A method and system for lung airway segmentation and repair

Abstract

This invention discloses a method and system for airway segmentation and repair in the lungs. The method includes: A. constructing a neural network to train a coarse segmentation model on CT images; B. uncertainty analysis and candidate region extraction. Multiple airway voxel coarse segmentation probability maps are obtained by repeatedly reasoning about the same CT influence using multiple sets of model weights W. These multiple airway voxel coarse segmentation probability maps are statistically analyzed at the voxel level to obtain an entropy map reflecting prediction consistency. High uncertainty regions Ω are extracted based on an entropy threshold. u ; Calculate the local tube diameter on the coarse segmentation probability map, and obtain the supplementary set Ω based on the different tube diameters and sampling rates of the airways. d The final candidate fracture region is Ω. c Ω c =Ω u ∪Ω d C. Fracture generation; D. Fracture enhancement; E. Fracture repair; This invention can automatically segment the airway tree in lung CT images and repair fractures in small airway segments that are easily missed, thereby improving the integrity and accuracy of airway segmentation.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing, and specifically relates to a method and system for lung airway segmentation and repair. Background Technology

[0002] Segmentation of the pulmonary airway structure plays a crucial role in the diagnosis and treatment of respiratory diseases. For example, in the diagnosis of COPD and asthma, airway morphology and patency are key indicators; in lung surgery planning and bronchoscopic navigation, accurate airway tree models help to develop safe and effective pathways. Therefore, how to automatically and accurately segment the complete airway tree from lung CT images has always been a key focus in the field of medical image processing.

[0003] Currently, deep learning methods have been widely applied to airway segmentation tasks, significantly improving the segmentation performance of main bronchi and larger bronchi. However, for the small airway branches at the lung margins, existing algorithms often have limited effectiveness due to their small diameter, large number of small branches, and complex distribution. Conventional segmentation networks are prone to missing or breaking small airways, resulting in discontinuous airway trees and missing terminal branches in the segmentation results. This phenomenon of missing edge airways and airway breaks affects the accuracy of lung function assessment and disease analysis, limiting the application value of segmentation results in clinical diagnosis and treatment planning.

[0004] Patent application CN 119762419A discloses a two-stage active learning method for fine-grained bronchial airway segmentation. In its retraining process, it selectively filters and discards datasets, ultimately confirming the adoption of the dataset through expert annotation. This method reduces the true uncertainty of airway segmentation and results in low dataset value density. Therefore, there is an urgent need for an improved method that can enhance airway segmentation integrity and repair broken small airways. Summary of the Invention

[0005] The purpose of this invention is to provide a method for airway segmentation and repair in the lungs, which can automatically segment the airway tree in lung CT images and repair broken small airway segments that are easily missed, thereby improving the integrity and accuracy of airway segmentation.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a method for lung airway segmentation and repair, comprising the following steps:

[0007] A. A neural network is constructed to train a coarse segmentation model on CT images. During training, the model parameters are continuously updated in consecutive training rounds. The model weights W are saved every m training iterations for uncertainty analysis. The model weights obtained in the m-th training iteration are Wm. mThe model weights obtained from the 2mth training iteration are W. 2m And so on, where m is a positive integer;

[0008] B. Uncertainty Analysis and Candidate Region Extraction: By repeatedly inferring the same CT influence using multiple sets of model weights W, multiple coarse airway voxel segmentation probability maps can be obtained. These multiple coarse airway voxel segmentation probability maps are statistically analyzed at the voxel level to obtain an entropy map reflecting prediction consistency. High uncertainty regions Ω are extracted based on the entropy threshold. u ; Calculate the local tube diameter on the last coarse segmentation probability map, and obtain the supplementary set Ω based on different tube diameters and sampling rates of the airway. d The final candidate fracture region is Ω. c Ω c =Ω u ∪Ωd;

[0009] C. Fracture formation, in the candidate fracture region Ω c Generate fracture template M break Furthermore, three-dimensional corrosion on the pipe is applied at a scale proportional to the local pipe diameter to simulate structural defects;

[0010] D. Fracture enhancement, using fracture template M break Mapped to lung airway image I raw Generate fracture enhancement image I break ;

[0011] E. Fracture repair, I raw with I break The airway features F are obtained by inputting them into the airway feature encoder. raw F with lung fracture enhancement feature break The multi-source fusion features are obtained, and the multi-source fusion features are decoded step by step to generate a probability map P. coarse and P fine ,P coarse Prediction of segmentation of the main and medium-thickness airways, P fine For specific repair segmentation prediction targeting fracture location and small branches; P coarse and P fine Linear fusion is performed, and the final airway segmentation result is obtained through the Sigmoid activation function.

[0012] In another implementation, step B utilizes multiple sets of model weights {W} m W 2m ... W final} Image of the same lung airway I raw Perform multiple inferences to obtain the probability set {P1, P2, ..., P}. final}, calculate the average probability Then calculate the voxel entropy map:

[0013]

[0014] Voxels in the entropy map that are above the threshold X are marked as high uncertainty regions Ω. u .

[0015] In another implementation, X is greater than 0.5.

[0016] In another implementation, in step B, when calculating the local tube diameter, a 70% sampling rate is used for small airway regions with d(x) ≤ 2 mm, and a 20% sampling rate is used for small airway regions with d(x) > 2 mm in medium-sized airways. An additional 10% of the airways are randomly sampled to obtain a supplementary set Ω. d .

[0017] In another implementation, in step D, the normal region voxel γ norm A contrast mapping of 1.0 was used, with γ applied to voxels in the fracture region. crack Adjusted by a power factor of 0.7, and introducing a smoothing factor β = 0.1, the enhanced formula is as follows:

[0018]

[0019] In another implementation, in step E, the lung airway features F raw F with lung fracture enhancement feature break Spatial fusion feature F is obtained through spatial fusion. space Further enhancement of lung fracture features F break Channel features F are obtained by semantic reweighting through a channel attention mechanism. c Channel feature F c Spatial integration feature F space After channel splicing and convolutional compression, the final multi-source fusion feature is output.

[0020] In another implementation, lung airway characteristics F raw This includes capturing the structural features of the main lung trunk and medium-to-large airways.

[0021] In another implementation, the lung fracture enhancement feature F break Emphasis should be placed on the location of the break and information about the small branch airways.

[0022] In another implementation, lung airway characteristics F raw F with lung fracture enhancement feature break During spatial fusion, spatial alignment and weighted fusion of the two features are achieved through spatial attention weight α(x) to compensate for and correct the positional information deviation of the fracture region.

[0023] In another implementation, lung airway characteristics F rawF with lung fracture enhancement feature break When spatial integration

[0024] First, feature concatenation: combine features F of the same scale. raw With F break By concatenating along the channel dimension, a fused input is generated to obtain F. cat ;

[0025] Reconvolution mapping: for Fc at Perform convolution to obtain the intermediate mapping feature map F A ;

[0026] Then activate normalization: for F A Apply the Sigmoid activation function to generate a spatial attention map α(x);

[0027] Final fusion output: Calculate the final spatial fusion feature F using a spatial weighting strategy. space :

[0028] F space =α(x)⊙F raw +(1-α(x))⊙F break …………(3).

[0029] Another implementation involves generating channel weights β(c) to enhance the lung fracture feature F. break Channel weighting is performed to suppress redundant information and highlight high-discrimination channels associated with small airways.

[0030] Another implementation involves generating channel weights β(c) to enhance the lung fracture feature F. break When performing channel weighting,

[0031] First, channel compression is performed: enhancing the lung fracture feature F break Global average pooling is performed to obtain the channel extraction feature S. c ;

[0032] Then perform feature mapping: extract features S from the channels. c Two fully connected layers are input sequentially, first reducing the dimensionality and then increasing it, with a ReLU activation function embedded in between, to obtain the updated channel features;

[0033] Then activate scaling: apply Sigmoid activation to the updated channel features to obtain the channel weights β(c);

[0034] Finally, feature weighting: the channel weight β(c) is weighted by the lung fracture enhancement feature F. break Multiplying each channel sequentially yields the weighted feature map Fc:

[0035]

[0036] The present invention also provides a system for the above-described method of pulmonary airway segmentation and repair, comprising:

[0037] The coarse segmentation module is configured to continuously update the model parameters during training, saving the model weights W every m training iterations for uncertainty analysis; the model weights obtained in the m-th training iteration are Wm. m The model weights obtained from the 2mth training iteration are W. 2m And so on, where m is a positive integer;

[0038] The airway segmentation prediction module is configured to repeatedly infer the same CT effect using multiple sets of model weights W to obtain multiple coarse airway voxel segmentation probability maps. These multiple coarse airway voxel segmentation probability maps are statistically analyzed at the voxel level to obtain an entropy map reflecting prediction consistency. High uncertainty regions Ω are extracted based on an entropy threshold. u ; Calculate the local tube diameter on the last coarse segmentation probability map, and obtain the supplementary set Ω based on different tube diameters and sampling rates of the airway. d The final candidate fracture region is Ω. c Ω c =Ω u ∪Ω d ;

[0039] The fracture generation module is configured to generate fractures in the candidate fracture region Ω. c Generate fracture template M break Furthermore, three-dimensional corrosion on the pipe is applied at a scale proportional to the local pipe diameter to simulate structural defects;

[0040] Fracture reinforcement module, which is configured to place fracture template M break Mapped to lung airway image I raw Generate fracture enhancement image I break ;

[0041] The fracture repair module is configured to I raw with I break The airway features F are obtained by inputting them into the airway feature encoder. raw F with lung fracture enhancement feature break The multi-source fusion features are obtained, and the multi-source fusion features are decoded step by step to generate a probability map P. coarse and P fine ,P coarse Prediction of segmentation of the main and medium-thickness airways, P fine For specific repair segmentation prediction targeting fracture location and small branches; P coarse and P fine Linear fusion is performed, and the final airway segmentation result is obtained through the Sigmoid activation function.

[0042] The beneficial effects of this invention are as follows: 1. Precise localization: Uncertainty estimation utilizes weighted inference across multiple training rounds, which is more stable and reproducible than random Dropout, and can accurately pinpoint potential missed detection areas; 2. Controllable fracture simulation: The corrosion nucleus is adaptively matched with the airway diameter, avoiding excessive damage to large airways while increasing the learning difficulty for small airways; 3. Efficient fusion: The dual encoder, combined with a channel attention module and a spatial fusion module, takes into account both spatial context and channel discriminability, significantly improving the detection rate of small airways; 4. Topological integrity: The dual-branch decoding strategy simultaneously ensures global continuity and local details. Attached Figure Description

[0043] Figure 1 A flowchart of the lung airway segmentation and repair method in this invention (showing the coarse segmentation network, the break generation module, and the break repair module);

[0044] Figure 2 Schematic diagram of the coarse segmentation stage;

[0045] Figure 3 Fracture generation flowchart;

[0046] Figure 4 Fracture repair flowchart;

[0047] Figure 5 Flowchart of the multi-source feature fusion module algorithm. Detailed Implementation

[0048] The present invention will now be described in detail with reference to the embodiments shown in the accompanying drawings.

[0049] The methods for lung airway segmentation and repair include the following steps:

[0050] A. such as Figure 2 As shown, a coarse segmentation model is obtained by training a neural network using U-Net on CT images. During training, the model parameters are continuously updated in consecutive training rounds. The model weights W are saved once every m training iterations for uncertainty analysis. The model weights obtained in the m-th training iteration are Wm. m The model weights obtained from the 2mth training iteration are W. 2m And so on, to obtain multiple sets of model weights {W}. m W 2m ... W final}, where m is a positive integer; for the same lung airway image I raw Perform multiple inferences to obtain the probability set {P1, P2, ..., P}. final}, calculate the average probability Then calculate the voxel entropy map:

[0051]

[0052] Voxels in the entropy map that are above the threshold X are marked as high uncertainty regions Ω. u X is greater than 0.5, and is taken as 0.6 in this embodiment;

[0053] B. For example Figure 3 As shown, uncertainty analysis and candidate region extraction utilize multiple sets of model weights W to repeatedly infer the influence of the same CT scan, resulting in multiple coarse airway voxel segmentation probability maps. These multiple coarse airway voxel segmentation probability maps are then statistically analyzed at the voxel level to obtain an entropy map reflecting prediction consistency. High uncertainty regions Ω are then extracted based on the entropy threshold. u In the coarse segmentation probability map P final The local tube diameter is calculated above. For the small airway region with d(x)≤2mm, a sampling rate of 70% is used, and for the medium airway region with d(x)>2mm, a sampling rate of 20% is used. An additional 10% of the airways are randomly sampled to obtain the supplementary set Ω. d The final candidate fracture region is Ω. c Ω c =Ω u ∪Ω d ;

[0054] C. For example Figure 3 As shown, fracture formation occurs in the candidate fracture region Ω. c Generate fracture template M break Furthermore, three-dimensional corrosion on the pipe is applied at a scale proportional to the local pipe diameter to simulate structural defects;

[0055] D. such as Figure 3 As shown, fracture enhancement is achieved by using fracture template M. break Mapped to lung airway image I raw Generate fracture enhancement image I break Among them, normal region voxel γ norm A contrast mapping of 1.0 was used, with γ applied to voxels in the fracture region. crack Adjusted by a power factor of 0.7, and introducing a smoothing factor β = 0.1, the enhanced formula is as follows:

[0056]

[0057] E. such as Figure 4 As shown, fracture repair, I raw with I break The airway features F are obtained by inputting them into the airway feature encoder. raw F with lung fracture enhancement feature break The multi-source fusion features are obtained, and the multi-source fusion features are decoded step by step to generate a probability map P. coarse and Pfine ,P coarse Prediction of segmentation of the main and medium-thickness airways, P fine For specific repair segmentation prediction targeting fracture location and small branches; P coarse and P fine Linear fusion, followed by Sigmoid activation function to obtain the final airway segmentation result;

[0058] Specifically: Lung airway characteristics F raw Includes capturing structural features of the main lung trunk and medium-to-large airways; enhanced features of lung fractures (F). break Emphasis should be placed on the location of the break and information on the small branch airways; lung airway characteristics F raw F with lung fracture enhancement feature break Spatial fusion feature F is obtained through spatial fusion. space Further enhancement of lung fracture features F break Channel features F are obtained by semantic reweighting through a channel attention mechanism. c Channel feature F c Spatial integration feature F space After channel splicing and 1×1 convolution compression, the final multi-source fusion feature is output.

[0059] Lung airway characteristics F raw F with lung fracture enhancement feature break During spatial fusion, spatial alignment and weighted fusion of the two feature streams are achieved through spatial attention weight α(x) to compensate for and correct the positional information deviation of the fracture region. Specifically, the steps include:

[0060] a. Feature concatenation: Concatenate the same-scale features Fraw and Fbreak along the channel dimension to generate a fused input, which is Fcat;

[0061] b. Convolutional mapping: Perform a 3×3 convolution on Fcat to obtain the intermediate mapping feature map FA;

[0062] c. Activation Normalization: Apply the Sigmoid activation function to the FA to generate a spatial attention map α(x);

[0063] d. Fusion Output: Calculate the final spatial fusion feature F using a spatial weighting strategy. space :

[0064] F space =α(x)⊙F raw +(1-α(x))⊙F break …………(3).

[0065] Enhancement features of lung fracture F break Channel features F are obtained by semantic reweighting through a channel attention mechanism.c At that time, the lung fracture enhancement feature F was enhanced by generating channel weight β(c). break Channel weighting is performed to suppress redundant information and highlight high-discrimination channels associated with small airways, specifically including the following steps:

[0066] I. Channel Compression: Enhancing Lung Fracture Features F break Global average pooling is performed to obtain the channel extraction feature S. c ;

[0067] II. Feature Mapping: Extracting Features S from Channels c Two fully connected layers are input sequentially, first reducing the dimensionality and then increasing it, with a ReLU activation function embedded in between, to obtain the updated channel features;

[0068] 3. Activation scaling: Perform Sigmoid activation on the updated channel features to obtain the channel weights β(c);

[0069] IV. Feature Weighting: The channel weight β(c) is weighted by the lung fracture enhancement feature F. break Multiplying each channel sequentially yields the weighted feature map Fc:

[0070]

[0071] A system for the above-mentioned airway segmentation and repair method for the lungs includes:

[0072] The coarse segmentation module is configured to continuously update the model parameters during training, saving the model weights W every m training iterations for uncertainty analysis; the model weights obtained in the m-th training iteration are Wm. m The model weights obtained from the 2mth training iteration are W. 2m And so on, where m is a positive integer;

[0073] The airway segmentation prediction module is configured to repeatedly infer the same CT effect using multiple sets of model weights W to obtain multiple coarse airway voxel segmentation probability maps. These multiple coarse airway voxel segmentation probability maps are statistically analyzed at the voxel level to obtain an entropy map reflecting prediction consistency. High uncertainty regions Ω are extracted based on an entropy threshold. u ; Calculate the local tube diameter on the last coarse segmentation probability map, and obtain the supplementary set Ω based on different tube diameters and sampling rates of the airway. d The final candidate fracture region is Ω. c Ω c =Ω u ∪Ω d ;

[0074] The fracture generation module is configured to generate fractures in the candidate fracture region Ω. cGenerate fracture template M break Furthermore, three-dimensional corrosion on the pipe is applied at a scale proportional to the local pipe diameter to simulate structural defects;

[0075] Fracture reinforcement module, which is configured to place fracture template M break Mapped to lung airway image I raw Generate fracture enhancement image I break ;

[0076] The fracture repair module is configured to I raw with I break The airway features F are obtained by inputting them into the airway feature encoder. raw F with lung fracture enhancement feature break The multi-source fusion features are obtained, and the multi-source fusion features are decoded step by step to generate a probability map P. coarse and P fine ,P coarse Prediction of segmentation of the main and medium-thickness airways, P fine For specific repair segmentation prediction targeting fracture location and small branches; P coarse and P fine Linear fusion is performed, and the final airway segmentation result is obtained through the Sigmoid activation function.

[0077] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for lung airway segmentation and repair, characterized in that, It includes the following steps: A. A neural network is constructed to train a coarse segmentation model on CT images. During training, the model parameters are continuously updated in consecutive training rounds. The model weights W are saved every m training iterations for uncertainty analysis. The model weights obtained in the m-th training iteration are Wm. m The model weights obtained from the 2mth training iteration are W. 2m And so on, where m is a positive integer; B. Uncertainty Analysis and Candidate Region Extraction: Using multiple sets of model weights W, repeated inference on the same CT image yields multiple coarse airway voxel segmentation probability maps. These multiple coarse airway voxel segmentation probability maps are statistically analyzed at the voxel level to obtain an entropy map reflecting prediction consistency. High uncertainty regions Ω are extracted based on the entropy threshold. u ; Calculate the local tube diameter on the last coarse segmentation probability map, and obtain the supplementary set Ω based on different tube diameters and sampling rates of the airway. d The final candidate fracture region is Ω. c Ω c =Ω u ∪Ω d ; C. Fracture formation, in the candidate fracture region Ω c Generate fracture template M break Furthermore, three-dimensional corrosion on the pipe is applied at a scale proportional to the local pipe diameter to simulate structural defects; D. Fracture enhancement, using fracture template M break Mapped to lung airway image I raw Generate fracture enhancement image I break ; E. Fracture repair, I raw with I break The airway features F are obtained by inputting them into the airway feature encoder. raw F with lung fracture enhancement feature break The multi-source fusion features are obtained, and the multi-source fusion features are decoded step by step to generate a probability map P. coarse and P fine ,P coarse Prediction of segmentation of the main and medium-thickness airways, P fine For specific repair segmentation prediction targeting fracture location and small branches; P coarse and P fine Linear fusion is performed, and the final airway segmentation result is obtained through the Sigmoid activation function.

2. The method for lung airway segmentation and repair according to claim 1, characterized in that: In step B, multiple sets of model weights {W} are used. m W 2m ... W final } Image I of the same lung airway raw Perform multiple inferences to obtain the probability set {P1, P2, ..., P}. final } Calculate the average probability `P`, and then calculate the voxel entropy map: …………（1）， Voxels in the entropy map that are above the threshold X are marked as high uncertainty regions Ω. u .

3. The method for lung airway segmentation and repair according to claim 2, characterized in that: X is greater than 0.

5.

4. The method for lung airway segmentation and repair according to claim 1, characterized in that: In step B, when calculating the local tube diameter, a 70% sampling rate is used for the small airway region with d(x) ≤ 2 mm, and a 20% sampling rate is used for the small airway region with d(x) > 2 mm. An additional 10% of the airways are randomly sampled to obtain the supplementary set Ω. d .

5. The method for lung airway segmentation and repair according to claim 1, characterized in that: In step D, voxel γ of the normal region is... norm A contrast mapping of 1.0 was used, and γ was applied to the voxels in the fracture region. crack Adjusted by a power factor of 0.7, and with a smoothing factor β = 0.1, the enhanced formula is as follows: …………（2）。 6. The method for lung airway segmentation and repair according to claim 1, characterized in that: In step E, lung airway characteristics F raw F with lung fracture enhancement feature break Spatial fusion feature F is obtained through spatial fusion. space Further enhancement of lung fracture features F break Channel features F are obtained by semantic reweighting through a channel attention mechanism. c Channel feature F c Spatial integration feature F space After channel concatenation and compression, the final multi-source fusion feature is output.

7. The method for lung airway segmentation and repair according to claim 6, characterized in that: Lung airway characteristics F raw This includes capturing the structural features of the main lung trunk and medium-to-large airways.

8. The method for lung airway segmentation and repair according to claim 6, characterized in that: Lung fracture enhancement feature F break Emphasis should be placed on the location of the break and information about the small branch airways.

9. The method for lung airway segmentation and repair according to claim 6, characterized in that: Lung airway characteristics F raw F with lung fracture enhancement feature break During spatial fusion, spatial alignment and weighted fusion of the two features are achieved through spatial attention weight α(x) to compensate for and correct the positional information deviation of the fracture region.

10. The method for lung airway segmentation and repair according to claim 9, characterized in that: Lung airway characteristics F raw F with lung fracture enhancement feature break When spatial integration First, feature concatenation: combine features F of the same scale. raw With F break By concatenating along the channel dimension, a fused input is generated to obtain F. cat ; Reconvolution mapping: for Fc at Perform convolution to obtain the intermediate mapping feature map F A ; Then activate normalization: for F A Apply the Sigmoid activation function to generate spatial attention weights α(x); Final fusion output: Calculate the final spatial fusion feature F using a spatial weighting strategy. space : …………（3）。 11. The method for lung airway segmentation and repair according to claim 9, characterized in that: The lung fracture enhancement feature F is enhanced by generating channel weight β(c). break Channel weighting is performed to suppress redundant information and highlight high-discrimination channels associated with small airways.

12. The method for lung airway segmentation and repair according to claim 11, characterized in that: The lung fracture enhancement feature F is enhanced by generating channel weight β(c). break When performing channel weighting, First, channel compression is performed: enhancing the lung fracture feature F break Perform global average pooling to obtain the channel extraction feature S. c ; Then perform feature mapping: extract features S from the channels. c Two fully connected layers are input sequentially, first reducing the dimensionality and then increasing it, with a ReLU activation function embedded in between, to obtain the updated channel features; Then activate scaling: apply Sigmoid activation to the updated channel features to obtain the channel weights β(c); Finally, feature weighting: the channel weight β(c) is weighted by the lung fracture enhancement feature F. break Multiplying each channel sequentially yields the weighted feature map F. c : …………（4）。 13. A system for the lung airway segmentation and repair method according to any one of claims 1-12, characterized in that, It includes: The coarse segmentation module is configured to continuously update the model parameters during training, saving the model weights W every m training iterations for uncertainty analysis; the model weights obtained in the m-th training iteration are Wm. m The model weights obtained from the 2mth training iteration are W. 2m And so on, where m is a positive integer; The airway segmentation prediction module is configured to repeatedly infer multiple airway voxel coarse segmentation probability maps by using multiple sets of model weights W on the same CT image. These multiple airway voxel coarse segmentation probability maps are statistically analyzed at the voxel level to obtain an entropy map reflecting prediction consistency. High uncertainty regions Ω are extracted based on an entropy threshold. u ; Calculate the local tube diameter on the last coarse segmentation probability map, and obtain the supplementary set Ω based on different tube diameters and sampling rates of the airway. d The final candidate fracture region is Ω. c Ω c =Ω u ∪Ω d ; The fracture generation module is configured to generate fractures in the candidate fracture region Ω. c Generate fracture template M break Furthermore, three-dimensional corrosion on the pipe is applied at a scale proportional to the local pipe diameter to simulate structural defects; Fracture reinforcement module, which is configured to place fracture template M break Mapped to lung airway image I raw Generate fracture enhancement image I break ; The fracture repair module is configured to I raw with I break The airway features F are obtained by inputting them into the airway feature encoder. raw F with lung fracture enhancement feature break The multi-source fusion features are obtained, and the multi-source fusion features are decoded step by step to generate a probability map P. coarse and P fine , P coarse Prediction of segmentation of the main and medium-thickness airways, P fine For specific repair segmentation prediction targeting fracture location and small branches; P coarse and P fine Linear fusion is performed, and the final airway segmentation result is obtained through the Sigmoid activation function.

Images