A two-dimensional cardiac ultrasound section marking data automatic acquisition method and system

Abstract

The application discloses a two-dimensional heart ultrasound section marking data automatic acquisition method and system. By inputting the ultrasound section image of the doctor video freeze frame into the trained size detection model and the auxiliary model, the size detection model outputs the key size marking in the ultrasound section image, including the end point, the line segment, the text area, the ellipse, the straight line and the angle; the auxiliary model outputs the section, the chamber and the geometric direction of the key size in the ultrasound section image; the corresponding numerical value is extracted from the key size marking by using the OCR technology, the line segment length in the key size marking is sorted, the detected size marking is converted from the pixel unit to the actual physical size of the corresponding size table according to the scale information in the image, and the sorting is performed; according to the sorting result, a plurality of groups of size information of the ultrasound section image are outputted, a measurement report is generated, and the acquisition of the size data in the two-dimensional heart ultrasound image is automatically completed.

Description

Technical Field

[0001] This application relates to the field of automatic data acquisition technology, and in particular to a method and system for automatic acquisition of two-dimensional cardiac ultrasound cross-sectional annotation data. Background Technology

[0002] Echocardiography is a crucial tool for assessing cardiac structure and function, widely used in clinical diagnosis. Echocardiographic images not only allow physicians to view key physiological indicators such as the morphology and wall thickness of the heart chambers, as well as the internal diameters of the ventricles and atria, in real time, but also enable effective assessment of the heart's pumping function by measuring these parameters. In clinical practice, sonographers typically need to perform detailed annotation and dimensional measurements on two-dimensional echocardiographic images to provide a basis for judging cardiac health. However, the manual annotation and measurement process is cumbersome and susceptible to human error, thus limiting the accuracy and consistency of the measurement results.

[0003] Currently, although ultrasound image analysis technology has made significant progress, and ultrasound artificial intelligence (AI) technology has achieved some advancements in image classification and chamber segmentation, its automation in cardiac ultrasound measurement remains insufficient. Ultrasound physicians still need to manually annotate and measure dimensions, inputting the calibrated dimensions and data into the system. Existing technologies largely focus on image classification and recognition, lacking support for automatic acquisition of dimensional data, real-time correction of physician annotation, and outlier identification. Furthermore, the operation between physicians and the system remains disconnected, failing to achieve automated correction and intelligent interaction, resulting in low measurement efficiency and a high risk of human error in numerous diagnostic tasks, further impacting diagnostic accuracy.

[0004] Therefore, automating cardiac ultrasound image measurement and improving the standardization and efficiency of data acquisition are important requirements in clinical applications. Summary of the Invention

[0005] This application provides a method, system, computer equipment, and storage medium for automatic acquisition of two-dimensional cardiac ultrasound section annotation data, aiming to solve the problem that in existing cardiac ultrasound examination procedures, ultrasound doctors usually rely on manual operation, the whole process is highly dependent on the doctor's experience, and the accuracy of measurement is easily affected by personal judgment deviations or operational errors.

[0006] In a first aspect, a method for automatically acquiring two-dimensional cardiac ultrasound cross-sectional annotation data, the method comprising:

[0007] The ultrasound cross-sectional image captured by the doctor is input into a trained size detection model and an auxiliary model. The size detection model outputs key size annotations in the ultrasound cross-sectional image, including endpoints, line segments, text regions, ellipses, lines, and angles. The auxiliary model outputs the cross-section, chamber, and geometric orientation of the chamber containing the key dimensions in the ultrasound cross-sectional image.

[0008] OCR technology is used to extract corresponding values ​​from the key dimension annotations, and the line segment lengths in the key dimension annotations are sorted. At the same time, based on the scale information in the image, the detected dimension annotations are converted from pixel units to actual physical dimensions and sorted.

[0009] Based on the sorting results, multiple sets of size information of the ultrasonic cross-section image are output, and a measurement report is generated.

[0010] Optionally, the above scheme may further include converting the detected dimension annotations from pixel units to actual physical dimensions and sorting them:

[0011] The detected key dimension annotations are matched to ensure that each annotation corresponds correctly to the actual physical size, and the matching results are verified to check for errors or omissions.

[0012] Optionally, in the above scheme, the training process of the trained size detection model includes:

[0013] Collect pre-annotated ultrasound image data, which covers various ultrasound examination scenarios, obtain ultrasound images annotated by doctors, the annotation content includes key dimension information, and preprocess the ultrasound image data, wherein the key dimension information includes: line segments, text regions, ellipses, straight lines and angles;

[0014] Detailed annotations are performed on the preprocessed ultrasound images, including endpoints, line segments, text regions, ellipses, lines, and angles.

[0015] The YOLOv11 model was used to train the recognition of the dimensions of the labeled ultrasound images. During the training process, the model parameters were optimized, and the trained model was evaluated using a validation set to check the model's ability to recognize different labeled content. Based on the evaluation results, the model parameters were adjusted or the training strategy was optimized.

[0016] Optionally, the preprocessing of the ultrasound image data in the above scheme includes: extracting images containing key size information from the ultrasound image data, preprocessing the images containing key size information to ensure consistent image quality, and converting the images into a uniform format and resolution.

[0017] In the above scheme, optionally, the auxiliary model includes an ultrasonic section classification model and an ultrasonic chamber segmentation model;

[0018] The ultrasound section classification model is used to automatically identify the section type of an ultrasound image;

[0019] The ultrasound chamber segmentation model is used to segment the chamber regions in the ultrasound image.

[0020] Optionally, in the above scheme, the OCR technology uses a convolutional neural network model combined with a long short-term memory network structure for training, in order to improve the accuracy of extracting text region size data in images.

[0021] In the above scheme, optionally, the labeled size information included in the training dataset includes multiple size measurement labels of the heart chambers, including key dimensions such as the inner diameter and wall thickness of the ventricles and atria.

[0022] Optionally, in the above scheme, the size annotation information output by the target detection model may also include a confidence score for the annotated region, which is used to represent the reliability of the detection result.

[0023] Secondly, an automatic data acquisition system for two-dimensional cardiac ultrasound cross-sectional annotation data, the system comprising:

[0024] The detection module is used to input the ultrasound cross-sectional image captured by the doctor into the trained size detection model and auxiliary model. The size detection model outputs key size annotations in the ultrasound cross-sectional image, including endpoints, line segments, text regions, ellipses, lines, and angles. The auxiliary model outputs the cross-section, chamber, and geometric orientation of the chamber containing the key dimensions in the ultrasound cross-sectional image.

[0025] Recognition module: Used to extract corresponding values ​​from the key dimension annotations using OCR technology, sort the line segment lengths in the key dimension annotations, and convert the detected dimension annotations from pixel units to actual physical dimensions and sort them according to the scale information in the image;

[0026] Output module: Used to output multiple sets of size information of the ultrasonic cross-section image according to the sorting results, and generate a measurement report.

[0027] Compared with the prior art, this application has at least the following beneficial effects:

[0028] This application, based on further analysis and research into the problems of existing technologies, recognizes that in current cardiac ultrasound examination procedures, sonographers typically rely on manual operation, and the entire process is highly dependent on the doctor's experience. This makes it susceptible to errors in judgment or operation, which can affect measurement accuracy. The application addresses this by inputting the ultrasound cross-sectional image captured by the doctor into a trained size detection model and an auxiliary model. The size detection model outputs key dimension annotations in the ultrasound cross-sectional image, including endpoints, line segments, text regions, ellipses, lines, and angles. The auxiliary model outputs the cross-section, chamber, and geometric orientation of the key dimensions in the ultrasound cross-sectional image. OCR technology is used to extract corresponding values ​​from the key dimension annotations, and the line segment lengths in the key dimension annotations are sorted. Simultaneously, based on the scale information in the image, the detected dimension annotations are converted from pixel units to actual physical dimensions and sorted. Based on the sorting results, multiple sets of dimension information from the ultrasound cross-sectional image are output, generating a measurement report. This automates the acquisition of dimension data from two-dimensional cardiac ultrasound images, effectively solving the problems of low measurement efficiency, large human error, and cumbersome operation mentioned in the background technology.

[0029] By automatically identifying key dimensions in ultrasound images using target detection technology and combining this with OCR technology to read the dimensional values, the system achieves automatic acquisition of labeled data in ultrasound images, significantly improving measurement efficiency and accuracy. Furthermore, the automated data acquisition process reduces human intervention, avoiding errors caused by physician fatigue or experience differences, and ensuring the accuracy and consistency of every measurement. Simultaneously, the system can correct the labeled data in real time, intelligently reporting abnormal dimensions to assist physicians in making corrections, further enhancing the reliability and standardization of diagnostic results. Therefore, this invention not only solves the problems of low efficiency and high error in traditional ultrasound measurement but also improves the overall accuracy and efficiency of cardiac ultrasound diagnosis, promoting the automation and intelligent development of cardiac ultrasound examinations. Attached Figure Description

[0030] Figure 1 A flowchart illustrating an automatic acquisition method for two-dimensional cardiac ultrasound section annotation data provided in one embodiment of this application;

[0031] Figure 2 This is a schematic diagram of two-dimensional cardiac ultrasound cross-section annotation provided in one embodiment of this application;

[0032] Figure 3 A flowchart illustrating an embodiment of the method for automatically obtaining dimension values ​​provided in this application;

[0033] Figure 4 This is a schematic diagram of an image showing the automatically obtained size values ​​provided in one embodiment of this application.

[0034] Figure 5 This is a schematic diagram of the size values ​​of an image automatically obtained according to an embodiment of this application;

[0035] Figure 6 A flowchart illustrating an embodiment of the method for automatically obtaining size names provided in this application;

[0036] Figure 7 This is a schematic diagram of chamber segmentation for an automatic acquisition method of two-dimensional cardiac ultrasound cross-sectional annotation data provided in one embodiment of this application;

[0037] Figure 8 A schematic diagram illustrating the overlap determination of an automatic acquisition method for two-dimensional cardiac ultrasound section annotation data provided in an embodiment of this application;

[0038] Figure 9 This is a schematic diagram illustrating the rectangular method for determining the major and minor axes in an automatic acquisition method for two-dimensional cardiac ultrasound section annotation data provided in one embodiment of this application. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0040] In one embodiment, such as Figure 1 As shown, an automatic acquisition method for two-dimensional cardiac ultrasound cross-sectional annotation data is provided, including the following steps:

[0041] The ultrasound cross-sectional image captured by the doctor is input into a trained size detection model and an auxiliary model. The size detection model outputs key size annotations in the ultrasound cross-sectional image, including endpoints, line segments, text regions, ellipses, lines, and angles. The auxiliary model outputs the cross-section, chamber, and geometric orientation of the chamber containing the key dimensions in the ultrasound cross-sectional image.

[0042] OCR technology is used to extract corresponding values ​​from the key dimension annotations, and the line segment lengths in the key dimension annotations are sorted. At the same time, based on the scale information in the image, the detected dimension annotations are converted from pixel units to actual physical dimensions and sorted.

[0043] Based on the sorting results, multiple sets of size information of the ultrasonic cross-section image are output, and a measurement report is generated.

[0044] In this embodiment, after converting the detected size annotations from pixel units to actual physical dimensions and sorting them, the process further includes:

[0045] The detected key dimension annotations are matched to ensure that each annotation corresponds correctly to the actual physical size, and the matching results are verified to check for errors or omissions.

[0046] In this embodiment, the training process of the trained size detection model includes:

[0047] Collect pre-annotated ultrasound image data, which covers various ultrasound examination scenarios, obtain ultrasound images annotated by doctors, the annotation content includes key dimension information, and preprocess the ultrasound image data, wherein the key dimension information includes: line segments, text regions, ellipses, straight lines and angles;

[0048] Detailed annotations are performed on the preprocessed ultrasound images, including endpoints, line segments, text regions, ellipses, lines, and angles.

[0049] The YOLOv11 model was used to train the recognition of the dimensions of the labeled ultrasound images. During the training process, the model parameters were optimized, and the trained model was evaluated using a validation set to check the model's ability to recognize different labeled content. Based on the evaluation results, the model parameters were adjusted or the training strategy was optimized.

[0050] In this embodiment, preprocessing the ultrasound image data includes: extracting images containing key size information from the ultrasound image data, preprocessing the images containing key size information to ensure consistent image quality, and converting the images into a uniform format and resolution.

[0051] In this embodiment, the auxiliary model includes an ultrasonic section classification model and an ultrasonic chamber segmentation model;

[0052] The ultrasound section classification model is used to automatically identify the section type of an ultrasound image;

[0053] The ultrasound chamber segmentation model is used to segment the chamber regions in the ultrasound image.

[0054] In this embodiment, the OCR technology uses a convolutional neural network model combined with a long short-term memory network structure for training, in order to improve the accuracy of extracting text region size data in images.

[0055] In this embodiment, the labeled size information included in the training dataset includes multiple size measurement labels of the heart chambers, including key dimensions such as the inner diameter and wall thickness of the ventricles and atria.

[0056] In this embodiment, the size annotation information output by the target detection model also includes a confidence score for the annotated region, which is used to represent the reliability of the detection result.

[0057] In one embodiment, an automatic acquisition system for two-dimensional cardiac ultrasound cross-sectional annotation data is provided, comprising:

[0058] The detection module is used to input the ultrasound cross-sectional image captured by the doctor into the trained size detection model and auxiliary model. The size detection model outputs key size annotations in the ultrasound cross-sectional image, including endpoints, line segments, text regions, ellipses, lines, and angles. The auxiliary model outputs the cross-section, chamber, and geometric orientation of the chamber containing the key dimensions in the ultrasound cross-sectional image.

[0059] Recognition module: Used to extract corresponding values ​​from the key dimension annotations using OCR technology, sort the line segment lengths in the key dimension annotations, and convert the detected dimension annotations from pixel units to actual physical dimensions and sort them according to the scale information in the image;

[0060] Output module: Used to output multiple sets of size information of the ultrasonic cross-section image according to the sorting results, and generate a measurement report.

[0061] The specific implementation details of each module can be found in the above description of the limitations of the automatic acquisition method for two-dimensional cardiac ultrasound cross-sectional annotation data, and will not be repeated here.

[0062] In one embodiment, an automatic size acquisition method for two-dimensional ultrasound sections is proposed, enabling size determination and numerical reading within a single image. In cardiac ultrasound diagnostic scenarios, sonographers typically take 30-50 measurements, of which approximately 20-30 are two-dimensional ultrasound measurements. Currently, there are no well-developed technologies for automatically matching the labeled dimensions and automatically reading their values. This invention, based on image recognition technology, can automatically detect key cardiac anatomical structures and accurately obtain the doctor's measurement results, improving measurement efficiency. Simultaneously, this method can run synchronously with the doctor's measurement process, providing intelligent measurement assistance. The system can automatically compare the doctor's annotation results, identify potential errors, and provide correction suggestions to ensure measurement accuracy and consistency. Furthermore, for abnormal dimensions, the system automatically alerts the doctor for verification and supports interactive adjustments to enhance the reliability of measurement results. This intelligent assistance method not only optimizes the clinical data acquisition process but also provides more efficient and accurate support for cardiac ultrasound diagnosis.

[0063] Current ultrasound AI technology primarily focuses on image classification, structural segmentation, and chamber identification, but lacks sufficient support for automating the measurement process. Doctors still need to manually mark measurement points, draw dimension lines, and input values. In particular, the doctor's measurement process is disconnected from AI analysis, and the system cannot synchronously correct errors or intelligently identify outliers. Realizing real-time reading and intelligent interaction of doctor's dimensional measurements based on existing image recognition technology, along with automatic report generation, represents a gap in current ultrasound AI applications.

[0064] This invention relates to the field of medical image analysis and intelligent measurement technology, specifically to an automatic acquisition method for two-dimensional cardiac ultrasound section annotation data. Based on image recognition technology, this method acquires dimensional data from the doctor's annotation process within a single ultrasound section, reducing the doctor's data entry work and improving the accuracy and efficiency of data acquisition. It is suitable for dimensional measurement and diagnostic assistance in freeze-frame examinations of continuous ultrasound video.

[0065] This embodiment mainly addresses two key technical issues in the automatic acquisition of two-dimensional cardiac ultrasound cross-sectional dimension data:

[0066] Automatic acquisition of dimensional values: Ultrasound images typically contain multiple dimensions that need to be labeled. Ensuring these dimensions accurately correspond to specific values ​​in a dimension table is essential. This invention employs target detection technology to detect the location of the labeled dimensions and the location of the dimension table. Dimension values ​​are obtained through OCR, and a sorting method is used to complete the dimensional value acquisition process.

[0067] Automatic size name acquisition: In traditional measurement processes, doctors rely on their experience to determine the dimensions to be measured, and the definitions of these dimensions depend on the specific cardiac chamber structure. This invention automatically identifies cardiac chamber structures in ultrasound images through section classification and chamber segmentation, and intelligently matches the dimensions to be measured based on the geometry of the chambers, thereby determining the specific measurement content and achieving standardized automatic size determination. Figure 2 , Figure 3 and Figure 4 As shown.

[0068] The first key technical issue mentioned above, the automatic acquisition of size values, includes the following three steps: Step 1: Train an object detection model to detect the acquired size information, including the rectangular bounding box and size label. Add all sizes, including the pixel length of the longer side of the rectangle, to list1; Step 2: Use OCR technology to recognize the contents of the size table and add the values ​​to list2; Step 3: Use a sorting algorithm to match sizes; Obtain labeled size information through the object detection model: Size information includes: the rectangular coordinates of the ruler, label name, and size table. First, train an object detection model for size; Second: In a normal doctor's ultrasound work scenario, extract the doctor's still image and use the object detection model to automatically recognize the size information in the ultrasound image. Calculate the length of the rectangle with line segment labels (edge) and add multiple length line segments to list1;

[0069] Using OCR to read dimension tables from images: OCR (Optical Character Recognition) technology is used to extract values ​​from the dimension table (box labels) of the target detection results. OCR can read dimension information from images, including numbers, units, and annotations. This technology can obtain values ​​from the ultrasound machine and add length values ​​(unit: mm) to list2.

[0070] Dimension length sorting and correspondence with the dimension table: All line segments in list1 are sorted according to the corresponding rectangle length, where the rectangle length is the pixel length of the rectangle. Simultaneously, the actual dimension values ​​in list2 are also sorted. This process ensures that all dimension annotations correspond accurately and that each annotated dimension corresponds one-to-one with the values ​​given by the ultrasonic machine. Figure 5 As shown.

[0071] The second key issue is the automatic acquisition of dimension names: using medical image classification and segmentation methods, the dimensions labeled by the doctor are assigned to various cavities on different sections, and then the specific dimension content is determined based on the geometry of the cavity. All dimensions uniquely correspond to a feature table, and each dimension must have three attributes: the section it belongs to, the cavity it belongs to, and the geometric orientation of the cavity. The automatic determination of dimension names consists of the following three key components: such as... Figure 6 As shown, the main function of the section classifier is to automatically classify different ultrasound section types based on the chambers and structures in the cardiac ultrasound image. Cardiac ultrasound images can present a variety of different sectional views, such as the four-chamber view, the left ventricular outflow island, and the right ventricular outflow tract. Each section contains different anatomical information; therefore, accurate section classification is fundamental to precisely acquiring specific dimensions. It can distribute a large number of dimensions across various sections, and matching candidate dimensions within each section reduces the difficulty of matching.

[0072] The section classifier uses image classification technology to divide input ultrasound images into different section categories. Based on a deep learning model, this classifier learns a large number of cardiac image section features during training, enabling it to accurately identify and classify different types of sections. For example, when the input image is a four-chamber view of the left ventricle, the section classifier can identify this section and label it as a four-chamber view. This classification process provides the foundation for subsequent chamber segmentation and keypoint recognition.

[0073] Chamber Segmentation Component: The core task of the chamber segmenter is to accurately segment the heart chambers in ultrasound images and extract the contour of each chamber. The heart has multiple chambers: the left ventricle, right ventricle, left atrium, and right atrium. The size, shape, and location of each chamber are crucial for assessing cardiac health.

[0074] The chamber segmenter employs advanced image segmentation techniques, particularly deep convolutional neural networks (ResNet), to automatically segment cardiac chambers in ultrasound images. This segmenter can identify the boundaries of each chamber and provide accurate regional information for the measurement of each chamber. For example, when segmenting the left ventricle, the segmenter can accurately delineate the inner and outer walls of the left ventricle, avoiding interference from other anatomical structures. The chamber segmenter can process complex ultrasound images, ensuring high accuracy and reliability in chamber segmentation, thus providing a basis for the extraction of specific dimensions. Figure 7 As shown, the overlap detection component matches the dimensional information annotated by the doctor with the chamber segmentation results, ensuring that the measured dimensions correspond to the correct cardiac chamber structure. Specifically, when the doctor annotates dimensions on the ultrasound image, the chamber segmenter automatically identifies the location of the annotated line segment and calculates its overlap with the chamber boundary. By comparing the matching of the annotated dimensions with the chamber contour, the system can accurately determine the chamber to which that dimension belongs, thereby achieving automatic classification and optimization of measurement data.

[0075] like Figure 8 As shown, the determination of geometric structure includes: in a single chamber structure, the dimensions that need to be measured are very few. The above process essentially completes the determination of most dimensions, especially for cases where a chamber has two dimensions, such as the major and minor axes. To accurately determine the dimensions, this invention uses the rectangular method or the aforementioned relative dimension method for differentiation:

[0076] Rectangular method: Based on the segmented contour of the chamber, fit the minimum bounding rectangle, and use its major and minor axes as the preliminary criteria for determining the major and minor axes, such as... Figure 9 As shown.

[0077] This embodiment provides an automatic size measurement method for cardiac ultrasound images, the core of which lies in the design and implementation of the size detection model and the involvement of the size location acquisition model. Specifically, it includes the following steps:

[0078] Data acquisition and preprocessing: Acquire ultrasound image data with doctor annotations, extract images containing key size information, and perform standardization processing.

[0079] Training of the size detection model: The YOLO11 model is used to identify and train the size annotations of doctors. The core objective is to accurately obtain the annotation endpoints and determine the size annotation pair to which the endpoints belong. The specific annotation content includes: endpoints, line segments, text regions, ellipses, lines, and angles.

[0080] Auxiliary model training: mainly for training ultrasonic section classification and ultrasonic chamber segmentation models.

[0081] Dimension detection of fixed-section: The trained model is used to automatically detect key dimensions in the image and convert them into actual physical dimensions by combining the scale bar to ensure measurement accuracy.

[0082] Measurement result output: Generates measurement reports and automatically enters the results into the system, reducing human transcription errors and improving diagnostic efficiency.

[0083] This embodiment includes a method for size correspondence of a single page: it includes a target detection model that detects all size positions and size labels marked by doctors; and it makes an ordered correspondence between the size pixel length and the real values ​​in the size table recognized by OCR.

[0084] Determining the size of content on a single page: Using a classification and segmentation model, the size is assigned to the chamber, and the size content is determined by the geometry or length of the chamber.

[0085] This embodiment is used for the automatic acquisition of annotation content of two-dimensional cardiac ultrasound fixed-frame sections; the automatic acquisition of measurement annotation of fixed-frame sections of ultrasound scenes; and the automatic data acquisition of manual annotation of medical images in dynamic and static frames.

[0086] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for automatic acquisition of two-dimensional cardiac ultrasound slice annotation data, characterized in that, The method comprises: Input the doctor's fixed ultrasonic section image into the trained size detection model and the auxiliary model, the size detection model outputs the key size label in the ultrasonic section image, including endpoints, line segments, text regions, ellipses, straight lines and angles; The auxiliary model outputs the section, chamber and geometric direction of the key size in the ultrasonic section image; Use OCR technology to extract the corresponding numerical value from the key size label, and sort the length of the line segment in the key size label, and according to the scale information in the image, the detected size label is converted from pixel unit to actual physical size and sorted; According to the sorting result, output multiple sets of size information of the ultrasonic section image, and generate a measurement report; The training process of the trained size detection model comprises: Collect pre-labeled ultrasonic image data, which covers multiple ultrasonic examination scenarios, obtain the ultrasonic image labeled by doctors, and the labeling content includes key size information, and pre-process the ultrasonic image data, wherein the key size information includes line segments, text regions, ellipses, straight lines and angles; Detailed labeling is performed on the pre-processed ultrasonic image, and the labeling content includes endpoints, line segments, text regions, ellipses, straight lines and angles; YOLOv11 model is used to identify and train the labeled ultrasonic image size, in the training process, the parameters of the model are optimized, and the trained model is evaluated using a validation set to check the recognition ability of the model for different labeling contents, and according to the evaluation result, the model parameters are adjusted or the training strategy is optimized; The auxiliary model includes an ultrasonic section classification model and an ultrasonic chamber segmentation model; The ultrasonic section classification model is used to automatically identify the section type of the ultrasonic image; The ultrasonic chamber segmentation model is used to segment the chamber region in the ultrasonic image; The OCR technology uses a convolutional neural network model combined with a long short-term memory network structure for training to improve the accuracy of text region size data extraction in images; The size label information output by the target detection model also includes a confidence score of the labeled region, which is used to represent the reliability of the detection result.

2. The method of claim 1, wherein, After converting the detected size label from pixel unit to actual physical size and sorting, it also includes: Match the detected key size label to ensure that each labeling content can correctly correspond to the actual physical size, and verify the matching result to check whether there is an error or omission.

3. The method of claim 1, wherein, The pre-processing of the ultrasonic image data comprises: extracting the image containing key size information from the ultrasonic image data, pre-processing the image containing key size information to ensure consistent image quality, and converting the image to a unified format and resolution.

4. The method of claim 1, wherein, The labeled size information contained in the training data set includes multiple size measurement labels of the heart chamber, including the internal diameter and wall thickness key size of the ventricle and atrium.

5. A two-dimensional cardiac ultrasound slice annotation data automatic acquisition system, characterized in that, It comprises: The detection module is configured to input the doctor's fixed ultrasonic section image into a trained size detection model and an auxiliary model, the size detection model outputs key size labels in the ultrasonic section image, including endpoints, line segments, text regions, ellipses, straight lines and angles, and the auxiliary model outputs the section, chamber and geometric direction of the key size in the ultrasonic section image; The recognition module is configured to extract corresponding numerical values from the key size labels by using an OCR technique, sort the lengths of line segments in the key size labels, and convert the detected size labels from pixel units to actual physical sizes according to scale information in the image and sort the sizes; The output module is configured to output multiple groups of size information of the ultrasonic section image according to the sorting results, and generate a measurement report; The training process of the trained size detection model includes: Collecting pre-labeled ultrasonic image data, the ultrasonic image data covers multiple ultrasonic examination scenarios, obtaining ultrasonic images labeled by doctors, the labeling content includes key size information, and the ultrasonic image data is preprocessed, wherein the key size information includes line segments, text regions, ellipses, straight lines and angles; Detailed labeling is performed on the preprocessed ultrasonic images, and the labeling content includes endpoints, line segments, text regions, ellipses, straight lines and angles; A YOLOv11 model is used to identify and train the size of the labeled ultrasonic images, in the training process, the parameters of the model are optimized, and the trained model is evaluated using a validation set to check the recognition ability of the model for different labeling contents, according to the evaluation results, the model parameters are adjusted or the training strategy is optimized; The auxiliary model includes an ultrasonic section classification model and an ultrasonic chamber segmentation model; The ultrasonic section classification model is used to automatically identify the section type of the ultrasonic image; The ultrasonic chamber segmentation model is used to segment the chamber region in the ultrasonic image; The OCR technique uses a convolutional neural network model combined with a long short-term memory network structure for training to improve the accuracy of text region size data extraction in images; The size label information output by the target detection model also includes a confidence score of the labeled region, and the confidence score is used to represent the reliability of the detection result.

Images