Training system for artificial intelligence-based lung nodule detection model

Abstract

The present application relates to the technical field of image feature recognition, in particular to a training system of a lung nodule detection model based on artificial intelligence. The system compares and analyzes the CT density difference after aligning two images detected at different times, extracts the suspected lesion area, and screens out the accurate lung nodule area from the alveolar feature, vascular feature and lesion feature. According to the morphological change of the lung nodule area in two examinations, the accurate nodule change state can be judged, and rich and effective information is marked in the contrast image. The neural network can be trained based on the marked image, so that the neural network can accurately identify the nodule area and its growth state in the lung CT image by comparing the lung CT images of different examinations.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image feature recognition, and particularly relates to a training system of a lung nodule detection model based on artificial intelligence. BACKGROUND

[0002] Lung nodules have significant features in lung CT images, and the training of a lung nodule recognition neural network can be realized through lung CT images with lung nodule regions. The input of a conventional neural network is a lung CT image of a target patient, and the output is a labeled nodule region. In order to ensure accuracy, the neural network needs to be trained with a large amount of data, and the training data is a lung CT image labeled with a nodule region. The labeling of training images requires a large amount of manpower, and the process of labeling by personal experience may result in mislabeling or incorrect labeling, affecting the recognition accuracy of the neural network, and thus the neural network cannot be used to detect and prompt the nodule, and cannot effectively assist the doctor in judgment. SUMMARY

[0003] In order to solve the technical problem that the existing technology cannot effectively detect and prompt the nodule region due to inaccurate labeling of the nodule region in the lung CT image, the purpose of the present application is to provide a training system of a lung nodule detection model based on artificial intelligence, and the technical solution adopted is as follows:

[0004] The present application provides a training system of a lung nodule detection model based on artificial intelligence, which comprises:

[0005] A lung CT detection image acquisition module is configured to obtain an initial image of a patient's lung under the previous CT detection before review and a contrast image under review detection;

[0006] A region analysis module is configured to, after aligning the initial image and the contrast image, screen out a suspected lesion region in the contrast image and an initial region in the initial image matched with the suspected lesion region according to the CT density difference between the same positions of the lung regions on the two images;

[0007] A nodule recognition module is configured to obtain the alveolar features of the suspected lesion region according to the area change and the CT density change between the suspected lesion region and the initial region; obtain the vascular features of the suspected lesion region according to the blood vessel skeleton and the edge definition in the suspected lesion region; obtain the lesion features of the suspected lesion region according to the uneven distribution of the CT density of the suspected lesion region; and screen out the lung nodule region according to the alveolar features, the vascular features and the lesion features;

[0008] The information labeling module is configured to determine the nodule change state according to the CT density difference and the area difference between the lung nodule region and the matched initial region, and the edge burr degree of the lung nodule region, and label the contrast image and be used for model training.

[0009] Further, the method for aligning the initial image and the contrast image comprises:

[0010] The spine contour in the initial image and the contrast image is obtained by using an edge detection algorithm; a plurality of non-collinear inflection points on the spine contour are selected as matching feature points; and the matching feature points between the initial image and the contrast image are matched and aligned based on a sequential constraint matching algorithm.

[0011] Further, the method for screening the suspected lesion region in the contrast image comprises:

[0012] The lung region in the contrast image and the initial image is extracted by using the CT density value of the pixel point; the lung regions in the contrast image and the initial image are compared to obtain the CT density difference between the pixel points at the same position, and the CT density difference is subjected to statistical significance test to obtain the significance value of each pixel point in the lung curve in the contrast image; and the region formed by the continuous pixel points satisfying the preset numerical condition of the significance value is taken as the suspected lesion region in the contrast image.

[0013] Further, the matching method of the initial region comprises:

[0014] In the initial image, the pixel points are clustered based on the CT density of each pixel point to obtain a plurality of clustering regions; for each suspected lesion region in the contrast image, the suspected lesion region is mapped to the initial image to obtain a mapping region, and the clustering region with the largest overlap of the mapping region is selected as the initial region matched with the suspected lesion region.

[0015] Further, the method for obtaining the alveolar feature comprises:

[0016] The area difference and the average CT density value difference between the suspected lesion region and the matched initial region are obtained; and the ratio of the area difference to the average CT density value difference is taken as the alveolar feature.

[0017] Further, the method for obtaining the blood vessel feature comprises:

[0018] The suspected lesion region is expanded according to a preset size to obtain an expanded region; and the ratio of the total length of the blood vessel skeleton in the expanded region to the area of the expanded region is taken as the blood vessel density index.

[0019] The gradient average value of each pixel point on the boundary of the suspected lesion region is obtained, and the gradient average value is subjected to normalization processing to obtain the edge definition.

[0020] Fusing the blood vessel density index and the edge definition positively to obtain the blood vessel feature.

[0021] Further, the method for obtaining the lesion feature comprises:

[0022] Taking the CT density standard deviation in the suspected lesion region as the lesion feature.

[0023] Further, the method for screening the lung nodule region according to the alveolus feature, the blood vessel feature and the lesion feature comprises:

[0024] According to the alveolus feature, the alveolus region in the suspected lesion region is removed, according to the blood vessel feature, the blood vessel region in the suspected lesion region is removed, and in the remaining suspected lesion region, the suspected lesion region with the lesion feature meeting the preset lesion condition is selected as the lung nodule region.

[0025] Further, the method for judging the nodule change state according to the CT density difference and the area difference between the lung nodule region and the matched initial region, and the edge burr degree of the lung nodule region comprises:

[0026] The following judgment processes are sequentially executed:

[0027] The CT density value is used to judge whether the initial region matched with the lung nodule region is a lung parenchyma region, if the initial region is a lung parenchyma region, and the average CT density of the lung nodule region is greater than the average CT density of the initial region, the lung nodule region is identified as a new nodule region;

[0028] If the edge burr degree of the lung nodule region is greater than a preset burr degree threshold, the lung nodule region is identified as a malignant lesion nodule region;

[0029] If the area of the lung nodule region relative to the matched initial region becomes smaller, and the CT density increases by at least a preset increase threshold, the lung nodule region is judged as a benign growth nodule region;

[0030] If the area of the lung nodule region relative to the matched initial region becomes larger, the lung nodule region is judged as a nodule enlargement region.

[0031] Further, the method for obtaining the edge burr degree comprises:

[0032] The lung nodule region is subjected to an opening operation to obtain a burr-free region, the lung nodule region is subtracted from the burr-free region, and the remaining region is a burr region; the area ratio of the burr region to the lung nodule region is taken as the edge burr degree.

[0033] The present application has the following beneficial effects:

[0034] The present application considers that the nodule will produce growth characteristics with time change, so the lung nodule area in the contrast image under the review detection can be determined by comparing the images detected at different times, and the change state of the lung nodule area can be identified, and then the automatic annotation of information-rich is realized. In the comparison analysis process, after aligning the two images, the CT density difference is used for comparison analysis, the suspected lesion area is extracted, and further considering the various tissue forms of the lung, the accurate lung nodule area is screened out from the alveolar characteristics, vascular characteristics and lesion characteristics. Further, according to the morphological change of the lung nodule area in two examinations, the accurate nodule change state can be judged, and then the rich and effective information is annotated in the contrast image. Based on the annotated image, the neural network can be trained, so that the neural network can identify the nodule area and its growth state in the lung CT image by comparing the lung CT images of different examinations, and assist the doctor to make effective judgment on the patient. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art and the advantages thereof, a brief introduction will be given to the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0036] Figure 1 A training system block diagram of a lung nodule detection model based on artificial intelligence provided by an embodiment of the present application;

[0037] Figure 2 A nodule comparison schematic diagram in different lung examination CT images provided by an embodiment of the present application;

[0038] Figure 3 A nodule comparison schematic diagram of the first type of growth state provided by an embodiment of the present application;

[0039] Figure 4 A nodule comparison schematic diagram of the second type of growth state provided by an embodiment of the present application;

[0040] Figure 5 A nodule comparison schematic diagram of the third type of growth state provided by an embodiment of the present application;

[0041] Figure 6 A nodule comparison schematic diagram of the fourth type of growth state provided by an embodiment of the present application;

[0042] Figure 7 An initial image and a comparison image alignment schematic diagram provided by an embodiment of the present application;

[0043] Figure 8 A lung region diagram provided by one embodiment of the present application. DETAILED DESCRIPTION

[0044] In order to further clarify the technical means and effects taken by the present application to achieve the predetermined inventive purpose, the following describes in detail the specific implementation, structure, features and effects of a training system of an artificial intelligence-based lung nodule detection model according to the present application in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0046] The specific scheme of the training system of the artificial intelligence-based lung nodule detection model provided by the present application is described in detail below in combination with the accompanying drawings.

[0047] Please refer to Figure 1 which shows a block diagram of the training system of the artificial intelligence-based lung nodule detection model provided by one embodiment of the present application. The system includes a lung CT detection image acquisition module 101, a region analysis module 102, a nodule recognition module 103, and an information labeling module 104.

[0048] The embodiment of the present application aims to determine the lung nodule region in the image under review detection and the growth characteristics of the region by comparing the detection images of different lung CT detection periods. Therefore, first, the lung CT detection image acquisition module 101 is used to obtain the initial image of the patient's lung under the previous CT detection before review and the comparison image under review detection. It should be noted that each pixel position in the CT image corresponds to a CT density value. For the lung region, the CT density of a position represents the air ratio at that position.

[0049] Please refer to Figure 2 which shows a nodule comparison diagram in different lung examination CT images provided by one embodiment of the present application. Wherein Figure 2 the left image of is a local part of the initial image under the previous CT detection before review, and the right image is a local part of the comparison image under review detection. By comparing the two images, it can be seen that the region in the marked circle belongs to the lung nodule region, and the nodule has grown to produce a relatively significant region. The comparison image and the initial image have obvious CT density difference at the same position.

[0050] It should be noted that, because the embodiment of the present application aims to obtain a large number of images with annotations for model training of a neural network, a large number of images need to be automatically annotated. A group of initial images and contrast images of the same patient can be obtained as a group of training data by calling images stored in a hospital internal database. According to the relevant regulations for diagnosis and treatment of pulmonary nodules, the CT review interval of pulmonary nodules is 3-6 months, so each group of training data can be determined according to the time interval.

[0051] As shown in Figure 2 The pulmonary nodules between images of different periods will have significant CT density changes due to growth. However, because there is a long time interval between the two images, and the respiratory motion and body position difference during CT scanning will cause spatial dislocation of anatomical structures, the lesion may be offset in position due to respiration, and the projection of structures such as ribs will be deformed due to body position changes. These differences will directly lead to inaccurate measurement of the area of the nodule in subsequent analysis. The dynamic evolution of the nodule needs to be compared in a unified spatial coordinate system, otherwise subtle changes may be masked by respiratory artifacts or body position deviation, affecting the accuracy of nodule region identification. Therefore, the region analysis module 102 first aligns the initial image and the contrast image. After alignment, the suspected lesion region in the contrast image can be screened by comparing the CT density difference between the same positions in the lung region. That is, the suspected lesion region is the region that has obvious changes compared with the last time in this review. Such a change region may be due to alveolar motion or vascular motion, so it cannot be directly judged as a pulmonary nodule region and needs to be further compared and analyzed. Therefore, the initial region matching the suspected lesion region in the initial image needs to be obtained. Because the two images have been aligned, the initial region can be obtained by matching the region corresponding to the position of the suspected lesion region in the initial image.

[0052] The nodule recognition module 103 further analyzes the contrast features between the suspected lesion region and the matched initial region, and the morphological features of the suspected lesion region itself, so as to screen out the accurate lung nodule region in the suspected lesion region. In the suspected lesion region, the main influence of false recognition is the alveolar region and the blood vessel region. The alveolus is the basic unit of the lung for gas exchange. When the alveolus expands due to respiration, the area increases, and the internal air proportion increases significantly. When the lung nodule increases, the internal soft tissue component increases, and the air proportion changes little. The air proportion change can be reflected by the CT density change, so the alveolar features of the suspected lesion region can be obtained according to the area change and the CT density change between the suspected lesion region and the initial region. The blood vessel region has a significant skeleton feature, showing a long strip-like spread feature. The normal blood vessel region usually has dense blood vessel branches within 3 mm around the region due to the anatomical running of the bronchial vascular bundle. The benign lung nodule usually grows independently, and there are fewer blood vessels around. The malignant lung nodule has a blurred edge due to tumor infiltration, and is adhered to the surrounding tissue or blood vessels, so the blood vessel features of the suspected lesion region can be obtained according to the blood vessel skeleton and the edge definition in the suspected lesion region.

[0053] Considering that the lung nodule is a lesion caused by lung inflammation, it contains exudate, necrotic tissue and residual gas-containing alveoli, resulting in mixed density, so the lesion features of the suspected lesion region can be further obtained according to the unevenness of the CT density distribution of the suspected lesion region. The lung nodule region can be screened out according to the alveolar features, the blood vessel features and the lesion features. That is, if a suspected lesion region is not an alveolus, is not a blood vessel and has the lesion features of a lung nodule, the suspected lesion region is obviously a lung nodule region.

[0054] The region in the circle in Figures 3 to 6 , Figures 3 to 6 represents the nodule region. For lung nodules, the growth process in the CT image mainly includes four categories. The first category: from nothing to something, that is, the nodule region appears in the CT image in the review stage, as shown in Figure 3 . The second category: the edge of the nodule region in the CT image is blurred due to malignant lesions. When the nodule tissue cells grow, they will invade the surrounding lung tissue, resulting in burrs, blur or irregularity on the edge of the nodule, as shown in Figure 4 . The third category: the nodule region caused by benign growth decreases from large to small, but still shows a high CT density feature, as shown in Figure 5 . The fourth category: the region caused by free growth of the nodule increases from small to large, as shown in Figure 6 .

[0055] After the accurate lung nodule region is determined, in order to enrich the labeling information, the growth type of the lung nodule region can be labeled in the contrast image. In combination with Figures 3 to 6The features in the information labeling module 104 can further judge the nodule change state according to the CT density difference and the area difference between the lung nodule region and the matched initial region, and the edge burr degree of the lung nodule region, and label to the contrast image and be used for model training.

[0056] It should be noted that the neural network in the embodiment of the present application can adopt a deep convolutional network, for example, a DSSD_Xception model, which realizes multi-size convolution kernel feature response operation through a deep separable convolution. The model input in the embodiment of the present application is the CT images at two different CT examination times, and the output is the image at the latest examination, and the image contains the labeling of the lung nodule region and the growth state labeling of each lung nodule region. The training method of the specific deep convolutional network is a technical means familiar to those skilled in the art, which will not be described here.

[0057] In summary, the present application determines the lung nodule region in the contrast image under review detection by comparing the images detected at different times, and can identify the change state of the lung nodule region, and then realize automatic labeling of information rich. In the process of contrast analysis, after aligning the two images, the CT density difference is used for contrast analysis, the suspected lesion region is extracted, and further consideration is given to the various tissue forms of the lung, and the accurate lung nodule region is selected from the alveolar features, vascular features and lesion features. Further, according to the morphological change of the lung nodule region in the two examinations, the accurate nodule change state can be judged, and then the rich and effective information is labeled in the contrast image. The neural network can be trained based on the labeled image, so that the neural network can identify the nodule region and its growth state in the lung CT image by comparing the lung CT images at different examinations, and assist the doctor in making effective judgment on the patient.

[0058] Preferably, in the embodiment of the present application, it is considered that the position such as blood vessels is easily affected by human activities such as breathing and displacement changes occur, while the spine is the most stable skeletal structure in the chest, and its shape can remain relatively rigid in the breathing change, and only a small deformation can occur under the action of gravity. In the CT image, the high-density calcified edge of the outer contour of the spine section is clear and identifiable, and is suitable as a reference for CT image registration. Therefore, the stable structure of the spine is selected as the registration reference in the embodiment of the present application, and the edge detection algorithm is used to obtain the spine contour in the initial image and the contrast image; a plurality of non-collinear inflection points on the curvature of the spine contour are selected as matching feature points; and the matching feature points between the initial image and the contrast image are matched and aligned based on the sequential constraint matching algorithm.

[0059] Specifically, as an example, in the embodiment of the present application, the edge detection is performed by using the canny edge detection algorithm, and the spine profile is determined by matching the detected edge with a standard spine edge. The matching algorithm can be a feature point matching algorithm, which is a well-known technical means to those skilled in the art, and will not be described here. Please refer to Figure 7 Fig. 1 shows an alignment diagram of an initial image and a contrast image provided by an embodiment of the present application. Figure 7 The lowermost profile convex corner point is selected as the origin in the image, and the curvature values of each profile point are calculated in the clockwise direction to determine the inflection points, and then two other non-collinear inflection points are selected as matching feature points. The alignment between the two images can be realized based on the sequential constraint matching algorithm.

[0060] Preferably, in the embodiment of the present application, it is considered that the CT image obtained by CT detection includes mixed ribs, mediastinum and other structures in addition to the lung region. The existence of these non-lung tissues will seriously interfere with the subsequent analysis. By using segmentation technology to accurately extract the lung region, the high-density bone artifacts and low-density fat interference can be effectively excluded, the bone artifacts can be avoided to cover the small lesions in the lung, and the boundary between the lung tissue and the pleural cavity can be clearly defined, so that the sensitivity of lesion detection is significantly improved. For space-occupying lesions such as pulmonary nodules, the morphological changes of the lesions can be accurately calculated after segmentation, and the measurement results can be avoided to be interfered by external factors such as pleural adhesion. Therefore, in the embodiment of the present application, the suspected lesion region in the contrast image is selected, which includes:

[0061] The lung region in the contrast image and the initial image is extracted by using the CT density value of the pixel point. The lung regions in the contrast image and the initial image are compared to obtain the CT density difference between the pixel points at the same position. The CT density difference is subjected to statistical significance test to obtain the significance value of each pixel point in the lung curve in the contrast image. The purpose of the statistical significance test is to suppress noise interference, and the expression of the specific significance value is:

[0062] ; wherein is the significance value at the coordinate position, is the CT density difference at the coordinate position obtained by comparison and analysis, is the average CT density difference at all coordinates, is the standard deviation of the CT density difference at all coordinates.

[0063] The region composed of continuous pixel points with significance values satisfying the preset numerical conditions is regarded as the suspected lesion region in the contrast image.

[0064] In one embodiment of the present application, the CT density difference is the CT density of a pixel in the contrast image minus the CT density of the same position in the initial image, so the value is positive or negative, and the sign is also retained after statistical significance test, so the value condition is set as: if the absolute value of the significant value is greater than a preset value threshold, it is determined that the condition is met. The value threshold can be set to 2.58 in the embodiment of the present application according to experimental results.

[0065] As an example, in the embodiment of the present application, the lung region is obtained by clustering the image using the K-means clustering algorithm, where K is set to 2, representing a high-CT-density cluster formed by blood vessels and bones, and a low-CT-density cluster formed by lung parenchyma and background regions. Morphological opening operation is performed on the high-CT-density cluster to eliminate the blood vessel and bone regions, generate a mask, and then extract the lung image. Please refer to Figure 8 which shows a lung region diagram provided by one embodiment of the present application.

[0066] Preferably, in the embodiment of the present application, the matching method of the initial region includes:

[0067] In the initial image, the pixels are clustered based on the CT density of each pixel to obtain a plurality of clustering regions; for each suspected lesion region in the contrast image, the suspected lesion region is mapped to the initial image to obtain a mapping region, and the clustering region with the largest overlap with the mapping region is selected as the initial region matched with the suspected lesion region. The clustering method in the initial image can use the DBSCAN clustering method, and the distance metric is the absolute value of the CT density difference between the pixels. The specific clustering algorithm process is not described again, each clustering cluster obtained will form a corresponding clustering region, and then a plurality of clustering regions are obtained.

[0068] Preferably, in the embodiment of the present application, the lung cavity feature acquisition method includes:

[0069] Obtain the area difference between the suspected lesion region and the matched initial region and the average CT density value difference; take the ratio of the area difference and the average CT density value difference as the alveolus feature. The average CT density value difference is the average CT density of the suspected lesion region minus the average CT density in the initial region; the area difference is the area of the suspected lesion region minus the area of the initial region. Because the average CT density value difference can represent the air proportion of the tissue, if the suspected lesion region is an alveolus region, the area difference will become larger while the average CT density value difference will also become larger; but if the suspected lesion region is a blood vessel or nodule region, the area difference will become larger while the average CT density value difference will not change or will have a weak change, so the smaller the alveolus feature obtained, the more likely the suspected lesion region is an alveolus region. After the area difference and the average CT density value difference are normalized in their respective dimensions in the embodiment of the present application, the alveolus feature threshold is set to 1, and the suspected lesion region with the alveolus feature smaller than the alveolus feature threshold is an alveolus region.

[0070] Preferably, in the embodiment of the present application, the blood vessel feature acquisition method comprises:

[0071] The suspected lesion region is expanded according to a preset size to obtain an expanded region; and the ratio of the total length of the blood vessel skeleton in the expanded region to the area of the expanded region is taken as the blood vessel density index. In the embodiment of the present application, considering the actual situation, the normal blood vessel region usually has dense blood vessel branches within a range of 3 mm around the bronchial blood vessel bundle due to the anatomical running, so the range obtained by scaling 3 mm in the image is taken as the preset size.

[0072] The gradient average value of each pixel point on the boundary of the suspected lesion region is obtained, and the gradient average value is normalized to obtain the edge sharpness.

[0073] The blood vessel density index and the edge sharpness are positively fused to obtain the blood vessel feature. The positive fusion in the embodiment of the present application adopts the multiplication method.

[0074] It should be noted that the normalization and other means adopted in the embodiment of the present application can adopt range standardization, that is, normalization is realized by statistics of the maximum and minimum values in their respective dimensions, which will not be described and limited in detail.

[0075] In the embodiment of the present application, because the larger the blood vessel density index and the more clear the edge, the more likely the suspected lesion region is a blood vessel region, the larger the blood vessel feature obtained, the more likely it is a blood vessel region. After the blood vessel feature is normalized, the blood vessel feature threshold is set to 0.7, and if the normalized blood vessel feature is greater than 0.7, the corresponding suspected lesion region is taken as a blood vessel region.

[0076] Preferably, in the embodiments of the present application, the CT density standard deviation in the suspected lesion region is taken as the lesion feature.

[0077] Preferably, the lung nodule region is screened according to the alveolus feature, the blood vessel feature and the lesion feature, and the screening comprises:

[0078] The alveolus region in the suspected lesion region is removed according to the alveolus feature, the blood vessel region in the suspected lesion region is removed according to the blood vessel feature, and the suspected lesion region in which the lesion feature meets the preset lesion condition is selected as the lung nodule region from the remaining suspected lesion region. Because the greater the lesion feature is, the more significant the density inhomogeneity in the region is, and the lesion feature meets the lung inflammation lesion feature, the threshold value of the lesion feature is set to 20, and the case that the lesion feature is greater than the threshold value of the lesion feature is taken as the case that meets the preset lesion condition.

[0079] Preferably, in the embodiments of the present application, the nodule change state is judged according to the CT density difference and the area difference between the lung nodule region and the matched initial region, and the edge burr degree of the lung nodule region, and the judgment comprises:

[0080] The following judgment processes are sequentially performed:

[0081] The CT density value is used to judge whether the initial region matched with the lung nodule region is a lung parenchyma region, if the initial region is a lung parenchyma region, and the average CT density of the lung nodule region is greater than the average CT density of the initial region, the lung nodule region is identified as a new nodule region. Because the CT density value range of the lung parenchyma region is between-950 and-600, the average CT density value of the initial region can be directly compared with the density value range to judge whether it belongs to the lung parenchyma region.

[0082] If the edge burr degree of the lung nodule region is greater than the preset burr degree threshold value, the lung nodule region is identified as a malignant lesion nodule region. In the embodiments of the present application, the edge burr degree is normalized, and the burr degree threshold value is set to 0.1.

[0083] If the area of the lung nodule region relative to the matched initial region becomes smaller, and the CT density increases by at least a preset increase threshold value, the lung nodule region is judged as a benign growth nodule region. In the embodiments of the present application, the increase threshold value is set to 10.

[0084] If the area of the lung nodule region relative to the matched initial region becomes larger, the lung nodule region is judged as a nodule increase region.

[0085] It is to be noted that the sequential order of the above-described embodiments of the present application only for the purpose of description, but not the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0086] Each of the embodiments in the specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments.

Claims

1.A training system of an artificial intelligence-based lung nodule detection model, characterized by, The system comprises: a lung CT detection image acquisition module, configured to obtain an initial image of a patient's lung under a previous CT detection before a review and a contrast image under a review detection; a region analysis module, configured to, after aligning the initial image and the contrast image, screen a suspected lesion region in the contrast image and an initial region in the initial image matched with the suspected lesion region according to a CT density difference between the same positions of the lung regions on the two images; a nodule identification module, configured to obtain a lung cavity feature of the suspected lesion region according to an area change and a CT density change between the suspected lesion region and the initial region, obtain a blood vessel feature of the suspected lesion region according to a blood vessel skeleton in the suspected lesion region and an edge definition of the suspected lesion region, and obtain a lesion feature of the suspected lesion region according to a non-uniformity of a CT density distribution of the suspected lesion region; and screen a lung nodule region according to the lung cavity feature, the blood vessel feature and the lesion feature; an information labeling module, configured to judge a nodule change state according to a CT density difference and an area difference between the lung nodule region and the matched initial region and an edge burr degree of the lung nodule region, and label the nodule change state to the contrast image and use the nodule change state for model training; the screening of the suspected lesion region in the contrast image comprises: extracting the lung regions in the contrast image and the initial image by using CT density values of pixel points; comparing the lung regions between the contrast image and the initial image to obtain a CT density difference between the pixel points at the same positions; performing a statistical significance test on the CT density difference to obtain a significance value of each pixel point in a lung curve in the contrast image; and regarding a region formed by continuous pixel points with the significance value satisfying a preset numerical condition as the suspected lesion region in the contrast image; the matching method of the initial region comprises: in the initial image, clustering the pixel points based on CT densities of the pixel points to obtain a plurality of clustering regions; for each suspected lesion region in the contrast image, mapping the suspected lesion region to the initial image to obtain a mapping region, and selecting a clustering region with the largest overlap with the mapping region as the initial region matched with the suspected lesion region. 2.The training system of an artificial intelligence-based lung nodule detection model according to claim 1, wherein, the aligning method of the initial image and the contrast image comprises: obtaining a spine contour in the initial image and the contrast image by using an edge detection algorithm; selecting a plurality of non-collinear inflection points on the spine contour as matching feature points; and matching and aligning the matching feature points between the initial image and the contrast image based on a sequential constraint matching algorithm. 3.The training system of an artificial intelligence-based lung nodule detection model according to claim 1, wherein, the lung cavity feature acquisition method comprises: obtaining an area difference and an average CT density value difference between the suspected lesion region and the matched initial region; and regarding a ratio of the area difference to the average CT density value difference as the lung cavity feature. 4.The training system of an artificial intelligence-based lung nodule detection model according to claim 1, wherein, the blood vessel feature acquisition method comprises: expanding the suspected lesion region by a preset size to obtain an expanded region; regarding a ratio of a total length of a blood vessel skeleton in the expanded region to an area of the expanded region as a blood vessel density index; obtaining a gradient average value of each pixel point on a boundary of the suspected lesion region; and performing normalization processing on the gradient average value to obtain an edge definition. The blood vessel density index and the edge definition are positively fused to obtain the blood vessel feature. 5.The training system of an artificial intelligence-based lung nodule detection model according to claim 1, wherein, The method for obtaining the lesion feature comprises: A CT density standard deviation in the suspected lesion region is taken as the lesion feature. 6.The training system of an artificial intelligence-based lung nodule detection model according to claim 1, wherein, The lung nodule region is screened according to the alveolus feature, the blood vessel feature and the lesion feature, which comprises: The alveolus region in the suspected lesion region is removed according to the alveolus feature, the blood vessel region in the suspected lesion region is removed according to the blood vessel feature, and the suspected lesion region in which the lesion feature meets the preset lesion condition is selected as the lung nodule region from the remaining suspected lesion region. 7.The training system of an artificial intelligence-based lung nodule detection model according to claim 1, wherein, The change state of the nodule is judged according to the CT density difference and the area difference between the lung nodule region and the matched initial region, and the edge burr degree of the lung nodule region, which comprises: The following judgment processes are sequentially performed: It is judged whether the initial region matched with the lung nodule region is a lung parenchyma region by using the CT density value, if the initial region is a lung parenchyma region and the average CT density of the lung nodule region is greater than the average CT density of the initial region, the lung nodule region is identified as a new nodule region; If the edge burr degree of the lung nodule region is greater than a preset burr degree threshold, the lung nodule region is identified as a malignant lesion nodule region; If the area of the lung nodule region relative to the matched initial region is reduced and the CT density is increased by at least a preset increase threshold, the lung nodule region is a benign growth nodule region; If the area of the lung nodule region relative to the matched initial region is increased, the lung nodule region is a nodule increase region. 8.The training system of an artificial intelligence-based lung nodule detection model according to claim 1, wherein, The method for obtaining the edge burr degree comprises: The lung nodule region is subjected to an opening operation to obtain a burr-free region, the lung nodule region is subtracted from the burr-free region, and the remaining region is a burr region; and the area proportion of the burr region in the lung nodule region is taken as the edge burr degree.

Images