Lung Nodule Detection Method and Device Based on Dual-Channel Dense Connection
By adopting a dual-channel dense connection structure and SENet attention block in the detection of lung nodules, the problems of low feature map utilization and network degradation are solved, and the detection accuracy is significantly improved.
Patent Information
- Application Number
- CN202111522143.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-13
- Publication Date
- 2025-05-30
- Estimated Expiration
- 2041-12-13
AI Technical Summary
In the detection of lung nodules, the problem of network degradation and low detection accuracy in the detection of pulmonary nodules is low.
The lung nodule detection method based on dual-channel dense connection is adopted, and the dual-channel dense structure and SENet attention block are designed to improve the efficiency of feature extraction and detection accuracy, and solve the problem of network degradation.
The utilization rate of the network for feature maps is improved, combined with residual connection and attention mechanism, the problem of network degradation during deep learning training is solved, and the detection accuracy is significantly improved.
Smart Images
Figure CN114240873B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of target detection, and relates to target detection based on a deep learning network model and an attention mechanism. Specifically, it relates to a method and device for detecting pulmonary nodules based on dual-channel dense connection. Background Art
[0002] Currently, medical detection of the lungs mainly uses computed tomography to obtain CT images of the lungs. Pulmonary nodules have various types, diverse structures, uncertain sizes and volumes, and unfixed positions, and may adhere to the surrounding lung environment. These characteristics result in their morphological similarity to normal tissues, making it extremely easy to confuse the two. The above characteristics of nodules make it a very heavy task for doctors to view CT images. Therefore, using computer technology to detect pulmonary nodules has become a trend.
[0003] Currently, deep learning technology is widely used in the detection of pulmonary nodules. In the field of medical images, methods based on convolutional neural networks have achieved good results, and neural networks represented by U-Net have shown excellent performance. Zhu et al. proposed a DeepLung network that combines a three-dimensional dual-path network with a U-Net structure for the automatic detection and classification of pulmonary nodules. Cao et al. proposed a two-stage method for pulmonary nodule detection. The initial detection of candidate nodules uses an improved U-Net network, and a 3D-CNN classification network is constructed to reduce false positives. Gong et al. used a three-dimensional deep squeeze-and-excitation network to automatically detect pulmonary nodules. A three-dimensional U-shaped structure network and a 3D-CNN network are respectively used for candidate detection and false positive reduction of pulmonary nodules. A 3D-SE-ResNet module is added to the model to effectively learn nodule features and improve the detection performance of the model. Wang et al. proposed a dual-attention 3D-UNet network to fuse the attention module to improve the segmentation accuracy of pulmonary nodules.
[0004] The pulmonary nodule detection methods represented by the above achievements have greatly improved the detection efficiency of computers for pulmonary nodules and reduced the workload of doctors. However, there are still the following problems in the current detection of pulmonary nodules: the utilization rate of feature maps in the network is relatively low, the problem of network degradation occurs after the network depth is increased, and the detection accuracy is relatively low. Summary of the Invention
[0005] Aiming at the deficiencies of the prior art, the present invention proposes a method and device for detecting pulmonary nodules based on dual-channel dense connection, designs a dual-channel dense connection structure, improves the utilization rate of feature maps and the recognition accuracy of the detection model, and solves the problem of network degradation after training deepening.
[0006] The method for detecting pulmonary nodules based on dual-channel dense connection specifically includes the following steps:
[0007] Step 1: Image preprocessing
[0008] Input the lung CT image into the preprocessing module. After enhancing the feature dimension, perform feature extraction. The preprocessing module includes two 3D convolutional blocks, and each 3D convolutional block is sequentially connected by a convolutional layer with a kernel size of 3, a batch normalization layer, and a ReLU activation function layer.
[0009] Step 2: Feature Extraction
[0010] Construct a dual-path dense structure for feature extraction to enhance feature utilization. The dual-path dense structure includes two identical convolutional paths. Each convolutional path includes three sequentially connected 3D convolutional blocks and a SENet attention block. The input features of the convolutional path are concatenated and input into the 3D convolutional block in a dense connection manner. Then, the output features of the SENet attention block are added to the input features of the convolutional path at the element level as the output features of the convolutional path. After the output features of the two convolutional paths are concatenated in the channel dimension, they are used as the output of the dual-path dense structure.
[0011] s2.1. Construct the downsampling module group
[0012] The downsampling module group includes the first to fifth downsampling modules, the first to fourth pooling layers, and a special deconvolution layer. The output image of the preprocessing module sequentially passes through the first pooling layer, the first downsampling module, the second pooling layer, the second downsampling module, the third pooling layer, the third downsampling module, the fourth pooling layer, and the fourth downsampling module as the output of the downsampling module group. Among them, after the output of the third downsampling module passes through the special deconvolution layer, it is concatenated with the outputs of the first downsampling module and the second downsampling module and then input into the fifth downsampling module for feature enhancement. Among them, the downsampling module adopts a dual-path dense structure.
[0013] s2.2. Construct the upsampling module group
[0014] The upsampling module group includes the first and second upsampling modules and the first and second deconvolution layers. The output of the fourth downsampling module passes through the first deconvolution layer and is concatenated with the output of the third downsampling module in the channel dimension, and then sequentially passes through the first upsampling module and the second deconvolution layer, and then is concatenated with the output of the fifth downsampling module and input into the second upsampling module as the output of the upsampling module group. Among them, the upsampling module adopts a dual-path dense structure.
[0015] Step 3: Lung Nodule Detection
[0016] Perform a Dropout operation on the output of the upsampling module group to randomly inactivate the network structure with a probability of 0.5, and then input it into two convolutional layers with a kernel size of 1. The convolutional layer outputs the possible positions and probabilities of nodules in the lung CT image.
[0017] A lung nodule detection device based on dual-channel dense connection includes a data acquisition module, a detection module, and a display module.
[0018] The data acquisition module is used to acquire lung CT images and input them into the detection module. The detection module is based on a lung nodule detection method with dual-channel dense connection and outputs the possible positions and probabilities of nodules in the lung CT images. The display module is used to display the output results of the detection module.
[0019] The present invention has the following beneficial effects:
[0020] Feature extraction is performed through a dual-channel dense structure to improve the network's utilization rate of feature maps. Combining residual connections and attention mechanisms solves the problem of network degradation during deep learning training and improves the detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 It is a schematic diagram of the detection module structure;
[0022] Figure 2 It is a flow chart of lung nodule detection;
[0023] Figure 3 It is a schematic diagram of the dual-channel dense structure;
[0024] Figure 4 It is a schematic diagram of the SENet attention block structure. DETAILED DESCRIPTION OF THE INVENTION
[0025] The following further explains the present invention with reference to the accompanying drawings;
[0026] A lung nodule detection device based on dual-channel dense connection includes a data acquisition module, a detection module, and a display module.
[0027] The data acquisition module is used to acquire lung CT images and input them into the detection module. As Figure 1 shown, the detection module is based on a network structure with dual-channel dense connection and outputs the possible positions and probabilities of nodules in the lung CT images. The display module is used to display the output results of the detection module.
[0028] As Figure 2 shown, the lung nodule detection method based on dual-channel dense connection includes (1) preprocessing to increase the dimension of the feature map; (2) a downsampling module group to extract features; (3) an upsampling module group to extract features and restore the size; (4) a prediction module to obtain nodule probabilities and position information, specifically:
[0029] Step 1. Image preprocessing
[0030] Input the lung CT images in the LUNA16 dataset into the preprocessing module to enhance the feature dimension. The preprocessing module includes two 3D convolutional blocks, and each 3D convolutional block is sequentially connected by a convolutional layer with a kernel size of 3, a batch normalization layer, and a ReLU activation function layer. After passing through the preprocessing module, the channel dimension of the lung CT image is increased from 1 to 24.
[0031] Step 2: Feature extraction
[0032] Construct a dual-path dense structure for feature extraction to enhance feature utilization. As Figure 3 shown, the dual-path dense structure includes two identical convolutional paths. Each convolutional path includes three 3D convolutional blocks and a SENet attention block connected in sequence. The input features of the convolutional path are concatenated and input into the 3D convolutional block in a dense connection manner. Then, the output features of the SENet attention block are added to the input features of the convolutional path at the element level as the output features of the convolutional path. After the output features of the two convolutional paths are concatenated in the channel dimension, they serve as the output of the dual-path dense structure. As Figure 4 shown, the input of the SENet attention block is processed through a pooling layer, a fully connected layer, a ReLu activation function, a fully connected layer, and a Sigmoid layer in sequence and then concatenated with the original input as the output of the SENet attention block.
[0033] s2.1: Construct the downsampling module group
[0034] The downsampling module group includes the first to fifth downsampling modules, the first to fourth pooling layers, and a special deconvolution layer. The output image of the preprocessing module passes through the first pooling layer, the first downsampling module, the second pooling layer, the second downsampling module, the third pooling layer, the third downsampling module, the fourth pooling layer, and the fourth downsampling module in sequence as the output of the downsampling module group. The size of the pooling layer is 2*2. To fuse the feature maps of different network depths, the output of the third downsampling module is enlarged by a special deconvolution layer with a size of 2*2 and then concatenated with the outputs of the first downsampling module and the second downsampling module, and then input into the fifth downsampling module for feature enhancement. Among them, the downsampling module adopts a dual-path dense structure.
[0035] s2.2: Construct the upsampling module group
[0036] The upsampling module group packages the first and second upsampling modules and the first and second transposed convolution layers with a size of 2×2. The output of the fourth downsampling module passes through the first transposed convolution layer and is concatenated with the output of the third downsampling module in the channel dimension, and then sequentially passes through the first upsampling module and the second transposed convolution layer, and then is concatenated with the output of the fifth downsampling module and input into the second upsampling module as the output of the upsampling module group. Among them, the upsampling module adopts a dual-path dense structure. While extracting features, the upsampling module group can also restore the feature map to the size before being processed by the preprocessing module.
[0037] Step 3: Lung nodule detection
[0038] Perform a Dropout operation on the output of the upsampling module group to randomly deactivate the network structure with a probability of 0.5, and then input it into two convolution layers with a kernel size of 1. The convolution layer outputs the possible positions and probabilities of nodules in the lung CT image.
[0039] The embodiments described above are only descriptions of the preferred example modes of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention should all fall within the protection scope determined by the claims of the present invention.
Claims
1. A method for detecting pulmonary nodules based on dual-channel dense connection, characterized in that: Specifically, it includes the following steps: Step 1. Image preprocessing Input the lung CT image into the preprocessing module. After passing through two 3D convolutional blocks in sequence, the channel dimension is increased; Step 2. Feature extraction Construct a dual-path dense structure composed of two identical convolutional paths. Each convolutional path includes three 3D convolutional blocks and a SENet attention block connected in sequence. The input features of the convolutional path are spliced and input into the 3D convolutional block in a dense connection manner. Then, the output features of the SENet attention block are added to the input features of the convolutional path at the element level as the output features of the convolutional path. After the output features of the two convolutional paths are spliced in the channel dimension, they are used as the output of the dual-path dense structure; s2.
1. Construct the downsampling module group The downsampling module group includes the first to fifth downsampling modules, the first to fourth pooling layers, and a special deconvolution layer. The output image of the preprocessing module passes through four pooling layers and four downsampling modules arranged alternately for feature extraction; Among them, the output of the third downsampling module passes through the special deconvolution layer and is spliced with the outputs of the first downsampling module and the second downsampling module, and then input into the fifth downsampling module for feature enhancement. Among them, the downsampling module adopts the dual-path dense structure; s2.
2. Construct the upsampling module group The upsampling module group includes the first and second upsampling modules and the first and second deconvolution layers. The output of the fourth downsampling module passes through the first deconvolution layer and is spliced with the output of the third downsampling module in the channel dimension, and then passes through the first upsampling module and the second deconvolution layer in sequence, and then is spliced with the output of the fifth downsampling module and input into the second upsampling module to complete feature extraction and size restoration. Among them, the upsampling module adopts the dual-path dense structure; Step 3. Pulmonary nodule detection Perform Dropout operation on the output of the second upsampling module, randomly inactivate the network structure with a probability of 0.5, and then input it into two convolutional layers with a kernel size of 1 to output the position and probability of nodules in the lung CT image.
2. The method for detecting pulmonary nodules based on dual-channel dense connection according to claim 1, characterized in that: The shown 3D convolutional block is composed of a convolutional layer with a kernel size of 3, a batch normalization layer, and a ReLU activation function layer connected in sequence.
3. The method for detecting pulmonary nodules based on dual-channel dense connection according to claim 1, characterized in that: The input of the SENet attention block is processed by passing through a pooling layer, a fully connected layer, a ReLu activation function, a fully connected layer, and a Sigmoid layer in sequence and then spliced with the original input as the output of the SENet attention block.
4. The method for detecting pulmonary nodules based on dual-channel dense connection according to claim 1, characterized in that: The sizes of the pooling layers and deconvolution layers in the downsampling module group and the upsampling module group are both 2*2.
5. A device for detecting pulmonary nodules based on dual-channel dense connection, characterized in that: It includes a detection module and a display module; The detection module extracts the features of the lung CT image and outputs the possible positions and probabilities of nodules in the CT image; the display module is used to display the output results of the detection module; The usage method of the device includes the following steps: Step 1. Image preprocessing Input the lung CT image into the preprocessing module. After passing through two 3D convolutional blocks in sequence, the channel dimension is increased; Step 2. Feature extraction Construct a dual-path dense structure composed of two identical convolutional paths. Each convolutional path includes three 3D convolutional blocks and one SENet attention block connected in sequence. The input features of the convolutional path are spliced and input into the 3D convolutional block in a densely connected manner. Then, the output features of the SENet attention block are added to the input features of the convolutional path at the element level as the output features of the convolutional path. After the output features of the two convolutional paths are spliced in the channel dimension, they are used as the output of the dual-path dense structure; s2.
1. Construct the downsampling module group The downsampling module group includes the first to fifth downsampling modules, the first to fourth pooling layers, and a special deconvolution layer; the output image of the preprocessing module passes through four pooling layers and four downsampling modules arranged alternately for feature extraction; among them, the output of the third downsampling module passes through the special deconvolution layer and is spliced with the outputs of the first downsampling module and the second downsampling module, and then input into the fifth downsampling module for feature enhancement; among them, the downsampling module adopts the dual-path dense structure; s2.
2. Construct the upsampling module group The upsampling module group includes the first and second upsampling modules and the first and second deconvolution layers; the output of the fourth downsampling module passes through the first deconvolution layer and is spliced with the output of the third downsampling module in the channel dimension, and then passes through the first upsampling module and the second deconvolution layer in sequence, and is spliced with the output of the fifth downsampling module and then input into the second upsampling module to complete feature extraction and size restoration; among them, the upsampling module adopts the dual-path dense structure; Step 3. Lung nodule detection Perform Dropout operation on the output of the second upsampling module, randomly inactivate the network structure with a probability of 0.5, and then input it into two convolutional layers with a kernel size of 1 to output the possible positions and probabilities of nodules in the lung CT image.
6. A lung nodule detection device based on dual-channel dense connection, characterized in that: It further includes a data acquisition module for acquiring lung CT images and inputting them into the detection module.