Method for training a segmentation model for brain vessel segmentation and related products

Abstract

The application discloses a method for training a segmentation model for cerebral vascular segmentation and related products. The method comprises: obtaining a three-dimensional gray image and a corresponding three-dimensional blood vessel segmentation mask based on a three-dimensional angiogram; generating a three-dimensional gray entropy feature map according to the three-dimensional gray image; generating a three-dimensional label entropy map according to the three-dimensional blood vessel segmentation mask; performing feature fusion on the three-dimensional gray image and the three-dimensional gray entropy feature map to obtain fused features; inputting the fused features into the segmentation model for blood vessel segmentation and calculating a weighted loss based on the three-dimensional label entropy map; and training the segmentation model for cerebral vascular segmentation based on the weighted loss. The scheme of the application can improve the recognition ability of the model for blood vessel boundaries, small blood vessels and complex structure regions, effectively reduce the blood vessel adhesion phenomenon, and enhance the robustness and clinical applicability of the segmentation result.

Description

Technical Field

[0001] This application generally relates to the field of medical image segmentation technology. More specifically, this application relates to a method for training a segmentation model for cerebral blood vessel segmentation, a method for cerebral blood vessel segmentation, an electronic device, and a computer-readable storage medium. Background Technology

[0002] Digital subtraction angiography (DSA) vascular segmentation is a key technology in the diagnosis and treatment of cerebrovascular diseases. Its core value lies in providing physicians with precise spatial structural information of cerebral blood vessels, supporting crucial clinical aspects such as disease diagnosis, treatment planning, and postoperative follow-up. Accurate cerebral vascular segmentation results help physicians clearly identify the course of blood vessels, branching relationships, and locations of stenosis or malformation, serving as an important imaging basis for the diagnosis and treatment of diseases such as intracranial aneurysms and cerebral infarction.

[0003] Existing methods for vascular segmentation are mainly based on traditional image processing techniques (such as thresholding, morphological manipulation, and region growing) and conventional deep learning techniques (such as traditional convolutional neural networks). However, these methods face significant technical bottlenecks in practical clinical applications: On the one hand, due to noise interference, uneven distribution of contrast agents, and weak visualization of small blood vessels in DSA images, vascular boundaries are often blurred, leading to adhesion between different vascular segments during segmentation, making it difficult to accurately distinguish adjacent or overlapping vascular structures; on the other hand, existing methods mostly rely on pixel- or voxel-level local features, and do not fully utilize prior knowledge such as the complexity of vascular boundaries and structural uncertainties. In the segmentation of complex cerebral vascular networks, they are prone to problems such as decreased segmentation accuracy and poor structural consistency, failing to meet the clinical demand for precise segmentation.

[0004] In view of this, this application provides a scheme for training a segmentation model for cerebral blood vessel segmentation. By combining label entropy weighted loss with gray-level entropy prior input, the model's ability to identify blood vessel boundaries, small blood vessels and structurally complex regions is improved, effectively reducing blood vessel adhesion and enhancing the robustness and clinical applicability of the segmentation results. Summary of the Invention

[0005] In order to at least address one or more of the technical problems mentioned above, this application proposes a scheme for training a segmentation model for cerebral blood vessel segmentation in several aspects.

[0006] In a first aspect, this application provides a method for training a segmentation model for cerebral blood vessel segmentation, the method comprising: acquiring a three-dimensional grayscale image and a corresponding three-dimensional blood vessel segmentation mask based on a three-dimensional angiography image; generating a three-dimensional grayscale entropy feature map based on the three-dimensional grayscale image; generating a three-dimensional label entropy map based on the three-dimensional blood vessel segmentation mask; performing feature fusion on the three-dimensional grayscale image and the three-dimensional grayscale entropy feature map to obtain fused features; inputting the fused features into the segmentation model for blood vessel segmentation, and calculating a weighted loss based on the three-dimensional label entropy map; and training the segmentation model for cerebral blood vessel segmentation based on the weighted loss.

[0007] In some embodiments, the three-dimensional grayscale image is obtained by performing denoising processing on the three-dimensional angiography image to obtain the three-dimensional grayscale image, wherein the denoising processing includes Gaussian filtering and / or nonlocal mean filtering.

[0008] In some embodiments, generating a three-dimensional grayscale entropy feature map based on the three-dimensional grayscale image includes: calculating the grayscale distribution within a target neighborhood centered on each voxel in the three-dimensional grayscale image; and calculating Shannon entropy based on the grayscale distribution to obtain the three-dimensional grayscale entropy feature map.

[0009] In some embodiments, the Shannon entropy is calculated based on the grayscale distribution to obtain the three-dimensional grayscale entropy feature map using the following formula: ,in, Indicated by voxels The grayscale distribution within the target neighborhood centered on the target. Represents a minimal constant. This indicates the grayscale level.

[0010] In some embodiments, generating a three-dimensional label entropy map based on the three-dimensional vessel segmentation mask includes: calculating the ratio of foreground voxels and background voxels in the target neighborhood centered on each voxel in the three-dimensional vessel segmentation mask; and calculating Shannon entropy based on the ratio values ​​to obtain the three-dimensional label entropy map.

[0011] In some embodiments, the Shannon entropy is calculated based on the scale value using the following formula to obtain the three-dimensional label entropy map: ,in, This represents the scale value of the foreground voxels. The scale value of the background voxels, P0 = 1 - P1, This represents a very small constant.

[0012] In some embodiments, feature fusion of the three-dimensional grayscale image and the three-dimensional grayscale entropy feature map to obtain fused features includes: using a feature network to perform feature fusion of the three-dimensional grayscale image and the three-dimensional grayscale entropy feature map to obtain the fused features; or performing feature concatenation of the three-dimensional grayscale image and the three-dimensional grayscale entropy feature map to obtain the fused features.

[0013] In some embodiments, the feature network includes a first feature extraction unit, a second feature extraction unit, and a feature fusion unit. The feature network is used to fuse the three-dimensional grayscale image and the three-dimensional grayscale entropy feature map to obtain the fused features. This includes: using the first feature extraction unit and the second feature extraction unit to extract features from the three-dimensional grayscale image and the three-dimensional grayscale entropy feature map respectively, obtaining corresponding first features and second features; and using the feature fusion unit to fuse the first features and the second features to obtain the fused features.

[0014] In some embodiments, the weighted loss is calculated using the following formula: Where L represents the weighted loss, This represents the tag entropy value corresponding to the i-th voxel in the three-dimensional tag entropy map. Indicates predicted value and the true value The cross-entropy, where i represents the voxel index and N represents the total number of voxels.

[0015] In a second aspect, this application provides a method for cerebral blood vessel segmentation, comprising: acquiring a three-dimensional angiography image to be segmented; inputting the three-dimensional angiography image into a segmentation model trained according to the method described in the first aspect for cerebral blood vessel segmentation, and obtaining cerebral blood vessel segmentation results.

[0016] In some embodiments, the method further includes: performing post-processing operations on the cerebral blood vessel segmentation results to obtain the final cerebral blood vessel segmentation results, wherein the post-processing operations include boundary smoothing and / or isolated region removal.

[0017] In a third aspect, this application provides an electronic device, comprising: a processor; and a memory having computer instructions stored thereon for training a segmentation model for cerebral blood vessel segmentation, wherein when the computer instructions are executed by the processor, the method described in the first aspect is implemented; or having computer instructions stored thereon for cerebral blood vessel segmentation, wherein when the computer instructions are executed by the processor, the method described in the second aspect is implemented.

[0018] In a fourth aspect, this application provides a computer-readable storage medium having stored thereon computer program instructions for training a segmentation model for cerebral blood vessel segmentation, wherein when executed by one or more processors, the computer program instructions cause the method described in the first aspect above to be implemented; or having stored thereon computer program instructions for cerebral blood vessel segmentation, wherein when executed by one or more processors, the computer program instructions cause the method described in the second aspect above to be implemented.

[0019] By employing the training scheme for the segmentation model used for cerebral blood vessel segmentation provided above, this embodiment of the application uses a label entropy weighted loss mechanism to enable the model to automatically focus on blood vessel boundaries and structurally complex regions. Combined with gray-level entropy prior input to guide the network to perceive the features of difficult-to-segment regions, it significantly improves the boundary accuracy and structural consistency of blood vessel segmentation, effectively reduces blood vessel adhesion, enhances the robustness of the model to complex cerebral blood vessel networks, and provides reliable image support for the diagnosis and treatment of cerebrovascular diseases. Attached Figure Description

[0020] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of this application are illustrated by way of example and not limitation, and the same or corresponding reference numerals denote the same or corresponding parts, wherein: Figure 1 This is an exemplary flowchart illustrating a method 100 for training a segmentation model for cerebral blood vessel segmentation according to an embodiment of this application; Figure 2 This is another exemplary flowchart illustrating the training of a segmentation model for cerebral blood vessel segmentation according to an embodiment of this application; Figure 3 This is yet another exemplary flowchart illustrating the training of a segmentation model for cerebral blood vessel segmentation according to an embodiment of this application; Figure 4 This is an exemplary flowchart illustrating a brain vessel segmentation method according to an embodiment of this application; Figure 5 An exemplary structural block diagram of an electronic device 500 according to an embodiment of this application is shown. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] It should be understood that the terms "comprising" and "including" used in the specification and claims of this application indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0023] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this specification and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this specification and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0024] As used in this specification and claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0025] The specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0026] Figure 1 This is an exemplary flowchart illustrating a method 100 for training a segmentation model for cerebral blood vessel segmentation according to an embodiment of this application.

[0027] like Figure 1 As shown, the method 100 may include: step S101: obtaining a three-dimensional grayscale image and a corresponding three-dimensional blood vessel segmentation mask based on a three-dimensional angiography image; step S102: generating a three-dimensional grayscale entropy feature map based on the three-dimensional grayscale image; step S103: generating a three-dimensional label entropy map based on the three-dimensional blood vessel segmentation mask; step S104: performing feature fusion on the three-dimensional grayscale image and the three-dimensional grayscale entropy feature map to obtain fused features; step S105: inputting the fused features into a segmentation model for blood vessel segmentation, and calculating a weighted loss based on the three-dimensional label entropy map; and step S106: training the segmentation model for cerebral blood vessel segmentation based on the weighted loss.

[0028] First, in step S101, a 3D grayscale image and the corresponding 3D vascular segmentation mask are acquired based on the 3D angiography image. In some implementation scenarios, 3D angiography images of patients with cerebrovascular diseases can be acquired using clinical DSA equipment. The image format can be .nii.gz, which is the standard format in the medical imaging field, supporting 3D data storage and fast reading. During acquisition, it is necessary to ensure that the image covers the complete cerebral vascular network, including major branches such as the Circle of Willis, the anterior cerebral artery, and the middle cerebral artery, as well as small vessels, to avoid insufficient model learning due to incomplete field of view.

[0029] In other implementation scenarios, a combination of manual and semi-automatic annotation can be used to generate the corresponding 3D blood vessel segmentation mask. First, a radiologist with clinical experience performs preliminary blood vessel region annotation on the 3D grayscale image. Then, medical image annotation tools (such as 3D Slicer) are used for fine-tuning to ensure accuracy. Blood vessel boundaries are aligned with the actual blood vessel edges in the image, small blood vessels (diameter ≥ 0.5mm) are not missed, and background areas are accurately annotated. The final generated 3D blood vessel segmentation mask is a binary matrix with the same spatial dimensions as the 3D grayscale image, where a voxel value of 0 represents the background (such as bone, soft tissue, or air), and 1 represents the blood vessel foreground.

[0030] Next, in step S102, a three-dimensional grayscale entropy feature map is generated based on the three-dimensional grayscale image. In some embodiments, the three-dimensional grayscale image can be obtained by performing denoising processing on the three-dimensional angiography image to obtain the three-dimensional grayscale image, wherein the denoising processing includes Gaussian filtering and / or nonlocal mean filtering.

[0031] Specifically, the Gaussian kernel size can be chosen, for example, a 3×3×3 three-dimensional neighborhood window, balancing noise reduction and detail preservation. The standard deviation σ can be set to 0.8~1.2. A weighted average of the neighborhood gray values ​​of each voxel is applied through convolution operations to smooth noise while preserving as much detail as possible at the vessel edges. The local search window size can be 7×7×7, the similarity window size can be 3×3×3, and the attenuation factor h can be set to 10~15. This method, by finding a group of voxels similar to the target voxel in the image and applying a weighted average to their gray values, can better preserve the fine branching structure of blood vessels while removing noise, avoiding the detail blurring caused by traditional filtering.

[0032] Based on the generated 3D grayscale image, in some embodiments, the Shannon entropy can be calculated by calculating the grayscale distribution in the target neighborhood centered on each voxel in the 3D grayscale image, and then calculating the Shannon entropy based on the grayscale distribution to obtain a 3D grayscale entropy feature map.

[0033] In some implementation scenarios, the target neighborhood can be a 3×3×3 three-dimensional neighborhood window set for each voxel p (coordinates (x,y,z)) in the 3D grayscale image. This can accurately capture the local grayscale distribution while ensuring computational efficiency. Next, the grayscale values ​​of all voxels within the target neighborhood window are counted to determine the grayscale level range (0~255). For each grayscale level i (i∈[0,255]), its frequency of occurrence within the neighborhood window is calculated, and then the frequency is divided by the total number of voxels in the neighborhood window to obtain the grayscale distribution, i.e., the probability P(i) of the grayscale level. Then, the Shannon entropy is calculated to obtain the 3D grayscale entropy feature map.

[0034] In some embodiments, the Shannon entropy can be calculated based on the grayscale distribution using the following formula to obtain a three-dimensional grayscale entropy feature map: ,in, Indicated by voxels The grayscale distribution (probability of grayscale levels) within the target neighborhood centered on the target. Represents a minimal constant. This represents the grayscale level. It can be understood that in the 3D grayscale entropy feature map, regions with irregular grayscale distribution, such as vessel boundaries, small branches, and areas with uneven contrast agent distribution, have higher entropy values ​​due to their chaotic grayscale distribution; while pure background or large areas of uniformly distributed blood vessels have concentrated grayscale distributions and lower entropy values. This 3D grayscale entropy feature map can guide the model to automatically focus on difficult-to-segment regions, providing structural prior information for subsequent feature fusion.

[0035] In step S103, a three-dimensional label entropy map is generated based on the three-dimensional vessel segmentation mask. In some embodiments, the three-dimensional label entropy map can be obtained by calculating the proportion values ​​of foreground voxels and background voxels in the target neighborhood centered on each voxel in the three-dimensional vessel segmentation mask, and calculating the Shannon entropy based on the proportion values.

[0036] Consistent with the generation of the 3D grayscale entropy feature map, a 3×3×3 3D neighborhood window can also be set for each voxel p in the 3D blood vessel segmentation mask to ensure the spatial scale of the two entropy maps is consistent, facilitating subsequent model processing. Next, the number of voxels with voxel values ​​of 0 (background) and 1 (foreground, blood vessel) within the target neighborhood window is counted: let the number of background voxels be... The number of foreground voxels is ,but (Total number of voxels in the neighboring window). Background proportions in this scene. and foreground ratio They are respectively And satisfy Then, Shannon entropy is calculated based on the ratio to obtain a three-dimensional label entropy map.

[0037] In some embodiments, the Shannon entropy can be calculated based on a scale value using the following formula to obtain a three-dimensional label entropy map: ,in, This represents the scale value of the foreground voxels. The scale value of the background voxels, P0 = 1 - P1, This represents a minimal constant. It can be understood that in the 3D label entropy map, the entropy values ​​are higher in areas such as vessel boundaries, small branches, and overlapping vessel intersections because the foreground-to-background ratio is close to 1:1; while the entropy values ​​are lower in pure background areas or large areas of continuous vessels. This 3D label entropy map can accurately locate difficult-to-segment regions, providing a basis for regional weighting in the weighted loss calculation.

[0038] Furthermore, feature fusion is performed on the 3D grayscale image and the 3D grayscale entropy feature map to obtain fused features. That is, by organically combining the original image features with the prior structural features, the feature representation capability of the model is improved.

[0039] In some embodiments, a feature network can be used to fuse the features of a 3D grayscale image and a 3D grayscale entropy feature map to obtain fused features; or the features of a 3D grayscale image and a 3D grayscale entropy feature map can be stitched together to obtain fused features.

[0040] In some embodiments, the feature network may include a first feature extraction unit, a second feature extraction unit, and a feature fusion unit. Specifically, the first and second feature extraction units are used to extract features from the 3D grayscale image and the 3D grayscale entropy feature image, respectively, to obtain corresponding first and second features. Then, the feature fusion unit fuses the first and second features to obtain a fused feature.

[0041] In some implementation scenarios, the first and second feature extraction units can be convolutional layers or small networks. The feature fusion unit can be set in the initial stage or intermediate layer of the network, and the first and second features are fused using methods such as concatenation, weighting, and attention mechanisms to obtain fused features.

[0042] In other words, the embodiments of this application can fuse the 3D grayscale image and the 3D grayscale entropy feature map through a dual-branch feature extraction and fusion method. Alternatively, the 3D grayscale image and the 3D grayscale entropy feature map can be directly concatenated to obtain fused features. Through dual-branch feature extraction, both features can be specifically enhanced, and after integration by the fusion unit, the resulting fused features are more discriminative than single features, significantly improving the model's accuracy in recognizing blood vessel boundaries and small blood vessels. The direct channel concatenation method is simple, completely preserves the original information of both features, and is suitable for direct fusion at the model input layer, allowing the model to learn both image information and structural priors simultaneously in the early stages of feature extraction.

[0043] Based on the obtained fusion features, in step S105, the fusion features are input into the segmentation model for blood vessel segmentation, and a weighted loss is calculated based on the 3D label entropy map. In some implementation scenarios, the segmentation model can be, for example, a 3D U-Net or V-Net model, which has encoding and decoding structures and skip connections, and can effectively extract multi-scale features and preserve spatial details, making it suitable for 3D segmentation tasks of cerebral blood vessels.

[0044] In the implementation scenario, fused features are input into the segmentation model. The encoding layer progressively downsamples to extract deep features, while the decoding layer progressively upsamples to restore spatial resolution and fuses multi-scale features. Finally, a probability map is obtained through the output layer. The probability map is then thresholded (e.g., 0.5), and voxels with a probability ≥ 0.5 are identified as blood vessels (predicted value). Voxels with a probability <0.5 are classified as background (predicted value). ), to generate the initial segmentation results.

[0045] The core logic of weighted loss is to assign higher loss weights to high-entropy, hard-to-segment regions during model training, thereby focusing on optimizing the segmentation accuracy of these regions. In some embodiments, the weighted loss can be calculated using the following formula: Where L represents the weighted loss, This represents the tag entropy value corresponding to the i-th voxel in the 3D tag entropy map. Indicates predicted value and the true value The cross-entropy, where i represents the voxel index and N represents the total number of voxels.

[0046] Finally, in step S106, the segmentation model for cerebral blood vessel segmentation is trained based on a weighted loss. That is, the weighted loss is minimized through an optimization algorithm, allowing the model to gradually learn the ability to accurately segment cerebral blood vessels. The fused features of the training set, the 3D blood vessel segmentation mask, and the 3D label entropy map are input into the model. Forward propagation yields the segmentation results and the weighted loss, followed by backpropagation to calculate the gradient and update the model parameters. This process is repeated until the training epochs are completed or an early stopping condition is triggered. After training, the model parameters, including the weights and biases of the segmentation model, are saved for subsequent cerebral blood vessel segmentation inference. Through weighted loss training, the model's segmentation accuracy in difficult-to-segment regions such as blood vessel boundaries and small branches is significantly improved, and the model's robustness to complex cerebral vascular networks is significantly enhanced.

[0047] As described above, this embodiment obtains a high-quality 3D grayscale image by denoising the 3D angiography image. The grayscale entropy feature map generated from the 3D grayscale image provides prior information on structural complexity. Then, the label entropy map generated based on the 3D blood vessel segmentation mask achieves regional adaptive weighting of the loss function. Simultaneously, the dual-branch feature extraction and fusion strategy allows the model to fully utilize the original image information and prior features. Combined with end-to-end training of the segmentation model, this not only guides the model to automatically focus on difficult-to-segment regions such as blood vessel boundaries and small branches, effectively solving the problems of blurred and adhered blood vessel boundaries in traditional methods, but also enhances the model's adaptability to interference from real-world scenarios such as uneven contrast agent distribution and image noise. This significantly improves the boundary accuracy and structural consistency of cerebral blood vessel segmentation, enhancing the model's robustness in segmenting complex cerebral vascular networks.

[0048] Figure 2 This is another exemplary flowchart illustrating the training of a segmentation model for cerebral blood vessel segmentation according to an embodiment of this application. Figure 2 As shown, in step S201, a three-dimensional angiography image is acquired. This three-dimensional angiography image may be annotated with a three-dimensional vessel segmentation mask. In step S202, the three-dimensional angiography image is denoised to obtain a three-dimensional grayscale image in step S203. In some embodiments, the denoising process may include Gaussian filtering and / or non-local mean filtering.

[0049] Next, in step S204, a three-dimensional gray-level entropy feature map is generated based on the three-dimensional gray-level image. As mentioned earlier, this can be based on calculating the gray-level distribution (probability of gray levels) within the target neighborhood centered on each voxel in the three-dimensional gray-level image. A three-dimensional grayscale entropy feature map is generated. Based on the three-dimensional grayscale image and the three-dimensional grayscale entropy feature map, in steps S205 and S206, the first feature extraction unit and the second feature extraction unit are used to extract features from the three-dimensional grayscale image and the three-dimensional grayscale entropy feature map, respectively, to obtain the corresponding first feature and second feature.

[0050] As an example, the first feature extraction unit and the second feature extraction unit may each contain, for example, two 3D convolutional layers with a kernel size of 3×3×3 and a stride of 1. The first layer has 16 kernels and the second layer has 32 kernels. A batch normalization (BN) layer and an activation function (ReLU) are used to extract the local texture and morphological features corresponding to the original image and the 3D grayscale entropy feature map, respectively.

[0051] Next, in step S207, a feature fusion unit is used to fuse the first feature and the second feature to obtain fused features. The feature fusion unit may include, for example, weighted fusion, assigning different weights to the first and second features; attention mechanism fusion, which dynamically allocates the contribution of the two features by introducing a 3D spatial attention module to calculate the spatial attention weights of the feature maps. Convolutional fusion after concatenation concatenates the first and second features along the channel dimension, and can further perform feature fusion through a 3×3×3 convolutional layer (32 kernels) as shown in Example 1, compressing the number of channels while extracting cross-feature correlation information to obtain fused features.

[0052] Further, in step S208, the fused features are input into the segmentation model for training. During training, combined with step S209, a weighted loss is calculated based on the 3D label entropy map for training. The 3D label entropy map can be calculated by determining the ratio of foreground voxels to background voxels within the target neighborhood centered on each voxel in the 3D blood vessel segmentation mask, based on the above... The calculation was obtained. For more details on the calculation, please refer to the above. Figure 1 The descriptions made will not be repeated here.

[0053] Figure 3 This is yet another exemplary flowchart illustrating the training of a segmentation model for cerebral blood vessel segmentation according to an embodiment of this application. Figure 3 As shown, in step S301, a three-dimensional angiography image is acquired. This three-dimensional angiography image may be annotated with a three-dimensional vessel segmentation mask. In step S302, the three-dimensional angiography image is denoised to obtain a three-dimensional grayscale image in step S303. In some embodiments, the denoising process may include Gaussian filtering and / or non-local mean filtering.

[0054] Next, in step S304, a three-dimensional gray-level entropy feature map is generated based on the three-dimensional gray-level image. As mentioned earlier, this can be based on calculating the gray-level distribution (probability of gray levels) within the target neighborhood centered on each voxel in the three-dimensional gray-level image. A 3D grayscale entropy feature map is generated. Based on the 3D grayscale image and the 3D grayscale entropy feature map, in step S305, the 3D grayscale image and the 3D grayscale entropy feature map are concatenated to obtain a fused feature. Feature concatenation involves stacking two single-channel feature maps along the channel dimension to form a dual-channel feature map. For example, if the dimension of the 3D grayscale image is H×W×D×1 (H is height, W is width, D is depth, and 1 is the number of channels), and the dimension of the 3D grayscale entropy feature map is also H×W×D×1, the dimension of the fused feature after concatenation is H×W×D×2.

[0055] Similarly, in step S306, the fused features are input into the segmentation model for training. During training, combined with step S307, a weighted loss is calculated based on the 3D label entropy map for training. The 3D label entropy map can be calculated by determining the ratio of foreground voxels to background voxels within the target neighborhood centered on each voxel in the 3D vessel segmentation mask, based on the above... The calculation was obtained. For more details on the calculation, please refer to the above. Figure 1 The descriptions made will not be repeated here.

[0056] Based on the trained segmentation model described above, new image data can be automatically segmented. In some embodiments, a three-dimensional angiography image to be segmented is acquired; the three-dimensional angiography image is then input into the trained segmentation model to perform cerebral blood vessel segmentation, thereby obtaining the cerebral blood vessel segmentation result.

[0057] In some embodiments, post-processing operations can be performed on the cerebral blood vessel segmentation results to obtain the final cerebral blood vessel segmentation results, wherein the post-processing operations may include boundary smoothing and / or isolated region removal.

[0058] Figure 4 This is an exemplary flowchart illustrating a method for cerebral blood vessel segmentation according to an embodiment of this application. Figure 4 As shown, in step S401, a three-dimensional angiography image to be segmented is acquired. As mentioned earlier, this image can be acquired using a clinical DSA device. Next, in step S402, the three-dimensional angiography image is input into a trained segmentation model for cerebral blood vessel segmentation to obtain the segmentation result. Before inputting the three-dimensional angiography image into the segmentation model, denoising can be performed on the image to generate a three-dimensional grayscale image and a three-dimensional grayscale entropy feature map. These two features can then be fused, and the fused features are input into the segmentation model.

[0059] Finally, in step S403, post-processing operations are performed on the cerebral blood vessel segmentation results to obtain the final cerebral blood vessel segmentation result. As mentioned earlier, post-processing operations may include boundary smoothing and / or isolated region removal. Smoothed vessel boundaries more closely resemble the anatomical morphology of real cerebral blood vessels, avoiding clinical misjudgments caused by irregular boundaries. After removing isolated regions, the false positive rate in the segmentation results is reduced, and the integrity and accuracy of the vascular network are significantly improved. In some implementation scenarios, three-dimensional Gaussian filtering can be used for boundary smoothing; connected component analysis algorithms, for example, can be used to identify and remove isolated regions. The final output cerebral blood vessel segmentation result includes a three-dimensional segmentation mask of the cerebral blood vessels.

[0060] Figure 5An exemplary structural block diagram of an electronic device 500 according to an embodiment of this application is shown. It will be understood that the device implementing the solution of this application may be a single device (e.g., a computing device) or a multifunctional device including various peripheral devices.

[0061] like Figure 5 As shown, the electronic device of this application may include a central processing unit (“CPU”) 511, which may be a general-purpose CPU, a dedicated CPU, or other information processing and program execution unit. Further, the electronic device 500 may also include a mass storage device 512 and a read-only memory (“ROM”) 513. The mass storage device 512 may be configured to store various types of data, including various data related to three-dimensional angiography images, three-dimensional grayscale, three-dimensional label entropy maps, fusion features, weighted losses, algorithm data, intermediate results, and various programs required to run the electronic device 500. The ROM 513 may be configured to store data and instructions required for the power-on self-test of the electronic device 500, the initialization of various functional modules in the system, the drivers for the system's basic input / output, and the booting of the operating system.

[0062] Optionally, the electronic device 500 may also include other hardware platforms or components, such as the tensor processing unit (“TPU”) 514, graphics processing unit (“GPU”) 515, field-programmable gate array (“FPGA”) 516, and machine learning unit (“MLU”) 517 shown. It is understood that although various hardware platforms or components are shown in the electronic device 500, they are merely exemplary and not limiting, and those skilled in the art can add or remove appropriate hardware as needed. For example, the electronic device 500 may include only a CPU, associated storage devices, and interface devices to implement the method for training a segmentation model for cerebral blood vessel segmentation or the method for cerebral blood vessel segmentation of this application.

[0063] In some embodiments, to facilitate data transmission and interaction with external networks, the electronic device 500 of this application further includes a communication interface 518, through which it can connect to a local area network / wireless local area network (“LAN / WLAN”) 505, and further through the LAN / WLAN to connect to a local server 506 or to the Internet (“Internet”) 507. Alternatively or additionally, the electronic device 500 of this application can also directly connect to the Internet or cellular network via the communication interface 518 based on wireless communication technology, such as wireless communication technology based on 3G (“3G”), 4G (“4G”), or 5G (“5G”). In some application scenarios, the electronic device 500 of this application can also access the server 508 and database 509 of an external network as needed to obtain various known algorithms, data, and modules, and can remotely store various data, such as various data or instructions for presenting, for example, three-dimensional angiography images, three-dimensional grayscale, three-dimensional label entropy maps, fusion features, weighted losses, etc.

[0064] Peripherals of the electronic device 500 may include a display device 502, an input device 503, and a data transmission interface 504. In one embodiment, the display device 502 may include, for example, one or more speakers and / or one or more visual displays, configured to train the segmentation model for cerebral vessel segmentation of this application or to provide voice prompts and / or image / video display for cerebral vessel segmentation. The input device 503 may include, for example, a keyboard, mouse, microphone, gesture capture camera, or other input buttons or controls, configured to receive audio data input and / or user commands. The data transmission interface 504 may include, for example, a serial interface, parallel interface, or Universal Serial Bus interface (“USB”), Small Computer System Interface (“SCSI”), Serial ATA, FireWire (“FireWire”), PCI Express, and High Definition Multimedia Interface (“HDMI”), configured for data transmission and interaction with other devices or systems. According to the scheme of this application, the data transmission interface 504 can receive acquired three-dimensional angiography images from a DSA device and transmit three-dimensional angiography images or various other types of data or results to the electronic device 500.

[0065] The aforementioned CPU 511, mass storage 512, ROM 513, TPU 514, GPU 515, FPGA 516, MLU 517, and communication interface 518 of the electronic device 500 of this application can be interconnected via bus 519, and can interact with peripheral devices through this bus. In one embodiment, the CPU 511 can control other hardware components and peripheral devices in the electronic device 500 through bus 519.

[0066] The above combination Figure 5 An electronic device that can be used to execute this application is described. It should be understood that the device structure or architecture described herein is merely exemplary, and the implementation method and implementation entity of this application are not limited thereto, but can be changed without departing from the spirit of this application.

[0067] Based on the foregoing description in conjunction with the accompanying drawings, those skilled in the art will understand that the embodiments of this application can also be implemented by software programs. Therefore, this application also provides a computer-readable storage medium storing computer-readable instructions for training a segmentation model for cerebral blood vessel segmentation or for cerebral blood vessel segmentation. When executed by one or more processors, these computer-readable instructions can be used to implement the embodiments of this application in conjunction with the accompanying drawings. Figure 1 The description includes methods for training segmentation models for cerebral blood vessel segmentation or Figure 4 The method described is for cerebral blood vessel segmentation.

[0068] It should be noted that although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart can be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0069] It should be understood that when the terms "first," "second," "third," and "fourth," etc., are used in the claims, specification, and drawings of this application, they are used only to distinguish different objects and not to describe a specific order. The terms "comprising" and "including" as used in the specification and claims of this application indicate the presence of the described features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof.

[0070] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this specification and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this specification and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0071] While numerous embodiments of this application have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will arise for those skilled in the art without departing from the spirit and intent of this application. It should be understood that various alternatives to the embodiments of this application described herein may be employed in the practice of this application. The appended claims are intended to define the scope of protection of this application and therefore cover equivalents or alternatives within the scope of these claims.

Claims

1. A method for training a segmentation model for cerebral blood vessel segmentation, the method comprising: Three-dimensional grayscale images and corresponding three-dimensional blood vessel segmentation masks are obtained from three-dimensional angiography images. A three-dimensional grayscale entropy feature map is generated based on the three-dimensional grayscale image; A three-dimensional label entropy map is generated based on the three-dimensional blood vessel segmentation mask; The three-dimensional grayscale image and the three-dimensional grayscale entropy feature map are fused to obtain fused features; The fused features are input into the segmentation model for blood vessel segmentation, and a weighted loss is calculated based on the three-dimensional label entropy map. The segmentation model for cerebral blood vessel segmentation is trained based on the weighted loss training. The generation of a 3D label entropy map based on the 3D blood vessel segmentation mask includes: Calculate the ratio of foreground voxels and background voxels in the target neighborhood centered on each voxel in the three-dimensional blood vessel segmentation mask; Shannon entropy is calculated based on the stated ratio to obtain the three-dimensional label entropy map; The Shannon entropy is calculated based on the stated ratio using the following formula to obtain the three-dimensional label entropy map: in, This represents the scale value of the foreground voxels. The scale value of the background voxels, P0 = 1 - P1, Represents a minimal constant; The weighted loss is calculated using the following formula: Where L represents the weighted loss, This represents the tag entropy value corresponding to the i-th voxel in the three-dimensional tag entropy map. Indicates predicted value and the true value The cross-entropy, where i represents the voxel index and N represents the total number of voxels.

2. The method of claim 1, wherein the three-dimensional grayscale image is obtained by the following operations: The three-dimensional angiography image is denoised to obtain the three-dimensional grayscale image, wherein the denoising process includes Gaussian filtering and / or nonlocal mean filtering.

3. The method according to claim 1, wherein generating a three-dimensional grayscale entropy feature map based on the three-dimensional grayscale image comprises: Calculate the grayscale distribution within the target neighborhood centered on each voxel in the three-dimensional grayscale image; Shannon entropy is calculated based on the grayscale distribution to obtain the three-dimensional grayscale entropy feature map.

4. The method according to claim 3, wherein the three-dimensional gray-level entropy feature map is obtained by calculating the Shannon entropy based on the gray-level distribution using the following formula: in, Indicated by voxels The grayscale distribution within the target neighborhood centered on the target. Represents a minimal constant. This indicates the grayscale level.

5. The method according to claim 1, wherein feature fusion of the three-dimensional grayscale image and the three-dimensional grayscale entropy feature map to obtain fused features includes: The feature network is used to fuse the 3D grayscale image and the 3D grayscale entropy feature image to obtain the fused features; or The fused feature is obtained by concatenating the 3D grayscale image and the 3D grayscale entropy feature image.

6. The method according to claim 5, wherein the feature network comprises a first feature extraction unit, a second feature extraction unit, and a feature fusion unit, and the feature network is used to perform feature fusion on the three-dimensional grayscale image and the three-dimensional grayscale entropy feature image to obtain the fused features, including: The first feature extraction unit and the second feature extraction unit are used to extract features from the three-dimensional grayscale image and the three-dimensional grayscale entropy feature image, respectively, to obtain the corresponding first feature and second feature; The feature fusion unit is used to fuse the first feature and the second feature to obtain the fused feature.

7. A method for segmenting cerebral blood vessels, comprising: Acquire three-dimensional angiography images to be segmented; The three-dimensional angiography image is input into the segmentation model trained by the method according to any one of claims 1-6 to perform cerebral blood vessel segmentation and obtain cerebral blood vessel segmentation results.

8. The method according to claim 7, further comprising: The cerebral blood vessel segmentation results are post-processed to obtain the final cerebral blood vessel segmentation results, wherein the post-processing operations include boundary smoothing and / or isolated region removal.

9. An electronic device, comprising: processor; as well as A memory having computer instructions stored thereon for training a segmentation model for cerebral vessel segmentation, wherein, when executed by a processor, the computer instructions cause the implementation of the method according to any one of claims 1-6; or having computer instructions stored thereon for cerebral vessel segmentation, wherein, when executed by a processor, the computer instructions cause the implementation of the method according to any one of claims 7-8.

10. A computer-readable storage medium storing thereon computer program instructions for training a segmentation model for cerebral vessel segmentation, wherein the computer program instructions, when executed by one or more processors, cause the implementation of the method according to any one of claims 1-6; or storing thereon computer program instructions for cerebral vessel segmentation, wherein the computer program instructions, when executed by one or more processors, cause the implementation of the method according to any one of claims 7-8.

Images