Low-quality data training method and system in ultrasonic image deep learning feature modeling

Abstract

The invention relates to a low-quality data training method and system in ultrasonic image deep learning feature modeling, and belongs to the technical field of ultrasonic images. The method comprises the following steps: firstly, acquiring ultrasonic image information, and carrying out joint labeling on image quality, disease types and regions of interest; the ultrasonic image quality is automatically evaluated and classified, low-quality images are input into the low-quality data training module, and qualified images are input into the enhancement training module. The low-quality data training module performs feature modeling on a low-quality image by using a deep learning algorithm and an attention mechanism, and the enhancement training module performs enhancement training on a qualified data set to optimize a model effect. And finally, the decision fusion module fuses the model results of the plurality of regions of interest to obtain a final decision result. According to the method, the low-quality ultrasonic image data is utilized, the waste of medical resources is avoided, the generalization ability of the model is enhanced, and a new thought is provided for a clinical feature modeling task.

Description

Technical Field

[0001] The present invention belongs to the technical field of ultrasonic imaging, and relates to a low-quality data training method and system in deep learning feature modeling of ultrasonic imaging. Background Art

[0002] Feature modeling of ultrasonic imaging plays a key role in the process of establishing the association between ultrasonic imaging and disease information. Researchers use advanced deep learning technologies to extract certain patterns or rules by training and learning a large number of samples. However, the characteristics of low standardization and low quality of ultrasonic imaging seriously affect the data accumulation and modeling quality. In the process of feature modeling, how to effectively utilize the discharged and abandoned medical resources is an urgent problem to be solved.

[0003] As high-dimensional unstructured data, ultrasonic imaging faces challenges in data quality and standardization in deep learning feature modeling. To overcome these challenges, the prior art mainly adopts three strategies. 1. Data screening and inclusion / exclusion: By setting strict inclusion and exclusion criteria to screen high-quality ultrasonic imaging data to ensure the knowledge content of the training set; 2. Image enhancement technology: Such as homomorphic filtering algorithm, to improve the clarity of ultrasonic imaging; 3. Data generation: Using deep learning technologies to convert low-quality data into high-quality data that meets the standards to meet the training needs of deep learning models.

[0004] To overcome the challenges brought by the data quality and data standardization of ultrasonic imaging data to deep learning feature modeling, the prior art uses data screening, image enhancement, and data generation methods to process ultrasonic data. However, the following problems are faced: 1. The most widely used data screening and inclusion / exclusion methods result in a relatively high proportion of data being excluded from feature modeling, and the fragmented and scattered knowledge contained in low-quality or retrospective data is difficult to be directly utilized; 2. Image enhancement technology can, to a certain extent, eliminate noise and artifact information in ultrasonic imaging, but it is difficult to pay attention to and utilize the disease information in non-standard sections; 3. Data generation technology requires a large number of samples for training, has the problem of "artificial intelligence hallucination", and has a low practical application value.

[0005] In the process of feature modeling, how to effectively utilize the discharged and abandoned medical resources is a problem worthy of in-depth study. From the perspective of medical resource utilization, this large amount of neglected data actually contains valuable information. These data may cover the diversity of different cases, disease stages, and treatment methods, and can provide more comprehensive training samples for the model, thereby improving the generalization performance of the model. Moreover,

[0006] These data may also contain rare cases or special situations, which are of great significance for medical research and clinical practice. Therefore, how to fully explore and utilize the disease information contained in low-quality and unqualified data during the feature modeling process will become an important task in existing research. However, it is not easy to achieve this goal. These excluded data may have problems such as noise and artifacts, so the reliability and effectiveness of the data need to be considered during utilization. Summary of the Invention

[0007] In view of this, the purpose of the present invention is to provide a method and system for training with low-quality data in ultrasonic image deep learning feature modeling. Existing methods for ultrasonic image deep learning feature modeling discard a large amount of low-quality ultrasonic image data, resulting in waste and insufficient utilization of medical data resources. Therefore, in order to effectively utilize the disease knowledge in low-quality ultrasonic images and optimize the ultrasonic image deep learning feature modeling task, this patent provides a method for training and utilizing low-quality data for ultrasonic image deep learning feature modeling. Transfer the knowledge in the clinical image information resources that are ignored due to not meeting the experimental standards to a qualified specific data set to optimize the feature modeling task. This method not only solves the problems of small samples in medical images and waste of information resources to a certain extent, but also expands the application scope through multi-granularity feature modeling, providing a new solution idea for the clinical feature modeling task.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A method for training with low-quality data in ultrasonic image deep learning feature modeling, comprising the following steps:

[0010] Obtain ultrasonic image information;

[0011] Perform joint annotation on image quality, disease category, and region of interest;

[0012] Automatically evaluate the quality of ultrasonic images and classify them;

[0013] Perform feature modeling on low-quality ultrasonic images;

[0014] Perform enhanced training on qualified ultrasonic image data;

[0015] Fuse the enhanced training module to obtain the model results for each region of interest.

[0016] Furthermore, the obtaining of ultrasonic image information includes:

[0017] Obtain ultrasonic images including both sequence videos and image signals;

[0018] When the input is a sequence video signal, perform frame decomposition on it to obtain image signals.

[0019] Furthermore, the joint annotation includes:

[0020] Annotating the image quality and classifying the image quality into two labels: qualified and unqualified;

[0021] Annotating different possible disease types in the image;

[0022] Marking multiple regions of interest in each image.

[0023] Furthermore, the automatic evaluation and classification of ultrasound image quality include:

[0024] Automatically identifying the regions of interest in the ultrasound image;

[0025] Automatically evaluating the quality category of the image based on the quantity and distribution characteristics of these regions.

[0026] Furthermore, the feature modeling of low-quality ultrasound images includes:

[0027] Performing feature modeling of low-quality data based on the number of regions of interest in the low-quality ultrasound image data;

[0028] The feature modeling algorithms include VGG, DenseNet, ResNet, Inception, GoogleNet, and Xception model algorithms;

[0029] The attention mechanisms include CBAM, SKNet, SENet, and ECANet.

[0030] Furthermore, the enhanced training on qualified ultrasound image data includes:

[0031] Performing enhanced training based on the number of regions of interest in the qualified ultrasound image data;

[0032] The training algorithms include VGG, DenseNet, ResNet, Inception, GoogleNet, and Xception model algorithms;

[0033] The attention mechanisms include CBAM, SKNet, SENet, and ECANet.

[0034] Furthermore, the fusion enhanced training module obtaining the model results of each region of interest includes:

[0035] Using decision-level fusion algorithms, including learning method, voting method, averaging method, Stacking, Bagging, and Boosting.

[0036] An ultrasonic image deep learning feature modeling system, comprising:

[0037] A data acquisition module, configured to acquire ultrasonic image information;

[0038] An aggregation annotation module, configured to perform joint annotation on image quality, disease category, and region of interest;

[0039] A quality recognition and classification module, configured to automatically evaluate and classify the quality of ultrasonic images;

[0040] A low-quality data training module, configured to perform feature modeling on low-quality ultrasonic images;

[0041] An enhancement training module, configured to perform enhancement training on qualified ultrasonic image data;

[0042] A decision fusion module, configured to fuse the model results of each region of interest obtained by the enhancement training module.

[0043] Furthermore, the data acquisition module includes:

[0044] An acquisition module, configured to acquire ultrasonic images including both sequence video and image signals;

[0045] A frame decomposition module, configured to perform frame decomposition processing on the input sequence video signal to obtain image signals when the input is a sequence video signal.

[0046] Furthermore, the quality recognition and classification module includes:

[0047] An object detection module, configured to automatically identify the region of interest in the ultrasonic image;

[0048] A quality evaluation module, configured to automatically evaluate the quality category of the image based on the quantity and distribution characteristics of these regions.

[0049] The beneficial effects of the present invention are as follows:

[0050] The low-quality data training method and system in the deep learning feature modeling of ultrasonic images of the present invention for creative inventions have significant advantages and positive effects compared with the prior art. First of all, it effectively utilizes the excluded low-quality ultrasonic image data, avoiding the waste of medical resources. Although these data are insufficient in standardization and quality, they still contain rich medical knowledge and can provide more training samples for the model. Secondly, by integrating diverse case and disease stage information, the present invention enhances the generalization ability of the model, making it perform better in different clinical scenarios. In addition, the present invention not only solves the small sample problem but also provides new ideas for the clinical feature modeling task, broadening the application scope of ultrasonic images in medical research and clinical practice. By migrating the knowledge in the clinical image information resources that do not meet the experimental standards to the qualified data sets, the present invention optimizes the feature modeling task and improves the data utilization efficiency. Finally, this method can mine and utilize the neglected low-quality data, especially the data of rare cases or special situations, providing a new perspective and data support for medical research. In summary, through the innovative low-quality data training method, the present invention significantly improves the efficiency and effect of the deep learning feature modeling of ultrasonic images, providing important technical support for research and applications in the medical field.

[0051] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent specification, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following specification. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0053] Figure 1 is a schematic flow chart of the low-quality data training method in the deep learning feature modeling of ultrasonic images in this application;

[0054] Figure 2 is a system block diagram of the low-quality data training method in the deep learning feature modeling of ultrasonic images in this application;

[0055] Figure 3 is an embodiment of cardiac ultrasound examination;

[0056] Figure 4 is an embodiment of fetal facial ultrasound examination. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] The following uses specific specific examples to illustrate the implementation manners of the present invention. Those skilled in the art can easily understand the other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the diagrams provided in the following embodiments only schematically illustrate the basic concept of the present invention. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0058] Among them, the attached drawings are only used for exemplary illustration, showing only schematic diagrams, rather than physical diagrams, and cannot be understood as a limitation to the present invention; in order to better illustrate the embodiments of the present invention, some components in the attached drawings will be omitted, enlarged or reduced, which does not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the attached drawings may be omitted.

[0059] In the attached drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc. indicating the orientation or positional relationship, it is based on the orientation or positional relationship shown in the attached drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the attached drawings are only used for exemplary illustration and cannot be understood as a limitation to the present invention. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0060] In order to effectively utilize the disease knowledge in low-quality ultrasound images and optimize the deep learning feature modeling task of ultrasound images, this patent provides a method and system for training and utilizing low-quality data for deep learning feature modeling of ultrasound images. The present application adopts the following technical solutions:

[0061] 1. Data acquisition module: Acquire ultrasound image information;

[0062] 2. Aggregated annotation module: Jointly annotate the image quality, disease category, and region of interest;

[0063] 3. Quality identification and classification module: Automatically evaluate and classify the quality of ultrasound images;

[0064] 4. Low-quality data training module: Perform feature modeling on low-quality ultrasound images;

[0065] 5. Enhanced training module: Perform enhanced training on qualified ultrasound image data;

[0066] 6. Decision Fusion Module: The enhanced training module obtains the model results for each region of interest.

[0067] The low-quality data training method and system in the deep learning feature modeling of ultrasonic images include an ultrasonic image data acquisition module, an aggregation annotation module, a quality recognition and classification module, a low-quality data training module, an enhancement training module, and a decision fusion module.

[0068] Data Acquisition Module: It is used to acquire ultrasonic images, which include two ways of sequence videos and image signals. If it is a sequence video, it will be frame-sliced into image signals. This module is responsible for acquiring ultrasonic image data, and the data forms include sequence video signals and static image signals. When the input is a sequence video signal, the module will perform frame decomposition processing on it, converting the video signal into a series of continuous static image signals, denoted as: {I 1 , I 2 ,..., I n}, where I i represents the i-th frame image. The converted image signal is sent to the multi-information aggregation annotation module for label annotation.

[0069] Aggregation Annotation Module: It is used to jointly annotate the image quality, disease category, and region of interest of ultrasonic images. The image quality is annotated with two labels; there can be multiple annotation categories for the disease category; and there can be multiple region annotations for the region of interest. For each image I i , the quality assessment is performed, and the classification label is denoted as Q(I i ) ∈ {q 1 , q 2}, where q 1 represents qualified quality, and q 2 represents unqualified quality; for different disease types that may be recognized in the image, the annotation module provides multiple disease annotation categories. For each image I i , the disease category annotation set is denoted as D(I i ) = {d 1 , d 2 ,..., d m}, where d j represents the j-th type of disease; the module allows marking multiple regions of interest in each image. For the image I i , these regions of interest are represented by R(I i ) = {r 1 , r 2 ,..., r k}, where r j represents the j-th region of interest. The output of the multi-information aggregation annotation module integrates the above multi-dimensional annotation information to form a comprehensive data output structure. This structure can be described as: OI = {Q, D, R}, where O I represents the fully annotated output of each image, which is sent to the region-driven image quality recognition module for image quality classification.

[0070] Quality recognition and classification module: This module aims to automatically identify the regions of interest in ultrasonic images and automatically evaluate the quality category of the images based on the quantity and distribution characteristics of these regions. In specific implementation, the automatic identification of regions of interest relies on advanced object detection algorithms, including models such as YOLO, R-CNN, EfficientDet, and RetinaNet. Assume that N regions of interest are identified in the processed image, and the maximum score for the quality of each image is 10 points. The calculation formula for the quality score is: Q = 10 - (10 - N) = N. When the quality score Q = 10, the image is regarded as a high-quality qualified category. When the quality score Q < 10, the image is regarded as a low-quality unqualified category. Through this module, ultrasonic images are divided into two categories: low quality and qualified quality. Among them, low-quality ultrasonic images are input into the low-quality ultrasonic image feature modeling module, and qualified-quality ultrasonic images are input into the low-quality model transfer learning module.

[0071] Low-quality data training module: This module learns relevant disease diagnosis and treatment knowledge in low-quality ultrasonic image data through deep learning feature modeling. This module performs feature modeling of low-quality data based on the number of regions of interest in low-quality ultrasonic image data. The modeling algorithms include model algorithms such as VGG, DenseNet, ResNet, Inception, GoogleNet, and Xception, and the attention mechanisms include CBAM, SKNet, SENet, ECANet, etc. For each region of interest, a corresponding model and training weights are obtained through feature modeling. Extract each region of interest from the low-quality ultrasonic image, denoted as R i , where i = 1, 2,..., N. Input each region R i into the selected model M K , and combine it with the attention mechanism A m , (such as CBAM, SENet, etc.) for feature extraction. The model weights obtained through modeling can be expressed as: P i = A m (M k (R i ))), where P i = {P 1 , P 2 ,......, P i} are the model weights extracted from the region of interest R i , which are used as the initial training weights for the input enhancement training module.

[0072] Enhanced Training Module: Based on the feature modeling of low-quality ultrasound images, this module conducts enhanced training on a qualified dataset to optimize the feature modeling effect in qualified domain data. This module performs enhanced training based on the number of regions of interest in the qualified ultrasound image data. The training algorithms include model algorithms such as VGG, DenseNet, ResNet, Inception, GoogleNet, Xception, etc., and the attention mechanisms include CBAM, SKNet, SENet, ECANet, etc. For each region of interest, a corresponding model and training weights are obtained through enhanced training. Each region of interest is extracted from the qualified ultrasound images, denoted as H i , where i = 1, 2,..., N. Each region H i is input into the selected model Z K , and feature extraction is performed in combination with the attention mechanism D m , (such as CBAM, SENet, etc.). Using the training method of transfer learning, for each region of interest i, there are initial weights P i = {P 1 , P 2 ,......, P i} from the low-quality ultrasound image feature modeling module, which are used for enhanced training to obtain the model PZ i = {PZ 1 , PZ 2 ,......, PZ i} for each region of interest. The training process is expressed as: PZ i = D m (Z k (R i ), Pi), where i = {1, 2,......, N}, D m is the selected type of attention mechanism, Z k is the selected feature modeling algorithm, and R i is the region of interest. The enhanced training module obtains the corresponding model PZ i for each region of interest i and inputs it into the decision fusion module. If the region of interest is 1, the module obtained through training is the final decision model.

[0073] Decision Fusion Module: This module is used to fuse the model results of each region of interest obtained by the enhanced training module and achieve decision-making through fusion (if the region of interest is 1, this module is not used for fusion). This module uses decision-level fusion algorithms, including learning methods (including SVM, XGBoost, random forest, etc.), voting method, averaging method, Stacking, Bagging, Boosting, etc. For the models PZ i={PZ 1 , PZ 2 , PZ 3 , ..., PZ i}, and the decision fusion process is expressed as: Among them, represents the final decision fusion result, and f fusion (*) represents the fusion algorithm, which is defined according to different decision fusion algorithms. The decision fusion module finally obtains the decision fusion result where i = {1, 2,......, N}, representing the number of regions of interest.

[0074] As Figure 1 and Figure 2 shown, the low-quality data training method and system in the ultrasonic image deep learning feature modeling of the present invention include six parts, which are composed of a data acquisition module, an aggregation annotation module, a quality recognition and classification module, a low-quality data training module, an enhanced training module, and a decision fusion module, and respectively execute the tasks of acquiring ultrasonic images, multi-label annotation, automatically evaluating and classifying the quality of ultrasonic images, feature modeling of low-quality ultrasonic images, enhanced training on qualified ultrasonic images, and fusing the training results of the model.

[0075] In the data acquisition module, this system adopts a variety of imaging methods, including two-dimensional (B-ultrasound), three-dimensional, and four-dimensional imaging technologies, to meet different ultrasonic image analysis requirements. These requirements cover various application scenarios such as obstetric ultrasound, cardiac ultrasound, abdominal ultrasound, urinary system ultrasound, breast ultrasound, soft tissue ultrasound, thyroid ultrasound, and emergency ultrasound. This module is compatible with a variety of data formats, including common image formats (such as JPG, PNG), DICOM format, and video format. During the imaging process, users can flexibly adjust the gain, frequency, depth, and focus position to optimize the imaging effect. In addition, the system also provides real-time processing functions such as denoising, smoothing, and enhancement processing to improve the clarity of the image and ensure the stability of the imaging process. Such a designed module not only meets diverse clinical needs but also improves the application efficiency of ultrasonic images.

[0076] This article provides two cases of one region of interest (i = 1) and multiple regions of interest (i > 1) for demonstration, that is, two typical embodiments of fetal facial ultrasound examination and cardiac ultrasound examination are given as examples.

[0077] In a cardiac ultrasound examination embodiment, step 1: First, obtain two-dimensional images or videos of cardiac ultrasound examinations through a data acquisition module. The videos are converted into images by frame cutting; step 2: In the aggregation annotation module, invite professional doctors to annotate the image quality, regions of interest, and disease categories. The image categories are low-quality categories and qualified-quality categories. In this case, only 1 region of interest in the cardiac region needs to be annotated; step 3: The quality recognition and classification module performs recognition training on the regions of interest based on the cardiac regions annotated by the aggregation annotation module. In this embodiment, only 1 region of interest, which is the overall cardiac structure, is annotated. The target model used for region-of-interest recognition is the YOLO network in this case. If the cardiac structure is recognized, it indicates that the section quality is qualified. The formula for the low-quality data training module can be expressed as P = A m (M k (R 1 ))), where R 1 is the region of interest (only 1 in this case), M k is the feature modeling algorithm. The algorithm used in this case is DenseNet. Other model algorithms such as VGG, ResNet, Inception, GoogleNet, and Xception can also be used. A m is the attention mechanism algorithm. The one used in this case is CBAM, but it is not limited to CBAM, SKNet, SENet, ECANet, etc.; The enhanced training module is expressed as PZ 1 = D 1 (Z 1 (R 1 ))), where P is the learning weight of the low-quality data, R 1 is the region of interest (only 1 in this case), Z 1 is the feature modeling algorithm. The algorithm used in this case is DenseNet, and D 1 is the attention mechanism algorithm. The one used in this case is CBAM; The calculation formula of the decision fusion module can be expressed as Since there is only 1 region of interest in this case, the fusion module only needs to consider the feature modeling results of this region.

[0078] In an embodiment of fetal facial ultrasound examination. Step 1, first, the two-dimensional image or video of the fetal facial ultrasound examination is obtained through the data acquisition module, and the video is converted into an image by frame cutting; Step 2, in the aggregation annotation module, professional doctors are invited to annotate the image quality, region of interest, and disease category. The image categories are low-quality category and qualified-quality category. In this case, 6 regions of interest need to be annotated, including the nose (the tip of the nose and nasal bone), chin, palate, nuchal translucency, cranial vertex, and head; Step 3, the quality recognition and classification module identifies and trains based on the 6 regions of interest of the fetal facial ultrasound annotated by the aggregation annotation module. In this embodiment, only 1 region of interest, the overall structure of the heart, is annotated. If 6 regions of interest are recognized, it indicates that the section quality is qualified; Step 4, the formula of the low-quality data training module can be expressed as P i =A m (M k (R i ))), where P i ={P 1 , P 2 ,......, P 6} is the model weight extracted from the region of interest R i , and is used as the initial training weight of the 6 regions of interest for input into the enhancement training module. M k is the feature modeling algorithm. In this case, the algorithm used is DenseNet. Other model algorithms such as VGG, ResNet, Inception, GoogleNet, Xception, etc. can also be used. A m is the attention mechanism algorithm. In this case, CBAM is used, but it is not limited to CBAM, SKNet, SENet, ECANet, etc.; Step 5, for the 6 regions of interest i = 6 in the enhancement training module, there is an initial weight P i ={P 1 , P 2 ,......, P 6} from the low-quality ultrasound image feature modeling module, which is used for enhancement training to obtain the model PZ i ={PZ 1 , PZ 2 ,......, PZ 6} of each region of interest. The training process is expressed as: PZ i =D m (Z k (R i ), Pi), where i = {1, 2,......, 6}, D m is the selected attention mechanism type CBAM, and Z kThe selected feature modeling algorithm is ResNet; Step 6, the decision fusion module obtains the model PZ for each region of interest for enhanced training i ={PZ 1 , PZ 2 , PZ 3 ,..., PZ 6}}, and the fusion process is expressed as where represents the final decision fusion result, and f fusion (*) represents the fusion algorithm XGboost. The decision fusion module finally obtains the decision fusion result

[0079] As Figure 3 and Figure 4 shown, through the 6 steps of the above 2 cases, the training of low-quality ultrasound image data in the process of deep learning feature modeling of ultrasound images is realized, and the data resources discarded in the previous methods are effectively utilized. The above content is only the preferred embodiment of the present application and does not limit the protection scope of the present application. Any feature disclosed in this specification (including the abstract and drawings), without special explanation, can be replaced by other alternative features with equivalent or similar purposes. In other words, unless otherwise specified, each feature is only an example of multiple equivalent or similar features.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the purpose and scope of the present technical solution, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A low-quality data training method for deep learning feature modeling of ultrasound images, characterized by: The following steps are involved: Obtaining ultrasound image information; Jointly annotate image quality, disease category, and regions of interest; Automatically assess and classify ultrasound image quality; Feature modeling of low-quality ultrasound images; Conduct enhanced training on qualified ultrasound image data; The fusion enhancement training module obtains the model results for each region of interest.

2. The low-quality data training method in ultrasound image deep learning feature modeling according to claim 1 is characterized in that: The obtaining of ultrasound image information comprises: Acquire ultrasound images in two ways including sequential video and image signals; When the input is a sequential video signal, it is frame-decomposed into image signals.

3. The low-quality data training method in ultrasound image deep learning feature modeling according to claim 1 is characterized in that: The joint annotation includes: Label the image quality and classify it into two labels: qualified and unqualified; Label the different types of diseases that may be identified in the image; Mark multiple regions of interest in each image.

4. The low-quality data training method in ultrasound image deep learning feature modeling according to claim 1, characterized in that: The automatic evaluation and classification of ultrasound image quality comprises: Automatically identify regions of interest in ultrasound images; The quality category of the image is automatically assessed based on the number and distribution characteristics of these areas.

5. The low-quality data training method in ultrasound image deep learning feature modeling according to claim 1, characterized in that: The feature modeling of low-quality ultrasound images includes: Perform feature modeling of low-quality data based on the number of regions of interest in low-quality ultrasound image data; The feature modeling algorithms include VGG, DenseNet, ResNet, Inception, GoogleNet and Xception model algorithms; The attention mechanisms include CBAM, SKNet, SENet and ECANet.

6. The low-quality data training method in ultrasound image deep learning feature modeling according to claim 1, characterized in that: The enhanced training on qualified ultrasound image data includes: Enhanced training is performed based on the number of regions of interest in qualified ultrasound image data; The training algorithms include VGG, DenseNet, ResNet, Inception, GoogleNet and Xception model algorithms; The attention mechanisms include CBAM, SKNet, SENet and ECANet.

7. The low-quality data training method in ultrasound image deep learning feature modeling according to claim 1, characterized in that: The fusion enhancement training module obtains the model results of each region of interest including: Use decision layer fusion algorithms, including learning, voting, averaging, stacking, bagging, and boosting.

8. An ultrasound image deep learning feature modeling system, characterized by: include: A data acquisition module, used for acquiring ultrasound image information; Aggregate annotation module, used to jointly annotate image quality, disease category, and region of interest; Quality recognition and classification module, used to automatically evaluate and classify ultrasound image quality; Low-quality data training module, used for feature modeling of low-quality ultrasound images; An enhanced training module, used for performing enhanced training on qualified ultrasound image data; The decision fusion module is used to fuse the enhanced training module to obtain the model results of each region of interest.

9. The ultrasound image deep learning feature modeling system according to claim 8, characterized in that: The data acquisition module comprises: An acquisition module is used to acquire ultrasound images in two modes including sequential video and image signals; The frame decomposition module is used to decompose the input sequence video signal into an image signal.

10. The ultrasound image deep learning feature modeling system according to claim 8, characterized in that: The quality identification and classification module comprises: The object detection module is used to automatically identify regions of interest in ultrasound images; The quality assessment module is used to automatically assess the quality category of the image based on the quantity and distribution characteristics of these areas.

Images