Pulmonary nodule detection method based on single-step detection and cyclic generative adversarial network

Abstract

The invention discloses a pulmonary nodule detection method based on single-step detection and a cyclic generative adversarial network, and the method comprises the steps: cutting out a cubic region from the interior of a lung cavity of a lung CT scanning image, and enabling the cubic region to comprise a nodule image formed by a nodule region and a background region around the nodule region, and an empty background image of a nodule-free region, marking a nodule region bounding box in the nodule image and coordinates corresponding to the nodule region; processing the cubic region by applying a cyclic generative adversarial network model, randomly simulating and converting the nodule image and the empty background image into enhanced samples, and taking the generated enhanced samples and the cut cubic region as training samples; detecting an image with a pulmonary nodule area from the training sample, and classifying the detected pulmonary nodules; according to the method, the detection capability of the nodules is enhanced, and the condition of misrecognition is reduced, so that the specificity is improved, the diversity of samples is enhanced, and the method has stronger capability of distinguishing different types of background regions.

Description

Technical Field

[0001] The present invention relates to the technical field of image enhancement classification, and particularly relates to a pulmonary nodule detection method based on single-step detection and cyclic generative adversarial network. Background Art

[0002] Currently, low-dose CT scans (LDCT) play an important role in pulmonary nodule detection. However, traditional manual reading requires a large amount of training for professional radiologists, with heavy workload and low efficiency. Therefore, developing a computer-aided detection system (CAD) that can improve the sensitivity and specificity of pulmonary nodule detection has important medical value.

[0003] Existing methods usually adopt a two-stage detection framework, but there are still trade-offs between sensitivity and specificity in small target detection. In addition, the imbalance in data distribution between pulmonary nodules and the background further increases the detection difficulty. Summary of the Invention

[0004] The purpose of the present invention is to provide a pulmonary nodule detection method based on single-step detection and cyclic generative adversarial network to solve the technical problem that the imbalance in data distribution between pulmonary nodules and the background in the prior art further increases the detection difficulty.

[0005] To solve the above technical problems, the present invention specifically provides the following technical solutions:

[0006] The present invention provides a pulmonary nodule detection method based on single-step detection and cyclic generative adversarial network, including the following steps:

[0007] Step 100: Crop a cubic region from the lung cavity of the pulmonary CT scan image, where the cubic region includes a nodule image formed by a nodule region and its surrounding background region and an empty background image of a nodule-free region, and mark the boundary box of the nodule region in the nodule image and the coordinates corresponding to the nodule region;

[0008] Step 200: Apply the cyclic generative adversarial network model to process the cubic region, randomly simulate and transform the nodule image and the empty background image into enhanced samples, where the enhanced samples include simulating and transforming the nodule image into a pseudo-background image without nodules, and simulating and generating a pseudo-nodule image with pseudo-nodules from the empty background image, and use the generated enhanced samples and the cropped cubic region as training samples;

[0009] Step 300: Detect the images with pulmonary nodule regions from the training samples and classify the detected pulmonary nodules.

[0010] As a preferred embodiment of the present invention, in the step 200, the specific implementation method of using the cycle generative adversarial network model to process the nodule image in the cube region and generate a pseudo-background image is as follows: the background generator in the cycle generative adversarial network model locates the nodule region according to the morphology and texture features of the nodule, determines the set of real nodule pixels in the nodule region, removes the set of real nodule pixels from the nodule image, and fills the removed position with the pixel values of the background region of the nodule image to generate a pseudo-background image;

[0011] The specific implementation method of using the cycle generative adversarial network model to process the empty background image in the cube region and generate a pseudo-nodule image is as follows: use the pulmonary nodule generator in the cycle generative adversarial network model to generate a pseudo-nodule at the center position in the empty background image, and make the generated pseudo-nodule image consistent with the real nodule in the nodule image in terms of morphology and gray-scale distribution.

[0012] As a preferred embodiment of the present invention, in the step 200, the cycle consistency loss in the cycle generative adversarial network model is used to constrain the pseudo-nodule to ensure that the newly generated nodule in the pseudo-nodule sample remains at the center position of the image, so that the newly generated nodule in the pseudo-nodule sample is spatially consistent with the original nodule region of the cube region.

[0013] As a preferred embodiment of the present invention, in the step 200, an image of a cube region centered on the cropped nodule is input into the cycle generative adversarial network model, where the background region of the cube region image is a fragment of a pulmonary CT image without nodules;

[0014] The generator is used to convert the pixel information in the background region into a region with nodules, that is, to simulate pseudo-nodules in the background region to generate a new virtual image;

[0015] The discriminator is used to classify the new virtual image and judge the authenticity of the new virtual image;

[0016] Through adversarial learning, the generator and the discriminator are optimized so that the generator can generate pseudo-nodules that are visually similar to real nodules;

[0017] The position constraint loss is used to calculate the deviation between the position of the generated pseudo-nodule and the center point of the virtual image, and the cycle consistency loss is used to constrain the position of the newly generated pseudo-nodule to be fixed at the center position of the virtual image.

[0018] As a preferred embodiment of the present invention, in step 200, an image of a cubic region centered on the cropped nodule is input into the cyclic generative adversarial network model, and a pseudo-background is generated in the nodule region of the cubic region image. The specific implementation method is as follows:

[0019] Use the generator to eliminate the features of the nodule region of the cubic region image and generate a pseudo-background region;

[0020] Use the discriminator to learn to distinguish the real background region of the cubic region image and the pseudo-background region generated by the generator.

[0021] Optimize the generator and discriminator through adversarial learning, and constrain the position of the pseudo-background region through the cyclic consistency loss, so that the pseudo-background region generated by the generator is consistent with the real background region.

[0022] As a preferred embodiment of the present invention, the specific implementation steps for optimizing the performance of the generator and discriminator through adversarial learning are as follows:

[0023] Initialize the generator and discriminator, and use random weights to initialize the parameters of the generator and discriminator;

[0024] Define an adversarial loss function, where the adversarial loss function combines the generator loss, discriminator loss, and cyclic consistency loss;

[0025] Train the adversarial loss function, alternately optimize the generator and discriminator until convergence;

[0026] Evaluate the performance of the generator and discriminator, evaluate the quality of the virtual samples generated by the generator, form a test set with real samples and newly generated virtual samples, and test the classification accuracy of the discriminator so that the discriminator can distinguish real samples and newly generated virtual samples.

[0027] As a preferred embodiment of the present invention, the implementation method for training the adversarial loss function is as follows:

[0028] Extract training samples from the real dataset as training data;

[0029] Calculate the discriminator loss and update the parameters of the discriminator;

[0030] Optimize the parameters of the generator by combining the generator loss and the cyclic consistency loss;

[0031] Alternately optimize the generator and discriminator until convergence.

[0032] As a preferred embodiment of the present invention, in the step 300, first perform pixel-level segmentation on the nodule region of the training sample through the encoding-decoding structure of U-Net, and then process the feature map of the nodule region output by the decoder by adding an additional classification branch in the decoder structure of U-Net to identify the nodule types of different nodule regions, so as to classify the nodule regions according to the nodule types;

[0033] Among them, the output image of the pixel-level segmentation of the nodule region of the training sample is a binary classification image, where the nodule region is marked as 1 and the background region is 0.

[0034] As a preferred embodiment of the present invention, use a multi-task loss function to calculate the accuracy of the segmentation result of the nodule region and calculate the segmentation error of the nodule region;

[0035] Use a multi-task loss function to calculate the accuracy of the classification result of the nodule region and calculate the classification error of the nodule region;

[0036] Based on the segmentation error and the classification error, adjust the encoding-decoding structure parameters of the U-Net until the segmentation error and the classification error meet the expectations.

[0037] As a preferred embodiment of the present invention, the multi-task loss function includes an anchor box classification loss function, a nodule classification loss function, and an anchor box regression loss function;

[0038] Among them, the anchor box classification loss function is a segmentation result loss function for segmenting the nodule region and the background region of the cube region;

[0039] The nodule classification loss function is a classification result loss function for feature classification of the nodule region.

[0040] The present invention has the following beneficial effects compared with the prior art:

[0041] The enhanced samples generated by the present invention through the CycleGAN model can expose to more types of data distributions during the training process. This method not only enhances the nodule detection ability but also enables the model to learn more background features, reducing the occurrence of false positives (misjudging normal regions as nodules), thereby improving the specificity. The diversity of the enhanced samples enables the model to have a stronger discrimination ability for different types of background regions. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only exemplary. For those of ordinary skill in the art, without creative efforts, other implementation drawings can also be obtained based on the provided drawings.

[0043] Figure 1 Schematic flow chart of the lung nodule detection method according to an embodiment of the present invention; Specific embodiments

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0045] As Figure 1 shown, the present invention provides a lung nodule detection method based on single-step detection and cyclic generative adversarial network, including the following steps:

[0046] Step 100: Crop a cubic region from the lung cavity of the lung CT scan image, where the cubic region includes a nodule image formed by the nodule region and its surrounding background region and an empty background image of the nodule-free region, and mark the boundary box of the nodule region in the nodule image and the coordinates corresponding to the nodule region.

[0047] When marking the nodule region in the lung CT scan image, specifically obtain the three-dimensional coordinates and boundary box of the nodule region in the lung CT scan image. These nodules may be lesion tissues of various sizes and shapes. Therefore, during the marking process, it is necessary to provide the position and size information of the nodules. The main purpose of the marking is to provide an accurate target region for the subsequent training process, so as to train the nodule detection model and the cyclic generative adversarial network model.

[0048] Step 200: Apply the cyclic generative adversarial network model to process the cubic region, randomly simulate and transform the nodule image and the empty background image into enhanced samples. The enhanced samples include simulating and transforming the nodule image into a pseudo-background image without nodules, and simulating and generating a pseudo-nodule image with pseudo-nodules from the empty background image. Use the generated enhanced samples and the cropped cubic region as training samples;

[0049] In step 200, the implementation method of using the cycle generative adversarial network model to process the nodule image in the cubic region and generate a pseudo-background image is as follows: The background generator in the cycle generative adversarial network model locates the nodule region according to the morphology and texture features of the nodule, determines the set of real nodule pixels within the nodule region, removes the set of real nodule pixels from the nodule image, and fills the removed position with the pixel values of the background region of the nodule image to generate a pseudo-background image;

[0050] The implementation method of using the cycle generative adversarial network model to process the empty background image in the cubic region and generate a pseudo-nodule image is as follows: The pulmonary nodule generator in the cycle generative adversarial network model generates a pseudo-nodule at the central position in the empty background image, and makes the generated pseudo-nodule image consistent with the real nodules in the nodule image in terms of morphology and gray distribution.

[0051] A pseudo-background sample refers to removing or eliminating the real nodules in the "nodule image" and replacing them with a nodule-free background. The goal of this process is to eliminate the nodules and maintain the background structure. Through this method, samples containing pure background regions can be generated.

[0052] The background region refers to the nodule-free regions in all the cropped cubic regions, that is, the part of the CT image without any nodules. In this background region, the 3D cycle generative adversarial network will generate a pseudo-nodule, that is, simulate the appearance of a nodule in the background region and simulate the possible pseudo-nodule samples in the real situation.

[0053] It should be further noted that the cycle consistency loss constraint in the cycle generative adversarial network model is used to ensure that the newly generated nodules in the pseudo-nodule samples remain at the central position of the image, so that the newly generated nodules in the pseudo-nodule samples are spatially consistent with the original nodule regions in the cubic region.

[0054] The cycle consistency loss plays a constraining role, that is, when the generator generates a pseudo-nodule, the generated nodule must be spatially consistent with the original nodule (for example, in terms of position, size, etc.). This constraint helps to reduce the possible errors in the image conversion process, such as the pseudo-nodule shifting to the wrong position.

[0055] Specifically in step 200, the cubic region image centered on the cropped nodule is input into the cycle generative adversarial network model. The background region of this cubic region image is a fragment of the pulmonary CT image without nodules. The method of simulating a pseudo-nodule from the background region in the real cubic region image centered on the nodule is as follows:

[0056] The generator is used to convert the pixel information in the background region into a region with nodules, that is, simulate a pseudo-nodule in the background region and generate a new virtual image;

[0057] Use a discriminator to classify the new virtual image and determine the authenticity of the new virtual image;

[0058] Optimize the generator and discriminator through adversarial learning so that the generator can generate pseudo nodules that are visually similar to real nodules;

[0059] Use a position constraint loss to calculate the deviation between the position of the generated pseudo nodule and the center point of the virtual image, and constrain the position of the newly generated pseudo nodule to be fixed at the center position of the virtual image through a cycle consistency loss.

[0060] In step 200, input the cropped cube region image of the nodule center into the cycle generative adversarial network model, generate a pseudo background in the nodule region of the cube region image, remove the nodule from the nodule region in the real cube region image of the nodule center, and generate a pseudo background without nodule features. The specific implementation method is as follows:

[0061] Use the generator to eliminate the features of the nodule region of the cube region image and generate a pseudo background region;

[0062] Use the discriminator to learn to distinguish between the real background region of the cube region image and the pseudo background region generated by the generator.

[0063] Optimize the generator and discriminator through adversarial learning, and constrain the position of the pseudo background region through a cycle consistency loss, so that the pseudo background region generated by the generator is consistent with the real background region.

[0064] That is, after inputting the original sample (such as a cube region containing a nodule region and its surrounding background region) into the generator, the generated target sample (a virtual sample image containing a pseudo background) can be output. After inputting the real sample (a cube image containing a nodule region and its surrounding background region) and the virtual sample (a virtual sample image containing a pseudo background) generated by the generator into the discriminator, the authenticity probability of all samples can be output.

[0065] Among them, the specific implementation steps for optimizing the performance of the generator and discriminator through adversarial learning are as follows:

[0066] (1) Initialize the generator and discriminator, and use random weights to initialize the parameters of the generator and discriminator;

[0067] (2) Define an adversarial loss function. Among them, the adversarial loss function combines the generator loss, discriminator loss, and cycle consistency loss;

[0068] (3) Train the adversarial loss function, alternately optimize the generator and discriminator until convergence;

[0069] The implementation method for training the adversarial loss function is as follows:

[0070] Extract training samples from the real dataset as training data;

[0071] Calculate the discriminator loss and update the parameters of the discriminator;

[0072] Optimize the parameters of the generator by combining the generator loss and the cycle consistency loss;

[0073] Alternately optimize the generator and the discriminator until convergence.

[0074] (4) Evaluate the performance of the generator and the discriminator, evaluate the quality of the virtual samples generated by the generator, form a test set with the real samples and the newly generated virtual samples, and test the classification accuracy of the discriminator so that the discriminator can distinguish between real samples and newly generated virtual samples.

[0075] Among them, the goal of the generator is to generate realistic virtual samples so that the discriminator cannot correctly classify them. The specific generator loss function is:

[0076]

[0077] Among them, z is the input of the generator (such as the background area), G(z) is the generated pseudo-nodule sample, and D(G(z)) is the discriminator's evaluation of the authenticity of the generated pseudo-nodule sample.

[0078] The goal of the discriminator is to maximize the ability to distinguish between real samples and generated samples. The specific discriminator loss function is:

[0079]

[0080] Among them, x is the real sample (such as the cube area of the nodule area and its surrounding background area), D(x) is the discriminator's evaluation result of the real sample, and G(z) is the generated sample of the generator.

[0081] To ensure that the generated samples of the generator are consistent with the original real samples in structure and features, a cycle consistency loss function is introduced. The specific cycle consistency loss function is:

[0082]

[0083] Among them, F is the inverse generator, ensuring that x → G(X) → F(G(x)) ≈ x.

[0084] Combining the above generator loss function, discriminator loss function, and cycle consistency loss function, a complete adversarial loss function is formed. The specific one is:

[0085]

[0086] Among them, λ1, λ2, and λ3 are loss weight coefficients.

[0087] Pseudo-nodules are generated in the background region of the cropped cube region image of the nodule center, and the real nodule region is transformed into pseudo-background. When generating pseudo-nodules from the background region, specifically, a Cycle Generative Adversarial Network (CycleGAN) is used to generate fake nodules in the background region. These fake nodules look like nodules in the image but are not actually nodules. The goal of generating fake nodules is to enable the model to recognize and ignore the pseudo-nodules generated in these background regions. The background region refers to the region in the image without real nodules. The goal is to transform the features of these regions into pseudo-nodule regions similar to nodules through CycleGAN, thereby enhancing the robustness of the model.

[0088] When transforming the nodule region into pseudo-background, the features of the nodule region are transformed into a "background" pattern through a Cycle Generative Adversarial Network (CycleGAN), which means "removing" the nodule region in the image so that the model can learn the background features independent of the nodules themselves.

[0089] In this way, the model will learn the contrast features between nodules and the background, rather than relying on the intuitive features of the nodule region. The generated "nodule transformed into background" images will be used to train the model to learn to distinguish nodules from other regions.

[0090] In summary, the advantages of this conversion method are as follows:

[0091] First, enhance the robustness of the model: By generating pseudo-nodules, the model can learn how to recognize and distinguish real nodules from virtual generated nodules. The generation of pseudo-nodules can help the model cope with nodules of different sizes, shapes, and types, enhancing its recognition ability.

[0092] Second, enhance the background recognition ability: By transforming the nodule region into background, the model can learn how to recognize the normal background region of the lung CT image in the absence of nodules. This helps improve the performance of the model in images without nodules, reducing false positives and false negatives.

[0093] Third, improve the data imbalance problem: In many medical image datasets, nodule samples are scarce, especially small nodules or early nodules. By generating pseudo-nodule and background samples, the training set can be effectively expanded, and the sample numbers of nodules and background can be balanced, thereby improving the performance of the model in the case of scarce data.

[0094] Fourth, improve the generalization ability of the model: Through this conversion method, the model can not only recognize nodules but also process the background information around nodules, enhancing the generalization ability of the model in different scenarios and different types of lung nodules.

[0095] Step 300: Detect real images with pulmonary nodules from the training samples and classify the detected pulmonary nodules.

[0096] In step 300, first perform pixel-level segmentation on the nodule regions of the training samples through the encoding-decoding structure of U-Net, and then process the feature maps of the nodule regions output by the decoder by adding an additional classification branch in the decoder structure of U-Net to identify the nodule types in different nodule regions, so as to classify the nodule regions according to the nodule types.

[0097] Among them, the output image of the pixel-level segmentation of the nodule regions of the training samples is a binary classification image, where the nodule regions are labeled as 1 (or other positive classes), and the background regions are 0 (or other negative classes).

[0098] First perform pixel-level segmentation on the nodule regions of the training samples through the encoding-decoding structure of U-Net to achieve pulmonary nodule detection. The specific implementation method for detecting pulmonary nodules is as follows:

[0099] Pulmonary nodule segmentation: Use models such as U-Net to perform the segmentation task, and achieve pixel-level segmentation of the nodule regions in the CT images through the encoding-decoding structure of U-Net. The output of the segmentation is usually a binary classification image, where the nodule regions are labeled as 1 (or other positive classes), and the background regions are 0 (or other negative classes).

[0100] Pulmonary nodule classification: Add an additional classification branch in the decoder structure of U-Net to process the feature maps output by the decoder, and convert the feature maps into classification results (such as benign or malignant) through convolutional layers and fully connected layers.

[0101] Among them, in the decoder part of U-Net, add a network branch for classification through additional convolutional layers or fully connected layers to process the feature maps of the nodule regions output by the decoder. In this way, this embodiment can perform pixel-level detection tasks and region-level classification tasks simultaneously. During the training process, the loss functions of the two tasks are jointly optimized by means of weighted sum.

[0102] For the detection task, use cross-entropy loss and L2 to calculate the detection error.

[0103] For the classification task, use cross-entropy loss to calculate the classification error.

[0104] Use the multi-task loss function to calculate the accuracy of the segmentation results of the nodule regions and calculate the segmentation error of the nodule regions.

[0105] Use the multi-task loss function to calculate the accuracy of the classification results of the nodule regions and calculate the classification error of the nodule regions.

[0106] Adjust the encoding-decoding structure parameters of the U-Net based on the segmentation error and classification error until the segmentation error and classification error meet the expectations.

[0107] The multi-task loss function includes an anchor box classification loss function, a nodule classification loss function, and an anchor box regression loss function; among them, the anchor box classification loss function is the segmentation result loss function for segmenting the nodule region and the background region in the cube region; the nodule classification loss function is the classification result loss function for classifying the features of the nodule region.

[0108] The nodule classification loss function performs binary classification on the cube region centered on the input nodule, and the specific calculation formula is:

[0109]

[0110] where p Z is the true label, and w1 and w0 are class weights used to adjust the imbalance between the nodule region and the background region.

[0111] The anchor box classification loss function L cls performs binary classification on each anchor box (i.e., the detection region) to determine whether it is a nodule region, so as to detect pulmonary nodules in the training samples. The specific calculation formula is:

[0112]

[0113] where p i is the predicted probability of the anchor box i output by the model, indicating the probability that the anchor box contains a nodule. p i * is the true label of the anchor box i, usually 0 or 1, indicating whether it is a nodule region.

[0114] The anchor box regression loss function L reg regresses and predicts the three-dimensional coordinates and diameter of the anchor box (i.e., the detection region), and uses the smooth L1 loss to calculate the difference between the predicted value and the true value of the anchor box (i.e., the detection region).

[0120] In this embodiment, enhanced samples are generated through a CycleGAN (Cycle Generative Adversarial Network). The CycleGAN can access more types of data distributions during the training process. This method not only enhances the nodule detection ability but also enables the model to learn more background features, reduces the occurrence of false positives (misjudging normal regions as nodules), and thus improves the specificity. The diversity of the enhanced samples enables the model to have a stronger discrimination ability for different types of background regions.

[0121] Therefore, this embodiment uses a CycleGAN model to solve the data imbalance problem. The pseudo nodule samples generated by the CycleGAN model can balance the proportion of positive and negative samples, avoid overlearning of the background area by the model, thereby improving the detection sensitivity. Moreover, the pseudo samples generated by the CycleGAN model have diverse features (such as pseudo nodules of different sizes, shapes, and positions), which can expand the distribution of the training data, improve the performance of the model in various scenarios, and enhance the generalization ability of the model. The small-sized pseudo nodule samples generated by the CycleGAN model can guide the model to pay more attention to the detailed features of nodules, improve the detection accuracy of small nodules, and enhance the ability of the model to capture details.

[0122] The above embodiments are only exemplary embodiments of the present application and are not used to limit the present application. The protection scope of the present application is defined by the claims. Those skilled in the art can make various modifications or equivalent replacements within the essence and protection scope of the present application, and such modifications or equivalent replacements should also be regarded as falling within the protection scope of the present application.

Claims

1. A lung nodule detection method based on single-step detection and cyclic generative adversarial network, characterized in that It includes the following steps: Step 100: Crop a cubic region from the lung cavity of the lung CT scan image. The cubic region includes a nodule image formed by a nodule region and its surrounding background region, and an empty background image of a nodule-free region. Mark the bounding box of the nodule region in the nodule image and the coordinates corresponding to the nodule region; Step 200: Apply a cyclic generative adversarial network model to process the cubic region. Randomly simulate and transform the nodule image and the empty background image into enhanced samples. The enhanced samples include simulating and transforming the nodule image into a pseudo-background image without nodules, and simulating and generating a pseudo-nodule image with pseudo-nodules from the empty background image. Use the generated enhanced samples and the cropped cubic region as training samples; Step 300: Detect images with lung nodule regions from the training samples and classify the detected lung nodules.

2. A lung nodule detection method based on single-step detection and cyclic generative adversarial network according to claim 1, characterized in that In step 200, when applying the cyclic generative adversarial network model to process the nodule image in the cubic region, the specific implementation method for generating a pseudo-background image is as follows: The background generator in the cyclic generative adversarial network model locates the nodule region according to the shape and texture features of the nodule, and determines the set of real nodule pixels within the nodule region. Remove the set of real nodule pixels from the nodule image, and use the pixel values of the background region of the nodule image to fill the removed position to generate a pseudo-background image; When applying the cyclic generative adversarial network model to process the empty background image in the cubic region, the specific implementation method for generating a pseudo-nodule image is as follows: Use the lung nodule generator in the cyclic generative adversarial network model to generate a pseudo-nodule at the center position in the empty background image, and make the generated pseudo-nodule image consistent with the real nodule in the nodule image in terms of shape and gray-scale distribution.

3. A lung nodule detection method based on single-step detection and cyclic generative adversarial network according to claim 2, characterized in that In step 200, use the cyclic consistency loss in the cyclic generative adversarial network model to constrain the pseudo-nodule samples to ensure that the newly generated nodules in the pseudo-nodule samples remain at the center position of the image, so that the newly generated nodules in the pseudo-nodule samples are spatially consistent with the original nodule region of the cubic region.

4. A lung nodule detection method based on single-step detection and cyclic generative adversarial network according to claim 2, characterized in that In step 200, input the cubic region image of the nodule center cropped into the cyclic generative adversarial network model, where the background region of the cubic region image is a fragment of a lung CT image without nodules; Use the generator to convert the pixel information in the background region into a region with nodules, that is, simulate pseudo-nodules in the background region to generate a new virtual image; Use a discriminator to classify the new virtual image and determine the authenticity of the new virtual image; Optimize the generator and discriminator through adversarial learning so that the generator can generate pseudo nodules that are visually similar to real nodules; Use a position constraint loss to calculate the deviation between the position of the generated pseudo nodule and the center point of the virtual image, and constrain the position of the newly generated pseudo nodule to be fixed at the center position of the virtual image through a cycle consistency loss.

5. The method for detecting pulmonary nodules based on single-step detection and cyclic generative adversarial network according to claim 2, wherein, In step 200, input the cropped cube region image of the nodule center into the cyclic generative adversarial network model, and generate a pseudo background in the nodule region of the cube region image. The specific implementation method is: Use the generator to eliminate the features of the nodule region of the cube region image and generate a pseudo background region; Use the discriminator to learn to distinguish the real background region of the cube region image and the pseudo background region generated by the generator. Optimize the generator and discriminator through adversarial learning, and constrain the position of the pseudo background region through a cycle consistency loss, so that the pseudo background region generated by the generator is consistent with the real background region.

6. The method for detecting pulmonary nodules based on single-step detection and cyclic generative adversarial network according to claim 4 or 5, wherein, The specific implementation steps for optimizing the performance of the generator and discriminator through adversarial learning are: Initialize the generator and discriminator, and use random weights to initialize the parameters of the generator and discriminator; Define an adversarial loss function, wherein the adversarial loss function combines the generator loss, discriminator loss, and cycle consistency loss; Train the adversarial loss function, alternately optimize the generator and discriminator until convergence; Evaluate the performance of the generator and discriminator, evaluate the quality of the virtual samples generated by the generator, form a test set with real samples and newly generated virtual samples, and test the classification accuracy of the discriminator so that the discriminator can distinguish real samples and newly generated virtual samples.

7. The method for detecting pulmonary nodules based on single-step detection and cyclic generative adversarial network according to claim 6, wherein, The implementation method for training the adversarial loss function is: Extract training samples from the real dataset as training data; Calculate the discriminator loss and update the parameters of the discriminator; Optimize the parameters of the generator by combining the generator loss and cycle consistency loss; Alternately optimize the generator and discriminator until convergence.

8. The method for detecting pulmonary nodules based on single-step detection and cyclic generative adversarial network according to claim 2, wherein, In step 300, first perform pixel-level segmentation on the nodule region of the training sample through the encoding-decoding structure of U-Net, and then process the feature map of the nodule region output by the decoder by adding a classification branch in the decoder structure of U-Net to identify the nodule types of different nodule regions, so as to classify the nodule regions according to the nodule types; Among them, the output image of the pixel-level segmentation of the nodule region of the training sample is a binary classification image, where the nodule region is labeled as 1 and the background region is labeled as 0.

9. A pulmonary nodule detection method based on single-step detection and cyclic generative adversarial network according to claim 8, wherein the multi-task loss function is used to calculate the accuracy of the segmentation result of the nodule region, and the segmentation error of the nodule region is calculated; the multi-task loss function is used to calculate the accuracy of the classification result of the nodule region, and the classification error of the nodule region is calculated; based on the segmentation error and the classification error, the encoding-decoding structure parameters of the U-Net are adjusted until the segmentation error and the classification error meet the expectations.

10. A pulmonary nodule detection method based on single-step detection and cyclic generative adversarial network according to claim 9, wherein the multi-task loss function includes an anchor box classification loss function, a nodule classification loss function, and an anchor box regression loss function; among them, the anchor box classification loss function is the segmentation result loss function for segmenting the nodule region and the background region of the cube region; the nodule classification loss function is the classification result loss function for classifying the features of the nodule region.

Images