Fractional calibration method and device based on ultrasonic medical image
By constructing anomaly score data sources using generative adversarial networks and graph algorithms, and combining convolutional neural networks and graph structure models, the accuracy and stability issues of lesion detection in ultrasound images were resolved, achieving precise quantification of abnormal areas and consistency in scoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGZHOU FIRST PEOPLES HOSPITAL (GUANGZHOU DIGESTIVE DISEASE CENT GUANGZHOU FIRST PEOPLES HOSPITAL GUANGZHOU MEDICAL UNIV THE SECOND AFFILIATED HOSPITAL OF SOUTH CHINA UNIV OF TECH)
- Filing Date
- 2026-01-08
- Publication Date
- 2026-04-17
AI Technical Summary
Existing fractional calibration methods for ultrasound medical images are inadequate to capture the distribution differences between lesion tissue and normal tissue in deep structures, fail to utilize the spatial proximity relationship between images, lack modeling of temporal changes in lesions, resulting in insufficient recall and accuracy of abnormality detection, and the scoring is not stable or reliable enough.
Generative adversarial networks and graph algorithms are used to construct anomaly score data sources. By reconstructing difference assessment and topological correlation assessment, combined with convolutional neural networks and graph structure models, the scores of image data are calibrated. Generators and discriminators are used to calculate reconstructed differences, screen outomaly regions, and construct a unified scoring system through topological correlation assessment and temporal evolution analysis.
It enables precise quantification of abnormal areas in ultrasound images, improves the accuracy of identification in complex anatomical structures and the ability to detect early lesions, enhances the consistency and reliability of scoring, and provides a higher quality data foundation.
Smart Images
Figure CN121883284A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of data calibration, specifically to a fractional calibration method and device based on ultrasound medical images. Background Technology
[0002] With the widespread application of ultrasound medical imaging in clinical diagnosis, how to achieve automatic identification and quantitative scoring of lesion areas based on massive image data has become a hot research topic. However, at present, the score calibration of ultrasound medical images still faces the following challenges:
[0003] Current ultrasound image anomaly detection mainly relies on a few explicit features such as texture and grayscale, which makes it difficult to capture the distribution differences between lesion tissue and normal tissue in deep structures, resulting in insufficient recall and accuracy of anomaly detection.
[0004] Ultrasound images exhibit spatial continuity between different sections and anatomical layers, but traditional algorithms only perform calculations on single-frame images and do not utilize topological information such as spatial proximity relationships and texture mutual information between regions, which can easily lead to misjudgments in complex anatomical structures.
[0005] Most existing methods are based on a single examination, but many lesions in clinical practice have the characteristics of gradual evolution. There is a lack of modeling of abnormal trends in changes over time between historical images and current images, resulting in insufficient detection capabilities for early lesions and lesions with subtle changes.
[0006] Different imaging devices and different operators can produce significant differences. Most existing score calibration algorithms rely on empirical thresholds and fail to calibrate based on the depth features, statistical features, and prior probabilities of abnormal areas, resulting in unstable and unreliable final scores. Summary of the Invention
[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a fractional calibration method based on ultrasound medical images, comprising,
[0008] The process involves acquiring ultrasound medical image data and constructing anomaly score data sources using an anomaly score assessment model, specifically as follows:
[0009] The system acquires ultrasound probe scan image sequence data in real time through a deployed interface. An anomaly score training mechanism is constructed based on generative adversarial networks and graph algorithms, including reconstruction difference assessment and topological correlation assessment. The acquired ultrasound medical image data is sequentially processed through a two-layer assessment mechanism, and the detected abnormal regions are quantified and scored, then combined into an anomaly score data source. This also includes…
[0010] Historically diagnosed typical lesion image data and normal tissue image data are collected from the database. Convolutional neural network algorithm is used to determine the generator and discriminator of the generative adversarial network. Based on the determined generator and discriminator, the reconstruction difference of the acquired ultrasound medical image data is calculated.
[0011] Furthermore, the consistency of topological connections between anomaly candidate regions and the global anatomical context is analyzed using a graph structure model;
[0012] Furthermore, based on the constructed abnormal score data source, score calibration of ultrasound medical images is performed, specifically as follows:
[0013] By constructing a score calibration model, image data from abnormal score data sources is used as input data for the model, and the image data score is calibrated based on the model's output.
[0014] As a preferred embodiment of the fractional calibration method based on ultrasound medical images described in this invention, the specific steps of calculating the reconstructed differences of the acquired ultrasound medical image data based on the determined generator and discriminator are as follows:
[0015]
[0016] in, This indicates the currently acquired ultrasound medical image data. This represents a pre-trained encoder network. Indicates pairing A generator network is used, and after training, it generates reconstructed normal tissue image data. express The first paradigm is used to calculate the pixel-level absolute difference between the reconstructed image and the original image. This represents the mutual information between the acquired ultrasound medical image data and the corresponding coded features. This represents the coded features corresponding to the acquired ultrasound image data. Represents the balance coefficient. This indicates the degree of reconstruction difference in the acquired ultrasound medical image data, which is used for preliminary screening of abnormal areas.
[0017] As a preferred embodiment of the fractional calibration method based on ultrasound medical images described in this invention, the preliminary screening of abnormal areas specifically includes the following:
[0018] If the reconstruction difference of the currently acquired ultrasound medical image data, compared with the mean and standard deviation of the reconstruction difference of historical normal image data, satisfies the formula... If the value is 0, it means that the area corresponding to the currently acquired ultrasound medical image data is a normal tissue area.
[0019] If the reconstruction difference of the currently acquired ultrasound medical image data, compared with the mean and standard deviation of the reconstruction difference of historical normal image data, satisfies the formula... If , it means that the area corresponding to the currently acquired ultrasound medical image data is an abnormal tissue area;
[0020] If the reconstruction difference of the currently acquired ultrasound medical image data, compared with the mean and standard deviation of the reconstruction difference of historical normal image data, satisfies the formula... If the region corresponding to the currently acquired ultrasound medical image data is an abnormal candidate region, then the abnormal score data source is constructed through topological correlation evaluation.
[0021] As a preferred embodiment of the fractional calibration method based on ultrasound medical images described in this invention, the topological correlation evaluation is specifically as follows:
[0022] The selected abnormal candidate regions are sorted according to the spatial sequence of image acquisition;
[0023] The segmented cross-sectional sequences are input into a graph structure model as data nodes. Topological correlation analysis of the image region is then performed based on the correlation analysis results between nodes in the graph structure model. This also includes...
[0024] Based on the constructed graph structure model, calculate the topological consistency score of the image region in the current cross section;
[0025] Based on the graph structure model of the same anatomical site in the user's historical examination, calculate the topological evolution anomaly score of the current image region over the time span.
[0026] Based on the calculated topological consistency score and topological evolution anomaly score, the final quantitative score of the current image region in the graph structure model is calculated comprehensively.
[0027] Based on the calculated final quantitative score, an anomaly score data source is constructed.
[0028] As a preferred embodiment of the fractional calibration method based on ultrasound medical images described in this invention, the step of inputting data nodes into the graph structure model is as follows:
[0029] Input the segmented sequence of sections into the graph structure model. ,in, Representing a graph structure model, This represents the set of nodes in a graph structure model. This represents the edge connection relationship between nodes in a graph structure model. The nodes in the graph structure model are composed of image regions in the cross section. The edge connection relationship in the graph structure model is the correlation relationship between image regions. At the same time, the edge connection weight coefficient between regions is determined based on the spatial proximity and texture mutual information between regions.
[0030] As a preferred embodiment of the score calibration method based on ultrasound medical images described in this invention, the specific steps for comprehensively calculating the final quantitative score of the current image region in the graph structure model are as follows:
[0031]
[0032] in, This represents the adaptive fusion weight coefficients for the time and space dimensions. Indicates the first The first section in the th cross-section The topological consistency score corresponds to each image region. This represents the topological evolution anomaly score calculated from the set of historical inspection time points. Indicates the first The first section in the th cross-section The final quantitative score of each image region in the graph structure model.
[0033] As a preferred embodiment of the fraction calibration method based on ultrasound medical images described in this invention, the specific steps for constructing the abnormal fraction data source are as follows:
[0034] If the final quantitative score corresponding to the acquired ultrasound medical image data satisfies the formula... If the result is positive, it indicates that the current region is performing normally in both spatial and temporal topology. The reconstruction difference score corresponding to the current region is combined with the final quantitative score and marked as safety score data, which is then incorporated into the safety score data source.
[0035] If the final quantitative score corresponding to the acquired ultrasound medical image data satisfies the formula... If the result is negative, it indicates that the current region exhibits anomalies in both spatial and temporal topology. The reconstruction difference score corresponding to the current region is combined with the final quantitative score and marked as anomaly score data, which is then incorporated into the anomaly score data source.
[0036] As a preferred embodiment of the fractional calibration method based on ultrasound medical images described in this invention, the fractional calibration of ultrasound medical images based on the constructed abnormal fractional data source is specifically as follows:
[0037] Based on the image region data in the abnormal score data source, the abnormal score-related features of the region are extracted from the current data. At the same time, features with the same feature dimension are extracted from the historical database. Based on the features extracted from the historical database, a score standard feature vector is constructed, including features of confirmed lesions and features of confirmed normality.
[0038] Calculate the posterior probability of confirmed lesions in the current area to be calibrated, and use the calculated posterior probability as the calibrated score to perform score calibration of ultrasound medical images.
[0039] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described fractional calibration method based on ultrasound medical images.
[0040] The beneficial effects of this invention are:
[0041] This invention employs a reconstruction difference assessment technique based on generative adversarial networks to achieve precise quantification of pixel-level and feature-level differences between normal and abnormal tissues in ultrasound images, effectively improving the initial screening accuracy of abnormal areas.
[0042] By employing a graph convolutional neural network-based topological correlation evaluation technique, a comprehensive analysis of spatial proximity and texture mutual information between different cross-sectional image regions was achieved, significantly improving the accuracy of abnormal region identification under complex anatomical structures.
[0043] By employing time-topological evolution analysis technology combined with historical examinations, a quantitative assessment of the trend of lesion changes over time was achieved, enhancing the detection capability of early and progressive lesions.
[0044] By employing an adaptive fusion technique of spatial topology scoring and temporal evolution scoring, a unified anomaly scoring system was constructed, providing accurate and interpretable quantitative indicators for subsequent score calibration.
[0045] By employing Bayesian score calibration technology based on prior probability and likelihood distribution, dynamic calibration of quantitative scoring of abnormal areas was achieved, significantly improving the consistency and reliability of scoring under different equipment and inspection conditions.
[0046] By employing the typical lesion scoring interval (quartile) comparison technique, the standardized construction of safe score data sources and abnormal score data sources was achieved, providing a higher quality data foundation for subsequent diagnostic models. Attached Figure Description
[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:
[0048] Figure 1 This is a schematic diagram of the overall method steps of the fractional calibration method based on ultrasound medical images according to the present invention. Detailed Implementation
[0049] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.
[0050] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0051] Example 1
[0052] Reference Figure 1 As an embodiment of the present invention, a method for calibrating the fraction of a certain object based on ultrasound medical imaging is provided, comprising the following steps:
[0053] S1: Acquire ultrasound medical image data and construct anomaly score data source using an anomaly score assessment model.
[0054] Specifically, the acquisition of ultrasound medical image data and the construction of an anomaly score data source using the anomaly score assessment model involves acquiring real-time ultrasound image data and then constructing the anomaly score data source using the anomaly score assessment model. The specific implementation is as follows:
[0055] An image acquisition interface is deployed on the ultrasound diagnostic equipment, and image sequence data is generated in real time by acquiring ultrasound probe scans through the deployed interface. The acquired image data and user basic information are stored in the medical image database to ensure secure data storage, rapid retrieval, and standardized access.
[0056] An anomaly score data source is constructed using an anomaly score evaluation model. This model is based on a generative adversarial network and graph algorithm to build an anomaly score training mechanism. The anomaly score training mechanism consists of two layers: reconstruction difference evaluation and topological correlation evaluation. The acquired ultrasound medical image data is sequentially passed through the two-layer evaluation mechanism, and the detected abnormal regions are quantified and scored before being combined to form the anomaly score data source. The specific implementation is as follows:
[0057] For the acquired ultrasound medical image data, historically diagnosed typical lesion image data and normal tissue image data are collected from the database. A convolutional neural network algorithm is used to determine the generator and discriminator of the generative adversarial network. The generator takes the extracted normal tissue image data and inputs it into the convolutional neural network algorithm. After multiple training layers, it generates reconstructed normal tissue image data. The discriminator takes the extracted typical lesion image data and inputs it into the convolutional neural network algorithm. After multiple training layers, it generates reconstructed abnormal tissue image data. Based on the determined generator and discriminator, the reconstruction difference of the acquired ultrasound medical image data is calculated, thereby achieving reconstruction difference assessment, as detailed below:
[0058] Based on generative adversarial networks for reconstruction difference detection, we have:
[0059]
[0060] in, This indicates the currently acquired ultrasound medical image data. This represents a pre-trained encoder network. Indicates pairing A generator network is used, and after training, it generates reconstructed normal tissue image data. express The first paradigm is used to calculate the pixel-level absolute difference between the reconstructed image and the original image. This represents the mutual information between the acquired ultrasound medical image data and the corresponding coded features. This represents the coded features corresponding to the acquired ultrasound image data. This represents the balance coefficient, which is set by the implementers based on the actual application scenario. This indicates the degree of reconstruction difference in the acquired ultrasound medical image data, used for preliminary screening of abnormal areas, specifically:
[0061] Based on historical data from the historical database, a generative adversarial network (GAN) is trained on normal tissue image data. The GAN is then used to obtain the reconstruction difference between historical normal image data and historical lesion image data. Based on the calculated reconstruction difference, preliminary screening of abnormal regions is performed. Therefore,
[0062] If the reconstruction difference of the currently acquired ultrasound medical image data, compared with the mean and standard deviation of the reconstruction difference of historical normal image data, satisfies the formula... If the value is 0, it means that the area corresponding to the currently acquired ultrasound medical image data is a normal tissue area. The normal tissue area is marked as a normal area, and the reconstruction difference evaluation score corresponding to the normal area is set to 0.
[0063] If the reconstruction difference of the currently acquired ultrasound medical image data, compared with the mean and standard deviation of the reconstruction difference of historical normal image data, satisfies the formula... If the value is 0, it indicates that the area corresponding to the currently acquired ultrasound medical image data is an abnormal tissue area. The abnormal tissue area is marked as an abnormal area, and the reconstruction difference assessment score corresponding to the normal area is set to 1.
[0064] If the reconstruction difference of the currently acquired ultrasound medical image data, compared with the mean and standard deviation of the reconstruction difference of historical normal image data, satisfies the formula... If the region corresponding to the currently acquired ultrasound medical image data is an abnormal candidate region, then the abnormal score data source is constructed through topological correlation evaluation.
[0065] Furthermore, topological correlation assessment involves constructing corresponding feature map structures from ultrasound medical image data collected in anomaly candidate regions. These feature map structures are then input into a graph convolutional neural network model. By analyzing the consistency of topological connections between the anomaly candidate regions and the global anatomical context, the final quantitative assessment is completed, as detailed below:
[0066] For the selected abnormal candidate regions, sorting them according to the spatial sequence of image acquisition (different scanning sections) yields the following:
[0067]
[0068] in, This represents the sequence of sections that are divided. This represents the first section of the division. Indicates the first division One cross-section, Let represent the total number of slices, and each slice contains multiple image regions. Then, ,in, This indicates the first image region. Indicates the first Each image area This indicates the total number of image regions in the cross section, which can be set by the implementer based on the actual application.
[0069] Furthermore, the segmented cross-sectional sequences are input into a graph structure model in the form of data nodes. Based on the correlation analysis results between nodes in the graph structure model, topological correlation analysis of the image region is performed. The specific implementation is as follows:
[0070] Input the segmented sequence of sections into the graph structure model. ,in, Representing a graph structure model, This represents the set of nodes in a graph structure model. This represents the edge connection relationship between nodes in the graph structure model. The nodes in the graph structure model are composed of image regions in the cross section. The edge connection relationship in the graph structure model is the correlation relationship between image regions. At the same time, the edge connection weight coefficient between regions is determined based on the spatial proximity and texture mutual information between regions.
[0071] Based on the constructed graph structure model, the topological consistency score of the image region in the current cross section is calculated, and then...
[0072]
[0073] in, Indicates the first The first section in the th cross-section Each image area Indicates the first division One cross-section, Indicates the first Consistent representation of image regions in corresponding cross sections. This represents the activation function, used to map representations to intervals. middle, Indicates the first The first section in the th cross-section Topological consistency score corresponding to each image region;
[0074] Based on the graph structure model of the same anatomical site in the user's historical examinations, the topological evolution anomaly score of the current imaging region over the time span is calculated, specifically as follows:
[0075] Set a set of historical inspection time points Calculate the topological evolution anomaly score in the set of historical inspection time points, then we have:
[0076]
[0077] in, This indicates the time node selected from the set of historical inspection time points. This represents the set of historical inspection time points. Indicates the first The first section in the th cross-section Each image area This indicates the area corresponding to the anatomical location during historical examinations. Indicates the first The first section in the th cross-section The topological consistency score corresponds to each image region. This indicates the number corresponding to the anatomical location in the historical examination. The first section in the th cross-section The topological consistency score corresponds to each image region. This represents the time decay weighting coefficient, which is set by the implementers based on the actual application scenario. This represents the topological evolution anomaly score in the set of historical inspection time nodes.
[0078] Based on the calculated topological consistency score and topological evolution anomaly score, a comprehensive calculation of the global anomaly score in the graph structure model of the current image region yields the following:
[0079]
[0080] in, The adaptive fusion weighting coefficients representing the time and spatial dimensions are set by the implementers based on the actual application scenario. Indicates the first The first section in the th cross-section The topological consistency score corresponds to each image region. This represents the topological evolution anomaly score calculated from the set of historical inspection time points. Indicates the first The first section in the th cross-section The final quantitative score of each image region in the graph structure model.
[0081] Based on the calculated final quantitative score, an anomaly score data source is constructed, as specifically implemented below:
[0082] The lesion regions corresponding to typical lesion images from the historical database are input into the constructed graph structure model, and the corresponding final quantitative scores are determined. By comparing the final quantitative score corresponding to the acquired ultrasound medical image data with the final quantitative score corresponding to the lesion area, an abnormal score data source is constructed based on the comparison results. Specifically:
[0083] If the final quantitative score corresponding to the acquired ultrasound medical image data satisfies the formula... If the result is positive, it indicates that the current region is performing normally in both spatial and temporal topology. The reconstruction difference score corresponding to the current region is combined with the final quantitative score and marked as safety score data, which is then incorporated into the safety score data source.
[0084] If the final quantitative score corresponding to the acquired ultrasound medical image data satisfies the formula... If the result is negative, it indicates that the current region exhibits anomalies in both spatial and temporal topology. The reconstruction difference score corresponding to the current region is combined with the final quantitative score and marked as anomaly score data, which is then incorporated into the anomaly score data source.
[0085] It should be noted that the process of constructing the abnormal score data source... , These represent the first and third quartiles of the final quantitative score corresponding to the lesion area, respectively. The first quartile represents the 25th percentile of the final quantitative score corresponding to the lesion area, and the third quartile represents the 75th percentile of the final quantitative score corresponding to the lesion area.
[0086] The constructed safe score data sources and abnormal score data sources are indexed and sorted according to the acquired image sequences and spatial locations to provide an accurate data foundation for subsequent score calibration of ultrasound medical images.
[0087] S2: Score calibration of ultrasound medical images based on the constructed abnormal score data source.
[0088] Specifically, score calibration of image data from outlier score data sources is achieved by constructing a score calibration model. The image data from these outlier score data sources is used as input to the model, and the model's output is used to calibrate the image data's score. The specific implementation is as follows:
[0089] For image region data in the abnormal score data source, extract abnormal score-related features of the region from the current data, including the abnormal score of the region, the area of the region, the average gray value of the region, the texture entropy value of the region, and the frequency of the region in the historical inspection sequence.
[0090] The extracted abnormal score-related features are defined as follows: ,in, This represents the feature set formed by the combination of features related to outlier scores. This represents the first feature in the feature set. Represents the first in the feature set The feature set represents the total number of features in the feature set. The specific number is set by the implementer according to the actual application scenario. In this embodiment, the feature set includes five features: regional anomaly score, regional area, regional average echo intensity, echo intensity gradient between the region and adjacent regions, and regional shape irregularity index.
[0091] Meanwhile, features with the same feature dimensions are extracted from the historical database, and a score standard feature vector is constructed based on the features extracted from the historical database, including features of confirmed lesions and features of confirmed normality.
[0092] Furthermore, the score calibration model calculates the posterior probability of confirmed lesions in the current calibration area based on the feature data of the current calibration region, combined with the prior probability and likelihood distribution calculated from historical data. The calculated posterior probability is then used as the calibrated score. The specific implementation is as follows:
[0093] Set the feature vector of the current region to be calibrated as ,in, This indicates the outlier score in the current region to be calibrated. This indicates the area of the region currently to be calibrated. This represents the average gray value of the area to be calibrated. This represents the region texture entropy value of the area currently to be calibrated. This indicates the frequency of the region to be calibrated in the historical inspection sequence. This represents a feature vector constructed from features related to anomaly scores;
[0094] At the same time, the standard feature vector of the score is set as ,in, Indicates the characteristics of the diagnosed lesions. This indicates confirmation of normal characteristics. This represents the constructed score standard feature vector.
[0095] Based on the constructed feature vector, the posterior probability of confirming lesions in the current region to be calibrated is calculated, and the calculation result is used as the calibrated score. The specific implementation is as follows:
[0096] The prior probability is calculated based on the eigenvectors of the score standard, specifically as follows:
[0097]
[0098] in, This represents the features of diagnosed lesions in the standard feature vector of the score. This represents the prior probability of confirming normal features. This represents the total amount of data extracted from the historical database. This represents the number of confirmed lesions in the total data.
[0099] The likelihood distribution, on the other hand, uses the kernel density function to calculate the likelihood distribution of each continuous feature in the feature vector of the current region to be calibrated.
[0100] Furthermore, based on the calculated prior probabilities, the posterior probability of confirming lesions in the current region to be calibrated is calculated using a Bayesian algorithm, as follows:
[0101]
[0102] in, This represents the probability that a feature in the feature vector constructed from the current region to be calibrated is correct, assuming that the confirmed lesion features in the current region to be calibrated are correct. This represents the prior probability of the diagnostic lesion feature in the standard feature vector of the fraction. This represents the probability that a feature in the feature vector constructed from the current region to be calibrated is correct, assuming that the confirmed normal features in the current region to be calibrated are correct. This represents the prior probability of identifying normal features in the standard feature vector of the score. This represents the posterior probability of confirming a lesion in the current area to be calibrated, used for fractional calibration of ultrasound medical images, as follows:
[0103] Set an abnormal score calibration threshold, and perform score calibration of ultrasound medical images based on the set abnormal score calibration threshold, specifically as follows:
[0104] If the calculated posterior probability of confirmed lesions in the current region to be calibrated exceeds the set abnormal score calibration threshold, it indicates that the current region to be calibrated is confirmed as abnormal after calibration and is retained in the abnormal score data source after calibration.
[0105] If the calculated posterior probability of the confirmed lesion in the current region to be calibrated is lower than the set abnormal score calibration threshold, it means that the current region to be calibrated is confirmed to be normal after calibration, and the abnormal score data corresponding to the current region is removed from the abnormal score data source.
[0106] It should be noted that the pre-trained encoder network is a deep convolutional neural network. The corresponding structure and weights have been trained and optimized on a large and diverse normal ultrasound image dataset. The input two-dimensional or three-dimensional ultrasound images are nonlinearly transformed and dimensionality reduced to extract low-dimensional feature vectors that can characterize their core anatomical structures and tissue textures. The extracted feature vectors discard redundant information at the original pixel level and retain key semantic features such as tissue boundaries, echo intensity distribution, and local patterns, laying the foundation for subsequent reconstruction and comparison. The specific information to be discarded and retained is determined by the implementers according to the actual application scenario.
[0107] Furthermore, if the aforementioned function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0108] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.
[0109] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.
[0110] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A score calibration method based on ultrasound medical images, characterized in that: include, The process involves acquiring ultrasound medical image data and constructing anomaly score data sources using an anomaly score assessment model, specifically as follows: The system acquires ultrasound probe scan image sequence data in real time through a deployed interface. An anomaly score training mechanism is constructed based on generative adversarial networks and graph algorithms, including reconstruction difference assessment and topological correlation assessment. The acquired ultrasound medical image data is sequentially processed through a two-layer assessment mechanism, and the detected abnormal regions are quantified and scored, then combined into an anomaly score data source. This also includes… Historically diagnosed typical lesion image data and normal tissue image data are collected from the database. Convolutional neural network algorithm is used to determine the generator and discriminator of the generative adversarial network. Based on the determined generator and discriminator, the reconstruction difference of the acquired ultrasound medical image data is calculated. Furthermore, the consistency of topological connections between anomaly candidate regions and the global anatomical context is analyzed using a graph structure model; Furthermore, based on the constructed abnormal score data source, score calibration of ultrasound medical images is performed, specifically as follows: By constructing a score calibration model, image data from abnormal score data sources is used as input data for the model, and the image data score is calibrated based on the model's output.
2. The score calibration method based on medical ultrasound images of claim 1, wherein: The specific differences in the reconstruction of the acquired ultrasound medical image data calculated based on the determined generator and discriminator are as follows: wherein, represents the current collected ultrasound medical image data, represents a pre-trained encoder network, represents a paired generator network, and generates reconstructed normal tissue image data after training, represents 1 norm, used to calculate the pixel-level absolute difference between the reconstructed image and the original image, represents the mutual information between the collected ultrasound medical image data and the corresponding encoded features, represents the encoded features corresponding to the collected ultrasound image data, represents a balance coefficient, represents the reconstruction difference degree of the collected ultrasound medical image data, used for preliminary screening of abnormal areas.
3. The fractional calibration method based on ultrasound medical images as described in claim 2, characterized in that: The specific steps for preliminary screening of abnormal areas are as follows: If the reconstruction difference of the currently acquired ultrasound medical image data, compared with the mean and standard deviation of the reconstruction difference of historical normal image data, satisfies the formula... If the value is 0, it means that the area corresponding to the currently acquired ultrasound medical image data is a normal tissue area. If the reconstruction difference of the currently acquired ultrasound medical image data, compared with the mean and standard deviation of the reconstruction difference of historical normal image data, satisfies the formula... If , it means that the area corresponding to the currently acquired ultrasound medical image data is an abnormal tissue area; If the reconstruction difference of the currently acquired ultrasound medical image data, compared with the mean and standard deviation of the reconstruction difference of historical normal image data, satisfies the formula... If the region corresponding to the currently acquired ultrasound medical image data is an abnormal candidate region, then the abnormal score data source is constructed through topological correlation evaluation.
4. The fractional calibration method based on ultrasound medical images as described in claim 3, characterized in that: The specific topological association assessment is as follows: The selected abnormal candidate regions are sorted according to the spatial sequence of image acquisition; The segmented cross-sectional sequences are input into a graph structure model as data nodes. Topological correlation analysis of the image region is then performed based on the correlation analysis results between nodes in the graph structure model. This also includes... Based on the constructed graph structure model, calculate the topological consistency score of the image region in the current cross section; Based on the graph structure model of the same anatomical site in the user's historical examination, calculate the topological evolution anomaly score of the current image region over the time span. Based on the calculated topological consistency score and topological evolution anomaly score, the final quantitative score of the current image region in the graph structure model is calculated comprehensively. Based on the calculated final quantitative score, an anomaly score data source is constructed.
5. The fractional calibration method based on ultrasound medical images as described in claim 4, characterized in that: The specific steps for inputting data nodes into the graph structure model are as follows: Input the segmented sequence of sections into the graph structure model. ,in, Representing a graph structure model, This represents the set of nodes in a graph structure model. This represents the edge connection relationship between nodes in a graph structure model. The nodes in the graph structure model are composed of image regions in the cross section. The edge connection relationship in the graph structure model is the correlation relationship between image regions. At the same time, the edge connection weight coefficient between regions is determined based on the spatial proximity and texture mutual information between regions.
6. The fractional calibration method based on ultrasound medical images as described in claim 5, characterized in that: The specific details of calculating the final quantitative score in the in-map structure model of the current image region are as follows: in, This represents the adaptive fusion weight coefficients for the time and space dimensions. Indicates the first The first section in the th cross-section The topological consistency score corresponds to each image region. This represents the topological evolution anomaly score calculated from the set of historical inspection time points. Indicates the first The first section in the th cross-section The final quantitative score of each image region in the graph structure model.
7. The fractional calibration method based on ultrasound medical images as described in claim 6, characterized in that: The specific details for constructing the abnormal score data source are as follows: If the final quantitative score corresponding to the acquired ultrasound medical image data satisfies the formula... If the result is positive, it indicates that the current region is performing normally in both spatial and temporal topology. The reconstruction difference score corresponding to the current region is combined with the final quantitative score and marked as safety score data, which is then incorporated into the safety score data source. If the final quantitative score corresponding to the acquired ultrasound medical image data satisfies the formula... If the result is negative, it indicates that the current region exhibits anomalies in both spatial and temporal topology. The reconstruction difference score corresponding to the current region is combined with the final quantitative score and marked as anomaly score data, which is then incorporated into the anomaly score data source.
8. The fractional calibration method based on ultrasound medical images as described in claim 7, characterized in that: The specific steps for fractional calibration of ultrasound medical images based on the constructed abnormal fractional data source are as follows: Based on the image region data in the abnormal score data source, the abnormal score-related features of the region are extracted from the current data. At the same time, features with the same feature dimension are extracted from the historical database. Based on the features extracted from the historical database, a score standard feature vector is constructed, including features of confirmed lesions and features of confirmed normality. Calculate the posterior probability of confirmed lesions in the current area to be calibrated, and use the calculated posterior probability as the calibrated score to perform score calibration of ultrasound medical images.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 8.