An ultrasound data set construction method and apparatus
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHENGDU TIANTIAN JINCHENG MEDICAL DEVICE TECHNOLOGY RESEARCH CO LTD
- Filing Date
- 2026-04-16
- Publication Date
- 2026-05-29
Smart Images
Figure CN122116048A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of ultrasound image processing and artificial intelligence technology, and in particular to a method and apparatus for constructing an ultrasound dataset. Background Technology
[0002] Ultrasound examinations are widely used in disease screening, diagnosis, and efficacy evaluation across various clinical departments due to their advantages such as being non-invasive, convenient, real-time, and low-cost. With the rapid development of artificial intelligence technology in the field of medical imaging, intelligent diagnostic models based on ultrasound images are emerging in large numbers. High-quality, standardized ultrasound datasets are the core foundation for the training, validation, and optimization of such models, directly determining their diagnostic accuracy and clinical applicability.
[0003] Currently, existing methods for constructing ultrasound datasets generally suffer from the following shortcomings: First, they only collect the ultrasound images themselves, ignoring crucial auxiliary information that is simultaneously displayed on the ultrasound screen during the ultrasound examination and directly affects the image characteristics. This includes information such as the brand and model of the ultrasound machine, and imaging parameters like probe frequency, image gain, scanning depth, and dynamic range. This information directly impacts the imaging quality, texture features, and grayscale distribution of the ultrasound images. The lack of such information leads to insufficient dataset completeness, thus affecting the adaptability of intelligent models to ultrasound images from different machine models and with different parameters. Second, existing methods lack a standardized process from data acquisition, information recognition, feature extraction to dataset organization. Therefore, existing ultrasound image datasets suffer from inconsistent formats, non-standard annotations, and data redundancy, resulting in poor universality and reusability, making it difficult to meet the training needs of intelligent models in different scenarios. Therefore, there is an urgent need for a method and apparatus that can overcome these shortcomings and achieve comprehensive ultrasound data acquisition, accurate identification of key information, efficient feature extraction, and standardized dataset construction. Summary of the Invention
[0004] To address the shortcomings of the existing technology, this invention provides a method and apparatus for constructing ultrasound datasets, enabling the efficient construction of standardized, high-quality ultrasound datasets and providing reliable support for the development of intelligent ultrasound diagnostic models.
[0005] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows: A method for constructing an ultrasound dataset, comprising the following steps: S1. Acquire the screen data of the ultrasound diagnostic instrument. The screen data includes an ultrasound image area and a synchronously displayed auxiliary information area. The auxiliary information area includes at least one key text information, such as the device model and imaging parameters. S2. Process the screen data to separate the ultrasound image area and the auxiliary information area; S3. Use a text recognition model to perform text recognition on the auxiliary information area to extract the key text information; use a deep learning feature extraction model to extract the deep features of the ultrasound image in the ultrasound image area; combine the key text information to annotate and correct the deep features of the ultrasound image to obtain an ultrasound image feature set that integrates multi-dimensional information. S4. Dataset Organization: Associate and store the original ultrasound images, extracted key text information, and ultrasound image feature sets to establish the correspondence between "ultrasound image - machine information - imaging parameters - image features" and generate a standardized ultrasound dataset.
[0006] As a preferred technical solution, acquiring the screen data of the ultrasound diagnostic instrument includes: directly connecting a high-definition screen acquisition card to the screen output interface of the ultrasound diagnostic instrument for real-time data acquisition.
[0007] As a preferred technical solution, in step S2, the ultrasound image region and the auxiliary information display region are segmented by an image thresholding algorithm.
[0008] As a preferred technical solution, in step S3, the text recognition model is used to perform text recognition on the auxiliary information area to extract the key text information, including: using an OCR text recognition model based on a convolutional neural network to recognize the text in the auxiliary information area and extracting information such as ultrasound machine model, probe model, probe frequency, scanning depth, gain, dynamic range, focus position, and frame rate; using a character correction model based on the Transformer architecture to correct the recognition results of the OCR text recognition model, and the character correction model is matched and verified by an ultrasound key information corpus.
[0009] As a preferred technical solution, in step S3, a deep learning feature extraction model is used to extract deep features of the ultrasound image in the ultrasound image region, including: The segmented ultrasound image regions were preprocessed as follows: median filtering algorithm was used to remove image noise, and linear gray-level normalization was used to map the image gray values to the [0,1] interval; the ultrasound image size was standardized to 512×512 pixels. The preprocessed ultrasound images are input into a trained improved deep learning feature extraction model to extract deep features, which include texture features, grayscale features, morphological features and lesion-related features of the ultrasound images. The improved deep learning feature extraction model is based on the SE-ResNet architecture and integrates an SE channel attention mechanism after each residual block.
[0010] As a preferred technical solution, step S4 also includes: cleaning the associated data and removing unqualified samples; classifying and labeling the data samples according to the preset labeling specifications and performing cross-validation; encapsulating the sorted samples using a unified data format to generate a standardized ultrasound dataset, and establishing a dataset index based on anatomical location, machine model, and lesion type.
[0011] As a preferred technical solution, step S5, dataset verification and optimization, is also included: selecting a portion of samples to verify the standardized ultrasound dataset and evaluate the dataset quality; for any issues discovered during verification, returning to the corresponding steps for optimization and adjustment; and performing version management after optimization to ensure dataset traceability.
[0012] As a preferred technical solution, the verification dimensions include completeness, accuracy, and standardization. Completeness verification checks whether the sample contains ultrasound images, key information, image features, and annotation information to ensure no information is missing. Accuracy verification evaluates recognition accuracy and annotation accuracy by comparing manual recognition results with machine recognition results, and manual annotation results with model feature extraction results. Standardization verification checks whether the sample format, size, and feature dimensions are consistent. The optimization and adjustment include optimizing the OCR text recognition model and fine-tuning the parameters of the deep learning feature extraction model. Version management includes labeling the version number, recording the update content and reasons, and forming a dataset update log.
[0013] An ultrasound dataset construction apparatus, comprising: The acquisition module is used to acquire the screen data of the ultrasound diagnostic instrument. The screen data includes an ultrasound image area and a synchronously displayed auxiliary information area. The auxiliary information area includes at least one key text information, such as the device model and imaging parameters. The processing module is used to process the screen data to separate the ultrasound image area and the auxiliary information area; The recognition module is used to perform text recognition on the auxiliary information area using a text recognition model in order to extract the key text information; The extraction module is used to extract deep features of the ultrasound image in the ultrasound image region using a deep learning feature extraction model. The association module is used to combine key text information to annotate and correct the deep features of ultrasound images, thereby obtaining an ultrasound image feature set that integrates multi-dimensional information. The processing module is used to associate and store the original ultrasound images, extracted key text information, and ultrasound image feature sets, establish the correspondence between "ultrasound image - machine information - imaging parameters - image features", and generate a standardized ultrasound dataset.
[0014] As a preferred technical solution, the recognition module is configured to: use a convolutional neural network-based OCR text recognition model to recognize the text in the auxiliary information region, and use a character correction model based on the Transformer architecture to correct the recognition result of the OCR text recognition model; the extraction module is configured to: use a deep learning feature extraction model based on the SE-ResNet architecture to extract deep features.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: The ultrasound dataset construction method of the present invention realizes comprehensive acquisition of ultrasound data. It not only acquires the ultrasound images themselves, but also simultaneously acquires key auxiliary information such as machine model and imaging parameters on the ultrasound screen. This solves the problem of insufficient completeness of existing datasets. By associating these key text information with the deep features of the images, the constructed dataset can reflect the impact of different machine models and imaging parameters on ultrasound images, thereby improving the universality and practicality of the dataset.
[0016] The ultrasound dataset construction method of this invention integrates text recognition technology and deep learning methods. On the one hand, it uses an OCR text recognition model combined with a character correction model to achieve accurate and automated recognition of key information on the ultrasound screen, replacing the traditional manual input method, greatly improving recognition efficiency and reducing labor costs. On the other hand, it uses an improved deep learning model to extract deep features from ultrasound images. Compared with traditional image processing algorithms, feature extraction is more accurate and comprehensive, effectively capturing subtle details in the images and providing more reliable feature support for intelligent diagnostic models.
[0017] The ultrasound dataset construction method of this invention establishes a standardized ultrasound dataset construction process. From data acquisition, information identification, feature extraction to data processing, verification and optimization, each step has clear operating specifications and quality control standards. This effectively solves the problems of inconsistent dataset formats, non-standard annotation, and data redundancy in existing datasets. The constructed dataset has a high degree of standardization and reusability and can be widely used for the training, verification and optimization of different types of ultrasound intelligent models.
[0018] The ultrasound dataset construction method of the present invention ensures the quality and traceability of the dataset through dataset verification and optimization steps and version management mechanism. It can continuously update and optimize the dataset according to clinical needs and technological development, extend the life of the dataset, and provide stable and reliable data support for the long-term development of ultrasound intelligent diagnostic technology. It has important clinical application value and industrialization prospects. Attached Figure Description
[0019] 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.
[0020] Figure 1 This is a flowchart illustrating the ultrasound dataset construction method of the present invention; Figure 2 This is a schematic diagram of ultrasonic screen data acquisition and region segmentation in this invention. Detailed Implementation
[0021] 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.
[0022] A method for constructing an ultrasound dataset, such as Figure 1 As shown, the steps include: S1. Acquire the screen data of the ultrasound diagnostic instrument. The screen data includes an ultrasound image area and a synchronously displayed auxiliary information area. The auxiliary information area includes at least one key text information, such as the device model and imaging parameters. S2. Process the screen data to separate the ultrasound image area and the auxiliary information area; S3. Use a text recognition model to perform text recognition on the auxiliary information area to extract the key text information; use a deep learning feature extraction model to extract the deep features of the ultrasound image in the ultrasound image area; combine the key text information to annotate and correct the deep features of the ultrasound image to obtain an ultrasound image feature set that integrates multi-dimensional information. S4. Dataset Organization: Associate and store the original ultrasound images, extracted key text information, and ultrasound image feature sets to establish the correspondence between "ultrasound image - machine information - imaging parameters - image features" and generate a standardized ultrasound dataset.
[0023] In a preferred embodiment, step S1, acquiring the screen data of the ultrasound diagnostic instrument, specifically includes: capturing the complete screen image displayed in real time during the examination of different models of ultrasound diagnostic instruments using a high-definition data acquisition device to obtain the raw ultrasound screen data. Preferably, a high-definition screen acquisition card with a resolution of not less than 3840×2160 is selected as the acquisition device, and the high-definition screen acquisition card is directly connected to the screen output interface of the ultrasound diagnostic instrument to achieve real-time, lossless acquisition of ultrasound screen data. Commonly used clinical ultrasound diagnostic instruments are selected to acquire ultrasound examination screen data of different anatomical sites such as the abdomen, heart, breast, and thyroid. During the acquisition process, the acquisition time, examination site, patient basic information (desensitized), ambient light intensity, and operating parameters of the screen acquisition card (such as sampling frequency and transmission rate) are recorded simultaneously. The raw ultrasound screen data is stored in PNG lossless format to ensure that ultrasound image details and auxiliary information are not lost. In this embodiment, 10,000 raw data samples are acquired.
[0024] In step S2, preprocessing the screen data includes: converting the image to grayscale, denoising, and enhancing contrast; segmenting the ultrasound image region and auxiliary information display region using an image thresholding algorithm, such as... Figure 2 As shown, the auxiliary information display area is located at the edge of the ultrasound screen (usually at the top, bottom, or right side of the image) and contains text information such as machine model, probe model, and imaging parameters. Morphological operations or convolutional neural networks (CNNs) are used to locate the text boxes, and the detected auxiliary information area is rotated, corrected, and binarized.
[0025] In a preferred embodiment, in step S3, the text recognition model is used to perform text recognition on the auxiliary information area to extract the key text information. This includes: using an OCR text recognition model based on a convolutional neural network (CNN) to recognize the text in the auxiliary information area and extracting key information such as the ultrasound machine model, probe model, probe frequency, scanning depth, gain, dynamic range, focus position, and frame rate. To address potential misrecognition issues caused by blurring, character overlap, or glare on the ultrasound screen text, a Transformer-based character correction model is introduced to correct the OCR text recognition model's results. The character correction model uses a constructed ultrasound key information corpus to match and verify the recognition results, ensuring that the accuracy of key information recognition is no less than 98%. For example, for a GE Logiq E9 ultrasound machine, the machine model is accurately identified as "Logiq E9," the probe model as "ML6-15," the probe frequency as 5MHz, and the scanning depth as 8cm. The recognition results are then associated and stored with the corresponding original ultrasound screen data.
[0026] In step S3, a deep learning feature extraction model is used to extract deep features of the ultrasound image in the ultrasound image region, including: The segmented ultrasound image regions are preprocessed: median filtering algorithm is used to further remove image noise, and linear gray-level normalization is used to map the image gray values to the [0,1] interval to eliminate gray-level differences under different acquisition conditions; all ultrasound image sizes are standardized to 512×512 pixels to ensure the uniformity of subsequent feature extraction.
[0027] Preprocessed ultrasound images are input into a trained improved deep learning feature extraction model. 512-dimensional deep features are extracted from the model's GAP layer. These deep features include multi-dimensional information such as texture, grayscale, morphological, and lesion-related features of the ultrasound images. The improved deep learning feature extraction model is based on an optimized SE-ResNet architecture. It introduces 3D convolutional layers to replace traditional 2D convolutional layers on top of ResNet18, adapting to the spatial feature extraction of ultrasound images. Simultaneously, an SE module (channel attention mechanism) is integrated after each residual block to enhance the ability to capture subtle lesion features. During model training, the PyTorch framework is used, with the acquired preprocessed ultrasound images as training samples. The batch size is set to 32, and the initial learning rate is 1×10⁻⁶. -4 The Adam optimizer and Cox negative biased likelihood loss function are used. The training lasts for 100 rounds. If the validation loss stabilizes for 15 consecutive rounds, an early stopping mechanism is triggered to avoid model overfitting.
[0028] Step S3 further includes converting the extracted key textual information such as ultrasound machine model and imaging parameters into feature vectors, and then fusing them with the deep feature vectors of the image to obtain an ultrasound image feature set with fused multi-dimensional information. The feature set of each sample contains deep image features and key information features. In this embodiment, a total of 10,000 feature samples are obtained.
[0029] Furthermore, step S4 includes cleaning the associated data: removing blurry, distorted, and redundant samples, deleting incorrectly identified samples, and finally retaining valid samples. In this embodiment, a total of 800 blurry, distorted, and redundant samples were removed, 200 incorrectly identified samples were deleted, and 9000 valid samples were retained. Following clinical ultrasound diagnostic standards and dataset annotation specifications, at least two senior ultrasound physicians were organized to classify and annotate the valid samples. The annotation content included anatomical location (abdomen, heart, etc.), lesion type (normal, benign lesion, malignant lesion), and image quality level (excellent, good, qualified). Cross-validation was performed after annotation to ensure that the annotation accuracy rate was not less than 99%. The processed samples were encapsulated in JSON format to generate a standardized ultrasound dataset. Simultaneously, a dataset index based on anatomical location, machine model, and lesion type was established to facilitate subsequent retrieval and access.
[0030] Furthermore, the process includes dataset validation and optimization. A subset of samples is selected to validate the standardized ultrasound dataset and assess its quality. For any issues discovered during validation, the corresponding steps are reviewed for optimization and adjustment. After optimization, version management is implemented to ensure dataset traceability. Preferably, 10% of the samples are randomly selected from the standardized ultrasound dataset as the validation set. Validation is performed across three dimensions: completeness, accuracy, and standardization. Completeness validation checks whether the samples contain ultrasound images, key information, image features, and annotation information to ensure no information is missing. Accuracy validation compares human recognition results with machine recognition results, and human annotation results with model feature extraction results to evaluate recognition and annotation accuracy. Standardization validation checks whether the sample format, size, and feature dimensions are consistent. In this embodiment, the validation results show that the dataset completeness is 100%, the key information recognition accuracy is 98.2%, the annotation accuracy is 99.1%, and the sample format, size, and feature dimensions are consistent, meeting the preset quality standards. For the 18 samples found to have imaging parameter recognition biases during the verification process, the ultrasound key information corpus was optimized using the character correction model in step S3, and the biases were corrected after re-identification. For the 12 samples with insufficiently accurate deep feature extraction, the parameters of the deep learning feature extraction model were fine-tuned in step S3, and deep features were extracted again. After optimization, the dataset was version-managed, labeled with version number V1.0, and the optimization content and reasons were recorded to form a dataset update log to ensure traceability.
[0031] The standardized ultrasound dataset constructed in this embodiment contains 9,000 valid samples, covering 3 ultrasound machine models, 4 anatomical locations, and 3 lesion types. It can be directly used for the training, verification, and optimization of ultrasound intelligent diagnostic models, effectively improving the diagnostic accuracy and clinical adaptability of the models.
[0032] The present invention also provides an ultrasound dataset construction apparatus, comprising: The acquisition module is used to acquire the screen data of the ultrasound diagnostic instrument. The screen data includes an ultrasound image area and a synchronously displayed auxiliary information area. The auxiliary information area includes at least one key text information, such as the device model and imaging parameters. The processing module is used to process the screen data to separate the ultrasound image area and the auxiliary information area; The recognition module is used to perform text recognition on the auxiliary information area using a text recognition model in order to extract the key text information; the recognition module is configured to: use an OCR text recognition model based on a convolutional neural network to recognize the text in the auxiliary information area, and use a character correction model based on a Transformer architecture to correct the recognition result of the OCR text recognition model; An extraction module is used to extract deep features of ultrasound images in the ultrasound image region using a deep learning feature extraction model; the extraction module is configured to extract deep features using a deep learning feature extraction model based on the SE-ResNet architecture. The association module is used to combine key text information to annotate and correct the deep features of ultrasound images, thereby obtaining an ultrasound image feature set that integrates multi-dimensional information. The processing module is used to associate and store the original ultrasound images, extracted key text information, and ultrasound image feature sets, establish the correspondence between "ultrasound image - machine information - imaging parameters - image features", and generate a standardized ultrasound dataset.
[0033] 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 constructing an ultrasound dataset, characterized by: steps include: S1. Acquire the screen data of the ultrasound diagnostic instrument. The screen data includes an ultrasound image area and a synchronously displayed auxiliary information area. The auxiliary information area includes at least one key text information, such as the device model and imaging parameters. S2. Process the screen data to separate the ultrasound image area and the auxiliary information area; S3. Use a text recognition model to perform text recognition on the auxiliary information area to extract the key text information; use a deep learning feature extraction model to extract the deep features of the ultrasound image in the ultrasound image area; combine the key text information to annotate and correct the deep features of the ultrasound image to obtain an ultrasound image feature set that integrates multi-dimensional information. S4. Dataset Organization: Associate and store the original ultrasound images, extracted key text information, and ultrasound image feature sets to establish the correspondence between "ultrasound image - machine information - imaging parameters - image features" and generate a standardized ultrasound dataset.
2. The method for constructing an ultrasound dataset according to claim 1, characterized in that: The acquisition of screen data from the ultrasound diagnostic instrument includes: real-time data acquisition via a high-definition screen acquisition card directly connected to the screen output interface of the ultrasound diagnostic instrument.
3. The method for constructing an ultrasound dataset according to claim 1, characterized in that: In step S2, the ultrasound image region and the auxiliary information display region are obtained by segmenting the image using an image thresholding algorithm.
4. The method for constructing an ultrasound dataset according to claim 1, characterized in that: In step S3, the text recognition model is used to perform text recognition on the auxiliary information area to extract the key text information, including: using an OCR text recognition model based on a convolutional neural network to recognize the text in the auxiliary information area and extracting information such as ultrasound machine model, probe model, probe frequency, scanning depth, gain, dynamic range, focus position, and frame rate; and using a character correction model based on the Transformer architecture to correct the recognition results of the OCR text recognition model, with the character correction model being matched and verified by an ultrasound key information corpus.
5. The method for constructing an ultrasound dataset according to claim 1, characterized in that: In step S3, a deep learning feature extraction model is used to extract deep features of the ultrasound image in the ultrasound image region, including: The segmented ultrasound image regions were preprocessed as follows: median filtering algorithm was used to remove image noise, and linear gray-level normalization was used to map the image gray values to the [0,1] interval; the ultrasound image size was standardized to 512×512 pixels. The preprocessed ultrasound images are input into a trained improved deep learning feature extraction model to extract deep features, which include texture features, grayscale features, morphological features and lesion-related features of the ultrasound images. The improved deep learning feature extraction model is based on the SE-ResNet architecture and integrates an SE channel attention mechanism after each residual block.
6. The method for constructing an ultrasound dataset according to claim 1, characterized in that: Step S4 also includes: cleaning the associated data and removing unqualified samples; classifying and labeling the data samples according to the preset labeling specifications and performing cross-validation; encapsulating the sorted samples using a unified data format to generate a standardized ultrasound dataset, and establishing a dataset index based on anatomical location, machine model, and lesion type.
7. The method for constructing an ultrasound dataset according to claim 1, characterized in that: It also includes step S5, dataset validation and optimization: select a portion of the samples to validate the standardized ultrasound dataset and evaluate the dataset quality; for any issues found during validation, return to the corresponding step for optimization and adjustment; After optimization, version management is implemented to ensure the traceability of the dataset.
8. The method for constructing an ultrasound dataset according to claim 7, characterized in that: The verification dimensions include completeness, accuracy, and standardization. Completeness verification checks whether the sample contains ultrasound images, key information, image features, and annotation information to ensure that no information is missing. Accuracy verification evaluates the recognition accuracy and annotation accuracy by comparing the results of manual recognition with machine recognition, and the results of manual annotation with the results of model feature extraction. Standardization verification involves checking whether the sample format, size, and feature dimensions are consistent; optimization and adjustment include optimizing the OCR text recognition model and fine-tuning the parameters of the deep learning feature extraction model; version management includes labeling the version number, recording the update content and reasons, and forming a dataset update log.
9. An ultrasound dataset construction device, characterized in that, include: The acquisition module is used to acquire the screen data of the ultrasound diagnostic instrument. The screen data includes an ultrasound image area and a synchronously displayed auxiliary information area. The auxiliary information area includes at least one key text information, such as the device model and imaging parameters. The processing module is used to process the screen data to separate the ultrasound image area and the auxiliary information area; The recognition module is used to perform text recognition on the auxiliary information area using a text recognition model in order to extract the key text information; The extraction module is used to extract deep features of the ultrasound image in the ultrasound image region using a deep learning feature extraction model. The association module is used to combine key text information to annotate and correct the deep features of ultrasound images, thereby obtaining an ultrasound image feature set that integrates multi-dimensional information. The processing module is used to associate and store the original ultrasound images, extracted key text information, and ultrasound image feature sets, establish the correspondence between "ultrasound image - machine information - imaging parameters - image features", and generate a standardized ultrasound dataset.
10. The ultrasound dataset construction apparatus according to claim 9, characterized in that, The recognition module is configured to: use a convolutional neural network-based OCR text recognition model to recognize the text in the auxiliary information area, and use a Transformer-based character correction model to correct the recognition result of the OCR text recognition model; the extraction module is configured to: use a deep learning feature extraction model based on the SE-ResNet architecture to extract deep features.