Cerebral vessel segmentation method, system and device based on TOF-MRA image and storage medium

Abstract

The invention discloses a cerebrovascular segmentation method, system and device based on a TOF-MRA image and a storage medium, and relates to the field of artificial intelligence, the method comprises the following steps: collecting TOF-MRA image data, the TOF-MRA image data comprising a cerebrovascular image and a label; preprocessing the TOF-MRA image data; constructing an image segmentation network; training an image segmentation network by using the preprocessed TOF-MRA image data; storing the optimal image segmentation network; and predicting a cerebrovascular segmentation result by using the optimal image segmentation network. According to the method, the defects of poor segmentation performance, low model training reasoning speed and the like caused by a complex topological structure of the brain blood vessel are overcome, and the accuracy of brain blood vessel segmentation and the integrity of a blood vessel map are ensured and improved.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and information technology, and in particular to a cerebral blood vessel segmentation method, system, device and storage medium based on TOF-MRA images. Background Art

[0002] Cerebrovascular diseases, such as cerebral aneurysms, cerebral thrombosis, cerebral hemorrhage, and cerebral vascular malformations, constitute a major challenge facing the global health field and are one of the main causes of disability and death. The first step to deeply understand cerebrovascular lesions and their association with neurological diseases is to achieve fast and accurate cerebrovascular segmentation. Clinically, commonly used cerebrovascular imaging technologies include: Magnetic Resonance Angiography (MRA), Computed Tomography Angiography (CTA), and Digital Subtraction Angiography (DSA). MRA can obtain high-resolution cerebrovascular images without radiation and non-invasively. Among them, TOF-MRA (Time-of-Flight Magnetic Resonance Angiography) is suitable for multiple examinations of diseases such as aneurysms, stenosis, and arteriovenous malformations because of its characteristic of obtaining clear vascular images without the injection of contrast agents.

[0003] Although the contrast between cerebral vessels and surrounding tissues in TOF-MRA images is high, efficient and high-precision automated segmentation is still a challenge. Convolutional neural network methods have gradually become the mainstream method for cerebral vascular segmentation because they can automatically learn complex and abstract feature representations from large amounts of data. Deep learning models require sufficient high-resolution imaging data and accurately segmented and labeled data to obtain better clinical results, which consumes a lot of computing resources. At the same time, cerebral vascular segmentation models need to have a high model complexity to obtain more accurate segmentation results, which makes it a challenge to adjust and select appropriate network parameters.

[0004] Current segmentation models all perform feature extraction in grid image space, without considering feature extraction for non-grid irregular structures. The spatial topology of cerebral blood vessels is complex, so it is necessary to design convolution operations that can capture spatial heterogeneity distribution to better learn tubular orientation features.

[0005] Therefore, there is an urgent need on the market for a dynamic configuration method and system for cerebrovascular segmentation based on TOF-MRA images to overcome the defects of existing technologies and models that cannot achieve optimal performance under existing computing resources, do not consider the differences in blood vessels in the x, y, and z directions, and do not consider the balance between multi-task objectives, so as to achieve more accurate cerebrovascular segmentation. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention provides a cerebral vascular segmentation method, system, device and storage medium based on TOF-MRA images, so as to overcome the defects that the prior art and model cannot obtain the optimal performance under the existing computing resources, do not consider the differences of blood vessels in the x, y and z directions, and do not consider the balance between multi-task objectives, so as to achieve more accurate cerebral vascular segmentation.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] A cerebral blood vessel segmentation method based on TOF-MRA images comprises the following steps:

[0009] Collecting TOF-MRA image data, including cerebral vascular images and labels;

[0010] Preprocess TOF-MRA image data;

[0011] Build an image segmentation network;

[0012] Use the preprocessed TOF-MRA image data to train the image segmentation network; save the optimal image segmentation network;

[0013] Use the optimal image segmentation network to predict cerebral vascular segmentation results.

[0014] To optimize the above technical solutions, the specific measures taken also include:

[0015] Furthermore, the preprocessing of TOF-MRA image data specifically includes:

[0016] The skull part in the cerebrovascular image is removed, and the cerebrovascular image and label are resampled respectively; the cerebrovascular image is subjected to Z-score normalization.

[0017] Furthermore, the image segmentation network includes an encoder and a decoder, each layer of the encoder is a dynamic feature perception module, and each layer of the decoder is a dynamic feature perception module; the dynamic feature perception module includes a multi-scale feature fusion module, a dynamic snake convolution module in the x direction, a dynamic snake convolution module in the y direction, and a dynamic snake convolution module in the z direction; the TOF-MRA image data is input into each layer of the encoder, and the convolution kernels of three sizes of 1, 3, and 5 of the multi-scale feature fusion module are respectively extracted and fused to obtain a first feature map, the dynamic snake convolution module in the x direction, the dynamic snake convolution module in the y direction, and the dynamic snake convolution module in the z direction respectively output a second feature map, a third feature map, and a fourth feature map, the first feature map, the second feature map, the third feature map, and the fourth feature map are fused to obtain a fused feature, the fused feature output by each layer of the encoder is input into the corresponding layer of the decoder, and the decoder outputs a cerebrovascular segmentation result.

[0018] Furthermore, the dynamic snake convolution module is used to extract the curvature features of the cerebral vascular image, which is implemented based on the offset strategy of the convolution kernel. The offset strategy is specifically:

[0019] The deformation offset is Δ, given a 7×7 convolution kernel K, whose center coordinate is marked as K i =(x i ,y i ), the offset changes of the convolution kernel K in the x-axis and y-axis directions are expressed as follows:

[0020]

[0021]

[0022] In the formula, K i+m represents the change of the convolution kernel K after the shift in the x direction, (c i+m ,y i+m ) represents the center coordinate K i The coordinates at a distance m along the positive x and y directions, (x i-m ,y i-m ) represents the center coordinate K i The coordinates at a distance m along the negative x and y directions, x i +m means an offset of m is generated along the positive direction of the x-axis. i -m means an offset of m is generated along the negative direction of the x-axis. It represents the cumulative sum of all offsets from position i to position i+m in the y direction. represents the cumulative sum of all offsets from position im to position i, K j+m Indicates the change in the convolution kernel offset in the y direction, (x j+m ,y j+m0 means the center coordinate K j =(x j ,y j ) is the coordinate at a distance m along the positive x and y directions, (x j-m ,y j-m ) represents the center coordinate K i The coordinates at a distance m along the negative x and y directions, It represents the cumulative sum of all offsets from position j to position j+m in the x direction. Represents the cumulative sum of all offsets from position jm to position j in the x direction.

[0023] Furthermore, the resampling of the cerebral vascular image and the label is specifically as follows:

[0024] The cerebrovascular images are three-dimensional images. For anisotropic cerebrovascular images, cubic spline interpolation is used in the xy plane, and nearest neighbor interpolation is used outside the xy plane. For isotropic cerebrovascular images, cubic spline interpolation is used for overall resampling.

[0025] For labels, they are first converted into one-hot encoding. For anisotropic labels, linear interpolation is used in the xy plane, and nearest neighbor interpolation is used outside the xy plane. For isotropic labels, linear interpolation is performed.

[0026] The cubic spline interpolation is specifically:

[0027] The interpolation function S(x,y,z) of three-dimensional cubic spline interpolation is expressed as:

[0028]

[0029] In the formula, l, m, n represent the control point indexes in the x, y, and z directions respectively, and N x ,N y ,N z Represents the total number of control points in three directions, c lmn is located at the control point (x l ,y m ,z n ), B l (x),B m (y),B n (z) represents the cubic spline basis functions along the x, y, and z directions respectively;

[0030] The linear interpolation is specifically:

[0031] Based on the linear combination of the eight neighboring points where the voxel is located, the three-dimensional linear interpolation estimates the intermediate value by connecting the straight lines between the adjacent data points in three directions. The intensity values ​​of the eight neighboring points are V1, V2, V3, V4, V5, V6, V7, and V8 respectively;

[0032] The interpolation ratios used to describe the positional relationship of the target points in the three-dimensional cube are defined in three directions as:

[0033]

[0034] Where (x, y, z) represents the coordinates of the target point to be interpolated, (x0, y0, z0) represents the coordinates of the minimum boundary vertex in the target point cube, (x1, y1, z1) represents the coordinates of the maximum boundary vertex in the target point cube, u represents the interpolation ratio in the x direction, v represents the interpolation ratio in the y direction, and w represents the interpolation ratio in the z direction;

[0035] Based on the interpolation ratio, the interpolation result of the target point is expressed as the weighted sum of the intensity values ​​of the eight adjacent points:

[0036] V(x,y,z)=(1-u)(1-v)(1-w)V1+(1-u)(1-v)wV2+(1-u)v(1-w)V3+(1-u)vwV4+u(1-v)(1-w)V5+u(1-v)wV6+uv(1-w)V7+uvwV8

[0037] Where V(x,y,z) represents the interpolation result of the target point.

[0038] Furthermore, the total loss in the process of training the image segmentation network is a weighted sum of the segmentation loss, the centerline loss and the category loss;

[0039] The total loss function expression is as follows:

[0040] Loss = λ1Loss Dice +λ2Loss cl +λ3Loss CE

[0041] In the formula, Loss represents the total loss, and λ1 represents the segmentation loss Loss Dice The weight of λ2 represents the center line loss Loss cl The weight of λ3 represents the category loss Loss CE The weight of

[0042] The weight value selection rule is the dynamic weight strategy, specifically:

[0043] From the standard normal distribution Sampling weight And the sampling sequence is mapped by the mapping function f(·) Normalization is performed, and the formula is as follows:

[0044]

[0045] In the formula, λ represents the final weight value obtained by function mapping, represents a set of all sample values, and the mapping function adopts the softmax function.

[0046] Furthermore, the segmentation loss is measured using Dice loss, the centerline loss is measured using centerline Dice loss, and the category loss is measured using cross entropy loss;

[0047] The calculation formula of the Dice loss is as follows:

[0048]

[0049] In the formula, Loss Dice represents the Dice loss, N represents the total number of voxels in the image, and p k represents the predicted value of the kth voxel, g k represents the true value of the k-th voxel;

[0050] The calculation formula of the centerline Dice loss is as follows:

[0051]

[0052] In the formula, Loss cl represents the centerline Dice loss, Tprec(S P ,V L ) represents the topological accuracy, Tsens(S L ,V P ) represents the topological sensitivity, V L and V P Represent the true segmentation mask and the predicted segmentation result respectively, S L and S P Represent the skeletons of the true segmentation mask and the predicted segmentation result respectively;

[0053] The calculation formula of the cross entropy loss is as follows:

[0054]

[0055] Among them, Loss CE Represents the cross entropy loss, which is used to measure the difference between the predicted probability distribution and the true label distribution, p k represents the predicted value of the kth voxel, g krepresents the true value of the kth voxel, and N represents the total number of voxel points in the image.

[0056] The present invention also proposes a cerebral vascular segmentation system based on TOF-MRA images, comprising:

[0057] A data acquisition module, used for acquiring TOF-MRA image data, the TOF-MRA image data including cerebral vascular images and labels;

[0058] Data preprocessing module, used to preprocess TOF-MRA image data;

[0059] Network construction module, used to build image segmentation network;

[0060] A training module is used to train the image segmentation network using the pre-processed TOF-MRA image data and save the optimal image segmentation network;

[0061] The inference module is used to predict the cerebrovascular segmentation results using the optimal image segmentation network.

[0062] The present invention also proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the cerebral blood vessel segmentation method based on TOF-MRA images as described above is implemented.

[0063] The present invention further provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the cerebral blood vessel segmentation method based on TOF-MRA images as described above.

[0064] The beneficial effects of the present invention are:

[0065] (1) Considering the slender and curved topological structure of blood vessels, a dynamic feature perception module consisting of dynamic snake convolution and multi-scale feature fusion is designed. The dynamic snake convolution can dynamically perceive the spatial distribution heterogeneity of blood vessels. The multi-scale feature fusion can take into account the advantages of small-size convolution kernels in capturing local details and edge information and the advantages of large-size convolution kernels in capturing a wider range of contextual information, thereby obtaining richer feature expressions.

[0066] (2) A dynamically weighted loss function with random initialization is designed. During training, both the mini-batch data sampling and the weights of each part of the loss function are random. The double randomness makes it more effective in escaping the local optimum than the method with fixed loss weights.

[0067] (3) Resample the image and label separately and perform Z-score normalization on the image. The image block size and batch size of the input network can be automatically configured according to the median size of the resampled data and the memory size of the GPU.

[0068] (4) An automatic parameter setting framework for the adaptive graphics computing unit was designed to enable the cerebrovascular segmentation model to achieve better segmentation accuracy on existing hardware devices.

[0069] (5) The higher segmentation accuracy and lower false positives make this segmentation method a reference option for cerebral vascular segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 A flowchart of a cerebral blood vessel segmentation method based on TOF-MRA images provided by the present invention;

[0071] Figure 2 and Figure 3 Schematic diagram of dynamic snake convolution offset strategy;

[0072] Figure 4 It is a schematic diagram of the results of the present invention and other comparative methods;

[0073] Figure 5 A schematic diagram of the structure of an image processing device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0074] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0075] Embodiment 1

[0076] The present invention proposes a cerebral vascular segmentation method based on TOF-MRA images. The overall process is as follows: Figure 1 As shown, the following steps are included:

[0077] TOF-MRA image data were collected, including cerebrovascular images and labels; TOF-MRA image data collected by the Biomedical Image Analysis Group, Imperial College London from three hospitals in London, UK were screened, and data from two of them with the same resolution and voxel spacing were selected;

[0078] Preprocessing of TOF-MRA image data; including: using FreeSufer software to remove the skull part in the cerebrovascular image, resampling the cerebrovascular image and labels respectively; performing Z-score standardization on the cerebrovascular image.

[0079] The specific resampling of cerebrovascular images and labels is as follows:

[0080] The cerebrovascular images are three-dimensional images. For anisotropic cerebrovascular images, cubic spline interpolation is used in the xy plane, and nearest neighbor interpolation is used outside the xy plane. For isotropic cerebrovascular images, cubic spline interpolation is used for overall resampling.

[0081] For labels, they are first converted into one-hot encoding. For anisotropic labels, linear interpolation is used in the xy plane, and nearest neighbor interpolation is used outside the xy plane. For isotropic labels, linear interpolation is performed.

[0082] The cubic spline interpolation is specifically:

[0083] The interpolation function S(x,y,z) of three-dimensional cubic spline interpolation is expressed as:

[0084]

[0085] In the formula, l, m, n represent the control point indexes in the x, y, and z directions respectively, and N x ,N y ,N z Represents the total number of control points in three directions, c lmn is located at the control point (x l ,y m ,z n ), B l (x),B m (y),B n (z) represents the cubic spline basis functions along the x, y, and z directions respectively;

[0086] The linear interpolation is specifically:

[0087] Based on the linear combination of the eight neighboring points where the voxel is located, the three-dimensional linear interpolation estimates the intermediate value by connecting the straight lines between the adjacent data points in three directions. The intensity values ​​of the eight neighboring points are V1, V2, V3, V4, V5, V6, V7, and V8 respectively;

[0088] The interpolation ratios used to describe the positional relationship of the target points in the three-dimensional cube are defined in three directions as:

[0089]

[0090] Where (x, y, z) represents the coordinates of the target point to be interpolated, (x0, y0, z0) represents the coordinates of the minimum boundary vertex in the target point cube, (x1, y1, z1) represents the coordinates of the maximum boundary vertex in the target point cube, u represents the interpolation ratio in the x direction, v represents the interpolation ratio in the y direction, and w represents the interpolation ratio in the z direction;

[0091] Based on the interpolation ratio, the interpolation result of the target point is expressed as the weighted sum of the intensity values ​​of the eight adjacent points:

[0092] V(x,y,z)=(1-u)(1-v)(1-w)V1+(1-u)(1-v)wV2+(1-u)v(1-w)V3+(1-u)vwV4+u(1-v)(1-w)V5+u(1-v)wV6+ub(1-w)V7+uvwV8

[0093] Where V(x,y,z) represents the interpolation result of the target point.

[0094] The calculation process of the Z-score standardization is:

[0095]

[0096] Among them, K new is the standardized data, K is the original data, specifically the intensity value of the MRA image voxel, μ is the mean of the original data, and σ is the standard deviation of the original data. Each image is standardized using its own mean and standard deviation. This method is very effective for possible imaging intensity changes between different patients and helps to improve the adaptability of the model.

[0097] An image segmentation network is constructed; the image segmentation network includes an encoder and a decoder, each layer of the encoder is a dynamic feature perception module, and each layer of the decoder is a dynamic feature perception module; the dynamic feature perception module includes a multi-scale feature fusion module, a dynamic snake convolution module in the x direction, a dynamic snake convolution module in the y direction, and a dynamic snake convolution module in the z direction; the TOF-MRA image data is input into each layer of the encoder, and the convolution kernels of three sizes of 1, 3, and 5 of the multi-scale feature fusion module are respectively used for feature extraction and fusion to obtain a first feature map, a small-sized convolution kernel is better at capturing local details and edge information, which is helpful for extracting vascular detail features; and a large-sized convolution kernel is more suitable for capturing a larger range of contextual information, which is helpful for maintaining the integrity and continuity of the vascular topological structure. The dynamic snake convolution module in the x-direction, the dynamic snake convolution module in the y-direction, and the dynamic snake convolution module in the z-direction output the second feature map, the third feature map, and the fourth feature map respectively. The first feature map, the second feature map, the third feature map, and the fourth feature map are fused to obtain the fused feature. The fused feature output by each layer of the encoder is input into the corresponding layer of the decoder, and the decoder outputs the cerebrovascular segmentation result.

[0098] The dynamic snake convolution module is used to extract the curvature features of cerebrovascular images. It is implemented based on the offset strategy of the convolution kernel. The schematic diagram of the dynamic snake convolution offset strategy is shown in Figure 2 and 3 The specific offset strategy is:

[0099] The deformation offset is Δ, and the limit range of Δ is [-1, 1]. Given a 7×7 convolution kernel K, its center coordinate is marked as K i =(x i ,y i ), the offset changes of the convolution kernel K in the x-axis and y-axis directions are expressed as follows:

[0100]

[0101] In the formula, K i+m represents the change of convolution kernel K after shifting in the x direction, (x i+m ,y i+m ) represents the center coordinate K i The coordinates at a distance m along the positive x and y directions, (x i-m ,y i-m ) represents the center coordinate K i The coordinates at a distance m along the negative x and y directions, x i +m means an offset of m is generated along the positive direction of the x-axis. i -m means an offset of m is generated along the negative direction of the x-axis. It represents the cumulative sum of all offsets from position i to position i+m in the y direction. represents the cumulative sum of all offsets from position im to position i, K j+m Indicates the change in the convolution kernel offset in the y direction, (x j+m ,y j+m ) represents the center coordinate K j =(x j ,y j ) is the coordinate at a distance m along the positive x and y directions, (x j-m ,y j-m ) represents the center coordinate K i The coordinates at a distance m along the negative x and y directions, It represents the cumulative sum of all offsets from position j to position j+m in the x direction. Represents the cumulative sum of all offsets from position jm to position j in the x direction.

[0102] The image segmentation network was trained using the preprocessed TOF-MRA image data. Five-fold cross validation was used for training and validation, and the optimal image segmentation network was saved. The total loss during the image segmentation network training process was the weighted sum of the segmentation loss, centerline loss, and category loss.

[0103] The total loss function expression is as follows:

[0104] Loss = λ1Loss Dice +λ2Loss cl +λ3Loss CE

[0105] In the formula, Loss represents the total loss, and λ1 represents the segmentation loss Loss Dice The weight of λ2 represents the center line loss Loss cl The weight of λ3 represents the category loss Loss CE The weight of

[0106] The weight value selection rule is the dynamic weight strategy, specifically:

[0107] From the standard normal distribution Sampling weight And the sampling sequence is mapped by the mapping function f(·) Normalization is performed, and the formula is as follows:

[0108]

[0109] In the formula, λ represents the final weight value obtained by function mapping, Represents a set of all sampled values, and the mapping function uses a softmax function. The double randomness generated by mini-batch data sampling during training and the weights of each part of the loss function through standard normal distribution sampling can more effectively assist the function to be optimized to jump out of the local minimum, thereby improving the generalization performance compared to fixed weights.

[0110] The segmentation loss is measured using Dice loss (Dice Loss), the centerline loss is measured using centerline Dice loss (clDiceLoss), and the category loss is measured using cross-entropy loss (Cross-Entropy Loss);

[0111] The calculation formula of the Dice loss is as follows:

[0112]

[0113] In the formula, Loss Dice represents the Dice loss, N represents the total number of voxels in the image, and p k represents the predicted value of the kth voxel, g k represents the true value of the k-th voxel;

[0114] The predicted binary segmentation results and the true labels are converted into corresponding skeleton representations through the morphological skeleton algorithm. Skeletonization is a result of thinning, and only the center line of the object is retained. The calculation formula of the center line Dice loss is as follows:

[0115]

[0116] In the formula, Loss cl represents the centerline Dice loss, which is suitable for processing targets with complex topological structures. P ,V L ) represents the topological accuracy, Tsens(S L ,V P ) represents the topological sensitivity, V L and V P Represent the true segmentation mask and the predicted segmentation result respectively, S L and S P Respectively represent the skeleton of the real segmentation mask and the predicted segmentation result; the topological accuracy Tprec(S P ,V L ) and topological sensitivity Tsens(S L ,V P ) are calculated as follows:

[0117]

[0118] Among them, VL and V P Represent the true segmentation mask and the predicted segmentation result respectively, S L and S P They represent the skeleton of the true segmentation mask and the predicted segmentation result respectively, and ε is the smoothing factor, which is set to 10-4.

[0119] The calculation formula of the cross entropy loss is as follows:

[0120]

[0121] Among them, Loss CE Represents the cross entropy loss, which is used to measure the difference between the predicted probability distribution and the true label distribution, p k represents the predicted value of the kth voxel, g k represents the true value of the kth voxel, and N represents the total number of voxel points in the image.

[0122] Use the optimal image segmentation network to predict cerebral vascular segmentation results.

[0123] Embodiment 2

[0124] The present invention proposes a cerebral blood vessel segmentation system based on TOF-MRA images corresponding to the method of embodiment 1, comprising:

[0125] A data acquisition module, used for acquiring TOF-MRA image data, the TOF-MRA image data including cerebral vascular images and labels;

[0126] Data preprocessing module, used to preprocess TOF-MRA image data;

[0127] Network construction module, used to build image segmentation network;

[0128] A training module is used to train the image segmentation network using the pre-processed TOF-MRA image data and save the optimal image segmentation network;

[0129] The inference module is used to predict the cerebrovascular segmentation results using the optimal image segmentation network.

[0130] The implementation methods of each module and module function in the system are completely consistent with the steps of the method in Example 1, so they will not be repeated here.

[0131] Embodiment 3

[0132] The present invention proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the cerebral blood vessel segmentation method based on TOF-MRA images as described in Example 1 is implemented.

[0133] Embodiment 4

[0134] The present invention provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the cerebral blood vessel segmentation method based on TOF-MRA images as described in the first embodiment.

[0135] The effect of the technical solution of the present invention is further verified and explained below in conjunction with specific experiments.

[0136] In order to verify the accuracy and excellence of the method of the present invention, the method of the present invention and the classic segmentation method CS 2 The performance of Net, DeepVesselNet and ResUNet in segmentation indicators, the experimental data results are compared as follows:

[0137]

[0138] The model segmentation prediction results of the present invention and the classic segmentation method CS 2 Net, DeepVesselNet and ResUNet sample segmentation results are shown in Figure 4 As shown. In the segmentation of sample (a), the Dy-SnakeUNet method proposed in the present invention predicts segmentation more accurately in the more peripheral areas. The three classic segmentation methods have a large number of false positive predictions (such as around the basilar artery BA segment), and the partial structure of the Willis circle is more accurate than the three comparison methods. The comparison methods have different degrees of false positive predictions in the area of ​​the Willis circle, while the method of the present invention basically has no false positive voxels there. In the segmentation of sample (b), the method of the present invention has a more obvious advantage in false positives, and maintains relatively correct connectivity and segmentation details in the magnified area. From the overall level, the blood vessels segmented by using Dy-SnakeUNet have a relatively complete topological structure, and in the local small and curved parts, Dy-SnakeUNet has higher segmentation accuracy and connectivity than the other three methods, which is due to the feature perception capabilities of the dynamic snake convolution and multi-scale feature fusion modules. At the same time, compared to CS 2 Net, DeepVesselNet, and ResUNet have lower false positives and higher specificity, which has a positive effect on automatic segmentation models to assist clinical diagnosis.

[0139] In the embodiments disclosed in the present application, the computer storage medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. The computer storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the above. More specific examples of computer storage media can include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.

[0140] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0141] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A cerebral vascular segmentation method based on TOF-MRA images, characterized in that: The following steps are involved: Collecting TOF-MRA image data, including cerebral vascular images and labels; Preprocess TOF-MRA image data; Build an image segmentation network; Use the preprocessed TOF-MRA image data to train the image segmentation network; save the optimal image segmentation network; Use the optimal image segmentation network to predict cerebral vascular segmentation results.

2. The method for cerebral blood vessel segmentation based on TOF-MRA images according to claim 1, characterized in that: The preprocessing of TOF-MRA image data specifically includes: The skull part in the cerebrovascular image is removed, and the cerebrovascular image and label are resampled respectively; the cerebrovascular image is subjected to Z-score normalization.

3. The method for cerebral blood vessel segmentation based on TOF-MRA images according to claim 1, characterized in that: The image segmentation network includes an encoder and a decoder, each layer of the encoder is a dynamic feature perception module, and each layer of the decoder is a dynamic feature perception module; the dynamic feature perception module includes a multi-scale feature fusion module, a dynamic snake convolution module in the x direction, a dynamic snake convolution module in the y direction, and a dynamic snake convolution module in the z direction; the TOF-MRA image data is input into each layer of the encoder, and the convolution kernels of three sizes of 1, 3, and 5 of the multi-scale feature fusion module are respectively extracted and fused to obtain a first feature map, the dynamic snake convolution module in the x direction, the dynamic snake convolution module in the y direction, and the dynamic snake convolution module in the z direction respectively output a second feature map, a third feature map, and a fourth feature map, and the first feature map, the second feature map, the third feature map, and the fourth feature map are fused to obtain a fusion feature, the fusion feature output by each layer of the encoder is input into the corresponding layer of the decoder, and the decoder outputs a cerebrovascular segmentation result.

4. The method for cerebral blood vessel segmentation based on TOF-MRA images according to claim 3, characterized in that: The dynamic snake convolution module is used to extract the curvature features of the cerebral vascular image, and is implemented based on the offset strategy of the convolution kernel. The specific offset strategy is: The deformation offset is Δ, given a 7×7 convolution kernel K, whose center coordinate is marked as K i =(x i ,y i ), the offset changes of the convolution kernel K in the x-axis and y-axis directions are expressed as follows: In the formula, K i+m represents the change of the convolution kernel K after the shift in the x direction, (x i+m ,y i+m ) represents the center coordinate K i The coordinates at a distance m along the positive x and y directions, (x i-m ,y i-m ) represents the center coordinate K i The coordinates at a distance m along the negative x and y directions, x i +m means an offset of m is generated along the positive direction of the x-axis. i -m means an offset of m is generated along the negative direction of the x-axis. Δy represents the cumulative sum of all offsets from position i to position i+m in the y direction. Δy represents the cumulative sum of all offsets from position im to position i, K j+m Indicates the change in the convolution kernel offset in the y direction, (x j+m ,y j+m ) represents the center coordinate K j =(x j ,y j 0 is the coordinate at a distance m from the positive x and y directions, (x j-m ,y j-m ) represents the center coordinate K i The coordinates at a distance m along the negative x and y directions, Δx represents the cumulative sum of all offsets from position j to position j+m in the x direction. Δx represents the cumulative sum of all offsets from position jm to position j in the x direction.

5. The method for cerebral blood vessel segmentation based on TOF-MRA images according to claim 2, characterized in that: The resampling of the cerebral vascular image and the label is specifically as follows: The cerebrovascular images are three-dimensional images. For anisotropic cerebrovascular images, cubic spline interpolation is used in the xy plane, and nearest neighbor interpolation is used outside the xy plane. For isotropic cerebrovascular images, cubic spline interpolation is used for overall resampling. For labels, they are first converted into one-hot encoding. For anisotropic labels, linear interpolation is used in the xy plane, and nearest neighbor interpolation is used outside the xy plane. For isotropic labels, linear interpolation is performed. The cubic spline interpolation is specifically: The interpolation function S(x,y,z) of three-dimensional cubic spline interpolation is expressed as: In the formula, l, m, n represent the control point indexes in the x, y, and z directions respectively, and N x ,N y ,N z Represents the total number of control points in three directions, c lmn is located at the control point (x l ,y m ,z m ), B l (x),B m (y),B n (z) represents the cubic spline basis functions along the x, y, and z directions respectively; The linear interpolation is specifically: Based on the linear combination of the eight neighboring points where the voxel is located, the three-dimensional linear interpolation estimates the intermediate value by connecting the straight lines between the adjacent data points in three directions. The intensity values ​​of the eight neighboring points are V1, V2, V3, V4, V5, V6, V7, and V8 respectively; The interpolation ratios used to describe the positional relationship of the target points in the three-dimensional cube are defined in three directions as: Where (x, y, z) represents the coordinates of the target point to be interpolated, (x0, y0, z0) represents the coordinates of the minimum boundary vertex in the target point cube, (x1, y1, z1) represents the coordinates of the maximum boundary vertex in the target point cube, u represents the interpolation ratio in the x direction, v represents the interpolation ratio in the y direction, and w represents the interpolation ratio in the z direction; Based on the interpolation ratio, the interpolation result of the target point is expressed as the weighted sum of the intensity values ​​of the eight adjacent points: V(x,y,z)=(1-u)(1-v)(1-w)V1+(1-u)(1-v)wV2+(1-u)v(1-w)V3+(1-u)vwV4+u(1-v)(1-w)V5+u(1-v)wV6+uv(1-w)V7+uvwV8 Where V(x,y,z) represents the interpolation result of the target point.

6. The method for cerebral blood vessel segmentation based on TOF-MRA images according to claim 1, characterized in that: The total loss in the process of training the image segmentation network is the weighted sum of the segmentation loss, the centerline loss and the category loss; The total loss function expression is as follows: Loss=λ1Loss Dice +λ2Loss cl +λ3Loss CE In the formula, Loss represents the total loss, and λ1 represents the segmentation loss Loss Dice The weight of λ2 represents the center line loss Loss cl The weight of λ3 represents the category loss Loss CE The weight of The weight value selection rule is the dynamic weight strategy, specifically: Sample weights from a standard normal distribution N(0,1) And the sampling sequence is mapped by the mapping function f(·) Normalization is performed, and the formula is as follows: In the formula, λ represents the final weight value obtained by function mapping, represents a set of all sample values, and the mapping function adopts the softmax function.

7. The method for cerebral blood vessel segmentation based on TOF-MRA images according to claim 6, characterized in that: The segmentation loss is measured using Dice loss, the centerline loss is measured using centerline Dice loss, and the category loss is measured using cross entropy loss; The calculation formula of the Dice loss is as follows: In the formula, Loss Dice represents the Dice loss, N represents the total number of voxels in the image, and p k represents the predicted value of the kth voxel, g k represents the true value of the k-th voxel; The calculation formula of the centerline Dice loss is as follows: In the formula, Loss cl represents the centerline Dice loss, Tprec(S P ,V L ) represents the topological accuracy, Tsens(S L ,V P ) represents the topological sensitivity, V L and V P Represent the true segmentation mask and the predicted segmentation result respectively, S L and S P Represent the skeletons of the true segmentation mask and the predicted segmentation result respectively; The calculation formula of the cross entropy loss is as follows: Among them, Loss CE Represents the cross entropy loss, which is used to measure the difference between the predicted probability distribution and the true label distribution, p k represents the predicted value of the kth voxel, g k represents the true value of the kth voxel, and N represents the total number of voxel points in the image.

8. A cerebral vascular segmentation system based on TOF-MRA images, characterized in that: include: A data acquisition module, used for acquiring TOF-MRA image data, the TOF-MRA image data including cerebral vascular images and labels; Data preprocessing module, used to preprocess TOF-MRA image data; Network construction module, used to build image segmentation network; A training module is used to train the image segmentation network using the pre-processed TOF-MRA image data and save the optimal image segmentation network; The inference module is used to predict the cerebrovascular segmentation results using the optimal image segmentation network.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for cerebral vessel segmentation based on TOF-MRA images as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: The computer program enables a computer to execute the cerebral blood vessel segmentation method based on TOF-MRA images as described in any one of claims 1 to 7.

Images