Method and system for recognizing examination site and probe posture based on ultrasound volume marker labeling map, and storage medium
By constructing an ultrasound target dataset and training a lightweight model, the target regions and probe orientation in ultrasound images are automatically identified, solving the problems of time-consuming and labor-intensive ultrasound image annotation and insufficient data utilization, and achieving efficient and unified annotation and data utilization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-27
- Publication Date
- 2026-05-29
AI Technical Summary
Current ultrasound image annotation relies on manual operation, which is time-consuming and labor-intensive, and is subject to subjective differences. Furthermore, the lack of standardized annotation for ultrasound images makes it difficult to effectively utilize the data, and body landmark information is not fully utilized.
By constructing a multi-vendor ultrasound target dataset, a lightweight target detection network and a probe posture key point regression model are trained to automatically identify target regions, anatomical locations, and probe postures, generating standardized labels.
It has achieved automation and consistency in ultrasound image annotation, improved annotation efficiency, fully explored the value of massive amounts of data, adapted to equipment from multiple manufacturers, and reduced the cost of clinical implementation.
Smart Images
Figure CN122115942A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing technology, specifically to a method, system, and storage medium for recognizing examination sites and probe postures based on ultrasound body marker mapping. Background Technology
[0002] Ultrasound imaging, as a non-invasive, real-time, and low-cost medical imaging technique, is widely used in various clinical examination areas, including the abdomen, breast, heart, and blood vessels. With the popularization of ultrasound equipment and the development of digital storage, a large amount of ultrasound image data has been accumulated in the clinical and research fields. This data is of great value in medical research, artificial intelligence model training, and the development of auxiliary diagnostic systems.
[0003] The current technology for annotating and utilizing ultrasound images presents the following problems: 1. Annotation of ultrasound images relies on manual operation: experienced sonographers or annotators examine each image individually, subjectively judging its corresponding anatomical location, standard section, and probe orientation, and then inputting the data into an annotation platform. This method is not only time-consuming, labor-intensive, and costly, but also suffers from significant subjective differences among annotators, making it difficult to ensure consistency and objectivity in annotation on large-scale datasets. 2. Massive amounts of data are difficult to utilize effectively: Deep learning models increasingly demand a greater quantity and diversity of training samples. Existing ultrasound image data, such as historical data stored in hospital PACS systems or ultrasound images from publicly available online platforms, are often saved only in static image format, retaining only the image itself and lacking clear and standardized annotations of examination sites, scanning sections, or body surface locations. This makes it difficult to effectively utilize these images in subsequent data processing, analysis, and model training. 3. Ultrasound body marker information is not fully utilized: In standard ultrasound examination procedures, to clarify the spatial attributes of the recorded image, the operator usually overlays ultrasound body markers (Body Marks) on specific areas of the ultrasound image. This body landmark primarily contains two key pieces of visualized information: first, an anatomical map, which uses simplified lines to outline the anatomical structures of specific examination sites (such as the thyroid, liver, kidneys, and breasts), providing a macroscopic anatomical background for subsequent image interpretation; and second, a probe indicator, superimposed on the anatomical map, precisely characterizing the ultrasound probe's contact position and orientation on the body surface at the moment of scanning. While this body landmark information provides crucial reference for annotators to understand the spatial attributes of images in clinical practice, in current technology, it is mainly used for manual interpretation. Its informational value has not been fully realized, and it has not been applied to tasks such as automatic image analysis, examination site identification, or data re-annotation, resulting in a waste of resources. Summary of the Invention
[0004] To address the shortcomings of the prior art, this invention provides a method, system, and storage medium for identifying examination sites and probe postures based on ultrasound body marker maps. This method automatically locates body marker regions, identifies anatomical marker categories, and analyzes the position and orientation of probe indicators using only the body marker information superimposed on ultrasound images. This allows for the automatic determination of corresponding examination sites and / or scanning sections without relying on grayscale image content, reducing the manual cost of large-scale ultrasound image data annotation and processing, and improving annotation consistency and utilization efficiency.
[0005] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows: On the one hand, the present invention provides a method for recognizing examination sites and probe postures based on ultrasound body marker maps, including a training phase and an inference and automatic annotation phase, the specific steps of which are as follows: (1) Training phase: S101. Construct a multi-vendor ultrasound landmark dataset: Collect raw ultrasound images containing landmark maps, with data sources covering major ultrasound equipment manufacturers; perform manual cleaning and fine annotation on the dataset, annotating each image with the landmark region bounding box, anatomical location category label, and probe posture key point label, wherein the probe posture key point label includes the positioning point P. A and direction point P B ; S102. Design and train the "integrated model of body landmark localization and anatomical classification": Use a lightweight target detection network, input a pre-processed complete ultrasound image, output the coordinates of the body landmark bounding box, the confidence level of the body landmark presence and the anatomical part category label, and crop out the body landmark ROI image; S103. Design and train the "Probe Attitude Keypoint Regression Model": For the target ROI image cropped in step S102, use a multi-scale feature extractor to extract features, and output the keypoint coordinate vector through regression prediction task head. V ; S104. Joint Model Evaluation and Deployment: The two models obtained in steps S102 and S103 are called in series to perform end-to-end inference on ultrasound images containing body targets, evaluate core indicators and determine the confidence threshold for body target detection. (2) Reasoning and automatic annotation stage: S201. Image Acquisition and Preprocessing: Batch acquire ultrasound images containing body marker maps and perform preprocessing operations consistent with the training phase. S202, Joint Inference: Execute two pre-trained models in series to complete body landmark localization, anatomical part classification and probe pose key point extraction; S203, Bodily Target Pose Analysis and Standard View Label Generation: Calculating Direction Vectors and Direction Angles Based on Keypoint Coordinates Map semantic direction labels, combining anatomical location categories and keypoint coordinate vectors. V Generate standard view labels; S204. Structured data generation and storage: Linking and storing image data with standardized annotation information.
[0006] As a preferred technical solution, the mainstream ultrasound equipment manufacturers include, but are not limited to, GE, Philips, and Mindray, and the anatomical site category labels include, but are not limited to, liver, kidney, breast, thyroid, and carotid artery.
[0007] As a preferred technical solution, in step S102, the lightweight target detection network is YOLO or MobileNet.
[0008] As a preferred technical solution, the total loss of the lightweight target detection network consists of the object category classification loss. Confidence loss and bounding box regression loss The weighted composition and the total loss expression are as follows: in, , , The weighting coefficients for each loss; the classification loss for the target category. Cross-entropy loss and confidence loss are used. Binary classification cross-entropy loss; bounding box regression loss. Use CIoU, GIoU, or Smooth L1 loss.
[0009] As a preferred technical solution, in step S103, the multi-scale feature extractor is a convolutional neural network or a Vision Transformer network, which extracts a high-dimensional geometric feature vector containing the position and orientation of the probe indicator icon. The regression prediction task head is used to connect a global average pooling layer and several fully connected layers after the output feature vector, and output a keypoint coordinate vector containing four floating-point numbers. V = (x1, y1; x2, y2), where (x1, y1) and (x2, y2) represent the positioning points P indicated by the probe markers. A and direction point P B Normalized position in the ROI image coordinate system.
[0010] As a preferred technical solution, in step S203, the direction vector and direction angle are calculated based on the coordinates of the key points. The method for mapping semantic direction labels is as follows: Based on key point coordinate vectorsV = (x1, y1; x2, y2), construct the direction vector in the normalized coordinate system of the target ROI image: Calculate the direction angle: The mapping rules for angle θ include: θ = 0° is mapped to a cross section, θ = 90° is mapped to a longitudinal section. The mapping is such that the probe is pointed to the patient's left side. The mapping is such that the probe is pointing to the right side of the patient.
[0011] As a preferred technical solution, in step S203, the anatomical site category label and key point coordinate vector are combined. V The method for generating standard view labels is as follows: S2031, Based on the location point P A (x1, y1) Determine the relative position of the probe in the human body outline and infer the scanning area of the body surface; S2032, According to the direction vector and angle Determine the direction of the cut; S2033, Label the anatomical part category y part The combination of the surface scan area and the section direction is mapped to the final standard view label y. view .
[0012] As a preferred technical solution, in step S204, the annotation information includes organ category, body surface area, and probe direction, and is stored using structured fields.
[0013] On the other hand, the present invention also provides a system for recognizing examination sites and probe postures based on ultrasound body markers, used to implement the aforementioned method for recognizing examination sites and probe postures based on ultrasound body markers, including: The dataset construction module collects raw ultrasound images with body landmarks covering major ultrasound equipment manufacturers. Each image is labeled with a bounding box for the body landmark region, an anatomical location category label, and probe posture key point labels. The probe posture key point labels include the location point P. A and direction point P B ; The body landmark localization and classification model training module is used to build a lightweight target detection network. It takes a pre-processed complete ultrasound image as input and outputs the coordinates of the body landmark bounding box, the confidence level of the body landmark presence, and the anatomical location category label, and crops the body landmark ROI image. The probe attitude key point regression model training module receives the target ROI image output by the target localization and classification model training module, constructs a multi-scale feature extractor based on convolutional neural networks or Vision Transformer to extract features, and outputs the localization point P through the regression prediction task head. A and direction point P B The normalized coordinate vector; The model evaluation and deployment module is used to connect the target localization and classification model and the probe posture key point regression model to perform end-to-end inference, evaluate core indicators, and determine the confidence threshold for target detection. The image preprocessing module is used to acquire ultrasound images containing body marker maps in batches and perform preprocessing operations. The joint inference module is used to link the pre-trained body landmark localization and classification model and the probe posture key point regression model to complete body landmark localization, anatomical site classification and probe posture key point extraction. The pose parsing and label generation module calculates the direction vector and angle θ based on the extracted probe pose keypoint coordinates, maps the angle θ to a semantic direction label, and combines the anatomical site category label and the keypoint coordinate vector. V Generate standardized view labels; The structured storage module is used to associate and store ultrasound image data with standardized annotation information, including organ category, body surface area, and probe orientation.
[0014] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon for implementing the aforementioned method for recognizing examination sites and probe postures based on ultrasound body markers, or for implementing the aforementioned system for recognizing examination sites and probe postures based on ultrasound body markers.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention provides a method and system for identifying examination sites and probe postures based on ultrasound body marker mapping, which can significantly reduce labor costs and improve annotation consistency: there is no need to manually annotate examination sites and scan sections one by one. The annotation process is automated through automatic parsing of body marker information, which not only improves annotation efficiency but also avoids subjective differences in manual annotation and improves annotation consistency.
[0016] The present invention provides a method and system for identifying examination sites and probe postures based on ultrasound body marker mapping, which can fully tap the value of massive ultrasound data: it solves the problem of lack of standardized annotations in historical ultrasound images, and supplements massive unlabeled data with standardized labels through automated annotation, making it directly usable for medical research and artificial intelligence model training, maximizing the utilization rate of data resources; it is the first to realize the automated value transformation of ultrasound body marker information: it breaks through the limitation that body marker information is only used for manual interpretation, and uses it as the core basis for automated annotation, and realizes the full-link automation of body marker positioning, anatomical classification and probe posture analysis through dual-model joint reasoning.
[0017] The present invention provides a method and system for identifying examination sites and probe postures based on ultrasound body marker mapping. It has strong compatibility with multiple manufacturers and high clinical versatility: the dataset covers mainstream ultrasound equipment manufacturers, and the model can be adapted to different body marker styles after being trained with data from multiple manufacturers. It does not require customized development for a single manufacturer, thus reducing the cost of clinical implementation. Attached Figure Description
[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 This is a flowchart illustrating the method of the present invention; Figure 2 This is a schematic diagram indicating the position of the body marker in a longitudinal section ultrasound image of the thyroid gland; Figure 3 This refers to the probe indicator in the body marker. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0021] A method for recognizing examination sites and probe postures based on ultrasound body marker mapping includes a training phase and an inference and automatic annotation phase, such as... Figure 1 As shown, the specific steps are as follows: (1) Training phase: Step 101: Construct a multi-vendor ultrasound landmark dataset Collect a large number of raw ultrasound images containing body markers. Data sources should cover major ultrasound equipment manufacturers, such as GE, Philips, and Mindray, to ensure sample diversity. The dataset will be manually cleaned and finely annotated; each image should include the following annotation information: The bounding box (BBox) of the volumetric region: indicates the coordinate range (x, y) of the volumetric region in the original drawing. min , y min , x max ,y max ), used to train the detection model; Anatomical Site Category Labels (y part Categories such as liver, kidneys, breast, thyroid, and carotid artery; Probe attitude key point label (y pose For the "probe indicator icon" in the body marker pattern, mark two key feature points, positioning point P. A and direction point P B Location point P A Point P represents the center point of contact between the probe icon and the human body. B The directional indicator endpoint of the probe icon (such as the light spot or marker end), such as Figure 3 As shown.
[0022] Step 102: Design and train an integrated model for body landmark localization and anatomical classification. A lightweight object detection network, such as YOLO or MobileNet, is used as input. The input is a preprocessed complete ultrasound image, and the output includes the coordinates of the object's bounding box, the object's presence confidence score, and the anatomical location category label. The bounding box coordinates represent the object's position in the image, the presence confidence score is the probability that the region contains the object, and the anatomical location category label is the anatomical location category corresponding to the object. Preprocessing includes grayscale conversion and noise reduction.
[0023] Preferably, the total loss of the lightweight object detection network consists of the object category classification loss. Confidence loss and bounding box regression loss Weighted composition; loss of classification by body category Cross-entropy loss and confidence loss can be applied to the predicted and true class labels. Binary cross-entropy loss (BCELoss) can be used to distinguish between "void labels" and "non-void labels"; bounding box regression loss can also be used. Losses such as CIoU, GIoU, and SmoothL1 can be used to constrain the overlap between the predicted bounding box and the labeled bounding box.
[0024] The total loss expression is: in, , , These are the weighting coefficients for each type of loss.
[0025] The training objective of the integrated model for body landmark localization and anatomical classification is to enable the model to accurately locate and crop out ROI images of various types of organs containing only line patterns from a complex ultrasound echo background.
[0026] S103. Design and train the "Probe Attitude Keypoint Regression Model": For the target ROI image cropped in step S102, construct an end-to-end keypoint regression network, focusing on resolving the precise geometric attitude of the probe within the target. The specific method is as follows: S1031. Input the ROI image cropped in step S102; S1032. Employ multi-scale feature extractors, such as convolutional neural networks or VisionTransformer network structures based on self-attention mechanisms, to extract high-dimensional geometric feature vectors containing the position and orientation of the probe indicator icon; S1033. Output keypoint coordinate vectors from the regression prediction task header. V ; The structure of the regression prediction task is as follows: after the feature vector output by S1032, a global average pooling layer and several fully connected layers are connected to output a keypoint coordinate vector containing four floating-point numbers. V = (x1, y1; x2, y2), (x1, y1) and (x2, y2) represent the positioning points P indicated by the probe. A and direction point P B Normalized position in the ROI image coordinate system.
[0027] The probe attitude key point regression model uses Smooth L1 Loss to minimize the Euclidean distance error between the predicted coordinate points and the manually labeled real coordinate points.
[0028] Step S104: Joint Model Evaluation and Deployment On independent validation and test sets, the two models obtained in steps S102 and S103 are cascaded and used to perform end-to-end inference on ultrasound images containing body landmarks. Key metrics such as body landmark detection recall, anatomical site classification accuracy, and regression error of probe posture key points are evaluated. Based on the evaluation results, operational parameters such as the confidence threshold for body landmark detection are determined.
[0029] Meanwhile, based on the key point prediction results under different body style in multi-vendor data, statistical direction angles are calculated. Based on the distribution of the data, the angle ranges used in step S203 are calibrated and corrected to form a standardized body pose resolution threshold suitable for the target deployment environment.
[0030] (2) Reasoning and automatic annotation stage Step S201: Image Acquisition and Preprocessing Batch acquire ultrasound images containing body marker maps from an ultrasound workstation, PACS system, or public dataset; input images can be raw screenshots or exported static frame images. Perform the same preprocessing operations as during the training phase on each image to generate a normalized image tensor that conforms to the model input specifications.
[0031] Step S202: Automatic localization of body landmark area, anatomical recognition, and joint inference of probe posture key points The ultrasound image to be labeled is input into the deep learning system, and two pre-trained models are executed in series to perform a full-link analysis of the body landmark map in the image. In one forward inference process, the automatic localization of the body landmark region and the classification of anatomical parts are completed first, and then the localized region is subjected to fine feature extraction, outputting the key point coordinates of the probe posture.
[0032] Step S203: Body landmark pose analysis and standardized label generation First, determine the rules for resolving the pose of the bodily target: Based on the key point coordinate vector predicted in step S103 V = (x1, y1; x2, y2), where (x1, y1) is the location point P. A (x2, y2) is the direction point P. B Construct direction vectors in the normalized coordinate system of the ROI image: Further calculation of direction and angle: Based on the coordinate system conventions of different manufacturers' body icons and the standard human orientation settings, several angle intervals are pre-divided and mapped to discrete semantic direction labels, for example: Methods for generating standard view labels: In the anatomical location category label y part and key point coordinate vector V Based on this, semantic parsing of the probe's body landmark posture is performed, and standardized examination site / section labels are generated. Specifically, this includes: S2031, Based on the location point P A (x1, y1) determines the relative position of the probe in the body contour of the target human body, thereby inferring the body surface scanning area, such as the upper right abdomen, upper left abdomen, anterior neck, and parasternal region; S2032, According to the direction vector and angle The direction of the cut is determined as "longitudinal cut / transverse cut / oblique cut" and semantic direction such as "pointing towards the head / foot / left / right side"; S2033, Label the anatomical part category y part The combination of the surface scan area and the section direction is mapped to the final standard view label y. view For example, "longitudinal thyroidectomy", etc. Figure 2 .
[0033] Step S204: Structured Data Generation and Storage Based on the determination result obtained in step S203, corresponding annotation information for the body surface location or examination site is automatically generated for the current ultrasound image and stored in association with the image data. The annotation information may include structured fields such as organ category, body surface region, and probe direction, for subsequent image management, retrieval, or analysis.
[0034] This invention also provides a system for recognizing examination sites and probe postures based on ultrasound body marker mapping, used to implement the above method, specifically including: The dataset construction module is used to collect raw ultrasound images containing body landmarks from major ultrasound equipment manufacturers (such as GE, Philips, Mindray, etc.). Each image is labeled with the body landmark region bounding box, anatomical location category label (such as liver, kidney, breast, thyroid, carotid artery, etc.), and probe posture key point label. The probe posture key point label includes the positioning point P. A and direction point P B Location point P A Point P represents the center point of contact between the probe icon and the human body. B The directional indicator endpoint of the probe icon (such as a light spot or marker end).
[0035] The body landmark localization and classification model training module is used to build a lightweight object detection network, such as YOLO or MobileNet. It takes a pre-processed complete ultrasound image as input and outputs the coordinates of the body landmark bounding box, the confidence level of the body landmark presence, and the anatomical location category label. It also crops out the ROI images of various types of organs containing only line patterns. The coordinates of the body landmark bounding box are the position of the body landmark in the image, the confidence level of the body landmark presence is the probability that the region contains the body landmark, and the anatomical location category label is the anatomical location category corresponding to the body landmark.
[0036] The probe posture keypoint regression model training module receives the target ROI image output by the target localization and classification model training module, constructs a multi-scale feature extractor based on a convolutional neural network or Vision Transformer to extract features, and outputs keypoint coordinate vectors through the regression prediction task head. The multi-scale feature extractor extracts a high-dimensional geometric feature vector containing the position and orientation of the probe indicator icon. The structure of the regression prediction task is as follows: after the feature vector output by S1032, a global average pooling layer and several fully connected layers are connected to output a keypoint coordinate vector containing four floating-point numbers. V = (x1, y1; x2, y2), (x1, y1) and (x2, y2) represent the positioning points P indicated by the probe. A and direction point P B Normalized position in the ROI image coordinate system.
[0037] The model evaluation and deployment module is used to connect the object localization and classification model and the probe posture keypoint regression model for end-to-end inference, evaluate core metrics, and determine the confidence threshold for object detection. Core metrics include object detection recall, anatomical site classification accuracy, and regression error of probe posture keypoints. It is also used to statistically analyze orientation and angle based on keypoint prediction results from different object styles across multiple vendors' data. Based on the distribution of the data, the angle ranges used in step S203 are calibrated and corrected to form a standardized body pose resolution threshold suitable for the target deployment environment.
[0038] The image preprocessing module is used to batch acquire ultrasound images containing body marker maps from ultrasound workstations, PACS systems, or public datasets. The input images can be original screenshots or exported static frame images; perform preprocessing operations; and generate standardized image tensors that conform to the model input specifications.
[0039] The joint inference module is used to link the pre-trained body landmark localization and classification model and the probe posture key point regression model to complete the automatic localization of body landmark regions, classification of anatomical parts, and extraction of probe posture key points. The pose parsing and label generation module calculates the direction vector and angle θ based on the extracted probe pose keypoint coordinates, maps the angle θ to a semantic direction label, and combines the anatomical site category label and the keypoint coordinate vector. V Generate standardized view labels, such as "thyroid longitudinal section".
[0040] The structured storage module is used to associate and store ultrasound image data with standardized annotation information, which includes structured fields such as organ category, body surface region, and probe direction, for subsequent image management, retrieval, or analysis.
[0041] The present invention also provides a computer-readable storage medium having a computer program stored thereon for implementing the aforementioned method for recognizing examination sites and probe postures based on ultrasound body markers, or for implementing the aforementioned system for recognizing examination sites and probe postures based on ultrasound body markers.
[0042] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.
Claims
1. A method for identifying examination sites and probe posture based on ultrasound body marker mapping, characterized in that: The process includes a training phase and an inference and automatic annotation phase, with the following specific steps: (1) Training phase: S101. Construct a multi-vendor ultrasound landmark dataset: Collect raw ultrasound images containing landmark markers, with data sources covering major ultrasound equipment manufacturers; perform manual cleaning and fine annotation on the dataset, annotating each image with the landmark region bounding box, anatomical location category label, and probe posture key point label, wherein the probe posture key point label includes the positioning point P. A and direction point P B ; S102. Design and train the "integrated model of body landmark localization and anatomical classification": Use a lightweight target detection network, input a pre-processed complete ultrasound image, output the coordinates of the body landmark bounding box, the confidence level of the body landmark presence and the anatomical part category label, and crop out the body landmark ROI image; S103. Design and train the "Probe Attitude Keypoint Regression Model": For the target ROI image cropped in step S102, use a multi-scale feature extractor to extract features, and output the keypoint coordinate vector through regression prediction task head. V ; S104. Joint Model Evaluation and Deployment: The two models obtained in steps S102 and S103 are called in series to perform end-to-end inference on ultrasound images containing body targets, evaluate core indicators and determine the confidence threshold for body target detection. (2) Reasoning and automatic annotation stage: S201. Image Acquisition and Preprocessing: Batch acquire ultrasound images containing body markers and perform preprocessing operations consistent with the training phase. S202, Joint Inference: Execute two pre-trained models in series to complete body landmark localization, anatomical part classification and probe pose key point extraction; S203, Bodily Target Pose Analysis and Standard View Label Generation: Calculating Direction Vectors and Direction Angles Based on Keypoint Coordinates Map semantic direction labels, combining anatomical location categories and keypoint coordinate vectors. V Generate standard view labels; S204. Structured data generation and storage: Linking and storing image data with standardized annotation information.
2. The method for identifying examination sites and probe postures based on ultrasound body marker mapping according to claim 1, characterized in that: The mainstream ultrasound equipment manufacturers include, but are not limited to, GE, Philips, and Mindray, and the anatomical site category labels include, but are not limited to, liver, kidney, breast, thyroid, and carotid artery.
3. The method for identifying examination sites and probe postures based on ultrasound body marker mapping according to claim 1, characterized in that: In step S102, the lightweight target detection network is YOLO or MobileNet.
4. The method for identifying examination sites and probe postures based on ultrasound body marker mapping according to claim 3, characterized in that: The total loss of the lightweight object detection network consists of the target category classification loss. Confidence loss and bounding box regression loss The weighted composition and the total loss expression are as follows: in, , , The weighting coefficients for each loss; the classification loss for the target category. Cross-entropy loss and confidence loss are used. Binary classification cross-entropy loss; bounding box regression loss. Use CIoU, GIoU, or Smooth L1 loss.
5. The method for identifying examination sites and probe postures based on ultrasound body marker mapping according to claim 1, characterized in that: In step S103, the multi-scale feature extractor is a convolutional neural network or a Vision Transformer network, which extracts a high-dimensional geometric feature vector containing the position and orientation of the probe indicator icon. The regression prediction task head is used to connect a global average pooling layer and several fully connected layers after the output feature vector, and output a keypoint coordinate vector containing four floating-point numbers. V = (x1, y1; x2, y2), where (x1, y1) and (x2, y2) represent the positioning points P indicated by the probe. A and direction point P B Normalized position in the ROI image coordinate system.
6. The method for identifying examination sites and probe postures based on ultrasound body marker mapping according to claim 5, characterized in that: In step S203, the direction vector and direction angle are calculated based on the key point coordinates. The method for mapping semantic direction labels is as follows: Based on key point coordinate vectors V = (x1, y1; x2, y2), construct the direction vector in the normalized coordinate system of the target ROI image: Calculate the direction angle: The mapping rules for angle θ include: θ = 0° is mapped to a cross section, θ = 90° is mapped to a longitudinal section. The mapping is such that the probe is pointed to the patient's left side. The mapping is such that the probe is pointing to the right side of the patient.
7. The method for identifying examination sites and probe postures based on ultrasound body marker mapping according to claim 1, characterized in that: In step S203, the anatomical site category labels and key point coordinate vectors are combined. V The method for generating standard view labels is as follows: S2031, Based on the location point P A (x1, y1) Determine the relative position of the probe in the human body outline of the body target, and infer the body surface scanning area; S2032, According to the direction vector and angle Determine the direction of the cut; S2033, Label the anatomical part category y part The combination of the surface scan area and the section direction is mapped to the final standard view label y. view .
8. The method for identifying examination sites and probe postures based on ultrasound body marker mapping according to claim 1, characterized in that: In step S204, the annotation information includes organ category, body surface region, and probe orientation, and is stored using structured fields.
9. A system for recognizing examination sites and probe postures based on ultrasound body marker mapping, characterized in that, The system is used to implement the examination site and probe posture recognition method based on ultrasound body marker mapping as described in claim 1, including: The dataset construction module collects raw ultrasound images with body landmarks covering major ultrasound equipment manufacturers. Each image is labeled with a bounding box for the body landmark region, an anatomical location category label, and probe posture key point labels. The probe posture key point labels include the location point P. A and direction point P B ; The body landmark localization and classification model training module is used to build a lightweight target detection network. It takes a pre-processed complete ultrasound image as input and outputs the coordinates of the body landmark bounding box, the confidence level of the body landmark presence, and the anatomical location category label, and crops the body landmark ROI image. The probe attitude key point regression model training module receives the target ROI image output by the target localization and classification model training module, constructs a multi-scale feature extractor based on convolutional neural networks or Vision Transformer to extract features, and outputs the localization point P through the regression prediction task head. A and direction point P B The normalized coordinate vector; The model evaluation and deployment module is used to connect the target localization and classification model and the probe posture key point regression model to perform end-to-end inference, evaluate core indicators, and determine the confidence threshold for target detection. The image preprocessing module is used to acquire ultrasound images containing body markers in batches and perform preprocessing operations. The joint inference module is used to link the pre-trained body landmark localization and classification model and the probe posture key point regression model to complete body landmark localization, anatomical site classification and probe posture key point extraction. The pose parsing and label generation module calculates the direction vector and angle θ based on the extracted probe pose keypoint coordinates, maps the angle θ to a semantic direction label, and combines the anatomical site category label and the keypoint coordinate vector. V Generate standardized view labels; The structured storage module is used to associate and store ultrasound image data with standardized annotation information, including organ category, body surface area, and probe orientation.
10. A computer-readable storage medium, characterized in that: It stores a computer program for implementing the examination site and probe posture recognition method based on ultrasound body markers as described in any one of claims 1-8, or a computer program for implementing the examination site and probe posture recognition system based on ultrasound body markers as described in claim 9.