Blood vessel wall image segmentation method, device, computer equipment and storage medium

Abstract

The present application relates to a method, device, computer equipment and storage medium for segmenting a vascular wall image. The method comprises: obtaining an intracranial vascular wall image of a three-dimensional black blood T1-weighted magnetic resonance image; performing lumen segmentation on the intracranial vascular wall image to obtain a three-dimensional vascular lumen segmentation result; then, extracting a center line according to the three-dimensional vascular lumen segmentation result, recutting perpendicularly to the center line to obtain a two-dimensional vascular wall cross-sectional image; performing segmentation on the two-dimensional vascular wall cross-sectional image to obtain a vascular lumen area and an outer wall area. The method first segments the three-dimensional lumen in the magnetic resonance image, and then segments the two-dimensional wall on the two-dimensional vascular wall cross-sectional image according to the three-dimensional lumen segmentation result. By accurately segmenting the lumen and the wall, the accuracy and efficiency of vascular wall image segmentation are improved.

Description

Technical Field

[0001] The present application relates to the field of medical imaging technology, and in particular to a method, device, computer equipment and storage medium for segmenting a vascular wall image. Background Art

[0002] Intracranial Atherosclerosis is one of the most common causes of stroke. Vascular stenosis determined by traditional angiography is still the main method for diagnosing intracranial atherosclerosis. However, it is impossible to accurately judge intracranial atherosclerosis based on vascular stenosis alone. According to studies, the degree of vascular wall stenosis and vascular wall thickness / vascular lumen area are significantly correlated with plaque symptoms, so it is of great significance to accurately segment the intracranial vascular wall.

[0003] At present, the intracranial vessel wall image segmentation scheme usually requires manual outlining of the vascular centerline, or registration with the angiography image to obtain the vascular centerline mapping. The additional manual intervention or registration operation process is cumbersome and prone to introduce errors. There is still no method for fully automated intracranial vessel wall image segmentation based only on three-dimensional black-blood T1-weighted images. Summary of the invention

[0004] Based on this, it is necessary to provide a vascular wall image segmentation method, device, computer equipment and storage medium that can improve the accuracy and efficiency of vascular wall image segmentation in order to address the above technical problems.

[0005] In a first aspect, a pipe wall image segmentation method is provided, the method comprising:

[0006] Acquire three-dimensional black-blood T1-weighted MRI images of intracranial vessel walls;

[0007] Perform lumen segmentation on the intracranial vascular wall image to obtain a three-dimensional vascular lumen segmentation result;

[0008] Extracting a center line according to the three-dimensional blood vessel lumen segmentation result, and recutting perpendicularly to the center line to obtain a two-dimensional blood vessel wall cross-sectional image;

[0009] Segmentation is performed on the two-dimensional blood vessel wall cross-sectional image to obtain the blood vessel lumen region and the outer vessel wall region.

[0010] In one embodiment, performing lumen segmentation on the intracranial vascular wall image to obtain a three-dimensional vascular lumen segmentation result includes:

[0011] The intracranial blood vessel wall image is used to generate an image patch using a cubic sliding window;

[0012] Perform lumen segmentation on the image patch to obtain a binary three-dimensional vascular lumen segmentation result;

[0013] The center line is extracted according to the vascular lumen segmentation result, including:

[0014] The centerline is extracted based on the binarized three-dimensional blood vessel lumen segmentation result.

[0015] In one embodiment, after obtaining the vascular lumen region and the outer vessel wall region, the method further includes:

[0016] The morphological information of the blood vessel is calculated according to the blood vessel lumen area and the outer vessel wall area.

[0017] In one of the embodiments, the intracranial blood vessel wall segmentation model includes a three-dimensional lumen segmentation sub-model, a centerline extraction sub-model, and a two-dimensional vessel wall segmentation sub-model;

[0018] The method further includes:

[0019] The intracranial blood vessel wall image is segmented by the three-dimensional blood vessel lumen segmentation sub-model to obtain a three-dimensional blood vessel lumen segmentation result;

[0020] The centerline is extracted according to the three-dimensional blood vessel lumen segmentation result by the centerline extraction sub-model, and a two-dimensional blood vessel wall cross-sectional image is obtained by recutting perpendicularly to the centerline;

[0021] Segmentation is performed on the two-dimensional blood vessel wall cross-sectional image by using the two-dimensional vessel wall segmentation sub-model to obtain a blood vessel lumen region and an outer vessel wall region;

[0022] The training process of the intracranial vascular wall segmentation model includes:

[0023] Acquire a three-dimensional black blood T1-weighted magnetic resonance intracranial blood vessel wall sample image and a labeled image of the intracranial blood vessel wall sample image;

[0024] The intracranial blood vessel wall sample image is input into the initial three-dimensional lumen segmentation sub-model to obtain a three-dimensional blood vessel lumen sample segmentation result;

[0025] The three-dimensional blood vessel cavity sample segmentation result is input into the centerline extraction sub-model to obtain the sample centerline, and recut perpendicularly to the sample centerline to obtain a two-dimensional blood vessel wall sample cross-sectional image;

[0026] Inputting the two-dimensional blood vessel wall sample cross-sectional image into the initial two-dimensional vessel wall segmentation sub-model to obtain a sample blood vessel lumen region and a sample outer vessel wall region;

[0027] Calculating the 3D model loss according to the 3D blood vessel lumen sample segmentation result and the marked image, and obtaining a trained 3D lumen segmentation sub-model when the 3D model segmentation loss is lower than a preset 3D model loss threshold or the number of training times reaches a preset maximum value;

[0028] The two-dimensional model loss is calculated based on the sample blood vessel cavity area, the sample outer wall area and the marked image. When the two-dimensional model segmentation loss is lower than a preset two-dimensional model loss threshold or the number of training times reaches a preset maximum value, a trained two-dimensional wall segmentation sub-model is obtained.

[0029] In one embodiment, the intracranial blood vessel wall sample image is input into an initial three-dimensional lumen segmentation sub-model to obtain a three-dimensional blood vessel lumen sample segmentation result, including:

[0030] The sample image of the intracranial blood vessel wall is used to generate a sample image patch using a cubic sliding window;

[0031] The sample image patch is input into the initial three-dimensional lumen segmentation sub-model to obtain a binary three-dimensional vascular lumen sample segmentation result.

[0032] In one embodiment, after re-cutting perpendicular to the sample centerline to obtain a two-dimensional cross-sectional image of the blood vessel wall sample, the method further includes:

[0033] The two-dimensional blood vessel wall sample cross-sectional image is data enhanced to obtain an enhanced two-dimensional blood vessel wall sample cross-sectional image; the enhanced two-dimensional blood vessel wall sample cross-sectional image is used to train an initial two-dimensional vessel wall segmentation sub-model to obtain a trained two-dimensional vessel wall segmentation model.

[0034] In one embodiment, after obtaining the sample blood vessel lumen area and the sample outer vessel wall area, the method further includes:

[0035] The sample blood vessel lumen region and the sample outer vessel wall region are used as the sample blood vessel segmentation result;

[0036] Calculate the segmentation accuracy based on the blood vessel segmentation result of the sample and the manual labeling result of the sample;

[0037] The morphological information of the sample blood vessel is calculated according to the sample blood vessel lumen area and the sample outer vessel wall area, and the segmentation consistency between the morphological information of the sample blood vessel and the gold standard is verified.

[0038] In a second aspect, a blood vessel wall image segmentation device is provided, the device comprising:

[0039] A magnetic resonance image acquisition module, used for acquiring three-dimensional black blood T1-weighted magnetic resonance images of intracranial blood vessel walls;

[0040] A three-dimensional lumen segmentation module is used to perform lumen segmentation on the intracranial vascular wall image to obtain a three-dimensional vascular lumen segmentation result;

[0041] A blood vessel centerline extraction module is used to extract the centerline according to the three-dimensional blood vessel lumen segmentation result, and re-cut perpendicularly to the centerline to obtain a two-dimensional blood vessel wall cross-sectional image;

[0042] The two-dimensional vessel wall segmentation module is used to perform segmentation on the two-dimensional vessel wall cross-sectional image to obtain a vessel lumen area and an outer vessel wall area.

[0043] In a third aspect, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0044] Acquire three-dimensional black-blood T1-weighted MRI images of intracranial vessel walls;

[0045] Perform lumen segmentation on the intracranial vascular wall image to obtain a three-dimensional vascular lumen segmentation result;

[0046] Extracting a center line according to the three-dimensional blood vessel lumen segmentation result, and recutting perpendicularly to the center line to obtain a two-dimensional blood vessel wall cross-sectional image;

[0047] Segmentation is performed on the two-dimensional blood vessel wall cross-sectional image to obtain the blood vessel lumen region and the outer vessel wall region.

[0048] In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0049] Acquire three-dimensional black-blood T1-weighted MRI images of intracranial vessel walls;

[0050] Perform lumen segmentation on the intracranial vascular wall image to obtain a three-dimensional vascular lumen segmentation result;

[0051] Extracting a center line according to the three-dimensional blood vessel lumen segmentation result, and recutting perpendicularly to the center line to obtain a two-dimensional blood vessel wall cross-sectional image;

[0052] Segmentation is performed on the two-dimensional blood vessel wall cross-sectional image to obtain the blood vessel lumen region and the outer vessel wall region.

[0053] The above-mentioned vascular wall image segmentation method, device, computer equipment and storage medium obtain an intracranial vascular wall image of a three-dimensional black blood T1-weighted magnetic resonance image; perform lumen segmentation on the intracranial vascular wall image to obtain a three-dimensional vascular lumen segmentation result; then, extract a center line based on the three-dimensional vascular lumen segmentation result, recut perpendicularly to the center line to obtain a two-dimensional vascular wall cross-sectional image; perform segmentation on the two-dimensional vascular wall cross-sectional image to obtain a vascular lumen area and an outer wall area. The method first segments the three-dimensional lumen in the magnetic resonance image, and then segments the two-dimensional wall on the two-dimensional vascular wall cross-sectional image based on the three-dimensional lumen segmentation result. By accurately segmenting the lumen and the wall, the accuracy and efficiency of vascular wall image segmentation are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1A diagram showing an application environment of a blood vessel wall image segmentation method in one embodiment;

[0055] Figure 2 is a flow chart of a blood vessel wall image segmentation method in one embodiment;

[0056] Figure 3 is a schematic diagram of a framework of a blood vessel wall image segmentation method in one embodiment;

[0057] Figure 4 is a schematic diagram of a lumen segmentation result in one embodiment;

[0058] Figure 5 is a schematic diagram of the outer tube wall segmentation result in one embodiment;

[0059] Figure 6 is a schematic diagram of the blood vessel wall image segmentation performance in one embodiment;

[0060] Figure 7 is a structural block diagram of a blood vessel wall image segmentation device in one embodiment;

[0061] Figure 8 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0063] The segmentation of vascular cavities based on MRI (Magnetic Resonance Imaging) is a very challenging task. The noise and artifacts in MR (Magnetic Resonance) images easily cause low contrast between the vessel wall and surrounding tissues. The complex vascular cavity shape structure is also a technical difficulty to be considered in segmentation. At present, traditional vascular wall image segmentation algorithms such as region growing algorithm, fuzzy connection algorithm and multi-scale filter based on Hessian matrix have certain difficulties in application in intracranial vessel wall image segmentation. The neural network-based method can segment the intracranial vascular cavity area and the vessel wall area in an end-to-end manner, which can greatly improve the accuracy and efficiency of vascular wall image segmentation. At present, some studies have developed a method for segmenting the intracranial vascular wall area on 2D cross-sectional slice images generated based on manually drawn vascular centerlines, which introduces human intervention. Another study extracted the lumen based on angiography TOF (Time of Flight) images, and then mapped the lumen segmentation and centerline extraction results to the black vascular wall image for further wall segmentation. This method requires additional registration operations between multiple image sequences, which is time-consuming, cumbersome and prone to introduce additional registration errors.

[0064] In response to the limitations and improvement needs of related technologies, the present application provides a method for intracranial vascular wall image segmentation based on three-dimensional black blood T1-weighted magnetic resonance vascular wall images, which can achieve accurate and efficient segmentation of the vascular cavity area and outer wall area of ​​the intracranial artery, thereby improving the accuracy of segmentation.

[0065] The blood vessel wall image segmentation method provided in this application can be applied to Figure 1 In the application environment shown in the figure, the terminal 102 communicates with the server 104 through the network. The terminal 102 obtains the image collected by the magnetic resonance scanning device through the network. The terminal 102 can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers and portable wearable devices, and the server 104 can be implemented by an independent server or a server cluster composed of multiple servers.

[0066] In one embodiment, Figure 2 As shown, a blood vessel wall image segmentation method is provided, and the method is applied to Figure 1 The terminal in is used as an example to illustrate, including the following steps:

[0067] Step 202: Acquire a three-dimensional black blood T1-weighted magnetic resonance image of the intracranial blood vessel wall.

[0068] Among them, the intracranial vascular wall image refers to the intracranial 3D image collected by a magnetic resonance scanning device.

[0069] Specifically, the 3T magnetic resonance scanning device uses a 3D VISTA (Volume Isotropic Turbo spin echo Acquisition, isotropic fast spin echo volume acquisition) sequence to collect intracranial images, and the terminal obtains the intracranial 3D image collected by the magnetic resonance scanning device. VISTA is a typical black-blood magnetic resonance imaging sequence (Black-Blood MRImaging), which is an effective method for measuring the thickness and area of ​​vascular walls and determining intracranial vascular information. As an effective black-blood magnetic resonance imaging sequence, 3D VSITA not only provides high isotropic spatial resolution, but also has an excellent inhibitory effect on blood flow, can accurately measure the thickness of intracranial vascular walls, and can quickly image. For example, a 3.0T magnetic resonance machine can be used within 7 minutes to obtain intracranial images with a wide coverage range, which can provide plaque loads that cannot be diagnosed by traditional angiography, and provide effective parameters for determining vascular information. The intracranial image results scanned by the 3T magnetic resonance scanning device are layers of brain slices, that is, MR images. A series of original MR images constitute the intracranial 3D image, and the terminal obtains a series of original MR images.

[0070] Step 204: perform lumen segmentation on the intracranial vascular wall image to obtain a three-dimensional vascular lumen segmentation result.

[0071] Specifically, the terminal inputs a series of acquired original MR images into the 3D lumen segmentation sub-model of the trained intracranial vascular wall segmentation model, extracts lumen features from a series of original MR images, and performs lumen segmentation based on the lumen features to obtain 3D lumen segmentation results. The intracranial vascular wall segmentation model refers to a model that first segments the vascular lumen based on the intracranial vascular wall image, then extracts the center line, and then segments the vascular lumen area and the outer wall area on a 2D vascular wall cross-sectional image perpendicular to the center line.

[0072] Step 206: extract a center line based on the three-dimensional blood vessel lumen segmentation result, and re-cut perpendicularly to the center line to obtain a two-dimensional blood vessel wall cross-sectional image.

[0073] Among them, re-cutting refers to cutting perpendicular to the center line of the blood vessel to obtain a set of two-dimensional blood vessel wall cross-sectional images.

[0074] Specifically, the centerline extraction submodel of the intracranial vascular wall segmentation model integrates the automatic skeletonization method to extract the vascular centerline. Other centerline extraction algorithms can also be used to extract the vascular centerline, such as interactive detection algorithms, matching detection algorithms, etc., which are not limited in this embodiment. After obtaining the vascular centerline, recutting is performed perpendicular to the centerline to obtain a set of two-dimensional lumen cross-sectional images perpendicular to the vascular centerline.

[0075] Step 208: segment the two-dimensional blood vessel wall cross-sectional image to obtain a blood vessel lumen region and an outer vessel wall region.

[0076] Specifically, the two-dimensional vessel wall segmentation sub-model of the intracranial vessel wall segmentation model further segments the two-dimensional vessel lumen cross-sectional image to obtain a vessel lumen region and an outer vessel wall region in the intracranial vessel wall image.

[0077] The segmented vascular lumen area and outer wall area are taken as the segmentation results. After obtaining the segmentation results, the conditional random field (CRF) method is used to post-process the segmentation results to remove the isolated areas with smaller areas in the segmentation result image, optimize the rough and uncertain marks in the segmented image, correct the fragmented misclassified areas, and obtain more detailed segmentation boundaries, thereby obtaining more accurate vascular lumen area and outer wall area and improving the accuracy of wall segmentation.

[0078] In the above-mentioned vascular wall image segmentation method, an intracranial vascular wall image of a three-dimensional black blood T1-weighted magnetic resonance is obtained; the intracranial vascular wall image is segmented to obtain a three-dimensional vascular lumen segmentation result; then, a center line is extracted according to the three-dimensional vascular lumen segmentation result, and a two-dimensional vascular wall cross-sectional image is obtained by recutting perpendicular to the center line; segmentation is performed on the two-dimensional vascular wall cross-sectional image to obtain a vascular lumen area and an outer wall area. This method first segments the three-dimensional lumen in the magnetic resonance image, and then segments the two-dimensional wall on the two-dimensional vascular wall cross-sectional image according to the three-dimensional lumen segmentation result. By accurately segmenting the lumen and the wall, the accuracy and efficiency of vascular wall image segmentation are improved.

[0079] In an optional embodiment, performing lumen segmentation on the intracranial vascular wall image to obtain a three-dimensional vascular lumen segmentation result includes:

[0080] The intracranial blood vessel wall image is used to generate an image patch using a cubic sliding window;

[0081] The image patch is input into the trained intracranial vascular wall segmentation model to obtain a binary three-dimensional vascular lumen segmentation result;

[0082] The center line is extracted according to the vascular lumen segmentation result, including:

[0083] The centerline is extracted based on the binarized three-dimensional blood vessel lumen segmentation result.

[0084] The three-dimensional lumen segmentation result is a binary segmentation result. In a possible implementation, the lumen is marked as a first mark in the original MR image, and the background area is marked as a second mark, so that the lumen is segmented from the MR image. The first mark and the second mark are different. For example, the first mark is 1 and the second mark is 0. Then, the lumen is marked as 1 in the original MR image, and the background area is marked as 0, so that the lumen is segmented from the image.

[0085] The cubic sliding window is a three-dimensional sliding window used to divide the intracranial vascular wall image into multiple image blocks of the same size as the cubic sliding window. Image patches refer to multiple image blocks obtained after division using the cubic sliding window.

[0086] Specifically, a cubic sliding window of a preset size is used to divide the intracranial 3D image into multiple image patches, and the multiple image patches are input into the three-dimensional lumen segmentation submodel of the trained intracranial vascular wall segmentation model. The lumen is marked as 1 in the image patch, and the background area is marked as 0, so that the lumen is segmented from the image patch. The lumen results output by the multiple image patches together constitute a complete binary three-dimensional vascular lumen segmentation result. Afterwards, the vascular centerline is extracted based on the binary three-dimensional vascular lumen segmentation result through the automatic skeletonization method of the centerline extraction submodel. The size of the cubic sliding window can be determined according to the size of the intracranial 3D image, and can be 128*128*128 or other sizes, which is not limited in this embodiment.

[0087] The original intracranial vascular wall image is a 3D stereo image, which occupies a large memory space. It takes a long time to obtain the 3D lumen segmentation result by directly inputting the 3D lumen segmentation sub-model of the trained intracranial vascular wall segmentation model.

[0088] In this embodiment, by dividing the intracranial vascular wall image into multiple image patches, and then inputting the multiple image patches into the three-dimensional lumen segmentation sub-model of the trained intracranial vascular wall segmentation model, a binary three-dimensional vascular lumen segmentation result is obtained, which can accelerate the lumen segmentation speed and thus improve the vascular wall segmentation efficiency.

[0089] In an optional embodiment, after obtaining the vascular lumen region and the outer vessel wall region, the method further includes:

[0090] The morphological information of the blood vessel is calculated according to the blood vessel lumen area and the outer vessel wall area.

[0091] Among them, the vascular information includes lumen area, wall area, wall thickness and standardized wall index.

[0092] Specifically, the two-dimensional vessel wall segmentation sub-model can also calculate the vessel lumen area according to the vessel lumen area; calculate the outer wall area according to the outer wall area, and subtract the outer wall area from the vessel lumen area to obtain the wall area; subtract the outer wall radius from the lumen radius to obtain the wall thickness; determine the standardized wall index according to the lumen area and the wall area, and the specific determination formula is as follows: standardized wall index = 100% * wall area / (wall area + lumen area). The morphological information of the blood vessel can be used to evaluate the accuracy of the segmentation result.

[0093] In an optional embodiment, the intracranial blood vessel wall segmentation model includes a three-dimensional lumen segmentation sub-model, a centerline extraction sub-model and a two-dimensional vessel wall segmentation sub-model;

[0094] The method further includes:

[0095] The intracranial blood vessel wall image is segmented by the three-dimensional blood vessel lumen segmentation sub-model to obtain a three-dimensional blood vessel lumen segmentation result;

[0096] The centerline is extracted according to the three-dimensional blood vessel lumen segmentation result by the centerline extraction sub-model, and a two-dimensional blood vessel wall cross-sectional image is obtained by recutting perpendicularly to the centerline;

[0097] Segmentation is performed on the two-dimensional blood vessel wall cross-sectional image by using the two-dimensional vessel wall segmentation sub-model to obtain a blood vessel lumen region and an outer vessel wall region;

[0098] The training process of the intracranial vascular wall segmentation model includes:

[0099] Acquire a three-dimensional black blood T1-weighted magnetic resonance intracranial blood vessel wall sample image and a labeled image of the intracranial blood vessel wall sample image;

[0100] The intracranial blood vessel wall sample image is input into the initial three-dimensional lumen segmentation sub-model to obtain a three-dimensional blood vessel lumen sample segmentation result;

[0101] The three-dimensional blood vessel cavity sample segmentation result is input into the centerline extraction sub-model to obtain the sample centerline, and recut perpendicularly to the sample centerline to obtain a two-dimensional blood vessel wall sample cross-sectional image;

[0102] Inputting the two-dimensional blood vessel wall sample cross-sectional image into the initial two-dimensional vessel wall segmentation sub-model to obtain a sample blood vessel lumen region and a sample outer vessel wall region;

[0103] Calculating the 3D model loss according to the 3D blood vessel lumen sample segmentation result and the marked image, and obtaining a trained 3D lumen segmentation sub-model when the 3D model segmentation loss is lower than a preset 3D model loss threshold or the number of training times reaches a preset maximum value;

[0104] The two-dimensional model loss is calculated based on the sample blood vessel cavity area, the sample outer wall area and the marked image. When the two-dimensional model segmentation loss is lower than a preset two-dimensional model loss threshold or the number of training times reaches a preset maximum value, a trained two-dimensional wall segmentation sub-model is obtained.

[0105] Specifically, Figure 3 As shown, the intracranial vascular wall segmentation model includes a 3D lumen segmentation sub-model, a centerline extraction sub-model and a 2D wall segmentation sub-model. The 3D lumen segmentation sub-model is used to perform lumen segmentation on the intracranial vascular wall image to obtain a 3D vascular lumen segmentation result; the centerline is extracted from the 3D vascular lumen segmentation result by the centerline extraction sub-model, and a 2D vascular wall cross-sectional image is obtained by recutting perpendicular to the centerline; the 2D wall segmentation sub-model is used to perform segmentation on the 2D vascular wall cross-sectional image to obtain a vascular lumen area and an outer wall area.

[0106] The 3D lumen segmentation sub-model is a 3D UNet model. The 3D UNet model has an encoding path and a decoding path, each of which has 4 resolution levels. Each layer on the analysis path contains two 3×3×3 convolutional layers, each of which is followed by an instance normalization and a leaky rectified linear unit (Leaky ReLU) as an activation layer, followed by a 2×2×2 maximum pooling layer with a stride of 2 in each direction. In the synthesis path, each layer contains a 2×2×2 deconvolution layer with a stride of 2, followed by two 3×3×3 convolutional layers, and then an activation layer. The feature layer of the same resolution in the encoding path is spliced ​​with the feature layer of the decoding path through jump connections, providing multi-scale and multi-level feature information for subsequent segmentation. The loss function of the 3D lumen segmentation sub-model is defined as the sum of the cross-entropy loss and the dice loss.

[0107] Each layer is normalized using its mean and standard deviation, and these values ​​are used to update global statistics. Instance normalization can avoid gradient vanishing and gradient exploding, reduce the dependence of gradients on model parameters or their initial scales, and allow the network to be trained at a larger learning rate, thereby accelerating network convergence; since the mean and variance used in instance batch normalization during training are calculated on small batches of samples such as sample image patches, rather than on the entire dataset, there will be some small noise in the mean and variance, and the scaling process will also generate a little noise because of the use of noisy standardized values, which forces each subsequent layer of neurons not to be overly dependent on the previous neurons. Therefore, instance normalization can also be seen as a regularization method that improves the generalization ability of the network.

[0108] The 2D tube wall segmentation sub-model is a 2D Attention Unet model. The 2D Attention Unet model with soft attention is used to supervise the features of the previous level through the features of the next level to implement the attention mechanism, that is, to limit the activated part to the area to be segmented, reduce the activation value of the background to optimize the segmentation, and achieve end-to-end segmentation. As the number of iterations increases, the activated area is increasingly concentrated in the area to be segmented, which can improve the effect of tube wall segmentation. The loss function of the 2D tube wall segmentation sub-model is defined as the Dice loss.

[0109] In the training sample acquisition stage, 50 healthy volunteers were selected as group 1 and 3 patients with intracranial atherosclerosis were selected as group 2. The intracranial vascular wall sample images were obtained by using a 32-channel head coil on a 3T magnetic resonance scanner (Achieva CX, Philips Healthcare, The Netherlands), with an acquisition sequence of 3D VISTA and a spatial resolution of 0.6 mm isotropic.

[0110] The data of 50 healthy volunteers were randomly divided into a training set (80%, 40 / 50) and a test set (20%, 10 / 50), where the training set was used to train the intracranial vascular wall segmentation model and the test set was used to evaluate the intracranial vascular wall segmentation model. The lumen and outer wall boundaries of the intracranial arteries of each healthy volunteer, including the basilar artery, the middle cerebral artery M1 segment, the anterior cerebral artery A1 segment, and the posterior cerebral artery P1 segment, were manually outlined by experienced doctors. For patients, their non-stenotic segments are basically consistent with the normal walls of healthy people, so only the lumen and outer wall boundaries of the stenotic segments are outlined to further verify the performance of the segmentation network on abnormal walls.

[0111] Obtain a sample image of the intracranial vascular wall of a three-dimensional black blood T1-weighted magnetic resonance image, and obtain a labeled image of the sample image of the intracranial vascular wall obtained by outlining the lumen and outer wall boundaries of the intracranial artery by the doctor. Further, the doctor outlines each MR sample image in a series of original MR sample images to obtain a series of labeled images, and this series of two-dimensional labeled images constitutes a three-dimensional lumen surface as a three-dimensional lumen labeled image; the three-dimensional lumen labeled image is input into the initial three-dimensional lumen segmentation sub-model to obtain a three-dimensional vascular lumen sample segmentation result; the sample centerline is extracted from the three-dimensional vascular lumen sample segmentation result using the automatic skeletonization method of the centerline extraction sub-model, and the sample is recut perpendicular to the sample centerline to obtain a two-dimensional vascular wall sample cross-sectional image; the two-dimensional vascular wall sample cross-sectional image is input into the initial two-dimensional wall segmentation sub-model to obtain a sample vascular lumen area and a sample outer wall area.

[0112] The 3D vascular lumen sample segmentation results and the labeled image are input into the loss function of the 3D lumen segmentation sub-model to obtain the 3D model loss. When the 3D model segmentation loss is lower than the preset 3D model loss threshold or the number of training times reaches the preset maximum value, the trained 3D lumen segmentation sub-model is obtained.

[0113] The loss function of the two-dimensional vessel wall segmentation sub-model is input according to the sample vessel lumen area and the sample outer vessel wall area and the marked image to obtain the two-dimensional model loss. When the two-dimensional model segmentation loss is lower than the preset two-dimensional model loss threshold or the number of training times reaches the preset maximum value, a trained two-dimensional vessel wall segmentation sub-model is obtained. The trained three-dimensional lumen segmentation sub-model, the automated skeletonization method, the re-cutting method, and the trained two-dimensional vessel wall segmentation sub-model constitute a trained intracranial vessel wall segmentation model.

[0114] In this embodiment, the instance batch normalization in the 3D lumen segmentation sub-model can avoid gradient vanishing and gradient explosion, accelerate the convergence of the model, and improve the generalization of the model. The accuracy of the vessel wall segmentation is improved by the attention mechanism in the 2D vessel wall segmentation sub-model. The intracranial vessel wall segmentation model greatly reduces the workload and subjective errors of manual delineation, and the vessel wall segmentation is more accurate and robust.

[0115] In an optional embodiment, the intracranial blood vessel wall sample image is input into an initial three-dimensional lumen segmentation sub-model to obtain a three-dimensional blood vessel lumen sample segmentation result, including:

[0116] The sample image of the intracranial blood vessel wall is used to generate a sample image patch using a cubic sliding window;

[0117] The sample image patch is input into the initial three-dimensional lumen segmentation sub-model to obtain a binary three-dimensional vascular lumen sample segmentation result.

[0118] Specifically, a cubic sliding window of a preset size is used to divide the intracranial vascular wall sample image into multiple sample image patches, and the multiple sample image patches are input into the three-dimensional lumen segmentation sub-model in the initial intracranial vascular wall segmentation model to obtain a binary three-dimensional vascular lumen sample segmentation result. Then, the sample centerline is extracted according to the binary three-dimensional vascular lumen sample segmentation result, and the sample is recut perpendicular to the sample centerline to obtain a two-dimensional vascular wall sample cross-sectional image. The size of the cubic sliding window can be determined according to the size of the intracranial vascular wall sample image, which is not limited in this embodiment.

[0119] In this embodiment, by dividing the intracranial vascular wall sample image into multiple sample image patches and then inputting the multiple image patches into a trained intracranial vascular wall segmentation model, the lumen segmentation speed can be accelerated, thereby improving the stage vascular wall segmentation efficiency and speeding up the training process.

[0120] In an optional embodiment, after recutting perpendicularly to the sample centerline to obtain a two-dimensional cross-sectional image of the blood vessel wall sample, the process further includes:

[0121] The two-dimensional blood vessel wall sample cross-sectional image is data enhanced to obtain an enhanced two-dimensional blood vessel wall sample cross-sectional image; the enhanced two-dimensional blood vessel wall sample cross-sectional image is used to train an initial two-dimensional vessel wall segmentation sub-model to obtain a trained two-dimensional vessel wall segmentation model.

[0122] Specifically, the data of the two-dimensional blood vessel wall cross-sectional image is enhanced. In addition to translation, flipping and rotation, elastic deformation is used to simulate abnormal vessel walls that may appear in atherosclerosis to obtain an enhanced two-dimensional blood vessel wall cross-sectional image. The initial two-dimensional vessel wall segmentation sub-model is trained according to the enhanced two-dimensional vessel wall sample cross-sectional image until a trained two-dimensional vessel wall segmentation model is obtained.

[0123] In this embodiment, by performing data enhancement on the two-dimensional blood vessel wall cross-sectional image, a richer two-dimensional blood vessel wall cross-sectional image data set can be obtained, which effectively simulates the abnormal vessel wall appearing in the lesion and improves the accuracy and generalization of vessel wall segmentation.

[0124] In an optional embodiment, after obtaining the sample blood vessel lumen area and the sample outer vessel wall area, the method further includes:

[0125] The sample blood vessel lumen region and the sample outer vessel wall region are used as the sample blood vessel segmentation result;

[0126] Calculating the segmentation accuracy rate according to the sample blood vessel segmentation result and the sample manual labeling result;

[0127] The morphological information of the sample blood vessel is calculated according to the sample blood vessel lumen area and the sample outer vessel wall area, and the segmentation consistency between the morphological information of the sample blood vessel and the gold standard is verified.

[0128] The sample blood vessel morphological information includes information such as the sample lumen area, the sample wall area, the sample wall thickness and the sample standardized wall index.

[0129] Specifically, the two-dimensional tube wall segmentation sub-model can further calculate the sample tube lumen area, sample tube wall area, and sample tube wall thickness according to the sample blood vessel lumen area and the sample outer tube wall area, and calculate the sample standardized tube wall index according to the sample tube lumen area and the sample tube wall area. The specific calculation process is described above and will not be repeated here.

[0130] The automatic segmentation results of the intracranial 3D lumen and 2D cross-sectional wall were compared with the manual markings, and the parameters such as sensitivity, specificity, precision, Dice coefficient and intersection-over-union were calculated to quantitatively evaluate the segmentation performance. The specific formula is as follows:

[0131] Sensitivity=TP / (TP+FN)

[0132] Specificity = TN / (TN + FP)

[0133] Precision=TP / (TP+FP)

[0134] Dice=2|Pre∩GT| / (|Pre|+|GT|)

[0135] IoU=|Pre∩GT| / |Pre∪GT|

[0136] Among them, TP represents true positive examples, FN represents false negative examples, TN represents true negative examples, and FP represents false positive examples. Sensitivity represents how many samples that are actually positive examples are predicted as positive. Specificity represents the proportion of samples predicted as negative examples to all negative examples. Precision represents how many samples predicted as positive examples are actually positive. Pre in the Dice coefficient and the Intersection-Union ratio represents the prediction result, and GT represents the result of manual labeling.

[0137] Afterwards, the Pearson correlation coefficient was used to evaluate the consistency of vascular morphological measurement results, that is, vascular information (including lumen area, wall area, wall thickness, and standardized wall index) between manual and automatic two-dimensional vessel wall segmentation results, and the segmentation results were further evaluated.

[0138] like Figure 4 As shown, an example result of the 3D lumen segmentation result with manual marking as reference is shown, where the line in the lumen is the centerline of the blood vessel extracted using the automatic skeletonization algorithm. It can be observed that the blood vessel segmentation is good and complete, and the calculated dice coefficient of the 3D lumen segmentation result is 0.872.

[0139] The automatic segmentation results of normal vessel wall and stenosis vessel wall are compared with the visual results of manual labeling, such as Figure 5As shown in Figure 1, the consistency between the 2D vessel wall segmentation results and the manual labeling results for normal and stenotic arteries is shown. For the healthy group 1 dataset, the Dice coefficient of the 2D vessel wall segmentation submodel reached high segmentation Dice coefficients of 0.941 (lumen) and 0.894 (wall). For the stenotic group 2 dataset, as shown in Table 1, the Dice coefficients for lumen and vessel wall segmentation reached 0.922 and 0.896, respectively.

[0140] Table 1 Quantitative evaluation of 3D lumen and 2D vessel wall segmentation performance

[0141]

[0142] DICE, dice coefficient; Sen, sensitivity; Spe, specificity; Prec, precision; MIoU, mean intersection over union.

[0143] For vascular morphology measurements, such as Figure 6 The figure shows a scatter plot comparison of the two-dimensional lumen area, vascular wall area, and standardized wall index of the intracranial vascular wall segmentation model segmentation results and the manual labeling results. The Pearson correlation coefficient is shown in the lower right corner. The Pearson correlation coefficients of the segmentation results of the manual segmentation and the intracranial vascular wall segmentation model are 0.983 for lumen area, 0.929 for vascular wall area, and 0.906 for standardized wall index, respectively. Regression analysis shows that the probability that the difference between samples is caused by sampling error is less than 0.001, and the results are statistically significant, proving the accuracy and robustness of the intracranial vascular wall segmentation model.

[0144] It should be understood that although Figure 2 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 2 At least part of the steps may include multiple steps or multiple stages. These steps or stages are not necessarily performed at the same time, but can be performed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed in turn or alternately with other steps or at least part of the steps or stages in other steps.

[0145] In one embodiment, Figure 7 As shown, a blood vessel wall image segmentation device is provided, comprising: a magnetic resonance image acquisition module 702, a three-dimensional lumen segmentation module 704, a blood vessel centerline extraction module 706 and a two-dimensional vessel wall segmentation module 708, wherein:

[0146] The magnetic resonance image acquisition module 702 is used to acquire three-dimensional black blood T1-weighted magnetic resonance images of intracranial blood vessel walls.

[0147] The three-dimensional lumen segmentation module 704 is used to perform lumen segmentation on the intracranial vascular wall image to obtain a three-dimensional vascular lumen segmentation result.

[0148] The blood vessel centerline extraction module 706 is used to extract the centerline according to the three-dimensional blood vessel lumen segmentation result, and re-cut perpendicularly to the centerline to obtain a two-dimensional blood vessel wall cross-sectional image.

[0149] The two-dimensional vessel wall segmentation module 708 is used to perform segmentation on the two-dimensional vessel wall cross-sectional image to obtain a vessel lumen region and an outer vessel wall region.

[0150] In an optional embodiment, the three-dimensional lumen segmentation module 704 is further used to generate an image patch using a cubic sliding window on the intracranial blood vessel wall image;

[0151] and performing lumen segmentation on the image patch to obtain a binary three-dimensional vascular lumen segmentation result;

[0152] The blood vessel centerline extraction module 706 is further used to extract the centerline according to the binarized three-dimensional blood vessel lumen segmentation result.

[0153] In an optional embodiment, the blood vessel wall image segmentation device further includes a blood vessel information determination module, which is used to calculate the morphological information of the blood vessel according to the blood vessel cavity area and the outer vessel wall area.

[0154] In an optional embodiment, the intracranial blood vessel wall segmentation model includes a three-dimensional lumen segmentation sub-model, a centerline extraction sub-model and a two-dimensional vessel wall segmentation sub-model;

[0155] The three-dimensional lumen segmentation module 704 is further used to perform lumen segmentation on the intracranial vascular wall image by using the three-dimensional lumen segmentation sub-model to obtain a three-dimensional vascular lumen segmentation result;

[0156] The blood vessel centerline extraction module 706 is further used to extract the centerline according to the three-dimensional blood vessel lumen segmentation result through the centerline extraction sub-model, and re-cut perpendicularly to the centerline to obtain a two-dimensional blood vessel wall cross-sectional image;

[0157] The two-dimensional vessel wall segmentation module 708 is further used to perform segmentation on the two-dimensional vessel wall cross-sectional image by using the two-dimensional vessel wall segmentation sub-model to obtain a vessel lumen region and an outer vessel wall region;

[0158] The vascular wall image segmentation device also includes a training module for acquiring a sample image of the intracranial vascular wall of a three-dimensional black blood T1-weighted magnetic resonance image and a labeled image of the sample image of the intracranial vascular wall;

[0159] and inputting the intracranial blood vessel wall sample image into an initial three-dimensional lumen segmentation sub-model to obtain a three-dimensional blood vessel lumen sample segmentation result;

[0160] and inputting the three-dimensional blood vessel cavity sample segmentation result into the centerline extraction sub-model to obtain the sample centerline, and recutting perpendicularly to the sample centerline to obtain a two-dimensional blood vessel wall sample cross-sectional image;

[0161] and inputting the two-dimensional blood vessel wall sample cross-sectional image into the initial two-dimensional vessel wall segmentation sub-model to obtain a sample blood vessel lumen region and a sample outer vessel wall region;

[0162] and calculating the 3D model loss according to the 3D blood vessel lumen sample segmentation result and the marked image, and obtaining a trained 3D lumen segmentation sub-model when the 3D model segmentation loss is lower than a preset 3D model loss threshold or the number of training times reaches a preset maximum value;

[0163] And the two-dimensional model loss is calculated according to the sample blood vessel cavity area, the sample outer wall area and the marked image, and when the two-dimensional model segmentation loss is lower than a preset two-dimensional model loss threshold or the number of training times reaches a preset maximum value, a trained two-dimensional wall segmentation sub-model is obtained.

[0164] In an optional embodiment, the training module is further used to generate a sample image patch using a cubic sliding window from the sample image of the intracranial blood vessel wall;

[0165] And the sample image patch is input into the initial three-dimensional lumen segmentation sub-model to obtain a binary three-dimensional vascular lumen sample segmentation result.

[0166] In an optional embodiment, the training module is also used to perform data enhancement on the two-dimensional vascular wall sample cross-sectional image to obtain an enhanced two-dimensional vascular wall sample cross-sectional image; the enhanced two-dimensional vascular wall sample cross-sectional image is used to train the initial two-dimensional vessel wall segmentation sub-model to obtain a trained two-dimensional vessel wall segmentation model.

[0167] In an optional embodiment, the training module further includes a sample blood vessel information determination unit, which is used to use the sample blood vessel lumen region and the sample outer vessel wall region as the sample blood vessel segmentation result;

[0168] Calculate the segmentation accuracy based on the blood vessel segmentation result of the sample and the manual labeling result of the sample;

[0169] The morphological information of the sample blood vessel is calculated according to the sample blood vessel lumen area and the sample outer vessel wall area, and the segmentation consistency between the morphological information of the sample blood vessel and the gold standard is verified.

[0170] The specific definition of the vascular wall image segmentation device can refer to the definition of the vascular wall image segmentation method above, which will not be repeated here. Each module in the above-mentioned vascular wall image segmentation device can be implemented in whole or in part by software, hardware and their combination. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0171] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 8 As shown. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a blood vessel wall image segmentation method is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a key, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0172] Those skilled in the art will understand that Figure 8 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0173] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0174] Acquire three-dimensional black-blood T1-weighted MRI images of intracranial vessel walls;

[0175] Perform lumen segmentation on the intracranial vascular wall image to obtain a three-dimensional vascular lumen segmentation result;

[0176] Extracting a center line according to the three-dimensional blood vessel lumen segmentation result, and recutting perpendicularly to the center line to obtain a two-dimensional blood vessel wall cross-sectional image;

[0177] Segmentation is performed on the two-dimensional blood vessel wall cross-sectional image to obtain the blood vessel lumen region and the outer vessel wall region.

[0178] In one embodiment, when the processor executes the computer program, the following steps are also implemented: continuing lumen segmentation on the intracranial vascular wall image to obtain a three-dimensional vascular lumen segmentation result, including: using a cubic sliding window to generate an image patch on the intracranial vascular wall image; performing lumen segmentation on the image patch to obtain a binary three-dimensional vascular lumen segmentation result; extracting a centerline based on the vascular lumen segmentation result, including: extracting a centerline based on the binary three-dimensional vascular lumen segmentation result.

[0179] In one embodiment, when the processor executes the computer program, the following steps are also implemented: after obtaining the blood vessel lumen area and the outer tube wall area, the following steps are also included: calculating the morphological information of the blood vessel based on the blood vessel lumen area and the outer tube wall area.

[0180] In one embodiment, when the processor executes the computer program, the following steps are also implemented: the intracranial blood vessel wall segmentation model includes a three-dimensional lumen segmentation sub-model, a centerline extraction sub-model and a two-dimensional wall segmentation sub-model; the method also includes: performing lumen segmentation on the intracranial blood vessel wall image by the three-dimensional lumen segmentation sub-model to obtain a three-dimensional blood vessel lumen segmentation result; extracting a centerline according to the three-dimensional blood vessel lumen segmentation result by the centerline extraction sub-model, and recutting perpendicularly to the centerline to obtain a two-dimensional blood vessel wall cross-sectional image; performing segmentation on the two-dimensional blood vessel wall cross-sectional image by the two-dimensional wall segmentation sub-model to obtain a blood vessel lumen area and an outer wall area; the training process of the intracranial blood vessel wall segmentation model includes: obtaining a sample image of the intracranial blood vessel wall of a three-dimensional black blood T1-weighted magnetic resonance and a labeled image of the sample image of the intracranial blood vessel wall; inputting the sample image of the intracranial blood vessel wall into an initial three-dimensional lumen segmentation sub-model; The three-dimensional vascular cavity sample segmentation result is obtained by cutting the sub-model; the three-dimensional vascular cavity sample segmentation result is input into the centerline extraction sub-model to obtain the sample centerline, and re-cut perpendicularly to the sample centerline to obtain a two-dimensional vascular wall sample cross-sectional image; the two-dimensional vascular wall sample cross-sectional image is input into the initial two-dimensional tube wall segmentation sub-model to obtain a sample vascular cavity area and a sample outer tube wall area; the three-dimensional model loss is calculated according to the three-dimensional vascular cavity sample segmentation result and the marked image, and when the three-dimensional model segmentation loss is lower than the preset three-dimensional model loss threshold or the number of training times reaches the preset maximum value, a trained three-dimensional tube cavity segmentation sub-model is obtained; the two-dimensional model loss is calculated according to the sample vascular cavity area and the sample outer tube wall area and the marked image, and when the two-dimensional model segmentation loss is lower than the preset two-dimensional model loss threshold or the number of training times reaches the preset maximum value, a trained two-dimensional tube wall segmentation sub-model is obtained.

[0181] In one embodiment, when the processor executes the computer program, the following steps are also implemented: the sample image of the intracranial blood vessel wall is input into an initial three-dimensional lumen segmentation sub-model to obtain a three-dimensional blood vessel lumen sample segmentation result, including: using a cubic sliding window to generate a sample image patch from the sample image of the intracranial blood vessel wall; the sample image patch is input into the initial three-dimensional lumen segmentation sub-model to obtain a binarized three-dimensional blood vessel lumen sample segmentation result.

[0182] In one embodiment, the processor further implements the following steps when executing the computer program: after recutting perpendicular to the sample centerline to obtain a two-dimensional blood vessel wall sample cross-sectional image, it also includes: performing data enhancement on the two-dimensional blood vessel wall sample cross-sectional image to obtain an enhanced two-dimensional blood vessel wall sample cross-sectional image; the enhanced two-dimensional blood vessel wall sample cross-sectional image is used to train an initial two-dimensional vessel wall segmentation sub-model to obtain a trained two-dimensional vessel wall segmentation model.

[0183] In one embodiment, the processor further implements the following steps when executing the computer program: after obtaining the sample blood vessel lumen area and the sample outer wall area, it also includes: taking the sample blood vessel lumen area and the sample outer wall area as the sample blood vessel segmentation result; calculating the segmentation accuracy based on the sample blood vessel segmentation result and the manual marking result of the sample; calculating the morphological information of the sample blood vessel based on the sample blood vessel lumen area and the sample outer wall area, and verifying the segmentation consistency of the morphological information of the sample blood vessel with the gold standard.

[0184] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0185] Acquire three-dimensional black-blood T1-weighted MRI images of intracranial vessel walls;

[0186] Perform lumen segmentation on the intracranial vascular wall image to obtain a three-dimensional vascular lumen segmentation result;

[0187] Extracting a center line according to the three-dimensional blood vessel lumen segmentation result, and recutting perpendicularly to the center line to obtain a two-dimensional blood vessel wall cross-sectional image;

[0188] Segmentation is performed on the two-dimensional blood vessel wall cross-sectional image to obtain the blood vessel lumen region and the outer vessel wall region.

[0189] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: performing lumen segmentation on the intracranial vascular wall image to obtain a three-dimensional vascular lumen segmentation result, including: using a cubic sliding window to generate an image patch on the intracranial vascular wall image; inputting the image patch into a trained intracranial vascular wall segmentation model to obtain a binarized three-dimensional vascular lumen segmentation result; extracting a centerline based on the vascular lumen segmentation result, including: extracting a centerline based on the binarized three-dimensional vascular lumen segmentation result.

[0190] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented: after obtaining the blood vessel lumen area and the outer vessel wall area, the following steps are also included: calculating the morphological information of the blood vessel based on the blood vessel lumen area and the outer vessel wall area.

[0191] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: the intracranial blood vessel wall segmentation model includes a three-dimensional lumen segmentation sub-model, a centerline extraction sub-model and a two-dimensional wall segmentation sub-model; the method also includes: performing lumen segmentation on the intracranial blood vessel wall image by the three-dimensional lumen segmentation sub-model to obtain a three-dimensional blood vessel lumen segmentation result; extracting a centerline according to the three-dimensional blood vessel lumen segmentation result by the centerline extraction sub-model, and recutting perpendicularly to the centerline to obtain a two-dimensional blood vessel wall cross-sectional image; performing segmentation on the two-dimensional blood vessel wall cross-sectional image by the two-dimensional wall segmentation sub-model to obtain a blood vessel lumen area and an outer wall area; the training process of the intracranial blood vessel wall segmentation model includes: obtaining a sample image of the intracranial blood vessel wall of a three-dimensional black blood T1-weighted magnetic resonance and a labeled image of the sample image of the intracranial blood vessel wall; inputting the sample image of the intracranial blood vessel wall into an initial three-dimensional lumen The segmentation sub-model is used to obtain a three-dimensional vascular cavity sample segmentation result; the three-dimensional vascular cavity sample segmentation result is input into the centerline extraction sub-model to obtain a sample centerline, and recut perpendicularly to the sample centerline to obtain a two-dimensional vascular wall sample cross-sectional image; the two-dimensional vascular wall sample cross-sectional image is input into the initial two-dimensional tube wall segmentation sub-model to obtain a sample vascular cavity area and a sample outer tube wall area; the three-dimensional model loss is calculated according to the three-dimensional vascular cavity sample segmentation result and the marked image, and when the three-dimensional model segmentation loss is lower than a preset three-dimensional model loss threshold or the number of training times reaches a preset maximum value, a trained three-dimensional tube cavity segmentation sub-model is obtained; the two-dimensional model loss is calculated according to the sample vascular cavity area and the sample outer tube wall area and the marked image, and when the two-dimensional model segmentation loss is lower than a preset two-dimensional model loss threshold or the number of training times reaches a preset maximum value, a trained two-dimensional tube wall segmentation sub-model is obtained.

[0192] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: the sample image of the intracranial blood vessel wall is input into an initial three-dimensional lumen segmentation sub-model to obtain a three-dimensional blood vessel lumen sample segmentation result, including: using a cubic sliding window to generate a sample image patch from the sample image of the intracranial blood vessel wall; the sample image patch is input into the initial three-dimensional lumen segmentation sub-model to obtain a binarized three-dimensional blood vessel lumen sample segmentation result.

[0193] In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: after recutting perpendicular to the sample centerline to obtain a two-dimensional blood vessel wall sample cross-sectional image, the following steps are also included: data enhancement is performed on the two-dimensional blood vessel wall sample cross-sectional image to obtain an enhanced two-dimensional blood vessel wall sample cross-sectional image; the enhanced two-dimensional blood vessel wall sample cross-sectional image is used to train the initial two-dimensional vessel wall segmentation sub-model to obtain a trained two-dimensional vessel wall segmentation model. In one embodiment, when the computer program is executed by the processor, the following steps are also implemented: after obtaining the sample vessel lumen area and the sample outer vessel wall area, the following steps are also included: taking the sample vessel lumen area and the sample outer vessel wall area as the sample vessel segmentation result; calculating the segmentation accuracy based on the sample vessel segmentation result and the manual labeling result of the sample; calculating the sample vessel morphological information based on the sample vessel lumen area and the sample outer vessel wall area, and verifying the segmentation consistency of the sample vessel morphological information with the gold standard.

[0194] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0195] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0196] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A blood vessel wall image segmentation method, characterized in that: The method comprises: Acquire three-dimensional black-blood T1-weighted MRI images of intracranial vessel walls; Using a cubic sliding window to generate an image patch for the intracranial blood vessel wall image; The image patch is input into a 3D lumen segmentation sub-model of a trained intracranial vascular wall segmentation model, lumen features in the image patch are extracted, and lumen segmentation is performed according to the lumen features to obtain a binary 3D vascular lumen segmentation result; a centerline is extracted according to the binary 3D vascular lumen segmentation result, and a 2D vascular wall cross-sectional image is obtained by recutting perpendicular to the centerline; the centerline extraction sub-model of the intracranial vascular wall segmentation model incorporates an automatic skeletonization method to extract the vascular centerline; Segmentation is performed on the two-dimensional blood vessel wall cross-sectional image to obtain a blood vessel lumen region and an outer vessel wall region.

2. The method according to claim 1, characterized in that After obtaining the vascular cavity region and the outer tube wall region, the method further includes: The morphological information of the blood vessel is calculated according to the blood vessel lumen area and the outer vessel wall area.

3. The method according to claim 1, characterized in that The intracranial vascular wall segmentation model includes a 3D lumen segmentation sub-model, a centerline extraction sub-model and a 2D vessel wall segmentation sub-model. The method further comprises: Performing lumen segmentation on the intracranial vascular wall image using the three-dimensional lumen segmentation sub-model to obtain a three-dimensional vascular lumen segmentation result; Extracting a centerline according to the three-dimensional blood vessel lumen segmentation result by using the centerline extraction sub-model, and recutting perpendicularly to the centerline to obtain a two-dimensional blood vessel wall cross-sectional image; Segmentation is performed on the two-dimensional blood vessel wall cross-sectional image by using the two-dimensional vessel wall segmentation sub-model to obtain a blood vessel lumen region and an outer vessel wall region; The training process of the intracranial vascular wall segmentation model includes: Acquire a three-dimensional black blood T1-weighted magnetic resonance intracranial blood vessel wall sample image and a labeled image of the intracranial blood vessel wall sample image; Inputting the intracranial blood vessel wall sample image into an initial three-dimensional lumen segmentation sub-model to obtain a three-dimensional blood vessel lumen sample segmentation result; Input the three-dimensional blood vessel cavity sample segmentation result into the centerline extraction sub-model to obtain the sample centerline, and re-cut perpendicularly to the sample centerline to obtain a two-dimensional blood vessel wall sample cross-sectional image; Inputting the two-dimensional blood vessel wall sample cross-sectional image into the initial two-dimensional vessel wall segmentation sub-model to obtain a sample blood vessel lumen region and a sample outer vessel wall region; Calculating the 3D model loss according to the 3D blood vessel lumen sample segmentation result and the marked image, and obtaining a trained 3D lumen segmentation sub-model when the 3D model segmentation loss is lower than a preset 3D model loss threshold or the number of training times reaches a preset maximum value; The two-dimensional model loss is calculated based on the sample blood vessel cavity area, the sample outer wall area and the marked image, and when the two-dimensional model segmentation loss is lower than a preset two-dimensional model loss threshold or the number of training times reaches a preset maximum value, a trained two-dimensional wall segmentation model is obtained.

4. The method according to claim 3, characterized in that The step of inputting the intracranial blood vessel wall sample image into the initial three-dimensional lumen segmentation sub-model to obtain a three-dimensional blood vessel lumen sample segmentation result includes: Using a cubic sliding window to generate a sample image patch from the intracranial blood vessel wall sample image; The sample image patch is input into the initial three-dimensional lumen segmentation sub-model to obtain a binarized three-dimensional vascular lumen sample segmentation result.

5. The method according to claim 3, characterized in that: After recutting perpendicularly to the sample centerline to obtain a two-dimensional cross-sectional image of the blood vessel wall sample, the method further includes: The two-dimensional vascular wall sample cross-sectional image is data enhanced to obtain an enhanced two-dimensional vascular wall sample cross-sectional image; the enhanced two-dimensional vascular wall sample cross-sectional image is used to train the initial two-dimensional vessel wall segmentation sub-model to obtain a trained two-dimensional vessel wall segmentation model.

6. The method according to claim 3, characterized in that After obtaining the sample blood vessel cavity area and the sample outer tube wall area, the method further includes: The sample blood vessel lumen region and the sample outer vessel wall region are used as the sample blood vessel segmentation result; Calculating the segmentation accuracy rate according to the sample blood vessel segmentation result and the manual marking result of the sample blood vessel; The morphological information of the sample blood vessel is calculated according to the sample blood vessel lumen area and the sample outer vessel wall area, and the segmentation consistency between the morphological information of the sample blood vessel and the gold standard is verified.

7. A blood vessel wall image segmentation device, characterized in that: The device comprises: A magnetic resonance image acquisition module, used for acquiring three-dimensional black blood T1-weighted magnetic resonance images of intracranial blood vessel walls; A three-dimensional lumen segmentation module is used to generate an image patch from the intracranial vascular wall image using a cubic sliding window; the image patch is input into a three-dimensional lumen segmentation sub-model of a trained intracranial vascular wall segmentation model, lumen features in the image patch are extracted, and lumen segmentation is performed according to the lumen features to obtain a binary three-dimensional vascular lumen segmentation result; A blood vessel centerline extraction module is used to extract the centerline according to the binarized three-dimensional blood vessel lumen segmentation result, and re-cut perpendicularly to the centerline to obtain a two-dimensional blood vessel wall cross-sectional image; the centerline extraction sub-model of the intracranial blood vessel wall segmentation model incorporates an automatic skeletonization method to extract the blood vessel centerline; The two-dimensional vessel wall segmentation module is used to perform segmentation on the two-dimensional vessel wall cross-sectional image to obtain a vessel cavity region and an outer vessel wall region.

8. The device according to claim 7, characterized in that The device also includes a blood vessel information determination module, which is used to: The morphological information of the blood vessel is calculated according to the blood vessel lumen area and the outer vessel wall area.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images