inner" and "outer" and other terms indicating direction or positional relationship are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and does not indicate or imply that the device or component must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this application.

[0137] In the description of this application, the references to terms such as "an embodiment," "some embodiments," "in this embodiment," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0138] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An automatic segmentation method of an aortic dissection image, characterized by, The method comprises the following steps: Step 1, obtaining a three-dimensional CT image with labels, processing to obtain a two-dimensional CT image sequence composed of two-dimensional CT images with labels, the labels including three label categories of an aortic wall label, a true cavity label and a false cavity label; Step 2, identifying and repairing broken points and discontinuous regions of labels in each two-dimensional CT image by using morphological modeling, and grouping two-dimensional CT images with the same label category combination into a subset; Step 3, randomly selecting a two-dimensional CT image from the subset containing all label categories as an anchor frame; Step 4, based on the anchor frame, combining a bidirectional propagation mechanism to match image frames adjacent to the anchor frame as query frames, and using a feature extractor to extract anchor frame features, query frame features and anchor frame segmentation label features; Step 5, calculating a similarity matrix of the query frame features and the anchor frame features; Step 6, based on the similarity matrix and the anchor frame segmentation label features, using a weighted propagation mechanism to generate high-dimensional features of the query frame; Step 7, based on the high-dimensional features and the query frame features, using a spatial attention mechanism to generate a spatial attention weight map, and then using a channel attention mechanism to generate a channel attention feature map; Step 8, inputting the channel attention feature map into a decoder to obtain the segmentation label of the query frame, and iteratively obtaining a new anchor frame according to the segmentation label, and returning to step 4.

2. The method of automatic segmentation of aortic dissection imagery according to claim 1, wherein, The step 1 specifically comprises: Step 101, obtaining a three-dimensional CT image with labels; Step 102, performing resampling, standardization and noise reduction processing on the three-dimensional CT image with labels; Step 103, extracting continuous slices along the Z-axis direction of the three-dimensional CT image to obtain a plurality of two-dimensional CT images with labels to form a two-dimensional CT image sequence.

3. The method of claim 1, wherein the method further comprises: The step 2 specifically comprises: Step 201, obtaining a binary mask of each label category in each two-dimensional CT image; Step 202, identifying and repairing broken points and discontinuous regions of labels in the binary mask by using morphological modeling; Step 203, based on the repaired binary mask, identify the label class combination on each frame of two-dimensional CT image, the first The label class combination of the frame of two-dimensional CT image is represented as: ； wherein represents the first a binary mask of the label class in the frame two-dimensional CT image, represents the spatial pixel coordinate; represents the aorta wall label, represents the true lumen label, represents the false lumen label; Step 204, dividing the two-dimensional CT images having the same label class combination into subsets .​ 4. The method of automatic segmentation of aortic dissection imagery of claim 3, wherein, The step 3 specifically comprises: Step 301, define a subset containing all label categories as a candidate anchor set The expression is: ； In the formula, denotes an anchor frame randomly selected from the candidate anchor set denotes an anchor frame randomly selected from the candidate anchor set denotes a label class combination; Step 302, initializing a label propagation process, represented as: ； In the formula, represents a network propagation mechanism, from the anchor frame to the query frame propagates the segmentation result, represents the anchor frame segmentation label.

5. The method of claim 3, wherein the step of automatically segmenting the aortic dissection image is characterized by, The feature extractor in step 4 includes an initial convolutional layer and four residual layers connected in sequence, the step length of the initial convolutional layer is 1; and the channel configuration parameters of the feature extractor are modified according to the gray characteristics of the two-dimensional CT image.

6. The method of claim 3, wherein the step of automatically segmenting the aortic dissection image is characterized by, The bidirectional propagation mechanism in step 4 comprises: Forward propagation: from anchor frame to subsequent frames propagating label features; backpropagating, from anchor frames to previous frames propagating label features; Then, a complete prediction sequence is generated by weighted fusion.

7. The method of automatic segmentation of aortic dissection imagery of claim 6, wherein, The expression for calculating the similarity matrix of the query frame features and the anchor frame features in step 5 is: ； In the formula, The query frame characteristics are the location of the query frame. The feature vector at that location, The anchor frame feature is the position of the anchor frame. eigenvectors; This represents the cosine similarity between the query frame features and the anchor frame features.

8. The method of automatic segmentation of aortic dissection imagery of claim 7, wherein, The calculation formula for generating high-dimensional features of the query frame by using the weighted propagation mechanism in step 6 is: ； wherein denotes the segmentation label of the anchor frame, denotes the position of the query frame in the sequence, predicted high-dimensional features.

9. The method of automatic segmentation of aortic dissection imagery of claim 8, wherein, The step 7 specifically comprises: Step 701, based on high-dimensional features and query frame features , a spatial attention mechanism is used to generate a spatial attention weight map ; Step 702, multiply the spatial attention weight map and the query frame feature to obtain a filtered low-level feature attention map ; At step 703, a channel attention mechanism is used to filter the low-level feature attention map The channel attention weight is obtained by processing The calculation formula is: ； wherein denotes a global average pooling operation, denotes a global max pooling operation, denotes a one-dimensional convolution with kernel ; Step 704, filter low-level feature attention map and channel attention weight multiply to get channel attention feature map .

10. An automated segmentation system, characterized by, Comprise: A data acquisition and processing unit for converting the obtained three-dimensional CT image with labels into a two-dimensional CT image sequence with labels; An anchor frame selection unit connected with the data acquisition and processing unit, for performing morphological identification and label repair on the two-dimensional CT image sequence, grouping two-dimensional CT images with the same label category combination into a subset, and randomly selecting an anchor frame from the subset containing all label category combinations; A memory bank connected with the anchor frame selection unit, for storing the two-dimensional CT image of the current anchor frame and its segmentation label; A data acquisition and processing unit for converting the obtained three-dimensional CT image with labels into a two-dimensional CT image sequence with labels; An anchor frame selection unit connected with the data acquisition and processing unit, for performing morphological identification and label repair on the two-dimensional CT image sequence, grouping two-dimensional CT images with the same label category combination into a subset, and randomly selecting an anchor frame from the subset containing all label category combinations; A memory bank connected with the anchor frame selection unit, for storing the two-dimensional CT image of the current anchor frame and its segmentation label; The bidirectional propagation unit is connected with the anchor point frame selection unit and the data acquisition and processing unit, and is configured to match the query frame adjacent to the anchor point frame by using a bidirectional propagation mechanism. The feature extraction unit is connected with the bidirectional propagation unit, and is configured to configure a feature extractor and extract anchor point frame features, query frame features and anchor point frame segmentation label features. The query frame high-dimensional feature generation unit is connected with the feature extraction unit and the memory bank, and is configured to calculate a similarity matrix and generate high-dimensional features of the query frame by using a weighted propagation mechanism. The joint attention unit is connected with the feature extraction unit and the query frame high-dimensional feature generation unit, and includes cascaded spatial attention units and channel attention units, and is configured to generate a channel attention feature map based on a spatial attention mechanism and a channel attention mechanism. The decoder is connected with the joint attention module, and is configured to decode the channel attention feature map to generate segmentation labels of the query frame. An output end of the decoder is connected with the memory bank, and is configured to iterate new anchor point frames.

Images