A key point positioning method based on tree pipeline topology constraint and cross-task cooperation

Abstract

The application belongs to the technical field of medical image processing, and particularly relates to a key point positioning method based on tree pipeline topology constraint and cross-task cooperation. Firstly, image segmentation is performed on the tree pipeline structure, a segmentation mask is used to impose spatial priori constraint on a key point heat map, and the key point prediction range is limited in the tree pipeline structure area. Then, cross-task prompt features between segmentation and key points are used for multi-round iteration interaction optimization, the tree pipeline structure is provided, the key point positioning position is adjusted, and the robustness of key point positioning is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a key point localization method based on tree-like pipeline topology constraints and cross-task collaboration. Background Technology

[0002] In medical image processing, the localization of key points in tree-like tubular structures (such as tracheal trees, vascular trees, and bile duct trees) is a fundamental step supporting many clinical diagnostic and surgical navigation tasks. Accurate extraction of anatomical key points in these tubular systems can effectively characterize the connections between branches, and is therefore a core prerequisite for achieving topological structure analysis and global structure modeling.

[0003] Traditional keypoint localization methods typically rely on image intensity or local texture features, making predictions by directly regressing coordinates or optimizing heatmaps point-by-point. However, these methods often ignore the inherent topological continuity and morphological priors of tree-like pipes, leading to localization errors or false keypoints in areas with small pipe diameters, low contrast, or deformities. While deep learning-based keypoint detection methods have made some progress in recent years, most works still treat keypoint localization as an independent task, failing to fully utilize the complete prior information inherent in the pipe structure itself to constrain the geometric consistency between keypoints.

[0004] In fact, segmentation of the pipeline structure can provide the overall morphology and topological relationships between branches, which is of great guiding significance for distinguishing anatomically real key points (such as arterial bifurcation points). Therefore, how to effectively integrate the two tasks of pipeline segmentation and key point localization, and use the topological information obtained from segmentation to guide the accurate identification and precise localization of key points, remains a key problem to be solved in current research. To this end, this invention proposes a key point localization method based on tree-like pipeline topological constraints and cross-task collaboration. Summary of the Invention

[0005] The purpose of this invention is to provide a key point localization method based on tree-like pipeline topology constraints and cross-task collaboration. This method can achieve key point localization by segmenting the tree-like pipeline structure, using segmentation masks to apply spatial prior constraints to the key point heatmap, limiting the prediction range of key points to the area of ​​the tree-like pipeline structure, and then using cross-task cue features between segmentation and key points for multi-round iterative interactive optimization to improve the robustness of key point localization. The tree-like pipeline structure is provided to adjust the position of key point localization.

[0006] The specific technical solution adopted by this invention is as follows: A key point localization method based on tree-like pipeline topology constraints and cross-task collaboration includes the following steps: Step 1: Initial segmentation and keypoint prediction; First, data preprocessing is performed. After data preprocessing, the original 3D medical image is input into the constructed network model. Through the segmentation branch and the key point localization branch, the initial segmentation prediction map and the initial key point heat map are generated. Step 2: Cross-task cue feature extraction; The design collaborative optimization module extracts spatial segmentation cue features and key point structure cue features of the tree-like pipeline topology from the segmentation prediction map of the segmentation branch and the key point heatmap of the key point localization branch, respectively, through a cue encoder. Each of these features is encoded and passed to the other branch as input to the cross-attention module. The cross-attention module performs feature optimization based on cross-attention. Specifically, it inputs the spatial segmentation cue features of the tree-like pipeline topology into the keypoint network, and guides the optimization of the keypoint feature map through the cross-attention module; and inputs the keypoint structure cue features into the segmentation network; after multiple iterations of optimization, the segmentation network outputs an updated segmentation prediction map, and the keypoint network outputs an updated keypoint heatmap. Step 3: The graph structure is further optimized to locate key points; based on the mask and key point heatmap of the tree-like pipeline topology, the coordinates of key points are finally refined by the graph convolution module, and the refined three-dimensional coordinates are output to complete the key point recognition and localization of medical images.

[0007] The technical effects achieved by this invention are as follows: The keypoint localization method based on tree-like pipeline topology constraints and cross-task collaboration provided by this invention can effectively solve the problem of keypoints easily deviating from blood vessels into the background or non-target tissues by segmenting the tree-like pipeline structure, using masks to apply spatial prior constraints to the keypoint heatmap, and then limiting the response region of the keypoints to within the tree-like pipeline structure. Keypoint cue features can be fed back to the segmentation network, using the fine location information provided by the keypoints to correct the segmented tree-like pipeline regions, improving the segmentation integrity of small branches of the pipeline structure. Furthermore, constraints can be applied through the anatomical connectivity relationships of graph convolution modules to constrain the anatomical relationships between multiple keypoints (e.g., six keypoints designed in the renal artery), ensuring that the predicted keypoint coordinates are reasonably distributed across the anatomical structure. This method improves the keypoint localization capability by linking the tree-like pipeline topology segmentation with the keypoint task, allowing them to guide each other. Attached Figure Description

[0008] Figure 1 This is a framework diagram of a key point localization method based on tree-like pipeline topology constraints and cross-task collaboration according to the present invention; Figure 2 This is a framework diagram of the collaborative key point localization network model TVC-Net with tree-like pipeline constraints of the present invention; Figure 3 This is a framework diagram of the graph convolution module, namely the KpGraphRefiner module, of this invention. Detailed Implementation

[0009] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0010] like Figures 1-3 As shown, a key point localization method based on tree-like pipeline topology constraints and cross-task collaboration includes the following steps: Step 1: Initial segmentation and keypoint prediction; First, data preprocessing is performed. After data preprocessing, the original 3D medical image is input into the constructed network model. Through the segmentation branch and the key point localization branch, the initial segmentation prediction map and the initial key point heat map are generated. Step 2: Cross-task cue feature extraction; The design collaborative optimization module extracts spatial segmentation cue features and key point structure cue features of the tree-like pipeline topology from the segmentation prediction map of the segmentation branch and the key point heatmap of the key point localization branch, respectively, through a cue encoder. Each of these features is encoded and passed to the other branch as input to the cross-attention module. The cross-attention module performs feature optimization based on cross-attention. Specifically, it inputs the spatial segmentation cue features of the tree-like pipeline topology into the keypoint network, and guides the optimization of the keypoint feature map through the cross-attention module; and inputs the keypoint structure cue features into the segmentation network; after multiple iterations of optimization, the segmentation network outputs an updated segmentation prediction map, and the keypoint network outputs an updated keypoint heatmap. Step 3: The graph structure is further optimized to locate key points; based on the mask and key point heatmap of the tree-like pipeline topology, the coordinates of key points are finally refined by the graph convolution module, and the refined three-dimensional coordinates are output to complete the key point recognition and localization of medical images.

[0011] Preferably, in step 1, during image preprocessing, the collected raw data, including medical image data and segmentation labels of tree-like pipelines, are preprocessed to enhance the data, and a training and testing dataset is constructed simultaneously; specific operations: First: Image preprocessing; grayscale truncation is performed on the collected image data to limit the CT value to the [0, 2048] HU range, eliminating extreme grayscale interference from bones and air, and then linear mapping is performed to the [0,1] range to complete grayscale normalization and eliminate differences between different devices and scanning parameters.

[0012] Secondly: size standardization and center padding; the image is fixed to a size of (128, 128, 128) by interpolation or resampling.

[0013] Then: data annotation; under the guidance of 2-3 clinicians, the ITK-SNAP software was used to label the key points of the dataset; the training set, validation set and test set were randomly divided in an 8:1:1 ratio to ensure a balanced distribution of different data.

[0014] Finally: Training set data augmentation; spatial transformations for data augmentation, including random rotation and random flipping along three axes to adapt to different anatomical orientations; random scaling to simulate organ size differences between different patients or scaling changes during scanning, scaling to 75%-125%; all transformations are applied synchronously to the image and mask to ensure spatial alignment, thereby enhancing the model's ability to generalize to irregular shapes and size differences.

[0015] Preferably, in step 1, the network model construction process is as follows: Figure 2 As shown, the designed tree-structured pipeline-constrained collaborative keypoint localization network model TVC-Net includes two branches: a segmentation branch and a keypoint localization branch. The segmentation branch uses DenseBiasNet as the backbone network, which enhances the feature capture capability of slender pipe structures through a dense bias connection mechanism. The input is the original 3D medical image, and the output is the initial segmentation prediction map of the tree-like pipe structure. The keypoint localization branch uses a 3D U-Net that integrates contrastive attention and spatial attention as the backbone network to enhance the feature response to keypoint regions. It takes the same original image as input and outputs an initial keypoint heatmap.

[0016] Preferably, in step 2, the preliminary prediction results of the two branches do not exist independently, but rather exchange information through subsequent cross-task prompts and cross-attention modules. The prior spatial information of the tree-like pipeline structure provided by the segmentation branch can constrain the key points to focus on the tree-like pipeline region, and the bifurcation point location information provided by the key point branch can guide the segmentation network to focus on the easily broken bifurcation points, thereby forming a two-way collaborative optimization mechanism. Preferably, in step 2, the collaborative optimization module achieves mutual guidance between segmentation and key point localization tasks through bidirectional 3D cross-attention and iteratively optimizes multiple times; firstly, cross-task prompt features are extracted, and the segmentation prediction map and key point heatmap are input into the prompt encoder, which is then encoded into spatial segmentation prompt features and key point structure prompt features through a two-layer 3DU-Net convolutional localization network and the LeakyReLU activation function, respectively.

[0017] Preferably, in step 2, the cross-attention module performs cross-attention optimization. The key point structure cue features are input into the cross-attention module of the segmentation network to guide the segmentation network to focus on the bifurcation region and improve the segmentation accuracy at the bifurcation. At the same time, the segmentation tree-like pipeline structure cue features are input into the cross-attention module of the keypoint network to guide the key points to focus on the pipeline structure region and suppress false background responses. After multiple iterations, the segmentation network outputs an updated segmentation prediction map, and the keypoint network outputs an updated keypoint heatmap.

[0018] Preferably, in step 3, specifically, the design graph convolution module is as follows: Figure 3 As shown, the graph convolution module, based on the mask and key point heatmap of the tree-like pipeline topology output by the collaborative optimization module, further refines the key point coordinates through graph convolution. First, the key point heatmap is mapped to a high-dimensional feature space after soft argmax normalization. Second, a directed graph is constructed based on anatomical connections, with the main key point as the root node and the branch key points as child nodes, clarifying the connection relationships. The message passing mechanism of graph convolution is used to pass the features of the main node to the branch node, inferring the reasonable offset of each branch key point relative to the main node, and outputting the refined three-dimensional coordinates after range constraints.

[0019] In this invention, the network model training settings are as follows: preprocessed images and segmentation labels are input into the network, and the training batch size is set to 2-4 to adapt to the GPU memory capacity. The Adam optimizer is used, and the initial learning rate is set to 1e-4. A loss function is designed based on the difference between the predicted heatmap and the ground truth heatmap. The loss function is a multi-task weighted combination, including segmentation mixture loss, keypoint weighted MSE loss, and spatial consistency loss.

[0020] In this invention, the loss function is designed as follows: the key point heatmap loss is designed. Weighted mean square error is used, and the preliminary heatmap is monitored. Heatmap of key refining points The median value of the actual heatmap H is greater than the threshold. The voxels are considered as positive samples, forming a weighted MSE, as shown in equation (1).

[0021] (1) (2) (3) in This indicates the weight of the initial predicted loss, while To refine the weights of the prediction loss, an empirical structural design from the fields of medical image processing and 3D keypoint detection is used to avoid excessively large gradients in the refined branches that could overwhelm the initial branches; N represents the total number of voxels, and the threshold... Voxels with values ​​greater than this value are considered keypoint regions; otherwise, they are considered background. As a balancing factor, it is a better solution selected after repeated comparisons on a specific dataset. It assigns higher weights to positive samples, forcing the model to focus on the neighborhood of key points; if the weights are too low, the samples will be overwhelmed by the background. If the background of the dataset is more complex, it needs to be increased. When replacing different backbone networks, their initial gradient scales are different, and can be appropriately reduced. .

[0022] Loss of spatial consistency in design A tree-like pipe structure is used to segment and constrain key point heatmaps via a mask, ensuring that predicted values ​​fall near the tree-like pipe structure. The tree-like pipe mask is then applied. Max pooling expansion yields a neighborhood mask for a tree-like pipe structure. A squared penalty is applied to heatmap values ​​falling outside this region. Spatial consistency loss. As shown in equation (4): (4) Where N represents the total number of all voxels, This represents the probability value of the key point heatmap at position i in the refined output. This indicates that the voxel is located within the tree-like pipe structure or its expanded neighborhood.

[0023] Furthermore, the segmentation loss, denoted by cross-entropy and Dice loss, measures the difference between the predicted segmentation map and the ground truth label. Ultimately, the loss function L of the entire network is composed of the segmentation loss. Key point loss and spatial consistency loss Weighted summation: (5) in , , The contribution intensity of each loss is controlled separately.

[0024] This invention performs multi-dimensional verification of the results, specifically verifying the algorithm's performance in terms of localization accuracy, segmentation effect, and processing efficiency to ensure that the algorithm's performance meets clinical needs. Key point localization verification: Statistical analysis of localization errors on the test set (such as 3D Euclidean distance error, mean radial error (MRE)); Quantitative segmentation verification: Dice similarity coefficient (DSC), mean Hausdorff distance (HD95), and volume overlap error (VOE) were used. Processing efficiency verification: The processing time of a single medical image (including image preprocessing and model inference) was statistically analyzed. The processing time of a single 3D CT image with 300-800 layers was ≤5 minutes. Compared with the traditional multi-step method, the processing time was reduced by more than 60%, which meets the clinical needs for efficient processing.

[0025] This invention optimizes keypoint localization through prior constraints of a tree-like pipeline topology, and simultaneously optimizes segmentation accuracy at branching points of the pipeline topology through feature information feedback from keypoint localization, forming a virtuous cycle. This solves the problem of traditional keypoint localization methods easily drifting into the background and improves the segmentation quality at branching points of the pipeline topology. The final output effectively characterizes the branching connections of the pipeline, which is a core prerequisite for topology analysis and global structure modeling. The proposed method has strong versatility and clinical translational value.

[0026] Preferably, the 3D U-Net convolutional localization network can be replaced by classic 3D heatmap regression models such as HRNet-3D and the improved 3DU-Net++, or 3D-UNETR or Swin UNETR based on Transformer can be used as the feature extraction backbone, and the output head can be adjusted to adapt the key point heatmap generation.

[0027] Preferably, the DenseBiasNet tree-like pipeline topology segmentation network can be replaced by a segmentation architecture of 3D U-Net, V-Net, or nnU-Net, or by an Attention U-Net architecture that introduces an attention mechanism.

[0028] Preferably, the graph convolution module can be replaced by a graph attention network or a graph transformer neural network to perform inference about the positional offset of relative key points.

[0029] In the specific implementation of this invention, Figure 1 An algorithmic framework for key point localization based on tree-like pipeline topology constraints and cross-task collaboration.

[0030] The basic steps of the proposed algorithm framework are as follows: Figure 1 As shown, after data preprocessing, the original CT images are first segmented into a tree-like pipe topology and coarsely predicted for key points using a designed collaborative keypoint localization network model with tree-like pipe constraints (TVC-Net). Secondly, a cross-attention module guides the optimization of keypoint features and multiple iterative updates of segmentation features. Finally, based on the pipe structure mask and keypoint heatmap, the keypoint coordinates are further optimized using a graph structure.

[0031] Figure 2The diagram shows the specific structure of the collaborative keypoint localization network model (TVC-Net) with tree-like pipeline constraints. The designed tree-structured pipeline constraint collaborative key point localization network (TVC-Net) is as follows: Figure 2 As shown, two complementary branches, pipeline structure segmentation and key point localization, are constructed. Feature interaction is achieved through an iterative cross-task prompting mechanism, and a tree-like pipeline topology is introduced to constrain the spatial relationship of key points.

[0032] The inputs for both branches are Where B is the batch size, D represents the 3D voxel data in the depth direction, H represents the height direction, and W represents the number of voxels in the width direction. The first branch performs initial segmentation of the tree-like pipe topology based on the DenseBiasNet segmentation network, obtaining the predicted mask. .

[0033] (6) Where D represents the feature transformation of the DenseBiasNet tree-like pipeline topology segmentation network.

[0034] The second branch is based on an improved 3D U-Net convolutional localization network, which obtains the initial predicted heatmap of key points. .

[0035] (7) Where F represents the feature transformation of the 3D U-Net convolutional localization network.

[0036] Splitting branches will predict the result S pre The encoded cue features are passed to the keypoint branch, and feedback information from the keypoint branch is received to refine the segmentation results. The segmentation cue features P are extracted using equation (8). seg , (8) in, This is a segmentation prompt encoder.

[0037] and segmentation prompt feature P seg Inject into the keypoint network, and simultaneously receive cue features P from the keypoint branches. kp Subsequently, feature optimization guided by equation (9) makes the segmentation network focus on the bifurcation region, and finally outputs the updated tree-like pipeline segmentation mask. .

[0038] (9) Among them, S pre These are preliminary predictions. For cross attention, Use the segmentation mask for the updated tree pipeline.

[0039] The keypoint branch encodes the heatmap into cue features and passes them to the segmentation branch. Simultaneously, it combines the tree-like pipeline topology provided by the segmentation branch to guide keypoints to focus on the pipeline structure region. The keypoint branch extracts its own cue features P using equation (10). kp It is injected into the segmentation network and simultaneously receives segmentation cue features.

[0040] (10) in, Initial prediction heatmap for key points, Provides key point hints for the encoder.

[0041] Subsequently, keypoint features are optimized using Equation (11) to suppress false background responses. The keypoint network outputs an updated keypoint heatmap. Then it is passed to the graph convolution module, that is, to the KpGraphRefiner module.

[0042] (11) in, Segmentation hint features, Initial prediction heatmap for key points, This is for cross-attention.

[0043] Figure 3 This is a schematic diagram of the graph convolution module designed for this invention, specifically a diagram showing the input to the KpGraphRefiner module.

[0044] The keypoint coordinates are offset and corrected using a graph convolution module, resulting in refined keypoint coordinates. The structure of the graph convolution module, specifically the KpGraphRefiner module, is as follows: Figure 3 As shown. For each key point, first apply constraints to the pipeline structure region to eliminate background responses, and obtain the constraint heatmap of each key point through equation (12). .

[0045] (12) in This indicates element-wise multiplication. For the pipe structure mask, The updated keypoint heatmap for k keypoints.

[0046] The initial coordinates H0 of each key point are extracted by the spatial constraint soft argmax, as shown in equation (13): (13) Where argmax is the soft argmax normalization. This is a heatmap of key point constraints, with H0 as the initial coordinate.

[0047] The initial coordinates are input into the offset prediction head Offset via two branches, the first of which is a jump connection and serves as the first input to the offset prediction head Offset.

[0048] The second branch first maps all keypoint coordinates H0 to a high-dimensional feature space using a coordinate encoder (LRL) to obtain node features. .

[0049] (14) In the LRL coordinate encoder, H0 is the initial coordinate.

[0050] Secondly, based on the tree-like pipeline topology, a directed graph structure G(V,E) is constructed to represent the topological relationships between key points, where the vertex set is V and the edge set is E=(v0,v...). t For each edge (v0, v) t The source node and target node features are concatenated and then input into the graph structure inference network. The graph structure carrying the topological relationships is obtained, and the transformation formula is shown in equation (15) below: (15) Where || denotes vector concatenation. Represents a graph-structured reasoning network. The node features of the main key points. The node characteristics of the child nodes.

[0051] Finally, the offset prediction head is used to refine the coarsely located keypoint coordinates, ultimately outputting the refined keypoint coordinates. : (16) Where Offset is the offset prediction header, H0 is the initial coordinates, and G is the graph structure.

[0052] This invention relies on the cross-task collaboration of a segmentation network and a keypoint localization network. It uses a tree-like pipeline topology segmentation mask to impose spatial prior constraints on the keypoint response region, and achieves iterative mutual optimization of segmentation and localization through a bidirectional cross-attention mechanism. The keypoint coordinates are refined based on the anatomical topological relationships using a graph convolution module, forming a complete algorithm system encompassing segmentation prediction, spatial constraints, cross-task cue fusion, iterative interactive optimization, and graph topology coordinate refinement. Its innovation lies in: This invention utilizes the segmentation of a tree-like pipeline topology to guide key point localization. The tree-like pipeline topology segmentation mask is multiplied element-by-element with the key point heatmap, forcing the key point response range to be limited to the area of ​​the tree-like pipeline topology, thus fundamentally solving the problem of key points drifting to the background or non-target tissues.

[0053] This invention designs a bidirectional cross-iterative collaborative optimization architecture for segmentation and keypoint networks. It alternately injects dense shape cues of the tree-like pipeline topology and sparse position cues of keypoints into each other's networks, forming a bidirectional closed-loop optimization in which the segmentation of the tree-like pipeline topology guides keypoint localization and keypoint localization corrects segmentation. This breaks through the limitation of traditional methods that treat segmentation and localization as independent tasks.

[0054] This invention introduces a graph convolution module based on anatomical topology to refine coordinates. It utilizes the tree-like pipeline topology mask obtained from segmentation and the graph convolution network to optimize the tree-like connection relationship between the main key points and the branch key points. It performs relative position reasoning and offset correction on the coordinates of each key point to ensure that the predicted multiple key points maintain a geometrically reasonable distribution in the anatomical structure.

[0055] The algorithm framework proposed in this invention is compatible with 3D medical image data from different individuals. This method can be widely applied to the localization of key bifurcation points of tree-like ductal structures, including renal arteries, coronary arteries, hepatic arteries, trachea and bronchi, and to provide anchor points for subsequent organ segmentation and partitioning. It also provides accurate anatomical basis for the localization of key vascular points and precise duct segmentation.

[0056] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A key point localization method based on tree-like pipeline topology constraints and cross-task collaboration, characterized in that: Includes the following steps: Step 1: Initial segmentation and keypoint prediction; First, data preprocessing is performed: the original 3D medical image is input into the constructed network model, and the initial segmentation prediction map and initial key point heatmap are generated through the segmentation branch and key point localization branch. Step 2: Cross-task prompt feature extraction and feature optimization; Design a collaborative optimization module. From the segmentation prediction map and the keypoint heatmap, respectively, extract spatial segmentation cue features of the tree-like pipeline topology and keypoint structure cue features through a cue encoder. Both feature encodings are then passed to the other branch as input to the cross-attention module. The cross-attention module optimizes features based on cross-attention. Specifically, it inputs the spatial segmentation cue features of the tree-like pipeline topology into the keypoint network, and simultaneously inputs the keypoint structural cue features into the segmentation network. Then, the cross-attention module guides the optimization of the keypoint feature map and the segmentation prediction map. After multiple iterations of optimization, it outputs the updated segmentation prediction map and keypoint heatmap. Step 3: Continue to optimize the graph structure and locate key points; Based on the mask and key point heatmap of the tree-like pipeline topology, the key point coordinates are finally refined by the graph convolution module, and the refined three-dimensional coordinates are output to optimize the key point recognition and localization of medical images.

2. The key point localization method based on tree-like pipeline topology constraints and cross-task collaboration according to claim 1, characterized in that: In step 1, during the image preprocessing process, the collected raw data, including medical image data and segmentation labels of tree-like pipelines, are preprocessed to enhance the data and construct a training and testing dataset. Specific operations: First, image preprocessing: The collected image data is truncated to grayscale, and the CT value is limited to the [0, 2048] HU range to eliminate extreme grayscale interference from bones and air. Then, it is linearly mapped to the [0,1] range to complete grayscale normalization and eliminate the differences between different devices and scanning parameters. Secondly, size normalization and center padding: The image is fixed to a size of (128, 128, 128) by interpolation or resampling; Then, data annotation: the ITK-SNAP software was used to label the key points of the dataset; the training set, validation set and test set were randomly divided in an 8:1:1 ratio; Finally, training set data augmentation: spatial transformations are used for data augmentation, including random rotation, random flipping along the three axes to adapt to different anatomical orientations; random scaling is used to simulate differences in organ size between different patients or scaling changes during scanning, scaling to 75%-125%.

3. The key point localization method based on tree-like pipeline topology constraints and cross-task collaboration according to claim 2, characterized in that: In step 1, the network model is built as follows: a tree-structured pipeline-constrained collaborative key point localization network model TVC-Net is designed; the tree-structured pipeline-constrained collaborative key point localization network model TVC-Net includes two branches: a segmentation branch and a key point localization branch. The segmentation branch uses the DenseBiasNet tree-like pipe topology segmentation network as the backbone network. The input is the original 3D medical image, and the output is the initial segmentation prediction map of the tree-like pipe structure. The keypoint localization branch uses a 3D U-Net convolutional localization network that integrates contrastive attention and spatial attention as the backbone of the convolutional localization network. It takes the same original image as input and outputs an initial keypoint heatmap.

4. The key point localization method based on tree-like pipeline topology constraints and cross-task collaboration according to claim 3, characterized in that: The 3D U-Net convolutional localization network can be replaced by classic 3D heatmap regression models such as HRNet-3D and the improved 3DU-Net++, or 3D-UNETR or Swin UNETR based on Transformer can be used as the feature extraction backbone, and the output head can be adjusted to adapt to the generation of key point heatmaps.

5. The key point localization method based on tree-like pipeline topology constraints and cross-task collaboration according to claim 3, characterized in that: The DenseBiasNet tree-structured segmentation network can be replaced by segmentation architectures such as 3D U-Net, V-Net, nnU-Net, or the Attention U-Net architecture that introduces an attention mechanism.

6. A key point localization method based on tree-like pipeline topology constraints and cross-task collaboration as described in claim 4 or 5, characterized in that: In step 2, the preliminary prediction results of the two branches do not exist independently, but rather exchange information through subsequent cross-task prompts and cross-attention modules. The prior spatial information of the tree-like pipeline structure provided by the segmentation branch can constrain the key points to focus on the tree-like pipeline region, and the bifurcation point location information provided by the key point branch can guide the segmentation network to focus on the easily broken bifurcation points, thereby forming a two-way collaborative optimization mechanism.

7. The key point localization method based on tree-like pipeline topology constraints and cross-task collaboration according to claim 6, characterized in that: In step 2, the collaborative optimization module achieves mutual guidance between segmentation and key point localization tasks through bidirectional 3D cross attention and iterates and optimizes multiple times. First, cross-task prompt features are extracted, and the segmentation prediction map and key point heatmap are input into the prompt encoder. Through two layers of 3DU-Net convolutional localization network and LeakyReLU activation function, they are encoded into spatial segmentation prompt features and key point structure prompt features, respectively.

8. The key point localization method based on tree-like pipeline topology constraints and cross-task collaboration according to claim 7, characterized in that: In step 2, the cross-attention module performs cross-attention optimization, inputting key point structure cue features into the cross-attention module of the segmentation network to guide the segmentation network to focus on the bifurcation region; at the same time, the segmentation tree-like pipeline structure cue features are input into the cross-attention module of the keypoint network to guide the keypoints to focus on the pipeline structure region and suppress false background responses; after multiple iterations, the segmentation network outputs an updated segmentation prediction map, and the keypoint network outputs an updated keypoint heatmap.

9. The key point localization method based on tree-like pipeline topology constraints and cross-task collaboration according to claim 8, characterized in that: In step 3, the design graph convolution module is based on the mask and key point heatmap of the tree-like pipeline topology output by the collaborative optimization module, and refines the key point coordinates through graph convolution. Specifically, the key point heatmap is first mapped to a high-dimensional feature space after being normalized by soft argmax; then, a directed graph is constructed based on the anatomical connection relationship, with the main key point as the root node and the branch key points as the child nodes, the connection relationship is clarified, and the message passing mechanism of graph convolution is used to pass the features of the main node to the branch node, infer the reasonable offset of each branch key point relative to the main node, and output the refined three-dimensional coordinates after range constraint.

10. The key point localization method based on tree-like pipeline topology constraints and cross-task collaboration according to claim 9, characterized in that: The graph convolution module can be replaced by a graph attention network or a graph transformer neural network to complete the inference of key point position offsets.

Images