Pulmonary nodule identification and segmentation method and system based on deep learning, and storage medium

Abstract

The invention discloses a pulmonary nodule recognition and segmentation method and system based on deep learning and a storage medium, and belongs to the technical field of medical image processing, and the method comprises the steps: obtaining a lung CT image, recognizing a first region and a second region from the lung CT image through a deep learning model, intercepting the first region and the second region as a first target image, and obtaining a second target image; the first area is an area where pulmonary vessels are located, and the second area is an area where pulmonary nodules are located; the deep learning model marks a first label in the first target image, wherein the first label comprises the types of pulmonary vessels and pulmonary nodules; the verification model divides pixel points in the first target image into a first type and a second type, and a second label is acquired based on the distribution condition of the pixel points of the first type and the second type; and if the first tag is the same as the second tag, taking the first tag as an identification result. According to the invention, artificial identification errors are reduced through the deep learning model, and the identification reliability is improved through setting the verification model.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and specifically relates to a method, system and storage medium for lung nodule recognition and segmentation based on deep learning. Background Art

[0002] Lung nodules are one of the early manifestations of lung cancer, and their early detection and treatment are crucial to improving patients' survival rates. However, traditional lung nodule detection methods rely on doctors' experience, which is not only time-consuming and labor-intensive, but also easily affected by subjective factors, leading to missed detection or false detection. In recent years, with the rapid development of deep learning technology, its application in the field of medical image processing has become increasingly widespread, which provides a new solution for the automatic identification and segmentation of lung nodules.

[0003] In the prior art, for example, a Chinese patent document with publication number CN116030325A discloses a lung nodule CT image recognition method based on a deep hybrid learning framework. The method uses a CNN model and a ViT model to learn the local structural features and long-distance contextual relationships of the lung nodule CT image respectively and output the corresponding representation information and the second feature map. Then, a feature aggregation module based on an attention mechanism is designed to aggregate the representation information and the second feature map in a manner of adaptively allocating fusion weights to obtain aggregated features. The aggregated feature map is input into a fully connected layer classifier, and finally, a semantic label of whether the lung nodule tissue is benign or malignant is output. Another example is a Chinese patent document with publication number CN117649408A that discloses a lung nodule recognition and processing method based on lung CT images. The method mixes nodule image data with non-nodule images and inputs them into a neural network classification model for training. After the parameters of the neural network model converge, the lung image data to be analyzed is input into the classification model for lung nodule classification.

[0004] However, even deep learning networks trained with large amounts of training data may produce incorrect classification results. If the incorrect results are shown to doctors, it will interfere with their diagnosis. Summary of the invention

[0005] To solve the above problems, the present invention provides a deep learning-based lung nodule recognition and segmentation method, system and storage medium to solve the problems in the prior art.

[0006] In order to achieve the above-mentioned purpose of the invention, the present invention proposes a lung nodule recognition and segmentation method based on deep learning, comprising:

[0007] Establishing a deep learning model, acquiring an image set of lung CT images of the same patient, extracting a single lung CT image from the image set, and identifying a first region and a second region therefrom using the deep learning model;

[0008] intercepting the first region and the second region from the lung CT image as a first target image, wherein the first region is a region where pulmonary blood vessels are located, and the second region is a region where pulmonary nodules are located;

[0009] The deep learning model annotates a first label in the first target image, wherein the first label includes types of pulmonary blood vessels and pulmonary nodules;

[0010] Establishing a verification model, wherein a plurality of segmentation values ​​are set in the verification model, wherein the verification model selects the segmentation value corresponding to the first label, and divides the pixels in the first target image into a first type and a second type in combination with the pixel values ​​of the pixels in the first target image, so as to obtain a second target image;

[0011] Acquire a second label of the second target image based on the distribution of the first type and the second type of pixel points;

[0012] If the first label is the same as the second label, the first label is used as the recognition result of the second target image.

[0013] Further, obtaining the second label of the second target image includes the following steps:

[0014] Define the area covered by adjacent pixels of the same type in the second target image as a discrete area, define the discrete area composed of the first type of pixels as a first area, and define the discrete area composed of the second type of pixels as a second area;

[0015] Count the first number and the second number of the first area and the second area in the second target image, take the first number and the second number as the distribution, set a first condition for the first label representing pulmonary blood vessels, and set a second condition for the first labels representing different types of pulmonary nodules. If the second target image is under the first label of pulmonary blood vessels, the distribution satisfies the first condition, and the second label of the second target image is determined to be pulmonary blood vessels. If the distribution satisfies the second condition under the first label of pulmonary nodules, the second label is determined to be the type of pulmonary nodules corresponding to the second condition.

[0016] Further, setting the segmentation value comprises the following steps:

[0017] Select standardized images representing pulmonary blood vessels and various types of pulmonary nodules, and mark the corresponding first labels, set initial values, convert the standardized images with different first labels into the second target images based on the initial values, count the first quantity and the second quantity in the second target image, and use a preset formula to calculate the difference value Yk between the kth second target image and the standardized image, and the preset formula is: Ak and Bk are respectively the first number and the second number corresponding to the k-th standardized image, N is the number of images with different first labels from the k-th standardized image, Cn and Dn are the first number and the second number corresponding to the n-th standardized image among images with different first labels from the k-th standardized image, and the median of the difference values ​​of each standardized image under the same type of the first label is used as a separation value of the first label of one type;

[0018] After increasing the initial value by a predetermined value, the standardized image is converted into the second target image again, and the separation value of each type of the first label is calculated, a coordinate system is established, and a curve of the separation value of each type of the first label changing with the initial value is plotted in the coordinate system, and the maximum value of the separation value in each type of the change curve is located, and the corresponding initial value is defined as the segmentation value of the first label.

[0019] Furthermore, both the first condition and the second condition include a first range corresponding to the first quantity and a second range corresponding to the second quantity. If the first quantity of the second target image is within the first range of the first condition or the second condition, and the second quantity is within the second range under the same condition, it is determined that the distribution satisfies the first condition or the second condition.

[0020] Further, determining the first range and the second range of the first condition or the second condition comprises the following steps:

[0021] Obtain the first median of all the first quantities under the segmentation value corresponding to the first label, and the second median of the second quantity, generate a first numerical range based on the first median, generate a second numerical range based on the second median, and use the first numerical range and the second numerical range as the first range and the second range corresponding to the first condition or the second condition.

[0022] Furthermore, if the first label and the second label of the second target image are different, the second target image is defined as the first image, the appearance position of the first image in the lung CT image is located, and other lung CT images are re-extracted from the image set and defined as auxiliary images. After the auxiliary images are recognized using the deep learning model and the verification model, if the second target image is generated and corresponds to the appearance position, the second target image is defined as the second image, and the number of identical labels in the first label and the second label of the first image and the second image is counted, and the label with the largest number is determined as the recognition result.

[0023] Furthermore, establishing the deep learning model includes the following steps:

[0024] A training set is constructed, the lung CT images in the training set are preprocessed, and learning labels are annotated therein, wherein the learning labels include the types of pulmonary blood vessels and pulmonary nodules, a basic recognition model based on an FCN network is constructed, wherein the basic recognition model includes a full convolution part and a deconvolution part, and the basic recognition model is trained using the training set to obtain the deep learning model.

[0025] Furthermore, the preprocessing includes denoising, contrast enhancement, and standardized cropping.

[0026] The present invention also provides a deep learning-based lung nodule recognition and segmentation system, which is used to implement the above-mentioned deep learning-based lung nodule recognition and segmentation method, and the system includes:

[0027] A first recognition module, including a deep learning model, acquires an image set of lung CT images of the same patient, extracts a single lung CT image from the image set, and uses the deep learning model to identify a first region and a second region therefrom;

[0028] A segmentation module, which extracts the first region and the second region from the lung CT image as a first target image, wherein the first region is a region where pulmonary blood vessels are located, and the second region is a region where pulmonary nodules are located;

[0029] A second recognition module, wherein a verification model is provided, wherein a plurality of segmentation values ​​are set in the verification model, wherein the verification model selects the segmentation value corresponding to the first label, and divides the pixels in the first target image into a first type and a second type in combination with the pixel values ​​of the pixels in the first target image, thereby obtaining a second target image, and obtaining a second label of the second target image based on the distribution of the pixels of the first type and the second type;

[0030] The output module uses the first label as the recognition result of the second target image if the first label is the same as the second label.

[0031] The present invention also discloses a computer storage medium, wherein the computer storage medium stores program instructions, wherein when the program instructions are executed, the device where the computer storage medium is located is controlled to execute the above-mentioned method.

[0032] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0033] The present invention reduces data redundancy and processing time by aggregating lung CT images of the same patient as an image set and extracting a single image therefrom for processing; the deep learning model automatically segments lung nodules and pulmonary vascular areas from the CT image, extracts them as target images, and annotates them with a first label. Afterwards, the model is verified and multiple segmentation values ​​are set, and the pixel points in the target image are classified into types based on the pixel values ​​of the pixel points. After the division, a second label is generated by analyzing the distribution of different types of pixels. If the second label is the same as the first label, it indicates that the recognition result is correct, thereby improving the reliability of the recognition process and the accuracy of the diagnosis. The automated recognition and segmentation process reduces the time doctors spend on image analysis and improves work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flowchart of the steps of the lung nodule recognition and segmentation method based on deep learning of the present invention;

[0035] Figure 2 A schematic diagram of a first region and a second region in a lung CT image of the present invention;

[0036] Figure 3 This is a schematic diagram of the structure of the lung nodule recognition and segmentation system based on deep learning of the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0038] It is understood that the terms "first", "second", etc. used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are only used to distinguish a first element from another element. For example, without departing from the scope of this application, a first xx script may be referred to as a second xx script, and similarly, a second xx script may be referred to as a first xx script.

[0039] like Figure 1 As shown, a lung nodule recognition and segmentation method based on deep learning includes:

[0040] S1: Establish a deep learning model, obtain an image set of lung CT images of the same patient, extract a single lung CT image from the image set, and use the deep learning model to identify the first region and the second region therefrom.

[0041] The deep learning model is established in the following steps:

[0042] A training set was constructed, and the lung CT images in the training set were preprocessed and labeled with learning labels, including the types of pulmonary blood vessels and pulmonary nodules. A basic recognition model based on the FCN network was constructed, and the basic recognition model included a full convolution part and a deconvolution part. The basic recognition model was trained using the training set to obtain a deep learning model.

[0043] Preprocessing includes denoising, contrast enhancement, and standardized cropping.

[0044] The deep learning model of this embodiment is constructed based on the FCN network. The FCN network includes a full convolution part and a deconvolution part. The full convolution part includes six convolution layers, each of which includes a convolution kernel 1*1, a convolution kernel 3*3, a pooling layer and an excitation layer RELU. The deconvolution part includes four deconvolution layers, each of which includes a deconvolution kernel and an excitation layer RELU. The size of the deconvolution kernel is 8*8, and the number of deconvolution kernels is 3. Each pixel is classified through the FCN network to achieve regional division of the lung CT image. Before the specific division, the lung CT images taken of the same patient are first summarized as an image set, and then a lung CT image is extracted from it and semantic segmentation is performed using a deep learning model. Finally, the first region representing the pulmonary vascular part is obtained, such as Figure 2 Region Q1 in represents the second region of the pulmonary nodule. Figure 2 Region Q2 in the.

[0045] S2: A first region and a second region are captured from the lung CT image as a first target image, wherein the first region is a region where the pulmonary blood vessels are located, and the second region is a region where the pulmonary nodules are located.

[0046] S3: The deep learning model annotates a first label in the first target image, where the first label includes types of pulmonary blood vessels and pulmonary nodules.

[0047] Continue to refer to Figure 2, region Q1 is intercepted from the lung CT image as the first target image including pulmonary blood vessels, and region Q2 is intercepted as the first target image including pulmonary nodules. The types of pulmonary nodules in this embodiment include ground glass nodules, solid nodules, and mixed ground glass nodules. After the interception is completed, a first label is annotated for each first target image. For example, the first target image generated by intercepting region Q1 is annotated with the first label of pulmonary blood vessels, which means that the first target image includes pulmonary blood vessels; the first target image generated by intercepting region Q2 is annotated with mixed ground glass nodules, which means that the first target image includes pulmonary nodules of the type of mixed ground glass nodules.

[0048] S4: Establish a verification model, in which a plurality of segmentation values ​​are set. The verification model selects a corresponding segmentation value according to the first label, and divides the pixels in the first target image into a first type and a second type in combination with the pixel values ​​of the pixels in the first target image to obtain a second target image.

[0049] S5: Acquire a second label of the second target image based on the distribution of the first type and the second type of pixel points.

[0050] S6: If the first label and the second label are the same, use the first label as the recognition result of the second target image.

[0051] In this embodiment, the verification model includes 4 segmentation values, where segmentation value 1 corresponds to pulmonary blood vessels, and segmentation values ​​2-4 correspond to ground glass nodules, solid nodules, and mixed ground glass nodules. After receiving the first target image whose first label is pulmonary blood vessels, the verification model selects segmentation value 1, divides the pixels in the first target image whose pixels are less than segmentation value 1 into the first type, and divides the pixels whose pixels are greater than or equal to segmentation value 1 into the second type, and defines the divided first target image as the second target image. The verification model then generates a second label for the second target image by analyzing the distribution of the first type and second type pixels in the second target image. The second label also includes pulmonary blood vessels, ground glass nodules, solid nodules, and mixed ground glass nodules.

[0052] If the second label obtained by the verification model is the same as the first label marked by the deep learning model, it means that the recognition results of the two deep learning models and the verification model are consistent, and the content in the first label is output as the final recognition result. The processing of the first label and the second label in different situations will be introduced later. The results of the deep learning model are verified by the verification model, thereby ensuring the accuracy of the recognition results.

[0053] The present invention reduces data redundancy and processing time by aggregating lung CT images of the same patient as an image set and extracting a single image from it for processing; the deep learning model automatically segments lung nodules and pulmonary vascular areas from the CT image, extracts them as target images, and annotates them with a first label. After that, the model is verified and multiple segmentation values ​​are set, and the pixel points in the target image are classified into types based on the pixel values ​​of the pixels. After the division, the second label is generated by analyzing the distribution of different types of pixels. If the second label is the same as the first label, it indicates that the recognition result is correct, thereby improving the reliability of the recognition process. Improved diagnostic accuracy. The automated recognition and segmentation process reduces the time doctors spend on image analysis and improves work efficiency.

[0054] It is particularly noteworthy that the present invention reduces human recognition errors through the deep learning model, and improves the reliability of recognition by setting a verification model, thereby reducing the possibility of outputting erroneous results.

[0055] In this embodiment, obtaining the second label of the second target image includes the following steps:

[0056] The area covered by adjacent pixels of the same type in the second target image is defined as a discrete area, and the discrete area composed of pixels of the first type is defined as a first area, and the discrete area composed of pixels of the second type is defined as a second area;

[0057] The first quantity and the second quantity of the first area and the second area in the second target image are counted, and the first quantity and the second quantity are used as the distribution. A first condition is set for the first label representing the pulmonary blood vessels, and a second condition is set for the first labels representing different types of pulmonary nodules. If the second target image is under the first label of the pulmonary blood vessels, the distribution satisfies the first condition, and the second label of the second target image is determined to be the pulmonary blood vessels. If the distribution satisfies the second condition under the first label of the pulmonary nodules, the second label is determined to be the type of pulmonary nodules corresponding to the second condition.

[0058] For the first target image representing the pulmonary vessels, such as Figure 2 As shown in region Q1 in the figure, after using a reasonable segmentation value 1 to divide the pixel type inside it, the number of first regions and second regions will show a specific regularity. For example, after segmentation, there are 3 discrete regions of the first type and 3 discrete regions of the second type. If the first number and the second number meet the preset first condition, it is considered that the second label of the second target image is pulmonary blood vessels.

[0059] The solid nodules in the CT images are basically composed of white pixels, and their shape area is close to a circle. After using a reasonable segmentation value of 2 to divide the pixel types inside, the number of the first area and the second area shows a specific regularity. For example, there is 1 discrete area of ​​the first type and 4 discrete areas of the second type. If the first number and the second number meet the preset second condition, the second label of the second target image is considered to be a pulmonary nodule and a solid nodule.

[0060] For ground glass nodules, which include densely distributed gray pixels, since the pixel values ​​of the gray pixels fluctuate greatly, after the area is divided using the segmentation value 3, there may be a large number of discrete areas of the first type and a large number of discrete areas of the second type. For example, there are 7 discrete areas of the first type and 9 discrete areas of the second type. If the first number and the second number meet the preset second condition, it is considered that the second label of the second target image is a lung nodule and a ground glass nodule.

[0061] For mixed ground glass nodules, such as Figure 2 As shown in area Q2 in the figure, which includes scattered gray pixels and white pixels, after the area is divided using the segmentation value 4, there will be a large number of discrete areas of the first type and a large number of discrete areas of the second type. For example, there are 12 discrete areas of the first type and 15 discrete areas of the second type. If the first number and the second number meet the preset second condition, it is considered that the second label of the second target image is a lung nodule and a mixed ground glass nodule.

[0062] It can be seen that after using a reasonable segmentation value to divide the pixel types inside the first target image, the discrete regions under different first labels have different distribution patterns, and it is possible to determine whether the first target image is a pulmonary vessel or a pulmonary nodule based on the distribution pattern, and determine the type of pulmonary nodule, thereby verifying the recognition result of the deep model. The first condition and the second condition will be introduced later.

[0063] In this embodiment, setting the split value includes the following steps:

[0064] Select standardized images representing pulmonary blood vessels and various types of pulmonary nodules, and mark the corresponding first labels, set initial values, convert standardized images with different first labels into second target images based on the initial values, count the first quantity and the second quantity in the second target image, and use a preset formula to calculate the difference value Yk between the kth second target image and the standardized image. The preset formula is: Ak and Bk are the first number and the second number corresponding to the k-th standardized image, respectively; N is the number of images with different first labels from the k-th standardized image; Cn and Dn are the first number and the second number corresponding to the n-th standardized image among images with different first labels from the k-th standardized image; the median of the difference values ​​of each standardized image under the same type of first label is used as the separation value of a type of first label;

[0065] After increasing the initial value by a predetermined value, the standardized image is converted into a second target image again, and the separation value of each type of the first label is calculated, a coordinate system is established, and a curve of the separation value of each type of the first label changing with the initial value is plotted in the coordinate system, and the maximum value of the separation value in each type of change curve is located, and the corresponding initial value is defined as the segmentation value of the first label.

[0066] The standardized images are clear, complete and representative images, and the standardized images are screened out by manual selection. After the screening is completed, the corresponding first label is annotated for each standardized image. In particular, there are multiple standardized images of each type. Then, an initial value is set for the segmentation value, and the pixel type in the standardized image is divided using the initial value to obtain the second target image corresponding to each standardized image. After counting the first number and the second number in each second target image, the difference value between each standardized image and other standardized images is calculated using a preset formula. In the preset formula, the greater the difference between the first number or the second number of the standardized image to be calculated and other standardized images of different types, the greater the calculated difference value. Finally, the median of the difference values ​​of each standardized image under the same type of first label is used as the separation value of the first label of a type. For example, the median of the corresponding difference values ​​of all standardized images with the first label of pulmonary vascular is used as the separation value of the pulmonary vascular under the initial value. For pulmonary vessels, the separation value represents the difference between the first number and the second number in the second target image containing pulmonary vessels and the first number and the second number in the second target image of other types of pulmonary nodules after segmentation using the initial value, and the larger the separation value, the greater the degree of differentiation.

[0067] Then, by continuously increasing the initial value, the change of each type of first label separator is obtained, and finally the initial value corresponding to the maximum separation value is selected as the segmentation value. The effect to be achieved by this step is that, for example, for pulmonary blood vessels, after using the corresponding segmentation value for division, if the second target image is indeed a pulmonary blood vessel, the first number and the second number should be within the corresponding numerical range, and if it is not a pulmonary blood vessel, the first number and the second number will not be within the corresponding numerical range, thus achieving the distinction of the second target image.

[0068] In this embodiment, the first condition and the second condition both include a first range corresponding to the first quantity and a second range corresponding to the second quantity. If the first quantity of the second target image is within the first range of the first condition or the second condition, and the second quantity is within the second range under the same condition, it is determined that the distribution satisfies the first condition or the second condition.

[0069] For example, for lung nodules, the corresponding second condition has a first range of 1-3 and a second range of 3-5. When the first number in the second target image is 2 and the second number is 4, it is considered that the second condition is met.

[0070] In this embodiment, determining the first range and the second range of the first condition or the second condition includes the following steps:

[0071] Get the first median of all first quantities and the second median of the second quantities under the first label corresponding to the split value, generate a first numerical range based on the first median, generate a second numerical range based on the second median, and use the first numerical range and the second numerical range as the first range and the second range corresponding to the first condition or the second condition.

[0072] For example, to obtain the first range and the second range of the pulmonary vessels corresponding to the first condition, the median of all first quantities (i.e., the first median) is obtained in the second target image converted from the pulmonary vascular standard image under the corresponding segmentation value, for example, 2, and the median of all second quantities (i.e., the second median) is obtained, for example, 4, and then the first median is added or subtracted by 1 to generate the first numerical range, specifically 1-3, and the second median is added or subtracted by 1 to generate the second numerical range, specifically 3-5, and 1-3 and 3-5 are used as the first range and the second range corresponding to the first condition. The generation process of various other second conditions is similar.

[0073] In this embodiment, if the first label and the second label of the second target image are different, the second target image is defined as the first image, the appearance position of the first image in the lung CT image is located, and other lung CT images are re-extracted from the image set and defined as auxiliary images. After the auxiliary images are recognized using the deep learning model and the verification model, if the second target image is generated and corresponds to the appearance position, the second target image is defined as the second image, and the number of identical labels in the first label and the second label of the first image and the second image is counted, and the label with the largest number is determined as the recognition result.

[0074] If the first label of the first target image is pulmonary blood vessels, and the second label of the second target image is solid nodules, the second target image is defined as the first image, and the interception position of the second target image in the original lung CT image is located, that is, the existence position, and then another lung CT image of the same patient is extracted from the image set and defined as an auxiliary image. The deep learning model is used to obtain the first label of the auxiliary image, such as pulmonary blood vessels, and the verification model is used to obtain the second label of the auxiliary image, such as pulmonary blood vessels. Since three of the four labels are pulmonary blood vessels, which occupy the majority of the number, the pulmonary blood vessels are used as the recognition result.

[0075] The present invention also provides a deep learning-based lung nodule recognition and segmentation system, which is used to implement the above-mentioned deep learning-based lung nodule recognition and segmentation method, and the system includes:

[0076] A first recognition module, including a deep learning model, obtains an image set of lung CT images of the same patient, extracts a single lung CT image from the image set, and uses the deep learning model to identify a first region and a second region therefrom;

[0077] A segmentation module, which extracts a first region and a second region from the lung CT image as a first target image, wherein the first region is a region where the pulmonary blood vessels are located, and the second region is a region where the pulmonary nodules are located;

[0078] A second recognition module includes a verification model, wherein a plurality of segmentation values ​​are set in the verification model, the verification model selects a corresponding segmentation value according to the first label, and divides the pixels in the first target image into a first type and a second type in combination with the pixel values ​​of the pixels in the first target image, obtains a second target image, and obtains a second label of the second target image based on the distribution of the first type and the second type of pixels;

[0079] The output module uses the first label as the recognition result of the second target image if the first label and the second label are the same.

[0080] The present invention also discloses a computer storage medium, wherein the computer storage medium stores program instructions, wherein when the program instructions are executed, the device where the computer storage medium is located is controlled to execute the above method.

[0081] It should be understood that the various technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the various technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0082] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A lung nodule recognition and segmentation method based on deep learning, characterized in that: include: Establishing a deep learning model, acquiring an image set of lung CT images of the same patient, extracting a single lung CT image from the image set, and identifying a first region and a second region therefrom using the deep learning model; intercepting the first region and the second region from the lung CT image as a first target image, wherein the first region is a region where pulmonary blood vessels are located, and the second region is a region where pulmonary nodules are located; The deep learning model annotates a first label in the first target image, wherein the first label includes types of pulmonary blood vessels and pulmonary nodules; Establishing a verification model, wherein a plurality of segmentation values ​​are set in the verification model, wherein the verification model selects the segmentation value corresponding to the first label, and divides the pixels in the first target image into a first type and a second type in combination with the pixel values ​​of the pixels in the first target image, so as to obtain a second target image; Acquire a second label of the second target image based on the distribution of the first type and the second type of pixel points; If the first label is the same as the second label, the first label is used as the recognition result of the second target image.

2. The method according to claim 1, characterized in that Acquiring the second label of the second target image comprises the following steps: Define the area covered by adjacent pixels of the same type in the second target image as a discrete area, define the discrete area composed of the first type of pixels as a first area, and define the discrete area composed of the second type of pixels as a second area; Count the first number and the second number of the first area and the second area in the second target image, take the first number and the second number as the distribution, set a first condition for the first label representing pulmonary blood vessels, and set a second condition for the first labels representing different types of pulmonary nodules. If the second target image is under the first label of pulmonary blood vessels, the distribution satisfies the first condition, and the second label of the second target image is determined to be pulmonary blood vessels. If the distribution satisfies the second condition under the first label of pulmonary nodules, the second label is determined to be the type of pulmonary nodules corresponding to the second condition.

3. The method according to claim 2, characterized in that Setting the split value includes the following steps: Select standardized images representing pulmonary blood vessels and various types of pulmonary nodules, and mark the corresponding first labels, set initial values, convert the standardized images with different first labels into the second target images based on the initial values, count the first quantity and the second quantity in the second target image, and use a preset formula to calculate the difference value Yk between the kth second target image and the standardized image, and the preset formula is: Ak and Bk are respectively the first number and the second number corresponding to the k-th standardized image, N is the number of images with different first labels from the k-th standardized image, Cn and Dn are the first number and the second number corresponding to the n-th standardized image among images with different first labels from the k-th standardized image, and the median of the difference values ​​of each standardized image under the same type of the first label is used as a separation value of the first label of one type; After increasing the initial value by a predetermined value, the standardized image is converted into the second target image again, and the separation value of each type of the first label is calculated, a coordinate system is established, and a curve of the separation value of each type of the first label changing with the initial value is plotted in the coordinate system, and the maximum value of the separation value in each type of the change curve is located, and the corresponding initial value is defined as the segmentation value of the first label.

4. The method according to claim 3, characterized in that Both the first condition and the second condition include a first range corresponding to the first quantity and a second range corresponding to the second quantity. If the first quantity of the second target image is within the first range of the first condition or the second condition, and the second quantity is within the second range under the same condition, it is determined that the distribution satisfies the first condition or the second condition.

5. The method according to claim 4, characterized in that Determining the first range and the second range of the first condition or the second condition comprises the following steps: Obtain the first median of all the first quantities under the segmentation value corresponding to the first label, and the second median of the second quantity, generate a first numerical range based on the first median, generate a second numerical range based on the second median, and use the first numerical range and the second numerical range as the first range and the second range corresponding to the first condition or the second condition.

6. The method according to claim 1, characterized in that If the first label and the second label of the second target image are different, define the second target image as the first image, locate the appearance position of the first image in the lung CT image, re-extract other lung CT images from the image set and define them as auxiliary images, and after using the deep learning model and the verification model to recognize the auxiliary images, if the second target image is generated and corresponds to the appearance position, define the second target image as the second image, count the number of identical labels in the first label and the second label of the first image and the second image, and determine the label with the largest number as the recognition result.

7. The method according to claim 1, characterized in that Establishing the deep learning model includes the following steps: A training set is constructed, the lung CT images in the training set are preprocessed, and learning labels are annotated therein, wherein the learning labels include the types of pulmonary blood vessels and pulmonary nodules, a basic recognition model based on an FCN network is constructed, wherein the basic recognition model includes a full convolution part and a deconvolution part, and the basic recognition model is trained using the training set to obtain the deep learning model.

8. The method according to claim 7, characterized in that The preprocessing includes denoising, contrast enhancement, and standardized cropping.

9. A deep learning-based lung nodule recognition and segmentation system, used to implement the method according to any one of claims 1 to 8, characterized in that: include: A first recognition module, including a deep learning model, acquires an image set of lung CT images of the same patient, extracts a single lung CT image from the image set, and uses the deep learning model to identify a first region and a second region therefrom; A segmentation module, which extracts the first region and the second region from the lung CT image as a first target image, wherein the first region is a region where pulmonary blood vessels are located, and the second region is a region where pulmonary nodules are located; A second recognition module, wherein a verification model is provided, wherein a plurality of segmentation values ​​are set in the verification model, wherein the verification model selects the segmentation value corresponding to the first label, and divides the pixels in the first target image into a first type and a second type in combination with the pixel values ​​of the pixels in the first target image, thereby obtaining a second target image, and obtaining a second label of the second target image based on the distribution of the pixels of the first type and the second type; The output module uses the first label as the recognition result of the second target image if the first label is the same as the second label.

10. A computer storage medium, characterized in that: The computer storage medium stores program instructions, wherein when the program instructions are executed, the device where the computer storage medium is located is controlled to execute the method according to any one of claims 1 to 8.

Images