Artificial intelligence-based pulmonary nodule detection model training system

Abstract

The invention relates to the technical field of image feature recognition, in particular to a pulmonary nodule detection model training system based on artificial intelligence. According to the system, after two images detected in different periods are aligned, the CT density difference is used for comparative analysis, suspected lesion areas are extracted, and accurate lung nodule areas are screened out from alveolar features, blood vessel features and lesion features. The accurate nodule change state can be judged according to the morphological change of the lung nodule area in the two times of examination, and rich and effective information is marked in a comparison image. According to the method, the neural network can be trained based on the marked image, so that the neural network can accurately identify the nodule region in the lung CT image and the growth state of the nodule region by comparing the lung CT images examined for different times.

Description

Technical Field

[0001] This invention relates to the field of image feature recognition technology, and more specifically to a training system for a lung nodule detection model based on artificial intelligence. Background Technology

[0002] Pulmonary nodules are distinctive features in lung CT images. A neural network for identifying pulmonary nodules can be trained using CT images of lungs containing nodule regions. A conventional neural network takes a target patient's lung CT image as input and outputs labeled nodule regions. To ensure accuracy, the neural network requires extensive training data, specifically labeled lung CT images with nodule regions. Labeling these training images is labor-intensive, and errors can occur due to personal experience, affecting the neural network's accuracy and ultimately hindering its ability to detect nodules effectively, thus failing to adequately assist physicians in their diagnosis. Summary of the Invention

[0003] To address the technical problem in existing technologies where inaccurate labeling of nodule regions in lung CT images prevents nodule recognition neural networks from effectively detecting and identifying nodule regions, the present invention aims to provide a training system for an artificial intelligence-based lung nodule detection model. The specific technical solution adopted is as follows: This invention proposes a training system for an artificial intelligence-based lung nodule detection model, the system comprising: The lung CT image acquisition module is used to obtain the initial image of the patient's lungs under the previous CT scan, as well as the comparison image under the follow-up scan. The region analysis module is used to align the initial image and the comparison image, and then filter out the suspected lesion areas in the comparison image and the initial regions in the initial image that match the suspected lesion areas based on the CT density difference between the same location in the lung region on the two images. The nodule identification module is used to obtain alveolar features of suspected lesion areas based on the area changes and CT density changes between the suspected lesion areas and the initial areas; to obtain vascular features of suspected lesion areas based on the vascular skeleton and edge clarity in the suspected lesion areas; to obtain lesion features of suspected lesion areas based on the uneven distribution of CT density in the suspected lesion areas; and to screen out lung nodule areas based on alveolar features, vascular features, and lesion features. The information annotation module is used to determine the nodule change status based on the CT density and area differences between the lung nodule region and the matched initial region, as well as the degree of edge spiculation of the lung nodule region, and to annotate it to the comparison image for model training.

[0004] Furthermore, methods for aligning the initial image and the comparison image include: An edge detection algorithm is used to obtain the spinal contour in the initial image and the comparison image; multiple non-collinear inflection points are selected as matching feature points based on the curvature of the spinal contour; and the matching feature points between the initial image and the comparison image are matched and aligned based on the order constraint matching algorithm.

[0005] Furthermore, the process of filtering out suspected lesion areas in the comparative images includes: The lung region in the comparison image and the initial image is extracted using the CT density value of the pixels. The lung regions in the comparison image and the initial image are compared to obtain the CT density difference between pixels at the same location. The CT density difference is subjected to statistical significance test to obtain the significance value of each pixel in the lung curve in the comparison image. The region formed by consecutive pixels whose significance values ​​meet the preset numerical conditions is taken as the suspected lesion region in the comparison image.

[0006] Furthermore, the matching method for the initial region includes: In the initial image, pixels are clustered based on the CT density of each pixel to obtain multiple clustered regions. For each suspected lesion region in the comparison image, the suspected lesion region is mapped to the initial image to obtain a mapped region. The clustered region with the largest overlap of the mapped regions is selected as the initial region for matching the suspected lesion region.

[0007] Furthermore, the method for obtaining the alveolar features includes: Obtain the area difference and mean CT density value difference between the suspected lesion area and the matched initial area; use the ratio of area difference to mean CT density value difference as alveolar characteristics.

[0008] Furthermore, the method for obtaining the vascular features includes: The suspected lesion area is expanded according to a preset size to obtain an expanded area; the ratio of the total length of the vascular skeleton in the expanded area to the area of ​​the expanded area is used as the vascular density index. The gradient average value of each pixel on the boundary of the suspected lesion area is obtained, and the gradient average value is normalized to obtain the edge sharpness. The vascular features are obtained by positively fusing the vascular density index and the edge clarity.

[0009] Furthermore, the method for obtaining the lesion features includes: The standard deviation of CT density in the suspected lesion area is used as the lesion feature.

[0010] Furthermore, the screening of pulmonary nodule regions based on alveolar characteristics, vascular characteristics, and lesion characteristics includes: Based on alveolar characteristics, alveolar regions in the suspected lesion area are removed; based on vascular characteristics, vascular regions in the suspected lesion area are removed; among the remaining suspected lesion areas, those whose lesion characteristics meet the preset lesion conditions are selected as pulmonary nodule areas.

[0011] Furthermore, the step of determining the nodule change status based on the CT density and area differences between the lung nodule region and the matched initial region, as well as the degree of edge spiculation in the lung nodule region, includes: The following judgment process will be executed sequentially: The CT density value is used to determine 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, then the lung nodule region is identified as a newly added nodule region. If the edge spiculation of the lung nodule region is greater than a preset spiculation threshold, then the lung nodule region is identified as a malignant lesion nodule region. If the area of ​​the lung nodule region becomes smaller relative to the matched initial region, and the CT density increases by at least a preset increase threshold, then the lung nodule region is determined to be a benign growth nodule region. If the area of ​​the lung nodule region increases relative to the initial matching region, then the lung nodule region is determined to be an enlarged nodule region.

[0012] Furthermore, the method for obtaining the degree of edge burrs includes: The lung nodule region is opened to obtain the non-spiculated region; the non-spiculated region is subtracted from the lung nodule region, and the remaining region is the spiculated region; the proportion of the spiculated region to the lung nodule region is used as the degree of edge spiculation.

[0013] The present invention has the following beneficial effects: This invention considers that nodules exhibit growth characteristics over time. Therefore, by comparing images from different examination periods, it identifies lung nodule regions in comparative images from follow-up examinations and can recognize changes in these regions, thus achieving information-rich automatic annotation. During the comparative analysis, two images are aligned and compared using CT density differences to extract suspected lesion areas. Furthermore, considering various lung tissue morphologies, accurate lung nodule regions are selected from alveolar, vascular, and lesion features. Further, the precise nodule change status can be determined based on morphological changes in the lung nodule region between the two examinations, thereby annotating the comparative images with rich and effective information. Based on the annotated images, a neural network can be trained, enabling it to identify nodule regions and their growth status in lung CT images from different examinations, assisting doctors in making effective judgments about patients. Attached Figure Description

[0014] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A block diagram of a training system for an artificial intelligence-based lung nodule detection model provided in one embodiment of the present invention; Figure 2 This is a schematic diagram comparing nodules in different CT images of the lungs, provided as an embodiment of the present invention. Figure 3 This is a comparative schematic diagram of nodules in a first type of growth state provided in an embodiment of the present invention; Figure 4 A comparative schematic diagram of nodules in a second type of growth state provided in an embodiment of the present invention; Figure 5 A comparative schematic diagram of nodules in a third growth state provided in an embodiment of the present invention; Figure 6 A comparative schematic diagram of nodules in a fourth growth state provided in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the alignment of an initial image and a comparison image according to an embodiment of the present invention; Figure 8 This is a schematic diagram of a lung region provided in one embodiment of the present invention. Detailed Implementation

[0016] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a training system for an artificial intelligence-based lung nodule detection model proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0017] 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 this invention pertains.

[0018] The specific scheme of the training system for an artificial intelligence-based lung nodule detection model provided by the present invention will be described in detail below with reference to the accompanying drawings.

[0019] Please see Figure 1 The diagram illustrates a training system block diagram of an artificial intelligence-based lung nodule detection model provided in an embodiment of the present invention. The system includes: a lung CT image acquisition module 101, a region analysis module 102, a nodule recognition module 103, and an information annotation module 104.

[0020] This invention aims to determine the lung nodule region and its growth characteristics in follow-up images by comparing images from different lung CT scans. Therefore, the lung CT image acquisition module 101 is first used to obtain the initial image of the patient's lungs from the previous CT scan, and the comparison image from the follow-up scan. It should be noted that each pixel in a CT image corresponds to a CT density value; for the lung region, the CT density at a location represents the air content at that location.

[0021] Please see Figure 2 This illustration shows a comparative diagram of nodules in different CT images of the lungs, provided by an embodiment of the present invention. Figure 2 The left image is a partial view of the initial image from the previous CT scan, and the right image is a partial view of the comparison image from the previous scan. By comparing the two images, it can be seen that the area circled in the markings belongs to the lung nodule area. The nodule has grown and produced a more significant area. The comparison image and the initial image show a significant difference in CT density at the same location.

[0022] It should be noted that, since the purpose of this invention is to acquire a large number of labeled images for neural network model training, a large number of images need to be automatically labeled. This can be achieved by calling images stored in the hospital's internal database, using a set of initial images and comparison images of the same patient as a set of training data, thereby obtaining multiple sets of training data. According to the relevant regulations for the diagnosis and treatment of pulmonary nodules, the interval between CT scans for pulmonary nodules is 3-6 months; therefore, each set of training data can be determined according to this timeframe.

[0023] like Figure 2 As shown, lung nodules exhibit significant CT density changes between images from different periods due to growth. However, the long time interval between images and the spatial misalignment of anatomical structures caused by respiratory movements and body position during CT scans can lead to lesion displacement due to respiration. Furthermore, changes in body position can deform the projection of structures such as ribs. These differences directly result in inaccurate nodule area measurements in subsequent analyses. The dynamic evolution of nodules needs to be compared within a unified spatial coordinate system; otherwise, subtle changes may be masked by respiratory artifacts or positional deviations, affecting the accuracy of nodule region identification. Therefore, the region analysis module 102 first aligns the initial and comparison images. After alignment, by comparing the CT density differences at the same location in the lung region, suspected lesion areas in the comparison images can be screened. In other words, suspected lesion areas are regions that show significant changes in this follow-up examination compared to the previous one. These changes may be due to differences caused by alveolar or vascular movement, and therefore cannot be directly identified as lung nodule areas, requiring further comparative analysis. Therefore, it is necessary to simultaneously obtain the initial region in the initial image that matches the suspected lesion area. Since the two images have been aligned, the initial region can be obtained by matching the location of the suspected lesion area with the corresponding region in the initial image.

[0024] The nodule identification module 103 further analyzes the comparative features between the suspected lesion area and the matched initial area, as well as the morphological features of the suspected lesion area itself, to screen out accurate lung nodule areas within the suspected lesion area. Within the suspected lesion area, the alveolar and vascular regions are the main sources of false identification. Alveoli are the basic units of gas exchange in the lungs. When alveoli expand due to respiration, their internal air content increases significantly. However, when lung nodules enlarge, their internal structure mainly consists of soft tissue, with a smaller change in air content. This change in air content can be reflected by changes in CT density. Therefore, the alveolar characteristics of the suspected lesion area can be obtained based on the area changes and CT density changes between the suspected lesion area and the initial area. The vascular region has significant skeletal features, exhibiting a long, spreading pattern. Normal vascular regions follow the anatomical course of the bronchovascular bundle, and there are usually dense vascular branches within a 3mm radius around them. Benign pulmonary nodules usually grow independently with fewer surrounding blood vessels, while malignant pulmonary nodules have blurred edges due to tumor infiltration and adhere to surrounding tissues or blood vessels. Therefore, the vascular features of a suspected lesion region can be obtained based on the vascular skeleton and edge clarity in the suspected lesion region.

[0025] Considering that pulmonary nodules are lesions caused by lung inflammation, containing exudate, necrotic tissue, and residual aerated alveoli, resulting in mixed densities, the lesion characteristics of the suspected lesion area can be further obtained based on the uneven distribution of CT density. Pulmonary nodule areas can be screened based on alveolar, vascular, and lesion characteristics. That is, if a suspected lesion area is neither an alveolus nor a blood vessel, and its lesion characteristics meet the criteria for a pulmonary nodule, then this suspected lesion area is clearly a pulmonary nodule area.

[0026] Please see Figures 3 to 6 , Figures 3 to 6 The circled area represents the nodule region. For lung nodules, their growth process on CT images mainly includes four categories. The first category is from non-existence to presence, meaning the nodule region appears on the CT image during follow-up examinations, such as... Figure 3 As shown; the second type: Malignant lesions cause blurred edges in CT images of nodule regions. As nodule cells grow, they invade surrounding lung tissue, resulting in spiculated, blurred, or irregular nodule edges, such as... Figure 4 As shown; the third type: nodules caused by benign growth decrease in size but still exhibit high CT density characteristics, such as... Figure 5 As shown; Category 4: Areas of nodules growing freely, ranging from small to large, such as... Figure 6 As shown.

[0027] After accurately identifying the lung nodule regions, to enrich the annotation information, the lung nodule regions and their growth types were labeled in the comparative images. Combined with... Figures 3 to 6The features in the image can be further identified in the information annotation module 104 based on the differences in CT density and area between the lung nodule region and the matched initial region, as well as the degree of edge spiculation in the lung nodule region, to determine the nodule change status and annotate it to the comparison image for model training.

[0028] It should be noted that the neural network in this embodiment of the invention can be a deep convolutional network, such as the DSSD_Xception model, which achieves multi-size convolutional kernel feature response operations through depthwise separable convolution. The model in this embodiment takes CT images from two different CT examinations as input and outputs an image from the most recent examination, which includes annotations of lung nodule regions and the growth status of each lung nodule region. The specific training methods for the deep convolutional network are well-known to those skilled in the art and will not be elaborated upon here.

[0029] In summary, this invention identifies pulmonary nodule regions in follow-up CT images by comparing images from different examination periods, and can recognize the changing states of these nodule regions, thus achieving information-rich automatic annotation. During the comparative analysis, after aligning the two images, CT density differences are used for comparative analysis to extract suspected lesion areas. Furthermore, considering various lung tissue morphologies, accurate pulmonary nodule regions are selected from alveolar features, vascular features, and lesion features. Further, the morphological changes of the pulmonary nodule regions in the two examinations can determine the accurate nodule change state, thereby annotating rich and effective information in the comparative images. Based on the annotated images, this invention can train a neural network, enabling the network to identify nodule regions and their growth states in lung CT images by comparing lung CT images from different examinations, assisting doctors in making effective judgments about patients.

[0030] Preferably, in this embodiment of the invention, considering that blood vessels and other structures are easily displaced by human activities such as respiration, and that the spine is the most stable skeletal structure within the thoracic cavity, its shape maintains relative rigidity during respiration and may only undergo minor deformation under gravity. In CT images, the high-density calcified edges of the spine's cross-section are clearly discernible, making it suitable as a reference for CT image registration. Therefore, this embodiment of the invention selects the structurally stable spine as the registration reference, uses an edge detection algorithm to obtain the spine contour in the initial and comparison images, selects multiple non-collinear inflection points as matching feature points based on the curvature of the spine contour, and matches and aligns the matching feature points between the initial and comparison images based on a sequential constraint matching algorithm.

[0031] Specifically, as an example, in this embodiment of the invention, the Canny edge detection algorithm is used for edge detection. The spinal contour is determined by matching and comparing the detected edges with the standard spinal edge. The matching algorithm can be a feature point matching algorithm, which is a well-known technique and will not be elaborated here. Please refer to [link to relevant documentation]. Figure 7 This illustrates a schematic diagram of initial and contrast image alignment provided by an embodiment of the present invention. Figure 7 The origin is selected from the bottommost protruding corner of the contour. The curvature value of each contour point is calculated sequentially in a clockwise direction, and the inflection point is determined. Then, two other non-collinear inflection points are selected as matching feature points. Alignment between two images can be achieved based on the sequential constraint matching algorithm.

[0032] Preferably, in this embodiment of the invention, considering that CT images obtained from CT scans include not only the lung region but also mixed structures such as ribs and mediastinum, the presence of these non-lung tissues can seriously interfere with subsequent analysis, the lung region can be accurately extracted through segmentation technology. This effectively eliminates high-density bone artifacts and low-density fat interference, preventing bone artifacts from obscuring small lung lesions, while clearly defining the boundary between lung tissue and the pleural cavity, significantly improving the sensitivity of lesion detection. For space-occupying lesions such as pulmonary nodules, segmentation allows for accurate calculation of lesion morphological changes, avoiding interference from external factors such as pleural adhesions with measurement results. Therefore, in this embodiment of the invention, suspected lesion areas in the comparative images are screened, including: The lung region in both the comparison image and the initial image is extracted using the CT density values ​​of individual pixels. The lung regions in the comparison image and the initial image are compared to obtain the CT density differences between pixels at the same location. Statistical significance testing is performed on these CT density differences to obtain the significance value of each pixel in the lung curve of the comparison image. The purpose of the statistical significance test is to suppress noise interference; the specific expression for the significance value is as follows: ;in coordinates Significant value at location, Coordinates obtained through comparative analysis CT density differences at location The average CT density difference across all coordinates. The standard deviation of CT density difference across all coordinates.

[0033] The region consisting of consecutive pixels whose significant values ​​meet the preset numerical conditions is taken as the suspected lesion region in the comparison image.

[0034] In one embodiment of the present invention, the CT density difference is the CT density of pixels in the comparison image minus the CT density at the same location in the initial image. Therefore, this value can be positive or negative, and the sign is retained after statistical significance testing. Thus, the numerical condition is set as follows: if the absolute value of the significant value is greater than a preset numerical threshold, the condition is considered satisfied. The numerical threshold in this embodiment of the present invention can be set to 2.58 based on experimental results.

[0035] As an example, in this embodiment of the invention, the lung region is obtained by clustering the image using the K-means clustering algorithm, where K is set to 2, representing high CT density clusters formed by blood vessels and bones, and low CT density clusters formed by lung parenchyma and background regions. Morphological opening operations are performed on the high CT density clusters to eliminate the vascular and skeletal regions, generating a mask for extracting the lung image. Please refer to [link to relevant documentation]. Figure 8 This illustrates a schematic diagram of a lung region provided by an embodiment of the present invention.

[0036] Preferably, in this embodiment of the invention, the initial region matching method includes: In the initial image, pixels are clustered based on their CT density to obtain multiple clustered regions. For each suspected lesion region in the comparison image, the suspected lesion region is mapped onto the initial image to obtain a mapped region. The clustered region with the largest overlap of the mapped regions is selected as the initial region for matching the suspected lesion region. The clustering method in the initial image can be the DBSCAN clustering method, with the distance metric being the absolute value of the CT density difference between pixels. The specific clustering algorithm process will not be elaborated here. Each cluster will form a corresponding clustered region, thus obtaining multiple clustered regions.

[0037] Preferably, in this embodiment of the invention, the method for obtaining alveolar features includes: The method obtains the area difference and the average CT density value difference between the suspected lesion area and the matched initial area; the ratio of the area difference to the average CT density value difference is used as the alveolar feature. The average CT density value difference is the average CT density of the suspected lesion area minus the average CT density of the initial area; the area difference is the area of ​​the suspected lesion area minus the area of ​​the initial area. Because the average CT density value difference can characterize the air content of tissue, if the suspected lesion area is an alveolar region, both the area difference and the average CT density value difference will increase; however, if the suspected lesion area is a blood vessel or nodule, the area difference will increase while the average CT density value difference remains unchanged or changes only slightly. Therefore, the smaller the alveolar feature value, the more likely the suspected lesion area is an alveolar region. In this embodiment, after normalizing the area difference and the average CT density value difference in their respective dimensions using range standardization, an alveolar feature threshold of 1 is set. Suspected lesion areas with alveolar features less than the alveolar feature threshold are considered alveolar regions.

[0038] Preferably, in this embodiment of the invention, the method for obtaining vascular features includes: The suspected lesion area is expanded according to a preset size to obtain an expanded area; the ratio of the total length of the vascular skeleton in the expanded area to the area of ​​the expanded area is used as the vascular density index. In this embodiment of the invention, considering that in actual situations, normal vascular areas usually have dense vascular branches within a 3mm radius around them because they follow the anatomical course of the bronchovascular bundle, the area obtained by scaling 3mm proportionally in the image is used as the preset size.

[0039] The gradient average value of each pixel on the boundary of the suspected lesion area is obtained, and the gradient average value is normalized to obtain the edge sharpness.

[0040] The vascular features are obtained by forward fusing the vascular density index and the edge sharpness. In this embodiment of the invention, forward fusing employs a multiplication method.

[0041] It should be noted that the normalization and other methods used in the embodiments of the present invention can all be range standardization, that is, normalization is achieved by statistically analyzing the maximum and minimum values ​​in their respective dimensions. The specifics will not be elaborated or limited further.

[0042] In this embodiment of the invention, since a higher vascular density index and clearer edges indicate that the suspected lesion area is more likely to be a vascular area, a larger vascular feature value further suggests that it should be a vascular area. After normalizing the vascular features, a vascular feature threshold of 0.7 is set. If the normalized vascular feature value is greater than 0.7, the corresponding suspected lesion area is considered a vascular area.

[0043] Preferably, in this embodiment of the invention, the standard deviation of CT density in the suspected lesion area is used as the lesion feature.

[0044] Preferably, pulmonary nodule regions are selected based on alveolar characteristics, vascular characteristics, and lesion characteristics, including: Alveolar regions were removed from the suspected lesion area based on alveolar characteristics, and vascular regions were removed from the suspected lesion area based on vascular characteristics. Among the remaining suspected lesion areas, those whose lesion characteristics met the preset lesion criteria were selected as pulmonary nodule areas. Since a larger lesion characteristic indicates significantly uneven density within the area, consistent with pneumonia lesions, a lesion characteristic threshold of 20 was set. Cases with lesion characteristics greater than this threshold were considered to meet the preset lesion criteria.

[0045] Preferably, in this embodiment of the invention, determining the nodule change state based on the CT density and area differences between the lung nodule region and the matched initial region, as well as the degree of edge spiculation in the lung nodule region, includes: The following judgment process will be executed sequentially: CT density values ​​are used to determine whether the initial region matched to 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, then the lung nodule region is identified as a newly added nodule region. Since the CT density value range of the lung parenchyma region is between -950 and -600, it is possible to directly determine whether it belongs to the lung parenchyma region by comparing the average CT density value of the initial region with the density value range.

[0046] If the degree of edge spiculation in the lung nodule region is greater than a preset spiculation degree threshold, the lung nodule region is determined to be a malignant lesion nodule region. In this embodiment of the invention, the degree of edge spiculation is normalized, and the spiculation degree threshold is set to 0.1.

[0047] If the area of ​​the lung nodule region decreases relative to the matched initial region, and the CT density increases by at least a preset increase threshold, then the lung nodule region is determined to be a benign growth nodule region. In this embodiment of the invention, the increase threshold is set to 10.

[0048] If the area of ​​the lung nodule region increases relative to the initial matching region, then the lung nodule region is determined to be an enlarged nodule region.

[0049] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0050] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A training system for an artificial intelligence-based lung nodule detection model, characterized in that, The system includes: The lung CT image acquisition module is used to obtain the initial image of the patient's lungs under the previous CT scan, as well as the comparison image under the follow-up scan. The region analysis module is used to align the initial image and the comparison image, and then filter out the suspected lesion areas in the comparison image and the initial regions in the initial image that match the suspected lesion areas based on the CT density difference between the same location in the lung region on the two images. The nodule identification module is used to obtain alveolar features of suspected lesion areas based on the area changes and CT density changes between the suspected lesion areas and the initial areas; to obtain vascular features of suspected lesion areas based on the vascular skeleton and edge clarity in the suspected lesion areas; to obtain lesion features of suspected lesion areas based on the uneven distribution of CT density in the suspected lesion areas; and to screen out lung nodule areas based on alveolar features, vascular features, and lesion features. The information annotation module is used to determine the nodule change status based on the CT density and area differences between the lung nodule region and the matched initial region, as well as the degree of edge spiculation of the lung nodule region, and to annotate it to the comparison image for model training.

2. The training system for an artificial intelligence-based lung nodule detection model according to claim 1, characterized in that, Methods for aligning the initial image and the comparison image include: An edge detection algorithm is used to obtain the spinal contour in the initial image and the comparison image; multiple non-collinear inflection points are selected as matching feature points based on the curvature of the spinal contour; and the matching feature points between the initial image and the comparison image are matched and aligned based on the order constraint matching algorithm.

3. The training system for an artificial intelligence-based lung nodule detection model according to claim 1, characterized in that, The process of filtering out suspected lesion areas in the comparison images includes: The lung region in the comparison image and the initial image is extracted using the CT density value of the pixels. The lung regions in the comparison image and the initial image are compared to obtain the CT density difference between pixels at the same location. The CT density difference is subjected to statistical significance test to obtain the significance value of each pixel in the lung curve in the comparison image. The region formed by consecutive pixels whose significance values ​​meet the preset numerical conditions is taken as the suspected lesion region in the comparison image.

4. The training system for an artificial intelligence-based lung nodule detection model according to claim 1, characterized in that, The matching method for the initial region includes: In the initial image, pixels are clustered based on the CT density of each pixel to obtain multiple clustered regions. For each suspected lesion region in the comparison image, the suspected lesion region is mapped to the initial image to obtain a mapped region. The clustered region with the largest overlap of the mapped regions is selected as the initial region for matching the suspected lesion region.

5. The training system for an artificial intelligence-based lung nodule detection model according to claim 1, characterized in that, The method for obtaining the alveolar features includes: Obtain the area difference and mean CT density value difference between the suspected lesion area and the matched initial area; use the ratio of area difference to mean CT density value difference as alveolar characteristics.

6. The training system for an artificial intelligence-based lung nodule detection model according to claim 1, characterized in that, The method for obtaining the vascular features includes: The suspected lesion area is expanded according to a preset size to obtain an expanded area; the ratio of the total length of the vascular skeleton in the expanded area to the area of ​​the expanded area is used as the vascular density index. The gradient average value of each pixel on the boundary of the suspected lesion area is obtained, and the gradient average value is normalized to obtain the edge sharpness. The vascular features are obtained by positively fusing the vascular density index and the edge clarity.

7. The training system for an artificial intelligence-based lung nodule detection model according to claim 1, characterized in that, The methods for obtaining the lesion features include: The standard deviation of CT density in the suspected lesion area is used as the lesion feature.

8. The training system for an artificial intelligence-based lung nodule detection model according to claim 1, characterized in that, The process of screening pulmonary nodule regions based on alveolar characteristics, vascular characteristics, and lesion characteristics includes: Based on alveolar characteristics, alveolar regions in the suspected lesion area are removed; based on vascular characteristics, vascular regions in the suspected lesion area are removed; among the remaining suspected lesion areas, those whose lesion characteristics meet the preset lesion conditions are selected as pulmonary nodule areas.

9. The training system for an artificial intelligence-based lung nodule detection model according to claim 1, characterized in that, The method of determining the nodule change status based on the CT density and area differences between the lung nodule region and the matched initial region, as well as the degree of edge spiculation in the lung nodule region, includes: The following judgment process will be executed sequentially: The CT density value is used to determine 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, then the lung nodule region is identified as a newly added nodule region. If the edge spiculation of the lung nodule region is greater than a preset spiculation threshold, then the lung nodule region is identified as a malignant lesion nodule region. If the area of ​​the lung nodule region becomes smaller relative to the matched initial region, and the CT density increases by at least a preset increase threshold, then the lung nodule region is determined to be a benign growth nodule region. If the area of ​​the lung nodule region increases relative to the initial matching region, then the lung nodule region is determined to be an enlarged nodule region.

10. The training system for an artificial intelligence-based lung nodule detection model according to claim 1, characterized in that, The method for obtaining the degree of edge burrs includes: The lung nodule region is opened to obtain the non-spiculated region; the non-spiculated region is subtracted from the lung nodule region, and the remaining region is the spiculated region; the proportion of the spiculated region to the lung nodule region is used as the edge spiculation degree.

Images