Data augmentation system and method for small sample medical image labeling

Abstract

The present application relates to the field of data enhancement, and more particularly to a data enhancement system and method for small sample medical image labeling, which comprises a lesion analysis module, a data processing module, a labeling interference module, and a data enhancement module, determines a plurality of organs affected by diseases and performs lesion analysis, calculates lesion occurrence characteristic values, determines lesion analysis organs, determines texture discrete characteristic values and organ covering characteristic values, determines organ adhesion areas based on the texture discrete characteristic values, analyzes the edge texture direction of the organ adhesion areas, obtains a reference path belt, determines an interference organ area, calculates a labeling interference coefficient, calls a data enhancement model to obtain an enhanced organ area, and completes medical image picture labeling of the enhanced organ area. The present application divides the adhesion area of organs that may be affected by diseases, determines the interference organ area, calls the data enhancement model according to the labeling interference coefficient, completes data labeling, and improves the accuracy and efficiency of data labeling.

Description

Technical Field

[0001] This invention relates to the field of data augmentation, and more particularly to a data augmentation system and method for small-sample medical image annotation. Background Technology

[0002] In the field of medical image analysis, deep learning-based models have become core tools for lesion detection and disease diagnosis. However, their performance depends on large-scale, high-quality labeled data. Medical image labeling faces challenges such as data acquisition being constrained by privacy regulations and data scarcity due to the natural scarcity of rare disease cases, collectively creating a "small sample" scenario. At the same time, traditional data augmentation methods in the field of general computer vision cannot adapt to the medical specificity of medical images, easily destroying the rationality of anatomical structures and obscuring key features of lesions, failing to meet the needs of diverse clinical scenarios. With the evolution of technology, data augmentation has evolved from early use of traditional methods and manual constraints, to mid-term integration of medical prior knowledge to guided augmentation, and now to the mainstream of generative augmentation based on deep learning such as GANs and Diffusion Models, which learn the medical feature distribution of a small amount of labeled data to generate new samples that conform to real medical logic. This has given rise to data augmentation methods for small sample medical image labeling.

[0003] Chinese Patent Publication No. CN116092643A discloses an interactive semi-automatic annotation method based on medical images, comprising: Step 1. Collecting medical image data with lesions and classifying the medical image data into different lesion categories; Step 2. Selecting and training different Deep Grabcut algorithm pre-trained models for different lesion categories; Step 3. Annotating the bounding boxes of regions of interest for the medical image data of different lesion categories using the pre-trained models trained in Step 2; Step 4. Performing data augmentation on the annotated medical image data; Step 5. Performing interactive model training on the augmented data and network parameters, and validating the model after training. When the Dice of the interactive model is greater than a threshold, proceed to Step 6; Step 6. Selecting an image segmentation network to train on the assisted annotated images, saving the best image segmentation network training model on the validation set and the last image segmentation network training model, evaluating the model performance, and determining whether it can initially enter the practical application stage.

[0004] Chinese Patent Publication No. CN117372400A discloses an automatic annotation and acquisition system based on medical enhanced images, including an image import module, a parameter input module, a ROI generation module, and a data generation module. This invention can automatically annotate and acquire data of blood vessels, corresponding soft tissues, or other anatomical parts in transverse images of medical enhanced images. By automatically segmenting the transverse images of medical enhanced images and adaptively adjusting the window width and window level, it can quickly and accurately locate target blood vessels, soft tissues, or other anatomical parts, realize automatic annotation and parameter selection of target blood vessels, soft tissues, or other anatomical parts, adaptive adjustment of the ROI size of target blood vessels, soft tissues, or other anatomical parts, and automatic summarization and export of target data.

[0005] It is evident that the existing technology still has the following problems:

[0006] In medical image data processing, most of the methods used are global and non-selective enhancement strategies. While enhancing the effective target area, these strategies also amplify interference signals in the image. Especially in small sample scenarios, this indiscriminate enhancement operation can lead to blurred annotation boundaries, decreased confidence, and reduced accuracy and reliability of annotation. Summary of the Invention

[0007] To address this, the present invention provides a data augmentation system for small-sample medical image annotation, which overcomes the problem that most medical image data processing adopts global and non-selective augmentation strategies. While enhancing the effective target area, it also couples and amplifies interference signals in the image. Especially in small-sample scenarios, this indiscriminate augmentation operation can lead to blurred annotation boundaries, decreased confidence, and reduced accuracy and reliability of annotation.

[0008] To achieve the above objectives, the present invention provides a data augmentation system for small-sample medical image annotation, comprising,

[0009] The lesion analysis module is used to identify several organs affected by the disease based on the disease description corresponding to the medical image, perform lesion analysis on each of the disease-affected organs, calculate the lesion occurrence characterization value of each disease-affected organ, and determine the lesion analysis organ.

[0010] A data processing module, which is connected to the lesion analysis module, is used to obtain the edge texture of the lesion analysis organ to determine the texture discrete characterization value, obtain the occlusion point of the lesion analysis organ, and determine the organ occlusion characterization value.

[0011] An interference annotation module, which is connected to the data processing module, is used to determine the organ adhesion region based on the texture discrete characterization value, analyze the edge texture direction of the organ adhesion region to obtain a reference path band, thereby determining the interference organ region, and for the interference organ region, calculate the annotation interference coefficient based on the texture discrete characterization value and the organ occlusion characterization value to determine whether to call the data augmentation model.

[0012] The data augmentation module, which is connected to the annotation interference module, responds to the calling state of the data augmentation model, augments the data on the reference path band, obtains the augmented organ region, and completes the annotation of the medical image of the augmented organ region.

[0013] Furthermore, the lesion analysis module identifies several organs affected by the disease, including,

[0014] Used to determine several disease keywords based on the description of the disease;

[0015] This is used to traverse the disease database based on each of the disease keywords and determine the organs corresponding to each of the disease keywords;

[0016] Used to identify the organs mentioned as those affected by the disease.

[0017] Furthermore, the lesion analysis module calculates the lesion occurrence characterization values ​​of each disease-affected organ, and determines the lesion analysis organs, including,

[0018] The historical incidence rate of each organ affected by the disease is used as a characterization value for the occurrence of the disease.

[0019] This is used to sort the characteristic values ​​of each lesion in descending order;

[0020] The organ affected by the disease corresponding to the highest lesion occurrence characteristic value is used as the lesion analysis organ.

[0021] Further, the data processing module determines the discrete texture representation values, including,

[0022] Used to obtain the edge texture of the organ being analyzed for the lesion;

[0023] The discreteness of the edge texture is used to determine the texture discreteness characterization value.

[0024] Furthermore, the data processing module determines organ occlusion characterization values, including,

[0025] Used to obtain the number of pixels of the occluded point of the organ being analyzed for lesions;

[0026] The ratio of the number of pixels at the occlusion point to the total number of pixels in the organ analyzed for lesion is used to determine the organ occlusion characterization value;

[0027] The obscured point is a point that is different in color from the organ being analyzed for lesions.

[0028] Furthermore, the interference annotation module identifies organ adhesion regions, wherein,

[0029] If the texture discrete characterization value is greater than the texture discrete characterization value threshold, then the edge region of the lesion analysis organ is determined to be the organ adhesion region.

[0030] If the texture discrete characterization value is less than or equal to the texture discrete characterization value threshold, then the edge region of the lesion analysis organ is determined to be a normal region.

[0031] The edge region of the lesion analysis organ refers to the region formed by the edge texture of the lesion analysis organ.

[0032] Furthermore, the interference annotation module identifies the interfering organ region, including:

[0033] Used to obtain all pixels in the image of the organ adhesion region and determine the maximum response amplitude of each pixel;

[0034] The filter direction corresponding to the maximum response amplitude is determined as the dominant texture direction of each pixel to generate a global orientation field map;

[0035] Used to record candidate seed points on the orientation field map;

[0036] The path band formed by the candidate seed points along the dominant texture direction is used to determine the reference path band;

[0037] If a reference path band can be identified, the region formed by the reference path band is determined to be the interfering organ region;

[0038] If the reference path band cannot be identified, the region of the interfering organ is uncertain;

[0039] Among them, the candidate seed point is the pixel point whose maximum response amplitude is greater than the response amplitude threshold.

[0040] Furthermore, the annotation interference module calculates the annotation interference coefficient, including,

[0041] The ratio of the texture discrete representation value to the baseline texture discrete representation value is used to determine the texture influence factor;

[0042] The ratio of the organ coverage characterization value to the baseline organ coverage characterization value is used to determine the coverage influence factor;

[0043] The weighted sum of the texture influence factor and the occlusion influence factor is used to determine the annotation interference coefficient.

[0044] Furthermore, the interference annotation module determines whether to invoke the data augmentation model, wherein,

[0045] If the labeled interference coefficient is greater than the labeled interference coefficient threshold, then the data augmentation model is invoked to perform data augmentation on the medical image.

[0046] If the annotation interference coefficient is less than or equal to the annotation interference coefficient threshold, then the data augmentation model will not be invoked, and the medical image will be annotated normally.

[0047] Furthermore, methods for data augmentation systems for small-sample medical image annotation include:

[0048] Obtain the disease description corresponding to the medical image, identify several organs affected by the disease based on the disease description, perform lesion analysis on each of the organs affected by the disease, calculate the lesion occurrence characterization value of each organ affected by the disease, and determine the organs for lesion analysis.

[0049] Obtain the edge texture of the lesion analysis organ to determine the texture discrete characterization value, obtain the occlusion point of the lesion analysis organ, and determine the organ occlusion characterization value;

[0050] Based on the texture discrete characterization value, the organ adhesion region is determined, the edge texture direction of the organ adhesion region is analyzed to obtain the reference path band, so as to determine the interfering organ region. For the interfering organ region, the labeled interference coefficient is calculated based on the texture discrete characterization value and the organ occlusion characterization value to determine whether to call the data augmentation model.

[0051] In response to the call status of the data augmentation model, the data on the reference path band is augmented to obtain the augmented organ region, and the medical image annotation of the augmented organ region is completed.

[0052] Compared with existing technologies, this invention, by setting up a lesion analysis module, a data processing module, an interference annotation module, and a data augmentation module, identifies several organs affected by disease and performs lesion analysis, calculates lesion occurrence characterization values, determines the organs for lesion analysis, determines texture discrete characterization values ​​and organ occlusion characterization values, determines organ adhesion regions based on the texture discrete characterization values, analyzes the edge texture direction of the organ adhesion regions to obtain reference path bands, determines interfering organ regions, calculates annotation interference coefficients, and calls a data augmentation model to obtain enhanced organ regions, thus completing the annotation of medical images of the enhanced organ regions. This invention targets organs that may develop lesions, divides adhesion regions, determines interfering organ regions, and calls a data augmentation model based on the annotation interference coefficients to complete data annotation, thereby improving the accuracy and efficiency of data annotation.

[0053] In particular, by performing lesion analysis in advance to identify the organs involved in the lesion analysis, a data foundation is provided for subsequent targeted use of data augmentation models to achieve regional data augmentation. In practice, the entire image is usually processed globally to identify all visible organs and their global features. Although this method can achieve a comprehensive scan, it does not distinguish between the pathological and visual significance of organs, and allocates equal computing resources to organs and regions without obvious lesions, resulting in a large amount of computing power being wasted on non-critical areas. Analyzing the entire region of organs increases the complexity of feature extraction and computation, reducing the overall processing efficiency of the system. At the same time, the scattered attention may weaken the model's ability to represent truly critical lesion areas, affecting the accuracy and robustness of subsequent tasks. Based on this, this invention performs lesion analysis, prioritizing the identification and screening of clinically significant lesion organs, thereby concentrating computing resources on these key targets. This lays a data foundation for subsequent data augmentation and refined annotation targeting specific regions, improves the utilization efficiency of computing resources and the targeting of the analysis process, and enhances the accuracy and efficiency of data annotation.

[0054] In particular, by determining the discrete texture representation value and the organ occlusion representation value, a data foundation is provided for subsequent calculation of the annotation interference coefficient, and the data that needs data augmentation is identified. In reality, different organs present significantly different morphologies in images, and the distribution characteristics of their edge textures and the clarity of organ occlusion exhibit high individual variability and complexity. For some organs, due to the obvious discrete distribution characteristics of their edge textures, the regularity and integrity of the texture structure are weak, making it difficult for traditional edge detection algorithms to effectively capture continuous and clear organ boundary contours. This discreteness of edge textures directly leads to ambiguity in organ localization, causing deviations in the data augmentation range determined based on boundary information, thus affecting the data augmentation effect and making it impossible to achieve targeted optimization of key areas. When organs with discrete edge textures are simultaneously occluded, the annotation difficulty will be exacerbated. Not only will the discreteness of edge textures hinder boundary recognition, but the feature loss caused by the occlusion area will further reduce the accuracy and completeness of organ annotation. Based on this, this invention considers calculating the annotation interference coefficient based on the discrete texture representation value and the organ occlusion representation value, and specifically determines the areas that need data augmentation to improve the accuracy and efficiency of data annotation.

[0055] In particular, by analyzing the discrete texture representation values ​​and edge texture direction, a reference path band is determined to obtain the interfering organ region. This provides a data basis for the subsequent calculation of the interference coefficient for this region. In reality, organ edge textures often exhibit ambiguity or structural occlusion due to anatomical adhesions with adjacent tissues or organs, making it difficult for the model to accurately identify and segment the target region. If conventional non-selective data augmentation strategies are still used in such cases, the augmentation operation will simultaneously amplify the confusing features of the adhered region and the noise signal introduced by the occlusion, causing further distortion of the organ boundary semantics, reducing the model's ability to discriminate the real structure, and leading to data annotation errors. Based on this, this invention determines the interfering organ region by identifying the reference path band, providing a data basis for further calculation of the interference coefficient for this region, providing a theoretical foundation for the subsequent determination of the data augmentation model to be called, and improving the accuracy and efficiency of data annotation. Attached Figure Description

[0056] Figure 1 A schematic diagram of the data augmentation system for small-sample medical image annotation according to an embodiment of the invention;

[0057] Figure 2 This is a logic block diagram for determining the organ adhesion region according to an embodiment of the invention;

[0058] Figure 3 This is a logic block diagram for determining the interfering organ region according to an embodiment of the invention;

[0059] Figure 4 A logic block diagram illustrating whether to invoke the data augmentation model in an embodiment of the invention;

[0060] Figure 5 This is a schematic diagram illustrating the steps of a data augmentation method for small-sample medical image annotation, as described in an embodiment of the invention. Detailed Implementation

[0061] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0062] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0063] It should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the term "connection" should be interpreted broadly. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0064] Please see Figure 1 , Figure 1 This is a schematic diagram of the structure of a data augmentation system for small-sample medical image annotation according to an embodiment of the invention. The data augmentation system for small-sample medical image annotation of the present invention includes:

[0065] The lesion analysis module is used to identify several organs affected by the disease based on the disease description corresponding to the medical image, perform lesion analysis on each of the disease-affected organs, calculate the lesion occurrence characterization value of each disease-affected organ, and determine the lesion analysis organ.

[0066] A data processing module, which is connected to the lesion analysis module, is used to obtain the edge texture of the lesion analysis organ to determine the texture discrete characterization value, obtain the occlusion point of the lesion analysis organ, and determine the organ occlusion characterization value.

[0067] An interference annotation module, which is connected to the data processing module, is used to determine the organ adhesion region based on the texture discrete characterization value, analyze the edge texture direction of the organ adhesion region to obtain a reference path band, thereby determining the interference organ region, and for the interference organ region, calculate the annotation interference coefficient based on the texture discrete characterization value and the organ occlusion characterization value to determine whether to call the data augmentation model.

[0068] The data augmentation module, which is connected to the annotation interference module, responds to the calling state of the data augmentation model, augments the data on the reference path band, obtains the augmented organ region, and completes the annotation of the medical image of the augmented organ region.

[0069] Specifically, there are no restrictions on the source of medical images. For example, they can be public medical databases, anonymized data provided by clinical partner institutions, and other publicly available or authorized image resources that comply with ethical and legal regulations. The key point is that all image data must be obtained and used in accordance with privacy protection and data security regulations, such as obtaining informed consent from patients and authorization from institutions. This will not be elaborated further.

[0070] Understandably, there are many mature computer-aided diagnostic systems in the existing technology that can perform pathological identification on medical images and then generate a text describing the condition. By performing natural language processing and key information extraction on this text, several organs associated with the current lesion can be further identified, i.e., the organs affected by the condition.

[0071] Specifically, the lesion analysis module identifies several organs affected by the disease, including,

[0072] Used to determine several disease keywords based on the description of the disease;

[0073] This is used to traverse the disease database based on each of the disease keywords and determine the organs corresponding to each of the disease keywords;

[0074] Used to identify the organs mentioned as those affected by the disease.

[0075] Specifically, the disease database can be an existing database or a database constructed by a person skilled in the art by collecting a number of data. The disease database includes a number of disease keywords and a number of organs corresponding to the disease keywords. For example, if the disease keyword is "cirrhosis", the corresponding organ is "liver"; if the disease keyword is "systemic lupus erythematosus", the corresponding organs are "skin, joints, kidneys, blood system, nervous system, serous membranes, heart and lungs".

[0076] Specifically, the lesion analysis module calculates the lesion manifestation values ​​of each disease-affected organ, and identifies the organs for lesion analysis, including:

[0077] The historical incidence rate of each organ affected by the disease is used as a characterization value for the occurrence of the disease.

[0078] This is used to sort the characteristic values ​​of each lesion in descending order;

[0079] The organ affected by the disease corresponding to the highest lesion occurrence characteristic value is used as the lesion analysis organ.

[0080] Specifically, there are no restrictions on the method for determining historical incidence rates. For example, it can be done by searching and extracting structured information from public health databases, government-run public health statistics platforms, or peer-reviewed medical research websites. Alternatively, proprietary datasets provided by researchers or medical institutions in the field that have undergone data licensing procedures can be used, which will not be elaborated further here.

[0081] Specifically, by performing lesion analysis in advance to identify the organs to be analyzed, a data foundation is provided for subsequent targeted data augmentation models to achieve regional data augmentation. In practice, the entire image is usually processed globally to identify all visible organs and their global features. Although this method can achieve a comprehensive scan, it does not distinguish between the pathological and visual significance of organs, and allocates equal computing resources to organs and regions without obvious lesions, resulting in a large amount of computing power being wasted on non-critical areas. Analyzing the entire region of organs increases the complexity of feature extraction and computation, reducing the overall processing efficiency of the system. At the same time, the scattered attention may weaken the model's ability to represent truly critical lesion areas, affecting the accuracy and robustness of subsequent tasks. Based on this, this invention performs lesion analysis, prioritizing the identification and screening of clinically significant lesion organs, thereby concentrating computing resources on these key targets. This lays the data foundation for subsequent data augmentation and refined annotation targeting specific regions, improves the utilization efficiency of computing resources and the targeting of the analysis process, and improves the accuracy and efficiency of data annotation.

[0082] Specifically, the data processing module determines the discrete texture representation values, including,

[0083] Used to obtain the edge texture of the organ being analyzed for the lesion;

[0084] The discreteness of the edge texture is used to determine the texture discreteness characterization value.

[0085] Specifically, the data processing module determines organ coverage characterization values, including,

[0086] Used to obtain the number of pixels of the occluded point of the organ being analyzed for lesions;

[0087] The ratio of the number of pixels at the occlusion point to the total number of pixels in the organ analyzed for lesion is used to determine the organ occlusion characterization value;

[0088] The obscured point is a point that is different in color from the organ being analyzed for lesions.

[0089] Specifically, by determining the discrete texture representation value and the organ occlusion representation value, a data foundation is provided for subsequent calculation of the annotation interference coefficient, and the data that needs data augmentation is identified. In reality, different organs present significantly different morphologies in images, and the distribution characteristics of their edge textures and the clarity of organ occlusion exhibit high individual variability and complexity. For some organs, due to the obvious discrete distribution characteristics of their edge textures, the regularity and integrity of the texture structure are weak, making it difficult for traditional edge detection algorithms to effectively capture continuous and clear organ boundary contours. This discreteness of edge textures directly leads to ambiguity in organ localization, causing deviations in the data augmentation range determined based on boundary information, thus affecting the data augmentation effect and making it impossible to achieve targeted optimization of key areas. When organs with discrete edge textures are simultaneously occluded, the annotation difficulty will be exacerbated. Not only will the discreteness of edge textures hinder boundary recognition, but the feature loss caused by the occlusion area will further reduce the accuracy and completeness of organ annotation. Based on this, this invention considers calculating the annotation interference coefficient based on the discrete texture representation value and the organ occlusion representation value, and specifically determines the areas that need data augmentation to improve the accuracy and efficiency of data annotation.

[0090] Please see Figure 2 , Figure 2 This is a logic block diagram for determining organ adhesion regions according to an embodiment of the invention. Specifically, an interference annotation module determines organ adhesion regions, wherein...

[0091] If the texture discrete characterization value is greater than the texture discrete characterization value threshold, then the edge region of the lesion analysis organ is determined to be the organ adhesion region.

[0092] If the texture discrete characterization value is less than or equal to the texture discrete characterization value threshold, then the edge region of the lesion analysis organ is determined to be a normal region.

[0093] The edge region of the lesion analysis organ refers to the region formed by the edge texture of the lesion analysis organ.

[0094] Specifically, the texture discrete characterization value threshold represents a boundary where the diseased organ adheres to surrounding organs. It is pre-calculated by obtaining texture discrete characterization values ​​corresponding to adhesion of several organs in advance. The product of each texture discrete characterization value and the discrete precision is determined as the texture discrete characterization value threshold. The discrete precision is selected within the range [0.8, 1.0]. In practice, to improve the accuracy of the analysis, the discrete precision is set to 0.9.

[0095] Please see Figure 3 , Figure 3 This is a logic block diagram illustrating the determination of interfering organ regions according to an embodiment of the invention. Specifically, the interference labeling module determines the interfering organ regions, including:

[0096] Used to obtain all pixels in the image of the organ adhesion region and determine the maximum response amplitude of each pixel;

[0097] The filter direction corresponding to the maximum response amplitude is determined as the dominant texture direction of each pixel to generate a global orientation field map;

[0098] Used to record candidate seed points on the orientation field map;

[0099] The path band formed by the candidate seed points along the dominant texture direction is used to determine the reference path band;

[0100] If a reference path band can be identified, the region formed by the reference path band is determined to be the interfering organ region;

[0101] If the reference path band cannot be identified, the region of the interfering organ is uncertain;

[0102] Among them, the candidate seed point is the pixel point whose maximum response amplitude is greater than the response amplitude threshold.

[0103] Specifically, the reference path band refers to a strip of pixels located within the organ adhesion region, which is most likely to represent the correct organ boundary and is inferred based on orientation field analysis.

[0104] Specifically, the maximum response amplitude is determined by a pre-designed filter bank. The input image of the organ adhesion region is convolved with the filter in each specific direction of the filter bank to obtain a set of texture response maps in different directions. For each pixel in the image, its corresponding value on all directional response maps is traversed, and the magnitude of these response values ​​is compared. The response value with the largest absolute value is determined as the maximum response amplitude of that pixel. This will not be elaborated further.

[0105] Specifically, the response amplitude threshold is a pre-calculated response amplitude that is obtained in advance from several response amplitudes that can identify the reference path band, and the average value of each response amplitude is determined as the response amplitude threshold.

[0106] Specifically, by analyzing the discrete texture representation values ​​and edge texture direction, a reference path band is determined to obtain the interfering organ region. This provides data for the subsequent calculation of the interference coefficient for this region. In reality, organ edge textures often exhibit ambiguity or structural occlusion due to anatomical adhesions with adjacent tissues or organs, making it difficult for the model to accurately identify and segment the target region. If conventional non-selective data augmentation strategies are still used in such cases, the augmentation operation will simultaneously amplify the confusing features of the adhered region and the noise signal introduced by the occlusion, causing further distortion of the organ boundary semantics, reducing the model's ability to discriminate the real structure, and leading to data annotation errors. Based on this, this invention determines the interfering organ region by identifying the reference path band, providing data for further calculation of the interference coefficient for this region, providing a theoretical basis for subsequently determining the data augmentation model to be called, and improving the accuracy and efficiency of data annotation.

[0107] Specifically, the annotation interference module calculates the annotation interference coefficient, including:

[0108] The ratio of the texture discrete representation value to the baseline texture discrete representation value is used to determine the texture influence factor;

[0109] The ratio of the organ coverage characterization value to the baseline organ coverage characterization value is used to determine the coverage influence factor;

[0110] The weighted sum of the texture influence factor and the occlusion influence factor is used to determine the annotation interference coefficient.

[0111] Specifically, the baseline texture discrete characterization value is calculated in advance. The texture discrete characterization values ​​of reference path bands in several organ adhesion region images are obtained in advance, and the mean value of each texture discrete characterization value is determined as the baseline texture discrete characterization value.

[0112] Specifically, the baseline organ occlusion characterization value is calculated in advance. The organ occlusion characterization values ​​of several organ adhesion region images with reference path bands are obtained in advance, and the average value of each organ occlusion characterization value is determined as the baseline organ occlusion characterization value.

[0113] Specifically, the sum of the weighting coefficients of the texture influence factor and the occlusion influence factor is 1. When configuring the weighting coefficients, considering that edge textures directly affect specific areas and have a greater impact on data augmentation, the weighting coefficient of the texture influence factor is 0.6 and the weighting coefficient of the occlusion influence factor is 0.4.

[0114] Please see Figure 4 , Figure 4 This is a logic block diagram illustrating whether to invoke a data augmentation model according to an embodiment of the invention. Specifically, the interference labeling module determines whether to invoke the data augmentation model, wherein...

[0115] If the labeled interference coefficient is greater than the labeled interference coefficient threshold, then the data augmentation model is invoked to perform data augmentation on the medical image.

[0116] If the annotation interference coefficient is less than or equal to the annotation interference coefficient threshold, then the data augmentation model will not be invoked, and the medical image will be annotated normally.

[0117] Specifically, the labeled interference coefficient threshold characterizes a boundary where surrounding organs become interfering organs during data augmentation. It is pre-calculated, and several labeled interference coefficients that interfere with interfering organ regions are obtained in advance. The product of the mean of each labeled interference coefficient and the interference precision is determined as the labeled interference coefficient threshold. The interference precision is selected within the interval [0.85, 1.0]. In practice, in order to improve the accuracy of data augmentation, the interference precision is determined to be 0.9.

[0118] Understandably, normal annotation refers to calling a conventional image segmentation model (such as U-Net) or having the medical image annotated by a professional physician, without triggering data augmentation.

[0119] Please see Figure 5 , Figure 5 This is a schematic diagram illustrating the steps of a data augmentation method for small-sample medical image annotation according to an embodiment of the invention. Specifically, the method of the data augmentation system for small-sample medical image annotation includes:

[0120] Step S1: Obtain the disease description corresponding to the medical image, determine several organs affected by the disease based on the disease description, perform lesion analysis on each affected organ, calculate the lesion occurrence characterization value of each affected organ, and determine the lesion analysis organ.

[0121] Step S2: Obtain the edge texture of the lesion analysis organ to determine the texture discrete characterization value, obtain the occlusion point of the lesion analysis organ, and determine the organ occlusion characterization value;

[0122] Step S3: Based on the texture discrete characterization value, determine the organ adhesion region, analyze the edge texture direction of the organ adhesion region to obtain the reference path band, so as to determine the interfering organ region. For the interfering organ region, calculate the labeled interference coefficient based on the texture discrete characterization value and the organ occlusion characterization value, and determine whether to call the data augmentation model.

[0123] Step S4: In response to the call status of the data augmentation model, the data on the reference path band is augmented to obtain the augmented organ region, and the medical image annotation of the augmented organ region is completed.

[0124] It is understood that there are no limitations on the data augmentation method. For example, generative adversarial networks or diffusion models can be used to generate texture data in this area to augment the data on the reference path band of the interfering organ region. Of course, those skilled in the art can also use other methods, as long as data augmentation can be performed. This will not be elaborated further.

[0125] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0126] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A data enhancement system for small sample medical image annotation, characterized in that, The application comprises: a lesion analysis module configured to determine a plurality of disease-affected organs based on disease expressions corresponding to medical image pictures, perform lesion analysis on each of the disease-affected organs, calculate lesion occurrence representation values of each of the disease-affected organs, and determine a lesion analysis organ; a data processing module connected to the lesion analysis module and configured to obtain edge texture of the lesion analysis organ, determine a texture dispersion representation value, obtain occlusion points of the lesion analysis organ, and determine an organ occlusion representation value; a labeling interference module connected to the data processing module and configured to determine an organ adhesion region based on the texture dispersion representation value, analyze edge texture direction of the organ adhesion region to obtain a reference path band, determine an interference organ region, calculate a labeling interference coefficient based on the texture dispersion representation value and the organ occlusion representation value for the interference organ region, and determine whether to call a data enhancement model; a data enhancement module connected to the labeling interference module and configured to, in response to a calling state of the data enhancement model, enhance data on the reference path band to obtain an enhanced organ region, and complete labeling of the medical image pictures of the enhanced organ region.

2. The system for data augmentation for small sample medical image annotation of claim 1, wherein, The lesion analysis module determines a plurality of disease-affected organs, comprising: determining a plurality of disease keywords based on the disease expressions; traversing a disease database according to each of the disease keywords to determine a plurality of organs corresponding to each of the disease keywords; determining each of the organs as a disease-affected organ.

3. The data augmentation system for small sample medical image annotation of claim 1, wherein, The lesion analysis module calculates lesion occurrence representation values of each of the disease-affected organs and determines a lesion analysis organ, comprising: determining a historical incidence rate of each of the disease-affected organs as a lesion occurrence representation value; arranging each of the lesion occurrence representation values in descending order; arranging the disease-affected organ corresponding to the lesion occurrence representation value in the first place as the lesion analysis organ.

4. The system of claim 1, wherein, The data processing module determines a texture dispersion representation value, comprising: obtaining edge texture of the lesion analysis organ; determining a dispersion degree of the edge texture as the texture dispersion representation value.

5. The system of claim 1, wherein, The data processing module determines an organ occlusion representation value, comprising: obtaining a number of occlusion point pixels of the lesion analysis organ; determining a ratio of the number of occlusion point pixels to a total number of pixels of the lesion analysis organ as the organ occlusion representation value; wherein the occlusion points are points different in color from the lesion analysis organ.

6. The small sample medical image annotation data augmentation system of claim 1, wherein, The labeling interference module determines an organ adhesion region, wherein: if the texture dispersion representation value is greater than a texture dispersion representation value threshold, a lesion analysis organ edge region is determined as the organ adhesion region; if the texture dispersion representation value is less than or equal to the texture dispersion representation value threshold, the lesion analysis organ edge region is determined as a normal region; wherein the lesion analysis organ edge region refers to a region formed based on the edge texture of the lesion analysis organ.

7. The small sample medical image annotation data augmentation system of claim 1, wherein, The labeling interference module determines an interference organ region, comprising: obtaining all pixel points in an image of the organ adhesion region, and determining a maximum response amplitude of each of the pixel points; determining a filter direction corresponding to the maximum response amplitude as a dominant texture direction of each of the pixels to generate a global direction field map; recording a candidate seed point on the direction field map; determining a reference path band formed by the candidate seed point along the dominant texture direction as the reference path band; if the reference path band can be identified, determining an area composed of the reference path band as an interfered organ area; if the reference path band cannot be identified, not determining the interfered organ area; wherein the candidate seed point is a pixel point with a maximum response amplitude greater than a response amplitude threshold.

8. The small sample medical image annotation data augmentation system of claim 1, wherein, The labeling interference module calculates a labeling interference coefficient, including, determining a ratio of the texture discrete representation value to a reference texture discrete representation value as a texture influence factor; determining a ratio of the organ covering representation value to a reference organ covering representation value as a covering influence factor; determining a weighted sum of the texture influence factor and the covering influence factor as the labeling interference coefficient.

9. The small sample medical image annotation data augmentation system of claim 1, wherein, The labeling interference module determines whether to call a data enhancement model, wherein, if the labeling interference coefficient is greater than a labeling interference coefficient threshold, it is determined to call the data enhancement model to perform data enhancement on the medical image picture; if the labeling interference coefficient is less than or equal to the labeling interference coefficient threshold, it is determined not to call the data enhancement model to perform normal labeling on the medical image picture.

10. A method applied to the data augmentation system for small sample medical image labeling according to any one of claims 1-9, characterized in that, including: obtaining a medical image picture corresponding to a disease expression, determining a plurality of disease-affected organs based on the disease expression, performing lesion analysis on each of the disease-affected organs, calculating a lesion occurrence representation value of each disease-affected organ, and determining a lesion analysis organ; obtaining an edge texture of the lesion analysis organ to determine a texture discrete representation value, and obtaining a covering point of the lesion analysis organ to determine an organ covering representation value; based on the texture discrete representation value, determining an organ adhesion area, analyzing the edge texture of the organ adhesion area to obtain a reference path band, determining an interfered organ area, and based on the texture discrete representation value and the organ covering representation value, calculating a labeling interference coefficient to determine whether to call a data enhancement model; in response to the calling state of the data enhancement model, enhancing the data on the reference path band to obtain an enhanced organ area, and completing the medical image picture labeling of the enhanced organ area.

Images