Three-dimensional image blood vessel segmentation method and system
By pre-training a segmentation network and learning the optimal fine-tuning strategy through a proxy network, and combining a Gaussian signed distance graph-assisted target segmentation network to learn the geometric features of blood vessels, the problem of poor generalization of blood vessel segmentation on small sample datasets in existing technologies is solved, and fast and accurate cross-site blood vessel segmentation is achieved.
Patent Information
- Application Number
- CN202310302610.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-24
- Publication Date
- 2026-08-25
- Estimated Expiration
- 2043-03-24
AI Technical Summary
Existing 3D medical image blood vessel segmentation methods are prone to overfitting and have poor generalization when trained on small sample datasets, making it difficult to achieve fast and accurate cross-site adaptive blood vessel segmentation.
By employing segmentation network pre-training, surrogate network learning of optimal fine-tuning strategy, and Gaussian signed distance map-assisted target segmentation network learning of vascular geometric features, cross-site adaptive fine-tuning from source domain to target domain is achieved, and the parameters of convolutional filters are automatically selected for freezing or fine-tuning.
It improves the accuracy and robustness of vessel segmentation, reduces the need for large-scale labeled data, and achieves fast and effective target vessel segmentation results.
Smart Images
Figure CN118334046B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing technology, and in particular to a three-dimensional image blood vessel segmentation method and a three-dimensional image blood vessel segmentation system, which can automatically and accurately segment blood vessels from three-dimensional blood vessel images. Background Technology
[0002] Vascular diseases pose a serious threat to human health worldwide. Conditions such as atherosclerosis, vascular malformations, aneurysms, and aortic dissections are prevalent in blood vessels across various parts of the body, including the brain, heart, and liver. Analysis of vascular structure is essential for diagnosing vascular diseases. To alleviate patient suffering while effectively analyzing these lesions, non-invasive three-dimensional imaging modalities such as computed tomography angiography (CTA) and magnetic resonance angiography (MRA) have been increasingly used in clinical practice in recent years, replacing invasive imaging modalities such as X-ray angiography. The results of vascular segmentation represent the topological structure of the vascular network, and their completeness and accuracy affect the effectiveness and robustness of various vascular analysis applications.
[0003] Currently, vessel segmentation in 3D medical images is generally divided into two main categories: model-driven methods and data-driven methods. Model-driven methods rely on prior assumptions about the grayscale distribution and geometry of blood vessels in the image, extracting vessel structures through multi-scale filtering, minimum path methods, active contour models, and path tracking. This approach heavily depends on manually designed features and is ill-suited for complex 3D vessel segmentation tasks. Data-driven methods, particularly deep learning networks, can automatically extract features from medical images, thereby improving the accuracy and robustness of vessel segmentation. However, deep learning methods typically require large-scale labeled datasets for training; otherwise, training on small sample datasets can easily lead to overfitting and poor generalization of the trained model. Summary of the Invention
[0004] To overcome the shortcomings of the prior art, the technical problem to be solved by the present invention is to provide a three-dimensional image blood vessel segmentation method, which effectively realizes cross-site adaptive fine-tuning from the source domain to the target domain, thereby quickly and effectively obtaining accurate target blood vessel segmentation results.
[0005] The technical solution of this invention is: a three-dimensional image blood vessel segmentation method, which includes the following steps:
[0006] (1) The publicly available liver vascular CTA annotation dataset Task08_HepaticVessel was used as the source domain. A pre-trained model was generated through the source domain segmentation network to initialize the parameters of the target segmentation network.
[0007] (2) Using the target blood vessel image as the target domain, the agent network is used to learn the optimal fine-tuning strategy;
[0008] (3) Based on the fine-tuning strategy, the pre-training parameters of each convolutional filter in the target segmentation network are frozen or fine-tuned, and the target segmentation network is assisted in learning the geometric features of blood vessels through the Gaussian signed distance map.
[0009] This invention employs a segmentation network pre-training method to learn vascular anatomy from a source domain dataset. A proxy network learns the optimal fine-tuning strategy from the source to the target domain dataset, automatically selecting parameter freezes or fine-tunes for each filter in the target segmentation network, maximizing knowledge transfer efficiency. A Gaussian signed distance graph-assisted learning task encourages the target segmentation network to learn vascular geometry during transfer learning, improving the accuracy and robustness of vascular segmentation. Therefore, this invention's method can extract vascular segmentation results superior to other existing methods, alleviating the need for large amounts of labeled training data in deep learning methods. It effectively achieves cross-site adaptive fine-tuning from the source to the target domain, thus quickly and efficiently obtaining accurate target vascular segmentation results.
[0010] A three-dimensional image blood vessel segmentation system is also provided, which includes:
[0011] The segmentation network pre-training module is configured to use the publicly available liver vascular CTA annotation dataset Task08_HepaticVessel as the source domain, and generates a pre-trained model through the source domain segmentation network to initialize the parameters of the target segmentation network.
[0012] The fine-tuning strategy learning module is configured to use the target blood vessel image as the target domain and learn the optimal fine-tuning strategy using an agent network.
[0013] The segmentation-shape joint transfer learning module is configured to guide the selection of frozen or fine-tuned pre-trained parameters for each convolutional filter in the target segmentation network based on a fine-tuning strategy, and to assist the target segmentation network in learning vascular geometric features through a Gaussian signed distance map. Attached Figure Description
[0014] Figure 1 This is a flowchart of a three-dimensional image blood vessel segmentation method according to the present invention.
[0015] Figure 2 This is a flowchart of step (1) according to the present invention.
[0016] Figure 3 This is a flowchart of step (2) according to the present invention.
[0017] Figure 4 This is a flowchart of step (3) according to the present invention. Detailed Implementation
[0018] like Figure 1 As shown, this three-dimensional image blood vessel segmentation method includes the following steps:
[0019] (1) The publicly available liver vascular CTA annotation dataset Task08_HepaticVessel was used as the source domain. A pre-trained model was generated through the source domain segmentation network to initialize the parameters of the target segmentation network.
[0020] (2) Using the target blood vessel image as the target domain, the agent network is used to learn the optimal fine-tuning strategy;
[0021] (3) Based on the fine-tuning strategy, the pre-training parameters of each convolutional filter in the target segmentation network are frozen or fine-tuned, and the target segmentation network is assisted in learning the geometric features of blood vessels through the Gaussian signed distance map.
[0022] This invention employs a segmentation network pre-training method to learn vascular anatomy from a source domain dataset. A proxy network learns the optimal fine-tuning strategy from the source to the target domain dataset, automatically selecting parameter freezes or fine-tunes for each filter in the target segmentation network, maximizing knowledge transfer efficiency. A Gaussian signed distance graph-assisted learning task encourages the target segmentation network to learn vascular geometry during transfer learning, improving the accuracy and robustness of vascular segmentation. Therefore, this invention's method can extract vascular segmentation results superior to other existing methods, alleviating the need for large amounts of labeled training data in deep learning methods. It effectively achieves cross-site adaptive fine-tuning from the source to the target domain, thus quickly and efficiently obtaining accurate target vascular segmentation results.
[0023] Preferably, such as Figure 2 As shown, step (1) includes the following sub-steps:
[0024] (1.1) For the input source domain blood vessel image, the depth features of the blood vessel image are extracted using the convolutional encoder and convolutional decoder in the source domain segmentation network;
[0025] (1.2) Based on the extracted depth features, the source domain vessel segmentation prediction results are output using the Sigmoid activation function;
[0026] (1.3) Retain the model after the source domain segmentation network has been trained as a pre-trained model.
[0027] Preferably, such as Figure 3 As shown, step (2) includes the following sub-steps:
[0028] (2.1) Using the target domain blood vessel image as input, the image features are learned through a proxy network composed of convolutional blocks, average pooling layers and fully connected layers;
[0029] (2.2) The gradient of the proxy network during backpropagation is estimated using the Gumbel-Softmax strategy, so that the output of the proxy network can be discretized. Predict and classify each filter of the target segmentation network, and choose to freeze or fine-tune the parameters.
[0030] Preferably, such as Figure 4 As shown, step (3) includes the following sub-steps:
[0031] (3.1) Transfer the pre-trained model parameters to the target segmentation network as the initialization parameters of the target segmentation network;
[0032] (3.2) Using the target domain blood vessel image as input, the filter selection parameters of the convolutional encoder and convolutional decoder in the target segmentation network are adaptively frozen or fine-tuned through the optimal fine-tuning strategy to realize knowledge transfer from the source domain image to the target domain image.
[0033] (3.3) The depth features of the target domain blood vessel image are extracted by the convolutional encoder and convolutional decoder in the target segmentation network, and the Gaussian signed distance map prediction results and the target domain blood vessel segmentation prediction results are output by the Tanh activation function and the Sigmoid activation function, respectively.
[0034] Preferably, in step (3.3), the Gaussian signed distance map models the response of the blood vessel cross-section by introducing a Gaussian kernel, so that the blood vessel centerline has the highest and the same absolute response value. The farther other points within the blood vessel structure are from the centerline, the lower their absolute response value. The absolute response value of a pixel in the background is defined as the minimum Euclidean distance from that point to the blood vessel boundary. The response values of the blood vessel region and the background have opposite signs, as expressed by the following formula:
[0035]
[0036] Where p is any point on the gold standard image Ω for target domain vessel segmentation, Z represents the vessel centerline, and z represents a point on the vessel centerline. Ω in and Ω out Let p represent the vessel boundary, the internal region of the vessel, and the background region of the vessel image Ω, respectively. ||pz||² represents the Euclidean distance between points p and z, and ||pq||² is defined similarly. η = 2 is the coefficient of the Gaussian kernel, and σ... p It is the radius of the centerline point closest to point p, and ε = 10 represents the parameter of the cutoff range in the vascular background.
[0037] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium. When executed, the program includes the steps of the methods of the above embodiments. The storage medium can be ROM / RAM, magnetic disk, optical disk, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a three-dimensional image blood vessel segmentation system, which is typically represented in the form of functional modules corresponding to the steps of the method. The system includes:
[0038] The segmentation network pre-training module is configured to use the publicly available liver vascular CTA annotation dataset Task08_HepaticVessel as the source domain, and generates a pre-trained model through the source domain segmentation network to initialize the parameters of the target segmentation network.
[0039] The fine-tuning strategy learning module is configured to use the target blood vessel image as the target domain and learn the optimal fine-tuning strategy using an agent network.
[0040] The segmentation-shape joint transfer learning module is configured to guide the selection of frozen or fine-tuned pre-trained parameters for each convolutional filter in the target segmentation network based on a fine-tuning strategy, and to assist the target segmentation network in learning vascular geometric features through a Gaussian signed distance map.
[0041] Preferably, the segmentation network pre-training module performs:
[0042] (1.1) For the input source domain blood vessel image, the depth features of the blood vessel image are extracted using the convolutional encoder and convolutional decoder in the source domain segmentation network;
[0043] (1.2) Based on the extracted depth features, the source domain vessel segmentation prediction results are output using the Sigmoid activation function;
[0044] (1.3) Retain the model after the source domain segmentation network has been trained as a pre-trained model.
[0045] Preferably, the fine-tuning strategy learning module performs:
[0046] (2.1) Using the target domain blood vessel image as input, the image features are learned through a proxy network composed of convolutional blocks, average pooling layers and fully connected layers;
[0047] (2.2) Estimating the ladder during backpropagation of the agent network using the Gumbel-Softmax strategy
[0048] The degree allows the agent network to discretize its output, predict and classify each filter of the target segmentation network, and select parameters to freeze or fine-tune.
[0049] Preferably, the segmentation-shape joint transfer learning module performs:
[0050] (3.1) Transfer the pre-trained model parameters to the target segmentation network as the initialization parameters of the target segmentation network;
[0051] (3.2) Using the target domain blood vessel image as input, the filter selection parameters of the convolutional encoder and convolutional decoder in the target segmentation network are adaptively frozen or fine-tuned through the optimal fine-tuning strategy to realize knowledge transfer from the source domain image to the target domain image.
[0052] (3.3) The depth features of the target domain blood vessel image are extracted by the convolutional encoder and convolutional decoder in the target segmentation network, and the Gaussian signed distance map prediction results and the target domain blood vessel segmentation prediction results are output by the Tanh activation function and the Sigmoid activation function, respectively.
[0053] Preferably, in the segmentation-shape joint transfer learning module, the Gaussian signed distance map models the response of the blood vessel cross-section by introducing a Gaussian kernel, so that the blood vessel centerline has the highest and the same absolute response value. The farther other points within the blood vessel structure are from the centerline, the lower their absolute response value. The absolute response value of a pixel in the background is defined as the minimum Euclidean distance from that point to the blood vessel boundary. The response values of the blood vessel region and the background have opposite signs, as expressed by the following formula:
[0054]
[0055] Where p is any point on the gold standard image Ω for target domain vessel segmentation, Z represents the vessel centerline, and z represents a point on the vessel centerline. Ω in and Ω out Let p represent the vessel boundary, the internal region of the vessel, and the background region of the vessel image Ω, respectively. ||pz||² represents the Euclidean distance between points p and z, and ||pq||² is defined similarly. η = 2 is the coefficient of the Gaussian kernel, and σ... p It is the radius of the centerline point closest to point p, and ε = 10 represents the parameter of the cutoff range in the vascular background.
[0056] This invention learns vascular anatomy from source domain vascular images to assist in vascular structure segmentation in target domain images, alleviating the need for labeled training data in deep learning methods, improving the accuracy of target vascular segmentation, and providing more precise structural information for subsequent vascular lesion analysis.
[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A three-dimensional image blood vessel segmentation method, characterized in that: It includes the following steps: (1) The publicly available liver vascular CTA annotation dataset Task08_HepaticVessel is used as the source domain. A pre-trained model is generated through the source domain segmentation network to initialize the parameters of the target segmentation network. (2) Using the target blood vessel image as the target domain, the agent network is used to learn the optimal fine-tuning strategy; (3) Based on the fine-tuning strategy, the pre-training parameters of each convolutional filter in the target segmentation network are frozen or fine-tuned, and the target segmentation network is assisted in learning the geometric features of blood vessels through the Gaussian signed distance map; Step (3) includes the following sub-steps: (3.1) Transfer the pre-trained model parameters to the target segmentation network as the initialization parameters of the target segmentation network; (3.2) Using the target domain blood vessel image as input, the filter selection parameters of the convolutional encoder and convolutional decoder in the target segmentation network are adaptively frozen or fine-tuned through the optimal fine-tuning strategy to realize knowledge transfer from the source domain image to the target domain image; (3.3) The depth features of the target domain blood vessel image are extracted by the convolutional encoder and convolutional decoder in the target segmentation network, and the Gaussian signed distance map prediction results and the target domain blood vessel segmentation prediction results are output by the Tanh activation function and the Sigmoid activation function, respectively. In step (3.3), the Gaussian signed distance map models the response of the blood vessel cross-section by introducing a Gaussian kernel, so that the blood vessel centerline has the highest and the same absolute response value. The farther other points in the blood vessel structure are from the centerline, the lower their absolute response value. The absolute response value of a pixel in the background is defined as the minimum Euclidean distance from that point to the blood vessel boundary. The response values of the blood vessel region and the background have opposite signs, as expressed by the following formula: ， in, Gold standard image for target domain vessel segmentation any point on, Indicates the center line of the blood vessel. Points on the central line of the blood vessel , and Images of blood vessels are shown below. The vascular boundary, the internal region of the vascular body, and the background region. Point and points The Euclidean distance between them The definition is similar. These are the coefficients of the Gaussian kernel. Distance point The radius of the nearest centerline point A parameter indicating the cutoff range in the vascular background.
2. The three-dimensional image blood vessel segmentation method according to claim 1, characterized in that: Step (1) includes the following sub-steps: (1.1) For the input source domain blood vessel image, the depth features of the blood vessel image are extracted using the convolutional encoder and convolutional decoder in the source domain segmentation network; (1.2) Based on the extracted depth features, the source domain vessel segmentation prediction results are output using the Sigmoid activation function; (1.3) Retain the model after the source domain segmentation network has been trained as a pre-trained model.
3. The three-dimensional image blood vessel segmentation method according to claim 2, characterized in that: Step (2) includes the following sub-steps: (2.1) Using the target domain blood vessel image as input, the image features are learned through a proxy network composed of convolutional blocks, average pooling layers and fully connected layers; (2.2) The gradient of the proxy network during backpropagation is estimated by using the Gumbel-Softmax strategy, so that the output of the proxy network can be discretized. The prediction and classification of each filter of the target segmentation network are performed, and the parameters are frozen or fine-tuned.
4. A three-dimensional image blood vessel segmentation system, characterized in that: It includes: The segmentation network pre-training module is configured to use the publicly available liver vascular CTA annotation dataset Task08_HepaticVessel as the source domain, and generates a pre-trained model through the source domain segmentation network to initialize the parameters of the target segmentation network. The fine-tuning strategy learning module is configured to use the target blood vessel image as the target domain and learn the optimal fine-tuning strategy using an agent network. The segmentation-shape joint transfer learning module is configured to guide the selection of frozen or fine-tuned pre-trained parameters for each convolutional filter in the target segmentation network based on a fine-tuning strategy, and to assist the target segmentation network in learning vascular geometric features through a Gaussian signed distance map. The segmentation-shape joint transfer learning module performs the following: (3.1) Transfer the pre-trained model parameters to the target segmentation network as the initialization parameters of the target segmentation network; (3.2) Using the target domain blood vessel image as input, the filter selection parameters of the convolutional encoder and convolutional decoder in the target segmentation network are adaptively frozen or fine-tuned through the optimal fine-tuning strategy to realize knowledge transfer from the source domain image to the target domain image. (3.3) The depth features of the target domain blood vessel image are extracted by the convolutional encoder and convolutional decoder in the target segmentation network, and the Gaussian signed distance map prediction results and the target domain blood vessel segmentation prediction results are output by the Tanh activation function and the Sigmoid activation function, respectively. In the segmentation-shape joint transfer learning module, the Gaussian signed distance map models the response of the blood vessel cross-section by introducing a Gaussian kernel, ensuring that the blood vessel centerline has the highest and identical absolute response value. The further other points within the blood vessel structure are from the centerline, the lower their absolute response value. The absolute response value of a pixel in the background is defined as the minimum Euclidean distance from that point to the blood vessel boundary. The response values of the blood vessel region and the background have opposite signs, as expressed by the following formula: ， in, Gold standard image for target domain vessel segmentation any point on, Indicates the center line of the blood vessel. Points on the central line of the blood vessel , and Images of blood vessels are shown below. The vascular boundary, the internal region of the vascular body, and the background region. Point and points The Euclidean distance between them The definition is similar. These are the coefficients of the Gaussian kernel. Distance point The radius of the nearest centerline point A parameter indicating the cutoff range in the vascular background.
5. The three-dimensional image blood vessel segmentation system according to claim 4, characterized in that: The segmentation network pre-training module performs the following: (1.1) For the input source domain blood vessel image, the depth features of the blood vessel image are extracted using the convolutional encoder and convolutional decoder in the source domain segmentation network; (1.2) Based on the extracted depth features, the source domain vessel segmentation prediction results are output using the Sigmoid activation function; (1.3) Retain the model after the source domain segmentation network has been trained as a pre-trained model.
6. The three-dimensional image blood vessel segmentation system according to claim 5, characterized in that: The fine-tuning strategy learning module executes as follows: (2.1) Using the target domain blood vessel image as input, the image features are learned through a proxy network composed of convolutional blocks, average pooling layers and fully connected layers; (2.2) The gradient of the proxy network during backpropagation is estimated by using the Gumbel-Softmax strategy, so that the output of the proxy network can be discretized. Predict and classify each filter of the target segmentation network, and choose to freeze or fine-tune the parameters.