Vessel segmentation method, apparatus, electronic device, and computer program product

Abstract

Embodiments of the present application disclose a blood vessel segmentation method and device, electronic equipment and computer program product. The method comprises: obtaining a CT image of a target tissue; using a first encoder of a target neural network to perform down-sampling processing on the CT image to obtain first features of multiple scales corresponding to the CT image, wherein the first encoder comprises multiple down-sampling layers, each down-sampling layer comprises a feature extraction module and an Inception module, and at each down-sampling layer, the feature extraction module is used to extract features of the CT image or first features output by a previous down-sampling layer to obtain intermediate features, and the Inception module is used to perform parallel processing on the intermediate features in different processing branches; and using a decoder in the target neural network to perform up-sampling processing on the first features output by the first encoder to obtain a segmentation result of the CT image.

Description

Technical Field

[0001] The present application relates to the technical field of medical image segmentation, and in particular to a blood vessel segmentation method, device, electronic device, and computer program product. Background Art

[0002] During medical diagnosis and treatment, such as ablation and puncture procedures in the lungs, arteriovenous information can greatly assist doctors in their diagnosis and treatment. Therefore, the importance of automated arteriovenous extraction in the diagnosis and treatment process is increasing.

[0003] In related technologies, arteriovenous extraction is mostly achieved using deep learning image segmentation technology. However, the deep learning network for arteriovenous segmentation in related technologies has poor segmentation effect on CT (Computed Tomography) images, especially non-enhanced CT images. It has technical problems such as low segmentation precision and accuracy, and inability to completely extract arteries and veins.

[0004] To address the above-mentioned problems, no effective solutions have been proposed so far. Summary of the Invention

[0005] The embodiments of the present application provide a blood vessel segmentation method, device, electronic device, and computer program product to at least solve the technical problem in the related art that deep learning networks have poor effect in segmenting arteries and veins in CT images.

[0006] According to an aspect of the embodiments of the present application, a blood vessel segmentation method is provided, comprising: obtaining a CT image of a target tissue; performing down-sampling processing on the CT image by using a first encoder of a target neural network to obtain first features of multiple scales corresponding to the CT image, wherein the first encoder comprises multiple down-sampling layers, the first features output by different down-sampling layers are different in feature size and feature dimension, each down-sampling layer comprises a feature extraction module and an Inception module, in each down-sampling layer, the feature extraction module is used to extract features of the CT image or the first features output by a previous down-sampling layer to obtain intermediate features, and the Inception module is used to perform parallel processing on the intermediate features in different processing branches; performing up-sampling processing on the first features output by the last down-sampling layer by using a decoder of the target neural network to obtain a segmentation result of the CT image, wherein the segmentation result is used to indicate a artery-vein category of blood vessels in the CT image; wherein different processing branches correspond to different convolution kernels; the Inception module comprises a first Inception module and a second Inception module, wherein the convolution kernels of each processing branch in the first Inception module correspond to a convolution step length of 1, the processing branches of the second Inception module comprise the convolution kernels with the convolution step length of 1, and each down-sampling layer comprises at least one first Inception module.

[0007] Optionally, the parallel processing of the intermediate features in different processing branches by the Inception module comprises: segmenting the intermediate features according to the number of processing branches contained in the Inception module, and outputting the segmented intermediate features to different processing branches in the Inception module for convolution processing to obtain the first features corresponding to the current down-sampling layer, wherein each convolution kernel learns different feature categories, and the feature categories include at least one of the following: edge, texture, and color.

[0008] Optionally, the target neural network includes at least one Transformer encoder; the method also includes: using the Transformer encoder to process the first feature output by the target downsampling layer to obtain a second feature, and transmitting the second feature to a specified downsampling layer connected to the Transformer encoder, wherein the second feature includes processing results from different feature output layers of the Transformer encoder, the number of feature output layers in the Transformer encoder is the same as the number of processing branches in the Inception module, and the target downsampling layer is any downsampling layer in the first encoder except the last downsampling layer; in the specified downsampling layer, the first feature output by the previous downsampling layer is fused with the second feature output by the Transformer encoder, and the first feature output by the current downsampling layer is determined based on the fused feature.

[0009] Optionally, in a specified downsampling layer, the first feature output by the previous downsampling layer is fused with the second feature output by the Transformer encoder, and the first feature output by the current downsampling layer is determined based on the fused feature, including: when the specified downsampling layer includes several first Inception modules connected in sequence, the first Inception module connected to the feature extraction module contained in the specified downsampling layer fuses the intermediate feature from the feature extraction module with the second feature, and transmits the fused feature to the next first Inception module, and the next first Inception module continues to fuse the fused feature with the second feature until the first feature output by the specified downsampling layer is determined based on the fused feature output by the last first Inception module; when the specified downsampling layer includes a first Inception module and a second Inception module connected in sequence, the first Inception module fuses the intermediate feature from the feature extraction module with the second feature, and transmits the fused feature to the second Inception module, and the second Inception module continues to fuse the fused feature with the second feature, and determines the first feature output by the specified downsampling layer based on the fused feature transmitted by the second Inception module.

[0010] Optionally, the Inception module further includes: a feature fusion module; in a specified downsampling layer, the first feature output by the previous downsampling layer is fused with the second feature output by the Transformer encoder, including: using the feature fusion module to divide the target features into N groups, where N is the number of feature output layers of each Transformer encoder, and the target features include intermediate features or fused features from any Inception module; and the divided N groups of target features are respectively fused with the processing results output by each layer of the Transformer encoder to obtain N groups of fused features.

[0011] Optionally, the training steps of the target neural network include: obtaining an initial training data set, wherein the training data set contains multiple training images, the training images are CT images of the target tissue, the CT images contain target labels, and the target labels are used to indicate arteries and veins in the images; performing a preprocessing operation on the training images in the initial training data set to obtain a target training data set, wherein the preprocessing operation includes at least one of the following: a cropping operation, a normalization operation, a resampling operation for unifying the voxel spacing of each training image, and a mirroring operation; using the target training data set to train the initial neural network to obtain a target neural network, wherein the model structures of the initial neural network and the target neural network are consistent.

[0012] According to another aspect of the embodiment of the present application, a blood vessel segmentation device is also provided, including: an image acquisition module for acquiring a CT image of the target tissue; a feature extraction module for using a first encoder of the target neural network to downsample the CT image to obtain first features of multiple scales corresponding to the CT image, wherein the first encoder includes multiple downsampling layers, and the feature sizes and feature dimensions of the first features output by different downsampling layers are different. Each downsampling layer includes: a feature extraction module and an Inception module. In any downsampling layer, the feature extraction module is used to extract the first features output by the CT image or the previous downsampling layer to obtain intermediate features, and the intermediate features are processed in parallel on different processing branches through the Inception module. ; An image segmentation module is used to use the decoder in the target neural network to upsample the first feature finally output by the first encoder to obtain a segmentation result of the CT image, wherein the segmentation result is used to indicate the arteriovenous category of the blood vessels in the CT image; wherein different processing branches correspond to different convolution kernels; the Inception module includes: a first Inception module and a second Inception module, wherein the convolution step size corresponding to the convolution kernel of each processing branch in the first Inception module is 1, the processing branch of the second Inception module contains the convolution kernel whose convolution step size is not 1, and each of the downsampling layers contains at least one of the first Inception modules.

[0013] According to another aspect of the embodiments of the present application, an electronic device is provided, including: a memory and a processor, wherein the processor is configured to run a program stored in the memory, wherein the blood vessel segmentation method is executed when the program is run.

[0014] According to another aspect of the embodiments of the present application, a computer program product is provided, including a computer program, which implements the steps of the blood vessel segmentation method when executed by a processor.

[0015] In an embodiment of the present application, a CT image of a target tissue is acquired; a first encoder of a target neural network is used to downsample the CT image to obtain first features of multiple scales corresponding to the CT image, wherein the first encoder includes multiple downsampling layers, and the feature sizes and feature dimensions of the first features output by different downsampling layers are different. Each downsampling layer includes: a feature extraction module and an Inception module. In any downsampling layer, the feature extraction module extracts the first features output by the CT image or the previous downsampling layer to obtain intermediate features, and the Inception module processes the intermediate features in parallel on different processing branches; a decoder in the target neural network is used to upsample the first features finally output by the first encoder to obtain a segmentation result of the CT image, wherein the segmentation result is used to indicate the arteriovenous category of the blood vessels in the CT image. By introducing a channel-separable Inception structure in the feature extraction stage of the encoder-decoder network structure, the network is allowed to process information in parallel on different channels, thereby achieving the purpose of efficient and accurate segmentation of blood vessels and arteries and veins, thereby solving the technical problem of poor arteriovenous segmentation effect of deep learning networks in related technologies for CT images. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0017] Figure 1 is a schematic diagram of a blood vessel segmentation method process provided according to an embodiment of the present application;

[0018] Figure 2 is a schematic diagram of a blood vessel data annotation effect provided according to an embodiment of the present application;

[0019] Figure 3 is a schematic diagram of an example of a non-enhanced CT training image after data annotation provided in an embodiment of the present application;

[0020] Figure 4 is a schematic diagram of a structure of a target neural network provided according to an embodiment of the present application;

[0021] Figure 5 is a schematic diagram of an InceptionS structure provided according to an embodiment of the present application;

[0022] Figure 6 is a schematic diagram of a dual-stream deep coupling neural network structure provided according to an embodiment of the present application;

[0023] Figure 7is a schematic diagram of a workflow of a Transformer encoder provided according to an embodiment of the present application;

[0024] Figure 8 is a schematic diagram of a feature fusion module provided according to an embodiment of the present application;

[0025] Figure 9 is a structural schematic diagram of a blood vessel segmentation device provided according to an embodiment of the present application;

[0026] Figure 10 This is a hardware structure block diagram of a computer terminal (or electronic device) for implementing a method for blood vessel segmentation provided in an embodiment of the present application. DETAILED DESCRIPTION

[0027] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0028] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in a sequence other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0029] In related technologies, the extraction of arteries and veins in target tissues is often achieved using deep learning image segmentation techniques. For example, the design of convolutional neural networks (CNNs) is widely used in arteriovenous segmentation. However, using CNN models alone cannot fully utilize feature information at different levels, thus affecting the accuracy and reliability of segmentation.

[0030] Therefore, an appropriate feature fusion strategy is crucial for integrating shallow local details of the image with deep global features, which has a significant impact, especially in the field of medical image segmentation.

[0031] In order to solve the above problems, the embodiments of the present application provide relevant solutions. The technical solutions of the present application are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0032] According to an embodiment of the present application, an embodiment of a method for blood vessel segmentation is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0033] The present invention provides a blood vessel segmentation method. Figure 1 FIG. 1 is a schematic diagram of a blood vessel segmentation method according to an embodiment of the present application. Figure 1 As shown, the method includes the following steps:

[0034] Step S102, acquiring a CT image of the target tissue;

[0035] Step S104: Downsampling the CT image using the first encoder of the target neural network to obtain first features at multiple scales corresponding to the CT image. The first encoder includes multiple downsampling layers, and the feature sizes and feature dimensions of the first features output by different downsampling layers are different. Each downsampling layer includes a feature extraction module and an Inception module. In any downsampling layer, the feature extraction module extracts the first features output by the CT image or the previous downsampling layer to obtain intermediate features, and the Inception module processes the intermediate features in parallel on different processing branches.

[0036] In step S106, the decoder in the target neural network is used to upsample the first feature finally output by the first encoder to obtain a segmentation result of the CT image, wherein the segmentation result is used to indicate the arteriovenous category of the blood vessels in the CT image.

[0037] Among them, different processing branches correspond to different convolution kernels; the Inception module includes: a first Inception module and a second Inception module, wherein the convolution step corresponding to the convolution kernel of each processing branch in the first Inception module is 1, and the processing branch of the second Inception module contains the convolution kernel whose convolution step is not 1, and each downsampling layer contains at least one first Inception module.

[0038] Through the above steps, by introducing the channel-separable Inception structure into the encoder part of the encoder-decoder network structure, the network is allowed to process information in parallel on different processing branches, thereby achieving the goal of efficient and accurate segmentation of arteries and veins, thereby solving the technical problem of poor arteriovenous segmentation effect of deep learning networks in related technologies for CT images.

[0039] The following further introduces the blood vessel segmentation method in steps S102 to S106 of the embodiment of the present application.

[0040] In an embodiment of the present application, the target tissue may include the lungs, liver, or kidneys, etc. The vascular segmentation scheme of the target tissue is illustrated by taking the segmentation of the pulmonary artery, pulmonary vein, aorta and other arteries and veins of the lungs as an example. Specifically, in an embodiment of the present application, the CT pulmonary artery and vein segmentation process based on deep learning mainly includes: data labeling, data preprocessing, model training, automatic segmentation of pulmonary arteries and veins and other steps, which are described in detail below.

[0041] First, we can construct a training data set through data labeling and data preprocessing, and train the target neural network model. The specific steps are as follows.

[0042] In some embodiments of the present application, the training steps of the target neural network include: obtaining an initial training data set, wherein the initial training data set contains multiple training images, the training images are CT images of the target tissue, the CT images contain target labels, and the target labels are used to indicate arteries and veins in the image; performing a preprocessing operation on the training images in the initial training data set to obtain a target training data set, wherein the preprocessing operation includes at least one of the following: a cropping operation, a normalization operation, a resampling operation for unifying the voxel spacing of each training image, and a mirroring operation; using the target training data set to train the initial neural network to obtain a target neural network, wherein the model structure of the initial neural network and the target neural network are consistent.

[0043] In practical applications, non-enhanced CT images are a widely used medical imaging method in clinical practice. Compared with enhanced CT images, they have the advantages of being faster, more adaptable to a wider range of people, and not requiring contrast agents. Therefore, in the examples of this application, the CT images used to train the target neural network are illustrated using non-enhanced CT images as an example.

[0044] Specifically, the initial training data set can be obtained by first annotating the non-enhanced CT image. For example, in the data annotation process, the non-enhanced CT image can be used and the pulmonary artery, pulmonary vein and aorta in the image can be annotated to obtain the training image. Figure 2 As shown, Figure 2(A) is the data labeling result of the pulmonary artery, (B) is the data labeling result of the pulmonary vein, (C) is the data labeling result of the thoracic aorta, and (D) is the data labeling result of the pulmonary artery and vein. The training image sample obtained after the pulmonary artery and vein data labeling is as follows: Figure 3 shown.

[0045] To ensure that the model receives high-quality, standardized training image data during training, preprocessing operations can be performed on the training images in the initial training dataset to obtain the target training dataset. Specifically, operations such as cropping, normalization, resampling, and mirroring can be used during data preprocessing. Cropping is used to reduce unnecessary background data, normalization ensures consistency in the data feature scale across training images, resampling ensures consistent voxel spacing in the CT data, and mirroring enhances data and improves model generalization.

[0046] This preprocessing not only removes noise and irrelevant information from the image, but also enhances image contrast and unifies image resolution, providing the model with clearer and more consistent vascular features. This training strategy addresses the problem of large variations in image quality affecting model training when processing CT images. By training the model on consistent image data, the accuracy and consistency of vascular segmentation are improved.

[0047] After obtaining the target training data set, the initial neural network is trained using the data set to obtain the target neural network (model) for automatic segmentation of pulmonary arteries and veins. The model structure and workflow of the neural network (model) in the embodiment of the present application are further introduced below, as follows.

[0048] In an embodiment of the present application, the target neural network includes a backbone network, which can extract the CNN feature stream of the CT image based on the convolutional neural network. For example, the backbone network can be a U-shaped ConvNet network or a ConvNet network, etc. In this embodiment, the backbone network is illustrated by taking the U-shaped ConvNet network as an example.

[0049] In the CNN feature flow, the U-shaped ConvNet serves as the backbone for feature extraction. Its encoder-decoder architecture uses skip connections to maintain information flow between features. Furthermore, the channel-separable Inception architecture is introduced into the encoder portion of the U-shaped ConvNet, allowing the network to process information in parallel across different channels. This reduces the number of parameters while increasing network capacity and complexity, as detailed below.

[0050] In some embodiments of the present application, the first encoder of the backbone network (U-shaped ConvNet network) included in the target neural network can be used to downsample the CT image to obtain first features of multiple scales corresponding to the CT image, wherein the first encoder includes multiple downsampling layers, and the feature sizes and feature dimensions of the first features output by different downsampling layers are different. Each downsampling layer includes: a feature extraction module and an Inception module; the decoder in the target neural network is used to upsample the first feature finally output by the first encoder to obtain the segmentation result of the blood vessels in the CT image.

[0051] Specifically, if Figure 4 As shown, it contains the first encoder (Encoder) and the decoder (Decoder). In the CNN feature stream, the first encoder contains multiple downsampling layers. Figure 4 Each downsampling layer is represented as a convolution block (ConvBlock). Multiple downsampling layers obtain high-level features by gradually downsampling the input image. During the downsampling process, the image size is gradually reduced and the feature channels are gradually increased. The decoder contains multiple upsampling layers (UPSample). Multiple upsampling layers gradually upsample the first feature output by the first encoder, and fuse the first features through jump connections during the upsampling process to finally obtain the segmentation result (Output).

[0052] Specifically, each downsampling layer (convolution block) contains a feature extraction module (CBAD (Conv (convolution), BatchNorm (batch normalization), Activate (activation layer), Dropout (dropout)) module) and an Inception module. Each of the convolution blocks can contain different configurations. For example, the number and type of Inception modules contained in the convolution block are different. For example, Figure 4 As shown, the first encoder of this embodiment mainly includes three convolution blocks: convolution block-A (ConvBlock-A), convolution block-B (ConvBlock-B) and convolution block-C (ConvBlock-C). Among them, ConvBlock-A includes a CBAD module and an InceptionS-A module, ConvBlock-B includes a CBAD module, an inceptionS-A module and an InceptionS-B module, and ConvBlock-C includes a CBAD module, an InceptionS-A module and an InceptionS-C module.

[0053] It should be noted that the above-mentioned ConvBlock-B1 and ConvBlock-B2 have the same network structure and are numbered 1 and 2 only for the convenience of subsequent description; ConvBlock-C1, ConvBlock-C2, and ConvBlock-C3 have the same network structure and are numbered 1, 2, and 3 only for the convenience of subsequent description.

[0054] The following describes the processing flow and functions of the feature extraction module (CBAD module) and the Inception module.

[0055] In some embodiments of the present application, at any downsampling layer, a feature extraction module is used to extract the first feature output by the CT image or the previous downsampling layer to obtain an intermediate feature, and the intermediate feature is processed in parallel on different processing branches by the Inception module; wherein, the parallel processing of the intermediate feature on different processing branches by the Inception module includes: using the Inception module to split the intermediate feature according to the number of processing branches it contains, and outputting the split intermediate features to different processing branches in the Inception module for convolution processing to obtain the first feature corresponding to the current downsampling layer, wherein different processing branches correspond to different convolution kernels, and each convolution kernel learns different feature categories, and the feature categories include at least one of the following: edge, texture, and color.

[0056] Specifically, the function of the CBAD module is feature extraction, which mainly adds channels to the first feature output by the previous downsampling layer; while InceptionS (generally referring to Inception-A / B / C) can increase the number of channels, so that the output results of the corresponding previous downsampling layer are input into different processing branches for processing. Each processing branch corresponds to a different convolution kernel (that is, a different filter), and each convolution kernel can learn different features, such as edges, textures, colors, etc.

[0057] The introduction of the feature extraction module in the embodiment of the present application can further refine and enhance the features processed by the Inception module, enabling it to capture key information more accurately. The design of the Inception module in the channel dimension is segmented and processed in parallel, allowing the model to simultaneously learn features of different scales and directions, thereby enhancing the understanding and segmentation capabilities of complex vascular networks.

[0058] It should be noted that the above-mentioned Inception module includes many different categories, as follows.

[0059] In some embodiments of the present application, the Inception module includes: a first Inception module and a second Inception module, wherein the convolution step corresponding to the convolution kernel of each processing branch in the first Inception module is 1, the processing branch of the second Inception module contains a convolution kernel with a convolution step not of 1, and each downsampling layer includes at least one first Inception module.

[0060] Specifically, if Figure 5 As shown in the figure, the difference between the first Inception module (such as Inception-A) and the second Inception module (such as Inception-B and Inception-C) is that the convolution step size corresponding to the convolution kernel of each processing branch of the first Inception module (such as Inception-A) is 1 (i.e., s=1), which is mainly used to increase the number of channels corresponding to the input features without changing the feature scale of the features; while the processing branches of the second Inception module (such as Inception-B and Inception-C) contain convolution kernels with a convolution step size not equal to 1 (such as Figure 5 As shown in the figure, it contains a convolution kernel of s=2), which is mainly used to increase the number of channels corresponding to the input features and reduce the feature scale. Different categories of second Inception modules have different degrees of reduction in feature scale. For example, Inception-B and Inception-C have different degrees of reduction in feature scale.

[0061] In the embodiment of the present application, the design of the Inception module in each convolution block includes but is not limited to the combined use of the first and second Inception modules. Each convolution block contains at least one first Inception module, which can ensure that the model increases the number of channels while maintaining the feature scale unchanged, and increases the number of channels by reducing the feature scale when necessary, thereby capturing the detailed features of blood vessels at different scales.

[0062] As an optional implementation, in order to enable the target neural network to effectively capture the long-distance dependency between features, a Transformer network structure for capturing the Transformer features of the image can also be introduced into the target neural network, that is, a 3D hybrid neural network architecture based on the above-mentioned U-shaped ConvNet and Transformer can be built, which can be named as a dual-stream deeply coupled neural network (DSDCNet) in the embodiments of the present application, wherein the Transformer network structure can include at least one Transformer encoder, and three Transformer encoders are taken as an example for illustration in the embodiments.

[0063] For example, as shown in FIG. 1, the DSDCNet can include a U-shaped ConvNet feature extraction backbone network and three parallel Transformer encoders (Transformer Encoder). It should be noted that in the network structure of the embodiments, the number of Transformer encoders is fixed to three in order to correspond to the ConvBlock, but the number of Transformer encoders can be adjusted according to the number of ConvBlock, and no limitation is made to this. Figure 6

[0064] The feature extraction backbone network and the Transformer encoder work together to capture local features and global context information of the image, and enhance the representation ability of the features through deep feature fusion strategy. Specifically as follows.

[0065] In some embodiments of the present application, the target neural network includes at least one Transformer encoder; the method further includes: using the Transformer encoder to process the first features output by the target down-sampling layer to obtain second features, and transmitting the second features to a specified down-sampling layer connected with the Transformer encoder, wherein the second features include processing results from different feature output layers of the Transformer encoder, the number of the feature output layers in the Transformer encoder is the same as the number of the processing branches in the Inception module, and the target down-sampling layer is any down-sampling layer except the last down-sampling layer in the first encoder; in the specified down-sampling layer, the first features output by the last down-sampling layer are fused with the second features output by the Transformer encoder, and the first features output by the current down-sampling layer are determined according to the fused features. ​

[0066] Specifically, in Figure 6 In the network structure shown, the Transformer feature stream consists of three parallel Transformer encoders of different patch sizes, all of which receive output from the same downsampling layer (i.e., the target downsampling layer mentioned above). Figure 6 The target downsampling layer is illustrated using Convolution Block-B1 (ConvBlock-B1) as an example. This embodiment does not impose any restrictions on this. The Transformer encoder can process the first feature after any downsampling except the last one. It only needs to adjust the number of patches (patchsize) so that its output is consistent with the output size of the corresponding downsampling layer. For example, to obtain the downsampling layer input earlier, the first feature after the first downsampling can be received. This allows for more original image information while the Transformer can obtain the same input. This design aims to capture feature representations at different scales to enhance the network's ability to fuse low-level details and high-level semantic information.

[0067] Among them, the internal workflow of each Transformer encoder is as follows Figure 7 (and Figure 6 As shown in the figure, corresponding to the Transformer encoder part on the far left), in this embodiment, the number of feature output layers of each Transformer encoder is set to 4, that is, each Transformer encoder consists of 4 identical layer stacks.

[0068] The workflow of the Transformer encoder is as follows: the three-dimensional input is transformed into Split and form n 1-dimensional sequences through input embedding. , H represents height, which refers to the number of pixels in the vertical direction of the image. ‌W represents width, which refers to the number of pixels in the horizontal direction of the image. D represents depth, which refers to the depth direction of the image, such as the number of slices in a medical image. The input features are added with position information after position embeddings. The LayerNorm layer normalizes the input sequence, and the multi-head attention mechanism module sets the number of attention heads to 4. Ultimately, each Transformer encoder can output a second feature, which contains a total of 4 layers (corresponding to the 4 feature output layers).

[0069] In this embodiment, the transformer encoder of three different patchsizes can output three different sizes of second features, which are fused with different stages (downsampling layers) of the backbone network respectively, thereby enhancing the feature expression capability of the network at different levels. Among them, the small size patch helps to capture fine texture and edge information, while the large size patch is better at capturing global context information. The patchsize design of the transformer is different from the corresponding stages of the Unet feature flow, that is, the first feature and the second feature with the same scale can be fused. For example, assuming that the first feature output by the downsampling layer has a channel number of 128, and the second feature output by the transformer encoder contains 4 layers in total, and each layer of feature has a channel of 32, then the 128 channels output by the convolution can be divided into 4 channels of 32 respectively, and then added to the corresponding transformer layer, and then the channels are combined to regenerate an output of 128 channels, thereby optimizing the fusion process of the transformer feature layer and the CNN backbone feature.

[0070] The process of fusing the first feature and the second feature in different types of downsampling layers (convolution blocks) is introduced as follows.

[0071] In some embodiments of the present application, at a specified downsampling layer, the first feature output by the previous layer of downsampling layer is fused with the second feature output by the transformer encoder, and the first feature output by the current downsampling layer is determined according to the fused feature, including the following steps: in the case that the specified downsampling layer contains a plurality of first Inception modules connected in sequence, the intermediate feature from the feature extraction module is fused with the second feature by the first Inception module connected with the feature extraction module contained in the specified downsampling layer, and the fused feature is transmitted to the next first Inception module, the next first Inception module continues to fuse the fused feature with the second feature, until the first feature output by the specified downsampling layer is determined according to the fused feature output by the last first Inception module;

[0072] For example, assuming that the specified downsampling layer contains a feature extraction module (CBAD module) and a first Inception module (taking Inception-A as an example), the first Inception module will fuse the intermediate features output by the feature extraction module with the second features output by the Transformer encoder, and output the fused features. Since the specified downsampling layer has only one first Inception module, that is, the first Inception module is the last first Inception module, the fused features output by it are directly determined as the first features output by the specified downsampling layer.

[0073] In another case, when the specified downsampling layer includes a first Inception module and a second Inception module connected sequentially, the first Inception module fuses the intermediate features from the feature extraction module with the second features, and transmits the fused features to the second Inception module. The second Inception module continues to fuse the fused features with the second features, and determines the first feature output by the specified downsampling layer based on the fused features transmitted by the second Inception module.

[0074] For example, Figure 6 The convolution block -C2 (ConvBlock-C2) shown in the figure is used as an example of a specified downsampling layer, which includes a feature extraction module (CBAD module), a first Inception module (Inception-A) and a second Inception module (Inception-C) connected in sequence. First, the first Inception module (Inception-A) will fuse the intermediate features output by the feature extraction module with the second features output by the Transformer encoder, and output the fused features to the second Inception module (Inception-C). Then, the second Inception module will continue to fuse the fused features input by the first Inception module (Inception-A) with the second features, and finally obtain the fused features output by the second Inception module (Inception-C) (that is, the first features output by the specified downsampling layer).

[0075] In order to achieve the fusion of the first feature and the second feature, the first encoder in the embodiment of the present application provides an access port for fusing the Transformer hierarchical features (i.e., the second feature). Specifically, the above-mentioned Inception module can also include an FBM (Feature Button Module, also known as a "feature button"). The feature fusion module can divide the processing results from the CNN feature channel according to the number of layers of the Transformer encoder (for example, proportional division), thereby achieving the fusion of feature streams, as follows.

[0076] In some embodiments of the present application, the Inception module further includes: a feature fusion module; in a specified downsampling layer, the first feature output by the previous downsampling layer is fused with the second feature output by the Transformer encoder, including: using the feature fusion module to divide the target features into N groups, where N is the number of feature output layers of each Transformer encoder, and the target features include intermediate features or fused features from any Inception module; the divided N groups of target features are respectively fused with the processing results output by each layer of the Transformer encoder to obtain N groups of fused features.

[0077] Specifically, if Figure 8 As shown in the figure, during feature fusion using the feature fusion module, the Transformer hierarchical features (i.e., the second features) form a hierarchical structure containing continuous semantic information, including low-level fine-grained information and high-level semantic information. At the output stage, these hierarchical features are resized to the same size as the CNN features (target features), and their number of channels is set to 1 / N of the number of CNN feature channels. That is, the number of channels of the features output by each layer of the Transformer encoder is 1 / N of the number of target feature channels. Since the number of feature output layers of each Transformer encoder is set to 4 in this embodiment, N is set to 4. This division allows the CNN features and hierarchical features to be effectively fused, achieving the complementarity between low-level fine-grained information and high-level semantic information. Here, D represents the depth, which refers to the depth direction of the image, such as the number of slices in a medical image. H represents the height, which refers to the number of pixels in the vertical direction of the image. W represents the width, which refers to the number of pixels in the horizontal direction of the image. C represents the number of channels. This fusion strategy can effectively retain key low-level details while strengthening high-level semantic features, thereby improving the performance of the model.

[0078] In addition, in order to enable the model to automatically focus on the feature parts that are more discriminative and informative when processing features, in an embodiment of the present application, the Inception module may also include a SimAM attention mechanism (i.e., an attention module), as follows.

[0079] Specifically, in the Inception module, the intermediate features output by the feature extraction module, after processing by the FBM module, are divided into N (for example, 4) processing branches: x1, x2, x3, and x4. By adjusting the channel preference of each branch, the number of model parameters can be effectively reduced. The features of all branches are then concat-fused, normalized using InstanceNorm (Norm module), and further activated using Leaky ReLU (Activate module) to enhance feature representation.

[0080] like Figure 5 As shown in the figure, in the Inception module (InceptionS structure), the number of channels in each processing branch is different. Often, more channels can be allocated to branches with fewer parameters to reduce the number of parameters. For example, in this embodiment, the channel setting can follow the following rules: for a 128-channel input, if the output channels are set to 256, the input features are divided into four equal parts of 32-channel features, sent to the four channels, and the channel with fewer parameters is set to 160 (128 + 32). The number of channels in other branches does not increase.

[0081] Through this attention mechanism, the model can more efficiently utilize computing resources and achieve accurate recognition and classification of vascular structures.

[0082] The following combination Figure 5 and Figure 6 Taking ConvBlock-B2 as an example, the overall process of dual feature stream fusion is explained. The CBAD module in ConvBlock-B2 extracts features from the processing result (i.e., the first feature) from ConvBlock-B1, and then inputs the feature extraction result (intermediate feature) into the SimAM attention mechanism module of InceptionS-A for processing. In the absence of output from the Transformer Encoder, the processing result of the SimAM attention mechanism module is directly input into the FBM for processing. The FBM does not need to perform feature fusion, and the processing result obtained by the InceptionS-A module continues to be input into InceptionS-B.

[0083] Taking ConvBlock-C2 as an example, the output result of ConvBlock-C1 (first feature) is input into the CBAD module for feature extraction, and the feature extraction result (intermediate feature) output by the CBAD module is input into the SimAM attention mechanism module of InceptionS-A for processing. The processing result output by the SimAM attention mechanism module and the output of TransformerEncoder (second feature) are input into FBM at the same time for feature fusion, and then processed by "concat", "norm" and "act" to obtain the output result of InceptionS-A (fused feature), and the output result of InceptionS-A is further output to InceptionS-C for processing. The processing process of InceptionS-C is similar to that of InceptionS-A.

[0084] The embodiment of the present application proposes a dual-stream deeply coupled neural network structure for non-enhanced CT pulmonary artery and vein segmentation. The network architecture includes two feature streams, Transformer and CNN. That is, in this network, the U-shaped ConvNet feature extraction backbone network and the Transformer encoder work in parallel, which can capture long-range spatial dependencies through the Transformer while retaining the ConvNet inductive bias. It can effectively integrate the advantages of the Transformer and the convolutional neural network to achieve efficient extraction of CT image features. At the same time, in the feature fusion stage, a fine-grained fusion strategy using the FBM module to achieve channel division and hierarchical semantic features is proposed, which can integrate feature information at different levels, that is, fuse the convolutional neural network output (first feature) with the Transformer feature (second feature). In addition, in the U-shaped ConvNet structure, an optimized Inception convolution structure is proposed, which can achieve parallel processing of features in the channel dimension by channel division. It not only maintains the network's deep feature fusion capability, but also significantly reduces the number of parameters and computational complexity of the Inception structure.

[0085] In an embodiment of the present application, after training an initial neural network having the above-described network structure using a target training data set, a target neural network capable of automatically segmenting pulmonary arteries and veins can be obtained. During the model application process, the trained weights are used to infer the input non-enhanced CT, and the pulmonary artery and vein segmentation results (dividing the blood vessels into arteries, veins, and aorta) can be obtained as the model output.

[0086] In practical applications, the target neural network (DSDCNet model) in the embodiments of this application can effectively segment fine blood vessels in the peripheral lung region, with segmentation speed far exceeding that of traditional segmentation methods in related technologies and segmentation accuracy superior to those used in related technologies, with segmented vessels as thin as 1 mm. This significantly improves the accuracy and robustness of blood vessel segmentation, providing more precise imaging support for clinical diagnosis and treatment, and has significant medical value and application prospects.

[0087] According to an embodiment of the present application, an embodiment of a blood vessel segmentation device is also provided. Figure 9 FIG. 1 is a schematic diagram of the structure of a blood vessel segmentation device according to an embodiment of the present application. Figure 9 As shown, the device includes:

[0088] An image acquisition module 90 is used to acquire a CT image of the target tissue;

[0089] A feature extraction module 92 is configured to downsample the CT image using a first encoder of the target neural network to obtain first features at multiple scales corresponding to the CT image. The first encoder includes multiple downsampling layers, and the feature sizes and dimensions of the first features output by different downsampling layers are different. Each downsampling layer includes a feature extraction module and an Inception module. In any downsampling layer, the feature extraction module extracts the first features output by the CT image or the previous downsampling layer to obtain intermediate features, and the Inception module processes the intermediate features in parallel on different processing branches.

[0090] The image segmentation module 94 is used to use the decoder in the target neural network to upsample the first feature finally output by the first encoder to obtain a segmentation result of the CT image, wherein the segmentation result is used to indicate the arteriovenous category of the blood vessels in the CT image.

[0091] Among them, different processing branches correspond to different convolution kernels; the Inception module includes: a first Inception module and a second Inception module, wherein the convolution step corresponding to the convolution kernel of each processing branch in the first Inception module is 1, and the processing branch of the second Inception module contains the convolution kernel whose convolution step is not 1, and each downsampling layer contains at least one first Inception module.

[0092] Optionally, performing parallel processing of intermediate features on different processing branches through the Inception module includes: using the Inception module to split the intermediate features according to the number of processing branches they contain, and outputting the split intermediate features to different processing branches in the Inception module for convolution processing to obtain the first feature corresponding to the current downsampling layer, wherein the feature categories learned by each convolution kernel are different, and the feature categories include at least one of the following: edge, texture, and color.

[0093] Optionally, the target neural network includes at least one Transformer encoder; the blood vessel segmentation device is further used to: use the Transformer encoder to process the first feature output by the target downsampling layer to obtain a second feature, and transmit the second feature to a specified downsampling layer connected to the Transformer encoder, wherein the second feature includes processing results from different feature output layers of the Transformer encoder, the number of feature output layers in the Transformer encoder is the same as the number of processing branches in the Inception module, and the target downsampling layer is any downsampling layer in the first encoder except the last downsampling layer; in the specified downsampling layer, the first feature output by the previous downsampling layer is fused with the second feature output by the Transformer encoder, and the first feature output by the current downsampling layer is determined based on the fused feature.

[0094] Optionally, in the specified down-sampling layer, the first feature output by the previous down-sampling layer is fused with the second feature output by the Transformer encoder, and the first feature output by the current down-sampling layer is determined according to the fused feature, including: in the case that the specified down-sampling layer contains a plurality of first Inception modules connected in sequence, the intermediate feature from the feature extraction module is fused with the second feature by the first Inception module connected with the feature extraction module in the specified down-sampling layer, and the fused feature is transmitted to the next first Inception module, the next first Inception module continues to fuse the fused feature with the second feature, until the first feature output by the specified down-sampling layer is determined according to the fused feature output by the last first Inception module; in the case that the specified down-sampling layer contains a first Inception module and a second Inception module connected in sequence, the intermediate feature from the feature extraction module is fused with the second feature by the first Inception module, and the fused feature is transmitted to the second Inception module, the second Inception module continues to fuse the fused feature with the second feature, and the first feature output by the specified down-sampling layer is determined according to the fused feature transmitted by the second Inception module.

[0095] Optionally, the Inception module further comprises: a feature fusion module; in the specified down-sampling layer, the first feature output by the previous down-sampling layer is fused with the second feature output by the Transformer encoder, including: using the feature fusion module to divide the target feature into N groups, wherein N is the number of feature output layers of each Transformer encoder, and the target feature includes the intermediate feature or the fused feature from any Inception module; and performing feature fusion on the N groups of divided target features and the processing results output by each layer of the Transformer encoder respectively to obtain N groups of fused features.

[0096] Optionally, the training step of the target neural network comprises: obtaining an initial training data set, wherein the training data set comprises a plurality of training images, the training images are CT images of the target tissue, and the CT images comprise a target label indicating the arterial and venous blood vessels in the images; performing a preprocessing operation on the training images in the initial training data set to obtain a target training data set, wherein the preprocessing operation comprises at least one of the following: a cropping operation, a normalization operation, a resampling operation for unifying the voxel spacing of each training image, and a mirror operation; training the initial neural network using the target training data set to obtain the target neural network, wherein the model structure of the initial neural network and the target neural network is consistent.

[0097] It should be noted that each module in the above blood vessel segmentation device can be a program module (for example, a program instruction set for implementing a certain specific function) or a hardware module. For the latter, it can be in the form of, but not limited to, a processor, or the functions of the above modules are implemented by a processor.

[0098] It should be noted that the blood vessel segmentation device provided in the embodiment can be used to execute Figure 1 the blood vessel segmentation method shown in the figure, and the related explanations of the above blood vessel segmentation method also apply to the embodiments of the present application, which will not be repeated here.

[0099] According to the embodiments of the present application, an embodiment of an electronic device is also provided. The method embodiments provided by the embodiments of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 10 A hardware structure block diagram of a computer terminal (or electronic device) for implementing the blood vessel segmentation method is shown. As Figure 10 shown, the computer terminal 100 (or electronic device) can include one or more (in the figure, 1002a, 1002b, …, 1002n are shown) processors 1002 (the processor 1002 can include but not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 1004 for storing data, and a transmission device 1006 for communication function. In addition, it can also include a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which can be included as one of the ports of the BUS bus), a network interface, a power supply and / or a camera. Those skilled in the art can understand, Figure 10 that the structure shown is only for illustration, and does not limit the structure of the above electronic device. For example, the computer terminal 100 can include more or less components than Figure 10 shown, or have a different configuration than Figure 10 shown.

[0100] It should be noted that the one or more processors 1002 and / or other data processing circuits described above can be referred to as "data processing circuits" herein. The data processing circuit can be embodied in whole or in part as software, hardware, firmware or any combination thereof. In addition, the data processing circuit can be a single independent processing module, or all or part of any one of the other elements combined into the computer terminal 100 (or electronic device). As referred to in the embodiments of the present application, the data processing circuit as a kind of processor control (for example, the selection of the variable resistance terminal path connected with the interface).

[0101] The memory 1004 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the blood vessel segmentation method in the embodiments of the present application. The processor 1002 executes the software programs and modules stored in the memory 1004 to perform various functional applications and data processing, thereby implementing the above-mentioned blood vessel segmentation method. The memory 1004 may include high-speed random access memory and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 1004 may further include memory remotely located relative to the processor 1002, and these remote memories may be connected to the computer terminal 100 via a network. Examples of such networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0102] Transmission device 1006 is configured to receive or transmit data via a network. A specific example of the aforementioned network may include a wireless network provided by the communications provider of computer terminal 100. In one embodiment, transmission device 1006 includes a network interface controller (NIC), which can be connected to other network devices via a base station to enable communication with the Internet. In another embodiment, transmission device 1006 may be a radio frequency (RF) module configured to communicate with the Internet wirelessly.

[0103] The display may be, for example, a touch screen liquid crystal display (LCD), which may enable a user to interact with a user interface of the computer terminal 100 (or electronic device).

[0104] An embodiment of the present application also provides a non-volatile storage medium, which includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the following blood vessel segmentation method by running the computer program: obtaining a CT image of the target tissue; using the first encoder of the target neural network to downsample the CT image to obtain first features of multiple scales corresponding to the CT image, wherein the first encoder includes multiple downsampling layers, and the feature sizes and feature dimensions of the first features output by different downsampling layers are different, and each downsampling layer includes: a feature extraction module and an Inception module, in any downsampling layer, the feature extraction module is used to extract the first features output by the CT image or the previous downsampling layer to obtain intermediate features, and the Inception module is used to parallelly process the intermediate features on different processing branches; using the decoder in the target neural network, the first feature finally output by the first encoder is upsampled to obtain a segmentation result of the CT image, wherein the segmentation result is used to indicate the arteriovenous category of the blood vessels in the CT image.

[0105] An embodiment of the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the blood vessel segmentation method described in each embodiment of the present application: obtaining a CT image of the target tissue; using the first encoder of the target neural network to downsample the CT image to obtain first features of multiple scales corresponding to the CT image, wherein the first encoder includes multiple downsampling layers, and the feature sizes and feature dimensions of the first features output by different downsampling layers are different, and each downsampling layer includes: a feature extraction module and an Inception module, in any downsampling layer, the feature extraction module extracts the first features output by the CT image or the previous downsampling layer to obtain intermediate features, and the Inception module processes the intermediate features in parallel on different processing branches; using the decoder in the target neural network, upsamples the first features finally output by the first encoder to obtain a segmentation result of the CT image, wherein the segmentation result is used to indicate the arteriovenous category of the blood vessels in the CT image.

[0106] The serial numbers of the above embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0107] In the above embodiments of the present application, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0108] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only exemplary. For example, the division of the units can be a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0109] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0110] In addition, each function unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software function unit.

[0111] When the integrated unit is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in the form of a contribution to the prior art, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present application. The foregoing storage medium includes various media that can store program codes, such as a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, etc.

[0112] The above is only the preferred embodiment of the present application, and it should be pointed out that, for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A blood vessel segmentation method, characterized in that: include: Acquire a CT image of the target tissue; A first encoder of a target neural network is used to downsample the CT image to obtain first features of multiple scales corresponding to the CT image, wherein the first encoder includes multiple downsampling layers, and the feature sizes and feature dimensions of the first features output by different downsampling layers are different. Each downsampling layer includes: a feature extraction module and an Inception module. In any of the downsampling layers, the feature extraction module is used to extract the first features output by the CT image or the previous downsampling layer to obtain intermediate features, and the Inception module is used to parallelly process the intermediate features on different processing branches; a decoder in the target neural network is used to upsample the first features finally output by the first encoder to obtain a segmentation result of the CT image, wherein the segmentation result is used to indicate the arteriovenous category of the blood vessels in the CT image; Among them, different processing branches correspond to different convolution kernels; the Inception module includes: a first Inception module and a second Inception module, wherein the convolution step corresponding to the convolution kernel of each processing branch in the first Inception module is 1, and the processing branch of the second Inception module contains the convolution kernel whose convolution step is not 1, and each downsampling layer contains at least one first Inception module.

2. The blood vessel segmentation method according to claim 1, characterized in that: The parallel processing of the intermediate features on different processing branches by the Inception module includes: The Inception module is used to segment the intermediate features according to the number of processing branches they contain, and the segmented intermediate features are respectively output to different processing branches in the Inception module for convolution processing to obtain the first features corresponding to the current downsampling layer, wherein each convolution kernel learns a different feature category, and the feature category includes at least one of the following: edge, texture, color, and HU value.

3. The blood vessel segmentation method according to claim 2, characterized in that: The target neural network includes at least one Transformer encoder; the method further includes: Using the Transformer encoder, processing the first feature output by the target downsampling layer to obtain a second feature, and transmitting the second feature to a specified downsampling layer connected to the Transformer encoder, wherein the second feature includes processing results from different feature output layers of the Transformer encoder, the number of feature output layers in the Transformer encoder is the same as the number of processing branches in the Inception module, and the target downsampling layer is any downsampling layer in the first encoder except the last downsampling layer; In the designated downsampling layer, the first feature output by the previous downsampling layer is fused with the second feature output by the Transformer encoder, and the first feature output by the current downsampling layer is determined based on the fused feature.

4. The blood vessel segmentation method according to claim 3, characterized in that: The step of fusing, at the designated downsampling layer, the first feature output by the previous downsampling layer with the second feature output by the Transformer encoder, and determining, based on the fused features, the first feature output by the current downsampling layer, includes: In a case where the designated downsampling layer includes a plurality of sequentially connected first Inception modules, the first Inception module connected to the feature extraction module contained in the designated downsampling layer fuses the intermediate features from the feature extraction module with the second features, and transmits the fused features to the next first Inception module, and the next first Inception module continues to fuse the fused features with the second features until the first features output by the designated downsampling layer are determined based on the fused features output by the last first Inception module; When the designated downsampling layer includes a first Inception module and a second Inception module connected sequentially, the first Inception module fuses the intermediate features from the feature extraction module with the second features, and transmits the fused features to the second Inception module. The second Inception module continues to fuse the fused features with the second features, and determines the first feature output by the designated downsampling layer based on the fused features transmitted by the second Inception module.

5. The blood vessel segmentation method according to claim 4, characterized in that: The Inception module further includes: a feature fusion module; in the designated downsampling layer, fusing the first feature output by the previous downsampling layer with the second feature output by the Transformer encoder, including: Using the feature fusion module, the target features are divided into N groups, where N is the number of feature output layers of each Transformer encoder, and the target features include the intermediate features or the fused features from any of the Inception modules; The divided N groups of target features are respectively fused with the processing results output by each layer of the Transformer encoder to obtain N groups of fused features.

6. The blood vessel segmentation method according to any one of claims 1 to 5, characterized in that: The training steps of the target neural network include: Acquire an initial training data set, wherein the training data set includes a plurality of training images, the training images are CT images of target tissues, the CT images include target labels, and the target labels are used to indicate arteries and veins in the images; Performing a preprocessing operation on the training images in the initial training dataset to obtain a target training dataset, wherein the preprocessing operation includes at least one of the following: a cropping operation, a normalization operation, a resampling operation for unifying voxel spacing of each training image, and a mirroring operation; The target training data set is used to train the initial neural network to obtain the target neural network, wherein the model structures of the initial neural network and the target neural network are consistent.

7. A blood vessel segmentation device, characterized in that: include: An image acquisition module, used for acquiring a CT image of a target tissue; a feature extraction module, configured to use a first encoder of a target neural network to downsample the CT image to obtain first features at multiple scales corresponding to the CT image, wherein the first encoder includes multiple downsampling layers, and the feature sizes and feature dimensions of the first features output by different downsampling layers are different. Each downsampling layer includes: a feature extraction module and an Inception module. In any downsampling layer, the feature extraction module extracts the first features output by the CT image or the previous downsampling layer to obtain intermediate features, and the Inception module processes the intermediate features in parallel on different processing branches; an image segmentation module, configured to use a decoder in the target neural network to upsample the first feature finally output by the first encoder to obtain a segmentation result of the CT image, wherein the segmentation result is used to indicate the arteriovenous category of the blood vessels in the CT image; Among them, different processing branches correspond to different convolution kernels; the Inception module includes: a first Inception module and a second Inception module, wherein the convolution step corresponding to the convolution kernel of each processing branch in the first Inception module is 1, and the processing branch of the second Inception module contains the convolution kernel whose convolution step is not 1, and each downsampling layer contains at least one first Inception module.

8. An electronic device, characterized in that: include: A memory and a processor, wherein the processor is configured to run a program stored in the memory, wherein the blood vessel segmentation method according to any one of claims 1 to 6 is executed when the program is run.

9. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the blood vessel segmentation method according to any one of claims 1 to 6 are implemented.

Images