A lung nodule image segmentation method, medium, device and product

Abstract

The application discloses a lung nodule image segmentation method, medium, equipment and product, relates to the technical field of image processing, and comprises the following steps: constructing a lung nodule image segmentation model based on a VM-UNet structure; an encoder of each stage is composed of a frequency domain decoupling module and a global-local interaction module; a decoder of each stage is composed of a frequency domain reconstruction module and a global-local interaction module; the frequency domain decoupling module separates high-frequency and low-frequency features of an input in the frequency domain; the global-local interaction module of the encoder extracts global features from the low-frequency features, extracts local features from the high-frequency features, and fuses the global and local features; the frequency domain reconstruction module reconstructs an input in the frequency domain; the global-local interaction module of the decoder further reconstructs and fuses frequency domain reconstruction features; and a lung nodule image to be measured is input into the trained model to obtain a lung nodule image segmentation result. The application realizes effective extraction of local features while global modeling.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, medium, device and product for segmenting lung nodules. Background Technology

[0002] With the development of artificial intelligence technology in recent years, deep learning-based segmentation methods, with their powerful automatic feature extraction and complex pattern recognition capabilities, have gradually replaced traditional manually designed feature segmentation methods and become the mainstream direction of lung nodule segmentation research.

[0003] Currently, deep learning-based image segmentation methods are generally divided into CNN-based, Transformer-based, and SSM-based methods. UNet, as a representative of CNN models, is known for its simple structure and strong feature extraction capabilities. However, during its downsampling process, due to pooling and setting stride convolutions to reduce the feature map size, it inevitably loses feature information. Furthermore, due to its inherent network structure, UNet tends to focus more on local information while neglecting global information, resulting in low nodule segmentation accuracy.

[0004] To address the weakness of CNNs in global modeling, TransUNet first introduced the Transformer mechanism. It uses a Vision Transformer for feature extraction during the encoding phase and a CNN during the decoding phase, demonstrating significant global information acquisition capabilities. Subsequently, Swin-UNet utilizes a hierarchical Swin Transformer to capture cross-scale global structure. Although window attention limits cross-window interactions, resulting in insufficient modeling of fine-grained edges, its global correlation ability is significantly improved compared to CNNs. While Transformer-based models exhibit superior performance in global modeling, their self-attention mechanism requires secondary computational complexity in terms of image size, significantly increasing the computational burden for tasks requiring dense predictions, such as medical image segmentation.

[0005] Recently, State Space Models (SSMs) have attracted considerable interest from researchers. Modern SSMs (such as Mamba) not only establish global dependencies but also exhibit linear complexity in terms of input size. Mamba-based segmentation methods such as VM-UNet improve long-range dependency modeling capabilities by introducing SSM units into the encoder. S4-UNet further enhances long sequence modeling capabilities by replacing convolutional layers with structured state space layers. However, due to the recursive computation and global modeling logic of SSMs, they perform poorly in capturing locally sensitive features such as lung nodule edges.

[0006] In recent years, thanks to the continuous efforts of researchers, a great deal of work has been done in the field of image segmentation applications, and certain research results have been achieved. In 2015, Zhou Zhihua systematically introduced deep convolutional neural networks for pixel-level visual tasks, verifying the significant advantage of CNNs in feature representation capabilities compared to traditional methods in semantic segmentation, laying the foundation for the transformation of domestic image segmentation research from manual features to end-to-end learning. In 2017, Zhang Tao introduced a multi-scale feature fusion mechanism based on the U-Net architecture for organ segmentation in medical images, effectively mitigating the impact of scale changes on segmentation performance, and achieving results superior to traditional CNNs on multiple medical image datasets. In 2019, Chen Li introduced an attention mechanism to improve the encoder-decoder structure, improving the model's ability to focus on key regions by explicitly modeling the importance weights of channels and space, achieving higher sensitivity in lung nodule and tumor segmentation tasks.

[0007] Existing models typically focus only on the high extraction of local features or on building global relationships between features. Few scholars consider both global modeling and the extraction of local features simultaneously. This often results in incomplete feature extraction by the model, which in turn reduces the effectiveness of lesion segmentation.

[0008] Therefore, this invention proposes an improved lung nodule segmentation algorithm for the VM-UNet network to address the aforementioned problems. First, a frequency domain decoupling and reconstruction module is introduced to preserve high- and low-frequency features of the nodule during upsampling and downsampling. High frequencies correspond to details such as edges and spurs, while low frequencies reflect the overall morphology. Second, a global-local interaction module integrating selective scanning and dynamic receptive field fusion is proposed as the foundational module of the model. The selected scanning channel receives low-frequency overall morphological information generated by frequency domain decoupling and constructs global domain modeling through selective scanning. The dynamic receptive field channel receives high-frequency edge detail features generated by frequency domain decoupling and then performs dynamic receptive field selection based on the characteristics of the high-frequency features to achieve effective extraction of local domain features. Through this method, the model can simultaneously achieve global modeling and effective extraction of local features. Summary of the Invention

[0009] The purpose of this invention is to address the problem that existing methods fail to focus on both global modeling and local feature extraction, and to propose a lung nodule image segmentation method, comprising the following steps: S1. Obtain images of lung nodules and divide them into training set, validation set and test set; S2. A lung nodule image segmentation model is constructed based on the VM-UNet structure, including a Patch Embedding layer, a four-stage encoder, a four-stage decoder, and a final layer. The encoder input is processed sequentially through the first to fourth stages of the encoder for feature extraction. The output features of the last stage encoder are input into the decoder, and the resolution is gradually restored by sequentially processing the first to fourth stages of the decoder. The encoder output of the nth stage (n=1,2,3) is added to the decoder output of the 4th to nth stages by using a skip connection and then input into the decoder of the 5th to nth stages. Each stage of the encoder consists of a frequency domain decoupling module and a global-local interaction module; Each stage of the decoder consists of a frequency domain reconstruction module and a global-local interaction module; The frequency domain decoupling module separates high- and low-frequency features from its input in the frequency domain to obtain low-frequency and high-frequency features; the encoder's global-local interaction module extracts global features from low-frequency features and local features from high-frequency features, and then fuses the global and local features to obtain fused features. The frequency domain reconstruction module performs frequency domain reconstruction on its input to obtain frequency domain reconstruction features. The global-local interaction module of the decoder further reconstructs and fuses the frequency domain reconstruction features. S3. Use the training set, validation set, and test set to train, validate, and test the lung nodule image segmentation model to obtain the trained model. Input the lung nodule image to be tested into the trained model to obtain the lung nodule image segmentation result.

[0010] The present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described lung nodule image segmentation method.

[0011] The present invention also proposes an electronic device, including a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to execute the above-described lung nodule image segmentation method.

[0012] The present invention also proposes a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described lung nodule image segmentation method.

[0013] The beneficial effects of the technical solution provided by this invention are: This invention constructs a lung nodule image segmentation model based on the VM-UNet architecture. The encoder at each stage consists of a frequency domain decoupling module and a global-local interaction module, while the decoder at each stage consists of a frequency domain reconstruction module and a global-local interaction module. The frequency domain decoupling and reconstruction modules preserve high- and low-frequency features of the nodule during upsampling and downsampling. High frequencies correspond to details such as edges and spurs, while low frequencies reflect the overall morphology. A global-local interaction module that integrates selective scanning and dynamic receptive field fusion is proposed. The selected scanning channel receives low-frequency overall morphological information generated by frequency domain decoupling, constructing a global domain model through selective scanning. The dynamic receptive field channel receives high-frequency edge detail features generated by frequency domain decoupling, and then dynamically selects the receptive field based on the characteristics of the high-frequency features to achieve effective extraction of local domain features. Through this method, the model can simultaneously achieve global modeling and effective extraction of local features. Attached Figure Description

[0014] Figure 1 This is a flowchart of a lung nodule image segmentation method according to an example of the present invention; Figure 2 This is a structural diagram of a lung nodule image segmentation model based on VM-UNet, an example of the present invention. Figure 3 This is a global-local interaction module structure of an example of the present invention; Figure 4 This is a structural diagram of a dynamic receptive field unit according to an embodiment of the present invention; Figure 5 This is a visualization analysis result of the model of this invention and other comparative models; Figure 6 This is a block diagram of an electronic device according to an exemplary embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0016] A flowchart of a lung nodule image segmentation method according to an example of the present invention is shown below. Figure 1 Specifically, it includes: S1. Obtain images of lung nodules and divide them into training, validation and test sets.

[0017] This invention uses the publicly available LIDC-IDRI lung nodule dataset. To prevent overfitting, data augmentation techniques are employed, including random flipping and random rotation. The processed dataset is then divided into training, validation, and test sets in a 7:2:1 ratio.

[0018] S2. Constructing a lung nodule image segmentation model based on VM-UNet architecture. A structural diagram of the lung nodule image segmentation model based on VM-UNet architecture, as an example of this invention, is provided below. Figure 2 It includes the Patch Embedding layer (input embedding layer), a four-stage encoder, a four-stage decoder, and the final layer.

[0019] The Patch Embedding layer divides the input lung nodule image into 16 patches using a convolution with stride=4, while mapping the number of channels to a specified number of channels. The spatial resolution is compressed to 1 / 16 of the original, resulting in unified basic features.

[0020] Each stage of the encoder consists of a frequency domain decoupling module (FDD) and a global-local interaction module (Con-Block). The output features are fed to the decoder for further feature extraction via the neck direction, and fed to the decoder for fusion via a skip connection to provide spatial features. After the encoder completes downsampling, the network obtains the processed bottleneck features at the deepest layer.

[0021] In each stage of the encoder, the frequency domain decoupling module separates the high and low frequency features of its input in the frequency domain to obtain low frequency features and high frequency features; the global-local interaction module of the encoder in this stage performs global feature extraction on the low frequency features, performs local feature extraction on the high frequency features, and fuses the global and local features to obtain fused features.

[0022] Each stage of the decoder consists of a frequency domain reconstruction module and a global-local interaction module. Starting from this bottleneck feature, the decoder symmetrically employs decoding operations to progressively restore resolution from the bottom up: in the decoding stage, the encoder output corresponding to the current scale is first extracted, passed through a linear mapping layer, and then skip connections are completed using simple element-wise addition. Each stage of the decoder first utilizes the frequency domain reconstruction module to complete frequency domain reconstruction, achieving spatial size enlargement and channel number reduction. The resulting upsampled reconstructed features are then further reconstructed and fused through Con-Block, refining object boundaries and local textures at a higher resolution. During the decoding stage, the spatial resolution of the features is gradually restored, and semantic and detail information gradually merges, ultimately resulting in a feature map with the same size as the input image at the end of the last decoding layer.

[0023] The skip connection is as follows: the encoder input is sequentially processed by the encoders of the first to fourth stages for feature extraction, and the output features of the encoder of the last stage are input into the decoder. The resolution is gradually restored by the decoders of the first to fourth stages. The encoder output of the nth stage (n=1,2,3) is added to the decoder output of the 4th to nth stages by skip connection and then input into the decoder of the 5th to nth stages.

[0024] Finally, the network performs a 4x upsampling through the final layer and uses convolution to compress the channel dimension from the intermediate feature dimension to the number of categories, thereby outputting a pixel-level segmentation prediction result with size H×W×number of categories.

[0025] In this invention, the frequency domain decoupling module uses discrete wavelet transform to capture both low-frequency and high-frequency features of the image. Discrete wavelet transform can extract different frequency components from the image, thereby achieving image enhancement and detail extraction. The discrete wavelet transform module outputs X. LH X HL X HH Three high-frequency features and X LL A low-frequency feature. X LL It is the result after two low-pass filtering steps (horizontal and vertical), containing the low-frequency information of the image, namely most of the image's energy and overall contour; X LH It is obtained through low-pass filtering (horizontal direction) and high-pass filtering (vertical direction), and contains the horizontal edge information of the image, that is, the high-frequency changes of the image in the vertical direction; X HL It is obtained through high-pass filtering (horizontal direction) and low-pass filtering (vertical direction), and includes the vertical edge information of the image, that is, the high-frequency changes of the image in the horizontal direction; X HH The result is achieved through two high-pass filtering processes (horizontal and vertical), containing high-frequency information along the diagonal of the image, i.e., the image's details and texture. By emphasizing the high-frequency components of the image, the details and texture information can be highlighted; by emphasizing the low-frequency components, the overall contour and structural information of the image can be highlighted.

[0026] The frequency domain reconstruction module uses inverse discrete wavelet transform for image reconstruction. Inverse discrete wavelet transform decomposes the input features into sub-bands of different scales and small angles, and the original image is reconstructed by processing the sub-bands.

[0027] Figure 3 It uses a global-local interaction (Con-Block) module structure, which includes two branches: a global feature extraction branch and a local feature extraction branch. Features processed by the frequency domain decoupling module are fed into the Con-Block module. Large-scale low-frequency components representing the overall contour and region are sent to the global feature extraction branch to obtain global feature associations; three high-frequency components representing local information such as edges and textures are sent to the local feature extraction branch for local feature extraction. The extracted local and global features are then weighted and fused by the channel weighting module.

[0028] The global feature extraction branch utilizes the Visual State Space (VSS) module in VM-UNet. In this branch, low-frequency features, after normalization, are split into two branches: in the first branch, the input passes through a linear layer. In the second branch, the input is processed through a linear layer, a depthwise separable convolution, and an activation function, before being fed into a 2D selective scan module for global contour modeling. Subsequently, the features are layer-normalized and element-wise added to the output of the first branch to merge the two paths. Finally, the features are blended using a linear layer, and this result is combined with a residual connection as the output of the global feature extraction branch.

[0029] The local feature extraction branch includes a layer normalization layer, a dynamic receptive field unit, a convolutional layer, and a feedforward layer. The high-frequency feature input of the local feature extraction branch passes through the normalization layer and the dynamic receptive field (DRF) unit in sequence to obtain the first local feature. The first local feature and the high-frequency feature input are added together and then fed into the convolutional layer. The outputs of the convolutional layer are added together and then fed into the feedforward layer. The outputs of the feedforward layer are added together to obtain the local feature extracted by the local feature extraction branch.

[0030] Figure 4 This is a structural diagram of a Dynamic Receptive Field (DRF) unit according to an embodiment of the present invention. The Dynamic Receptive Field unit divides its input features according to channels. The different heads, first of all and Feature fusion is performed to obtain features , will feature Other heads Perform recursive fusion operations. Taking a two-layer dynamic receptive field unit as an example, the head... The signal is divided into two paths. One path undergoes a 1x1 convolution operation and is activated by GELU, followed by depthwise separable convolution with an NxN kernel. The other path only undergoes 1x1 convolution for channel alignment. The two signals are then dot-producted on pixels at the same spatial location. This operation expands the receptive field and amplifies the influence of pixels on the receptive field. The other path, H1, undergoes depthwise separable convolution with an nxn kernel and then... The header is subjected to NxN convolution, and the resulting feature is fused. The fused feature is then concatenated with the dot product features for output. This output has two receptive fields. Compared to other heads... By performing recursive operations, features can be continuously integrated with the context in a progressive manner, eventually forming an asymptotic Gaussian distribution (AGD) that covers multiple scales, thus avoiding the AGD destruction problem commonly found in large kernel convolutional networks.

[0031] The formula for the recursive fusion operation of dynamic receptive field units is as follows:

[0032]

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039] in, , , , , , as well as They are respectively and Intermediate features generated during the fusion process This represents the i-th head feature. This represents the (i+1)th head feature. , express Convolution operation, express Activation function Indicates that the convolution kernel is In this invention, N=11 is used for depthwise convolution operations. This represents the dot product operation. Indicates addition. Indicates that the convolution kernel is In this invention, N=7 is used for depthwise convolution operations. and The i-th fused feature is obtained after splicing. .

[0040] S3. Use the training set, validation set, and test set to train, validate, and test the lung nodule image segmentation model to obtain the trained model. Input the lung nodule image to be tested into the trained model to obtain the lung nodule image segmentation result.

[0041] During training, the number of training epochs was set to 300, the batch size to 32, the model used the AdamW optimizer and cosine annealing learning scheduler, the loss function used the BceDice loss function, the maximum number of iterations was 50, the initial learning rate was 0.001, and the minimum learning rate was 0.00001.

[0042] The model begins training and prints parameters such as training epochs, loss, and learning rate on the terminal interface. Validation is performed every 5 epochs, printing the validation loss and evaluation metrics: Dice Similarity Coefficient (DSC), Mean Intersection over Union (MIoU), Accuracy, and Sensitivity (SE). After training, the best-performing model is saved. To test the model, input the test parameters and select the test model path and test result save path. The model then begins inference testing. After testing, the inference test results can be viewed in the preset test result save path.

[0043] To demonstrate the superior performance of the proposed model, a comparative experiment was conducted on the LIDC-IDRI dataset, comparing the proposed model with several popular and advanced medical image segmentation models, including U-Net, U-Net v2, U-Net++, Attention U-Net, SA-Net, TransNorm, SCR-Net, VM-UNet, and H-vmunet.

[0044] in: U-Net is the model in the paper "U-Net: Convolutional Networks for Biomedical Image Segmentation"; U-Net v2 is the model in the paper "U-Net v2: Rethinking the Skip Connections of U-Net for Medical Image Segmentation"; U-Net++ is the model in the paper "UNet++: A Nested U-Net Architecture for Medical ImageSegmentation"; Attention U-Net is the model in the paper "Attention U-Net: Learning Where to Look for the Pancreas"; SA-Net is the model in the paper "SA-Net: Shuffle Attention for Deep Convolutional Neural Networks"; TransNorm is the model in the paper "TransNorm: Transformer Provides a Strong SpatialNormalization Mechanism for a Deep Segmentation Model"; SCR-Net is the model in the paper "SCR-Net: A novel lightweight aquatic biological detection network"; VM-UNet is the paper "VM-UNet: Vision Mamba UNet for Medical Image Segmentation The model in 》; H-vmunet is the model in the paper "H-vmunet: High-order Vision Mamba UNet for Medical Image Segmentation".

[0045] Table 1 presents the comparative experimental results, showing that the model of this invention exhibits excellent segmentation performance in the task of segmenting lung nodules. The model of this invention achieves the best results in the three metrics of DSC, MIoU, and SE. As can be seen from the table, the performance of the improved VM-UNet of this invention is superior to the traditional UNet model, demonstrating the effectiveness of introducing selective scanning for global modeling. Furthermore, the table shows that compared to the original VM-UNet, the model of this invention performs better, achieving a 2.19% increase in DSC and a 3.26% increase in MIoU. This indicates that the dynamic receptive field proposed in this invention compensates for the insufficient local feature extraction of the Mamba model, especially for data with minute features, such as medical images.

[0046] Table 1

[0047] To verify the effectiveness of the frequency domain decoupling module and the dynamic receptive field module, four different models were set up. Model 4 is the model proposed in this invention; Model 3 is the model proposed in this invention with the dynamic receptive field branch (local feature extraction branch) removed, retaining only the selective scanning branch (global feature extraction branch); Model 2 is the model proposed in this invention with the frequency domain decoupling module and frequency domain reconstruction module replaced by traditional upsampling and downsampling; Model 1 is the model proposed in this invention with the dynamic receptive field branch removed, and the frequency domain decoupling module and frequency domain reconstruction module replaced by traditional upsampling and downsampling. The results of this ablation experiment are shown in the figure. In Table 2, the frequency domain decoupling module is referred to as a, and the dynamic receptive field module is referred to as b. From the data in Table 2, it can be concluded that the DSC and MIoU index values ​​are the lowest in Model 1, and the above indexes achieve the best results in Model 4, proving the effectiveness of the model proposed in this invention. Specifically, compared to Model 1, Model 3 improved DSC by 0.68% and MIoU by 0.21%, indicating that the frequency domain decoupling module explicitly separates low-frequency and high-frequency information through wavelet transform, enabling the model to capture boundary details and structural textures more fully. Compared to Model 1, Model 2 improved DSC by 0.35% and MIoU by 1.35%, indicating that the dynamic receptive field mechanism is more effective in covering multi-scale structures and can improve the discrimination of small targets and the differentiation of complex backgrounds in local areas. This can also be observed in the visualization comparison. On the LIDC dataset, the model of this invention is more sensitive to small targets. Compared with Models 2 and 3, Model 4 still maintains a significant index gain, proving that there is no redundancy between modules a and b in the model. Instead, they form a clear complementary relationship, jointly improving the detection and segmentation capabilities of the model.

[0048] Table 2

[0049] Figure 5 The results are visualized. The segmentation results of the model of this invention and other mainstream advanced models on the aforementioned LIDC-IDRI dataset are visualized. These results demonstrate the robustness and generalization ability of the model of this invention in various medical images. Furthermore, the visualization of the results clearly shows that the model of this invention has a greater sensitivity to lesion information. In contrast, VM-UNet still has shortcomings in learning lesion features. It can also be seen that although models such as U-Net++ and VM-UNet can correctly identify the location of lesions when segmenting lesion regions, the segmentation contours are still blurry.

[0050] In one exemplary embodiment, a computer-readable storage medium is included, which stores a computer program that, when executed by a processor, implements the lung nodule image segmentation method described above.

[0051] Please see Figure 6 In one exemplary embodiment, the device further includes an electronic device including at least one processor, at least one memory, and at least one communication bus.

[0052] The memory stores a computer program, which includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through the communication bus to execute the lung nodule image segmentation method described above.

[0053] In one exemplary embodiment, a computer program product is proposed, including a computer program / instructions that, when executed by a processor, implement the steps of the lung nodule image segmentation method described above.

[0054] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those 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 invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for segmenting lung nodule images, characterized in that, Includes the following steps: S1. Obtain images of lung nodules and divide them into training set, validation set and test set; S2. A lung nodule image segmentation model is constructed based on the VM-UNet structure, including a Patch Embedding layer, a four-stage encoder, a four-stage decoder, and a final layer. The encoder input is processed sequentially through the first to fourth stages of the encoder for feature extraction. The output features of the last stage encoder are input into the decoder, and the resolution is gradually restored by sequentially processing the first to fourth stages of the decoder. The encoder output of the nth stage (n=1,2,3) is added to the decoder output of the 4th to nth stages by using a skip connection and then input into the decoder of the 5th to nth stages. Each stage of the encoder consists of a frequency domain decoupling module and a global-local interaction module; Each stage of the decoder consists of a frequency domain reconstruction module and a global-local interaction module; The frequency domain decoupling module separates high- and low-frequency features from its input in the frequency domain to obtain low-frequency and high-frequency features; the encoder's global-local interaction module extracts global features from low-frequency features and local features from high-frequency features, and then fuses the global and local features to obtain fused features. The frequency domain reconstruction module performs frequency domain reconstruction on its input to obtain frequency domain reconstruction features. The global-local interaction module of the decoder further reconstructs and fuses the frequency domain reconstruction features. S3. Use the training set, validation set, and test set to train, validate, and test the lung nodule image segmentation model to obtain the trained model. Input the lung nodule image to be tested into the trained model to obtain the lung nodule image segmentation result.

2. The lung nodule image segmentation method according to claim 1, characterized in that, The frequency domain decoupling module uses discrete wavelet transform to capture the low-frequency and high-frequency features of the image.

3. The lung nodule image segmentation method according to claim 1, characterized in that, The frequency domain reconstruction module uses inverse discrete wavelet transform for image reconstruction.

4. The lung nodule image segmentation method according to claim 1, characterized in that, The global-local interaction module includes a global feature extraction branch and a local feature extraction branch; The global feature extraction branch performs global feature extraction on low-frequency features, while the local feature extraction branch performs local feature extraction on high-frequency features. The global feature extraction branch uses the visual state space module of VM-UNet; The local feature extraction branch includes a layer normalization layer, a dynamic receptive field unit, a convolutional layer, and a feedforward layer. The high-frequency feature input of the local feature extraction branch passes through the normalization layer and the dynamic receptive field unit in sequence to obtain the first local feature. The first local feature and the high-frequency feature input are added together and then input into the convolutional layer. The outputs of the convolutional layer are added together and then input into the feedforward layer. The outputs of the feedforward layer are added together to obtain the local feature extracted by the local feature extraction branch. The dynamic receptive field unit divides its input features according to channels into The different heads, first of all and Feature fusion is performed to obtain features , will feature Other heads Perform a recursive fusion operation.

5. The lung nodule image segmentation method according to claim 4, characterized in that, The formula for the recursive fusion operation of dynamic receptive field units is as follows: in, , , , , , as well as They are respectively and Intermediate features generated during the fusion process This represents the i-th head feature. This represents the (i+1)th head feature. , express Convolution operation, express Activation function Indicates that the convolution kernel is Depth convolution operation, This represents the dot product operation. Indicates addition. Indicates that the convolution kernel is The depthwise convolution operation will and The i-th fused feature is obtained after splicing. .

6. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 5.

7. An electronic device, characterized in that, The device includes a processor and a memory, the processor being interconnected with the memory, wherein the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to perform the method as described in any one of claims 1 to 5.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the method described in any one of claims 1 to 5.

Images