Trachea segmentation method based on connectivity analysis

Abstract

This disclosure provides a tracheal segmentation method based on connectivity analysis. The tracheal segmentation method based on connectivity analysis includes: acquiring a human lung CT image to be tested and a human lung CT image dataset; inserting a local attention mechanism module into a pre-defined convolutional neural network to determine a tracheal segmentation neural network; training the tracheal segmentation neural network based on the human lung CT image dataset to determine a tracheal segmentation neural network prediction model; inputting the human lung CT image to be tested into the tracheal segmentation neural network prediction model to determine the tracheal segmentation prediction result. Through this disclosure, the tracheal segmentation neural network prediction model is determined, thereby improving the robustness and accuracy of tracheal segmentation, increasing the connectivity of the tracheal segmentation prediction result, and thus enhancing its auxiliary reference value for clinical diagnosis and treatment.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a tracheal segmentation method based on connectivity analysis. Background Technology

[0002] Medical image segmentation is a common task in computer vision and medical image processing. It involves separating pixels belonging to various organs or lesion regions from the background of medical images. Common medical image modalities include CT, MRI, and X-rays. Medical image segmentation can be categorized into two types based on the segmentation target: the first type involves segmenting human organs or tissues from medical images, such as liver segmentation, heart segmentation, and trachea segmentation; the second type involves segmenting lesion regions from medical images, such as lung nodules and tumors. The results of medical image segmentation are used to assist doctors in the diagnosis and treatment process, improving the efficiency of medical visits.

[0003] Tracheal segmentation has significant clinical value in bronchoscopic intervention and lung disease diagnosis. Ensuring the connectivity and topological integrity of the segmentation results is crucial. However, manually labeling tracheal regions is time-consuming, labor-intensive, and requires medical expertise, making it unsuitable for practical applications. Traditional image processing-based tracheal segmentation algorithms rely on the accuracy of growth criteria. Due to the complexity of the tracheal structure, selecting a suitable growth criterion is difficult, potentially leading to leakage in the segmentation results. Deep learning techniques applied to tracheal segmentation lack consideration for tracheal connectivity, easily resulting in segmentation breaks and hindering accurate segmentation of terminal bronchioles. These breaks can mislead the diagnosis of lung diseases and reduce the efficiency of clinical interventional procedures. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a tracheal segmentation method based on connectivity analysis.

[0005] According to a first aspect of this disclosure, a tracheal segmentation method based on connectivity analysis is provided, comprising: Obtain the human lung CT image to be tested and the human lung CT image dataset, and divide the sample images in the human lung CT image dataset into an image training set, an image verification set and an image test set according to a preset ratio; A local attention mechanism module is inserted into a pre-defined convolutional neural network to determine the trachea segmentation neural network; The trachea segmentation neural network is trained based on the human lung CT image dataset to determine the trachea segmentation neural network prediction model. The CT image of the lungs to be tested is input into the trachea segmentation neural network prediction model to determine the trachea segmentation prediction result.

[0006] Optionally, inserting a local attention mechanism module into a preset convolutional neural network to determine the trachea segmentation neural network includes: A local attention mechanism module is inserted into each upsampling module of a pre-defined 3D U-Net convolutional neural network to determine the tracheal segmentation neural network.

[0007] Optionally, the step of training the trachea segmentation neural network based on the human lung CT image dataset to determine the trachea segmentation neural network prediction model includes: The training samples in the image training set divided by the human lung CT image dataset are preprocessed; The preprocessed training samples are input into the trachea segmentation neural network, and the trachea segmentation results are output to determine the trachea segmentation neural network model. Based on the actual labeling of the trachea and the trachea segmentation results, the tubular structure loss function is determined; The trachea segmentation neural network model is optimized based on the tubular structure loss function, and the model parameters of the trained trachea segmentation neural network model are optimized to convergence.

[0008] Optionally, the optimization of the trachea segmentation neural network model based on the tubular structure loss function optimizes the model parameters of the trained trachea segmentation neural network model until convergence. The minimization of the model parameters is defined as follows: in, This represents the parameters of the trachea segmentation neural network model. This indicates the number of training samples used in each model training iteration. This indicates the volume loss between the trachea segmentation result and the actual labeling of the trachea. This indicates the tubular structure loss between the trachea segmentation result and the actual labeling of the trachea. Represents the loss function of the tubular structure The corresponding weights.

[0009] Optionally, determining the volume loss between the trachea segmentation result and the actual trachea label includes: in, This indicates the result of the tracheal segmentation. This indicates that the trachea is actually labeled. This indicates the introduction of parameters.

[0010] Optionally, the step of inputting the preprocessed training samples into the trachea segmentation neural network and outputting trachea segmentation results includes: Determine the preset local neighborhood of the local attention mechanism module; The preprocessed training samples are input into the local attention mechanism module, and the aggregated features of each pixel of the training samples are determined according to the preset local neighborhood. Based on the aggregated features of each pixel in the training sample, determine the aggregated feature image of the training sample; The trachea segmentation result is determined based on the aggregated feature image.

[0011] Optionally, the step of inputting the preprocessed training samples into the local attention mechanism module, and determining the aggregated features of each pixel of the training samples based on the preset local neighborhood, includes: The training samples are input into the local attention mechanism module to determine a first feature map, a second feature map, and a third feature map, wherein the first feature map, the second feature map, and the third feature map are images of the same size; Based on the first feature map, the second feature map, and the third feature map, the pixel features corresponding to each pixel point on each feature map, the first feature region of the preset local neighborhood centered on each pixel point on each feature map, and the second feature region of the preset local neighborhood centered on each pixel point on each feature map are determined respectively. A local attention image is determined based on the pixel features corresponding to each pixel and the first feature region; Based on the local attention image and the second feature region, the aggregated feature of each pixel of the training sample is determined.

[0012] Optionally, determining the local attention image based on the pixel features corresponding to each pixel and the first feature region includes: in, The coordinates of the local attention image The pixels at each location are used to characterize the local similarity between the pixel features corresponding to each pixel point in each feature map and the first feature region. Represents the coordinates of the first feature region The features corresponding to the pixels at that location. This represents the pixel feature corresponding to each pixel point in each feature map. This represents the preset local neighborhood.

[0013] Optionally, determining the aggregated features of each pixel of the training sample based on the local attention image and the second feature region includes: in, The coordinates of the training sample are represented by ( Aggregation features of pixels at position ) This indicates that the second feature region is located at coordinates ( The features of the pixel at ().

[0014] Optionally, determining the trachea segmentation result based on the aggregated feature image includes: in, This represents the feature image used to generate the trachea segmentation result. Represents the balance coefficient. This represents the aggregated feature image. This refers to the preprocessed training samples.

[0015] The technical solution provided in this disclosure may include the following beneficial effects: By acquiring a dataset of human lung CT images, a trachea segmentation neural network is trained to determine the trachea segmentation neural network prediction model. This deep learning model enables the segmentation neural network prediction model to effectively recover the features of the input image. Inserting a local attention mechanism module into a pre-defined convolutional neural network reduces the occurrence of tracheal fragmentation and improves the connectivity of the trachea segmentation results, enabling the segmentation of a well-structured human trachea tree from lung CT images and improving the efficiency of clinical disease diagnosis. The human lung CT image to be tested is then input into the trachea segmentation neural network prediction model to determine the trachea segmentation prediction result, improving image processing efficiency and simplifying the operation steps.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a tracheal segmentation method based on connectivity analysis according to an exemplary embodiment.

[0019] Figure 2 This is a flowchart illustrating a method for determining a tracheal segmentation neural network prediction model according to an exemplary embodiment.

[0020] Figure 3 This is a flowchart illustrating a method for determining tracheal segmentation results according to an exemplary embodiment.

[0021] Figure 4 This is a flowchart illustrating an exemplary embodiment for determining the aggregated features of each pixel in a training sample. Detailed Implementation

[0022] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0023] It should be noted that all actions involving the acquisition of signals, information, or data in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization granted by the owner of the relevant device.

[0024] The tracheal segmentation method based on connectivity analysis can be applied to the surgical treatment of lung diseases, such as bronchoscopy.

[0025] Figure 1 This is a flowchart illustrating a tracheal segmentation method based on connectivity analysis, according to an exemplary embodiment. Figure 1 As shown, a tracheal segmentation method based on connectivity analysis includes: S11, acquire the human lung CT image to be tested and the human lung CT image dataset, and divide the sample images in the human lung CT image dataset into image training set, image validation set and image test set according to a preset ratio.

[0026] The preset ratio for dividing the human lung CT image dataset can be set according to the actual model building requirements, and this disclosure does not impose specific limitations.

[0027] The image training set is used to train the target model and build a model that meets the work requirements. The image validation set is used to validate the model trained on the image training set. The model with the best parameters is selected from the models built on the image training set as the final working model. The image test set is used to test the working performance of the final working model.

[0028] In this disclosure, the obtained human lung CT image dataset contains 70 sample images, of which 40 sample images are determined as the image training set, 10 sample images are determined as the image validation set, and 20 sample images are determined as the image test set.

[0029] Among them, the number of axial slices in each sample image is Between these, each slice is 512×512 in size, and the layer thickness of each sample image is between Between, the voxel spacing of each sample image is .

[0030] S12, Insert a local attention mechanism module into the preset convolutional neural network to determine the trachea segmentation neural network.

[0031] In some possible embodiments, this disclosure may employ a preset 3D U-Net convolutional neural network, inserting a local attention mechanism module into each upsampling module of the preset 3D U-Net convolutional neural network to determine the trachea segmentation neural network.

[0032] The 3D U-Net convolutional neural network can detect more terminal bronchioles, repair tracheal ruptures, and improve tracheal connectivity. Those skilled in the art should understand that the default convolutional neural network can also employ other deep learning-based neural networks, all of which fall within the protection scope of this disclosure.

[0033] Among them, the local attention mechanism module can aggregate similar features within a preset neighborhood to suppress the rupture phenomenon in the trachea and build a trachea segmentation neural network.

[0034] S13, The trachea segmentation neural network is trained based on the human lung CT image dataset to determine the trachea segmentation neural network prediction model.

[0035] The training process includes model training, validation, and testing.

[0036] The trachea segmentation neural network model was trained using training samples from a dataset of human lung CT images. The model was then validated using validation samples from a validation set. Multiple hyperparameters were set for the trachea segmentation neural network, and the validation samples were input into these models with different hyperparameters. The model with the optimal hyperparameters for trachea segmentation was determined and selected as the trachea segmentation neural network prediction model.

[0037] Test samples from the image test set are input into the trachea segmentation neural network prediction model for testing. The trachea segmentation neural network prediction model outputs the trachea segmentation results corresponding to the test samples, and the segmentation accuracy and connectivity of the trachea segmentation neural network prediction model are evaluated according to the evaluation index.

[0038] In this disclosure, the tracheal segmentation neural network prediction model is implemented using the PyTorch 1.7 framework and the computer hardware is an NVIDIA GeForce RTX 3090 GPU.

[0039] S14, input the CT image of the human lung to be tested into the trachea segmentation neural network prediction model to determine the trachea segmentation prediction result.

[0040] In some possible embodiments, the CT image of the human lung to be tested is preprocessed. Preprocessing can be done by cropping or normalizing the image to reduce the image segmentation range and improve the accuracy of tracheal segmentation.

[0041] The preprocessed CT image of the human lungs to be tested is input into the trachea segmentation neural network prediction model, and the trachea segmentation prediction result is output.

[0042] The above technical solution trains a trachea segmentation neural network using the acquired human lung CT image dataset to determine the trachea segmentation neural network prediction model. This deep learning model enables the segmentation neural network prediction model to effectively recover the features of the input image. Inserting a local attention mechanism module into the pre-defined convolutional neural network reduces the occurrence of tracheal fragmentation and improves the connectivity of the trachea segmentation results, enabling the segmentation of a well-structured human trachea tree from lung CT images and improving the efficiency of clinical disease diagnosis. The human lung CT image to be tested is then input into the trachea segmentation neural network prediction model to determine the trachea segmentation prediction result, improving image processing efficiency and simplifying the operation steps.

[0043] Figure 2 This is a flowchart illustrating a method for determining a neural network prediction model for tracheal segmentation according to an exemplary embodiment. Figure 2 As shown, in some possible embodiments, the step of training the trachea segmentation neural network based on the human lung CT image dataset to determine the trachea segmentation neural network prediction model includes S21 to S24.

[0044] S21, preprocess the training samples in the image training set divided by the human lung CT image dataset.

[0045] Preprocessing of training samples can be achieved through methods such as cropping, normalization, and data augmentation to narrow the image segmentation range of the training samples, enhance tracheal features, and improve the accuracy of tracheal segmentation.

[0046] Those skilled in the art should understand that the preprocessing described in this disclosure can also be performed in other ways, all of which fall within the protection scope of this disclosure. This disclosure does not specifically limit the methods of preprocessing.

[0047] As one possible embodiment, the training samples in the image training set are preprocessed. In this disclosure, the HU value of the training samples is truncated to the range of [-1000, 600] and normalized to [0, 1]. Image blocks of size 128×112×304 are cropped from the training samples through center sampling. Feature enhancement is performed by data augmentation methods such as flipping or rotating the image blocks.

[0048] S22, input the preprocessed training samples into the trachea segmentation neural network, output the trachea segmentation results, and determine the trachea segmentation neural network model.

[0049] As one possible implementation, when training the trachea segmentation neural network model using preprocessed training samples, the optimizer used is Adam, the training process is 30 generations, the batch size during training is set to 2, and the loss function is... We set the value to 0.05 and the learning rate to 0.002.

[0050] S23. Determine the tubular structure loss function based on the actual labeling of the trachea and the tracheal segmentation results.

[0051] In one possible embodiment, in a 3D image, the trachea segmentation result of the training sample is determined as P, and the actual trachea label is Y. Key points are identified in the actual trachea label Y, where the key points include the branch points and the endpoints of the trachea. The connectivity between each key point is determined, adjacent and connected key point pairs are marked, and the marked key point pairs are mapped to the trachea segmentation result P.

[0052] Among them, for key point pairs ( , Voxel loss = 1 - voxel probability value.

[0053] Key points in the tracheal segmentation result P To the key point Shortest path search to determine key points To the key point The shortest path is found, and the average loss of all voxels along the shortest path is used as the path loss. The path loss for each keypoint pair is then determined, and the average path loss for each keypoint pair is used as the tubular structure loss function. Where P represents the trachea segmentation result, and Y represents the actual labeling of the trachea. Let represent the set of all shortest paths. This indicates one of the key points. To the key point The shortest path, Represents the voxels on the shortest path. This represents the number of voxels in the shortest path set.

[0054] S24. The trachea segmentation neural network model is optimized based on the tubular structure loss function, and the model parameters of the trachea segmentation neural network model after model training are optimized to convergence.

[0055] The optimization process includes minimizing the model parameters by minimizing the loss function of the tubular structure until convergence. The minimization of the model parameters is defined as follows: in, This represents the parameters of the trachea segmentation neural network model. This indicates the number of training samples used in each model training iteration. This indicates the volume loss between the trachea segmentation result and the actual labeling of the trachea. This indicates the tubular structure loss between the trachea segmentation result and the actual labeling of the trachea. Represents the loss function of the tubular structure The corresponding weights.

[0056] In one possible embodiment, determining the volume loss between the trachea segmentation result and the actual trachea label includes: in, This indicates the result of the tracheal segmentation. This indicates that the trachea is actually labeled. This indicates the introduction of parameters.

[0057] The above technical solution employs a tubular structure loss function, and the tracheal segmentation neural network obtained through model training focuses on tubular structure features, thereby improving the connectivity of the tracheal segmentation results.

[0058] Figure 3This is a flowchart illustrating a method for determining tracheotomy results according to an exemplary embodiment. Figure 3 As shown, in some possible embodiments, the step of inputting the preprocessed training samples into the trachea segmentation neural network and outputting the trachea segmentation result includes S31 to S34.

[0059] S31, Determine the preset local neighborhood of the local attention mechanism module.

[0060] As an example, the preset local neighborhood is determined to be N.

[0061] S32, input the preprocessed training samples into the local attention mechanism module, and determine the aggregate features of each pixel of the training samples according to the preset local neighborhood.

[0062] Figure 4 This is a flowchart illustrating an exemplary embodiment for determining the aggregated features of each pixel in a training sample. Figure 4 As shown, in some possible embodiments, S32 includes S41 to S44.

[0063] S41, input the training samples into the local attention mechanism module to determine the first feature map, the second feature map and the third feature map. The first feature map, the second feature map and the third feature map are images of the same size.

[0064] Here, the preprocessed training samples are defined as The pixels in the training samples are defined as ,in, Represents the row coordinates of a pixel. Represents the column coordinates of a pixel.

[0065] The Local Attention Module (LAM) consists of a first convolutional layer, a second convolutional layer, and a third convolutional layer. The LAM is used to aggregate local similar features in the image, thereby enhancing the foreground features at the tracheal fracture site and suppressing the occurrence of tracheal fracture in the tracheal segmentation results.

[0066] The preprocessed training samples are input into the first convolutional layer of the local attention mechanism module, which outputs a first feature map. The preprocessed training samples are input into the second convolutional layer of the local attention mechanism module, which outputs a second feature map. The preprocessed training samples are input into the third convolutional layer of the local attention mechanism module, which outputs a third feature map. The first, second, and third feature maps are images of the same size.

[0067] S42, based on the first feature map, the second feature map and the third feature map, respectively determine the pixel feature corresponding to each pixel point on each feature map, the first feature region of the preset local neighborhood centered on each pixel point on each feature map, and the second feature region of the preset local neighborhood centered on each pixel point on each feature map.

[0068] Following the example above, the pixel features corresponding to each pixel are determined in the first feature map, and the size centered on each pixel is determined in the first feature map. The first feature region, in the first feature map, is defined with a size centered on each pixel. The second characteristic region.

[0069] The pixel feature corresponding to each pixel is defined as follows: coordinates The pixel features corresponding to the pixel at that location; the first feature region is defined as That is, coordinates on the first feature map The pixel at the center has a size of The feature region; the second feature region is defined as That is, coordinates on the first feature map The pixel at the center has a size of The characteristic region.

[0070] The method for determining the pixel features, first feature region, and second feature region corresponding to each pixel in the second feature map and the third feature map is the same as the method for determining the pixel features, first feature region, and second feature region corresponding to each pixel in the first feature map described above, and will not be repeated here.

[0071] S43, determine the local attention image based on the pixel features corresponding to each pixel and the first feature region.

[0072] Following the example above, the pixel features corresponding to each pixel point are... With the first feature region After performing product processing, followed by Softmax operation and normalization, the local attention image is determined: in, Represents a local attention image. The coordinates of the local attention image The pixels at each location are used to characterize the local similarity between the pixel features corresponding to each pixel point in each feature map and the first feature region. Represents the coordinates of the first feature region The features corresponding to the pixels at that location. This represents the pixel feature corresponding to each pixel point in each feature map. This represents the preset local neighborhood.

[0073] S44, based on the local attention image and the second feature region, determine the aggregated features of each pixel of the training sample.

[0074] Following the example above, the local attention image... With the second feature region Perform matrix multiplication to determine the aggregated features of each pixel: in, The coordinates of the training sample are represented by ( Aggregation features of pixels at position ) This indicates that the second feature region is located at coordinates ( The features of the pixel at ().

[0075] in, .

[0076] S33, determine the aggregated feature image of the training sample based on the aggregated features of each pixel of the training sample.

[0077] Following the example above, the aggregated features of each pixel in the training sample are determined according to the above process, and the aggregated features of each pixel are stitched together to form an aggregated feature image.

[0078] S34. Determine the trachea segmentation result based on the aggregated feature image.

[0079] Aggregate feature images and balance coefficients Multiply the samples by the product, then sum them with the preprocessed training samples to determine the output image. , Tracheal segmentation results can be generated by upsampling.

[0080] in, This represents the feature image used to generate the trachea segmentation result. Represents the balance coefficient. This represents the aggregated feature image. This refers to the preprocessed training samples.

[0081] By incorporating the local attention mechanism module into a pre-defined convolutional neural network, a tracheal segmentation neural network is constructed. When performing tracheal segmentation on the CT image of the human lungs to be tested, the local attention mechanism module aggregates similar features within a pre-defined local neighborhood to enhance the foreground features at the tracheal fracture site, thereby suppressing the tracheal fracture phenomenon during the tracheal cutting process. Furthermore, the connectivity of the trachea is enhanced through a tubular structure loss function, improving the accuracy of the tracheal segmentation results and enhancing the clinical diagnostic assistance capability.

[0082] In some possible embodiments, the tracheal segmentation neural network prediction model can be tested using an image test set from a human lung CT image dataset, and evaluation metrics can be determined to evaluate the tracheal segmentation neural network prediction model.

[0083] As an example, evaluation metrics may include the percentage of branches detected (BD), the percentage of tree-length detected (TD), and the Dice coefficient (DSC).

[0084] The detected branch percentage represents the proportion of tracheal branches detected in the tracheal segmentation prediction results to the total number of tracheal branches actually labeled. in, This indicates the number of tracheal branches detected in the tracheal segmentation prediction results. This indicates the actual number of tracheal branches that are labeled on the trachea.

[0085] The detection tree length ratio represents the ratio of the detected tracheal skeleton length in the tracheal segmentation prediction results to the actual tracheal skeleton length labeled with the tracheal segmentation tag. in, This indicates the length of the tracheal skeleton detected in the tracheal segmentation prediction results. This indicates the actual length of the tracheal skeleton as labeled on the trachea.

[0086] The Dice coefficient represents the degree of overlap between the predicted tracheal segmentation results and the actual tracheal labels: in, The set of voxels representing the tracheal segmentation prediction results. This represents the set of voxels representing the actual labels on the trachea. This represents the smoothing factor.

[0087] In this disclosure, tracheal segmentation is performed on the image test set using a 3D U-net convolutional neural network and the tracheal segmentation neural network prediction model determined in this disclosure, respectively. The evaluation metrics are shown in Table 1. Table 1 As shown in Table 1, compared with the 3D U-net convolutional neural network, the tracheal segmentation neural network prediction model disclosed in this paper can significantly improve the connectivity of the tracheal segmentation prediction results, while also ensuring the accuracy of the tracheal segmentation prediction results. The topological structure of the trachea is more complete, which is beneficial for clinical diagnosis and treatment of diseases.

[0088] The embodiments disclosed herein are merely preferred embodiments of the present invention. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the present invention, and are not intended to limit the present invention. Any modifications and variations made by those skilled in the art within the scope of this specification should fall within the protection scope of the present invention.

Claims

1. A tracheal segmentation method based on connectivity analysis, characterized in that, include: Obtain the human lung CT image to be tested and the human lung CT image dataset, and divide the sample images in the human lung CT image dataset into an image training set, an image verification set and an image test set according to a preset ratio; A local attention mechanism module is inserted into a pre-defined convolutional neural network to determine the trachea segmentation neural network; The trachea segmentation neural network is trained based on the human lung CT image dataset to determine the trachea segmentation neural network prediction model. The CT image of the human lung to be tested is input into the trachea segmentation neural network prediction model to determine the trachea segmentation prediction result. The step of training the trachea segmentation neural network based on the human lung CT image dataset to determine the trachea segmentation neural network prediction model includes: The training samples in the image training set divided by the human lung CT image dataset are preprocessed; The preprocessed training samples are input into the trachea segmentation neural network, and the trachea segmentation results are output to determine the trachea segmentation neural network model. Based on the actual labeling of the trachea and the trachea segmentation results, the tubular structure loss function is determined; The trachea segmentation neural network model is optimized based on the tubular structure loss function, and the model parameters of the trained trachea segmentation neural network model are optimized to convergence. The optimization process for the trachea segmentation neural network model based on the tubular structure loss function optimizes the model parameters of the trained trachea segmentation neural network model until convergence. The minimum model parameters are defined as follows: ； in, This represents the parameters of the trachea segmentation neural network model. This indicates the number of training samples used in each model training iteration. This indicates the volume loss between the trachea segmentation result and the actual labeling of the trachea. This indicates the tubular structure loss between the trachea segmentation result and the actual labeling of the trachea. Represents the loss function of the tubular structure The corresponding weights; Determining the volume loss between the trachea segmentation result and the actual trachea label includes: ； in, This indicates the result of the tracheal segmentation. This indicates that the trachea is actually labeled. Indicates the introduction of parameters; Identify key points in the actual tracheal label Y, where key points include the branch points and endpoints of the trachea, determine the connectivity between each key point, mark adjacent and connected key point pairs, and map the marked key point pairs to the tracheal segmentation result P. Among them, for key point pairs ( , ), identify key points To the key point The shortest path is found, and the average path loss for each keypoint pair is used as the tubular structure loss function: ； in, Let represent the set of all shortest paths. This indicates one of the key points. To the key point The shortest path, Represents the voxels on the shortest path. This represents the number of voxels in the shortest path set.

2. The method according to claim 1, characterized in that, The step of inserting a local attention mechanism module into a preset convolutional neural network to determine the trachea segmentation neural network includes: A local attention mechanism module is inserted into each upsampling module of a pre-defined 3D U-Net convolutional neural network to determine the tracheal segmentation neural network.

3. The method according to claim 1, characterized in that, The step of inputting the preprocessed training samples into the trachea segmentation neural network and outputting trachea segmentation results includes: Determine the preset local neighborhood of the local attention mechanism module; The preprocessed training samples are input into the local attention mechanism module, and the aggregated features of each pixel of the training samples are determined according to the preset local neighborhood. Based on the aggregated features of each pixel in the training sample, determine the aggregated feature image of the training sample; The trachea segmentation result is determined based on the aggregated feature image.

4. The method according to claim 3, characterized in that, The step of inputting the preprocessed training samples into the local attention mechanism module and determining the aggregated features of each pixel of the training samples based on the preset local neighborhood includes: The training samples are input into the local attention mechanism module to determine a first feature map, a second feature map, and a third feature map, wherein the first feature map, the second feature map, and the third feature map are images of the same size; Based on the first feature map, the second feature map, and the third feature map, the pixel features corresponding to each pixel point on each feature map, the first feature region of the preset local neighborhood centered on each pixel point on each feature map, and the second feature region of the preset local neighborhood centered on each pixel point on each feature map are determined respectively. A local attention image is determined based on the pixel features corresponding to each pixel and the first feature region; Based on the local attention image and the second feature region, the aggregated feature of each pixel of the training sample is determined.

5. The method according to claim 4, characterized in that, The step of determining the local attention image based on the pixel features corresponding to each pixel and the first feature region includes: ； in, The coordinates of the local attention image The pixels at each location are used to characterize the local similarity between the pixel features corresponding to each pixel point in each feature map and the first feature region. Represents the coordinates of the first feature region The features corresponding to the pixels at that location. This represents the pixel feature corresponding to each pixel point in each feature map. This represents the preset local neighborhood.

6. The method according to claim 5, characterized in that, The step of determining the aggregated feature of each pixel of the training sample based on the local attention image and the second feature region includes: ； in, The coordinates of the training sample are represented by ( Aggregation features of pixels at position ) This indicates that the second feature region is located at coordinates ( The features of the pixel at ().

7. The method according to claim 3, characterized in that, Determining the trachea segmentation result based on the aggregated feature image includes: ； in, This represents the feature image used to generate the trachea segmentation result. Represents the balance coefficient. This represents the aggregated feature image. This represents the preprocessed training samples.

Images