Brain blood vessel segmentation method and storage medium based on dual-structure constrained neural network

Abstract

The present invention relates to a cerebral blood vessel segmentation method and storage medium based on a dual-structure constrained neural network, and belongs to the technical field of medical image segmentation. The method first preprocesses MRA image data and divides the preprocessed MRA image data into a training set and a test set. Secondly, a cerebral blood vessel dual-structure constrained neural network is constructed, and the training set is used to train the cerebral blood vessel dual-structure constrained neural network to obtain a trained cerebral blood vessel dual-structure constrained neural network. Finally, the test set is input into the trained cerebral blood vessel dual-structure constrained neural network to obtain a segmented cerebral blood vessel structure. The present invention performs dual-structure constraints on the segmented cerebral blood vessels at two levels: network structure and inductive bias. The segmented image has strong anti-background interference ability, good connectivity of small cerebral blood vessels, and higher segmented image accuracy to assist subsequent pathological image diagnosis. The present invention reduces the dependence of the fully automatic method on data labeling accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image segmentation, and in particular relates to a cerebral blood vessel segmentation method and a storage medium based on a dual-structure constrained neural network. Background Art

[0002] Cerebrovascular image segmentation refers to the process of segmenting cerebral blood vessels from magnetic resonance angiography (MRA) images. Segmentation methods include manual segmentation, semi-automatic segmentation, and fully automatic segmentation methods. Manual segmentation methods require the operator to have rich prior knowledge and a lot of time. Semi-automatic segmentation methods mainly include active contour models and probabilistic methods. They usually require manual input and then use the manually inputted prior knowledge to achieve image segmentation through computers. Fully automatic segmentation methods overcome the shortcomings of manual and semi-automatic methods that require manual participation, further improving the efficiency of cerebrovascular image segmentation.

[0003] Obtaining complete cerebral vascular labeling is time-consuming and labor-intensive. If sparse labeling is used, the currently used fully automatic segmentation method will find it difficult to accurately segment dense and well-connected cerebral vessels. At the same time, the skull has a strong interference effect on cerebral vascular segmentation. Without skull removal, the existing fully automatic segmentation method generally contains some background noise in its segmentation results. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention proposes a cerebral blood vessel segmentation method based on a dual-structure constrained neural network, comprising the following steps:

[0005] Step 1: preprocess the MRA image data and divide the preprocessed MRA image data into a training set and a test set;

[0006] Step 2: Constructing a cerebral vascular dual-structure constraint neural network;

[0007] Step 3: Using the training set to train the cerebrovascular dual-structure constrained neural network to obtain a trained cerebrovascular dual-structure constrained neural network;

[0008] Step 4: Input the test set into the trained cerebrovascular dual-structure constrained neural network to obtain the segmented cerebrovascular structure.

[0009] The step 1 comprises the following steps:

[0010] Step 1.1: Determine whether the MRA image data has manual sparse markings. If so, proceed to step 1.3; otherwise, proceed to step 1.2.

[0011] Step 1.2: Set a resolution threshold, and perform sparse labeling on MRA images with resolutions lower than the threshold using a vascular information file containing cerebral vascular centerline and radius information;

[0012] Step 1.3: Set data precision, convert both the MRA image data and the sparse labeled data into a parallel training format, and normalize the data according to the set data precision to obtain the processed trainable data.

[0013] Step 1.4: Divide the trainable data into training and test sets in proportion according to actual needs;

[0014] Step 1.5: Perform data augmentation on the training set.

[0015] The data augmentation for the training set described in step 1.5 uses the random cropping method.

[0016] The cerebrovascular dual-structure constraint neural network includes an encoding module, a decoding module, a cerebrovascular structure attention module that constrains the network level structure, and a hybrid loss function that constrains the inductive bias level structure; the cerebrovascular structure attention module connects the lowest layer of the encoding module and the lowest layer of the decoding module;

[0017] Encoding module: used to encode the training set, obtain the high-level feature map of cerebral vascular, and send it to the cerebral vascular structure attention module;

[0018] Cerebrovascular structure attention module: used to input the vascular structure attention into the received cerebrovascular high-level feature map, obtain the cerebrovascular structure attention high-level feature map, and send it to the decoding module;

[0019] Decoding module: used to decode the high-level feature map of cerebrovascular structure attention to obtain the cerebrovascular segmentation result map;

[0020] Hybrid loss function: used to obtain the training error of the cerebrovascular dual-structure constrained neural network.

[0021] The encoding module is composed of N connected encoding residual convolution blocks; each encoding residual convolution block is composed of multiple 3D convolutions connected by batch normalization-ReLU activation function, and the input and output are connected using jump connections.

[0022] The decoding module is composed of N connected decoding convolution blocks; the decoding convolution blocks are connected through batch normalization.

[0023] The cerebrovascular structure attention module includes:

[0024] Structural self-attention module: used to receive the cerebrovascular high-level feature map sent by the encoding module, obtain the spatial similarity of the cerebrovascular structure, further multiply it with the original feature to obtain the spatial cerebrovascular attention feature map, and send it to the global attention feature compression module;

[0025] Global attention feature compression module: used to receive the cerebrovascular attention feature map at the spatial level, reduce the number of parameters in the map, and then obtain the cerebrovascular spatial attention feature map and send it to the bottleneck structure channel attention module;

[0026] Bottleneck structure channel attention module: used to receive the cerebrovascular spatial attention feature map, allocate attention to the cerebrovascular spatial attention feature map at the feature channel level, obtain the channel-level attention of the cerebrovascular structure, and further add the channel-level attention with the high-level feature map before inputting the cerebrovascular attention through the channel level to obtain the cerebrovascular structure attention high-level feature map, and send it to the decoding module.

[0027] The step 3 comprises the following steps:

[0028] Step 3.1, parameter initialization, including: setting the number of training times c, training batch size m, bottleneck rate r and network parameter learning rate lr of the dual structure constraint neural network, and randomly initializing the network weights;

[0029] Step 3.2: Set the update rule of the network parameter learning rate and obtain the current network parameter learning rate based on the network parameter learning rate of the previous round of training;

[0030] Step 3.3: Input the training set into the dual-structure constrained neural network and perform forward propagation to obtain the cerebral vascular segmentation result map;

[0031] Step 3.4, obtaining the training error of the dual-structure constrained neural network based on the sparsely labeled MRA image data and the cerebral vascular segmentation result map;

[0032] Step 3.5: Back propagate and update the optimized network parameters based on the obtained training error;

[0033] Step 3.6: Determine whether the number of training times reaches the set number of training times c. If so, the training is completed and the dual-structure constrained neural network is saved. Otherwise, return to step 3.2.

[0034] The training error of the dual-structure constrained neural network obtained in step 3.4 adopts the eclDice-Dice hybrid loss function, and the formula is as follows:

[0035]

[0036] L=L Dice +L eclDice(3)

[0037] Among them, L Dice represents the Dice similarity coefficient loss, N represents the sum of all voxels, i represents the i-th voxel in the MRA image, t i represents the i-th voxel value of the sparsely labeled MRA image, y i represents the i-th voxel value of the predicted cerebral vascular segmentation result image, ε represents the smoothing coefficient, which is used to smooth the loss value and gradient, L eclDice Indicates the Dice similarity coefficient loss of the balanced center line, p i represents the structural skeleton obtained by extracting the center line from the predicted cerebral vascular segmentation result image, m i represents the structural skeleton obtained from the centerline extraction of the i-th sparsely labeled MRA image, L represents the eclDice-Dice hybrid loss, and β represents the balance coefficient.

[0038] A storage medium storing a plurality of instructions suitable for loading by a processor to execute at least one step of the cerebral blood vessel segmentation method based on a dual-structure constraint neural network according to any one of claims 1 to 9.

[0039] Advantages of the present invention:

[0040] The present invention is based on a cerebral blood vessel segmentation method and storage medium using a dual-structure constraint neural network. Dual-structure constraints are applied to the segmented cerebral blood vessels at two levels: network structure and inductive bias. The segmented images have strong resistance to background interference, good connectivity of small cerebral blood vessels, and higher segmented image accuracy to assist in subsequent pathological image diagnosis. The present invention reduces the reliance of fully automatic methods on data labeling accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of a cerebral blood vessel segmentation method based on a dual-structure constrained neural network according to an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the structure of a cerebral vascular dual-structure constraint neural network according to an embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of the structure of a cerebral vascular structure attention module according to an embodiment of the present invention;

[0044] Figure 4 This is a method flow chart of step 3 of an embodiment of the present invention;

[0045] Figure 5 3D renderings of segmentation results before and after dual structure constraints according to an embodiment of the present invention;

[0046] Figure 6 This is a graph showing changes in performance indicators during the training of a cerebral vascular dual-structure constrained neural network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] An embodiment of the present invention will be further described below with reference to the accompanying drawings.

[0048] In the embodiment of the present invention, a cerebral blood vessel segmentation method based on a dual structure constrained neural network is shown in the flowchart of the method. Figure 1 As shown, the following steps are included:

[0049] Step 1: Preprocess the original magnetic resonance angiography (MRA) image data and divide the preprocessed MRA image data into a training set and a test set, as follows:

[0050] Step 1.1: Determine whether the MRA image data has manual sparse markings. If so, proceed to step 1.3; otherwise, proceed to step 1.2.

[0051] Step 1.2: Set the resolution threshold to 448 × 448 × 128, and perform sparse labeling on MRA images with resolutions below the threshold using a vascular information file containing cerebral vascular centerline and radius information;

[0052] Step 1.3: Set data precision, convert both the MRA image data and the sparsely labeled data into a parallel training data format for input into the neural network, and normalize the data according to the set data precision to obtain processed trainable data;

[0053] Step 1.4: Divide the trainable data into a training set and a test set in a ratio of 4:1;

[0054] Step 1.5: Use data augmentation methods to expand the data of the training set: Data augmentation uses a random cropping method to perform cropping. Randomly select positions in three directions of the three-dimensional voxel image, crop it according to the set volume, and obtain voxel data to input into the neural network for training.

[0055] Step 2: Construct a cerebral vascular dual-structure constraint neural network, as follows:

[0056] like Figure 2 As shown, the cerebrovascular dual-structure constraint neural network includes an encoding module, a decoding module, a cerebrovascular structure attention module that constrains the network level structure, and a hybrid loss function that constrains the inductive bias level structure; the cerebrovascular structure attention module connects the bottom layer of the encoding module and the bottom layer of the decoding module;

[0057] Encoding module: used to encode the training set, obtain the high-level feature map of cerebral vascular, and send it to the cerebral vascular structure attention module;

[0058] In an embodiment of the present invention, the encoding module is composed of N connected encoding residual convolution blocks; each encoding residual convolution block is composed of two 3D convolutions connected by a batch normalization-ReLU activation function, and the input and output are connected using a jump connection.

[0059] Cerebrovascular structure attention module: used to input the vascular structure attention into the received cerebrovascular high-level feature map, obtain the cerebrovascular structure attention high-level feature map, and send it to the decoding module;

[0060] like Figure 3 As shown, the cerebrovascular structure attention module includes:

[0061] Structural self-attention module: used to receive the cerebrovascular high-level feature map sent by the encoding module, obtain the spatial similarity of the cerebrovascular structure, further multiply it with the original feature to obtain the spatial cerebrovascular attention feature map, and send it to the global attention feature compression module;

[0062] Global attention feature compression module: used to receive the cerebrovascular attention feature map at the spatial level, reduce the number of parameters in the map, and then obtain the cerebrovascular spatial attention feature map and send it to the bottleneck structure channel attention module;

[0063] Bottleneck structure channel attention module: used to receive the cerebrovascular spatial attention feature map, allocate attention to the cerebrovascular spatial attention feature map at the feature channel level, obtain the channel-level attention of the cerebrovascular structure, and further add the channel-level attention with the high-level feature map before inputting the cerebrovascular attention through the channel level to obtain the cerebrovascular structure attention high-level feature map, and send it to the decoding module.

[0064] In the embodiment of the present invention, Figure 3 As shown in the figure, the structural self-attention module first obtains the high-level feature map F generated by the last encoding residual convolution block of the encoding module, uses a 3×1×1-1×1×3 3D convolution block to extract the sagittal high-level feature map S in the high-level feature map F, and uses a 1×3×1-1×1×3 3D convolution block to extract the coronal high-level feature map C in the high-level feature map F; the high-level feature map F, the sagittal high-level feature map S and the coronal high-level feature map C are deformed respectively to change the dimension of the feature map, and the height (H) width (W) and depth (D) of the feature map are deformed into one-dimensional space; the deformed feature map S and feature map C use dot product and softmax function to calculate the similarity between the sagittal high-level features and the coronal high-level features, and the obtained spatial attention features are applied to each element in the high-level feature map using matrix multiplication to obtain the high-level feature map F with spatial attention. s , this high-level feature map F sThe input is compressed into a global attention feature compression module consisting of a 1×1×1 3D convolution block and a softmax function, and the compressed cerebrovascular spatial attention feature map is input into the bottleneck structure channel attention module;

[0065] The bottleneck structure channel attention module consists of two layers of 1×1×1 3D convolution and the ReLU function. The ratio of the number of convolution channels in the first layer to the number of convolution channels in the second layer is set to r = 0.5, which is the bottleneck rate. The cerebrovascular spatial attention feature map is added to the high-level feature map F at the channel level to obtain the cerebrovascular structure attention high-level feature map F. SC ; Attention to high-level feature map of cerebrovascular structure F SC Input to the lowest layer of the decoding module;

[0066] Decoding module: used to decode the high-level feature map of cerebrovascular structure attention to obtain the cerebrovascular segmentation result map;

[0067] The decoding module is composed of N decoding convolution blocks connected together, and the decoding convolution blocks are connected through batch normalization;

[0068] In an embodiment of the present invention, for the corresponding Nth coding residual convolution block and decoding convolution block, a feature map connection operation is performed on the high-level feature map generated by the Nth coding residual convolution block and the high-level feature map generated by the N-1th decoding convolution block, and the spliced ​​high-level feature map is input into the Nth decoding convolution block.

[0069] Hybrid loss function: used to obtain the training error of the cerebrovascular dual-structure constrained neural network.

[0070] Step 3: Use the training set to train the cerebrovascular dual-structure constraint neural network to obtain the trained cerebrovascular dual-structure constraint neural network. The flowchart is as follows: Figure 4 As shown, the details are as follows:

[0071] Step 3.1, parameter initialization, including: setting the number of training times c, training batch size m, bottleneck rate r and network parameter learning rate lr of the dual structure constraint neural network, and randomly initializing the network weights;

[0072] Step 3.2: Set the update rule of the network parameter learning rate and obtain the current network parameter learning rate based on the network parameter learning rate of the previous round of training;

[0073] In the embodiment of the present invention, during the training of the cerebrovascular dual-structure constrained neural network, if the network parameter learning rate is too large, the network will be prone to oscillation during the training process and difficult to converge. If the network parameter learning rate is too small, the network convergence will be very slow, which will increase the time to find the optimal value or fall into the local optimum and be unable to find the optimal parameters. Therefore, the network parameter learning rate update formula used in the present invention is as follows:

[0074]

[0075] Among them, lr new Indicates the updated network parameter learning rate, lr initial represents the initial network parameter learning rate, epoch represents the current number of training rounds, max_epoch represents the number of training times c of the dual-structure constrained neural network, and the decay rate p is set to 0.9;

[0076] Step 3.3: Input the training set into the dual-structure constrained neural network and perform forward propagation to obtain the cerebral vascular segmentation result map;

[0077] Step 3.4, obtaining the training error of the dual-structure constrained neural network based on the sparsely labeled MRA image data and the cerebral vascular segmentation result map;

[0078] In the embodiment of the present invention, the training error of the dual-structure constrained neural network is obtained by using the eclDice-Dice hybrid loss function (eclDice is Equilibrium center line Dice, indicating the equilibrium center line), and the formula is as follows:

[0079]

[0080] L=L Dice +L eclDice (3)

[0081] Among them, L Dice represents the Dice similarity coefficient loss, N represents the sum of all voxels, i represents the i-th voxel in the MRA image, t i represents the i-th voxel value of the sparsely labeled MRA image, y i represents the i-th voxel value of the predicted cerebral vascular segmentation result image, ε represents the smoothing coefficient, which is used to smooth the loss value and gradient, L eclDice Indicates the Dice similarity coefficient loss of the balanced center line, p i represents the structural skeleton obtained by extracting the center line from the predicted cerebral vascular segmentation result image, m i represents the structural skeleton extracted from the centerline of the i-th sparsely labeled MRA image, L represents the eclDice-Dice hybrid loss, and β represents the balance coefficient, which is set to 2.

[0082] Step 3.5: Take the partial derivative of the obtained training error and use the chain rule to back-propagate and optimize the updated network parameters. The formula is as follows:

[0083]

[0084] Among them, θt represents the updated convolution weight parameter in the cerebrovascular dual structure constraint neural network, θ t-1 represents the convolution weight parameter in the cerebrovascular dual structure constraint neural network before updating, m t represents the correction of the first-order moment estimate of the gradient; v t represents the correction to the second moment estimate of the gradient;

[0085] Step 3.6: Determine whether the number of training times reaches the set number of training times c. If so, the training is completed and the dual-structure constrained neural network is saved. Otherwise, return to step 3.2.

[0086] Step 4: Input the test set into the trained cerebrovascular dual-structure constrained neural network to obtain the segmented cerebrovascular structure.

[0087] In the embodiment of the present invention, Figure 5 As shown in the figure, the two columns of images are the qualitative analysis and comparison of the segmentation results before and after the dual structure constraint in the horizontal, sagittal and coronal directions;

[0088] Before the dual-constraint structure was introduced, the segmentation results contained a large amount of noise from the skull, and many artifacts were over-segmented within the skull. Furthermore, the connectivity of the segmentation results was difficult to ensure in small blood vessels. After the dual-constraint structure was introduced, the segmentation results did not include skull vessels, and the over-segmentation of artifacts was greatly improved. This demonstrates that the proposed method has strong anti-interference capabilities against strong background noise.

[0089] Compared with the situation before the dual constraint structure was introduced, the vascular connectivity of the segmentation results after the constraint was greatly improved. This improvement was more obvious in small blood vessels. This shows that when sparse markers with poor connectivity are used as training sets, the connectivity of the segmentation results of the proposed method is stronger than that before the constraint, and has been greatly improved.

[0090] In the embodiment of the present invention, Figure 6 It reflects the changing trend of the accuracy and performance indicators during the training of the cerebrovascular dual-structure constraint neural network; as the number of iterations increases, the segmentation accuracy of the neural network continues to improve; in the embodiment of the present invention, the Dice Similarity Coefficient (DSC), Intersection of Union (IOU) and center line Dice (clDice) are respectively used to calculate the above indicators by comparing the segmentation results with the gold standard, and the segmentation results are quantitatively analyzed. By comparing the experimental results, the segmentation results in Table 1 are better than those in Table 2. Figure 6The performance indicators of the training process are better, which shows that the constraints at the network structure level and the inductive bias level enable the network to learn more vascular structures from the sparse labels, resulting in better segmentation results than when trained under sparse labels. This reduces the dependence of fully automatic methods on data labeling accuracy to a certain extent.

[0091] Table 1

[0092]

[0093] In an embodiment of the present invention, based on the same inventive concept, the present application also provides a storage medium, which stores multiple instructions, and the instructions are suitable for loading by a processor to execute steps 1-4 in the cerebral blood vessel segmentation method based on a dual-structure constraint neural network.

[0094] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. A cerebral blood vessel segmentation method based on a dual-structure constrained neural network, characterized by: The following steps are involved: Step 1: preprocess the MRA image data and divide the preprocessed MRA image data into a training set and a test set; Step 1 includes the following steps: Step 1.1: Determine whether the MRA image data has manual sparse markings. If so, proceed to step 1.3; otherwise, proceed to step 1.

2. Step 1.2: Set a resolution threshold, and perform sparse labeling on MRA images with resolutions lower than the threshold using a vascular information file containing cerebral vascular centerline and radius information; Step 1.3: Set data precision, convert both the MRA image data and the sparse labeled data into a parallel training format, and normalize the data according to the set data precision to obtain the processed trainable data. Step 1.4: Divide the trainable data into training and test sets in proportion according to actual needs; Step 1.5: Use random cropping method to enhance the training set data; Step 2: Constructing a cerebrovascular dual-structure constrained neural network; the cerebrovascular dual-structure constrained neural network includes an encoding module, a decoding module, a cerebrovascular structure attention module that constrains the network level structure, and a hybrid loss function that constrains the inductive bias level structure; the cerebrovascular structure attention module connects the lowest layer of the encoding module and the lowest layer of the decoding module; Encoding module: used to encode the training set, obtain the high-level feature map of cerebral vascular, and send it to the cerebral vascular structure attention module; Cerebrovascular structure attention module: used to input the vascular structure attention into the received cerebrovascular high-level feature map, obtain the cerebrovascular structure attention high-level feature map, and send it to the decoding module; Decoding module: used to decode the high-level feature map of cerebrovascular structure attention to obtain the cerebrovascular segmentation result map; Hybrid loss function: used to obtain the training error of the cerebrovascular dual-structure constrained neural network; Step 3: Using the training set to train the cerebrovascular dual-structure constrained neural network to obtain a trained cerebrovascular dual-structure constrained neural network; Step 4: Input the test set into the trained cerebrovascular dual-structure constrained neural network to obtain the segmented cerebrovascular structure.

2. The cerebral blood vessel segmentation method based on a dual-structure constrained neural network according to claim 1, characterized in that: The encoding module is composed of N connected encoding residual convolution blocks; each encoding residual convolution block is composed of multiple 3D convolutions connected by batch normalization-ReLU activation function, and the input and output are connected using jump connections.

3. The cerebral blood vessel segmentation method based on a dual-structure constrained neural network according to claim 1, characterized in that: The decoding module is composed of N connected decoding convolution blocks; the decoding convolution blocks are connected through batch normalization.

4. The cerebral blood vessel segmentation method based on a dual-structure constrained neural network according to claim 1, characterized in that: The cerebrovascular structure attention module includes: Structural self-attention module: used to receive the cerebrovascular high-level feature map sent by the encoding module, obtain the spatial similarity of the cerebrovascular structure, further multiply it with the original feature to obtain the spatial cerebrovascular attention feature map, and send it to the global attention feature compression module; Global attention feature compression module: used to receive the cerebrovascular attention feature map at the spatial level, reduce the number of parameters in the map, and then obtain the cerebrovascular spatial attention feature map and send it to the bottleneck structure channel attention module; Bottleneck structure channel attention module: used to receive the cerebrovascular spatial attention feature map, allocate attention to the cerebrovascular spatial attention feature map at the feature channel level, obtain the channel-level attention of the cerebrovascular structure, and further add the channel-level attention with the high-level feature map before inputting the cerebrovascular attention through the channel level to obtain the cerebrovascular structure attention high-level feature map, and send it to the decoding module.

5. The cerebral blood vessel segmentation method based on a dual-structure constrained neural network according to claim 1, characterized in that: The step 3 comprises the following steps: Step 3.1, parameter initialization, including: setting the number of training times c, training batch size m, bottleneck rate r and network parameter learning rate lr of the dual structure constraint neural network, and randomly initializing the network weights; Step 3.2: Set the update rule of the network parameter learning rate and obtain the current network parameter learning rate based on the network parameter learning rate of the previous round of training; Step 3.3: Input the training set into the dual-structure constrained neural network and perform forward propagation to obtain the cerebral vascular segmentation result map; Step 3.4, obtaining the training error of the dual-structure constrained neural network based on the sparsely labeled MRA image data and the cerebral vascular segmentation result map; Step 3.5: Back propagate and update the optimized network parameters based on the obtained training error; Step 3.6: Determine whether the number of training times reaches the set number of training times c. If so, the training is completed and the dual-structure constrained neural network is saved. Otherwise, return to step 3.

2.

6. The cerebral blood vessel segmentation method based on a dual-structure constrained neural network according to claim 5, characterized in that: The training error of the dual-structure constrained neural network obtained in step 3.4 adopts the eclDice-Dice hybrid loss function, and the formula is as follows: L=L Dice +L eclDice (3) Among them, L Dice represents the Dice similarity coefficient loss, N represents the sum of all voxels, i represents the i-th voxel in the MRA image, t i Represents the i-th voxel value of the sparsely labeled MRA image, y i represents the i-th voxel value of the predicted cerebral vascular segmentation result image, ε represents the smoothing coefficient, which is used to smooth the loss value and gradient, L eclDice Indicates the Dice similarity coefficient loss of the balanced center line, p i represents the structural skeleton obtained by extracting the center line from the predicted cerebral vascular segmentation result image, m i represents the structural skeleton obtained from the centerline extraction of the i-th sparsely labeled MRA image, L represents the eclDice-Dice hybrid loss, and β represents the balance coefficient.

7. A storage medium, characterized in that: The storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute at least one step of the cerebral blood vessel segmentation method based on a dual-structure constraint neural network according to any one of claims 1 to 6.

Images