Three-dimensional tubular structure segmentation method and system based on geometric prior and topology perception

Abstract

The invention discloses a three-dimensional tubular structure segmentation method and system based on geometric prior and topology perception. The method comprises the following steps: collecting and preprocessing a blood vessel image; constructing a MorVess segmentation network based on the SAM network and the frozen 2D visual encoder; the MorVess segmentation network comprises a multi-head mask decoder, a global-local fusion module and a 2.5 D slice fusion adapter used for capturing context between slices; inputting a blood vessel image into a MorVess segmentation network, generating a geometric prior graph comprising a blood vessel distance graph and a blood vessel thickness graph by a multi-head mask decoder, fusing the blood vessel distance graph, the blood vessel thickness graph and a multi-source feature graph generated by an SAM network by a global-local fusion module, and generating a final segmentation mask; the MorVess segmentation network is trained through a two-stage training strategy comprising a macroscopic adaptation stage and a topology refining stage; and inputting a to-be-segmented three-dimensional medical image into the trained MorVess segmentation network to obtain a voxel-level segmentation result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, and in particular to a three-dimensional tubular structure segmentation method and system based on geometric prior and topology perception. BACKGROUND

[0002] Three-dimensional vessel segmentation is a key task in medical image analysis, which is of great significance for the diagnosis, perfusion evaluation and surgical planning of various diseases (such as lung diseases). However, the blood vessel structure (especially the pulmonary vessels and renal vessels) presents a highly complex, sparse and fine tree-like distribution.

[0003] Existing segmentation methods based on deep learning mostly rely on standard binary masks as supervision signals. Such methods regard segmentation as an independent voxel-level classification problem, i.e., judging whether each voxel belongs to the blood vessel.

[0004] However, this method has obvious limitations in practical application:

[0005] 1. Topological discontinuity: Binary mask supervision lacks constraints on the global connectivity of blood vessels. When encountering noise, low contrast or partial volume effects, the model is prone to produce breaks (topological breaks) or false connections in small branches or terminal regions, which seriously affects the downstream morphological analysis.

[0006] 2. Insufficient geometric accuracy: The model lacks explicit perception of the geometric shape of blood vessels (such as vessel diameter and boundary). This leads to blurred edges and uneven tube thickness in the segmentation results, which do not match the real anatomical structure.

[0007] 3. Model adaptation difficulty: For three-dimensional (3D) medical images, directly using a 3D convolutional network (3D CNN) has extremely high computational cost. Using a 2D network for slice-by-slice processing completely loses the spatial context information between slices. How to efficiently migrate powerful 2D pre-trained models (such as SAM, Segment Anything Model) to 3D tasks is still a challenge.

[0008] Due to the above problems, there is an urgent need in the field of medical image segmentation for an innovative technical solution that can explicitly model the geometric and topological structure of blood vessels while maintaining computational efficiency. SUMMARY

[0009] To address the issues of topological discontinuities and insufficient geometric accuracy caused by relying solely on binary mask supervision in existing technologies, this invention provides a 3D tubular structure segmentation method and system based on geometric priors and topology awareness. This invention constructs continuously differentiable geometric prior supervision signals and designs an efficient 2.5D adaptive network architecture to jointly model geometric constraints and semantic segmentation tasks. This introduces explicit topological and geometric guidance beyond standard segmentation, significantly improving the topological integrity and geometric consistency of the segmentation results.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] The first aspect is a three-dimensional tubular structure segmentation method based on geometric priors and topology awareness, which includes the following steps:

[0012] S1. Acquire and preprocess vascular images; obtain the gold standard binary segmentation mask of the vascular images;

[0013] S2. Construct a MorVess segmentation network based on a SAM network and a frozen 2D visual encoder; the MorVess segmentation network includes a multi-head mask decoder, a global-local fusion module, and a 2.5D slice fusion adapter for capturing inter-slice context; input the vascular image into the MorVess segmentation network, the multi-head mask decoder generates a geometric prior map including a vascular distance map and a vascular thickness map, and the global-local fusion module fuses the vascular distance map, the vascular thickness map, and the multi-source feature map generated by the SAM network to generate the final segmentation mask;

[0014] S3. The MorVess segmentation network is trained using a two-stage training strategy that includes a macro-adaptation stage and a topology refinement stage.

[0015] S4. Input the three-dimensional medical image to be segmented into the trained MorVess segmentation network to obtain voxel-level segmentation results.

[0016] Preferably, in S2, the vascular image is input into the 2D visual encoder and reshaped along the depth axis into a sequence containing multiple slices; the 2.5D slice fusion adapter is integrated into each Transformer block of the 2D visual encoder; the slices are input into the 2.5D slice fusion adapter to generate image embeddings and a set of mask tokens.

[0017] Preferably, the 2.5D slice fusion adapter includes a downsampling layer for compressing the feature dimension, a 3D convolutional layer with a kernel size of 3*1*1 for capturing the cross-slice context, and an upsampling layer for recovering the feature dimension, and they are connected in a residual form.

[0018] As preferred, in S2, the multi-head mask decoder comprises a mask prediction head for generating an initial segmentation mask and two geometry prediction heads for generating the geometry prior map; the two geometry prediction heads share a hypernetwork mechanism; the average embedding of the output multiple sets of mask tokens is input into the two geometry prediction heads to generate the vessel distance map and the vessel thickness map in parallel.

[0019] As preferred, the average embedding generates dynamic weights respectively through the two geometry prediction heads independently, and acts on the up-sampled image features in the same way.

[0020] As preferred, the geometry prediction head generating the vessel distance map comprises:

[0021] Based on the set of vessel mask voxels in the gold standard binary segmentation mask, a morphological erosion is performed to generate a set of eroded vessel mask voxels; a blood vessel boundary layer is extracted through a set subtraction operation;

[0022] The shortest weighted Euclidean distance from any point in the three-dimensional discrete image space to the blood vessel boundary layer is calculated, and a boundary distance field is generated;

[0023] The boundary distance field is converted into a potential field with a value range of [0, 1] using an exponential decay function, and the vessel distance map for guiding the model to accurately locate the blood vessel edge is generated.

[0024] As preferred, the geometry prediction head generating the vessel thickness map comprises:

[0025] An internal distance field of the set of vessel mask voxels is calculated; a centerline skeleton of the set of vessel mask voxels is extracted through a topological refinement operator;

[0026] For any point in the set of vessel mask voxels, a projection point on the centerline skeleton is generated through a nearest neighbor projection operator; a skeleton projection point radius is generated according to the principle of the largest inscribed sphere and the internal distance field; the vessel thickness value of any point in the set of vessel mask voxels is twice the corresponding skeleton projection point radius, thereby generating the vessel thickness map.

[0027] As preferred, the global-local fusion module concatenates the generated decoder global semantic features, encoder shallow local texture features, encoder deep semantic abstract features, the vessel distance map and the vessel thickness map along the channel dimension; the concatenated tensor is input into a fusion network to generate fusion features; a dedicated vessel token is mapped into a set of dynamic channel weights through a small MLP, which is used to dynamically weight the fusion features to generate a fusion map; the fusion map is superimposed on the initial segmentation mask through a gated residual connection to generate the final segmentation mask.

[0028] As preferred, the two-stage training strategy in S3 comprises:

[0029] Macro-adaption stage: freeze the 2D visual encoder, only train the 2.5D slice fusion adapter, the multi-head mask decoder and the global-local fusion module, use 1x10 -3 learning rate to quickly adapt the 3D spatial structure and geometric prior tasks;

[0030] Topology refinement stage: freeze the converged 2.5D slice fusion adapter, use 5x10 -5 learning rate to fine-tune the multi-head mask decoder and the global-local fusion module to enhance the model's perception of blood vessel boundaries, thickness and topological connectivity.

[0031] In the second aspect, the three-dimensional tubular structure segmentation system based on geometric prior and topology perception comprises:

[0032] A data acquisition and preprocessing module is configured to acquire and preprocess the blood vessel image and obtain a gold standard binary segmentation mask of the blood vessel image;

[0033] A segmentation network model module based on the SAM network and the frozen 2D visual encoder comprises a multi-head mask decoder, a global-local fusion module and a 2.5D slice fusion adapter for capturing inter-slice context; the blood vessel image is input into the segmentation network model module, the multi-head mask decoder generates a geometric prior map comprising a blood vessel distance map and a blood vessel thickness map, the global-local fusion module fuses the blood vessel distance map, the blood vessel thickness map and a multi-source feature map generated by the SAM network to generate a final segmentation mask;

[0034] A model training module is configured to train the segmentation network model module through a two-stage training strategy comprising a macro-adaption stage and a topology refinement stage;

[0035] A segmentation execution module is configured to input a three-dimensional medical image to be segmented into the trained segmentation network model module to obtain a voxel-level segmentation result;

[0036] The three-dimensional tubular structure segmentation system based on geometric prior and topology perception is used to perform the three-dimensional tubular structure segmentation method based on geometric prior and topology perception as described in the first aspect.

[0037] Compared with the prior art, the beneficial effects of the present application are as follows:

[0038] (1) Multi-task learning with geometry and topology awareness: The present application innovatively proposes two continuous differentiable geometric priors, VDM and VTM, which extend the segmentation task from "pixel classification" to "geometry reconstruction". The model must learn "how thick" and "where the boundary is" while learning "where", fundamentally solving the problem of insufficient binary mask supervision information.

[0039] (2) Topological integrity is significantly improved: Through the global thickness constraint of VTM and the centerline constraint of LclDice loss, the method of the present application can explicitly perceive the tree-like topology of blood vessels, significantly reducing the breakage and false connection of small blood vessels, and the topological connectivity of the segmentation result is stronger.

[0040] (3) High geometric boundary precision: Through the boundary potential field guidance provided by VDM, the gradient signal of the model at the blood vessel edge is stronger, and the generated segmentation mask boundary is sharper and smoother, and the geometric precision (such as HD95 index) is greatly superior to traditional methods.

[0041] (4) Light-weight and efficient 2D pre-training model: The 2.5DAdapter scheme proposed by the present application completely freezes the 2D SAM encoder parameters, and only trains a small number of 3D convolution parameters, which realizes the effective mapping of 2D features to 3D space, greatly reduces the calculation and memory overhead while maintaining high performance.

[0042] (5) Fine multi-source feature fusion: The GLFB module of the present application not only fuses traditional multi-scale (global-local) features, but also innovatively uses the model's own predicted geometric prior (VDM, VTM) as an additional input to the fusion process, realizing the deep coupling of semantics and geometry, and significantly improving the reconstruction ability of peripheral blood vessels. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The method of embodiment 1 of the present application MorVess is shown in the schematic diagram;

[0044] Figure 2 The VDM generation process of embodiment 1 of the present application is shown in the schematic diagram;

[0045] Figure 3 The VTM generation process of embodiment 1 of the present application is shown in the schematic diagram;

[0046] Figure 4 The 2.5D slice fusion adapter and global-local fusion module (GLFB) structure of embodiment 1 of the present application is shown in the schematic diagram;

[0047] Figure 5Visualization comparison chart of the method (Ours) of embodiment 1 of the present application and other comparative methods on three-dimensional pulmonary vessel segmentation. DETAILED DESCRIPTION

[0048] In order to make the technical means, creative features, purposes and effects of the application easy to understand, the present application will be further described in conjunction with specific drawings. However, the present application is not limited to the following embodiments.

[0049] It should be noted that the structures, proportions, sizes, etc. shown in the drawings attached to the present specification are only used to cooperate with the content disclosed in the specification for understanding and reading by those skilled in the art, and do not define the limiting conditions for the implementation of the present application, so they do not have technical substantive significance. Any modification of structure, change of proportion relationship or adjustment of size, which does not affect the effects and purposes that can be achieved by the present application, should still fall within the scope of the technical content disclosed by the present application.

[0050] Embodiment 1: Taking pulmonary vessels as an example

[0051] The present embodiment provides a three-dimensional tubular structure segmentation method based on geometric prior and topological perception, as shown in Figure 1 The overall framework (MorVess) of the present application is shown in the figure. The method mainly includes the following steps:

[0052] S1, collect and pre-process the vessel image; obtain the gold standard binary segmentation mask of the vessel image;

[0053] S2, construct a MorVess segmentation network based on a SAM network and a frozen 2D visual encoder; the MorVess segmentation network includes a 2.5D slice fusion adapter for capturing inter-slice context, a multi-head mask decoder for generating an initial prediction, and a global-local fusion module; the initial prediction includes a vessel distance map and a vessel thickness map; input the vessel image into the MorVess segmentation network to generate a final segmentation mask;

[0054] The present application proposes a MorVess network architecture, which is based on a 2D SAM (ViT-Base) encoder backbone and loads its pre-training weights. The MorVess segmentation network includes:

[0055] (1) 2.5D slice fusion adapter (2.5DAdapter):

[0056] To capture the 3D spatial context while keeping the 2D ViT main encoder of SAM frozen, the invention integrates a 2.5D Adapter module in each Transformer block of the ViT encoder. The (vessel image) input is fed into a 2D vision encoder, which reshapes the input data along the depth (Z) axis into a sequence of N slices [B, N, H, W], where B is the batch size, N is the number of slices, H and W are the height and width of each slice, respectively, representing the dimension of data in the depth (Z axis).

[0057] The 2.5D Adapter module consists of a down-sampling layer (compressing feature dimension), a 3D convolution layer with kernel size 3*1*1 (capturing cross-slice context), and an up-sampling layer (restoring feature dimension), and adopts a residual connection form:

[0058] 2.5DAdapter(x) = Wup(Conv3D(Wdown(Norm(x)))) + x

[0059] Thus, the 2.5D slice fusion adapter outputs image embedding and a set of mask tokens. By fine-tuning only the parameters of the Adapter itself, SAM can organically integrate three-dimensional spatial context information into its two-dimensional feature representation without significantly increasing the additional computational burden.

[0060] (2) Multi-head mask decoder:

[0061] This module receives the image embedding and multiple sets of mask tokens output from the 2.5D slice fusion adapter, expands the original mask decoder of SAM, and outputs three initial prediction logits in parallel:

[0062] Mask prediction head (Mask): retains the original mechanism, interacts with the image embedding through mask_tokens, and generates the initial segmentation mask M through the hypernetwork initial .

[0063] Geometric prediction head: two dedicated prediction heads are introduced. These two heads share similar hypernetwork mechanisms, but their inputs are not a single mask token, but the average embedding of all output mask tokens The average embedding generates dynamic weights through two independent multi-layer perceptrons (MLP) of each geometric prediction head, and acts on the up-sampled image features in the same way, thereby generating the initial prediction logits of the vessel distance map (VDM) and the vessel thickness map (VTM) in parallel and efficiently.

[0064] The specific implementation of the geometric prediction head is as follows:

[0065] Based on the existing gold standard (Ground Truth) binary mask M gt Offline calculation of VDM and VTM.

[0066] As Figure 2 shown, to solve the problem that the standard segmentation loss has weak supervision signals for boundary voxels, the VDM geometric potential field, i.e., the vessel distance map, is constructed. It is assumed that the three-dimensional discrete image space is Vessel mask voxel set The distance between voxels is a vector S p First, the morphological erosion operation is performed to obtain the eroded vessel mask voxel set Then, the vessel boundary layer is extracted through set subtraction operation

[0067]

[0068] Next, the shortest weighted Euclidean distance of any point in the space to the vessel boundary layer The boundary distance field is defined as

[0069]

[0070] Finally, the boundary distance field is converted into a potential field with a value range of [0, 1] using an exponential decay function, i.e., VDM:

[0071]

[0072] Where λ is a hyperparameter that controls the decay rate. This VDM has a value of 1 on the boundary and rapidly decays inward and outward, providing a strong boundary supervision signal for network training.

[0073] As Figure 3 shown, to solve the problem that the model causes inconsistent pipe diameters and loss of connectivity due to the receptive field limitation, the VTM is constructed to provide global scale constraints. First, the internal distance field of the vessel mask voxel set Ω is calculated, which is denoted as The field gives each point x in the vessel mask voxel set Ω its shortest distance to the background Ω c .

[0074]

[0075] Second, the centerline skeleton S = τ(Ω) of the blood vessel is extracted through the topological refinement operator τ.

[0076] According to the principle of the largest inscribed sphere, the radius r(s) of any point s e S on the centerline skeleton is equal to the shortest distance from it to the blood vessel boundary, that is

[0077] Finally, for any point x inside the blood vessel mask voxel set Ω , find its projection point on the skeleton through the nearest neighbor projection operator , and define the blood vessel thickness value of point x as twice the radius of the skeleton projection point:

[0078] VTM(x) = 2r(π s (x))

[0079] (3) Global-Local Fusion Block (GLFB)

[0080] As shown in Figure 4 , to alleviate the information bottleneck caused by the SAM decoder only using the low-resolution feature map of the last layer of the 2D visual encoder, the present application designs GLFB as a parallel high-resolution refinement path.

[0081] This module integrates five kinds of complementary information:

[0082] F dec : global semantic context features from the standard decoder path;

[0083] F shallow : high-frequency features from the encoder shallow layer that reserve local texture;

[0084] F depth : high-level semantic abstract features from the encoder deep layer;

[0085] VDM_{pred}: blood vessel distance map predicted by the multi-head decoder itself;

[0086] VTM_{pred}: blood vessel thickness map predicted by the multi-head decoder itself.

[0087] These multi-source feature maps are upsampled to a uniform high resolution (such as 256x256), concatenated along the channel dimension to form F aggregated = concat[F dec ,F shallow ,F depth ,VDM,VTM].

[0088] This tensor is fed into a fusion network composed of 3x3 convolutions to generate fusion features F fused . At the same time, a dedicated blood vessel token h vessel is introduced, which is mapped to a set of dynamic channel weights w vessel=mlp(h vessel This algorithm dynamically weights the fused features to generate a final high-resolution fused map, which is then superimposed on the initial mask prediction, significantly improving detail reconstruction and topological integrity. These weights are used as dynamic 1x1 convolution kernels to weight the fused features F. fused Perform channel-wise weighted summation: logits vessel =w vessel ·F fused .

[0089] Finally, the fused graph is superimposed onto the initial segmentation mask M via gated residual connections. initial Above this, the final refined segmentation mask is generated:

[0090] M finnal =M initial +α·logits vessel )

[0091] Where α is a learnable gating parameter.

[0092] S3. Train the MorVess segmentation network using a two-stage training strategy that includes a macro-adaptation stage and a topology refinement stage.

[0093] All training was conducted on an Ubuntu 20.04, PyTorch 2.0, and NVIDIA L40 GPU hardware platform. The input data was processed using 2.5D slicing (e.g., 5 channels, target slice $\pm$2 adjacent slices).

[0094] (1) Phase 1: Macro-feature adaptation and 2.5D spatial modeling

[0095] Parameters: The 2DViT encoder with SAM completely frozen is trained only on the newly added lightweight modules (2.5DAdapter, multi-head mask decoder with GLFB and VDM / VTM prediction heads).

[0096] Optimization strategy: The AdamW optimizer is used, with an initial learning rate of 1×10⁻³. The learning rate is scheduled using a "warm-up" approach combined with polynomial decay.

[0097] Loss function: This stage does not include topological constraints. The goal is to enable the model to quickly adapt to 3D spatial structures and establish a basic geometric regression mapping.

[0098] (2) Phase Two: Topology Fine-tuning and Geometric Detail Optimization

[0099] Parameters: Freeze the well-trained 2.5D Adapter module in the pre-training phase, and only train the decoder-end core optimization module (GLFB, geometry prediction head, and new segmentation mask head).

[0100] Optimization strategy: The AdamW optimizer is used, and the initial learning rate is reduced to 5x10 -5 . The learning rate scheduling adopts "cosine annealing with periodic restart" to balance local detail optimization and global topology convergence.

[0101] Loss function: In order to optimize the accuracy of voxels and the consistency of topology, a composite loss function is designed to realize multi-level collaborative constraint through five weighted components:

[0102]

[0103] where λ 1-5 is a weight hyperparameter.

[0104] and are used to optimize the accuracy of voxel-level segmentation.

[0105] Centerline Dice loss) is used to measure the consistency between the predicted and real blood vessel skeletons, and to explicitly suppress fractures.

[0106] Distance map loss) uses L1 loss to calculate the difference between the predicted P dist and the gold standard G dist .

[0107] Thickness map loss) uses a scale-invariant L1 form to calculate the L1 loss between the predicted P thick and the gold standard G thick after maximum normalization within their respective samples, to ensure robustness to absolute pipe diameter:

[0108]

[0109] S4, input the three-dimensional medical image to be segmented into the trained MorVess segmentation network to obtain the voxel-level segmentation result.

[0110] After the model training is completed, a new three-dimensional CT image to be segmented is input into the trained MorVess model. The model performs forward propagation, passes through the architecture of S2, and finally outputs a three-dimensional blood vessel segmentation mask with high resolution and high topological consistency by the GLFB module.

[0111] For example Figure 5The 3D visualization comparison of the method of the present application (MorVess (Ours)) and a plurality of prior advanced methods (Methods 1-8) on the Parse2022 and AIIB data sets is shown. The experimental results show that the other competitive models all produce different degrees of broken connections of small branches in the blood vessel segmentation. The method (MorVess) proposed in the present application is far superior to other methods in the segmentation of local small blood vessels, and the segmentation results are more continuous and complete, and the cases of broken connections and false connections are the least, proving the significant superiority of the present application in maintaining topological integrity and geometric precision.

[0112] Embodiment 2:

[0113] The present embodiment provides a three-dimensional blood vessel segmentation system based on geometric prior regularization, which is used to implement the method of embodiment 1. The system comprises:

[0114] A data acquisition and preprocessing module is configured to acquire and preprocess the blood vessel image from a CT, MRI or other medical imaging device, and to acquire a gold standard binary segmentation mask of the blood vessel image.

[0115] A segmentation network model module comprises an instantiated MorVess network model. The model has a 2.5D slice fusion adapter, a multi-head decoder and a GLFB structure of embodiment 1S2. The module is configured to receive three-dimensional body data as input.

[0116] A model training module comprises an optimizer (such as AdamW), a learning rate scheduler (such as polynomial decay and cosine annealing) and a loss calculation unit. It is used to train the segmentation network model module through a two-stage training strategy including a macro adaptation stage and a topological refinement stage; the module calls the three-dimensional body data and the corresponding geometric prior graph from the data preprocessing module, executes the two-stage training strategy, and calculates the gradient through the composite loss function and updates the parameters of the segmentation network model module.

[0117] A segmentation execution module is configured to input the three-dimensional medical image to be segmented into the trained segmentation network model module to obtain the voxel-level segmentation result. The module receives new three-dimensional body data to be segmented, performs forward inference, and outputs the final voxel-level three-dimensional segmentation mask (such as NIfTI or NRRD format) for use by clinicians for diagnosis or surgical planning.

[0118] The above description is only the preferred embodiment of the present application and the explanation of the applied technical principles. It should be understood by those skilled in the art that the inventive scope involved in the embodiments of the present application is not limited to the technical solutions formed by the specific combinations of the above technical features, and should also cover other technical solutions formed by the combinations of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by the mutual replacement of the above features and the technical features with similar functions disclosed in the embodiments of the present application (but not limited to) form the technical solutions.

Claims

1. A method for three-dimensional tubular structure segmentation based on geometric prior and topology-awareness, characterized in that, The method comprises the following steps: S1, collecting and preprocessing the blood vessel image; obtaining a gold standard binary segmentation mask of the blood vessel image; S2, constructing a MorVess segmentation network based on a SAM network and a frozen 2D visual encoder; the MorVess segmentation network comprises a multi-head mask decoder, a global-local fusion module and a 2.5D slice fusion adapter for capturing inter-slice context; the blood vessel image is input into the MorVess segmentation network, the multi-head mask decoder generates geometric prior maps including a blood vessel distance map and a blood vessel thickness map, the global-local fusion module fuses the blood vessel distance map, the blood vessel thickness map and a multi-source feature map generated by the SAM network to generate a final segmentation mask; S3, training the MorVess segmentation network through a two-stage training strategy comprising a macro adaptation stage and a topology refinement stage; S4, inputting a three-dimensional medical image to be segmented into the trained MorVess segmentation network to obtain a voxel-level segmentation result.

2. The method of claim 1, wherein, In S2, the blood vessel image is input into the 2D visual encoder to be reshaped into a sequence comprising a plurality of slices along the depth axis; the 2.5D slice fusion adapter is integrated into each Transformer block of the 2D visual encoder; the slice is input into the 2.5D slice fusion adapter to generate image embedding and a set of mask tokens.

3. The method of claim 2, wherein, The 2.5D slice fusion adapter comprises a down-sampling layer for compressing feature dimensions, a 3D convolution layer with a kernel size of 3*1*1 for capturing cross-slice context and an up-sampling layer for restoring feature dimensions, and is connected in a residual form.

4. The method of claim 2, wherein, In S2, the multi-head mask decoder comprises a mask prediction head for generating an initial segmentation mask and two geometric prediction heads for generating the geometric prior maps; the two geometric prediction heads share a hypernetwork mechanism; the average embedding of the output multiple sets of mask tokens is input into the two geometric prediction heads to generate the blood vessel distance map and the blood vessel thickness map in parallel.

5. The method of claim 4, wherein, The average embedding generates dynamic weights through the respective independent multi-layer perceptrons of the two geometric prediction heads, and acts on the up-sampled image features in the same way.

6. The method of claim 4, wherein, The generation of the blood vessel distance map by the geometric prediction head comprises: performing morphological erosion on a blood vessel mask voxel set in the gold standard binary segmentation mask to generate an eroded blood vessel mask voxel set; extracting a blood vessel boundary layer through set subtraction operation; calculating the shortest weighted Euclidean distance from any point in a three-dimensional discrete image space to the blood vessel boundary layer, and generating a boundary distance field; using an exponential decay function to convert the boundary distance field into a potential field with a value range of [0, 1] to generate the blood vessel distance map for guiding the model to accurately locate the blood vessel edge.

7. The method of claim 6, wherein, The generation of the blood vessel thickness map by the geometric prediction head comprises: calculating an internal distance field of the blood vessel mask voxel set; extracting a centerline skeleton of the blood vessel mask voxel set through a topology refinement operator; For any point in the set of vessel mask voxels, a projection point of it on the centerline skeleton is generated by a nearest neighbor projection operator; a skeleton projection point radius is generated according to the principle of maximum inscribed sphere and the internal distance field; a vessel thickness value of any point in the set of vessel mask voxels is twice the skeleton projection point radius corresponding to it, so as to generate the vessel thickness map.

8. The method of claim 4, wherein, The global-local fusion module splices the generated decoder global semantic feature, the encoder shallow local texture feature, the encoder deep semantic abstract feature, the vessel distance map and the vessel thickness map along the channel dimension; inputs the spliced tensor into a fusion network to generate a fusion feature; maps a dedicated vessel token into a set of dynamic channel weights through a small MLP, which is used for dynamically weighting the fusion feature to generate a fusion map; and stacks the fusion map to the initial segmentation mask through a gated residual connection to generate a final segmentation mask.

9. The method of claim 1, wherein, The two-stage training strategy in S3 includes: Macro adaptation stage: freeze the 2D visual encoder, only train the 2.5D slice fusion adapter, the multi-head mask decoder and the global-local fusion module, use a learning rate of 1x10-3 to quickly adapt the 3D spatial structure and geometric prior task; Topology refinement stage: freeze the converged 2.5D slice fusion adapter, fine-tune the multi-head mask decoder and the global-local fusion module using a learning rate of 5x10 -5 to strengthen the model's perception of blood vessel boundaries, thickness, and topological connectivity.

10. A system for segmenting three-dimensional tubular structures based on geometric priors and topological awareness, the system comprising: Comprise: A data acquisition and preprocessing module for acquiring and preprocessing blood vessel images and obtaining a gold standard binary segmentation mask of the blood vessel images; a segmentation network model module based on a SAM network and a frozen 2D visual encoder, including a multi-head mask decoder, a global-local fusion module and a 2.5D slice fusion adapter for capturing inter-slice context; input the blood vessel image into the segmentation network model module, the multi-head mask decoder generates geometric prior maps including a vessel distance map and a vessel thickness map, and the global-local fusion module fuses the vessel distance map, the vessel thickness map and the multi-source feature map generated by the SAM network to generate a final segmentation mask; A model training module for training the segmentation network model module through a two-stage training strategy including a macro adaptation stage and a topology refinement stage; A segmentation execution module for inputting a three-dimensional medical image to be segmented into the trained segmentation network model module to obtain a voxel-level segmentation result; The three-dimensional tubular structure segmentation system based on geometric prior and topology perception is used to perform the three-dimensional tubular structure segmentation method based on geometric prior and topology perception as claimed in claim 1.

Images