Three-dimensional microscopic whole cerebrovascular image segmentation method and system

Abstract

The invention discloses a three-dimensional microscopic whole cerebrovascular image segmentation method and system, and the method comprises the steps: stacking microscopic cerebrovascular images which are shot layer by layer in a three-dimensional space according to a shooting sequence, segmenting the microscopic cerebrovascular images into a plurality of three-dimensional blocks which are suitable for training, and carrying out the necessary preprocessing; pre-training a teacher model through a reconstruction task on the three-dimensional microscopic cerebrovascular data which is not labeled, and meanwhile, obtaining a codebook containing a rich vascular structure distribution coding mode, and supervising and finely adjusting teacher model parameters through a label on the three-dimensional microscopic cerebrovascular data which is slightly labeled; training a student model by utilizing teacher model guidance and limited annotation data, and guiding the student model to perform high-quality extraction on the data by means of a vascular structure codebook; and the trained lightweight student model is used to complete segmentation of the whole brain microscopic image, and then the three-dimensional blocks are cut and reconstructed into a complete three-dimensional whole brain blood vessel network.

Description

Technical Field

[0001] The present invention belongs to the fields of computer vision and deep learning, and particularly relates to a three-dimensional microscopic whole cerebral vascular image segmentation method and system. Background Art

[0002] Cerebral vascular segmentation is of great significance in neuroscience research. Its core role is to visualize and extract fine information of complex cerebral vascular structures. This process is crucial for studying brain function, analyzing structural changes of cerebral blood vessels, and evaluating vascular abnormalities. Structural abnormalities of cerebral blood vessels, such as cerebral vascular malformations, small vessel diseases, and cerebral aneurysms, are usually closely related to various diseases and degenerative lesions. These abnormalities are manifested by changing the morphology or function of blood vessels, which also become key visual indicators for diagnosis and treatment.

[0003] High-resolution microscopic cerebral vascular images provide rich anatomical details, but at the same time bring many challenges in storage, annotation, and processing. Each microscopic section contains millions of pixels, and the three-dimensional reconstruction of the entire cerebral vascular network may generate data volumes of hundreds of gigabytes or even terabytes. This data scale leads to high requirements for storage and computing resources in deep learning models. In addition, due to the complexity of cerebral vascular morphology and the diversity of branches, the data annotation work is extremely time-consuming, significantly restricting the development of efficient automated segmentation technologies.

[0004] Currently, although some studies, such as CN110853051A, attempt to achieve cerebral vascular segmentation through imaging means such as MRA, due to the particularity of microscopic imaging, three-dimensional segmentation methods applicable to high-resolution microscopic cerebral vascular data are still very limited. These methods cannot fully exploit the characteristics of rich unlabeled data in microscopic images and are difficult to meet the precise segmentation requirements for studying small-diameter blood vessels (such as arterioles, capillaries, and venules).

[0005] Although labeled data of microscopic cerebral blood vessels is difficult to obtain, we can still utilize rich unlabeled microscopic cerebral blood vessel data to obtain additional information. For example, unsupervised and semi-supervised strategies can be used to prompt the model to focus on the distribution characteristics of microscopic cerebral blood vessels. In addition, designing corresponding three-dimensional microscopic cerebral vascular segmentation methods using knowledge distillation strategies will help enhance the segmentation effect of the model and thus improve the efficiency of blood vessel segmentation. Summary of the Invention

[0006] To solve the deficiencies of the prior art and achieve the purpose of improving blood vessel image segmentation, the present invention adopts the following technical solutions:

[0007] A three-dimensional microscopic blood vessel image segmentation method includes the following steps:

[0008] Step S1: Stack the microscopically taken blood vessel images layer by layer in three-dimensional space in sequence, and then cut them into a group of three-dimensional blocks;

[0009] Step S2: pre-training the teacher model through reconstruction tasks on unlabeled 3D microvascular data, and learning a codebook containing rich coding patterns of vascular structure distribution;

[0010] Step S3, using a small amount of labeled 3D microvascular data to supervise the teacher model and fine-tune the model parameters with a small learning rate;

[0011] Step S4, using the teacher model guidance and limited annotated data to train the student model, guiding the student model to extract high-quality data with the help of the codebook, and finally performing subsequent vascular image segmentation based on the trained student model;

[0012] Step S5: Use the trained lightweight student model to complete the segmentation of the whole brain microscopic image, and then reconstruct the obtained three-dimensional blocks into a complete three-dimensional whole brain vascular network according to the cutting method in step S1.

[0013] Furthermore, the step S2 comprises the following steps:

[0014] Step S21, feeding the unlabeled three-dimensional microvascular data obtained by step S1 into a teacher model SwinUNetR with more parameters and fine structure design, requiring the teacher model to reconstruct the original three-dimensional block after encoding, codebook quantization, and decoding, and constructing the reconstruction loss;

[0015] Step S22, constructing a vascular structure pattern codebook V at the encoder output position of the teacher model, wherein the dimension of each quantized vector is consistent with the vector dimension output by the encoder of the student model, and the codebook is continuously updated during the pre-training process of the teacher model, and a codebook loss is constructed;

[0016] Step S23: Based on the reconstruction loss and the codebook loss, construct an overall loss for updating the parameters of the teacher model.

[0017] Furthermore, in step S22, in order to provide guidance to the student model in the future, first, the encoder output vector z of the teacher model is made i The dimension is consistent with the vector dimension output by the encoder of the student model, and then the code vector with the closest Euclidean distance to each vector in the codebook is used to replace it to form a discretized code, and then the discretized code is upgraded to be consistent with that before dimensionality reduction and fed to the decoder of the teacher model.

[0018] Furthermore, the discrete sampling process is replaced by the Gumbel-Softmax technique, so that the sampling and encoding replacement process here is as follows:

[0019]

[0020] Among them, V represents the vascular structure pattern codebook, z i represents the encoder output vector of the teacher model, DA(*) represents dimension alignment, where a 1x1 convolutional kernel is used for dimension alignment, and ‖*‖2 represents the 2-norm. For each vector z i , find the quantization vector v with the smallest Euclidean distance from it in the codebook V j , where j is the index of the quantization vector corresponding to z i in this search in the codebook, and i is the vector serial number of the encoder in the teacher model. Then, use Gumbel-Softmax for approximate discretization to obtain the quantized vector Finally, replace the original encoder output vector z of the teacher model i , and feed it into the decoder of the teacher model

[0021] The codebook loss function is expressed as:

[0022]

[0023] Among them, represents the square of the 2-norm, sg[*] represents the stop gradient operation, and λ represents the balance coefficient.

[0024] Furthermore, in the step S21, calculate the average value of the absolute value of the brightness difference between the original three-dimensional block X and the reconstructed three-dimensional block pixel by pixel as the reconstruction loss function:

[0025]

[0026] Among them, N represents the number of three-dimensional blocks.

[0027] Furthermore, the step S4 includes the following steps:

[0028] Step S41: Freeze all the parameters of the teacher model and do not update them during the subsequent training process;

[0029] Step S42: Combine the unlabeled three-dimensional microscopic blood vessel blocks and the labeled three-dimensional microscopic cerebrovascular blocks in a certain proportion to form a batch, and feed them into the teacher model and the student model respectively;

[0030] Step S43: For the labeled vascular data, use the vascular label GT to supervise the vascular prediction result Y output by the student model, and construct the second segmentation loss of the labeled data; for the unlabeled data, use the vascular segmentation result Y of the teacher model for the same data tea as the pseudo label to guide the vascular prediction result of the student model, and construct the unlabeled semi-supervised loss;

[0031] Step S44: For the high-dimensional vector representation Z of the original 3D block output by the student model encoder stu , use the vector quantized by the teacher model through the codebook for constraint, and construct a feature distillation loss to prompt the student model to encode more high-quality features;

[0032] Step S45: Based on the second segmentation loss, semi-supervised loss, and feature distillation loss, construct the loss of the student model for training the student model.

[0033] Furthermore, in the said Step S3, the first segmentation loss function is used as follows:

[0034]

[0035] where Y tea represents the predicted result of blood vessel segmentation output by the teacher model, and GT represents the blood vessel label;

[0036] In the said Step S43, the second segmentation loss function used is as follows:

[0037]

[0038] where Y represents the predicted result of blood vessels output by the student model;

[0039] The semi-supervised loss function for unlabeled data is as follows:

[0040]

[0041] In the said Step S44, the L1 norm of the quantized feature encoding of the teacher model and the encoding output by the student model is used as the feature distillation loss function:

[0042]

[0043] In the said Step S45, the weighted sum of the second segmentation loss, semi-supervised loss, and feature distillation loss is used as the loss function of the student model, and the calculation method is:

[0044] Student model loss function = segmentation loss function + β * semi-supervised loss function + feature distillation loss function

[0045] where β is set as a Gaussian warm-up coefficient that increases with the number of training rounds, and the value of β in the t-th round is calculated as:

[0046]

[0047] Furthermore, the said Step S1 specifically includes the following steps:

[0048] Step S11: Stack the microvascular images taken layer by layer in three-dimensional space in the shooting order to form three-dimensional microscopic data, and read S images each time;

[0049] Step S12: Axially cut the large three-dimensional image into several three-dimensional cubes with side length S;

[0050] Step S13: Perform data preprocessing on the three-dimensional image. Conduct histogram statistics on the brightness of the complete data set to obtain the threshold quantile of the overall data set brightness, and then perform histogram truncation on the data set, truncating the excessive values to the threshold quantile. The threshold quantile is the minimum value that keeps 99% of the pixel brightness values unchanged.

[0051] A method for segmenting three-dimensional microscopic whole cerebral vascular images, which segments the three-dimensional microscopic whole cerebral vascular images through the above-mentioned method for segmenting three-dimensional microscopic vascular images.

[0052] A three-dimensional microscopic whole cerebral vascular image segmentation system includes a three-dimensional block generation module, a teacher module, and a student module. According to the above-mentioned method for segmenting three-dimensional microscopic whole cerebral vascular images, the three-dimensional microscopic whole cerebral vascular images are sequentially subjected to the generation of three-dimensional blocks, the generation and training of the teacher model codebook, and the training of the student model with the help of the codebook for subsequent vascular image segmentation.

[0053] The advantages and beneficial effects of the present invention are as follows:

[0054] Based on the high-resolution characteristics of microscopic cerebral vascular data, the present invention designs a codebook-based knowledge distillation paradigm. Under the condition of making full use of unlabeled data, it guides the student neural network model to learn diverse whole cerebral vascular morphological structure patterns during training. By focusing on the diverse whole cerebral vascular structure distribution characteristics, it enhances the segmentation effect of vascular images, thereby improving the efficiency of three-dimensional microscopic cerebral vascular segmentation. The cerebral vascular segmentation method designed by the present invention taking into account both segmentation speed and accuracy adapts to the additional expenses of computational resources and time consumption for the microscopic cerebral vascular segmentation task, ensuring a high-efficiency three-dimensional microscopic whole cerebral vascular segmentation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flowchart of the method of the embodiment of the present invention.

[0056] Figure 2 It is a flowchart of the data preprocessing stage in the embodiment of the present invention.

[0057] Figure 3a It is a schematic diagram of the teacher model pre-training stage in the embodiment of the present invention.

[0058] Figure 3b It is a schematic diagram of the teacher model fine-tuning stage in the embodiment of the present invention.

[0059] Figure 4a This is a schematic diagram of the student model training stage in the embodiments of the present invention.

[0060] Figure 4b This is a schematic diagram of the student model testing stage in the embodiments of the present invention.

[0061] Figure 5 This is a comparison diagram of the segmentation effects between the present invention and other methods in the embodiments of the present invention. Detailed implementation manners

[0062] The following will describe the detailed implementation manners of the present invention with reference to the accompanying drawings. It should be understood that the detailed implementation manners described herein are only for explaining and illustrating the present invention, and are not used to limit the present invention.

[0063] As Figure 1 shown, a three-dimensional microscopic whole cerebral vascular image segmentation method includes the following steps:

[0064] Step S1: Stack the microvascular cerebral images taken layer by layer in a three-dimensional space according to the shooting order, then cut them into several three-dimensional blocks suitable for training, and perform necessary preprocessing. As Figure 2 shown, the specific implementation method is:

[0065] Step S11: Stack the microvascular cerebral images taken layer by layer in a three-dimensional space according to the shooting order to form three-dimensional microscopic whole brain data. To reduce memory occupancy, read S images each time, where S corresponds to the side length of the subsequent three-dimensional blocks for training, and is generally set to 64;

[0066] Step S12: Cut the large three-dimensional brain image along the axial direction into several three-dimensional squares with a side length of S and store them in the hard disk;

[0067] Step S13: Perform histogram statistics on the brightness of the complete data set to obtain the 99th percentile of the overall data set brightness, and then perform histogram truncation on the data set, truncating the excessive values to this 99th percentile, which is actually the minimum value that keeps the brightness values of 99% of the pixel points unchanged;

[0068] Step S14: Further perform maximum-minimum normalization on the input data of the data set.

[0069] Step S2: Pre-train a teacher model through a reconstruction task on the unlabeled three-dimensional microscopic cerebral vascular data of rats, and simultaneously learn a codebook containing rich vascular structure distribution coding patterns. As Figure 3a 、 Figure 3b shown, the specific implementation method is:

[0070] Step S21: Feed the unlabeled three-dimensional microscopic cerebrovascular data obtained in Step S1 into a teacher model SwinUNetR with a large number of parameters and a fine-structured design, and require the teacher model to reconstruct the original three-dimensional block after encoding, codebook quantization, and decoding. Let the original three-dimensional block be X, and the reconstructed three-dimensional block be Calculate the average value of the absolute value of the brightness difference between the original three-dimensional block and the reconstructed three-dimensional block pixel by pixel as the reconstruction loss function, which is specifically expressed as:

[0071]

[0072] Step S22: Construct a vascular structure pattern codebook V at the output position of the encoder of the teacher model. The size of this codebook can be set to 1024 entries, and the dimension of each quantization vector is the same as the dimension of the vector output by the encoder of the student model, which is set to 48 here. During the pre-training process of the teacher model, continuously update this codebook. For subsequent guidance of the student model, first reduce the dimension of the output vector z of the encoder of the teacher model from 512 dimensions to the same as the dimension of the vector output by the encoder of the student model, which is 48, and then replace it with the code vector in the codebook that is closest to each vector in terms of Euclidean distance to form a discretized encoding. Then, dimension up the discretized encoding to the same as before the dimension reduction and feed it into the decoder of the teacher model. Replace the discrete sampling process with the Gumbel-Softmax technique, so the sampling and encoding replacement process here can be expressed as: i where DA(*) represents dimension alignment, and here a 1x1 convolutional kernel is used for dimension alignment.

[0073]

[0074] This part of the loss function can be expressed as:

[0075] where sg[*] represents the stop gradient operation, and λ is a balance coefficient, which is empirically set to 0.25.

[0076]

[0077]

[0078] Step S23: Thus, the overall loss function for updating the parameters of the teacher model is the sum of the reconstruction loss and the codebook loss.

[0079] Step S3: Supervise the teacher model through labels on a small amount of labeled three-dimensional microscopic cerebrovascular data of rats, and fine-tune its parameters with a small learning rate. Use the Dice loss function, and the calculation is expressed as:

[0080]

[0081] ​Step S4: Use the teacher model to guide and train the student model with limited labeled data, and use the vascular structure codebook to guide the student model to extract data of high quality. As Figure 4a , Figure 4b shown, the specific implementation method is as follows:

[0082] Step S41: Freeze all the parameters of the teacher model and do not update them during subsequent training;

[0083] Step S42: Compose unlabeled three-dimensional microscopic cerebrovascular blocks and labeled three-dimensional microscopic cerebrovascular blocks into a batch according to a certain ratio, and feed them into the teacher model and the student model respectively. Here, the ratio is selected according to the amount of labeled data, generally set to 1:1 (when the labeled data is relatively sufficient) or 1:3 (when the labeled data is less);

[0084] Step S43: For the labeled vascular data, use the label GT to supervise the output of the student model, that is, the vascular prediction result Y. The loss function is selected as Dice, and the calculation is expressed as:

[0085]

[0086] For the unlabeled data X ul , use the vascular segmentation result Y tea of the teacher model for the same data as the pseudo-label to guide the vascular prediction result of the student model. Since the result output by the model is generally a probability distribution, it is necessary to perform binary quantization on Y tea by Y tea =Y tea >0.5 before using it as the pseudo-label. Thus, the loss function of the unlabeled data here can be expressed as:

[0087]

[0088] Step S44: For the output Z stu of the encoder of the student model, that is, the high-dimensional vector representation of the original three-dimensional block, use the vector quantized by the teacher model through the codebook for constraint, so as to promote the student model to encode more high-quality. Take the L1 norm of the quantized feature encoding of the teacher model and the encoding output by the student model as the loss function, and the calculation method is:

[0089]

[0090] Step S45: Take the weighted sum of the label segmentation loss, the pseudo-label loss, and the encoding loss as the final loss function of the student model, and the calculation method is:

[0091] Student model loss function = segmentation loss function + β * semi-supervised loss function + feature distillation loss function

[0092] Among them, β is set as a Gaussian warm-up coefficient that increases with the number of training rounds according to experience. The value of β in the t-th round is calculated as follows:

[0093]

[0094] Step S5: Use the trained lightweight student model to complete the segmentation of the whole-brain microscopic images, and then reconstruct these three-dimensional blocks into a complete three-dimensional whole-brain blood vessel network according to the cutting method in step S1.

[0095] The present invention conducts ablation experiments on the VesSep2020 dataset. Through a large number of experimental results, it is proved that the knowledge distillation algorithm based on codebook quantization is beneficial to enhancing the analysis ability of the student model for microscopic blood vessel data and improving the segmentation efficiency of the model for three-dimensional microscopic blood vessel data. As Figure 5 shown, compared with other comparison methods, the segmentation results of the present invention not only have fewer misclassification and missed classification situations, but also only require a shorter inference time, which is beneficial to the subsequent computer vision algorithm based on codebook quantization to realize the segmentation, reconstruction and pathological analysis of the whole-brain blood vessels.

[0096] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A three-dimensional microscopic blood vessel image segmentation method, characterized in that It includes the following steps: Step S1: Stack the microvascular images taken layer by layer in a three-dimensional space in sequence, and then cut them into a group of three-dimensional blocks; Step S2: On the unlabeled three-dimensional microvascular data, pre-train the teacher model through a reconstruction task, and simultaneously learn a codebook containing the coding patterns of the vascular structure distribution; Step S3: On the labeled three-dimensional microvascular data, fine-tune the model parameters by label-supervising the teacher model; Step S4: Use the teacher model to guide and train the student model with limited labeled data, and use the codebook to guide the student model to extract data; Step S5: Use the trained student model to complete the segmentation of the whole-brain microscopic image, and then reconstruct the obtained three-dimensional blocks into a complete three-dimensional whole-brain vascular network.

2. The three-dimensional microscopic blood vessel image segmentation method according to claim 1, wherein: The said Step S2 includes the following steps: Step S21: Feed the unlabeled three-dimensional microvascular data processed in Step S1 into the teacher model. The teacher model reconstructs the original three-dimensional block after encoding, codebook quantization, and decoding, and constructs a reconstruction loss; Step S22: Construct a vascular structure pattern codebook at the output position of the encoder of the teacher model. The dimension of each quantization vector is the same as the dimension of the vector output by the encoder of the student model. During the pre-training process of the teacher model, continuously update this codebook, and construct a codebook loss; Step S23: Based on the reconstruction loss and the codebook loss, construct an overall loss for updating the parameters of the teacher model.

3. The three-dimensional microscopic blood vessel image segmentation method according to claim 2, wherein: In the said Step S22, first, make the dimension of the output vector of the encoder of the teacher model the same as the dimension of the output vector of the encoder of the student model. Then, use the code vector with the closest Euclidean distance to each vector in the codebook for replacement to form a discretized encoding. Then, raise the dimension of the discretized encoding to be the same as before the dimension reduction, and feed it into the decoder of the teacher model.

4. A three-dimensional microscopic blood vessel image segmentation method according to claim 3, characterized in that: The sampling and encoding replacement process is as follows: Among them, V represents the vascular structure pattern codebook, z i represents the encoder output vector of the teacher model, DA(*) represents dimension alignment, and ‖*‖2 represents the 2-norm; for each vector z i , find the quantization vector v with the smallest Euclidean distance from it in the codebook V j , where j is the index of the quantization vector corresponding to z i matched in this search in the codebook, and i is the vector serial number of the encoder in the teacher model; then, use Gumbel-Softmax for approximate discretization to obtain the quantized vector Finally, replace the original encoder output vector z i , and feed it into the decoder of the teacher model; The codebook loss function is expressed as: Among them, represents the square of the 2-norm, sg[*] represents the stop-gradient operation, and λ represents the balance coefficient.

5. A three-dimensional microscopic blood vessel image segmentation method according to claim 2, characterized in that: In the step S21, calculate the average value of the absolute values of the luminance differences between the original three-dimensional block X and the reconstructed three-dimensional block pixel by pixel, and use it as the reconstruction loss function: where N represents the number of three-dimensional blocks.

6. A three-dimensional microscopic blood vessel image segmentation method according to claim 1, characterized in that: The said Step S4 includes the following steps: Step S41: Freeze all the parameters of the teacher model and do not update them during the subsequent training process; Step S42: Combine the unlabeled three-dimensional blocks and the labeled three-dimensional blocks in a certain proportion to form a batch, and feed them into the teacher model and the student model respectively; Step S43: For the labeled vascular data, use the vascular label to supervise the vascular prediction result output by the student model, and construct a second segmentation loss for the labeled data; for the unlabeled data, use the vascular segmentation result of the teacher model for the same data as the pseudo label to guide the vascular prediction result of the student model, and construct an unlabeled semi-supervised loss; Step S44: For the high-dimensional vector representation of the original three-dimensional block output by the encoder of the student model, use the vector quantized by the teacher model through the codebook for constraint, and construct a feature distillation loss; Step S45: Based on the second segmentation loss, the semi-supervised loss, and the feature distillation loss, construct the loss of the student model for training the student model.

7. A three-dimensional microscopic blood vessel image segmentation method according to claim 6, characterized in that: In the said Step S3, the first segmentation loss function is as follows: Among them, Y tea represents the predicted result of vessel segmentation output by the teacher model, and GT represents the vessel label; In the said Step S43, the second segmentation loss function used is as follows: where Y represents the vascular prediction result output by the student model; The semi-supervised loss function for unlabeled data is as follows: In the step S44, the L1 norm of the quantization feature encoding of the teacher model and the encoding output by the student model is used as the feature distillation loss function: Among them, represents the vector obtained by quantizing the teacher model through the codebook, Z stu represents the high-dimensional vector of the original three-dimensional block; In the step S45, the weighted sum of the second segmentation loss, the semi-supervised loss, and the feature distillation loss is used as the student model loss function, and the calculation method is: Student model loss function = segmentation loss function + β * semi-supervised loss function + feature distillation loss function where β is set as a Gaussian warm-up coefficient that increases with the number of training rounds according to experience, and the calculation method for the value of β in the t-th round is:

8. A three-dimensional microscopic blood vessel image segmentation method according to claim 1, characterized in that: The step S1 specifically includes the following steps: Step S11: Stack the microvascular images taken layer by layer in a three-dimensional space in the shooting order to form three-dimensional microscopic data, and read S images each time; Step S12: Axially cut the large three-dimensional image into a number of three-dimensional cubes with side length S; Step S13: Perform data preprocessing on the three-dimensional image, perform histogram statistics on the brightness of the complete data set to obtain the threshold quantile of the overall data set brightness, and then perform histogram truncation on the data set, truncating the excessive values to the threshold quantile.

9. A three-dimensional microscopic whole cerebral vascular image segmentation method, characterized in that: Using the three-dimensional microvascular image segmentation method according to any one of claims 1 to 8, segment the three-dimensional microscopic whole cerebral vascular image.

10. A three-dimensional microscopic whole cerebral vascular image segmentation system, comprising a three-dimensional block generation module, a teacher module, a student module, and a reconstruction module, characterized in that: According to the three-dimensional microscopic whole cerebral vascular image segmentation method described in claim 9, for the three-dimensional microscopic whole cerebral vascular image, generate three-dimensional blocks in sequence, generate and train the teacher model codebook, and train the student model with the help of the codebook for subsequent vascular image segmentation, and reconstruct the three-dimensional whole cerebral vascular network through the reconstruction module.

Images