A two-stage network-based pulmonary nodule detection method for CT sequence images

Abstract

The present invention discloses a method for detecting lung nodules in CT sequence images based on a two-stage network. The method is specifically implemented as follows: (1) establishing a training dataset containing original CT sequence images and their corresponding lung nodule annotations; (2) constructing a two-stage network for lung nodule detection, comprising a U-shaped main network for extracting candidate lung nodules and a parasitic network for generating a false-positive suppression mask for lung nodules; (3) constructing a cross-entropy loss function with margins; (4) training the network using the training dataset; and (5) testing the trained network on the CT sequence images to be detected to obtain lung nodule detection results. By adopting a two-stage network, the present invention can effectively suppress false-positive nodules and improve the accuracy of lung nodule detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a method for detecting lung nodules in CT sequence images based on a two-stage network. Background Art

[0002] Lung cancer is the most common cancer worldwide, with the highest malignancy and mortality rate among all cancers. Early detection, diagnosis, and treatment of lung cancer are crucial for improving patient survival. Early diagnosis of lung cancer relies on the detection and localization of lung nodules in medical imaging. Compared to positron emission tomography (PET) and magnetic resonance imaging (MRI), chest computed tomography (CT) is faster, less expensive, and has higher density resolution, making it widely used for lung nodule detection. CT imaging uses a large number of slices, typically generating approximately 300 slices per patient. Manual visual inspection is labor-intensive and prone to missed or false positives. In recent years, the number of slices generated by lung CT has increased dramatically, placing a heavy burden on physicians in image interpretation and diagnosis. Therefore, developing efficient and accurate automatic methods for detecting lung nodules in CT images is crucial for improving the accuracy and efficiency of computer-assisted diagnosis and treatment of lung cancer.

[0003] Pulmonary nodules in CT images often exhibit blurred edges, heterogeneous grayscale, and diverse shapes. Some malignant nodules can be as small as 3–4 mm in diameter, making their automatic identification and detection extremely challenging. Existing methods for pulmonary nodule detection are primarily categorized as traditional, machine learning, and deep learning. Traditional methods are computationally cumbersome, rely on human-computer interaction across different programs, and generally have high false-positive rates. Machine learning relies on the selection of a large number of artificial features. Artificial features designed based on prior knowledge are often blind, incomplete, and inefficient, failing to effectively capture deep discriminative features in the image. Furthermore, most classifiers have limited fitting capabilities and perform poorly for linearly inseparable feature samples.

[0004] Deep learning can automatically learn efficient and more discriminative features from training data and achieve end-to-end training and testing. The deep learning networks currently used for lung nodule detection mainly include two-dimensional and three-dimensional networks. Two-dimensional networks have difficulty in acquiring three-dimensional texture and shape features, and are prone to misdetecting blood vessels as lung nodules, resulting in a high false positive rate. Therefore, most of the current lung nodule detection networks are designed based on three-dimensional convolution. However, almost all three-dimensional lung nodule detection methods use three-dimensional maximum or mean pooling. This pooling operation causes too much spatial compression of the feature map, which can easily lead to the loss of structural information of the image in shallow networks. Summary of the Invention

[0005] In response to the shortcomings and deficiencies of the existing technology, the present invention provides a method for detecting lung nodules in CT sequence images based on a two-stage network. Modules such as parallel downsampling, dense blocks, and hybrid attention are introduced into the two-stage network to improve the network's feature extraction capability and reduce the false positive rate of lung nodule detection.

[0006] A method for detecting lung nodules in CT sequence images based on a two-stage network comprises the following steps:

[0007] (1) Establish a training dataset A containing original CT sequence images and their corresponding pulmonary nodule annotations, and divide the training dataset A into a training subset and a validation subset in a 4:1 ratio;

[0008] (2) Construct a two-stage network for lung nodule detection, called TS-Net, which specifically includes:

[0009] (2-a) A U-shaped main network based on parallel downsampling and dense blocks is constructed, called PPD-UNet, for candidate lung nodule extraction from CT sequence images. The network specifically includes: a U-shaped structure with skip connections is used as the network benchmark framework, in which the encoding path is composed of a cascade of 5 encoding blocks, and the decoding path is composed of a cascade of 4 decoding blocks; the first encoding block is composed of two 3×3×3 convolutional layers, the second, third, and fourth encoding blocks have the same structure, and are all composed of a parallel downsampling block and a dense block, and the fifth encoding block is composed of a parallel downsampling block, a dense block and a deconvolution block with a stride of 2; the input of the first encoding block is the original CT sequence image, and the output of the previous encoding block is used as the input of the next encoding block; The structures of the first, second and third decoding blocks are the same, all consisting of a concatenation operation, a dense block and a deconvolution block with a stride of 2. The fourth decoding block consists of a concatenation operation, a 3×3×3 convolution and a Sigmoid activation function. Due to the introduction of skip connections, the outputs of the fourth and fifth encoding blocks are both used as the input of the first decoding block, the outputs of the third encoding block and the first decoding block are both used as the input of the second decoding block, the outputs of the second encoding block and the second decoding block are both used as the input of the third decoding block, and the outputs of the first encoding block and the third decoding block are both used as the input of the fourth decoding block. The original CT sequence image is the input of PPD-UNet, and the output of the fourth decoding block is the output of PPD-UNet.

[0010] (2-b) The parallel downsampling module described in step (2-a) has the following specific structure: Where C represents the number of channels of the input feature X, 2D, 2H, and 2W represent the depth, height, and width of the three-dimensional feature map of each channel of the input feature X, respectively. First, mean pooling, maximum pooling, and strided convolution with a step size of 2 are used to perform the convolution. Downsample and obtain features And splice X1, X2, and X3 to obtain Then X C Perform 3×3×3 convolution on the three branches to obtain features Then respectively Perform global mean pooling and Sigmoid activation to obtain three groups of channel vectors of length C And perform point multiplication based on the broadcast mechanism on X1 and W1, X2 and W2, and X3 and W3 respectively, that is, each pixel of the i-th channel feature map in X1 is multiplied by the i-th value in the channel vector W1, i = 1, 2, .., C, each pixel of the i-th channel feature map in X2 is multiplied by the i-th value in the channel vector W2, and each pixel of the i-th channel feature map in X3 is multiplied by the i-th value in the channel vector W3, thereby obtaining the feature Finally

[0011] Add up the elements of to get the output feature

[0012] (2-c) The dense block described in step (2-a) has a specific structure including:

[0013] (2-c-i) Construct a convolution combination consisting of "batch normalization + Relu activation + 3×3×3 convolution + batch normalization + Relu activation + 1×1×1 convolution";

[0014] (2-c-ⅱ) sequentially connect n convolution combinations described in step (2-ci), and add a splicing operation after all convolution combinations except the last one, wherein the splicing operation connected after the first convolution combination is used to splice the input features of the dense block with the output of the first convolution combination, and the remaining splicing operations are used to splice the output of the current convolution combination with the output of the previous convolution combination; the n is preferably a natural number between 3 and 8.

[0015] (2-d) Using the parallel downsampling described in step (2-b), the dense block described in step (2-c), and the hybrid attention module, a parasitic network is constructed to generate a false positive suppression mask for lung nodules. The parasitic network is called JS-Net, and its specific structure includes: first, the input CT sequence image is sequentially subjected to 3×3×3 convolution, dense block feature extraction, and parallel downsampling to obtain a feature map F1, and then F1 is concatenated with the output of the dense block in the third decoding block of PPD-UNet to obtain a feature map F2, and then F2 is sequentially subjected to hybrid attention, dense block, and deconvolution module with a stride of 2 to obtain a feature map F3, and F3 is concatenated with the output of the 3×3×3 convolution in the fourth decoding block of PPD-UNet to obtain a feature map F4, and finally F4 is sequentially subjected to hybrid attention, 3×3×3 convolution, and Sigmoid activation to obtain the output of the JS-Net network;

[0016] (2-e) The hybrid attention module described in step (2-d) has a specific structure including:

[0017] (2-e-ⅰ) Two parallel branches are used to obtain the channel attention weight and pixel attention weight of the input feature, specifically including: Where c1 represents the number of channels of the input feature f, d1, h1, and w1 represent the depth, height, and width of the three-dimensional feature map of each channel of the input feature f, respectively. Perform global mean pooling, 3×3×3 convolution, Relu activation, 3×3×3 convolution and Sigmoid activation in sequence to obtain channel attention weights The second branch pair Perform 3×3×3 convolution, Relu activation, 3×3×3 convolution and Sigmoid activation in sequence to obtain pixel attention weights

[0018] (2-e-ⅱ) According to the broadcast mechanism, the weight and Add together to get the mixed attention weight

[0019] (2-e-ⅲ) Input features With mixed attention weights Perform dot multiplication to obtain output features;

[0020] (2-f) Multiply the outputs of PPD-UNet and JS-Net to get the output of TS-Net;

[0021] (3) Construct a cross entropy loss function MCELoss with margins:

[0022]

[0023]

[0024] Among them, p j represents the probability of predicting the jth pixel as a lung nodule, g j represents the true probability that the jth pixel belongs to a lung nodule. α and β are preset parameters. α is used to balance positive and negative samples, and β is used to set the lossless interval. α is preferably a constant between 0.3 and 0.7, and β is preferably a constant between 0.4 and 0.8. Simple and easy-to-classify samples usually come from normal lung areas. Although the loss generated by simple samples is small, due to the large proportion of normal lung areas in CT images, a large amount of loss will still be generated. Considering the training difficulties caused by class imbalance, the MCELoss function is used to calculate the loss of |g j -p j For simple samples with |<β, no statistical loss is performed, and the cross entropy function is narrowed and mapped to the difficult sample interval, so that the network focuses on the identification of lung nodules and false positive areas;

[0025] (4) The TS-Net network is trained using the training dataset A. The specific training strategy includes: first, PPD-UNet is trained with an initial learning rate of 0.01. When the MCELoss value of the validation subset does not decrease for more than 5 consecutive epochs during training, the learning rate is reduced to 1 / 10 of the original value; in addition, JS-Net is trained after 10 epochs, and its learning rate adjustment strategy is consistent with that of PPD-UNet;

[0026] (5) The trained TS-Net is used to test the CT sequence images to be detected, and the probability of each pixel in the image belonging to a lung nodule is obtained. Finally, the lung nodule detection result can be obtained by introducing a threshold λ, and the λ is preferably a constant between 0.3 and 0.7. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram of the TS-Net network structure of the embodiment of the present invention

[0028] Figure 2 Schematic diagram of the parallel downsampling module structure of the embodiment of the present invention

[0029] Figure 3 Schematic diagram of dense block structure of an embodiment of the present invention

[0030] Figure 4 Schematic diagram of the hybrid attention module structure of the embodiment of the present invention

[0031] Figure 5 Some experimental results obtained by the embodiments of the present invention are shown in the following examples: Figure 5(a)~ Figure 5 (c) are three CT slice images randomly selected from the test data. Figure 5 (d)~ Figure 5 (f) is an example of a probability map of candidate lung nodules obtained by the PPD-UNet network. Figure 5 (g)~ Figure 5 (i) Example of false positive suppression mask for lung nodules obtained by JS-Net network, Figure 5 (g)~ Figure 5 (i) is an example of the result obtained by performing dot product on the outputs of PPD-UNet and JS-Net. Figure 5 (m)~ Figure 5 (o) is an example of the final node detection result obtained by the embodiment of the present invention DETAILED DESCRIPTION

[0032] Example 1

[0033] A method for detecting lung nodules in CT sequence images based on a two-stage network is proposed. The specific implementation steps are as follows: (1) Establish a training dataset A containing original CT sequence images and their corresponding lung nodule annotations, and

[0034] The training dataset A is divided into a training subset and a validation subset in a 4:1 ratio;

[0035] (2) Construct a two-stage network for lung nodule detection, called TS-Net, with the following structure: Figure 1 As shown, specifically including:

[0036] (2-a) A U-shaped main network based on parallel downsampling and dense blocks is constructed, called PPD-UNet, for candidate lung nodule extraction from CT sequence images. The network specifically includes: a U-shaped structure with skip connections is used as the network benchmark framework, in which the encoding path is composed of a cascade of 5 encoding blocks, and the decoding path is composed of a cascade of 4 decoding blocks; the first encoding block is composed of two 3×3×3 convolutional layers, the second, third, and fourth encoding blocks have the same structure, and are all composed of a parallel downsampling block and a dense block, and the fifth encoding block is composed of a parallel downsampling block, a dense block and a deconvolution block with a stride of 2; the input of the first encoding block is the original CT sequence image, and the output of the previous encoding block is used as the input of the next encoding block; The structures of the first, second and third decoding blocks are the same, all consisting of a concatenation operation, a dense block and a deconvolution block with a stride of 2. The fourth decoding block consists of a concatenation operation, a 3×3×3 convolution and a Sigmoid activation function. Due to the introduction of skip connections, the outputs of the fourth and fifth encoding blocks are both used as the input of the first decoding block, the outputs of the third encoding block and the first decoding block are both used as the input of the second decoding block, the outputs of the second encoding block and the second decoding block are both used as the input of the third decoding block, and the outputs of the first encoding block and the third decoding block are both used as the input of the fourth decoding block. The original CT sequence image is the input of PPD-UNet, and the output of the fourth decoding block is the output of PPD-UNet.

[0037] (2-b) The parallel downsampling module described in step (2-a) has the following structure: Figure 2 As shown, specifically including: for input features Where C represents the number of channels of the input feature X, 2D, 2H, and 2W represent the depth, height, and width of the three-dimensional feature map of each channel of the input feature X, respectively. First, mean pooling, maximum pooling, and strided convolution with a step size of 2 are used to perform the convolution. Downsample and obtain features And splice X1, X2, and X3 to obtain Then X C Perform 3×3×3 convolution on the three branches to obtain features Then respectively Perform global mean pooling and Sigmoid activation to obtain three groups of channel vectors of length C And perform point multiplication based on the broadcast mechanism on X1 and W1, X2 and W2, and X3 and W3 respectively, that is, each pixel of the i-th channel feature map in X1 is multiplied by the i-th value in the channel vector W1, i = 1, 2, .., C, each pixel of the i-th channel feature map in X2 is multiplied by the i-th value in the channel vector W2, and each pixel of the i-th channel feature map in X3 is multiplied by the i-th value in the channel vector W3, thereby obtaining the feature Finally Add up the elements of to get the output feature

[0038] (2-c) The dense block described in step (2-a) has a structure such as Figure 3 As shown, specifically including:

[0039] (2-c-i) Construct a convolution combination consisting of "batch normalization + Relu activation + 3×3×3 convolution + batch normalization + Relu activation + 1×1×1 convolution";

[0040] (2-c-ii) sequentially connect n convolution combinations described in step (2-ci), and add a splicing operation after all convolution combinations except the last one, wherein the splicing operation connected after the first convolution combination is used to splice the input features of the dense block with the output of the first convolution combination, and the remaining splicing operations are used to splice the output of the current convolution combination with the output of the previous convolution combination; in this embodiment, n=5 is preferred;

[0041] (2-d) Using the parallel downsampling described in step (2-b), the dense block described in step (2-c), and the hybrid attention module, a parasitic network is constructed to generate a false positive suppression mask for lung nodules. The parasitic network is called JS-Net, and its specific structure includes: first, the input CT sequence image is sequentially subjected to 3×3×3 convolution, dense block feature extraction, and parallel downsampling to obtain a feature map F1, and then F1 is concatenated with the output of the dense block in the third decoding block of PPD-UNet to obtain a feature map F2, and then F2 is sequentially subjected to hybrid attention, dense block, and deconvolution module with a stride of 2 to obtain a feature map F3, and F3 is concatenated with the output of the 3×3×3 convolution in the fourth decoding block of PPD-UNet to obtain a feature map F4, and finally F4 is sequentially subjected to hybrid attention, 3×3×3 convolution, and Sigmoid activation to obtain the output of the JS-Net network;

[0042] (2-e) The hybrid attention module described in step (2-d) has the following structure: Figure 4 As shown, specifically including:

[0043] (2-e-ⅰ) Two parallel branches are used to obtain the channel attention weight and pixel attention weight of the input feature, specifically including: Where c1 represents the number of channels of the input feature f, d1, h1, and w1 represent the depth, height, and width of the three-dimensional feature map of each channel of the input feature f, respectively. Perform global mean pooling, 3×3×3 convolution, Relu activation, 3×3×3 convolution and Sigmoid activation in sequence to obtain channel attention weights The second branch pair Perform 3×3×3 convolution, Relu activation, 3×3×3 convolution and Sigmoid activation in sequence to obtain pixel attention weights

[0044] (2-e-ⅱ) According to the broadcast mechanism, the weight and Add together to get the mixed attention weight

[0045] (2-e-ⅲ) Input features With mixed attention weights Perform dot multiplication to obtain output features;

[0046] (2-f) Multiply the outputs of PPD-UNet and JS-Net to get the output of TS-Net;

[0047] (3) Construct a cross entropy loss function MCELoss with margins:

[0048]

[0049]

[0050] Among them, p j represents the probability of predicting the jth pixel as a lung nodule, g j represents the true probability that the jth pixel belongs to a lung nodule. α and β are preset parameters. α is used to balance positive and negative samples. In this embodiment, α = 0.5 is preferred. β is used to set the lossless interval. In this embodiment, β = 0.7 is preferred. Samples that are easy to classify usually come from the normal lung area. Although the loss generated by simple samples is small, the normal lung area accounts for a large proportion in CT images, so a large amount of loss will still be generated. Considering the training difficulty caused by class imbalance, the MCELoss function is used to calculate the loss of |g j -p j For simple samples with |<β, no statistical loss is performed, and the cross entropy function is narrowed and mapped to the difficult sample interval, so that the network focuses on the identification of lung nodules and false positive areas;

[0051] (4) The TS-Net network is trained using the training dataset A. The specific training strategy includes: first, PPD-UNet is trained with an initial learning rate of 0.01. When the MCELoss value of the validation subset does not decrease for more than 5 consecutive epochs during training, the learning rate is reduced to 1 / 10 of the original value; in addition, JS-Net is trained after 10 epochs, and its learning rate adjustment strategy is consistent with that of PPD-UNet;

[0052] (5) The trained TS-Net is used to test the CT sequence images to be detected, and the probability of each pixel in the image belonging to a lung nodule is obtained. Finally, the lung nodule detection result can be obtained by introducing a threshold λ. In this embodiment, λ = 0.5 is preferred.

[0053] Example 2

[0054] The method described in Example 1 was used to conduct experiments on the LUNA16 public lung nodule dataset. LUNA16 contains 888 low-dose chest CT image sequences and 1,186 lung nodules annotated with their locations and diameters. In the experiment, these 888 CT sequences were divided into a test dataset and a training dataset in a ratio of 2:8.

[0055] Figure 5 Some experimental results on the test set are shown. Figure 5 (a)~ Figure 5 (c) Three CT slice images randomly selected from the test data. Figure 5 (d)~ Figure 5 (1) is to adopt the method in Example 1 to Figure 5 (a)~ Figure 5 (d) Intermediate results obtained from performing the test, wherein: Figure 5 (d)~ Figure 5 (f) is the probability map of candidate lung nodules obtained by the PPD-UNet network. It can be seen that although the lung nodules are effectively detected, there are also some misdetected nodules; Figure 5 (g)~ Figure 5 (i) False positive suppression mask of lung nodules obtained by JS-Net network. It can be seen that most of the misdetected nodules are effectively suppressed; Figure 5 (g)~ Figure 5 (i) is the result of performing dot product on the outputs of PPD-UNet and JS-Net. It can be seen that the lung nodule area has a higher probability value, while the normal area has a lower probability value. By introducing the threshold λ = 0.5, the final nodule detection result is as follows: Figure 5 (m)~ Figure 5 (o) shown.

[0056] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for detecting lung nodules in CT sequence images based on a two-stage network, characterized in that: The following steps are involved: (1) Establish a training dataset A containing original CT sequence images and their corresponding pulmonary nodule annotations, and divide the training dataset A into a training subset and a validation subset in a 4:1 ratio; (2) Construct a two-stage network for lung nodule detection, called TS-Net, which specifically includes: (2-a) A U-shaped main network based on parallel downsampling and dense blocks is constructed, called PPD-UNet, for candidate lung nodule extraction from CT sequence images. The network specifically includes: a U-shaped structure with skip connections is used as the network benchmark framework, in which the encoding path is composed of a cascade of 5 encoding blocks, and the decoding path is composed of a cascade of 4 decoding blocks; the first encoding block is composed of two 3×3×3 convolutional layers, the second, third, and fourth encoding blocks have the same structure, and are all composed of a parallel downsampling block and a dense block, and the fifth encoding block is composed of a parallel downsampling block, a dense block and a deconvolution block with a stride of 2; the input of the first encoding block is the original CT sequence image, and the output of the previous encoding block is used as the input of the next encoding block; The structures of the first, second and third decoding blocks are the same, all consisting of a concatenation operation, a dense block and a deconvolution block with a stride of 2. The fourth decoding block consists of a concatenation operation, a 3×3×3 convolution and a Sigmoid activation function. Due to the introduction of skip connections, the outputs of the fourth and fifth encoding blocks are both used as the input of the first decoding block, the outputs of the third encoding block and the first decoding block are both used as the input of the second decoding block, the outputs of the second encoding block and the second decoding block are both used as the input of the third decoding block, and the outputs of the first encoding block and the third decoding block are both used as the input of the fourth decoding block. The original CT sequence image is the input of PPD-UNet, and the output of the fourth decoding block is the output of PPD-UNet. (2-b) The parallel downsampling module described in step (2-a) has the following specific structure: Where C represents the number of channels of the input feature X, 2D, 2H, and 2W represent the depth, height, and width of the three-dimensional feature map of each channel of the input feature X, respectively. First, mean pooling, maximum pooling, and strided convolution with a step size of 2 are used to perform the convolution. Downsample and obtain features And splice X1, X2, and X3 to obtain Then X C Perform 3×3×3 convolution on the three branches to obtain features Then respectively Perform global mean pooling and Sigmoid activation to obtain three groups of channel vectors of length C And perform point multiplication based on the broadcast mechanism on X1 and W1, X2 and W2, and X3 and W3 respectively, that is, each pixel of the i-th channel feature map in X1 is multiplied by the i-th value in the channel vector W1, i = 1, 2, .., C, each pixel of the i-th channel feature map in X2 is multiplied by the i-th value in the channel vector W2, and each pixel of the i-th channel feature map in X3 is multiplied by the i-th value in the channel vector W3, thereby obtaining the feature Finally Add up the elements of to get the output feature (2-c) The dense block described in step (2-a) has a specific structure including: (2-c-i) Construct a convolution combination consisting of "batch normalization + Relu activation + 3×3×3 convolution + batch normalization + Relu activation + 1×1×1 convolution"; (2-c-ii) sequentially connect n convolution combinations described in step (2-ci), and add a splicing operation after all convolution combinations except the last one. The splicing operation connected after the first convolution combination is used to splice the input features of the dense block with the output of the first convolution combination, and the remaining splicing operations are used to splice the output of the current convolution combination with the output of the previous convolution combination; (2-d) Using the parallel downsampling described in step (2-b), the dense block described in step (2-c), and the hybrid attention module, a parasitic network is constructed to generate a false positive suppression mask for lung nodules. The parasitic network is called JS-Net, and its specific structure includes: first, the input CT sequence image is sequentially subjected to 3×3×3 convolution, dense block feature extraction, and parallel downsampling to obtain a feature map F1, and then F1 is concatenated with the output of the dense block in the third decoding block of PPD-UNet to obtain a feature map F2, and then F2 is sequentially subjected to hybrid attention, dense block, and deconvolution module with a stride of 2 to obtain a feature map F3, and F3 is concatenated with the output of the 3×3×3 convolution in the fourth decoding block of PPD-UNet to obtain a feature map F4, and finally F4 is sequentially subjected to hybrid attention, 3×3×3 convolution, and Sigmoid activation to obtain the output of the JS-Net network; (2-e) The hybrid attention module described in step (2-d) has a specific structure including: (2-e-ⅰ) Two parallel branches are used to obtain the channel attention weight and pixel attention weight of the input feature, specifically including: Where c1 represents the number of channels of the input feature f, d1, h1, and w1 represent the depth, height, and width of the three-dimensional feature map of each channel of the input feature f, respectively. Perform global mean pooling, 3×3×3 convolution, Relu activation, 3×3×3 convolution and Sigmoid activation in sequence to obtain channel attention weights The second branch pair Perform 3×3×3 convolution, Relu activation, 3×3×3 convolution and Sigmoid activation in sequence to obtain pixel attention weights (2-e-ⅱ) According to the broadcast mechanism, the weight and Add them together to get the mixed attention weight (2-e-ⅲ) Input features With mixed attention weights Perform dot multiplication to obtain output features; (2-f) Multiply the outputs of PPD-UNet and JS-Net to get the output of TS-Net; (3) Construct a cross entropy loss function MCELoss with margins: Among them, p j represents the probability of predicting the jth pixel as a lung nodule, g j represents the true probability that the jth pixel belongs to a lung nodule. α and β are preset parameters. α is used to balance positive and negative samples, and β is used to set the lossless interval. Simple and easy-to-classify samples usually come from normal lung areas. Although the loss generated by simple samples is small, due to the large proportion of normal lung areas in CT images, a large amount of loss will still be generated. Considering the training difficulties caused by class imbalance, the MCELoss function is used to calculate the probability of |g j -p j Simple samples with |<β do not have statistical loss, and the cross entropy function is narrowed and mapped to the difficult sample interval, so that the network focuses on the identification of lung nodules and false positive areas; (4) The TS-Net network is trained using the training dataset A. The specific training strategy includes: first, PPD-UNet is trained with an initial learning rate of 0.

01. When the MCELoss value of the validation subset does not decrease for more than 5 consecutive epochs during training, the learning rate is reduced to 1 / 10 of the original value; in addition, JS-Net is trained after 10 epochs, and its learning rate adjustment strategy is consistent with that of PPD-UNet; (5) The trained TS-Net is used to test the CT sequence images to be detected, and the probability that each pixel in the image belongs to a lung nodule is obtained. Finally, by introducing the threshold λ, the lung nodule detection result can be obtained.

2. The method for detecting lung nodules in CT sequence images based on a two-stage network as claimed in claim 1, wherein: The parameter n described in step (2-c-ⅱ) is preferably a natural number between 3 and 8, the parameter α described in step (3) is preferably a constant between 0.3 and 0.7, the parameter β is preferably a constant between 0.4 and 0.8, and the parameter λ described in step (5) is preferably a constant between 0.3 and 0.7.

Images