A method and system for automatic segmentation of honeycomb lung CT images

Abstract

The present invention discloses a method and system for automatically segmenting honeycomb lung CT images, relating to the field of image segmentation technology. The method comprises: preprocessing CT image data, establishing a SAA-Unet model comprising an encoder, a decoder, and skip connections, training the model, updating model parameters, obtaining a trained SAA-Unet model, performing performance evaluation on the trained SAA-Unet model, obtaining a final SAA-Unet model, and outputting a segmented CT image with a honeycomb lung lesion region. The present invention helps improve the network's ability to describe high-level features of honeycomb lung lesion regions in CT images, obtain more detailed features, and achieve better segmentation performance.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation, and in particular to a method and system for automatically segmenting honeycomb lung CT images. Background Art

[0002] Interstitial diseases (IPDs) are a group of conditions characterized by extensive fibrosis of the lung interstitium. Interstitial pulmonary fibrosis can destroy alveolar structures and dilate bronchi, resulting in numerous air sac-containing cavities. Because of its "honeycomb" appearance, researchers refer to it as the honeycomb pattern of interstitial lung disease, also known as "honeycomb lung." Its imaging features include irregular shapes, uneven grayscale, complex textures, and complex deformations. The most common of these fibrotic lung diseases is idiopathic pulmonary fibrosis (IPF), a disease characterized by a gradual decline in lung function. Early diagnosis and treatment of honeycomb lung is key to improving cure rates, but its early clinical symptoms are subtle. It often appears as tiny honeycomb patterns on imaging, making it easily overlooked by physicians during clinical diagnosis. Screening for early symptoms of honeycomb lung is the best way to prevent the disease. Computed tomography (CT) is currently the most commonly used method for early diagnosis. CT images clearly demonstrate subtle changes in the distribution and severity of honeycomb lung. In the treatment and follow-up of pulmonary fibrosis, accurate quantification of these changes will help evaluate the efficacy of anti-fibrotic therapy and predict the clinical prognosis of patients with interstitial lung disease.

[0003] Currently, the evaluation of honeycomb lungs relies primarily on the visual judgment of radiologists. This subjective visual evaluation method relies heavily on the physician's clinical experience and ability to recognize the symptom, resulting in only a qualitative analysis of the disease and failing to achieve accurate quantitative analysis. Previous research has shown that when different radiologists evaluate the same lesion image, the consistency of visual evaluation results is poor, both between and within observers. Therefore, in current clinical practice, there is an urgent need for an intelligent diagnostic method that is independent of the physician's subjective experience and can provide quantitative, objective, and repeatable evaluation and analysis of honeycomb lungs. To address this clinical problem, timely segmentation of the lesion area from CT images showing honeycomb signs is crucial. As the number of patients with honeycomb lungs continues to increase, the number of patient-related CT images of honeycomb lungs has also increased exponentially.

[0004] Therefore, it is an urgent problem for those skilled in the art to propose an automatic segmentation method and system for honeycomb lung CT images to solve the problems of low segmentation accuracy, large loss of feature diversity and easy redundancy in the existing technology. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for automatic segmentation of honeycomb lung CT images, which can achieve the effect of improving segmentation accuracy.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for automatically segmenting honeycomb lung CT images comprises the following steps:

[0008] Data acquisition steps: obtaining honeycomb lung CT images;

[0009] Data preprocessing steps: The obtained honeycomb lung CT images were cropped, and the cropped images were noised using a Gaussian function. The noisy images were divided into training and test datasets.

[0010] Model construction steps: Build a SAA-Unet model including encoder, decoder and skip connections;

[0011] Model training steps: Input the training dataset into the SAA-Unet model, output the predicted image of the honeycomb lung lesion area, use the output predicted image of the honeycomb lung lesion area and the actual honeycomb lung lesion area image to perform error analysis through the loss function, change the model weight value, and obtain the trained SAA-Unet model;

[0012] Model testing step: Input the test data set into the trained SAA-Unet model to determine whether the error between the output honeycomb lung CT image and the actual honeycomb lung CT image reaches the preset threshold. If yes, return to the model training step; if not, end the model testing step to obtain the tested SAA-Unet model; evaluate the tested SAA-Unet model and obtain the final SAA-Unet model based on the evaluation indicators;

[0013] Honeycomb lung CT image segmentation steps: The honeycomb lung CT image to be segmented is input into the final SAA-Unet model, and the segmented CT image with honeycomb lung lesion area is output.

[0014] In the above method, optionally, in the model construction step, the SAA-Unet model includes a 4-layer encoder structure, a 4-layer decoder structure, and a 4-layer skip connection channel;

[0015] Each encoder layer consists of a convolutional layer and a segmentation-attention module, which is used to extract features from the image and obtain simple features to obtain image feature maps;

[0016] Each decoder layer consists of a convolutional layer, an upsampling module, and an attention module, which is used to classify the pixels in the acquired image feature map and obtain deep feature restoration images;

[0017] The first jump connection channel connects the first encoder and the fourth decoder, fuses the shallow features with the deep features, obtains pixel position information and weight information, and generates the final pixel position encoding information.

[0018] In the above method, optionally, the segmentation-attention module includes a computational unit consisting of a set of feature maps and segmentation-attention operations.

[0019] The above method is optional, and the attention module specifically includes:

[0020] 1) The spatial information of the feature map output by the upsampling operation is extracted through the convolution operation and transmitted through the ReLU activation function to generate an attention gating signal to control the importance of features at different spatial locations;

[0021] 2) The feature map and attention gating signal output by the encoder at the same layer are simultaneously transmitted to the attention gating block for convolution operation, which fuses the structural information of the downsampling layer and the spatial information of the current layer into feature information;

[0022] 3) Use the Sigmoid function to normalize the feature map obtained after the convolution operation to obtain areas with strong correlation and generate an attention weight coefficient of 0 to 1;

[0023] 4) Multiply the attention weight coefficient by the input feature to get the final output.

[0024] In the above method, optionally, the upsampling module performs feature mapping using upsampling and convolution operations.

[0025] In the above method, the loss function in the model training step optionally uses the cross entropy loss function to measure the information difference between the ground truth and the predicted distribution, which is expressed as:

[0026] in, p i and y i are the binary predicted segmentation voxels and the ground truth voxels, N is the number of voxels; the cross entropy loss function directly measures the similarity of samples, and returns the weight coefficient in the model update decoder when the deviation exceeds the threshold.

[0027] A honeycomb lung CT image automatic segmentation system, applying any one of the above-mentioned honeycomb lung CT image automatic segmentation methods, comprising a data acquisition module, a data preprocessing module, a model construction module, a model training module, a model testing module and a honeycomb lung CT image segmentation module connected in sequence;

[0028] Data acquisition module: obtain honeycomb lung CT images;

[0029] Data preprocessing module: The obtained honeycomb lung CT images are cropped, noise is added to the cropped images using a Gaussian function, and the noisy images are divided into training and test datasets;

[0030] Model construction module: Build the SAA-Unet model including encoder, decoder and skip connection;

[0031] Model training module: The training dataset is input into the SAA-Unet model, and a predicted image of the honeycomb lung lesion area is output. The output predicted image of the honeycomb lung lesion area and the actual honeycomb lung lesion area image are used to perform error analysis through a loss function, and the model weight value is changed to obtain a trained SAA-Unet model.

[0032] Model testing module: Input the test data set into the trained SAA-Unet model and determine whether the error between the output honeycomb lung CT image and the actual honeycomb lung CT image reaches the preset threshold. If yes, return to the model training step; if not, end the model testing step and obtain the tested SAA-Unet model; evaluate the tested SAA-Unet model and obtain the final SAA-Unet model based on the evaluation indicators;

[0033] Honeycomb lung CT image segmentation module: The honeycomb lung CT image to be segmented is input into the final SAA-Unet model, and the segmented CT image with honeycomb lung lesion area is output.

[0034] It can be seen from the above technical solution that compared with the existing technology, the present invention provides a method and system for automatic segmentation of honeycomb lung CT images, which has the following beneficial effects: 1) The present invention proposes a segmentation attention module, which enables the network to learn more detailed features; 2) The present invention designs an attention mechanism to focus on channel information, learn the importance of each channel, realize the recalibration of feature channel responses, and improve the high-level feature description capability of the network; 3) The present invention has a slight improvement in segmentation evaluation indicators such as the IOU coefficient, and can achieve better segmentation performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0036] Figure 1This is a flow chart of a method for automatically segmenting honeycomb lung CT images disclosed in the present invention;

[0037] Figure 2 This is a structural diagram of the SAA-Unet model disclosed in this embodiment;

[0038] Figure 3 This is a structural block diagram of an automatic segmentation system for honeycomb lung CT images disclosed in the present invention;

[0039] Figure 4 This is a structural diagram of the segmentation-attention module disclosed in the present invention;

[0040] Figure 5 This is a structural diagram of the upsampling module disclosed in the invention. DETAILED DESCRIPTION

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0042] In this application, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. The terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or apparatus comprising the element.

[0043] Reference Figure 1 As shown, a method for automatically segmenting honeycomb lung CT images includes the following steps:

[0044] Data acquisition steps: obtaining honeycomb lung CT images;

[0045] Data preprocessing steps: The obtained honeycomb lung CT images were cropped, and the cropped images were noised using a Gaussian function. The noisy images were divided into training and test datasets.

[0046] Model construction steps: Build a SAA-Unet model including encoder, decoder and skip connections;

[0047] Model training steps: Input the training dataset into the SAA-Unet model, output honeycomb lung CT images, use the output honeycomb lung lesion area prediction image and the actual honeycomb lung lesion area image to perform error analysis using the loss function, change the model weight value, and obtain the trained SAA-Unet model;

[0048] Model testing step: Input the test data set into the trained SAA-Unet model to determine whether the error between the output honeycomb lung CT image and the actual honeycomb lung CT image reaches the preset threshold. If yes, return to the model training step; if not, end the model testing step to obtain the tested SAA-Unet model; evaluate the tested SAA-Unet model and obtain the final SAA-Unet model based on the evaluation indicators;

[0049] Honeycomb lung CT image segmentation steps: The honeycomb lung CT image to be segmented is input into the final SAA-Unet model, and the segmented CT image with honeycomb lung lesion area is output.

[0050] Further, refer to Figure 2 As shown in the figure (SA is the segmentation-attention module, A is the attention module, and UP is the upsampling module), in the model construction step, the SAA-Unet model includes a 4-layer encoder structure, a 4-layer decoder structure and a 4-layer skip connection channel;

[0051] Each encoder layer consists of a convolutional layer and a segmentation-attention module, which is used to extract features from the image and obtain simple features to obtain image feature maps;

[0052] Each decoder layer consists of a convolutional layer, an upsampling module, and an attention module, which is used to classify the pixels in the acquired image feature map and obtain deep feature restoration images;

[0053] The first jump connection channel connects the first encoder and the fourth decoder, fuses the shallow features with the deep features, obtains pixel position information and weight information, and generates the final pixel position encoding information.

[0054] Furthermore, the segmentation-attention module includes a set of computational units consisting of feature maps and segmentation-attention operations.

[0055] Furthermore, the attention module specifically includes:

[0056] 1) The spatial information of the feature map output by the upsampling operation is extracted through the convolution operation and transmitted through the ReLU activation function to generate an attention gating signal to control the importance of features at different spatial locations;

[0057] 2) The feature map and attention gating signal output by the encoder at the same layer are simultaneously transmitted to the attention gating block for convolution operation, which fuses the structural information of the downsampling layer and the spatial information of the current layer into feature information;

[0058] 3) Use the Sigmoid function to normalize the feature map obtained after the convolution operation to obtain areas with strong correlation and generate an attention weight coefficient of 0 to 1;

[0059] 4) Multiply the attention weight coefficient by the input feature to get the final output.

[0060] Furthermore, the upsampling module performs feature mapping using upsampling and convolution operations.

[0061] Furthermore, the loss function in the model training step uses the cross entropy loss function to measure the information difference between the ground truth and the predicted distribution, which is expressed as:

[0062]

[0063] in, p i and y i are the binary predicted segmentation voxels and the ground truth voxels, N is the number of voxels; the cross entropy loss function directly measures the similarity of samples, and returns the weight coefficient in the model update decoder when the deviation exceeds the threshold.

[0064] and Figure 1 Corresponding to the above method, the embodiment of the present invention also provides a honeycomb lung CT image automatic segmentation system for Figure 1 The specific implementation of the method is as follows: a schematic diagram of the structure of a honeycomb lung CT image automatic segmentation system provided by an embodiment of the present invention is as follows Figure 3 As shown, it includes a data acquisition module, a data preprocessing module, a model building module, a model training module, a model testing module and a honeycomb lung CT image segmentation module connected in sequence;

[0065] Data acquisition module: obtain honeycomb lung CT images;

[0066] Data preprocessing module: The obtained honeycomb lung CT images are cropped, noise is added to the cropped images using a Gaussian function, and the noisy images are divided into training and test datasets;

[0067] Model construction module: Build the SAA-Unet model including encoder, decoder and skip connection;

[0068] Model training module: The training dataset is input into the SAA-Unet model, and honeycomb lung CT images are output. The output honeycomb lung lesion area prediction image and the actual honeycomb lung lesion area image are used to perform error analysis through the loss function, and the model weight value is changed to obtain the trained SAA-Unet model.

[0069] Model testing module: Input the test data set into the trained SAA-Unet model and determine whether the error between the output honeycomb lung CT image and the actual honeycomb lung CT image reaches the preset threshold. If yes, return to the model training step; if not, end the model testing step and obtain the tested SAA-Unet model; evaluate the tested SAA-Unet model and obtain the final SAA-Unet model based on the evaluation indicators;

[0070] Honeycomb lung CT image segmentation module: The honeycomb lung CT image to be segmented is input into the final SAA-Unet model, and the segmented CT image with honeycomb lung lesion area is output.

[0071] In a specific embodiment, the following are included:

[0072] The data acquisition process involved collecting CT data from a large number of patients with honeycomb lung, taken using different CT scanners at Bethune Hospital in Shanxi Province. The patient data were collected from a single center to avoid significant discrepancies between the acquired data. Subsequently, experiments using the constructed honeycomb lung database were used to train and test the network. The dataset includes 2,350 CT images of honeycomb lung with image dimensions of 1896×878, 1896×790, 1684×1006, 1684×1006, 4684×1054, and 1684×878. All CT image data is anonymized. Each honeycomb lung image was annotated by an experienced thoracic radiologist. This was done to prevent different radiologists from evaluating the same lesion image, which could lead to inconsistent evaluations and affect the model's segmentation performance. The database provides significant variation in the shape and size of honeycomb lungs.

[0073] The data preprocessing steps include: first, cropping the obtained honeycomb lung CT images to reduce the impact of irrelevant areas on the model segmentation performance; then, we add noise to the cropped images using a Gaussian function and perform random contrast, brightness, hue, and saturation enhancement on the cropped images to reduce interference from external environmental factors; the above image enhancement methods can increase data capacity to a certain extent, enhance the network's generalization ability, and prevent overfitting problems.

[0074] Furthermore, all images are resized to a fixed size of 512×512 before being fed into the encoder, and finally normalized by subtracting the mean and dividing by the variance to normalize the intensity to [0, 1].

[0075] The SAA-Unet model in the model construction step includes a segmentation-attention module, an attention module, and an upsampling module.

[0076] Specifically, the split-attention module comes from ResNeSt. The split-attention module is a computational unit consisting of a set of feature maps and split-attention operations, which allows features to be mapped between different feature map groups. Figure 4 As shown, in deep neural networks, replacing convolutional modules with parallel stacks of tiny modules with the same topology can improve model feature utilization and accuracy without increasing model parameters. However, due to the small size of some lesions in honeycomb lung CT images, the process of grouping and dividing the number of input and output channels can lead to redundant feature extraction and an increase in model parameters, increasing computer and computational resource consumption and reducing model segmentation accuracy. Therefore, the cardinality K and radix R of the feature map group are set to 1 and the number of channels in the input feature map, respectively. The total number of feature groups G is R (G = KR).

[0077] In addition, the segmentation-attention module is a comprehensive representation of each red heart group, which can be obtained by element-wise addition across multiple segmentation layers. In this module, the input feature map is sliced ​​into K slices and then into R slices. The feature maps obtained from each slice are aggregated in this module through global averaging to collect global semantic information with channel features. The attention parameters of each slice are calculated by the segmentation attention module, and each segmentation attention module uses the overall features of the R slices in the process of calculating the attention parameters. Finally, the rsofmax function is used to aggregate the features and redistribute the weights of the feature map channels, increasing and decreasing the weights of the lesion area and background area respectively. In this way, the segmentation performance of the model is improved.

[0078] The attention module is based on the attention mechanism, which is essentially to let computers imitate the way humans observe things. Its core purpose is to select information that is more critical to the current task goal from a large amount of information. In a neural network, each input channel does not contain all the target information, and each channel contributes differently to the task. The attention mechanism focuses on the information on the channel, and its task goal is to learn the importance of each channel. Therefore, in order to improve the segmentation performance of the model, we apply the attention mechanism to the basic U-Net segmentation network to achieve attention to prominent areas and suppression of irrelevant background areas. Figure 5As shown in Figure 2, the attention module includes an attention gating signal (AGS) to control the importance of features at different spatial positions. Before merging the features of each resolution of the encoder with the corresponding features of the decoder, AGS is used to rescale the output features of the encoder.

[0079] The attention module learns the finest features through AGS and uses the attention gating block to globally focus on relevant features. First, a convolution operation extracts the spatial information of the feature map output by the upsampling operation and transmits it through the Reluctant Unit (ReLU) activation function to generate an attention gating signal, which is used to control the importance of features at different spatial locations. The feature map output by the encoder at the same layer and the attention gating signal are then simultaneously transmitted to the attention gating block for convolution, fusing the structural information of the downsampling layer with the spatial information of the current layer into feature information. We set the output features of the previous layer to X∈Rc×h×w, where c, h, and w represent the number of channels, height, and width of the previous layer output features, respectively. After convolution operations with Wg and Wl, the resulting features are Xg∈Rc / 2×h / 2×w / 2 and Xl∈Rc / 2×h / 2×w / 2. The fused feature information is then linearly activated using the Reluctant Unit (ReLU) activation function, and the resulting feature map is normalized using the Sigmoid function to identify regions with strong correlations and generate an attention weight coefficient ranging from 0 to 1. Finally, the output attention weight coefficient is multiplied by the input feature to obtain the final output, which makes the value of the irrelevant area smaller and the value of the target area larger, thereby improving the segmentation performance of the model. The attention weight coefficient is as follows:

[0080] ,

[0081] Among them, S is the Sigmoid activation function, R is the ReLU activation function; W, Wg and W l It is a convolution operation with a kernel size of 1×1, and g is the attention gate control signal.

[0082] The upsampling module is used to overcome feature redundancy and improve feature utilization, such as Figure 5As shown in Figure 2, during upsampling, deconvolution is often used directly within the U-Net framework to restore image resolution. However, in practice, deconvolution is prone to uneven overlap, especially when the kernel size is not divisible by the stride. This uneven overlap in the deconvolution operation can lead to redundant features, resulting in reduced feature utilization. While neural networks can theoretically avoid this flaw by carefully learning weight parameters, in reality, they cannot completely avoid it. This is because the overlap occurs in two dimensions. The uneven overlap along the two coordinate axes multiplies each other, resulting in features of varying sizes, similar to a chessboard. To mitigate the checkerboard effect, models typically add a deconvolution with a stride of 1 to the last layer. This has some effect, reducing the frequency of defects like the checkerboard effect, but it still cannot completely eliminate it. To overcome the checkerboard effect, upsampling and convolution operations are introduced to restore the feature map. The core idea of ​​the upsampling module is to replace the deconvolution operation in the decoding stage with upsampling and convolution, reducing information redundancy and expanding mapping capabilities.

[0083] The model training step includes the loss function,

[0084] Specifically, in the training process of the medical image segmentation neural network, cross entropy is uniformly used as the loss function for all experimental algorithms because cross entropy measures the information difference between the ground truth and the predicted distribution. The calculation formula of cross entropy is as follows.

[0085]

[0086] in, p i and y i are the binary predicted segmentation voxels and the ground truth voxels, N is the number of voxels. It directly measures the similarity between two samples and optimizes the network weights accordingly by minimizing the loss. Using the cross-entropy loss function for network training effectively avoids the problem of network training being unable to continue when the gradient is too small, ensuring that the neural network weights converge to their true values ​​during training.

[0087] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple. For relevant parts, refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0088] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for automatic segmentation of honeycomb lung CT images, characterized by: Data acquisition steps: obtaining honeycomb lung CT images; Data preprocessing steps: The obtained honeycomb lung CT images were cropped, and the cropped images were noised using a Gaussian function. The noisy images were divided into training and test datasets. Model construction steps: Build a SAA-Unet model including encoder, decoder and skip connections; Model training steps: The training dataset is input into the SAA-Unet model, and honeycomb lung CT images are output. The output honeycomb lung lesion area prediction map and the actual honeycomb lung lesion area map are used to perform error analysis through the loss function, update the model parameters, and obtain the trained SAA-Unet model. Model testing step: Input the test data set into the trained SAA-Unet model and determine whether the error between the output honeycomb lung lesion area prediction map and the actual honeycomb lung lesion area map reaches a preset threshold. If yes, return to the model training step; if not, end the model testing step and obtain the tested SAA-Unet model; evaluate the tested SAA-Unet model and obtain the final SAA-Unet model based on the evaluation indicators; Honeycomb lung CT image segmentation step: input the honeycomb lung CT image to be segmented into the final SAA-Unet model, and output a CT image with the segmented honeycomb lung lesion area; In the model construction step, the SAA-Unet model includes a 4-layer encoder structure, a 4-layer decoder structure and a 4-layer skip connection channel; Each encoder layer consists of a convolutional layer and a segmentation-attention module, which is used to extract features from the image and obtain simple features to obtain image feature maps; Each decoder layer consists of a convolutional layer, an upsampling module, and an attention module, which is used to classify the pixels in the acquired image feature map and obtain deep feature restoration images; The first skip connection channel connects the first encoder and the fourth decoder, fuses the shallow features with the deep features, obtains pixel position information and weight information, and generates the final pixel position encoding information; Note that the modules specifically include: 1) The spatial information of the feature map output by the upsampling operation is extracted through the convolution operation and transmitted through the ReLU activation function to generate an attention gating signal to control the importance of features at different spatial locations; 2) The feature map and attention gating signal output by the encoder at the same layer are simultaneously transmitted to the attention gating block for convolution operation, which fuses the structural information of the downsampling layer and the spatial information of the current layer into feature information; 3) Use the Sigmoid function to normalize the feature map obtained after the convolution operation to obtain areas with strong correlation and generate an attention weight coefficient of 0 to 1; 4) Multiply the attention weight coefficient by the input feature to get the final output.

2. The method for automatic segmentation of honeycomb lung CT images according to claim 1, characterized in that: The segmentation-attention module includes a set of computational units consisting of feature maps and segmentation-attention operations.

3. The method for automatic segmentation of honeycomb lung CT images according to claim 1, characterized in that: The upsampling module uses upsampling and convolution operations to perform feature mapping.

4. The method for automatic segmentation of honeycomb lung CT images according to claim 1, characterized in that: The loss function in the model training step uses the cross entropy loss function to measure the information difference between the ground truth and the predicted distribution. The expression is: in, p i and y i are the binary predicted segmentation voxels and the ground truth voxels, N is the number of voxels; the cross entropy loss function directly measures the similarity of samples, and returns the weight coefficient in the model update decoder when the deviation exceeds the threshold.

5. An automatic segmentation system for honeycomb lung CT images, characterized in that: A method for automatically segmenting a honeycomb lung CT image according to any one of claims 1 to 4 is applied, comprising a data acquisition module, a data preprocessing module, a model construction module, a model training module, a model testing module, and a honeycomb lung CT image segmentation module connected in sequence; Data acquisition module: obtain honeycomb lung CT images; Data preprocessing module: The obtained honeycomb lung CT images are cropped, noise is added to the cropped images using a Gaussian function, and the noisy images are divided into training and test datasets; Model construction module: Build the SAA-Unet model including encoder, decoder and skip connection; Model training module: The training dataset is input into the SAA-Unet model, and a predicted image of the honeycomb lung lesion area is output. The output predicted image of the honeycomb lung lesion area and the actual honeycomb lung lesion area image are used to perform error analysis through a loss function, and the model parameters are updated to obtain a trained SAA-Unet model. Model testing module: Input the test data set into the trained SAA-Unet model and determine whether the error between the output honeycomb lung CT image and the actual honeycomb lung CT image reaches the preset threshold. If yes, return to the model training step; if not, end the model testing step and obtain the tested SAA-Unet model; evaluate the tested SAA-Unet model and obtain the final SAA-Unet model based on the evaluation indicators; Honeycomb lung CT image segmentation module: The honeycomb lung CT image to be segmented is input into the final SAA-Unet model, and the segmented CT image with honeycomb lung lesion area is output; In the model construction module, the SAA-Unet model includes a 4-layer encoder structure, a 4-layer decoder structure, and a 4-layer skip connection channel; Each encoder layer consists of a convolutional layer and a segmentation-attention module, which is used to extract features from the image and obtain simple features to obtain image feature maps; Each decoder layer consists of a convolutional layer, an upsampling module, and an attention module, which is used to classify the pixels in the acquired image feature map and obtain deep feature restoration images; The first skip connection channel connects the first encoder and the fourth decoder, fuses the shallow features with the deep features, obtains pixel position information and weight information, and generates the final pixel position encoding information; Note that the modules specifically include: 1) The spatial information of the feature map output by the upsampling operation is extracted through the convolution operation and transmitted through the ReLU activation function to generate an attention gating signal to control the importance of features at different spatial locations; 2) The feature map and attention gating signal output by the encoder at the same layer are simultaneously transmitted to the attention gating block for convolution operation, which fuses the structural information of the downsampling layer and the spatial information of the current layer into feature information; 3) Use the Sigmoid function to normalize the feature map obtained after the convolution operation to obtain areas with strong correlation and generate an attention weight coefficient of 0 to 1; 4) Multiply the attention weight coefficient by the input feature to get the final output.

Images