Intracranial aneurysm automatic segmentation method and system

Abstract

The invention discloses an intracranial aneurysm automatic segmentation method and system, and the method comprises the steps: constructing a cerebral vessel segmentation pre-training network, obtaining the weight of a pre-training model, and constructing an intracranial aneurysm segmentation network; the aneurysm segmentation network is composed of a double-branch artery-aneurysm encoder and an aneurysm decoder, a grouping self-adaptive inquiry attention mechanism is constructed, aneurysm feature expression is enhanced by using artery features of the encoder, and interference of an artery structure on aneurysm segmentation is reduced; a depth separable cavity convolution module is constructed in a depth supervision framework of an aneurysm decoder, and multi-scale features are captured through three parallel depth separable cavity convolution branches. Modeling is carried out on the context relationship between the intracranial artery structure and the aneurysm, pre-trained blood vessel features are migrated to guide aneurysm segmentation, instability possibly caused by dependence on pre-segmentation of the artery structure is reduced, and accurate segmentation of aneurysms of various sizes is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of image processing, and in particular relates to an automatic segmentation method and system for intracranial aneurysms. Background Art

[0002] Intracranial aneurysms are vascular lesions caused by abnormal dilation of intracranial arteries. Their rupture may lead to subarachnoid hemorrhage, seriously threatening the patient's life and health. Intracranial aneurysm segmentation using three-dimensional medical images facilitates positioning and morphological assessment, and plays a vital role in aneurysm screening, hemodynamic analysis, risk assessment, and surgical planning. Manual segmentation of aneurysms by experienced physicians is clinically necessary, but is very time-consuming and labor-intensive. Therefore, an automatic and accurate three-dimensional intracranial aneurysm segmentation method is urgently needed. In recent years, deep learning has driven the development of lesion recognition technology in medical imaging. However, due to the small size, variable morphology, and close connection of aneurysms to arterial structures, accurate aneurysm segmentation remains a challenging task.

[0003] Existing methods have attempted to address the above challenges, but most methods fail to effectively utilize the contextual information of arterial structure and aneurysms or fail to set up a good fusion method, resulting in limited segmentation accuracy. For example, a dual-feature fusion MRA image intracranial aneurysm segmentation method with publication number CN115147404A only uses multiple feature enhancement fusions of the same input image, resulting in limited segmentation accuracy. Summary of the Invention

[0004] The present invention aims to provide an automatic segmentation method for intracranial aneurysms, which models the contextual relationship between the intracranial artery structure and the aneurysm structure, thereby improving the segmentation accuracy.

[0005] To solve the above technical problems, the present invention adopts the following technical solution: a method for automatic segmentation of intracranial aneurysms, comprising:

[0006] Step A: Construct a pre-trained network for cerebral vascular segmentation and obtain the pre-trained model weights;

[0007] Step B: Constructing intracranial aneurysm segmentation network;

[0008] In step B, the intracranial aneurysm segmentation network includes at least an artery feature encoder and an aneurysm feature encoder that are parallel and independently controllable;

[0009] Step B includes:

[0010] Step b1: using the aneurysm image to be segmented as input to the intracranial aneurysm segmentation network;

[0011] Step b2: passing the input image through the artery feature encoder and the aneurysm feature encoder to obtain artery coding features and aneurysm coding features;

[0012] Step b3: Adaptively modulating the artery coding features and the aneurysm coding features, and obtaining artery enhancement fusion features and artery suppression fusion features respectively through adaptive addition and subtraction calculations;

[0013] Step b4: linearly project the artery enhancement fusion features and the artery suppression fusion features respectively, and straighten them into vectors by reshaping to obtain the first query vector and the second query vector of the features;

[0014] Step b5: Perform multiple linear projections and reshapes on the aneurysm coding features, and take two sets of feature vectors, namely the first key vector and the first value vector, and the second key vector and the second value vector;

[0015] Step b6: Group the first query vector, the first key vector, and the first value vector into one group and perform attention calculation on them. Group the second query vector, the second key vector, and the second value vector into another group and perform attention calculation on them, to obtain two groups of attention feature vectors.

[0016] Step b7: Linearly project and reshape the two sets of attention feature vectors to restore them into feature tensors, and input them into the convolutional layer. Then, feature fusion is performed by adding them to the input aneurysm encoding features to obtain the encoder's aneurysm enhancement features.

[0017] Step b8: inputting the aneurysm enhancement feature of the encoder into the aneurysm decoder to obtain the aneurysm decoding feature;

[0018] Step b9: The aneurysm decoding features are subjected to multi-scale feature extraction through a depthwise separable dilated convolution module. Finally, the multi-scale features of each stage of the aneurysm feature decoder are deeply supervised to predict the aneurysm segmentation results.

[0019] In another embodiment, in step A, the cerebrovascular segmentation pre-training network adopts the Unet encoder-decoder architecture, wherein the encoder of the cerebrovascular segmentation pre-training network is composed of several convolutional layers and several maximum pooling layers, and the decoder of the cerebrovascular segmentation pre-training network is composed of a transposed convolutional layer and a convolutional layer.

[0020] In another embodiment, in step A, the cerebral vascular segmentation pre-training network is pre-trained based on the public TubeTK cerebral vascular dataset to predict the cerebral vascular segmentation results and obtain the cerebral vascular segmentation pre-training model weights.

[0021] In another embodiment, in step B, the initialization parameters of the artery feature encoder of the intracranial aneurysm segmentation network are obtained by migrating the encoder weights of the cerebral vascular segmentation pre-training model, and are frozen in the aneurysm segmentation training task, and the aneurysm feature encoder remains in a trainable state.

[0022] In another embodiment, in step B, the initialization parameters of the aneurysm feature encoder of the intracranial aneurysm segmentation network are initialized randomly.

[0023] In another embodiment, in step b3, the algorithm for obtaining the arterial enhancement fusion feature and the arterial suppression fusion feature is:

[0024] F Add =F Ar +Sigmoid(λ1)×F An … (1)

[0025] F sub =F An -Sigmoid(λ2)×F Ar … (2)

[0026] Among them, F Ar and F An Represent artery coding features and aneurysm coding features respectively; F Add and F sub They represent arterial enhancement fusion features and arterial suppression fusion features, respectively, to suppress the interference of arterial structure on aneurysms; λ1 and λ2 are learnable parameters used to control the modulation amplitude of the features; the Sigmoid function can normalize λ1 and λ2 to the range of [0,1].

[0027] In another embodiment, in step b9, when the aneurysm decoding features are subjected to multi-scale feature extraction by the depthwise separable atrous convolution module, the decoder features of each stage are used as input features, and the features are grouped according to the channel dimension to obtain multiple groups of features; each group of features is input into a depthwise separable atrous convolution branch that corresponds to each group of features and has a different receptive field to obtain multi-scale features of the corresponding number of groups.

[0028] In another implementation mode, in step b9, the decoder features of each stage are used as input features, and the features are grouped according to a certain dimensional ratio of the channel to obtain three groups of features; the three groups of features are respectively input into the first depth-separable dilated convolution, the second depth-separable dilated convolution, and the third depth-separable dilated convolution with different receptive fields to obtain multi-scale features of the corresponding number of groups, wherein the convolution kernel size of the first depth-separable dilated convolution is 3×3×3, 5×5×5 or 7×7×7, and the dilation rate is 1, 2 or 3; the convolution kernel size of the second depth-separable dilated convolution is 3×3×3, 5×5×5 or 7×7×7, and the dilation rate is 1, 2 or 3; the convolution kernel size of the third depth-separable dilated convolution is 3×3×3, 5×5×5 or 7×7×7, and the dilation rate is 1, 2 or 3.

[0029] In another implementation, the convolution kernel size of the first depth-separable dilated convolution is 5×5×5, and the dilation ratio is 2; the convolution kernel size of the second depth-separable dilated convolution is 5×5×5, and the dilation ratio is 1; the convolution kernel size of the third depth-separable dilated convolution is 3×3×3, and the dilation ratio is 1.

[0030] In another embodiment, the dimensional ratio is 1:1:2, 1:1:1, 1:2:2 or 1:2:3.

[0031] In another implementation, in step b9, the multi-scale features of the corresponding number of groups are merged in the channel dimension and added to the input features of the decoder to obtain the output features.

[0032] In another embodiment, the composition of the artery feature encoder and the aneurysm feature encoder is consistent with the encoder composition of the cerebral vascular segmentation pre-training network, and the composition of the aneurysm feature decoder is consistent with the decoder composition of the cerebral vascular segmentation pre-training network.

[0033] The present invention also provides an automatic segmentation system for intracranial aneurysms, which comprises:

[0034] A cerebral vascular segmentation pre-training module is configured to construct a cerebral vascular segmentation pre-training network and obtain pre-training model weights;

[0035] The intracranial aneurysm segmentation module is configured to take the aneurysm image to be segmented as the input of the intracranial aneurysm segmentation network; pass the input image through the artery feature encoder and the aneurysm feature encoder to obtain the artery coding feature and the aneurysm coding feature; perform adaptive feature modulation on the artery coding feature and the aneurysm coding feature, and obtain the artery enhancement fusion feature and the artery suppression fusion feature respectively through adaptive addition and subtraction calculation; linearly project the artery enhancement fusion feature and the artery suppression fusion feature respectively, and straighten them into vectors by reshaping to obtain the first feature query vector and the second feature query vector; perform multiple linear projections and reshaping on the aneurysm coding feature, and take two groups of feature vectors, namely the first key vector and the first value vector, and the second key vector and the second value vector; the first query vector, the first key vector and the first value vector are divided into one group and attention calculation is performed, the second query vector, the second key vector and the second value vector are divided into another group and attention calculation is performed to obtain two groups of attention feature vectors; the two groups of attention feature vectors are linearly projected and reshaped to recover the feature tensors, and input into the convolution layer, and then the feature fusion is performed by adding them with the input aneurysm encoding features to obtain the aneurysm enhancement features of the encoder; the aneurysm enhancement features of the encoder are input into the aneurysm decoder to obtain the aneurysm decoding features; the aneurysm decoding features are subjected to multi-scale feature extraction through a depthwise separable atrous convolution module, and finally the multi-scale features of each stage of the aneurysm feature decoder are deeply supervised to predict the aneurysm segmentation results.

[0036] The beneficial effects of the present invention are: designing an attention network guided by arterial knowledge to model the contextual relationship between intracranial arterial structure and aneurysm, migrating pre-trained vascular features to guide aneurysm segmentation, and reducing the instability that may be caused by relying on pre-segmented arterial structure; designing a group adaptive query attention mechanism, through an artery-aware adaptive group query strategy, calculating the attention of artery and aneurysm features, achieving aneurysm feature enhancement, and suppressing the interference of vascular features on aneurysm segmentation; designing a deep separable void convolution module, strengthening multi-scale feature learning at different resolution levels in deep supervision, and achieving accurate segmentation of aneurysms of various sizes; accurate intracranial aneurysm segmentation can effectively assist in the diagnosis and treatment of aneurysm diseases, such as rupture risk assessment, interventional path planning, surgical navigation, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Flowchart of the implementation method of the system of the present invention;

[0038] Figure 2 This is an algorithm flow chart of the group adaptive query attention mechanism of the system of the present invention;

[0039] Figure 3This is an algorithm flow chart of the depth-separable dilated convolution module of the system of the present invention. DETAILED DESCRIPTION

[0040] The present invention is described in detail below with reference to the embodiments shown in the accompanying drawings:

[0041] like Figure 1 As shown, the automatic segmentation method of intracranial aneurysms includes:

[0042] Step A: Construct a pre-trained network for cerebrovascular segmentation and obtain pre-trained model weights. The network uses the Unet encoder-decoder architecture. The encoder consists of several convolutional layers and several max-pooling layers, and the decoder consists of transposed convolutional layers and convolutional layers. The network is pre-trained on the public TubeTK cerebrovascular dataset to predict cerebrovascular segmentation results and obtain pre-trained model weights.

[0043] Step B: Construct an intracranial aneurysm segmentation network; wherein, the intracranial aneurysm segmentation network includes at least a parallel and independently controllable artery feature encoder and an aneurysm feature encoder, the composition of the artery feature encoder and the aneurysm feature encoder is consistent with the encoder composition of the cerebral vascular segmentation pre-training network, and the aneurysm feature decoder is consistent with the decoder composition of the cerebral vascular segmentation pre-training network; the initialization parameters of the artery feature encoder of the intracranial aneurysm segmentation network are obtained by migrating the encoder weights of the cerebral vascular segmentation pre-training model and are frozen in the aneurysm segmentation training task; the initialization parameters of the aneurysm feature encoder of the intracranial aneurysm segmentation network are randomly initialized, and the aneurysm feature encoder remains in a trainable state.

[0044] Step B specifically includes:

[0045] Step b1: using the aneurysm image to be segmented as input to the intracranial aneurysm segmentation network;

[0046] Step b2: passing the input image through the artery feature encoder and the aneurysm feature encoder to obtain artery coding features and aneurysm coding features;

[0047] Step b3: Figure 2 As shown in FIG, in order to fuse the artery features and aneurysm features of the dual-branch encoder, the artery coding features and the aneurysm coding features are adaptively modulated. The artery enhancement fusion feature and the artery suppression fusion feature can be obtained by adaptive addition and subtraction calculations, respectively. The algorithm for obtaining the artery enhancement fusion feature and the artery suppression fusion feature is as follows:

[0048] F Add =F Ar +Sigmoid(λ1)×FAn … (1)

[0049] F sub =F An -Sigmoid(λ2)×F Ar … (2)

[0050] Among them, F Ar and F An Represent artery coding features and aneurysm coding features respectively; F Add and F sub They represent arterial enhancement fusion features and arterial suppression fusion features, respectively, to suppress the interference of arterial structure on aneurysms; λ1 and λ2 are learnable parameters used to control the modulation amplitude of the features; the Sigmoid function can normalize λ1 and λ2 to the range of [0,1];

[0051] Step b4: linearly project the artery enhancement fusion features and the artery suppression fusion features respectively, and straighten them into vectors by reshaping to obtain the first query vector and the second query vector of the features;

[0052] Step b5: Perform multiple linear projections and reshapes on the aneurysm coding features, and take two sets of feature vectors, namely the first key vector and the first value vector, and the second key vector and the second value vector;

[0053] Step b6: Group the first query vector, the first key vector, and the first value vector into one group and perform attention calculation on them. Group the second query vector, the second key vector, and the second value vector into another group and perform attention calculation on them, to obtain two groups of attention feature vectors.

[0054] Step b7: Linearly project and reshape the two sets of attention feature vectors to restore them into feature tensors, and input them into the convolutional layer. Then, feature fusion is performed by adding them to the input aneurysm encoding features to obtain the encoder's aneurysm enhancement features.

[0055] Step b8: inputting the aneurysm enhancement feature of the encoder into the aneurysm decoder to obtain the aneurysm decoding feature;

[0056] Step b9: Figure 3As shown, in order to effectively extract the multi-scale aneurysm features, the aneurysm decoding features are subjected to multi-scale feature extraction by the depthwise separable atrous convolution module, and finally the multi-scale features of each stage of the aneurysm feature decoder are deeply supervised to predict the aneurysm segmentation results; when the aneurysm decoding features are subjected to multi-scale feature extraction by the depthwise separable atrous convolution module, the decoder features of each stage are used as input features, and the features are grouped according to the channel dimension to obtain multiple groups of features; each group of features is input into the depthwise separable atrous convolution branch corresponding to each group of features and having different receptive fields to obtain the corresponding number of multi-scale features; the decoder features of each stage are used as input features, and the features are grouped according to a certain dimension ratio of the channel to obtain three groups of features; the three groups of features are input into the first depthwise separable atrous convolution, the second depthwise separable atrous convolution, and the third depthwise separable atrous convolution with different receptive fields to obtain the corresponding number of multi-scale features, among which the first depthwise separable atrous convolution has different receptive fields. The convolution kernel size of the first depth-separable dilated convolution is 3×3×3, 5×5×5 or 7×7×7, and the dilation rate is 1, 2 or 3; the convolution kernel size of the second depth-separable dilated convolution is 3×3×3, 5×5×5 or 7×7×7, and the dilation rate is 1, 2 or 3; the convolution kernel size of the third depth-separable dilated convolution is 3×3×3, 5×5×5 or 7×7×7, and the dilation rate is 1, 2 or 3. In this embodiment, the convolution kernel size of the first depth-separable dilated convolution is 5× 5×5, with a dilation rate of 2; the convolution kernel size of the second depth-wise separable dilated convolution is 5×5×5, with a dilation rate of 1; the convolution kernel size of the third depth-wise separable dilated convolution is 3×3×3, with a dilation rate of 1; the dimension ratio can be 1:1:2, 1:1:1, 1:2:2 or 1:2:3, etc. In this embodiment, the dimension ratio is 1:1:2; the multi-scale features obtained in the corresponding groups are merged in the channel dimension and added to the input features of the decoder to obtain the output features.

[0057] Intracranial aneurysm automatic segmentation system, which includes:

[0058] A cerebral vascular segmentation pre-training module is configured to construct a cerebral vascular segmentation pre-training network and obtain pre-training model weights;

[0059] The intracranial aneurysm segmentation module is configured to take the aneurysm image to be segmented as the input of the intracranial aneurysm segmentation network; pass the input image through the artery feature encoder and the aneurysm feature encoder to obtain the artery coding feature and the aneurysm coding feature; perform adaptive feature modulation on the artery coding feature and the aneurysm coding feature, and obtain the artery enhancement fusion feature and the artery suppression fusion feature respectively through adaptive addition and subtraction calculation; linearly project the artery enhancement fusion feature and the artery suppression fusion feature respectively, and straighten them into vectors by reshaping to obtain the first feature query vector and the second feature query vector; perform multiple linear projections and reshaping on the aneurysm coding feature, and take two groups of feature vectors, namely the first key vector and the first value vector, and the second key vector and the second value vector; the first query vector, the first key vector and the first value vector are divided into one group and attention calculation is performed, the second query vector, the second key vector and the second value vector are divided into another group and attention calculation is performed to obtain two groups of attention feature vectors; the two groups of attention feature vectors are linearly projected and reshaped to recover the feature tensors, and input into the convolution layer, and then the feature fusion is performed by adding them with the input aneurysm encoding features to obtain the aneurysm enhancement features of the encoder; the aneurysm enhancement features of the encoder are input into the aneurysm decoder to obtain the aneurysm decoding features; the aneurysm decoding features are subjected to multi-scale feature extraction through a depthwise separable atrous convolution module, and finally the multi-scale features of each stage of the aneurysm feature decoder are deeply supervised to predict the aneurysm segmentation results.

[0060] The above embodiments are intended only to illustrate the technical concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the present invention and implement it accordingly. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made in accordance with the spirit of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A method for automatic segmentation of intracranial aneurysms, comprising: Step A: Construct a pre-trained network for cerebral vascular segmentation and obtain the pre-trained model weights; Step B: Constructing intracranial aneurysm segmentation network; The method is characterized in that: in step B, the intracranial aneurysm segmentation network includes at least an artery feature encoder and an aneurysm feature encoder that are parallel and independently controllable; Step B includes: Step b1: The aneurysm image to be segmented is used as the input of the intracranial aneurysm segmentation network; Step b2: passing the input image through the artery feature encoder and the aneurysm feature encoder to obtain artery coding features and aneurysm coding features; Step b3: Adaptively modulating the artery coding features and the aneurysm coding features, and obtaining artery enhancement fusion features and artery suppression fusion features respectively through adaptive addition and subtraction calculations; Step b4: linearly project the artery enhancement fusion features and the artery suppression fusion features respectively, and straighten them into vectors by reshaping to obtain the first query vector and the second query vector of the features; Step b5: Perform multiple linear projections and reshapes on the aneurysm coding features, and take two sets of feature vectors, namely the first key vector and the first value vector, and the second key vector and the second value vector; Step b6: Group the first query vector, the first key vector, and the first value vector into one group and perform attention calculation on them. Group the second query vector, the second key vector, and the second value vector into another group and perform attention calculation on them, to obtain two groups of attention feature vectors. Step b7: Linearly project and reshape the two sets of attention feature vectors to restore them into feature tensors, and input them into the convolutional layer. Then, feature fusion is performed by adding them to the input aneurysm encoding features to obtain the encoder's aneurysm enhancement features. Step b8: inputting the aneurysm enhancement feature of the encoder into the aneurysm decoder to obtain the aneurysm decoding feature; Step b9: The aneurysm decoding features are subjected to multi-scale feature extraction through a depthwise separable dilated convolution module. Finally, the multi-scale features of each stage of the aneurysm feature decoder are deeply supervised to predict the aneurysm segmentation results.

2. The method for automatic segmentation of intracranial aneurysms according to claim 1, characterized in that: In step A, the cerebrovascular segmentation pre-training network adopts the Unet encoder-decoder architecture, where the encoder of the cerebrovascular segmentation pre-training network consists of several convolutional layers and several maximum pooling layers, and the decoder of the cerebrovascular segmentation pre-training network consists of a transposed convolutional layer and a convolutional layer.

3. The method for automatic segmentation of intracranial aneurysms according to claim 1, characterized in that: In step A, the cerebrovascular segmentation pre-training network is pre-trained based on the public TubeTK cerebrovascular dataset to predict the cerebrovascular segmentation results and obtain the cerebrovascular segmentation pre-training model weights.

4. The method for automatic segmentation of intracranial aneurysms according to claim 1, characterized in that: In step B, the initialization parameters of the arterial feature encoder of the intracranial aneurysm segmentation network are obtained by migrating the encoder weights of the cerebral vascular segmentation pre-training model, and are frozen during the aneurysm segmentation training task, and the aneurysm feature encoder remains in a trainable state.

5. The method for automatic segmentation of intracranial aneurysms according to claim 1, characterized in that: In step B, the initialization parameters of the aneurysm feature encoder of the intracranial aneurysm segmentation network are randomly initialized.

6. The method for automatic segmentation of intracranial aneurysms according to claim 1, characterized in that: In step b3, the algorithm for obtaining arterial enhancement fusion features and arterial suppression fusion features is: F Add =F Ar +Si gmo i d(λ1) ×F An … (1) F sub =F An -Sigmoi d(λ2) ×F Ar … (2) Among them, F Ar and F An Represent artery coding features and aneurysm coding features respectively; F Add and F sub They represent arterial enhancement fusion features and arterial suppression fusion features respectively; λ1 and λ2 are learnable parameters used to control the modulation amplitude of the features; the Sigmoid function can normalize λ1 and λ2 to the range of [0,1].

7. The method for automatic segmentation of intracranial aneurysms according to claim 1, characterized in that: In step b9, when the aneurysm decoding features are subjected to multi-scale feature extraction by the depthwise separable dilated convolution module, the decoder features of each stage are used as input features, and the features are grouped according to the channel dimension to obtain multiple groups of features; Each group of features is input into a depth-wise separable dilated convolution branch with different receptive fields corresponding to each group of features to obtain multi-scale features of the corresponding number of groups.

8. The method for automatic segmentation of intracranial aneurysms according to claim 7, characterized in that: In step b9, the decoder features of each stage are used as input features, and the features are grouped according to a certain dimension ratio of the channel to obtain three groups of features; the three groups of features are respectively input into the first depth-separable dilated convolution, the second depth-separable dilated convolution, and the third depth-separable dilated convolution with different receptive fields to obtain multi-scale features of the corresponding number of groups, where the convolution kernel size of the first depth-separable dilated convolution is 3×3×3, 5×5×5, or 7×7×7, and the dilation rate is 1, 2, or 3; The convolution kernel size of the second depth-wise separable dilated convolution is 3×3×3, 5×5×5 or 7×7×7, and the dilation rate is 1, 2 or 3; the convolution kernel size of the third depth-wise separable dilated convolution is 3×3×3, 5×5×5 or 7×7×7, and the dilation rate is 1, 2 or 3.

9. The method for automatic segmentation of intracranial aneurysms according to claim 8, characterized in that: The convolution kernel size of the first depth-wise separable dilated convolution is 5×5×5, and the dilation ratio is 2; the convolution kernel size of the second depth-wise separable dilated convolution is 5×5×5, and the dilation ratio is 1; the convolution kernel size of the third depth-wise separable dilated convolution is 3×3×3, and the dilation ratio is 1.

10. The method for automatic segmentation of intracranial aneurysms according to claim 7, characterized in that: The dimension ratio is 1:1:2, 1:1:1, 1:2:2, or 1:2:

3.

11. The method for automatic segmentation of intracranial aneurysms according to claim 7, characterized in that: In step b9, the multi-scale features of the corresponding number of groups are merged in the channel dimension and added to the input features of the decoder to obtain the output features.

12. The method for automatic segmentation of intracranial aneurysms according to claim 1, characterized in that: The composition of the artery feature encoder and the aneurysm feature encoder is consistent with that of the encoder of the cerebral vascular segmentation pre-training network, and the composition of the aneurysm feature decoder is consistent with that of the decoder of the cerebral vascular segmentation pre-training network.

13. An automatic intracranial aneurysm segmentation system, comprising: A cerebral vascular segmentation pre-training module is configured to construct a cerebral vascular segmentation pre-training network and obtain pre-training model weights; An intracranial aneurysm segmentation module is configured to take the aneurysm image to be segmented as input to an intracranial aneurysm segmentation network; pass the input image through an artery feature encoder and an aneurysm feature encoder to obtain artery coding features and aneurysm coding features; perform adaptive feature modulation on the artery coding features and the aneurysm coding features, and obtain artery enhancement fusion features and artery suppression fusion features through adaptive addition and subtraction calculations, respectively; The artery enhancement fusion feature and the artery suppression fusion feature are linearly projected and straightened into vectors by reshaping to obtain the first query vector and the second query vector; the aneurysm coding feature is linearly projected and reshaped multiple times, and two groups of feature vectors are taken, namely the first key vector and the first value vector, and the second key vector and the second value vector; the first query vector, the first key vector and the first value vector are divided into one group and attention calculation is performed, and the second query vector, the second key vector and the second value vector are divided into another group and attention calculation is performed to obtain two groups of attention feature vectors; the two groups of attention feature vectors are linearly projected and reshaped to recover the feature tensor, and input into the convolution layer, and then the feature fusion is performed by adding the input aneurysm coding feature to obtain the encoder's aneurysm enhancement feature; Inputting the aneurysm enhancement features of the encoder into the aneurysm decoder to obtain the aneurysm decoding features; The aneurysm decoding features are subjected to multi-scale feature extraction through a depthwise separable dilated convolution module. Finally, the multi-scale features of each stage of the aneurysm feature decoder are deeply supervised to predict the aneurysm segmentation results.

Images