Aortic arch coarctation image segmentation method based on deep learning

Abstract

The present application relates to the technical field of coarctation of the aortic arch image segmentation, and discloses an aortic arch coarctation image segmentation method based on deep learning, which comprises the following steps: preprocessing aortic CT angiography three-dimensional images to obtain standardized three-dimensional images; acquiring key anatomical landmark point coordinates through a 3D encoding-decoding structure anatomical landmark point detection network, generating an aortic arch blood vessel centerline, constructing a tubular region of interest, and cutting to obtain an isthmus focusing image of the focusing isthmus lesion; using a double-branch collaborative segmentation network to extract overall vascular topological features and high-resolution vascular edge detail features, weighting and enhancing isthmus edge features, and outputting an initial three-dimensional segmentation result; calculating point-by-point pipe diameter based on the initial segmentation result, identifying pipe diameter mutation regions, and correcting the coarctation segment boundary in reverse to obtain a final segmentation result; and the present application solves the problems of the prior art, such as large background interference in aortic arch coarctation image segmentation, easy missed segmentation of the tiny coarctation of the isthmus, and insufficient accuracy of the blood vessel boundary.

Description

Technical Field

[0001] This invention relates to the field of aortic arch constriction image segmentation technology, and more particularly to a deep learning-based aortic arch constriction image segmentation method. Background Technology

[0002] Aortic arch coarctation is a common congenital cardiovascular malformation. Precise segmentation of blood vessels and lesions in 3D CT angiography is a core prerequisite for quantitatively assessing the degree of coarctation and developing preoperative treatment plans. Existing deep learning-based aortic segmentation technologies are mostly designed for the entire aorta or aortic dissection lesions, and are not adapted to the specific anatomical features of aortic arch coarctation, resulting in significant shortcomings in segmentation performance.

[0003] Chinese invention patent application with publication number CN117474878A uses a main network and sub-network collaborative architecture to achieve full-scale segmentation of the aorta. However, it does not introduce key anatomical landmarks of the aortic arch as prior guidance, and cannot achieve targeted feature enhancement for small narrowing areas in the isthmus. The risk of lesion missed is high. Furthermore, it does not combine the anatomical constraints of the vessel diameter to correct the segmentation boundary, and the boundary accuracy of the narrowing segment is difficult to meet the needs of clinical quantitative assessment.

[0004] Chinese invention patent application CN113947609A describes a segmentation network designed with sequence feature pyramid attention for aortic dissection lesions. However, it directly performs segmentation processing on the whole chest image without removing interference from the complex background of the thoracic cavity through region of interest cropping. The overall computational efficiency is low, and it does not design a dedicated segmentation architecture for the tubular anatomical characteristics of the aortic arch. The segmentation results are prone to vascular topology breakage and cannot adapt to the fine segmentation scenario of aortic arch stenosis.

[0005] In summary, existing technologies generally suffer from insufficient background interference suppression, weak response to isthmus narrowing lesion features, insufficient segmentation boundary accuracy, and poor vascular topological integrity, making it difficult to support accurate clinical diagnosis and quantitative assessment of aortic arch narrowing. Summary of the Invention

[0006] The present invention aims to provide a deep learning-based aortic arch constriction image segmentation method to solve the problems of aortic arch constriction image segmentation being greatly affected by background interference, easy omission of small constrictions in the isthmus, and insufficient accuracy of vessel boundaries in the prior art.

[0007] To achieve the above objectives, the present invention provides the following method:

[0008] The deep learning-based aortic arch constriction image segmentation method provided by this invention is as follows:

[0009] S1. Obtain the three-dimensional image of the aortic CT angiography to be segmented, and perform preprocessing on the three-dimensional image of the aortic CT angiography to obtain a standardized three-dimensional image.

[0010] S2. Input the standardized three-dimensional image into the pre-trained anatomical landmark detection network and output the coordinates of key anatomical landmarks of the aortic arch; the key anatomical landmarks include the opening point of the brachiocephalic trunk, the opening point of the left common carotid artery, the opening point of the left subclavian artery, the proximal landmark of the isthmus, and the distal landmark of the isthmus.

[0011] S3. Extract the aortic arch vessel centerline based on the coordinates of the key anatomical landmarks, generate a tubular region of interest along the aortic arch vessel centerline, and crop the isthmus focused image corresponding to the tubular region of interest from the standardized three-dimensional image;

[0012] S4. Input the focused image of the isthmus into a pre-trained dual-branch collaborative segmentation network, extract the overall topological features of the blood vessel through the tubular structure branch of the dual-branch collaborative segmentation network, and extract the edge detail features of the blood vessel through the boundary detail branch of the dual-branch collaborative segmentation network.

[0013] S5. Generate a spatial attention weight map based on the aortic arch centerline, and apply the spatial attention weight map to the feature map output by the boundary detail branch to weight and enhance the edge features of the isthmus region.

[0014] S6. Perform multi-scale feature fusion on the weighted enhanced vascular edge detail features and the overall vascular topology features to output the initial three-dimensional segmentation results of the aortic arch and the narrowing region.

[0015] S7. Based on the initial three-dimensional segmentation results, extract the vessel segmentation centerline and calculate the point-by-point vessel diameter. Based on the abrupt change in vessel diameter, reverse the narrowing segment boundary of the initial three-dimensional segmentation results to obtain the final aortic arch narrowing image segmentation result.

[0016] Furthermore, the preprocessing operation is performed in the order of resampling, gray-level truncation, background removal, and gray-level normalization. Specifically, the aortic CT angiography three-dimensional image is resampled to the same voxel spacing through trilinear interpolation, the image gray-level values ​​are truncated to the effective HU interval of the vascular tissue, and gray-level values ​​outside the interval are mapped to the interval endpoints. Threshold segmentation combined with three-dimensional morphological opening operation is used to remove non-vascular background areas of the bed plate and bones. Finally, gray-level linear normalization is performed on the remaining image to map the gray-level values ​​to the [0,1] interval to obtain the standardized three-dimensional image.

[0017] Furthermore, the anatomical landmark detection network is a 3D encoder-decoder keypoint detection model. The encoder of the network contains four sets of parallel 3D convolutional branches corresponding to different downsampling rates. Adjacent branches interact bidirectionally through cross-resolution convolution, maintaining high-resolution feature output throughout. The decoder adopts a stepwise trilinear upsampling structure, and fuses the corresponding resolution features of the encoder through skip connections, finally outputting a 5-channel heatmap with the same size as the standardized 3D image. During network training, a 3D Gaussian heatmap is generated centered on the marked landmarks as a supervision label. During inference, the 3D coordinates corresponding to the maximum probability of each channel heatmap are extracted to obtain the coordinates of the key anatomical landmarks.

[0018] Furthermore, the aortic arch vascular centerline is generated as follows: Frangi 3D vascular enhancement is performed on the standardized 3D image; the vascular probability of each voxel is calculated based on the eigenvalues ​​of the 3D Hessian matrix; a vascular cost map negatively linearly correlated with the vascular probability value is constructed; following the physiological course of the aortic arch from proximal to distal, the path starts at the brachiocephalic trunk opening point, and sequentially uses the left common carotid artery opening point, the left subclavian artery opening point, and the proximal isthmus marker as the path endpoint, ending at the distal isthmus marker, and the complete aortic arch vascular centerline is obtained by stitching together the path using the 3D Dijkstra shortest path algorithm; the tubular region of interest is generated with the aortic arch vascular centerline as the axis and a pre-defined ratio of the average diameter of the aortic arch as the cross-sectional radius to generate a 3D tubular mask; the isthmus focused image is the effective region image obtained by element-wise multiplication of the standardized 3D image and the 3D tubular mask.

[0019] Furthermore, the tubular structure branch of the dual-branch collaborative segmentation network adopts a four-level downsampling encoder structure. Each encoder level includes two levels of three-dimensional convolution, batch normalization layer, ReLU activation layer and max pooling layer with stride of 2. Three-dimensional dilated convolution with progressively increasing dilation rate is embedded between the two levels of convolution to expand the receptive field without reducing feature resolution and extract the overall topological features of the blood vessels, including vessel orientation, branch position and tubular morphology.

[0020] Furthermore, the boundary detail branch of the dual-branch collaborative segmentation network is composed of multi-level cascaded three-dimensional convolutional units, without performing downsampling throughout, and the feature resolution is consistent with the original resolution of the isthmus focused image; the spatial attention weight map is constructed with the aortic arch vessel centerline as the center, and a two-dimensional Gaussian distribution is constructed on a radial plane perpendicular to the centerline tangent, with the standard deviation of the Gaussian distribution being a set proportion of the preset average diameter of the aortic arch, and the center weight gradually decreases radially; the center weight of the isthmus vessel segment is multiplied by a gain coefficient to enhance the response, and the weighted enhancement is achieved by element-wise multiplication of the spatial attention weight map and the feature map.

[0021] Furthermore, the specific method of multi-scale feature fusion is as follows: the downsampled features of the four different scales output by the tubular structure branch are upsampled to the original resolution of the isthmus focused image through trilinear interpolation; the upsampled features of each scale are stitched together with the single-scale features output by the boundary detail branch along the channel dimension; after channel dimensionality reduction and feature integration are completed by a 1×1×1 three-dimensional convolutional layer, the features are upsampled to the original resolution step by step; and the segmentation probability of each voxel is output by the sigmoid activation function to form the initial three-dimensional segmentation result.

[0022] Furthermore, the vessel segmentation centerline is extracted using a three-dimensional thinning algorithm that maintains 26-neighbor topological connectivity; the point-by-point diameter is obtained by statistically analyzing the area of ​​the segmented vessel region on a cross-section perpendicular to the centerline tangent and calculating the equivalent circle diameter; the diameter mutation region is determined as a vessel segment where the absolute value of the rate of change of diameter at multiple consecutive centerline points exceeds a preset threshold. During correction, a preset search range is expanded radially with the centerline of the mutation region as the axis, and the gray-level radial gradient value is calculated within the range. The vessel boundary is corrected at the position with the maximum absolute gradient value.

[0023] Furthermore, both the anatomical landmark detection network and the bi-branch collaborative segmentation network are trained using a CT angiography image dataset labeled with aortic arch narrowing. During training, the input images are subjected to elastic deformation data enhancement through random cropping, random rotation, and B-spline free deformation. The bi-branch collaborative segmentation network uses a weighted combination of Dice loss and cross-entropy loss as the basic loss function, and adds a diameter consistency loss term to constrain the segmentation accuracy of the narrowed region.

[0024] Furthermore, before obtaining the final aortic arch stenosis image segmentation result, topology optimization post-processing is performed on the initial three-dimensional segmentation result; after 26-neighbor three-dimensional connected domain analysis, the largest aortic vessel connected domain is retained, and isolated mis-segmented regions with a volume ratio less than the preset proportion are removed; the endpoints of vascular branches are extracted and the branch lengths are counted. When the number of branches exceeds the normal anatomical range of the aortic arch, the shortest abnormal branch is removed to ensure that the final segmentation result conforms to the anatomical morphology of the aortic arch.

[0025] The beneficial effects of this invention are as follows: By combining anatomical landmark detection with tubular region of interest clipping guided by the aortic arch centerline, this invention effectively eliminates interference from the complex thoracic tissue background and accurately focuses on the isthmus lesion area, improving segmentation efficiency while enhancing the feature attention of the lesion area; it employs a dual-branch collaborative segmentation network with a centerline-driven spatial attention mechanism, simultaneously considering both the overall tubular topology of the aortic arch and the detailed features of the isthmus edge, specifically enhancing the feature response of the narrowed area and significantly reducing the probability of missing small narrowed lesions in the isthmus; supplemented by boundary correction based on vessel diameter feedback and anatomical topological constraint post-processing, it further improves the boundary segmentation accuracy of the narrowed segment, ensuring the integrity and continuity of the vascular tubular structure, and significantly improving the accuracy and robustness of aortic arch narrowing image segmentation, meeting the clinical needs for precise quantitative assessment of the degree of narrowing. Attached Figure Description

[0026] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0027] Figure 1 A schematic diagram of the aortic arch constriction image segmentation method based on deep learning provided in an embodiment of the present invention. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0030] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0031] Existing deep learning-based aortic segmentation technologies are mostly designed for the entire aorta or aortic dissection lesions, and are not adapted to the specific anatomical features of aortic arch constriction, resulting in significant shortcomings in segmentation performance.

[0032] Chinese invention patent application with publication number CN117474878A uses a main network and sub-network collaborative architecture to achieve full-scale segmentation of the aorta. However, it does not introduce key anatomical landmarks of the aortic arch as prior guidance, and cannot achieve targeted feature enhancement for small narrowing areas in the isthmus. The risk of lesion missed is high. Furthermore, it does not combine the anatomical constraints of the vessel diameter to correct the segmentation boundary, and the boundary accuracy of the narrowing segment is difficult to meet the needs of clinical quantitative assessment.

[0033] Chinese invention patent application CN113947609A describes a segmentation network designed with sequence feature pyramid attention for aortic dissection lesions. However, it directly performs segmentation processing on the whole chest image without removing interference from the complex background of the thoracic cavity through region of interest cropping. The overall computational efficiency is low, and it does not design a dedicated segmentation architecture for the tubular anatomical characteristics of the aortic arch. The segmentation results are prone to vascular topology breakage and cannot adapt to the fine segmentation scenario of aortic arch stenosis.

[0034] In summary, existing technologies generally suffer from insufficient background interference suppression, weak response to isthmus narrowing lesion features, insufficient segmentation boundary accuracy, and poor vascular topological integrity, making it difficult to support accurate clinical diagnosis and quantitative assessment of aortic arch narrowing.

[0035] The present invention aims to provide a deep learning-based aortic arch constriction image segmentation method to solve the problems of aortic arch constriction image segmentation being greatly affected by background interference, easy omission of small constrictions in the isthmus, and insufficient accuracy of vessel boundaries in the prior art.

[0036] like Figure 1 As shown in the figure, the specific embodiments of the present invention provide a deep learning-based aortic arch constriction image segmentation method, the specific steps of which are as follows:

[0037] S1: Acquire the three-dimensional CT angiography image of the aorta to be segmented, perform preprocessing operations on the three-dimensional CT angiography image of the aorta to obtain a standardized three-dimensional image.

[0038] In this embodiment of the invention, the preprocessing operation is performed sequentially in the order of resampling, gray-level truncation, background removal, and gray-level normalization. First, the three-dimensional aortic CT angiography image is resampled to an isotropic voxel spacing of 1mm×1mm×1mm using trilinear interpolation to unify the image spatial resolution and eliminate voxel size differences between different scanning devices. Then, gray-level truncation is performed on the resampled image, setting the effective range of vascular tissue to [-100, 500] HU, and uniformly assigning gray-level values ​​below -100 HU to -100 HU. Gray values ​​higher than 500 HU were uniformly assigned a value of 500 HU to suppress the interference of extreme gray values ​​from lung tissue, bone, etc., on the segmentation results. Next, a preliminary threshold segmentation was performed using 150 HU as the gray value threshold to extract high-density tissue regions from the image. Then, a three-dimensional morphological opening operation was performed using spherical structuring elements with a radius of 2 mm. The operation involved erosion followed by dilation to break the weak connections between blood vessels and bones / bed lamina, and to remove non-vascular background regions corresponding to bed lamina and bones. Finally, the remaining effective image regions were subjected to gray-level linear normalization processing according to the formula:

[0039] ;

[0040] The grayscale values ​​are linearly mapped to the [0,1] interval, where The original grayscale value after truncation. -100HU The value is 500HU, and a standardized 3D image is finally obtained.

[0041] S2: Input standardized 3D images into a pre-trained anatomical landmark detection network and output the coordinates of key anatomical landmarks of the aortic arch; key anatomical landmarks include the opening point of the brachiocephalic trunk, the opening point of the left common carotid artery, the opening point of the left subclavian artery, the proximal landmark of the isthmus, and the distal landmark of the isthmus.

[0042] In this embodiment of the invention, the anatomical landmark detection network is a 3D encoder-decoder keypoint detection model. The network's encoder contains four sets of parallel 3D convolutional branches, corresponding to downsampling resolutions of 1x, 2x, 4x, and 8x of the input image, respectively. The initial number of channels is set to 16, doubling with each downsampling pass. Branches at adjacent resolutions interact bidirectionally through a 1×1×1 cross-resolution 3D convolutional layer at each encoding stage, achieving continuous fusion of multi-scale features and maintaining high-resolution feature output throughout, ensuring the accuracy of landmark localization. The network's decoder employs a progressive trilinear upsampling structure with four upsampling levels. After each upsampling level, features from the encoder's corresponding resolution branches are fused via skip connections, ultimately restoring the feature resolution to match the standardized 3D image, with an output channel count of 5. During the network training phase, a training set of 200 annotated clinical CT angiography images of aortic arch constriction was used. A 3D Gaussian heatmap with a standard deviation of 3mm was generated, centered on the coordinates of manually annotated key anatomical landmarks, as the supervisory label. Iterative training was performed using the mean squared error loss function and the Adam optimizer, with an initial learning rate set to... The system was trained for 100 rounds, with an early stopping strategy used to prevent overfitting. During the network inference phase, a 5-channel heatmap with the same size as the standardized 3D image was output. Each channel corresponds to the probability distribution of a key anatomical landmark. The coordinates of the key anatomical landmark were obtained by extracting the 3D voxel coordinates corresponding to the maximum probability in each channel's heatmap.

[0043] S3: Extract the aortic arch centerline based on the coordinates of key anatomical landmarks, generate a tubular region of interest along the aortic arch centerline, and crop the isthmus focused image corresponding to the tubular region of interest from the standardized 3D image.

[0044] In this embodiment of the invention, the aortic arch centerline is generated as follows: First, Frangi 3D vascular enhancement processing is performed on the standardized 3D image. During processing, the Gaussian smoothing factor σ is set to a range of [1mm, 5mm] with a step size of 1mm. A 3D Hessian matrix is ​​calculated for each voxel, and three eigenvalues ​​are obtained. Based on the eigenvalue distribution characteristics of the tubular structure, the vascular similarity function value corresponding to each voxel is calculated to obtain the probability value of the vascular structure corresponding to each voxel, generating a vascular probability map consistent with the image size. After normalizing the probability values ​​in the vascular probability map to the [0,1] interval, the vascular cost map is calculated using the formula "cost value = 1 - normalized vascular probability value". Regions with higher vascular probability have lower cost values ​​and are given priority during path tracing.

[0045] Following the physiological course of the aortic arch from proximal to distal, the path starts at the brachiocephalic trunk opening, and sequentially sets the openings of the left common carotid artery, left subclavian artery, and the proximal isthmus as transit points. Finally, the path terminates at the distal isthmus. The three-dimensional Dijkstra shortest path algorithm is used to trace the optimal vascular path between adjacent landmarks. After splicing the paths, the complete aortic arch centerline is obtained, ensuring that the centerline conforms to the actual vascular anatomy. The tubular region of interest is defined with the aortic arch centerline as the axis. The average aortic arch diameter is pre-defined as 15 mm. A circular cross-section perpendicular to the tangent of the centerline is generated at each centerline point, with a radius set to 1.5 times the average diameter. All voxel regions covered by these cross-sections are combined to generate a 3D tubular mask. The standardized 3D image is then multiplied element-wise with the 3D tubular mask, retaining the grayscale values ​​of the mask-covered areas and assigning zero values ​​to the remaining areas. The minimum bounding box of the 3D tubular mask is then extracted. The multiplied image is cropped based on the bounding box coordinates, removing all-zero regions on the periphery to obtain a reduced-size focused image of the isthmus. This process removes background interference while reducing the computational load of subsequent segmentation networks.

[0046] S4: Input the focused image of the isthmus into the pre-trained bi-branch collaborative segmentation network. Extract the overall topological features of the blood vessel through the tubular structure branch of the bi-branch collaborative segmentation network, and extract the edge detail features of the blood vessel through the boundary detail branch of the bi-branch collaborative segmentation network.

[0047] In this embodiment of the invention, the tubular structure branch of the dual-branch collaborative segmentation network adopts a four-level downsampling encoder structure. The initial number of channels in the network is 32. Each level encoder consists of two 3×3×3 convolutional layers, a batch normalization layer, a ReLU activation layer, and a max pooling layer with a stride of 2. The number of channels doubles every time it is downsampled. A 3×3×3 dilated convolutional layer is embedded between the two convolutional layers of each level encoder. The dilation rates of the four levels encoder are 2, 4, 6, and 8, respectively. Through dilated convolution, the receptive field is gradually expanded without reducing the feature resolution, capturing the global morphological information of the aortic arch and extracting the overall topological features of the blood vessels, including the aortic arch vessel orientation, branch positions, and overall tubular morphology.

[0048] The boundary detail branch consists of three cascaded 3×3×3 convolutional layers, batch normalization layers, and ReLU activation layers. No downsampling is performed throughout, and the feature resolution remains consistent with the original resolution of the isthmus-focused image. The output channel count is 32, used to extract high-resolution vascular edge detail features, accurately capturing the boundary information of the narrowed region. During network training, the aortic arch narrowing CT angiography image dataset, which is from the same source as the marker detection network, is used. The annotation content includes 3D segmentation labels for the aortic arch vascular region and the narrowed region. During training, the input images undergo random cropping, random rotation, and B-spline free deformation elastic deformation data augmentation. The random cropping size is 96×96×96 voxels, and the random rotation angle range is... The elastic deformation control grid spacing is 16×16×16 voxels, and the deformation coefficient is 0.2. The Dice loss and cross-entropy loss are combined in a 1:1 weighted ratio as the basic loss function, with an additional pipe diameter consistency loss term constraining the segmentation results of the narrowed isthmus region. The pipe diameter consistency loss is obtained by calculating the mean square error between the point-by-point pipe diameter of the isthmus region in the predicted segmentation results and the point-by-point pipe diameter of the corresponding region in the labeled segmentation results. The weight coefficient of the pipe diameter consistency loss is 0.3. Iterative training is performed using the Adam optimizer, with an initial learning rate set to... The model was trained for a total of 200 rounds, and a learning rate decay strategy was used during training to improve the model's convergence performance.

[0049] S5: Generate a spatial attention weight map based on the aortic arch centerline, apply the spatial attention weight map to the feature map output by the boundary detail branch, and enhance the edge features of the isthmus region with weight.

[0050] In this embodiment of the invention, the spatial attention weight map is a three-dimensional weight map with the same size as the isthmus focused image. It is generated as follows: A two-dimensional Gaussian distribution is constructed on a radial plane perpendicular to the tangent direction of the centerline, centered on each voxel on the aortic arch centerline. Using a preset average aortic arch diameter of 15 mm as a benchmark, the standard deviation of the Gaussian distribution is set to 0.8 times the average diameter. The weight value at the center of the Gaussian distribution is 1, and the weight value gradually decreases radially outward with the Gaussian distribution. Simultaneously, for the isthmus segment between the proximal and distal landmarks, the center weight value of the corresponding Gaussian distribution is multiplied by a gain coefficient of 1.5. The weight map is then normalized to obtain the final spatial attention weight map. Weighted enhancement is achieved by element-wise multiplying the spatial attention weight map with the feature map output from the boundary detail branch in the spatial dimension. This allows the network to focus on the features of the vessel edge and the isthmus lesion area, strengthening the feature response intensity of the narrowed region.

[0051] S6: Perform multi-scale feature fusion on the weighted enhanced vascular edge details and overall vascular topology features to output the initial three-dimensional segmentation results of the aortic arch and constriction region.

[0052] In this embodiment of the invention, the specific implementation of multi-scale feature fusion is as follows: the tubular structure branch outputs downsampled features at four scales, with feature sizes of 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the input image size, respectively; the downsampled features at the four scales are upsampled to the original resolution of the isthmus focused image through trilinear interpolation, so that the topological features at each scale are consistent with the boundary detail features at the single scale; the topological features after upsampling at each scale are concatenated with the weighted enhanced vascular edge detail features along the channel dimension to obtain four sets of fused features; the four sets of fused features are concatenated and integrated along the channel dimension, and then feature refinement and channel dimensionality reduction are performed through two sets of cascaded 3×3×3 convolutional layers; finally, the probability value of each voxel belonging to the aortic arch and narrowing region is output through the sigmoid activation function; the probability map is binarized with a probability threshold of 0.5, and voxels with a probability value greater than or equal to 0.5 are determined to be vascular regions, forming the initial three-dimensional segmentation result.

[0053] S7: Extract the vessel segmentation centerline based on the initial three-dimensional segmentation results and calculate the point-by-point vessel diameter. Correct the narrowing segment boundary of the initial three-dimensional segmentation results in reverse according to the aortic arch narrowing region to obtain the final aortic arch narrowing image segmentation result.

[0054] In this embodiment of the invention, the vessel segmentation centerline is extracted using an iterative erosion-based topology-preserving 3D thinning algorithm. During the extraction process, boundary voxels of the vessel segmentation region are removed round by round, and each removal operation maintains the topological structure of the vessel region by determining the connectivity of the 26-neighborhood, thus avoiding vessel breakage during thinning, until a skeleton of a single voxel width is obtained as the vessel segmentation centerline. The point-by-point diameter is calculated as follows: at each point on the vessel segmentation centerline, a vessel cross-section perpendicular to the tangent direction of the centerline is extracted, and the total area of ​​voxels belonging to the vessel segmentation region on the cross-section is counted. The area of ​​the equivalent circle is then calculated using the formula... Calculate the equivalent circle radius r, and take twice the radius as the point-by-point diameter. The rule for determining the diameter abrupt change region is as follows: calculate the ratio of the diameter difference between adjacent centerline points to the diameter of the previous point as the diameter change rate. When the absolute value of the diameter change rate of three or more consecutive centerline points exceeds 20%, the continuous blood vessel segment is determined to be a diameter abrupt change region. During correction, take the centerline corresponding to the diameter abrupt change region as the axis and expand the search range of three voxels radially outward. Within the search range, calculate the first difference of gray level in the radial direction as the radial gradient value. Take the position with the largest absolute value of gray level gradient as the corrected blood vessel boundary and replace the segmentation boundary of the corresponding region in the initial 3D segmentation result to correct the boundary offset problem in the initial segmentation. After boundary correction, topology optimization post-processing was performed on the segmentation results: all three-dimensional connected domains in the segmentation results were analyzed and statistically analyzed using 26-neighbor three-dimensional connected domains, and only the main vessel connected domain with the largest volume was retained, while isolated mis-segmented regions with a volume less than 1% of the main vessel connected domain volume were removed; then, all vessel branch endpoints in the segmentation results were extracted, and the number of voxels along the centerline of each branch was counted as the branch length. When the number of branches exceeded the range of the three brachiocephalic branches in the normal anatomy of the aortic arch, the branch with the shortest branch length was identified as an abnormal branch region and removed, finally obtaining the aortic arch narrowing image segmentation result that conforms to the normal anatomical morphology of the aortic arch.

[0055] The beneficial effects of this invention are as follows: By combining anatomical landmark detection with tubular region of interest clipping guided by the aortic arch centerline, this invention effectively eliminates interference from the complex thoracic tissue background and accurately focuses on the isthmus lesion area, improving segmentation efficiency while enhancing the feature attention of the lesion area; it employs a dual-branch collaborative segmentation network with a centerline-driven spatial attention mechanism, simultaneously considering both the overall tubular topology of the aortic arch and the detailed features of the isthmus edge, specifically enhancing the feature response of the narrowed area and significantly reducing the probability of missing small narrowed lesions in the isthmus; supplemented by boundary correction based on vessel diameter feedback and anatomical topological constraint post-processing, it further improves the boundary segmentation accuracy of the narrowed segment, ensuring the integrity and continuity of the vascular tubular structure, and significantly improving the accuracy and robustness of aortic arch narrowing image segmentation, meeting the clinical needs for precise quantitative assessment of the degree of narrowing.

[0056] The above descriptions are merely embodiments of the present invention. Commonly known technical solutions or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A deep learning-based image segmentation method for aortic arch constriction, characterized in that, The method includes: S1. Obtain the three-dimensional image of the aortic CT angiography to be segmented, and perform preprocessing on the three-dimensional image of the aortic CT angiography to obtain a standardized three-dimensional image. S2. Input the standardized three-dimensional image into the pre-trained anatomical landmark detection network and output the coordinates of key anatomical landmarks of the aortic arch; the key anatomical landmarks include the opening point of the brachiocephalic trunk, the opening point of the left common carotid artery, the opening point of the left subclavian artery, the proximal landmark of the isthmus, and the distal landmark of the isthmus. S3. Extract the aortic arch vessel centerline based on the coordinates of the key anatomical landmarks, generate a tubular region of interest along the aortic arch vessel centerline, and crop the isthmus focused image corresponding to the tubular region of interest from the standardized three-dimensional image; S4. Input the focused image of the isthmus into a pre-trained dual-branch collaborative segmentation network, extract the overall topological features of the blood vessel through the tubular structure branch of the dual-branch collaborative segmentation network, and extract the edge detail features of the blood vessel through the boundary detail branch of the dual-branch collaborative segmentation network. S5. Generate a spatial attention weight map based on the aortic arch centerline, and apply the spatial attention weight map to the feature map output by the boundary detail branch to weight and enhance the edge features of the isthmus region. S6. Perform multi-scale feature fusion on the weighted enhanced vascular edge detail features and the overall vascular topology features to output the initial three-dimensional segmentation results of the aortic arch and the narrowing region. S7. Based on the initial three-dimensional segmentation results, extract the vessel segmentation centerline and calculate the point-by-point vessel diameter. Based on the abrupt change in vessel diameter, reverse the narrowing segment boundary of the initial three-dimensional segmentation results to obtain the final aortic arch narrowing image segmentation result.

2. The aortic arch constriction image segmentation method based on deep learning according to claim 1, characterized in that: The preprocessing operation is performed in the order of resampling, gray-level truncation, background removal, and gray-level normalization. Specifically, the aortic CT angiography three-dimensional image is resampled to the same voxel spacing in all directions by trilinear interpolation, the image gray-level value is truncated to the effective HU interval of the vascular tissue, and the gray-level values ​​outside the interval are mapped to the endpoints of the interval. Threshold segmentation combined with three-dimensional morphological opening operation is used to remove non-vascular background areas of the bed board and bones. Finally, grayscale linear normalization is performed on the remaining image to map the grayscale values ​​to the [0,1] interval to obtain the standardized three-dimensional image.

3. The aortic arch constriction image segmentation method based on deep learning according to claim 1, characterized in that: The anatomical landmark detection network is a 3D encoder-decoder keypoint detection model; the encoder of the network contains four sets of parallel 3D convolutional branches corresponding to different downsampling rates, and adjacent branches perform bidirectional feature interaction through cross-resolution convolution, maintaining high-resolution feature output throughout. The decoder adopts a stepwise trilinear upsampling structure, and fuses the corresponding resolution features of the encoder through skip connections, finally outputting a 5-channel heatmap with the same size as the standardized 3D image. During network training, a three-dimensional Gaussian heatmap is generated centered on the labeled marker points as a supervision label. During inference, the three-dimensional coordinates corresponding to the maximum probability of each channel heatmap are extracted to obtain the coordinates of the key anatomical marker points.

4. The aortic arch constriction image segmentation method based on deep learning according to claim 1, characterized in that: The aortic arch vascular centerline is generated as follows: Frangi 3D vascular enhancement is performed on the standardized 3D image, the vascular probability of each voxel is calculated based on the eigenvalues ​​of the 3D Hessian matrix, and a vascular cost map that is negatively linearly correlated with the vascular probability value is constructed. Following the physiological course of the aortic arch from the proximal to the distal end, the path starts at the opening point of the brachiocephalic trunk, and then passes through the opening points of the left common carotid artery, the left subclavian artery, and the proximal isthmus, ending at the distal isthmus. The complete aortic arch centerline is obtained by stitching together the path using the three-dimensional Dijkstra shortest path algorithm. The tubular region of interest is generated using the aortic arch centerline as the axis and a set ratio of the average diameter of the aortic arch as the cross-sectional radius to form a three-dimensional tubular mask. The isthmus focused image is the effective region image obtained by element-wise multiplication of the standardized three-dimensional image and the three-dimensional tubular mask.

5. The aortic arch constriction image segmentation method based on deep learning according to claim 1, characterized in that: The tubular structure branch of the dual-branch collaborative segmentation network adopts a four-level downsampling encoder structure. Each encoder level includes two levels of three-dimensional convolution, batch normalization layer, ReLU activation layer and max pooling layer with stride of 2. Three-dimensional dilated convolution with progressively increasing dilation rate is embedded between the two levels of convolution to expand the receptive field without reducing feature resolution and extract the overall topological features of the blood vessels, including vessel orientation, branch position and tubular morphology.

6. The aortic arch constriction image segmentation method based on deep learning according to claim 4, characterized in that: The boundary detail branch of the dual-branch collaborative segmentation network is composed of multi-level cascaded three-dimensional convolutional units, and no downsampling is performed throughout the process. The feature resolution is consistent with the original resolution of the isthmus focused image. The spatial attention weight map is constructed with the aortic arch vessel centerline as the center and a two-dimensional Gaussian distribution is constructed on a radial plane perpendicular to the tangent of the centerline. The standard deviation of the Gaussian distribution is a set proportion of the preset average diameter of the aortic arch, and the center weight gradually decreases along the radial direction. The response is enhanced by multiplying the center weight of the isthmus segment by a gain coefficient. This weighted enhancement is achieved by element-wise multiplication of the spatial attention weight map with the feature map.

7. The aortic arch constriction image segmentation method based on deep learning according to claim 1, characterized in that: The specific method of multi-scale feature fusion is as follows: the downsampled features of the four different scales output by the tubular structure branch are upsampled to the original resolution of the isthmus focused image by trilinear interpolation, and the upsampled features of each scale are stitched together with the single-scale features output by the boundary detail branch along the channel dimension. After channel dimensionality reduction and feature integration are completed by a 1×1×1 three-dimensional convolutional layer, the resolution is upsampled step by step to the original resolution, and the segmentation probability of each voxel is output by the sigmoid activation function to form the initial three-dimensional segmentation result. After channel dimensionality reduction and feature integration are completed by a 1×1×1 three-dimensional convolutional layer, the resolution is upsampled step by step to the original resolution. The segmentation probability of each voxel is output by the sigmoid activation function to form the initial three-dimensional segmentation result.

8. The aortic arch constriction image segmentation method based on deep learning according to claim 1, characterized in that: The vessel segmentation centerline is extracted using a three-dimensional thinning algorithm that maintains 26-neighbor topological connectivity. The point-by-point diameter is obtained by statistically analyzing the area of ​​the segmented vessel region on a cross-section perpendicular to the centerline tangent and calculating the equivalent circle diameter. The diameter mutation region is determined as a vessel segment where the absolute value of the rate of change of diameter at multiple consecutive centerline points exceeds a preset threshold. During correction, a preset search range is expanded radially with the centerline of the mutation region as the axis, and the gray-level radial gradient value is calculated within the range. The vessel boundary is corrected at the position with the maximum absolute gradient value.

9. The aortic arch constriction image segmentation method based on deep learning according to claim 1, characterized in that: Both the anatomical landmark detection network and the bi-branch collaborative segmentation network are trained using a CT angiography image dataset labeled with aortic arch constriction. During training, the input image is subjected to elastic deformation data augmentation through random cropping, random rotation, and B-spline free deformation. The dual-branch collaborative segmentation network uses a weighted combination of Dice loss and cross-entropy loss as the basic loss function, and adds a pipe diameter consistency loss term to constrain the segmentation accuracy of the narrowing region.

10. The aortic arch constriction image segmentation method based on deep learning according to claim 1, characterized in that: Before obtaining the final aortic arch constriction image segmentation result, topology optimization post-processing is performed on the initial three-dimensional segmentation result; Through 26-neighbor 3D connected domain analysis, the main blood vessel connected domain with the largest volume is retained, and isolated mis-segmented regions with a volume ratio less than the preset proportion are removed. The endpoints of the vascular branches are extracted and the branch lengths are counted. When the number of branches exceeds the normal anatomical range of the aortic arch, the shortest abnormal branch is removed to ensure that the final segmentation result conforms to the anatomical morphology of the aortic arch.

Images