A lung nodule segmentation method based on DMC-UNet

Abstract

The application provides a lung nodule segmentation method based on a DMC-UNet, which utilizes a DMC-UNet network model to realize image segmentation, so that the segmentation accuracy is higher, and early detection of lung cancer is facilitated, wherein the DMC-UNet network model is an improved network based on a UNet, which adds a lightweight residual structure, a multi-scale feature upsampling structure, a parallel fusion structure of a transpose convolution upsampling and an X / YCCA attention module on the basis of the UNet network, so that, under the premise of maintaining the main frame structure of the UNet network, the segmentation effect is more outstanding.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer technology and medical image analysis, and particularly relates to a lung nodule segmentation method based on DMC-UNet. BACKGROUND

[0002] In early lung cancer, it is often in the form of lung nodules. For early lung cancer patients, clinical surgical treatment can be used, and the survival rate is very high. Early detection and diagnosis of lung nodules can increase the chances of curing lung cancer patients, and different types of lung nodules, such as solid nodules, partial solid nodules, and calcified nodules, also have differences in severity and symptoms. Therefore, rapid and accurate screening and diagnosis of lung nodules are crucial for the diagnosis and treatment of lung cancer.

[0003] Nowadays, medical imaging technology has been widely used in early lung cancer detection, among which the prominent technologies are computed tomography (CT) and magnetic resonance imaging. In particular, the low-dose computed tomography (CT) technology has the advantages of low radiation and high resolution, but a large number of slice images are generated during each lung imaging, which brings a huge workload to radiologists. In addition, the types of lung nodules are various, the sizes are small, the shapes are changeable, and it is not easy to diagnose, which may lead to misdiagnosis. Therefore, with the development of CT technology, using computer-aided technology (CAD) to segment lung nodules has become a trend.

[0004] Traditional segmentation methods include threshold-based segmentation algorithms, region growing algorithms, edge detection algorithms, iterative local threshold segmentation algorithms, active contour models, clustering methods, etc. These methods mainly preprocess the image, such as grayscale, filtering, and binarization, and then apply specific algorithms to determine the boundaries or segmentation points of different regions in the image. These methods have advantages in computing speed, but usually require manual selection of appropriate parameters and adjustment of algorithms, and their accuracy and robustness are also limited. Segmentation methods based on deep learning have the advantages of higher accuracy, better robustness, automatic feature extraction, stronger transferability, and wider application range compared to traditional segmentation methods. 2D segmentation network is a neural network in deep learning, mainly used to solve image semantic segmentation problems. It is usually composed of an encoder and a decoder, where the encoder is used to extract the feature representation of the input image, and the decoder maps these features back to the original image size to generate labels or masks for each pixel, but the existing segmentation model still has the following shortcomings: 1. Low-contrast image processing is not accurate; 2. Poor detection of small-sized objects; 3. Long running time for large-scale image datasets. SUMMARY

[0005] In view of the above, it is necessary to provide a lung nodule segmentation method based on DMC-UNet, which uses a DMC-UNet network model to realize image segmentation, so that the segmentation accuracy is higher, wherein the DMC-UNet network model is an improved network based on UNet, which adds a lightweight residual structure, a multi-scale feature up-sampling, and a parallel fusion structure of a transposed convolution up-sampling and an X / Y CCA (X / Y Channel Attention Module and Coordinate Attention) attention module based on the UNet network. Through this improvement, the UNet network main frame structure is maintained, and the segmentation effect is more outstanding.

[0006] To achieve the above purpose, the technical scheme adopted by the present application is:

[0007] A lung nodule segmentation method based on DMC-UNet, comprising the following steps:

[0008] S1, pre-processing the CT image provided by the LIDC public data set to obtain a data set;

[0009] S2, adding a lightweight residual structure, a multi-scale feature up-sampling, and a parallel fusion structure of a transposed convolution up-sampling and an X / Y CCA attention module based on the UNet network to construct a DMC-UNet network model;

[0010] S3, training the DMC-UNet network model obtained in step S2 on the data set obtained in step S1 and testing;

[0011] S4, introducing the lung CT image to be segmented into the trained DMC-UNet network model to obtain the segmented structure.

[0012] Further, the pre-processing process in step S1 is as follows:

[0013] S1.1, download the LIDC public data set, and perform image preprocessing on the LIDC public data set, wherein each sample in the LIDC public data set contains four radiologists' annotation information; for each annotation information, perform OR operation to generate a new mask image; in each sample, if a doctor annotates a certain region as a nodule, the region is regarded as a nodule;

[0014] S1.2, cropping the pre-processed lung CT image, so that the nodule determined in step S1.2 is located at the center of the image, and at the same time, the resolution of the image is scaled to 128x128 during cropping; after processing the image, combine the low-resolution image with the high-resolution image into an information pair to make a data set.

[0015] Further, in step S2, the lightweight residual structure, the multi-scale feature upsampling, and the parallel fusion structure of the transposed convolution upsampling and the X / Y CCA attention module are added to the UNet network to construct a DMC-UNet network model, specifically:

[0016] S2.1, input the data set obtained in step S1 into the lightweight residual structure to obtain a preliminary feature map;

[0017] S2.2, perform a downsampling operation on the preliminary feature map obtained in step S2.1 to obtain a new feature map;

[0018] S2.3, after the feature map obtained in step S2.2 is subjected to the same downsampling operation four times in step S2.2, the features extracted from each layer containing abstract and high-level semantic information are subjected to an intermediate layer, and after the intermediate layer obtains high-level semantic information, it is subjected to upsampling processing in step S2.4;

[0019] S2.4, in the upsampling processing, the transposed convolution in the upsampling process of the UNet network and the improved ESPCN super-resolution technology are respectively subjected to the lightweight residual structure operation in step S2.1 in parallel, and then subjected to point multiplication operation, and the obtained feature map is subjected to processing in step S2.5;

[0020] S2.5, input the feature map obtained in step S2.4 into the CCA attention module, and the obtained feature map is subjected to processing in step S2.6;

[0021] S2.6, input the feature map obtained in step S2.5 into the CCA attention module again, and after three same operations, input the obtained feature map into the network of the last layer, and finally output the final feature map through the end convolution layer.

[0022] Further, in step S2, the lightweight residual structure is connected in residual connection by two DSC units and one DSC unit, and finally activated by a ReLU activation function, wherein the DSC unit is a depth separable convolution unit composed of a 3x3 channel-by-channel convolution and a 1x1 point-by-point convolution.

[0023] Further, in step S2.2, the downsampling operation specifically adopts a 2x2 max pooling operation to compress the resolution to 1 / 2 of the input feature map.

[0024] Further, in step S2.4, the improved ESPCN super-resolution technology is realized by replacing the original Hidden layers of the ESPCN with a multi-scale feature fusion module to extract features from the feature map, and the multi-scale feature fusion module is referred to as an ASPP module.

[0025] Further, the ASPP module is composed of four parallel branches, wherein the first branch is a 1x1 convolutional layer, and the remaining three branches adopt a 3x3 dilated convolution with dilated coefficients of 2, 4, and 8 respectively; the feature information of the four branches is spliced in the dimension, and the channel number is changed through a 1x1 convolutional layer to obtain the output of the ASPP module.

[0026] Further, in step S2.5, the CCA attention module is an X / Y CAM structure obtained by fusing the CAM attention and the CoordAttention attention; the X / Y CAM structure is specifically:

[0027] First, the operations in the X direction and in the Y direction are to fuse the input feature map into the CoordAttention, wherein,

[0028] In the X direction, the input feature map is fused into the CoordAttention through global average pooling and global maximum pooling in the horizontal direction to output and two-dimensional feature maps, wherein (H, W) is the resolution of the input feature map F, and C is the channel number of the feature map F; then, the obtained feature map is input into a shared weight double-layer neural network Shared MLP, in which the channels are compressed by a compression ratio r1 to obtain and , and then the channel number is restored to the channel number before input; finally, the results output by the Shared MLP are subjected to an Element-wise addition operation, and then subjected to a Sigmoid activation function to generate a vector;

[0029] In the Y direction, the operation is the same as that in the X direction, and finally, the operation in the Y direction generates a vector;

[0030] After being fused into the CoordAttention, the and are unified in the dimension through Reshape, and then concatenated in the spatial dimension to obtain a vector , which is then subjected to channel compression through a 1x1 convolution with a compression ratio r2 to obtain ; again, make the input features have the same mean and variance through BatchNormalization; then, help the network learn more complex features through the Non-linear function, which is also called the h_swish function; then, the vector obtained by splicing is restored through the Split operation Separation in the spatial dimension obtains And Vector, and then restore in the spatial dimension And Vector, and then each through the 1x1 convolution to restore the channel number consistent with the input feature map Consistent, and after the weighting operation, the feature map .

[0031] Further, the h_swish function is a nonlinear activation function, which consists of two parts: first, an "h-sigmoid" function, which is similar in form to the Sigmoid function, but uses ReLU6 as its approximation function, ReLU6 is a truncated ReLU; second, a linear transformation, which scales and translates the input data.

[0032] Further, in step S2.6, the output of each layer is fused with the same resolution features of the encoder and the decoder in a skip-link manner.

[0033] Compared with the prior art, the present application has the following beneficial effects:

[0034] 1、The lung nodule segmentation method based on DMC-UNet proposed in the present application uses a new network model, namely the DMC-Unet network model, which is improved on the basis of the UNet network model, has the UNet fast learning and model visualization capability, and under the use of the lightweight residual structure (DSC residual structure), the training time and memory consumption are reduced while the performance is maintained, two kinds of upsampling modes, namely transpose convolution and ASPP (Atrous Spatial Pyramid Pooling) +PixelShuffle, are fused in parallel form, so that the model can better restore the image details, and finally the CCA attention module is added after fusion, which helps the model to better recover the lost position information, thereby improving the accuracy of the semantic segmentation task.

[0035] 2、The method given in the present application shows high performance in segmenting lung nodules on the LIDC public data set, can effectively improve the segmentation accuracy, and is helpful for early detection of lung cancer.

[0036] 3、The segmentation method of the present application can effectively realize lung nodule detection, can improve the work efficiency of doctors, and also provides guarantee for early screening, diagnosis and treatment of lung cancer. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The whole flow chart of the present application.

[0038] Figure 2 The whole structure diagram of the DMC-Unet network model.

[0039] Figure 3 The schematic diagram of the lightweight residual structure.

[0040] Figure 4 The structure schematic diagram of the ASPP multi-scale feature fusion module.

[0041] Figure 5 The structure schematic diagram of the CCA attention mechanism.

[0042] Figure 6 The visualization multi-method comparison diagram on the LIDC public data set, in which, 'our' refers to the DMC-Unet network model of the present application.

[0043] The following specific embodiments will further illustrate the present application in combination with the above drawings. DETAILED DESCRIPTION

[0044] Please refer to Figures 1 to 6 In a preferred embodiment of the present application, a lung nodule segmentation method based on DMC-Unet comprises the following steps:

[0045] S1, pre-processing the CT image provided by the LIDC public data set to obtain a data set;

[0046] S2, adding lightweight residual structure, multi-scale feature upsampling, and parallel fusion structure of transposed convolution upsampling and X / Y CCA attention module to the UNet network to construct a DMC-Unet network model;

[0047] S3, training the DMC-Unet network model obtained in step S2 on the data set obtained in step S1 and testing;

[0048] S4, introducing the lung CT image to be segmented into the trained DMC-Unet network model to obtain the segmented structure.

[0049] That is, the general idea of the present application is to preprocess the CT images provided by the LIDC public dataset, reasonably divide the dataset, then put it into the constructed DMC-UNet network model for training, and finally test the trained model. The determined DMC-UNet network model can be used for lung CT image segmentation after testing. Such segmentation method can effectively improve the segmentation accuracy and help early detection of lung cancer.

[0050] Next, the lung nodule segmentation method provided by the present application will be described in detail.

[0051] In the present application, the LIDC public dataset used in the preprocessing process in step S1 includes 1018 cases. The preprocessing process in step S1 is as follows:

[0052] S1.1, download the LIDC public dataset, and perform image preprocessing on the LIDC public dataset, wherein each sample in the LIDC public dataset contains annotation information annotated by four radiologists; for each annotation information, perform or operation to generate a new mask image, that is, perform or operation on the annotation information annotated by the four radiologists in each sample to generate new label information, that is, fuse the lung nodule information annotated by the four doctors to generate a new mask image; in each sample, if a doctor annotates a specific region as a nodule, the region is considered as a nodule.

[0053] S1.2, crop the preprocessed lung CT image, so that the nodule determined in step S1.2 is located at the center of the image, and at the same time, the resolution of the image is scaled to 128x128. After processing the image, combine the low-resolution image with the high-resolution image into an information pair to make a dataset.

[0054] Further, in step S2, the UNet network is increased with a lightweight residual structure, a multi-scale feature up-sampling, and a structure of parallel fusion of transposed convolution up-sampling and X / Y CCA attention module to construct a DMC-UNet network model, specifically:

[0055] S2.1, input the dataset obtained in step S1 into the lightweight residual structure to obtain a preliminary feature map. That is, the images in the dataset first enter the lightweight residual structure. In the present application, the lightweight residual structure includes a residual unit, a DSC unit and a ReLu function. Specifically, the lightweight residual structure is connected in residual by two DSC units and one DSC unit, and finally activated by a ReLU activation function, wherein the DSC unit is a depth separable convolution unit composed of 3x3 channel-wise convolution and 1x1 point-wise convolution. After the above operation, the preliminary feature is obtained and processed in step S2.2.

[0056] S2.2, the preliminary feature map obtained in step S2.1 is subjected to a downsampling operation to obtain a new feature map. The downsampling operation is specifically a 2x2 max-pooling operation to compress the resolution to 1 / 2 of the input feature map, and the new feature map is subjected to processing in S2.3.

[0057] S2.3, the feature map obtained in step S2.2 is subjected to the same downsampling operation four times, and the extracted features containing abstract and high-level semantic information of each layer are subjected to intermediate layer processing, and the intermediate layer obtains high-level semantic information and is subjected to upsampling processing in step S2.4.

[0058] S2.4, in the upsampling processing, the transpose convolution in the UNet network upsampling process and the improved ESPCN super-resolution technology are respectively subjected to the lightweight residual structure operation in step S2.1, and then subjected to point multiplication operation, and the obtained feature map is subjected to processing in step S2.5. Specifically, the improved ESPCN super-resolution technology is based on the traditional ESPCN super-resolution technology, and the multi-scale feature fusion module is used to replace the original Hidden layers of ESPCN to realize feature extraction of the feature map. In this case, the multi-scale feature fusion module is called ASPP module. The ASPP module is composed of four parallel branches, wherein the first branch is a 1x1 convolution layer, and the other three branches use 3x3 dilated convolution with dilated coefficients of 2, 4 and 8 respectively; the feature information of the four branches is spliced in the dimension, and the channel number is changed through a 1x1 convolution layer to obtain the output of the ASPP module.

[0059] That is, in this step, the upsampling method is to fuse the transposed convolution in the UNet network's sampling process and the improved ESPCN super-resolution technique in parallel, replacing the transposed convolution used in the original U-Net. The improved ESPCN super-resolution technique is an improvement on the traditional ESPCN super-resolution technique. The traditional ESPCN super-resolution technique extracts features from the feature map through hidden layers, transforming the feature map of dimension (H, W, C) into (H, W, r2*C), and then upsampling the image through a sub-pixel convolution layer to obtain a feature map of dimension (r*H, r*W, C). The improved ESPCN super-resolution technique of this application replaces the original hidden layers in the ESPCN super-resolution technique with the ASPP module when the input is a feature map of resolution (H, W, C). The resulting image is then upsampled through a sub-pixel convolution layer, and the output new feature map is then processed in step S2.5.

[0060] S2.5, Input the feature map obtained in step S2.4 into the CCA attention module, and the resulting feature map is then processed in step S2.6.

[0061] The CCA attention module is an X / Y CAM structure obtained by fusing CAM attention and CoordAttention attention. This X / Y CAM structure addresses the fact that coordinate attention only considers the coordinate information of each position in the input feature map, neglecting the channel information of the features. CAM channel attention can compensate for this deficiency, thereby improving the robustness and performance of the model. Therefore, this invention uses an X / Y CAM structure to better integrate CAM attention into CoordAttention. The specific X / Y CAM structure is as follows:

[0062] First, operations are performed in the X and Y directions to incorporate the input feature map into CoordAttention, where...

[0063] In the X direction, the input feature map By using global average pooling and global max pooling in the horizontal direction respectively, the outputs are... and The feature map is generated in two dimensions, where (H, W) is the resolution of the input feature map F, and C is the number of channels in feature map F. The resulting feature map is then input into a two-layer shared weighted neural network, Shared MLP, where the channels are compressed at a compression ratio of r1. and , and then the number of channels is restored to the number of channels before input; finally, the results of the Shared MLP output are added element-wise and passed through a Sigmoid activation function to generate a vector. In this step, when selecting the compression ratio r1, it needs to be considered that a higher compression ratio may not be able to fully capture the complexity of the data, and the model may converge faster but may cause underfitting problems, while a too low compression ratio may cause too many parameters, increasing the computational complexity of the model, and may cause overfitting problems. Based on this, in the present embodiment, the compression ratio r1 is selected to be 32, so as to avoid underfitting or overfitting as much as possible under the premise of ensuring appropriate computational efficiency.

[0064] In the Y direction, the operation is the same as in the X direction, and finally, the operation in the Y direction generates a vector.

[0065] The above operations in the X direction and in the Y direction can be represented by the following formula 1 and formula 2:

[0066] (1)

[0067] (2)

[0068] In the formula, W0 and W1 are two weights allocated by two convolutions, and are feature maps obtained by horizontal global average pooling (AvgPool) and global maximum pooling (MaxPool), and are feature maps obtained by vertical global average pooling (AvgPool) and global maximum pooling (MaxPool).

[0069] After being integrated into CoordAttention, the and are unified in dimension, and then concatenated in the spatial dimension to obtain the vector , and then compressed by 1x1 convolution in the channel, with a compression ratio of r2, to obtain ; In this embodiment, the compression ratio r2 is set to 32, and it should be noted that when the number of channels is compressed to less than 8, the number of channels is 8. In addition, BatchNormalization is used to make the input features have the same mean and variance, so as to enhance the robustness of the model, speed up the convergence speed and improve the generalization ability of the model. Then, the Non-linear function helps the network to learn more complex features, where the Non-linear function is also called the h_swish function. Then, the Split operation is used to restore the vector obtained by splicing to and in the spatial dimension, and then restore and to the vector in the spatial dimension, and then restore each to the channel number consistent with the input feature map through 1x1 convolution, and after weighting operation, the feature map is obtained. The feature map obtained after the CCA attention structure enters step S2.6 for processing.

[0070] In this step, the h_swish function is a nonlinear activation function, which consists of two parts: first, an "h-sigmoid" function, which is similar in form to the Sigmoid function, but uses ReLU6 as its approximation function to effectively avoid the problem of gradient disappearance, and ReLU6 is a truncated ReLU; second, a linear transformation, which scales and translates the input data to control the range of output values, as shown in equation 3:

[0071] (3)

[0072] That is, the h_swish function is a nonlinear activation function that has a form similar to ReLU but is smoother than ReLU, which can effectively avoid the "dead zone" problem of the ReLU function when the input is negative.

[0073] S2.6, the feature map obtained in step S2.5 is input into the CCA attention module again, and after three times of the same operation, the obtained feature map is input into the last layer of the network, and finally the final feature map is output through the last convolutional layer. Specifically, in this embodiment, the last convolutional layer is a 1x1 convolution, and the output final feature map is a 128x128x1 feature map. Further, in step S2.6, the output of each layer is fused with the same resolution features of the encoder and the decoder in a skip-link manner.

[0074] In the present application, the training and testing network in step S3 is to train and test the processed data, specifically, the data set samples obtained in step S1 are divided into a training set, a validation set and a test set, wherein the training set is used to train the DMC-Unet network model obtained in step S2, the validation set is used to adjust the parameters of the DMC-Unet network model, and the test set is used to test whether the DMC-Unet network model is accurate.

[0075] In order to reflect the effect of the DMC-Unet network model of the present application, the present application performs lung nodule segmentation on the LIDC public data set. In the training stage of the lung nodule on the LIDC public data set, the maximum number of iterations is 100, the initial learning rate is 1e -4 , Adam is used as the optimizer, and the loss function is a combination of binary cross entropy loss (Binary Cross Entropy Loss) and Dice Loss (BCEDiceLoss) to accurately detect the area where the lung nodule exists. When testing, multiple related networks are compared, and evaluation indicators IoU, Precision, Recall and F1-score are used for comparative analysis, as shown in Table 1.

[0076] Table 1 Comparison results table of the DMC-Unet network model of the present application and multiple related networks

[0077] Method IoU Precision Recall F1-score U-Net 81.99 90.74 90.08 89.19 ResUNet++ 81.21 91.57 88.29 88.65 U-Net3+ 83.66 90.56 92.03 90.15 SAR-UNet 81.41 89.95 90.11 88.84 Attention-UNet 80.83 89.08 90.48 88.49 Swin-UNet 70.08 84.85 83.16 81.54 TransUNet 81.87 90.59 89.94 88.99 SMR-UNet 82.75 91.52 90.03 89.59 DMC-Unet 83.36 91 91.15 89.92

[0078] From Table 1, (1) the IoU and F1-score of the DMC-Unet network model of the present application are 83.36% and 89.92%, respectively, which are improved by 1.37% and 0.73% compared with the Unet network model.

[0079] (2) Through experiments, it is found that the DMC-Unet network model is not as good as the UNet3+ network model in IoU and F1-score, but in the training process, it can be clearly found that due to the introduction of the full-scale skip connection idea in the UNet3+ network model, the complexity of its structure is high, the parameters are more, and the training time is longer. In addition, from Figure 6 , it can be seen that the difference between the DMC-Unet network model and the UNet3+ network model in the segmentation results of the selected 6 images is also small, and even in the fifth row, it is better than the UNet3+ network model in detail processing.

[0080] (3) From Figure 6As can be seen, in the overall segmentation results, in addition to the edge profile of the Swin-Unet network model for part of the image segmentation being greatly different from the GT image, the lung nodule profiles of the U-Net, ResUNet++, SAR-UNet, Attention-UNet, UNet3+, TransUNet and SMR-Unet network models are similar to the GT image, but the SMR-Unet network model in the fifth row identifies the blood vessel part in the image as a lung nodule, which has a problem that other network models do not have, and the DMC-Unet network model of the application is closer to the GT image in the details, shape and size of the profile. In addition, it can also be found that the segmentation effect of all network models on the lung nodule picture in the second row is not ideal, and there are some differences in details, but the edge information of the DMC-Unet network model of the application is closest to the GT image, and the top highlighted part and the bottom profile are also closest to the GT image. In general, for lung nodules with simple edge information and less details, most networks can segment out relatively ideal results, but for irregular profile and more detailed lung nodules, some networks cannot achieve the desired effect, while the DMC-Unet network model of the application can obtain more accurate segmentation images, and the segmentation effect is relatively ideal.

[0081] The above description is a detailed description of the preferred embodiments of the application, but the embodiments are not intended to limit the scope of the patent application of the application. Any equivalent changes or modifications made under the technical spirit of the application should be included in the scope of the patent.

Claims

1. A lung nodule segmentation method based on DMC-UNet, characterized in that, Comprise the following steps: S1, pre-process the CT image provided by the LIDC public data set to obtain a data set; S2, on the basis of UNet network, increase the structure of light weight residual structure, multi-scale feature up-sampling and parallel fusion of transposed convolution up-sampling and X / Y CCA attention module, and construct DMC-UNet network model; S3, training the DMC-UNet network model obtained in step S2 on the data set obtained in step S1 and testing; S4, introducing the lung CT image to be segmented into the trained DMC-UNet network model to obtain the segmented structure; In step S2, on the basis of UNet network, increase the structure of light weight residual structure, multi-scale feature up-sampling and parallel fusion of transposed convolution up-sampling and X / Y CCA attention module, and construct DMC-UNet network model, specifically: S2.1, input the data set obtained in step S1 into the light weight residual structure to obtain the preliminary feature map; S2.2, the preliminary feature map obtained in step S2.1 is down-sampled to obtain a new feature map; S2.3, the feature map obtained in step S2.2 is subjected to the same down-sampling operation four times, and then the extracted features containing abstract and high-level semantic information of each layer are subjected to intermediate layer processing, and after the intermediate layer obtains high-level semantic information, it is subjected to up-sampling processing in step S2.4; S2.4, in the up-sampling processing, the transposed convolution in the up-sampling process of UNet network and the improved ESPCN super-resolution technology are parallelly subjected to the light weight residual structure operation of step S2.1, and then subjected to point multiplication operation, and the obtained feature map is subjected to processing in step S2.5; S2.5, input the feature map obtained in step S2.4 into the CCA attention module, and the obtained feature map is subjected to processing in step S2.6; S2.6, input the feature map obtained in step S2.5 into the CCA attention module again, and after three times of the same operation, input the obtained feature map into the network of the last layer, and finally output the final feature map through the end convolution layer.

2. The DMC-UNet-based lung nodule segmentation method of claim 1, wherein, The pre-processing process in step S1 is as follows: S1.1, download the LIDC public data set, and pre-process the LIDC public data set, wherein each sample in the LIDC public data set contains four radiologists' annotation information; for each annotation information, perform or operation to generate a new mask image; in each sample, if a doctor annotates a certain region as a nodule, the region is regarded as a nodule; S1.2, crop the pre-processed lung CT image, crop so that the nodule determined in step S1.2 is located at the center of the image, and at the same time, scale the resolution of the image to 128x128, after processing the image, combine the low resolution image and the high resolution image into information pairs to make a data set.

3. The DMC-UNet-based lung nodule segmentation method of claim 1, wherein, In step S2, the lightweight residual structure is connected in residual connection by two DSC units and one DSC unit, and finally activated by a ReLU activation function, wherein the DSC unit is a depth separable convolution unit composed of a 3x3 channel-wise convolution and a 1x1 point-wise convolution.

4. The DMC-UNet-based lung nodule segmentation method of claim 1, wherein, In step S2.2, the downsampling operation is specifically a 2x2 max pooling operation to compress the resolution to 1 / 2 of the input feature map.

5. The DMC-UNet-based lung nodule segmentation method of claim 1, wherein, In step S2.4, the improved ESPCN super-resolution technology is to replace the original Hidden layers of the ESPCN with a multi-scale feature fusion module to realize feature extraction of the feature map on the basis of the traditional ESPCN super-resolution technology. Herein, the multi-scale feature fusion module is referred to as an ASPP module.

6. The DMC-UNet-based lung nodule segmentation method of claim 5, wherein, The ASPP module is composed of four parallel branches, wherein the first branch is a 1x1 convolution layer, and the other three branches use 3x3 dilated convolution with dilated coefficients of 2, 4 and 8 respectively; the feature information of the four branches is spliced in the dimension, and the channel number is changed by a 1x1 convolution layer to obtain the output of the ASPP module.

7. The DMC-UNet-based lung nodule segmentation method of claim 1, wherein, In step S2.5, the CCA attention module is an X / Y CAM structure obtained by fusing CAM attention and CoordAttention attention; the X / Y CAM structure is specifically: First, the operations in the X direction and in the Y direction are to integrate the input feature map into the CoordAttention, wherein, In the X direction, the input feature map F is divided into two parts, F1 and F2, in the Y direction, and the feature map F1 is divided into two parts, F1a and F1b, in the X direction, and the feature map F2 is divided into two parts, F2a and F2b, in the X direction. , and and , where (H, W) is the resolution of the input feature map F, and C is the channel number of the feature map F; then, the obtained feature map is input into a shared weight double-layer neural network Shared MLP, in which the channels are compressed by a compression ratio r1 to obtain and , and then the channel number is restored to the channel number before input; finally, the results output by the Shared MLP are subjected to an Element-wise addition operation, and a vector of is generated after passing through a Sigmoid activation function. In the Y direction, the operation is the same as in the X direction, and finally, the operation in the Y direction generates a vector of ​ After being integrated into CoordAttention, the vector is reshaped to , and then is obtained by concatenating the vectors in the spatial dimension. Then, the channel compression is performed by 1x1 convolution, and the compression ratio is r2, and is obtained. Then, the input features have the same mean and variance by Batch Normalization. Then, the Non-linear function, also known as h_swish function, is used to help the network learn more complex features. Then, the vector is obtained by splitting the concatenated vector in the spatial dimension. Then, the vectors and are obtained by separating the vector in the spatial dimension. Then, the vectors and are obtained by restoring the vectors in the spatial dimension. Then, the channel number of the input feature map is restored by 1x1 convolution, and the feature map is obtained after weighting operation.

8. The DMC-UNet-based lung nodule segmentation method of claim 7, wherein, The h_swish function is a nonlinear activation function composed of two parts: first, a "h-sigmoid" function similar in form to the Sigmoid function, but using ReLU6 as its approximation function, and ReLU6 is a truncated ReLU; second, a linear transformation that scales and translates the input data.

9. The DMC-UNet-based lung nodule segmentation method of claim 1, wherein, In step S2.6, the output of each layer is fused with the features of the same resolution of the encoder and the decoder in a skip-link manner.

Images