Intelligent labeling method and system for medical CT image

Abstract

The embodiment of the invention provides an intelligent labeling method and system for a medical CT image, and belongs to the technical field of data labeling. The method comprises the steps of obtaining a user labeling result corresponding to a target CT image, and obtaining an exception description text corresponding to the target CT image; generating a corresponding initial template according to the exception description text by utilizing a template generation model; performing edge information extraction on the target CT image to obtain target edge information, and performing similarity calculation according to the initial template and the target edge information to obtain a similar result corresponding to the target edge information; obtaining a target template corresponding to the abnormal description text from the target edge information according to the similar result; performing labeling result supervision on the user labeling result according to the target template to obtain a target labeling supervision result corresponding to the target CT image; and performing data correction on the target CT image according to the target annotation supervision result to obtain a target annotation result. The problem of low labeling quality in related technologies is solved, and the quality of image labeling is also improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data annotation, and in particular to an intelligent annotation method and system for medical CT images. Background Art

[0002] Medical CT (computed tomography) images play a vital role in modern clinical diagnosis and treatment planning. These images can provide detailed information about the internal structure of the human body, helping doctors identify and locate lesions, plan surgical paths, and evaluate treatment effects. However, the analysis and annotation of CT images, especially the precise location, segmentation, and quantitative analysis of lesions, has long been a challenge.

[0003] With the development of artificial intelligence technology, especially the significant progress in deep learning and computer vision, users are usually required to perform manual annotation in the initial stage to support model training. However, if the quality of manual annotation is low, it will directly affect the effect of subsequent models. In current technology, in order to ensure the quality of annotation, cross-verification is usually adopted, which requires repeated annotation, thus greatly affecting the efficiency of image annotation. Summary of the invention

[0004] The main purpose of the embodiments of the present invention is to provide an intelligent annotation method and system for medical CT images, aiming to solve the problem of low annotation quality when annotating images in the related art.

[0005] In a first aspect, an embodiment of the present invention provides an intelligent annotation method for medical CT images, comprising:

[0006] Obtaining a user annotation result corresponding to a target CT image, and obtaining an abnormality description text corresponding to the target CT image;

[0007] Generate a corresponding initial template according to the abnormal description text using a template generation model;

[0008] Extracting edge information of the target CT image to obtain target edge information, and performing similarity calculation based on the initial template and the target edge information to obtain a similarity result corresponding to the target edge information;

[0009] Obtaining a target template corresponding to the abnormal description text from the target edge information according to the similarity result;

[0010] Performing labeling supervision on the user labeling result according to the target template to obtain a target labeling supervision result corresponding to the target CT image;

[0011] The target CT image is data corrected according to the target labeling supervision result to obtain a target labeling result.

[0012] In a second aspect, an embodiment of the present invention provides an intelligent annotation system for medical CT images, comprising:

[0013] A data acquisition module, used to obtain the user annotation result corresponding to the target CT image, and obtain the abnormality description text corresponding to the target CT image;

[0014] A template generation module, used to generate a corresponding initial template according to the abnormal description text using a template generation model;

[0015] A similarity calculation module is used to extract edge information of the target CT image to obtain target edge information, and perform similarity calculation based on the initial template and the target edge information to obtain a similarity result corresponding to the target edge information;

[0016] A template screening module, used for obtaining a target template corresponding to the abnormal description text from the target edge information according to the similarity result;

[0017] A data comparison module, used to supervise the user's annotation results according to the target template, and obtain a target annotation supervision result corresponding to the target CT image;

[0018] The data correction module is used to perform data correction on the target CT image according to the target labeling supervision result to obtain the target labeling result.

[0019] The embodiment of the present invention provides an intelligent annotation method and system for medical CT images, the method comprising: obtaining a user annotation result corresponding to a target CT image, and obtaining an abnormal description text corresponding to the target CT image; using a template generation model to generate a corresponding initial template according to the abnormal description text; extracting edge information of the target CT image to obtain target edge information, and performing similarity calculation based on the initial template and the target edge information to obtain a similar result corresponding to the target edge information; obtaining a target template corresponding to the abnormal description text from the target edge information according to the similar result; supervising the user annotation result according to the target template to obtain a target annotation supervision result corresponding to the target CT image; and performing data correction on the target CT image according to the target annotation supervision result to obtain a target annotation result. The problem of low annotation quality when annotating images in the related art is solved, and the overall quality of image annotation is further improved, thereby providing a more reliable data basis for various subsequent image processing applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0021] Figure 1 A schematic diagram of a process flow of an intelligent annotation method for medical CT images provided by an embodiment of the present invention;

[0022] Figure 2 A schematic diagram of the module structure of an intelligent annotation system for medical CT images provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0025] It should be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0026] The embodiment of the present invention provides a method and system for intelligent annotation of medical CT images. The method for intelligent annotation of medical CT images can be applied to a terminal device, which can be an electronic device such as a tablet computer, a laptop computer, a desktop computer, a personal digital assistant, and a wearable device. The terminal device can be a server or a server cluster.

[0027] Some embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0028] Please refer to Figure 1 , Figure 1A schematic diagram of a flow chart of an intelligent annotation method for medical CT images provided in an embodiment of the present invention.

[0029] like Figure 1 As shown, the intelligent annotation method for medical CT images includes steps S101 to S106.

[0030] Step S101: Obtain a user annotation result corresponding to a target CT image, and obtain an abnormality description text corresponding to the target CT image.

[0031] Exemplarily, a target CT image corresponding to a target patient and an abnormality description text of the target CT image by a target doctor corresponding to the target patient are obtained from a medical system. The abnormality description text is used to represent the target doctor's diagnostic information on the abnormality in the target CT image, and the corresponding user annotation result when the target user manually annotates the target CT image is obtained from a manual annotation system.

[0032] Step S102: Generate a corresponding initial template according to the abnormal description text using a template generation model.

[0033] Exemplarily, the template generation model includes CLIP (Contrastive Language-Image Pretraining) and a model based on a generative adversarial network for generating an initial template, thereby generating a corresponding initial template according to the abnormal description text according to the template generation model.

[0034] In some embodiments, the use of a template generation model to generate a corresponding initial template based on the exception description text includes: using an entity extraction network of the template generation model to perform keyword extraction on the exception description text to obtain target keywords; using a relationship recognition network of the template generation model to perform relationship analysis on the target keywords to obtain a target relationship corresponding to the target keywords; using a text generation network of the template generation model to generate a corresponding target text vector based on the target keywords and the target relationship; using a modal conversion network of the template generation model to obtain a target image vector corresponding to the exception description text based on the target text vector; and using an image generation network of the template generation model to generate the corresponding initial template based on the target image vector.

[0035] Exemplarily, the template generation model includes an entity extraction network, a relationship recognition network, a text generation network, a modality conversion network, and an image generation network.

[0036] For example, the anomaly description text is input into an entity extraction network, which aims to identify key entities in the anomaly description text. The entity extraction network analyzes the anomaly description text and extracts target keywords, such as anomaly name, anomaly size, anomaly location, anomaly shape, anomaly quantity, etc. The entity extraction network can be implemented through sequence labeling or other named entity recognition technologies.

[0037] Exemplarily, the extracted target keywords are input into the relationship recognition network. The relationship recognition network analyzes the contextual relationship between the target keywords to identify the relevance between them. The relationship recognition network returns the target relationship corresponding to the target keyword, such as the attribute relationship between the abnormal name and the abnormal size, the parallel relationship between the abnormal name and the abnormal name, etc., to provide a deeper understanding of the context.

[0038] For example, the target keywords and target relations are input into the text generation network. The text generation network will extract information from the context and generate a more structured target text vector. The text generation network uses embedding technology to combine the target keywords and target relations to form a high-dimensional feature vector to provide information for subsequent models.

[0039] Exemplarily, the generated target text vector is input into the modal conversion network. The purpose of the modal conversion network is to convert text information into image features for better image generation. The modal conversion network analyzes and generates a corresponding target image vector based on the target text vector, capturing the visual information conveyed by the text content. Finally, the target image vector is input into the image generation network. The image generation network generates the final initial template based on the target image vector. The image generation network may use techniques such as generative adversarial networks or variational autoencoders to generate images that conform to the characteristics of the input vector, ensuring that the image is clear and conforms to the description and structure of the input abnormal description text.

[0040] Step S103: extract edge information from the target CT image to obtain target edge information, and perform similarity calculation based on the initial template and the target edge information to obtain a similarity result corresponding to the target edge information.

[0041] Exemplarily, an appropriate edge detection algorithm is selected according to the characteristics of the target CT image, such as Canny edge detection, Sobel operator, Laplace operator, etc., so that the target CT image is subjected to edge detection according to the edge detection algorithm to obtain corresponding target edge information.

[0042] Exemplarily, the closed edge area corresponding to the target edge information is extracted to obtain the target closure information, and then the target closure information and the initial template are converted into the same coordinate system to obtain the first area corresponding to the target closure information and the second area corresponding to the initial template, thereby calculating the similarity between the first area and the second area according to the Jaccard similarity coefficient to obtain the similarity result between the first area and the second area.

[0043] In some embodiments, the extracting edge information of the target CT image to obtain the target edge information includes: processing the edge of the target CT image using inter-class variance to obtain the initial edge information corresponding to the target CT image; and performing edge correction on the initial edge information to obtain the target edge information.

[0044] Exemplarily, an image processing technique based on inter-class variance, such as the Otsu method, is used to extract initial edge information from the target CT image, and then edge detection operators such as Sobel, Prewitt, etc. are used to further refine the initial edge information extracted from the inter-class variance, and then morphological operations such as corrosion and expansion are performed on the initial edge information to remove noise points and enhance edge continuity. For example, for the extracted initial edge information, a suitable correction algorithm is selected to improve the accuracy of the edge, such as contour smoothing, edge enhancement, topology correction, etc., so that after the correction process, the improved initial edge information is integrated into the final target edge information.

[0045] In some embodiments, the edge processing of the target CT image using the inter-class variance to obtain the initial edge information corresponding to the target CT image includes: performing neighborhood mean calculation on the target CT image to obtain a target mean image corresponding to the target CT image; performing gradient calculation on the target CT image to obtain a target gradient image corresponding to the target CT image; determining a first probability corresponding to each first pixel in the target CT image according to the target mean image and the target gradient image; determining a second probability corresponding to each second pixel in the target mean image according to the target CT image and the target gradient image; determining a second probability corresponding to each second pixel in the target gradient image according to the target mean image and the target CT image. The method comprises the steps of: determining a third probability corresponding to three pixels; determining a first variance corresponding to the first pixel in the target CT image according to the first probability, and determining a first segmentation threshold corresponding to the target CT image according to the first variance; determining a second variance corresponding to the second pixel in the target mean image according to the second probability, and determining a second segmentation threshold corresponding to the target mean image according to the second variance; determining a third variance corresponding to the third pixel in the target gradient image according to the third probability, and determining a third segmentation threshold corresponding to the target gradient image according to the third variance; performing edge processing on the target CT image according to the first segmentation threshold, the second segmentation threshold and the third segmentation threshold to obtain the initial edge information.

[0046] Exemplarily, according to the target resolution and edge features of the target CT image, a suitable neighborhood size (such as a 3×3, 5×5 window) is determined. The choice of neighborhood size should ensure that local information can be captured without losing important details. The selected neighborhood window is then slid on the target CT image. The mean of the pixel values ​​is calculated for the pixel area covered by each window. Based on the mean of each window, a target mean image is generated. The value of each pixel in the target mean image corresponds to the average value of all pixels in the window in which it originally resides. The target mean image represents the average brightness or grayscale level of different areas of the image.

[0047] For example, a suitable gradient calculation method is selected according to the requirements. For example, the Sobel operator is used to calculate the gradient value of the target CT image in the horizontal direction and the vertical direction, so as to generate a target gradient image according to the calculated gradient value. The target gradient image highlights the edges and high contrast areas in the target CT image, and reflects the intensity changes of different areas in the target CT image.

[0048] Exemplarily, the target mean image, the target gradient image and the target CT image are established in a three-dimensional coordinate system, for example, the information of the target mean image is used as the x-axis, the information of the target gradient image is used as the y-axis and the information of the target CT image is used as the z-axis.

[0049] For example, a three-dimensional coordinate system is established and the coordinate axes are defined, and the X-axis represents the information in the target mean image. Specifically, the value on the X-axis can be the mean grayscale value in the target mean image. The Y-axis represents the information in the target gradient image. The value on the Y-axis is the gradient intensity of the corresponding pixel in the target gradient image. The Z-axis represents the information in the target CT image. The value on the Z-axis is the grayscale value of the corresponding pixel in the target CT image. For each pixel in the target CT image, its Z-axis value corresponds to the corresponding pixel value (X-axis value and Y-axis value) in the target mean image and the target gradient image. In this way, each pixel position will correspond to a point in the three-dimensional coordinate system, and the coordinates of the point are the values ​​of the pixel position in the three images. The position information of all pixels is filled into the three-dimensional coordinate system, thereby establishing a point cloud image containing all pixel information.

[0050] Exemplarily, the first pixel in the target CT image is selected, that is, the Z-axis value of the first pixel is fixed unchanged, that is, its Z coordinate in the three-dimensional coordinate system is kept unchanged. While keeping the Z-axis value unchanged, all points on the corresponding X-axis (information of the target mean image) and Y-axis (information of the target gradient image) are counted. The frequency of occurrence of these points on the XY plane is calculated, which is reflected in the target mean image and the target gradient image. When the first pixel in the target CT image is fixed, the frequency of occurrence of the target mean image information and the target gradient image information appears, thereby converting the statistically obtained frequency of occurrence into probability, that is, the frequency of occurrence of each corresponding pixel is used as the first probability corresponding to the pixel. Similarly, according to the above steps, the second probability corresponding to each second pixel in the target mean image and the third probability corresponding to each third pixel in the target gradient image are obtained.

[0051] Exemplarily, a first variance corresponding to a first pixel in a target CT image is determined using a first probability according to a maximum inter-class variance algorithm, thereby determining a first segmentation threshold corresponding to the target CT image according to a maximum value of the first variance. A second variance corresponding to a second pixel in a target mean image is determined using a second probability according to a maximum inter-class variance algorithm, thereby determining a second segmentation threshold corresponding to the target mean image according to a maximum value of the second variance. A third variance corresponding to a third pixel in a target gradient image is determined using a third probability according to a maximum inter-class variance algorithm, thereby determining a third segmentation threshold corresponding to the target gradient image according to a maximum value of the third variance.

[0052] Exemplarily, the target CT image is traversed, and when the pixel value in the target CT image is less than the first segmentation threshold, the pixel value of the target mean image at the same coordinates as the target CT image is less than the second segmentation threshold, and the pixel value of the target gradient image at the same coordinates as the target CT image is less than the third segmentation threshold, the pixel value at the corresponding coordinates in the target CT image is set to 0, otherwise the pixel value at the corresponding coordinates in the target CT image is set to 1, thereby obtaining the initial edge information corresponding to the target CT image.

[0053] Specifically, by combining the information of mean, gradient and pixel probability, we can more comprehensively understand and describe different regions in the image and improve the accuracy of target detection. By establishing different thresholds based on the variance of different regions, we can effectively achieve refined processing and better distinguish and identify different tissue structures when needed.

[0054] In some embodiments, the edge correction of the initial edge information to obtain the target edge information includes: determining a target window, extracting information from the target CT image according to the target window to obtain a window image corresponding to the target window; determining a first edge point, a second edge point, and a third edge point corresponding to the initial edge information from the window image; calculating a first approximation corresponding to a constant term of a first edge function corresponding to the first edge point and first error information corresponding to the first approximation; calculating a second approximation corresponding to a constant term of a second edge function corresponding to the second edge point and second error information corresponding to the second approximation; calculating a third approximation corresponding to a constant term of a third edge function corresponding to the third edge point and third error information corresponding to the third approximation; determining the first edge point, the second edge point, and the third edge point in the window image; The method comprises the steps of: determining a fourth approximation corresponding to the first approximation under the target edge function according to the relative position information, determining a fifth approximation corresponding to the second approximation under the target edge function according to the relative position information, and determining a sixth approximation corresponding to the third approximation under the target edge function according to the relative position information; determining a target slope and a target constant of the target edge function corresponding to the edge information in the window image according to the first error information, the second error information, the third error information, the fourth approximation, the fifth approximation and the sixth approximation; correcting the window image according to the target slope and the target constant to obtain a target correction result corresponding to the window image, and determining the target edge information according to the target correction result; wherein the target slope and the target constant are obtained according to the following formula:

[0055]

[0056] x1 =err 3 -1 +err 1 -1 +err 2 -1 ;

[0057] x 2 =err 3 -1 -err 1 -1 ;

[0058] x 3 =err 1 -1 *h 1 +err 2 -1 *h 2 +err 3 -1 *h 3 ;

[0059] x 4 =err 3 -1 +err 1 -1 ;

[0060] x 5 =err 3 -1 *h 3 -err 1 -1 *h 1 ;

[0061] Wherein, k represents the target slope, b represents the target constant, and x 1 、x 2 、x 3 、x 4 、x 5 Both represent intermediate variables, err 3 Indicates the third error information, err 2 Indicates the second error information, err 1 represents the first error information, h 1 represents the fourth approximation, h 2 represents the fifth approximation, h 3 represents the sixth approximation.

[0062] Exemplarily, to ensure that the target window includes at least 3 edge points, the target window can be a window of N rows and 3 columns, and then information is extracted from the target CT image according to the target window to obtain a window image corresponding to the target window; corresponding information is extracted from the initial edge information according to the window image, so as to determine whether the edge information corresponding to the window image contains 3 edge points, and if it does not include 3 edge points, the target window continues to slide in the target window until the edge information corresponding to the window image contains 3 edge points, and then the first edge point, the second edge point and the third edge point corresponding to the initial edge information are determined from the window image.

[0063] Exemplarily, a coordinate system is established with the first edge point as the coordinate origin, and the first edge function corresponding to the straight line across the first edge point, the second edge point and the third edge point is fitted, so as to obtain the constant term corresponding to the first edge function, that is, when the x-axis is 0, the y value corresponding to the first edge function is obtained, and then the first approximate value is obtained, and then the first error information is determined according to the fitting error corresponding to the fitting of the first edge function. Similarly, a coordinate system is established with the second edge point as the coordinate origin, and the second edge function corresponding to the straight line across the first edge point, the second edge point and the third edge point is fitted, so as to obtain the second approximate value and the second error information. And a coordinate system is established with the third edge point as the coordinate origin, and the third edge function corresponding to the straight line across the first edge point, the second edge point and the third edge point is fitted, so as to obtain the third approximate value and the third error information.

[0064] Exemplarily, according to the positions of the first edge point, the second edge point and the third edge point in the window image, the second edge point is taken as a reference point, and a first relative position of the first edge point relative to the second edge point is obtained, for example, the left position of the first edge point relative to the second edge point, and a second relative position of the third edge point relative to the second edge point is obtained, for example, the right upper position of the third edge point relative to the second edge point. Then, the relative position information is determined according to the first relative position and the second relative position.

[0065] Exemplarily, it is determined whether the first edge point, the second edge point and the third edge point are in the same row according to the relative position information. For example, when the first edge point is at the left side relative to the second edge point, the first edge point and the second edge point are in the same row. When the third edge point is at the upper right side relative to the second edge point, the third edge point and the second edge point are not in the same row. When the first edge point and the second edge point are in the same row, the first approximation can be directly determined as the fourth approximation corresponding to the target edge function; when the third edge point and the second edge point are not in the same row, the third approximation needs to be moved down, that is, it can be directly determined as the sixth approximation corresponding to the target edge function after subtracting 1. Since the second edge point is a reference point, the second approximation can be directly determined as the fifth approximation corresponding to the target edge function. That is, the first approximation point, the second approximation point and the third approximation point are transformed by judging whether they are in the same row, so that the first approximation point, the second approximation point and the third approximation point are on the same horizontal line, thereby providing support for the subsequent calculation of the target edge function corresponding to the first edge point, the second edge point and the third edge point.

[0066] Exemplarily, according to the first error information, the second error information, the third error information, and the fourth approximation, the fifth approximation, and the sixth approximation, the least squares principle is used to obtain the first edge function, the second edge function, and the third edge function to solve the average edge function. The corresponding error is minimized, so as to obtain the target slope and the target constant of the target edge function. The calculation formula is as follows:

[0067]

[0068] x 1 =err 3 -1 +err 1 -1 +err 2 -1 ;

[0069] x 2 =err 3 -1 -err 1 -1 ;

[0070] x 3 =err 1 -1 *h 1 +err 2 -1 *h 2 +err 3 -1 *h 3 ;

[0071] x 4 =err 3 -1 +err 1 -1 ;

[0072] x 5 =err 3 -1 *h 3 -err 1 -1 *h 1 ;

[0073] Where k represents the target slope, b represents the target constant, and x 1 、x 2 、x 3 、x 4 、x 5 Both represent intermediate variables, err 3 Indicates the third error information, err 2 Indicates the second error information, err 1 Indicates the first error information, h 1 represents the fourth approximation, h 2 represents the fifth approximation, h 3 Indicates a sixth approximation.

[0074] Exemplarily, according to the above formula, the first error information, the second error information, the third error information, and the fourth approximation, the fifth approximation, and the sixth approximation are substituted into the above formula to obtain the target slope and the target constant corresponding to the target edge function, so that the fitting straight lines corresponding to the first edge point, the second edge point, and the third edge point can be obtained according to the target slope and the target constant. Thus, the pixels in the fitting straight line in the window image are corrected to edge points using the fitting straight line corresponding to the target slope and the target constant, so as to obtain the target correction result corresponding to the window edge information of the window image.

[0075] Exemplarily, the initial edge information is traversed according to the target window to obtain all target correction results corresponding to the initial edge information, and then the initial edge information is corrected according to all target correction results to obtain the target edge information.

[0076] In some embodiments, performing similarity calculation based on the initial template and the target edge information to obtain a similar result corresponding to the target edge information includes: performing closed edge extraction on the target edge information to obtain first edge information; performing open edge extraction on the target edge information to obtain second edge information; determining a target range, and obtaining similar information corresponding to the second edge information in the target CT image based on the second edge information and the target range; performing data merging based on the similar information and the second edge information to obtain third edge information; performing similarity calculation based on the initial template and the first edge information to obtain a first similarity value; performing similarity calculation based on the initial template and the third edge information to obtain a second similarity value; and determining the similar result corresponding to the target edge information based on the first similarity value and the second similarity value.

[0077] Exemplarily, closed edges refer to edges that form closed contours, which are connected and enclose one or more regions. Based on pixel connectivity, check which edges are closed. This usually requires determining whether the edge points form a closed curve or contour, and then using a connected region marking algorithm to identify closed structures, thereby generating first edge information through the detected closed regions.

[0078] Exemplarily, open edges refer to edges that do not form closed contours, which are usually open and do not form closed areas. The distances between edge points are calculated to determine connectivity and edge opening positions, and the second edge information is generated by identifying non-closed features. For example, threshold segmentation can be used to clarify the boundaries and ensure that the extracted edges are not closed.

[0079] Exemplarily, a target range is set based on expert experience or historical experience, so that a pixel difference between a first pixel value corresponding to the second edge information in the target CT image and a second pixel value corresponding to the target range expanded outward from the second edge information is calculated within the target range; when the pixel difference is less than a preset difference, the pixel position corresponding to the corresponding second pixel value is also determined as an edge point, and then the pixel position corresponding to the second pixel value is determined as the similar information corresponding to the second edge information; finally, data merging is performed based on the similar information and the second edge information to obtain a closed edge, which is then determined as the third edge information.

[0080] Exemplarily, the first edge information and the initial template are converted to the same coordinate system to obtain a third region corresponding to the first edge information and a second region corresponding to the initial template, thereby performing similarity calculation on the third region and the second region according to the Jaccard similarity coefficient to obtain a first similarity value between the third region and the second region. And the third edge information and the initial template are converted to the same coordinate system to obtain a fourth region corresponding to the third edge information and a second region corresponding to the initial template, thereby performing similarity calculation on the fourth region and the second region according to the Jaccard similarity coefficient to obtain a second similarity value between the fourth region and the second region. Finally, the first similarity value and the second similarity value are both determined as similar results.

[0081] Specifically, by extracting the open edge in the target edge information to obtain the second edge information, and then closing the second edge information, it is possible to reduce the problem of the accuracy of the target CT image annotation caused by the subsequent abnormal matching with the initial template due to the presence of erroneous information in the target edge information, thereby providing support for the subsequent improvement of the accuracy of the target CT image annotation.

[0082] Step S104: Obtain a target template corresponding to the abnormal description text from the target edge information according to the similarity result.

[0083] Exemplarily, the edge region corresponding to the maximum similarity value in the similarity results is determined as the target template corresponding to the abnormal description text.

[0084] In some embodiments, obtaining the target template corresponding to the abnormal description text from the target edge information based on the similarity result includes: determining a preset value, and comparing the preset value with the similarity result to obtain a comparison result; obtaining the target template corresponding to the abnormal description text from the target edge information based on the comparison result.

[0085] Exemplarily, a preset value is determined based on expert experience or historical experience. When the similarity result is greater than or equal to the preset value, the comparison result is determined to satisfy the condition; when the similarity result is less than the preset value, the comparison result is determined to fail to satisfy the condition.

[0086] Exemplarily, when the comparison result satisfies the condition, the edge region corresponding to the similar result in the target edge information is determined as the target template corresponding to the abnormal description text, and the target template may be one or more. The target template is the edge region in the target edge information whose similarity result with the initial template is greater than or equal to a preset value.

[0087] Step S105: Supervise the user's annotation results according to the target template to obtain a target annotation supervision result corresponding to the target CT image.

[0088] Exemplarily, the target template is displayed in the target CT image to obtain the first area corresponding to the target template in the target CT image. And according to the second area corresponding to the user's annotation result in the target CT image, the deviation distance between the first area and the second area is calculated, and the deviation distance is compared with the preset distance to obtain the target annotation supervision result corresponding to the target CT image. Among them, the target template is obtained through the target edge information, and the target template and the target CT image are in the same coordinate system, so the target template can be directly added to the target CT image to obtain the first area.

[0089] For example, a preset distance is determined. When the deviation distance between the target template and the user annotation result is less than the preset distance, the target annotation supervision result corresponding to the target CT image is determined to be a normal annotation result; when the deviation distance between the target template and the user annotation result is greater than or equal to the preset distance, the target annotation supervision result corresponding to the target CT image is determined to be an abnormal annotation result.

[0090] Step S106: performing data correction on the target CT image according to the target labeling supervision result to obtain a target labeling result.

[0091] Exemplarily, when the target annotation supervision result is that the annotation result is normal, the user annotation result is directly determined as the target annotation result, which means that the annotation work of the target user is recognized and no further verification or modification is required. When the target annotation supervision result is that the annotation result is abnormal, the target template and the user annotation result are both displayed in the target CT image, so that another annotation user can re-annotate the data based on the target template and the user annotation result to obtain the corresponding target annotation result of another annotation user based on the target template and the user annotation result. That is, the new annotation user will undertake the task of re-annotating the data. At the same time, refer to the previous user annotation results and the target template to perform detailed and accurate data annotation. This process aims to correct or optimize the previous annotation to ensure the accuracy and reliability of the annotation results.

[0092] This dual verification and correction mechanism can greatly improve the quality of data annotation and provide more accurate training data for subsequent machine learning or deep learning models. At the same time, it also helps to improve the efficiency and accuracy of the entire data annotation process and provide strong data support for the development of artificial intelligence technology.

[0093] In some implementations, after obtaining the target annotation result, the method further includes: obtaining a target modification result of the target annotation result by the target user; and performing model correction on the template generation model according to the target modification result to obtain the corrected template generation model.

[0094] Exemplarily, the target modification results of the target user on the target annotation results are obtained from the annotation system, and then the template generation model is incrementally trained according to the target modification results and the exception description text, so as to achieve model correction of the template generation model, and then obtain the corrected template generation model, so that the template generation model can obtain a more accurate initial template in the subsequent initial template generation process, further improving the accuracy of data annotation.

[0095] In some embodiments, after obtaining the target labeling result, the method further includes: obtaining the number of targets corresponding to the target labeling result; when the number of targets is greater than a preset number, performing model training based on the target labeling result and the target CT image to obtain an abnormality recognition model.

[0096] For example, the target labeling results obtained are traversed to count the number of targets. According to actual needs and experience, a threshold representing a reasonable number of targets is set in advance, which is usually determined based on historical data and domain knowledge, so as to obtain a preset number.

[0097] Exemplarily, the calculated target quantity is compared with a preset quantity to determine whether it exceeds the preset quantity. If the target quantity is greater than the preset quantity, the subsequent model training step is triggered.

[0098] Exemplarily, select a suitable neural network architecture (such as a convolutional neural network, a deep learning model), and set the input and output layers of the model according to the task requirements. Ensure that the model can effectively process the target CT images and target annotation results and perform feature learning. Select a suitable training strategy (such as supervised learning, semi-supervised learning, etc.) according to specific needs. Set training parameters (such as learning rate, batch size, and number of training rounds, etc.). Use the target annotation results and target CT image data for model training. In the process, methods such as cross-validation can be used to evaluate the model performance, and the model parameters can be adjusted according to the loss function and evaluation indicators during the training process. After the training is completed, the generated anomaly recognition model is saved for subsequent use and deployment.

[0099] In addition, you can also use unseen test data sets to evaluate the performance of the trained anomaly recognition model and calculate key indicators such as recognition accuracy, recall rate, and F1 score to ensure the effectiveness of the anomaly recognition model. Based on the performance of the anomaly recognition model during the verification process, decide whether further data annotation is needed.

[0100] The embodiment of the present invention provides an intelligent annotation method and system for medical CT images, the method comprising: obtaining a target CT image and an abnormal description text corresponding to the target CT image; using a template generation model to generate a corresponding initial template according to the abnormal description text; extracting edge information of the target CT image to obtain target edge information, and performing similarity calculation based on the initial template and the target edge information to obtain a similarity result corresponding to the target edge information; obtaining a target template corresponding to the abnormal description text from the target edge information based on the similarity result; and annotating the target CT image data based on the target template to obtain a target annotation result. The problem of low annotation efficiency and low annotation quality when annotating images using existing models in the related art is solved, and the overall quality and efficiency of image annotation are further improved, thereby providing a more reliable data basis for various subsequent image processing applications.

[0101] See also Figure 2 , Figure 2 The present invention provides an intelligent annotation system 200 for medical CT images. The intelligent annotation system 200 for medical CT images includes a data acquisition module 201, a template generation module 202, a similarity calculation module 203, a template screening module 204, a data comparison module 205, and a data correction module 206. The data acquisition module 201 is used to obtain the user annotation result corresponding to the target CT image and obtain the abnormal description text corresponding to the target CT image; the template generation module 202 is used to generate the corresponding initial template according to the abnormal description text by using the template generation model; the similarity calculation module 203 is used to compare the target CT image with the user annotation result corresponding to the abnormal description text; the template generation module 202 is used to generate the corresponding initial template according to the abnormal description text by using the template generation model; the template screening module 204 is used to compare the target CT image with the user annotation result corresponding to the abnormal description text ... The target edge information is extracted from the target CT image, and the similarity is calculated based on the initial template and the target edge information to obtain a similarity result corresponding to the target edge information; a template screening module 204 is used to obtain a target template corresponding to the abnormal description text from the target edge information according to the similarity result; a data comparison module 205 is used to supervise the user annotation result according to the target template to obtain a target annotation supervision result corresponding to the target CT image; a data correction module 206 is used to correct the data of the target CT image according to the target annotation supervision result to obtain a target annotation result.

[0102] In some implementations, the template generation module 202, in the process of generating a corresponding initial template according to the abnormal description text using the template generation model, performs:

[0103] Using the entity extraction network of the template generation model to extract keywords from the abnormal description text to obtain target keywords;

[0104] Using the relationship recognition network of the template generation model to perform relationship analysis on the target keyword to obtain the target relationship corresponding to the target keyword;

[0105] Generate a corresponding target text vector according to the target keyword and the target relationship using the text generation network of the template generation model;

[0106] Obtaining a target image vector corresponding to the abnormal description text according to the target text vector using a modality conversion network of the template generation model;

[0107] The image generation network of the template generation model is used to generate the corresponding initial template according to the target image vector.

[0108] In some implementations, the similarity calculation module 203, in the process of extracting edge information from the target CT image to obtain target edge information, performs:

[0109] Processing the edge of the target CT image by using inter-class variance to obtain the initial edge information corresponding to the target CT image;

[0110] Perform edge correction on the initial edge information to obtain the target edge information.

[0111] In some implementations, the similarity calculation module 203 performs, during the process of processing the edge of the target CT image using the inter-class variance to obtain the initial edge information corresponding to the target CT image:

[0112] Performing neighborhood mean calculation on the target CT image to obtain a target mean image corresponding to the target CT image;

[0113] Performing gradient calculation on the target CT image to obtain a target gradient image corresponding to the target CT image;

[0114] Determine a first probability corresponding to each first pixel in the target CT image according to the target mean image and the target gradient image;

[0115] Determine a second probability corresponding to each second pixel in the target mean image according to the target CT image and the target gradient image;

[0116] Determine a third probability corresponding to each third pixel in the target gradient image according to the target mean image and the target CT image;

[0117] Determine a first variance corresponding to the first pixel in the target CT image according to the first probability, and determine a first segmentation threshold corresponding to the target CT image according to the first variance;

[0118] Determine a second variance corresponding to the second pixel in the target mean image according to the second probability, and determine a second segmentation threshold corresponding to the target mean image according to the second variance;

[0119] Determine a third difference corresponding to the third pixel in the target gradient image according to the third probability, and determine a third segmentation threshold corresponding to the target gradient image according to the third difference;

[0120] The target CT image is edge processed according to the first segmentation threshold, the second segmentation threshold, and the third segmentation threshold to obtain the initial edge information.

[0121] In some implementations, the similarity calculation module 203 performs, during the process of performing edge correction on the initial edge information to obtain the target edge information:

[0122] Determine a target window, and extract information from the target CT image according to the target window to obtain a window image corresponding to the target window;

[0123] Determine a first edge point, a second edge point, and a third edge point corresponding to the initial edge information from the window image;

[0124] Calculating a first approximate value corresponding to a constant term of a first edge function corresponding to the first edge point and first error information corresponding to the first approximate value;

[0125] Calculating a second approximate value corresponding to a constant term of a second edge function corresponding to the second edge point and second error information corresponding to the second approximate value;

[0126] Calculating a third approximate value corresponding to a constant term of a third edge function corresponding to the third edge point and third error information corresponding to the third approximate value;

[0127] Determine relative position information of the first edge point, the second edge point and the third edge point corresponding to the window image;

[0128] Determine a fourth approximation corresponding to the first approximation under the target edge function according to the relative position information, determine a fifth approximation corresponding to the second approximation under the target edge function according to the relative position information, and determine a sixth approximation corresponding to the third approximation under the target edge function according to the relative position information;

[0129] Determine a target slope and a target constant of the target edge function corresponding to the edge information in the window image according to the first error information, the second error information, the third error information, the fourth approximation, the fifth approximation, and the sixth approximation;

[0130] Correcting the window image according to the target slope and the target constant to obtain a target correction result corresponding to the window image, and determining the target edge information according to the target correction result;

[0131] Wherein, the target slope and the target constant are obtained according to the following formula:

[0132]

[0133] x 1 =err 3 -1 +err 1 -1 +err 2 -1 ;

[0134] x 2 =err 3 -1 -err 1 -1 ;

[0135] x 3 =err 1 -1 *h 1 +err 2 -1 *h 2 +err 3 -1 *h 3 ;

[0136] x 4 =err 3 -1 +err 1 -1 ;

[0137] x 5 =err 3 -1 *h 3 -err 1 -1 *h 1 ;

[0138] Wherein, k represents the target slope, b represents the target constant, and x 1 、x 2 、x 3 、x 4 、x 5 Both represent intermediate variables, err 3 Indicates the third error information, err 2Indicates the second error information, err 1 represents the first error information, h 1 represents the fourth approximation, h 2 represents the fifth approximation, h 3 represents the sixth approximation.

[0139] In some implementations, the similarity calculation module 203, in the process of performing similarity calculation based on the initial template and the target edge information to obtain a similarity result corresponding to the target edge information, executes:

[0140] Performing closed edge extraction on the target edge information to obtain first edge information;

[0141] Performing opening edge extraction on the target edge information to obtain second edge information;

[0142] determining a target range, and obtaining, in the target CT image, proximity information corresponding to the second edge information according to the second edge information and the target range;

[0143] Merge data according to the similar information and the second edge information to obtain third edge information;

[0144] Performing similarity calculation based on the initial template and the first edge information to obtain a first similarity value;

[0145] A similarity calculation is performed based on the initial template and the third edge information to obtain a second similarity value.

[0146] The similarity result corresponding to the target edge information is determined according to the first similarity value and the second similarity value.

[0147] In some implementations, the template screening module 204, in the process of obtaining the target template corresponding to the abnormal description text from the target edge information according to the similarity result, performs:

[0148] Determining a preset value, and comparing the preset value with the similar result to obtain a comparison result;

[0149] The target template corresponding to the abnormal description text is obtained from the target edge information according to the comparison result.

[0150] In some embodiments, after obtaining the target annotation result, the intelligent annotation system 200 for medical CT images further performs:

[0151] Obtaining a target modification result of the target marking result by a target user;

[0152] The template generation model is modified according to the target modification result to obtain the modified template generation model.

[0153] In some embodiments, after obtaining the target annotation result, the intelligent annotation system 200 for medical CT images further performs:

[0154] Obtaining the number of targets corresponding to the target labeling result;

[0155] When the number of targets is greater than a preset number, model training is performed according to the target labeling results and the target CT images to obtain an abnormality recognition model.

[0156] In some embodiments, the intelligent annotation system 200 for medical CT images may be applied to a terminal device.

[0157] It should be noted that technical personnel in the relevant field can clearly understand that, for the convenience and conciseness of description, the specific working process of the intelligent annotation system 200 for medical CT images described above can refer to the corresponding process in the aforementioned embodiment of the intelligent annotation method for medical CT images, and will not be repeated here.

[0158] An embodiment of the present invention also provides a storage medium for computer-readable storage, wherein the storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement any step of the intelligent annotation method for medical CT images provided in the description of the embodiment of the present invention.

[0159] The storage medium may be an internal storage unit of the terminal device described in the foregoing embodiment, such as a hard disk or memory of the terminal device. The storage medium may also be an external storage device of the terminal device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc., equipped on the terminal device.

[0160] It will be appreciated by those skilled in the art that all or some of the steps, systems, and functional modules / units in the methods disclosed above may be implemented as software, firmware, hardware, and appropriate combinations thereof. In a hardware embodiment, the division between the functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed by several physical components in cooperation. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or non-transitory medium) and a communication medium (or temporary medium). As known to those skilled in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer. In addition, it is well known to those of ordinary skill in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0161] It should be understood that the term "and / or" used in the present specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, including these combinations. It should be noted that, in this article, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "including a..." does not exclude the presence of other identical elements in the process, method, article or system including the element.

[0162] The serial numbers of the embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. An intelligent annotation method for medical CT images, characterized in that: The method comprises: Obtaining a user annotation result corresponding to a target CT image, and obtaining an abnormality description text corresponding to the target CT image; Generate a corresponding initial template according to the abnormal description text using a template generation model; Extracting edge information of the target CT image to obtain target edge information, and performing similarity calculation based on the initial template and the target edge information to obtain a similarity result corresponding to the target edge information; Obtaining a target template corresponding to the abnormal description text from the target edge information according to the similarity result; Performing labeling supervision on the user labeling result according to the target template to obtain a target labeling supervision result corresponding to the target CT image; The target CT image is data corrected according to the target labeling supervision result to obtain a target labeling result.

2. The method according to claim 1, characterized in that: The using of the template generation model to generate a corresponding initial template according to the abnormal description text includes: Using the entity extraction network of the template generation model to extract keywords from the abnormal description text to obtain target keywords; Using the relationship recognition network of the template generation model to perform relationship analysis on the target keyword to obtain the target relationship corresponding to the target keyword; Generate a corresponding target text vector according to the target keyword and the target relationship using the text generation network of the template generation model; Obtaining a target image vector corresponding to the abnormal description text according to the target text vector using a modality conversion network of the template generation model; The image generation network of the template generation model is used to generate the corresponding initial template according to the target image vector.

3. The method according to claim 1, characterized in that The step of extracting edge information from the target CT image to obtain target edge information includes: Processing the edge of the target CT image by using inter-class variance to obtain the initial edge information corresponding to the target CT image; Perform edge correction on the initial edge information to obtain the target edge information.

4. The method according to claim 3, characterized in that The step of processing the edge of the target CT image by using the inter-class variance to obtain the initial edge information corresponding to the target CT image includes: Performing neighborhood mean calculation on the target CT image to obtain a target mean image corresponding to the target CT image; Performing gradient calculation on the target CT image to obtain a target gradient image corresponding to the target CT image; Determine a first probability corresponding to each first pixel in the target CT image according to the target mean image and the target gradient image; Determine a second probability corresponding to each second pixel in the target mean image according to the target CT image and the target gradient image; Determine a third probability corresponding to each third pixel in the target gradient image according to the target mean image and the target CT image; Determine a first variance corresponding to the first pixel in the target CT image according to the first probability, and determine a first segmentation threshold corresponding to the target CT image according to the first variance; Determine a second variance corresponding to the second pixel in the target mean image according to the second probability, and determine a second segmentation threshold corresponding to the target mean image according to the second variance; Determine a third difference corresponding to the third pixel in the target gradient image according to the third probability, and determine a third segmentation threshold corresponding to the target gradient image according to the third difference; The target CT image is edge processed according to the first segmentation threshold, the second segmentation threshold, and the third segmentation threshold to obtain the initial edge information.

5. The method according to claim 3, characterized in that: The performing edge correction on the initial edge information to obtain the target edge information includes: Determine a target window, and extract information from the target CT image according to the target window to obtain a window image corresponding to the target window; Determine a first edge point, a second edge point, and a third edge point corresponding to the initial edge information from the window image; Calculating a first approximate value corresponding to a constant term of a first edge function corresponding to the first edge point and first error information corresponding to the first approximate value; Calculating a second approximate value corresponding to a constant term of a second edge function corresponding to the second edge point and second error information corresponding to the second approximate value; Calculating a third approximate value corresponding to a constant term of a third edge function corresponding to the third edge point and third error information corresponding to the third approximate value; Determine relative position information of the first edge point, the second edge point and the third edge point corresponding to the window image; Determine a fourth approximation corresponding to the first approximation under the target edge function according to the relative position information, determine a fifth approximation corresponding to the second approximation under the target edge function according to the relative position information, and determine a sixth approximation corresponding to the third approximation under the target edge function according to the relative position information; Determine a target slope and a target constant of the target edge function corresponding to the edge information in the window image according to the first error information, the second error information, the third error information, the fourth approximation, the fifth approximation, and the sixth approximation; Correcting the window image according to the target slope and the target constant to obtain a target correction result corresponding to the window image, and determining the target edge information according to the target correction result; Wherein, the target slope and the target constant are obtained according to the following formula: x1=err3 -1 +err1 -1 +err2 -1 ; x2=err3 -1 -err1 -1 ; <h2 style=";text-align:left;direction:ltr">x3=err1<h2 style=";text-align:left;direction:ltr"> -1 <h2 style=";text-align:left;direction:ltr"> *h1+err2<h2 style=";text-align:left;direction:ltr"> -1 <h2 style=";text-align:left;direction:ltr"> *h2+err3<h2 style=";text-align:left;direction:ltr"> -1 <h2 style=";text-align:left;direction:ltr"> *h3; x4=err3 -1 +err1 -1 ; <h2 style=";text-align:left;direction:ltr">x5=err3<h2 style=";text-align:left;direction:ltr"> -1 <h2 style=";text-align:left;direction:ltr"> *h3-err1<h2 style=";text-align:left;direction:ltr"> -1 <h2 style=";text-align:left;direction:ltr"> *h1; Among them, k represents the target slope, b represents the target constant, x1, x2, x3, x4, and x5 all represent intermediate variables, err3 represents the third error information, err2 represents the second error information, err1 represents the first error information, h1 represents the fourth approximate value, h2 represents the fifth approximate value, and h3 represents the sixth approximate value.

6. The method according to claim 1, characterized in that The performing similarity calculation based on the initial template and the target edge information to obtain a similarity result corresponding to the target edge information includes: Performing closed edge extraction on the target edge information to obtain first edge information; Performing opening edge extraction on the target edge information to obtain second edge information; determining a target range, and obtaining, in the target CT image, proximity information corresponding to the second edge information according to the second edge information and the target range; Merge data according to the similar information and the second edge information to obtain third edge information; Performing similarity calculation based on the initial template and the first edge information to obtain a first similarity value; A similarity calculation is performed based on the initial template and the third edge information to obtain a second similarity value. The similarity result corresponding to the target edge information is determined according to the first similarity value and the second similarity value.

7. The method according to claim 1, characterized in that The step of obtaining a target template corresponding to the abnormal description text from the target edge information according to the similarity result includes: Determining a preset value, and comparing the preset value with the similar result to obtain a comparison result; The target template corresponding to the abnormal description text is obtained from the target edge information according to the comparison result.

8. The method according to any one of claims 1 to 7, characterized in that After obtaining the target labeling result, the method further includes: Obtaining a target modification result of the target marking result by a target user; The template generation model is modified according to the target modification result to obtain the modified template generation model.

9. The method according to any one of claims 1 to 7, characterized in that: After obtaining the target labeling result, the method further includes: Obtaining the number of targets corresponding to the target labeling result; When the number of targets is greater than a preset number, model training is performed according to the target labeling results and the target CT images to obtain an abnormality recognition model.

10. An intelligent annotation system for medical CT images, characterized in that: include: A data acquisition module, used to obtain the user annotation result corresponding to the target CT image, and obtain the abnormality description text corresponding to the target CT image; A template generation module, used to generate a corresponding initial template according to the abnormal description text using a template generation model; A similarity calculation module is used to extract edge information of the target CT image to obtain target edge information, and perform similarity calculation based on the initial template and the target edge information to obtain a similarity result corresponding to the target edge information; A template screening module, used for obtaining a target template corresponding to the abnormal description text from the target edge information according to the similarity result; A data comparison module, used to supervise the user's annotation results according to the target template, and obtain a target annotation supervision result corresponding to the target CT image; The data correction module is used to perform data correction on the target CT image according to the target labeling supervision result to obtain the target labeling result.

Images