Pulmonary nodule segmentation method based on superpixel feature fusion

Abstract

The invention belongs to the technical field of image segmentation, and discloses a pulmonary nodule segmentation method based on superpixel feature fusion, and the method comprises the steps: obtaining a reference data set of pulmonary nodule segmentation, carrying out the preprocessing of the data set, and obtaining a preprocessed two-dimensional CT slice image of a pulmonary nodule; constructing a superpixel feature map based on the two-dimensional CT slice image of the pulmonary nodule; constructing a pulmonary nodule segmentation model, wherein the model input is a two-dimensional CT slice image and a superpixel feature map of the pulmonary nodule; the pulmonary nodule segmentation model takes U-Net of a double encoder as a reference network model, and comprises a boundary enhancement superpixel fusion module, a multi-scale non-local module and a bidirectional attention fusion module; and inputting the two-dimensional CT slice image of the pulmonary nodule and the superpixel feature map into the trained pulmonary nodule recognition model to obtain a pulmonary nodule segmentation result. The pulmonary nodule image segmentation method has a good pulmonary nodule image segmentation effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image segmentation, and particularly relates to a lung nodule segmentation method based on superpixel feature fusion. BACKGROUND

[0002] As a key indicator of early imaging of lung cancer, early and accurate diagnosis and treatment of lung nodules are crucial to improve survival rate. Computed tomography (CT) as the core imaging technology for lung nodule and lung cancer screening has irreplaceable clinical value. Traditional lung nodule image segmentation relies on manual annotation by radiologists, which is not only time-consuming and laborious, but also affected by the professional skills and experience of doctors, leading to inconsistent annotation and poor segmentation effect. Therefore, designing an efficient and accurate lung nodule segmentation method is of great significance for pathological analysis and clinical diagnosis.

[0003] In recent years, lung nodule segmentation methods based on deep convolutional neural networks have been rapidly developed. Among them, the U-Net architecture and its variants are widely used in this field. U-Net achieves feature extraction and image reconstruction through the symmetrical structure of the encoder and decoder, and alleviates the information loss problem in the downsampling process through the skip connection, showing superior performance in medical image segmentation (Reference 1: Ronneberger, O., Fischer, P., & Brox, T. (2015). U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention - MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 (pp. 234-241). Springer international publishing.). To further improve the trainability and expression ability of deep networks, ResUNet effectively alleviates the gradient vanishing and performance degradation problems in deep neural networks by introducing residual connections, enabling the successful training of very deep network structures and possessing stronger feature expression ability (Reference 2: Diakogiannis, F. I., Waldner, F., Caccetta, P., & Wu, C. (2020). ResUNet-a: A deep learning framework for semantic segmentation of remotely sensed data. ISPRS Journal of Photogrammetry and Remote Sensing, 162, 94-114.). With the widespread application of the Transformer architecture in the field of computer vision, some research attempts to combine it with U-Net to improve the model's ability to model global dependencies.Li et al. proposed CTBP-Net, which introduced a Cross-Transformer module and combined a bidirectional pyramid structure, effectively integrating local and global features and improving the segmentation accuracy of lung nodules (Reference 3: Li, X., Jiang, A., Wang, S., Li, F., & Yan, S. (2023). CTBP-Net: Lung nodule segmentation model based on the cross-transformer and bidirectional pyramid. Biomedical Signal Processing and Control, 82, 104528.). In addition, the role of boundary features in fine segmentation tasks has also been increasingly valued. Wang et al. proposed EANet, which enhanced boundary information extraction through an iterative edge attention mechanism, significantly improving the problem of fuzzy boundaries (Reference 4: Wang, K., Zhang, X., Zhang, X., Lu, Y., Huang, S., & Yang, D. (2022). EANet: Iterative edge attention network for medical image segmentation. Pattern Recognition, 127, 108636.).

[0004] Although existing methods have made some progress in lung nodule segmentation tasks, due to the diverse shapes and sizes of lung nodules, and the similarity of their image features to surrounding tissues, there are still the following technical difficulties that affect the overall performance of the model and its practical application:

[0005] (1) There is a problem of feature loss in the downsampling process. Current mainstream network structures generally use multi-level downsampling operations to extract high-level semantic features, but this process inevitably leads to a decrease in spatial resolution, especially causing significant loss of boundary and structural information for small-sized lung nodules, affecting the precision and accuracy of segmentation.

[0006] (2) Local receptive fields limit the model's ability to model multi-scale information. Traditional segmentation networks represented by U-Net mostly use standard convolution operations, which have limited receptive fields and can only extract local features, making it difficult to effectively capture global semantic information of lung nodules with scale differences, thus restricting the model's ability to represent multi-scale structures.

[0007] (3) The deep and shallow feature fusion mode has a semantic conflict problem. The simple splicing mode is used to fuse the deep and shallow features between the encoder and the decoder in the U-Net structure, which alleviates the information loss to a certain extent, but ignores the semantic inconsistency between the deep semantic features and the shallow edge details, resulting in interference of the fused features and affecting the accuracy and robustness of the segmentation result. SUMMARY

[0008] The present application aims at the problems of inaccurate segmentation and fuzzy boundary of the existing lung nodule segmentation method, and proposes a lung nodule segmentation method based on superpixel feature fusion.

[0009] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0010] A lung nodule segmentation method based on superpixel feature fusion comprises:

[0011] Step 1: Obtain a reference data set for lung nodule segmentation, and pre-process the data set to obtain a pre-processed two-dimensional CT slice image of the lung nodule.

[0012] Step 2: Construct a superpixel feature map based on the two-dimensional CT slice image of the lung nodule.

[0013] Step 3: Construct a lung nodule segmentation model, and the input of the model is the two-dimensional CT slice image of the lung nodule and the superpixel feature map; the lung nodule segmentation model takes a double-encoder U-Net as a reference network model, and comprises a boundary-enhanced superpixel fusion module, a multi-scale non-local module and a bidirectional attention fusion module; the boundary-enhanced superpixel fusion module is used to adaptively fuse the lung nodule image features extracted by the double encoder and the superpixel feature map; the multi-scale non-local module is used to fuse feature maps of different scales layer by layer, and introduce a non-local attention mechanism to model long-distance dependency; the bidirectional attention fusion module is used to combine the spatial attention of the shallow features and the channel attention of the deep features, so as to realize complementary fusion of the deep and shallow features.

[0014] Step 4: Input the two-dimensional CT slice image of the lung nodule and the superpixel feature map into the trained lung nodule recognition model to obtain a lung nodule segmentation result.

[0015] Further, the step 1 comprises:

[0016] The three-dimensional CT image of the lung nodule is converted into a two-dimensional CT slice image, and the image size is adjusted, then the image and the label are format-converted, and intensity normalization processing is performed to obtain the pre-processed two-dimensional CT slice image of the lung nodule.

[0017] Further, the step 2 comprises:

[0018] The simple linear iterative clustering algorithm is used for superpixel segmentation of the two-dimensional CT slice image of the lung nodule, to obtain a superpixel region composed of pixels with similar color and texture features; for each superpixel region, the average value of the internal pixels is calculated, and the average value is assigned to all pixels in the region to generate a superpixel feature map.

[0019] Further, the lung nodule recognition model, when constructed, embeds a boundary-enhanced superpixel fusion module in each layer of the encoder of the benchmark network model.

[0020] Further, in the boundary-enhanced superpixel fusion module:

[0021] First, the reverse attention is used to calculate the reverse weight of the superpixel feature S i , and the reverse weight is multiplied by the original feature X i to obtain the enhanced boundary feature X' i :

[0022] X i ′=(1-Sigmoid(S i ))⊙X i ,i∈{1,2,3,4}

[0023] Where ⊙ represents a pixel-by-pixel multiplication operation, and i represents the encoder layer index;

[0024] Meanwhile, the superpixel feature S i is subjected to a convolution operation to obtain the processed superpixel feature S' i :

[0025] S i ′=(ReLU(BN(Conv(S i )))),i∈{1,2,3,4}

[0026] Finally, X' i , S' i and X i are added pixel by pixel to obtain the final fusion feature F i :

[0027] F i =X′ i +S i ′+X i ,i∈{1,2,3,4}。

[0028] Further, in the multi-scale non-local module:

[0029] First, F2, F3, and F4 are upsampled and mapped to the same feature space as F1 through convolution, batch normalization, and ReLU activation function:

[0030] F i1 =Conv(BN(ReLU(Upsample(F i ))),i=2,3,4

[0031] wherein F1,F2,F3,F4 represent the first layer to the fourth layer encoder output feature map respectively, F i1 represents the feature obtained after the i-th layer encoder output feature map is mapped to the same feature space as F1, BN(.) represents batch normalization, ReLU(.) represents ReLU activation function, and Upsample(.) represents up-sampling;

[0032] Then, the element-wise multiplication is used to obtain the fusion feature F'1:

[0033] F'1=F1⊙F 21 ⊙F 31 ⊙F 41

[0034] wherein F 21 , F 31 , F 41 represent the features obtained after the second layer to the fourth layer encoder output feature map is mapped to the same feature space as F1, and ⊙ represents pixel-by-pixel multiplication operation;

[0035] F1' is spliced with the feature F 2_out processed by the second layer, and then the non-local attention mechanism is introduced to enhance the global dependence, and finally the residual connection is introduced with the input feature, and the formula is as follows:

[0036] F concat =Conv(Concat(F'1,F 2_out ))

[0037] F 1_out =NLB(F concat )+F concat

[0038] wherein F concdt represents the result of convolution processing of the spliced feature F'1 and F 2_out .

[0039] Further, in the bidirectional attention fusion module:

[0040] First, the deep layer feature F H is reversely weighted to obtain the feature F HThen, the deep layer feature and the shallow layer feature are processed by a standard convolution block respectively to obtain processed feature maps; then, the processed shallow layer feature and the deep layer feature are added element by element; next, the channel attention mechanism is used to weight the feature after element-wise addition, and the obtained channel attention is multiplied with the processed shallow layer feature of the standard convolution block element by element to obtain a feature S1; at the same time, the spatial attention mechanism is applied to the processed shallow layer feature of the standard convolution block, and the obtained spatial attention is multiplied with the feature after element-wise addition element by element to obtain a feature D1; finally, the F H , S1 and D1 are added element by element to obtain the final fusion feature.

[0041] Further, in the bidirectional attention fusion module, the deep layer feature F H is reversely weighted in the following manner to obtain a feature F H :

[0042]

[0043] Compared with the prior art, the present application has the beneficial effects that:

[0044] The present application uses an improved multi-encoder U-net network to segment the lesions of the lung nodule image. On the basis of the original network structure, a boundary-enhanced superpixel fusion module, a multi-scale non-local module and a bidirectional attention fusion module are constructed, which significantly improves the segmentation performance and robustness of the model. Through the boundary-enhanced superpixel fusion module, the lung nodule boundary information is effectively extracted, and the problem of small target feature loss in the downsampling process is alleviated. The multi-scale non-local module is introduced to realize cross-level feature fusion and non-local modeling, enhance the recognition ability of the model to different scale nodules, and improve the representation ability of complex structure background. The bidirectional attention fusion module combines the spatial attention and channel attention mechanisms, bridges the semantic gap between deep and shallow layer features, effectively alleviates the semantic conflict and information redundancy problem in feature fusion. Experiments prove that compared with other mainstream models, the performance of the present application in lung nodule image segmentation is better. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 A basic flowchart of a lung nodule segmentation method based on superpixel feature fusion provided for an embodiment of the present application is shown in the figure;

[0046] Figure 2 A lung nodule segmentation model network structure diagram provided for an embodiment of the present application is shown in the figure;

[0047] Figure 3 A boundary-enhanced superpixel fusion module (BESF) structure diagram provided for an embodiment of the present application is shown in the figure;

[0048] Figure 4A schematic diagram of the multi-scale nonlocal module (MSNL) structure provided in an embodiment of the present invention;

[0049] Figure 5 This is a schematic diagram of the bidirectional attention fusion module (BAF) provided in an embodiment of the present invention. Detailed Implementation

[0050] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments:

[0051] like Figure 1 As shown, a lung nodule segmentation method based on superpixel feature fusion includes:

[0052] S101: Obtain the baseline dataset for lung nodule segmentation, and preprocess the dataset to obtain the preprocessed two-dimensional CT slice images of lung nodules.

[0053] S102: Constructing superpixel feature maps based on 2D CT slice images of lung nodules.

[0054] S103: Construct a lung nodule segmentation model (an improved multi-encoder U-net network). The model input consists of 2D CT slice images of lung nodules and superpixel feature maps. The lung nodule segmentation model uses a dual-encoder U-Net as the baseline network model, including a boundary enhancement superpixel fusion module (BESF), a multi-scale nonlocal module (MSNL), and a bidirectional attention fusion module (BAF). The boundary enhancement superpixel fusion module adaptively fuses the lung nodule image features extracted by the dual encoders with the superpixel feature maps to enhance boundary information and improve spatial semantic expression. The multi-scale nonlocal module fuses feature maps of different scales layer by layer and introduces a nonlocal attention mechanism to model long-distance dependencies, thereby improving the model's ability to discriminate complex background regions. The bidirectional attention fusion module combines spatial attention of shallow features and channel attention of deep features to achieve complementary fusion of shallow and deep features and alleviate the problem of inconsistent semantic information. The network structure of the lung nodule segmentation model is as follows: Figure 2 As shown, E and SE represent dual encoders, and D represents a decoder.

[0055] S104: Input the two-dimensional CT slice image and superpixel feature map of the lung nodule into the trained lung nodule recognition model to obtain the lung nodule segmentation result.

[0056] The constructed boundary enhancement superpixel fusion module adaptively fuses lung nodule image feature maps and superpixel feature maps to enhance boundary features and improve the expressive power of spatial semantic information. The multi-scale nonlocal module fuses feature maps from different encoder layers layer by layer and introduces a nonlocal attention mechanism to enhance the model's global context awareness. The bidirectional attention fusion module optimizes key spatial information in shallow features through a spatial attention mechanism, strengthens semantic expression in deep features through a channel attention mechanism, and employs a deep feature back-weighting mechanism to suppress interference with shallow features, thereby improving the ability to focus on key regions.

[0057] Furthermore, S101 includes:

[0058] Download the public dataset from the Lung Image Database Consortium and the Image Database Resource Initiative (LIDC). This dataset includes 1018 chest CT scans from 1010 patients diagnosed with lung cancer. The 3D CT images of lung nodules were converted into 2D CT slice images, and the image size was adjusted to 96×96. Subsequently, the images and labels were formatted to meet the requirements of neural network input, and intensity normalization was performed to obtain preprocessed 2D CT slice images of lung nodules.

[0059] Furthermore, S102 includes:

[0060] Superpixel segmentation was performed on the preprocessed 2D CT slice images of lung nodules using the Simple Linear Iterative Clustering (SLIC) algorithm. For each generated superpixel region, the average value of all pixels within that region was calculated, and this value was used to replace the pixels within the region, thereby constructing a superpixel feature map, denoted as S.

[0061] Furthermore, when constructing the lung nodule recognition model, a boundary enhancement superpixel fusion module is embedded in each layer (dual) encoder of the baseline network model.

[0062] Furthermore, superpixel features S are extracted from each layer of the dual encoder in the baseline network model. i and original CT image features X i The structural details of the boundary enhancement superpixel fusion module are as follows: Figure 3 As shown. First, reverse attention is used to calculate the superpixel features S. i The inverse weights are then used in conjunction with the original feature X. i Multiplying yields the enhanced boundary feature X'. i As shown in formula (1), where ⊙ represents the pixel-wise dot product operation and i represents the encoder layer index.

[0063] X i= (1 - Sigmoid(S i )) o X i , i e {1, 2, 3, 4} (1)

[0064] To further refine the semantic information in the superpixel feature S i , a convolution operation is performed on the superpixel feature S i , and the processed superpixel feature is denoted as S' i . The formula is as follows.

[0065] S i ' = (ReLU(BN(Conv(S i ))), i e {1, 2, 3, 4} (2)

[0066] Finally, X' i , S' i , and X i are added pixel by pixel to obtain the final fusion feature F i , as shown in formula (3)

[0067] F i = X i ' + S i ', i e {1, 2, 3, 4} (3)

[0068] Further, the multi-scale non-local module (as shown in Figure 4 ) fuses feature maps of different scales layer by layer and introduces a non-local attention mechanism to model long-distance dependencies. Taking the first layer encoder output feature map F1 as an example, the remaining three layer encoder output feature maps (F2, F3, F4) are fused, and the feature space resolution is reduced step by step.

[0069] First, F2, F3, and F4 are upsampled (Upsample), and are mapped to the same feature space as F1 through convolution (Conv3x3), BN (batch normalization), and ReLU activation function, as shown in the following expression:

[0070] F i1 = Conv(BN(ReLU(Upsample(F i ))), for i = 2, 3, 4 (4)

[0071] where F2, F3, and F4 represent the second to fourth layer encoder output feature maps, and F i1 represents the feature obtained after the i-th layer encoder output feature map is mapped to the same feature space as F1.

[0072] Then, the fusion feature F′1 is obtained by multiplying element by element, as shown in formula (5).

[0073] F′1=F1⊙ F 21 ⊙F 31 ⊙F 41 (5)

[0074] Where F 21 F 31 F 41 These represent the features obtained after mapping the output feature maps of the second to fourth layers of encoders to the same feature space as F1.

[0075] Compare F1' with the feature F already processed in the second layer. 2_out The data is concatenated and processed, then a non-local attention (NLB) mechanism is introduced to enhance global dependencies, and finally concatenated with the input feature residuals, as shown in the following formula:

[0076] F concat =Conv(Concat(F′1,F 2_out (6)

[0077] F 1_out =NLB(F concat )+F concat (7)

[0078] Where F concat Represent F′1 and F 2_out The result of convolution processing on the concatenated features.

[0079] Furthermore, such as Figure 5 As shown, in the bidirectional attention fusion module, firstly, the deep features F are... H Perform inverse weighting to obtain feature F H Next, the deep and shallow features are processed separately using standard convolutional blocks to obtain processed feature maps. Then, the processed shallow features are summed element-wise with the deep features. Next, a channel attention mechanism is used to weight the element-wise summed features, and the resulting channel attention is multiplied element-wise with the shallow features processed by the standard convolutional block; simultaneously, a spatial attention mechanism is applied to the shallow features processed by the standard convolutional block, and the resulting spatial attention is multiplied element-wise with the element-wise summed features. Finally, the results of these three steps are summed element-wise to obtain the final fused feature. The specific calculation formula is as follows:

[0080]

[0081] S1 = CBR(F L )×(CA(CBR(F H )+CBR(FL ))) (9)

[0082] D1=(CBR(F H )+CBR(F L ))×(SA(CBR(F L ))) (10)

[0083] Out=S1+D1+F′ H (11)

[0084] Wherein, F L ,F H respectively represent shallow features and deep features, CBR represents Conv3x3 convolution, batch normalization and RELU activation function, CA represents channel attention mechanism, and SA represents spatial attention mechanism.

[0085] The improved multi-encoder U-Net network is used for lesion segmentation of the lung nodule image. On the basis of the original network structure, a boundary enhancement superpixel fusion module, a multi-scale non-local module and a bidirectional attention fusion module are constructed, which significantly improves the segmentation performance and robustness of the model. Through the boundary enhancement superpixel fusion module, the lung nodule boundary information is effectively extracted, and the problem of small target feature loss in the down-sampling process is alleviated. The multi-scale non-local module is introduced to realize cross-level feature fusion and non-local modeling, enhance the recognition ability of the model to different scale nodules, and improve the representation ability of complex structure background. The bidirectional attention fusion module combines spatial attention and channel attention mechanism, fills the semantic gap between deep and shallow features, effectively alleviates the semantic conflict and information redundancy problem in feature fusion. The experimental results are shown in Table 1, and according to the comparison of the experimental results, compared with other mainstream models, the performance of the present application on lung nodule image segmentation is better, and the task of clinical auxiliary diagnosis can be effectively completed.

[0086] Table 1 Comparison with other networks on LIDC-IDRI validation dataset

[0087]

[0088]

[0089] The above only shows the preferred embodiments of the present application, and it should be noted that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A lung nodule segmentation method based on superpixel feature fusion, characterized in that, include: Step 1: Obtain the baseline dataset for lung nodule segmentation and preprocess the dataset to obtain the preprocessed 2D CT slice images of lung nodules; Step 2: Construct superpixel feature maps based on 2D CT slice images of lung nodules; Step 3: Construct a lung nodule segmentation model. The model input consists of a 2D CT slice image of a lung nodule and a superpixel feature map. The lung nodule segmentation model uses a dual-encoder U-Net as the base network model, including a boundary enhancement superpixel fusion module, a multi-scale nonlocal module, and a bidirectional attention fusion module. The boundary enhancement superpixel fusion module is used to adaptively fuse the lung nodule image features extracted by the dual encoder with the superpixel feature map. The multi-scale nonlocal module is used to fuse feature maps of different scales layer by layer and introduces a nonlocal attention mechanism to model long-distance dependencies. The bidirectional attention fusion module is used to combine the spatial attention of shallow features and the channel attention of deep features to achieve complementary fusion of shallow and deep features. Step 4: Input the two-dimensional CT slice image and superpixel feature map of the lung nodule into the trained lung nodule recognition model to obtain the lung nodule segmentation result.

2. The lung nodule segmentation method based on superpixel feature fusion according to claim 1, characterized in that, Step 1 includes: The three-dimensional CT images of lung nodules were converted into two-dimensional CT slice images, and the image size was adjusted. Then, the image and label were converted into different formats and intensity was normalized to obtain the preprocessed two-dimensional CT slice images of lung nodules.

3. The lung nodule segmentation method based on superpixel feature fusion according to claim 1, characterized in that, Step 2 includes: A simple linear iterative clustering algorithm was used to perform superpixel segmentation on two-dimensional CT slice images of lung nodules to obtain superpixel regions composed of pixels with similar color and texture features. For each superpixel region, the average value of its internal pixels was calculated and assigned to all pixels in the region to generate a superpixel feature map.

4. The lung nodule segmentation method based on superpixel feature fusion according to claim 1, characterized in that, When constructing the lung nodule identification model, a boundary enhancement superpixel fusion module is embedded in each encoder layer of the baseline network model.

5. The lung nodule segmentation method based on superpixel feature fusion according to claim 1, characterized in that, In the boundary enhancement superpixel fusion module: First, reverse attention is used to compute superpixel features S. i The inverse weights are then used in conjunction with the original feature X. i Multiplying yields the enhanced boundary feature X'. i : X′ i =(1-Sigmoid(S i ))⊙X i ,i∈{1,2,3,4} Where ⊙ represents a pixel-wise dot product operation, and i represents the encoder layer index; Simultaneously, the superpixel feature S i Perform a convolution operation to obtain the processed superpixel features S' i : S′ i =(ReLU(BN(Conv(S i )))),i∈{1,2,3,4} Finally, X' i S' i With X i Pixel-by-pixel addition and fusion are performed to obtain the final fused feature F. i : F i =X′ i +S′ i +X i ,i∈{1,2,3,4}。 6. The lung nodule segmentation method based on superpixel feature fusion according to claim 1, characterized in that, In the multi-scale nonlocal module: First, F2, F3, and F4 are upsampled and mapped to the same feature space as F1 through convolution, batch normalization, and ReLU activation function: F i1 =Conv(BN(ReLU(Upsample(F i )))),i=2,3,4 Where F1, F2, F3, and F4 represent the output feature maps of the encoders from the first to the fourth layer, respectively. i1 This represents the feature obtained after mapping the output feature map of the i-th encoder to the same feature space as F1. BN(.) represents batch normalization, ReLU(.) represents the ReLU activation function, and Upsample(.) represents upsampling. Then, element-wise multiplication is used to obtain the fused feature F′1: F′1=F1⊙F 21 ⊙F 31 ⊙F 41 Where F 21 F 31 F 41 These represent the features obtained after mapping the output feature maps of the second to fourth layers of encoders to the same feature space as F1, and ⊙ represents the pixel-wise dot product operation. Compare F1' with the feature F already processed in the second layer. 2_out The data is concatenated and processed, then a non-local attention mechanism is introduced to enhance global dependencies, and finally connected to the input feature residuals, as shown in the following formula: F concat =Conv(Concat(F′1,F 2_out )) FAVORITE 1_out =NLB(F concat )+F concat Where F concat Represent F′1 and F 2_out The result of convolution processing on the concatenated features.

7. The lung nodule segmentation method based on superpixel feature fusion according to claim 1, characterized in that, In the bidirectional attention fusion module: First, for deep features F H Perform inverse weighting to obtain feature F H Next, deep and shallow features are processed separately using standard convolutional blocks to obtain processed feature maps. Then, the processed shallow features are summed element-wise with the deep features. Next, a channel attention mechanism is used to weight the element-wise summed features, and the resulting channel attention is multiplied element-wise with the shallow features processed by the standard convolutional block to obtain feature S1. Simultaneously, a spatial attention mechanism is applied to the shallow features processed by the standard convolutional block, and the resulting spatial attention is multiplied element-wise with the element-wise summed features to obtain feature D1. Finally, F... H S1 and D1 are added element by element to obtain the final fusion feature.

8. The lung nodule segmentation method based on superpixel feature fusion according to claim 8, characterized in that, In the bidirectional attention fusion module, the deep features F are processed in the following manner. H Perform inverse weighting to obtain feature F H ':

Images