Antenatal ultrasonic auxiliary diagnosis and quality control method and system based on artificial intelligence
Through the artificial intelligence module based on deep learning, the automated quality control and parameter measurement of fetal ultrasound images are realized, which solves the problem of relying on doctors in traditional methods, improves diagnostic accuracy and efficiency, and reduces the burden on doctors.
Patent Information
- Application Number
- CN202510466864.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-15
- Publication Date
- 2025-07-29
AI Technical Summary
Traditional fetal ultrasound image quality control and manual parameter measurement rely heavily on physician subjectivity, poor repetition, time-consuming and labor-consuming, and lack of automated quantitative quality control systems.
Using artificial intelligence modules based on deep learning, a quality control AI model and automatic measurement AI model are established to realize automated quality control and parameter measurement of fetal ultrasound images, and accurate measurement is completed through pixel length conversion.
It improves the accuracy and diagnostic efficiency of fetal growth and development status assessment, reduces the burden on physicians, and ensures consistency and comparable diagnostic results.
Smart Images
Figure CN120387993A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of ultrasonic image recognition, and in particular to a prenatal ultrasonic assisted diagnosis and quality control method and system based on artificial intelligence. Background Art
[0002] Prenatal ultrasound examination is an important method for evaluating the development status of the fetus and preventing the birth of severely defective infants. The main processes include: 1. Obtaining standard sections of key parts of the fetus; 2. Measuring fetal biological parameters based on the standard sections; 3. Evaluating the growth and development status of the fetus and screening for fetal malformations. It can be seen that, on the one hand, the quality of the standard sections collected by the ultrasound physician and the accuracy of the measured biological parameters will jointly affect the accuracy of the evaluation of the growth and development status of the fetus; on the other hand, according to the requirements of the "Ultrasound Prenatal Screening Guidelines (2022)" of the Ultrasound Medicine Branch of the Chinese Medical Association, multiple sections of the fetal whole body system must be collected and retained in a standardized manner according to the guidelines. Therefore, the quality control of fetal ultrasound images is of crucial importance and is involved in many aspects such as medical safety, avoiding missed and misdiagnosed cases, preventing medical disputes, department assessment, and remote consultation.
[0003] The traditional quality control of fetal ultrasound images and manual parameter measurement have the following problems: 1. It is highly dependent on the subjectivity and clinical work experience of ultrasound physicians, and the repeatability is poor; 2. This repetitive operation is time-consuming and laborious, seriously increasing the burden on ultrasound physicians. With the rapid development of computer vision and artificial intelligence technologies, prenatal ultrasound assisted diagnosis based on artificial intelligence technology has become an inevitable trend. Establishing a fully automated fetal ultrasound section quality control and parameter measurement solution can objectively evaluate the quality of ultrasound sections and accurately measure fetal biological indicators, assist ultrasound physicians in accurately evaluating the growth and development status of the fetus, improve the efficiency of prenatal ultrasound diagnosis, and reduce the burden on ultrasound physicians. However, at present, there is rarely a comprehensive automated quantitative quality control system proposed based on clinical quality control criteria for the quality evaluation of fetal sections obtained during prenatal ultrasound examinations. Summary of the Invention
[0004] The purpose of the present invention is to provide a prenatal ultrasonic assisted diagnosis and quality control method and system based on artificial intelligence. By establishing an artificial intelligence module based on deep learning, it realizes the automated quality control and parameter measurement of fetal ultrasound images collected by physicians, and establishes an artificial intelligence assisted diagnosis and quality control system that can assist ultrasound physicians in accurately evaluating the growth and development status of the fetus, improve the efficiency of prenatal ultrasound diagnosis, and reduce the burden on physicians.
[0005] To achieve the above purpose, the present invention provides the following technical solutions:
[0006] On the one hand, the present invention provides a prenatal ultrasonic assisted diagnosis and quality control method based on artificial intelligence, including the following steps:
[0007] S1. Obtain fetal ultrasound image data;
[0008] S2. Establish a quality control AI model, and input the fetal ultrasound image data in S1 into the quality control AI model for training;
[0009] S3. Establish an automatic measurement AI model, and input the fetal ultrasound images output by S2 into the automatic measurement AI model for training;
[0010] S4. Input the fetal ultrasound image to be measured into the quality control AI model that has completed training, and identify it through the automatic measurement AI model. The prediction result obtains the segmentation map of the contour of the area of the image to be measured, and through the conversion of pixel length and physical length, the automatic and accurate measurement of the image to be measured is completed.
[0011] In some embodiments, S1 includes the following steps:
[0012] Select an ultrasound image dataset of fetuses with a gestational age of 20 - 30 weeks;
[0013] Expand the ultrasound image dataset to 7 - 9 types of sections, and each type of section includes two categories: standard section and non - standard section;
[0014] Perform image translation, image rotation, and data augmentation on the expanded ultrasound image dataset.
[0015] In some embodiments, the sections are thalamic transverse section, cerebellar transverse section, spinal cord sagittal section, four - chamber heart section, abdominal circumference transverse section, double - kidney horizontal section, bladder horizontal section, umbilical cord abdominal wall entrance section, and femur long - axis section.
[0016] In some embodiments, the image translation is: randomly translate the ultrasound image in the horizontal and vertical directions; the image rotation is to randomly rotate the image, and the rotation angle range is ±15°; the data augmentation is to augment the data by adjusting the image brightness, contrast, and adding noise.
[0017] In some embodiments, S2 includes the following steps:
[0018] Extract the local phase feature map of the fetal ultrasound image and fuse it to obtain an RGB three - channel image;
[0019] Use the region - of - interest detector L - CNN based on a deep convolutional neural network to locate the region to be recognized in the fetal image;
[0020] Use the region - of - interest comprehensive classifier C - CNN to directly recognize the important anatomical structures in the region to be recognized in the fetal image.
[0021] In some embodiments, S3 includes the following steps:
[0022] Input the fetal ultrasound image recognized by S2, and use the level 0 object classifier to obtain the initial segmentation prediction probability map of the fetal area to be detected;
[0023] Add the probability map predicted at level 0 to the original image at a certain ratio to obtain a fused image, and obtain the segmentation prediction probability map of the fetal area to be detected at level 1;
[0024] Similarly, still based on the fetal ultrasound image recognized by S2, fuse it with the probability map predicted at level 1 to obtain the target segmentation probability map at level 2;
[0025] Finally, use the prediction result optimized by 2 iterations to obtain the segmentation map of the contour of the area to be measured, and complete the automatic measurement of the area to be measured through the conversion between pixel length and physical length.
[0026] On the other hand, the present invention provides an artificial intelligence-based prenatal ultrasound assisted diagnosis and quality control system, which uses the above method and includes the following modules:
[0027] Image data acquisition module: used to acquire fetal ultrasound image data;
[0028] Quality control AI module: used to establish a quality control AI model, and input the fetal ultrasound image data in S1 into the quality control AI model for training;
[0029] Automatic measurement AI module: used to establish an automatic measurement AI model, and input the fetal ultrasound image output by S2 into the automatic measurement AI model for training;
[0030] Recognition module: input the fetal ultrasound image to be measured into the quality control AI model that has completed training, and recognize it through the automatic measurement AI model to complete the measurement of the fetal ultrasound image.
[0031] Compared with the prior art, the present invention has the following beneficial effects:
[0032] The present invention automatically performs quality control and measurement on ultrasound images through an AI model, can objectively judge the standardity of the section and the integrity of the anatomical structure, effectively reduce misjudgment and missed judgment caused by factors such as insufficient doctor experience or fatigue, and improve the accuracy of diagnosis.
[0033] Based on the standardized analysis process of the AI model, the present invention can ensure that each examination is carried out according to the same criteria, avoid fluctuations in diagnosis results caused by human factors, make the diagnosis results between different doctors, different devices, and different hospitals have higher comparability and consistency, and contribute to clinical research and medical quality control.
[0034] The present invention deeply applies artificial intelligence technology to an innovative solution for fetal ultrasound examination, setting a model for the intelligent development in the field of medical imaging diagnosis. Its successful application will encourage more medical institutions and research institutions to explore the application of AI in other medical imaging fields, promoting the entire medical industry towards intelligent diagnosis and improving the overall quality and efficiency of medical services. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 It is a schematic diagram of the overall process of Embodiment 1 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0036] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.
[0037] Embodiment 1:
[0038] Please refer to Figure 1 , a prenatal ultrasound assisted diagnosis and quality control method based on artificial intelligence, comprising the following steps:
[0039] S1. Obtain fetal ultrasound image data.
[0040] Data collection: The fetal ultrasound image data covers fetal gestational weeks of 20 - 35 weeks. By detailed annotation of the ultrasound images, including information such as section type, fetal anatomical structure position, and whether it is a standard section, the accuracy and consistency of the data are ensured.
[0041] Data classification: Expand the fetal ultrasound data set (training set and test set) to 9 sections, including the transverse section of the thalamus, the transverse section of the cerebellum, the sagittal section of the spine, the four-chamber heart section, the transverse section of the abdominal circumference, the transverse section at the level of the double kidneys, the transverse section at the level of the bladder, the section of the umbilical cord abdominal wall entrance, and the long-axis section of the femur, totaling 18 categories (each section is divided into two categories: standard and non-standard) to meet the diverse needs of model training and testing.
[0042] Image translation: Randomly translate the ultrasound image in the horizontal and vertical directions, and the translation distance is a certain proportion (such as 5% - 10%) of the image width or height to simulate ultrasound sections at different positions and increase the diversity of the data set.
[0043] Image rotation: Randomly rotate the image, and the rotation angle range is ±15° to simulate scans at different angles of the ultrasound probe and enhance the model's recognition ability for sections at different angles.
[0044] Data augmentation: Other image processing techniques are adopted, such as adjusting image brightness, contrast, adding noise, etc., to further enrich the dataset and improve the generalization ability of the model.
[0045] S2. Establish a quality control AI model and input the fetal ultrasound image data in S1 into the quality control AI model for training.
[0046] Taking the fetal thalamus section of this embodiment as an example, the specific steps are as follows:
[0047] 1. Feature extraction and image fusion
[0048] Local phase feature extraction: Perform local phase feature extraction on the original ultrasound image to highlight key structural information such as edges and textures in the image, and enhance the recognition ability of important anatomical structures such as the fetal thalamus.
[0049] RGB three-channel image fusion: Fuse the extracted local phase feature map with the original image to generate an RGB three-channel image, providing a richer feature representation for the subsequent deep learning model.
[0050] 2. Detection of objects of interest
[0051] Detector based on deep convolutional neural network (L-CNN): Use a deep convolutional neural network to construct a detector for objects of interest, extract and analyze features from the fused RGB image, and accurately locate the fetal thalamus plane area. Through multi-layer convolution and pooling operations, gradually extract the high-level semantic features of the image, and finally output the position coordinates and confidence of the thalamus area.
[0052] 3. Recognition of important anatomical structures
[0053] Comprehensive classifier for objects of interest (C-CNN): For key anatomical structures (such as the septum pellucidum, thalamus, third ventricle, choroid plexus, etc.) within the fetal thalamus plane area, construct a comprehensive classifier. Based on a deep learning architecture, this classifier further subdivides and identifies the detected thalamus area, judging whether each anatomical structure exists and whether it is normal, etc., providing a basis for subsequent diagnosis.
[0054] Generalize the above method to the other 8 sections. For the specific anatomical structures and features of each section, adjust and optimize the design of feature extraction, object detection, and classifier to ensure the quality control ability of the system for all sections.
[0055] S3. Establish an automatic measurement AI model and input the fetal ultrasound images output by S2 into the automatic measurement AI model for training.
[0056] Taking the fetal abdominal circumference of this embodiment as an example, the specific steps are as follows:
[0057] 1. Initial segmentation prediction
[0058] Level 0 Object Classifier: Using the abdominal circumference plane ultrasound image as input, the Level 0 Object Classifier, based on the fully convolutional network architecture of deep learning, performs pixel-level classification on the image to obtain an initial segmentation prediction probability map of the fetal abdominal area. By learning a large amount of annotated data, this classifier preliminarily determines which pixels in the image belong to the fetal abdominal area.
[0059] 2. Iterative Optimization of Segmentation
[0060] Level 1 Segmentation Prediction: Add the probability map predicted at Level 0 to the original image in a certain ratio (such as 0.7:0.3) to obtain a fused image. The fused image retains the feature information of the original image and incorporates the preliminary segmentation prediction results, serving as the input for Level 1 segmentation. Use the fully convolutional network to process the fused image again to obtain a more accurate segmentation prediction probability map of the fetal abdominal area at Level 1.
[0061] Level 2 Segmentation Prediction: Based on the original image, fuse it with the probability map predicted at Level 1 in a certain ratio (such as 0.6:0.4) to obtain a new fused image. Continue to use the fully convolutional network for segmentation prediction to obtain the Level 2 object segmentation probability map. Through two iterations of optimization, gradually improve the accuracy and detail performance of the segmentation.
[0062] 3. Measurement Implementation
[0063] Contour Extraction and Measurement: Using the final Level 2 segmentation probability map, extract the accurate contour of the fetal abdominal area through image processing techniques such as threshold processing and edge detection. According to the conversion relationship between the pixel length and physical length of the ultrasound device (usually provided by device parameters or calibration data), convert the pixel coordinates of the contour into actual physical dimensions to complete the automated and accurate measurement of the abdominal circumference.
[0064] Generalize the above method to other important measurement indicators in fetal ultrasound examinations (such as head circumference, femur length, etc.). According to the characteristics and requirements of different measurement targets, adjust the architecture and training strategy of the fully convolutional network to achieve the automatic measurement function of multiple indicators.
[0065] The training and optimization of the quality control AI model and the automatic measurement AI model include the following steps:
[0066] (1) Training Data Preparation
[0067] Data Partitioning: Divide the augmented ultrasound data set into a training set, a validation set, and a test set, usually in a reasonable ratio such as 7:2:1 or similar, to ensure that each data set is representative and non-overlapping.
[0068] Data preprocessing: Normalize the training data to unify the pixel values within a specific range (such as [0, 1] or [-1, 1]) to accelerate model convergence. At the same time, encode and organize the annotation information to match the model output format.
[0069] (II) Model training strategy
[0070] Loss function selection: Select an appropriate loss function according to the model task (classification, detection, segmentation, etc.). For example, cross-entropy loss can be used for classification tasks, a combination of bounding box regression loss and classification loss for detection tasks, and Dice loss or cross-entropy loss for segmentation tasks, etc., to measure the difference between the model prediction results and the true annotations and guide the model optimization direction.
[0071] Application of optimization algorithms: Use advanced optimization algorithms (such as Adam, RMSprop, etc.) to update the model parameters. Set reasonable hyperparameters such as learning rate and momentum to balance the model convergence speed and stability. During the training process, the learning rate can be dynamically adjusted according to the change of the validation set loss. For example, when the validation set loss does not decrease for several consecutive epochs, reduce the learning rate to promote further optimization of the model.
[0072] Application of regularization techniques: To prevent model overfitting, apply regularization techniques such as L2 regularization and Dropout. L2 regularization limits the size of the model parameters by adding a weight decay term to the loss function, making the model smoother; Dropout randomly deactivates some neurons during training to enhance the generalization ability of the model and improve its performance on unknown data.
[0073] (III) Model evaluation and optimization
[0074] Model evaluation metrics: Select appropriate evaluation metrics according to the model task type. Classification models can use accuracy, recall, F1-score, etc.; detection models can adopt mean average precision (mAP), recall, etc.; segmentation models use metrics such as Dice coefficient and pixel accuracy to comprehensively evaluate the performance of the model on the test set.
[0075] Model optimization and adjustment: According to the evaluation results, optimize the model specifically. If the model performs poorly in certain categories or situations, analyze the reasons and adjust the model structure (such as adding feature extraction layers, changing the network depth, etc.), or supplement and enhance the training data, and retrain the model to improve the performance.
[0076] S4. Input the fetal ultrasound image to be measured into the trained quality control AI model, and complete the measurement of the fetal ultrasound image through the automatic measurement and recognition of the AI model.
[0077] Example 2
[0078] An artificial intelligence-based prenatal ultrasound assisted diagnosis and quality control system, using the above method, includes the following modules:
[0079] Image data acquisition module: used to acquire fetal ultrasound image data;
[0080] Quality control AI module: used to establish a quality control AI model and input the fetal ultrasound image data in S1 into the quality control AI model for training;
[0081] Automatic measurement AI module: used to establish an automatic measurement AI model and input the fetal ultrasound images output from S2 into the automatic measurement AI model for training;
[0082] Recognition module: input the fetal ultrasound image to be measured into the quality control AI model that has completed training, and recognize it through the automatic measurement AI model to complete the measurement of the fetal ultrasound image.
[0083] An artificial intelligence-based prenatal ultrasound assisted diagnosis and quality control system of the present invention can be installed in a computer device. The computer device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as an artificial intelligence-based prenatal ultrasound assisted diagnosis and quality control program. Among them, the memory includes at least one type of readable storage medium, and the readable storage medium includes flash memory, mobile hard disk, multimedia card, card-type memory (such as: SD or DX memory, etc.), magnetic memory, magnetic disk, optical disc, etc. The processor is the control core of the electronic device, connects various components of the entire computer device through various interfaces and lines, and executes various functions of the computer device and processes data by running or executing the programs or modules stored in the memory, and calling the data stored in the memory.
[0084] The module of the present invention refers to a series of computer program segments that can be executed by the processor of a computer device and can complete fixed functions, and are stored in the memory of the computer device.
[0085] After considering the specification and practicing the content disclosed herein, those skilled in the art will readily think of other embodiments of the present application. The present application aims to cover any variations, uses, or adaptive changes of the present application, and these variations, uses, or adaptive changes follow the general principles of the present application and include the common general knowledge or conventional technical means in the technical field not disclosed in the present application.
Claims
1. An artificial intelligence-based prenatal ultrasound assisted diagnosis and quality control method, characterized in that, It includes the following steps: S1. Obtain fetal ultrasound image data; S2. Establish a quality control AI model, and input the fetal ultrasound image data in S1 into the quality control AI model for training; S3. Establish an automatic measurement AI model, and input the fetal ultrasound images output by S2 into the automatic measurement AI model for training; S4. Input the fetal ultrasound image to be measured into the quality control AI model that has completed training, and identify it through the automatic measurement AI model to complete the measurement of the fetal ultrasound image.
2. The prenatal ultrasound-assisted diagnosis and quality control method based on artificial intelligence according to claim 1, characterized in that S1 includes the following steps: Select an ultrasound image dataset of fetuses with a gestational age of 20 - 30 weeks; Expand the ultrasound image dataset to 7 - 9 types of sections, and each type of section includes two categories: standard section and non-standard section; Perform image translation, image rotation, and data augmentation on the expanded ultrasound image dataset.
3. The prenatal ultrasound assisted diagnosis and quality control method based on artificial intelligence according to claim 2, characterized in that, The sections are thalamus transverse section, cerebellum transverse section, spinal cord sagittal section, four-chamber heart section, abdominal circumference transverse section, double kidney horizontal section, bladder horizontal section, umbilical cord abdominal wall entrance section, and femur long axis section.
4. The prenatal ultrasound-assisted diagnosis and quality control method based on artificial intelligence according to claim 2, wherein, The image translation is to randomly translate the ultrasound image in the horizontal and vertical directions; the image rotation is to randomly rotate the image, and the rotation angle range is ±15°; the data augmentation is to augment the data by adjusting the image brightness, contrast, and adding noise.
5. The prenatal ultrasound-assisted diagnosis and quality control method based on artificial intelligence according to claim 1, wherein, S2 includes the following steps: Extract local phase feature maps from the fetal ultrasound images and fuse them to obtain an RGB three-channel image; Use the region of interest (ROI) detector L-CNN based on a deep convolutional neural network to locate the region to be recognized in the fetal image; Use the ROI comprehensive classifier C-CNN to directly recognize the important anatomical structures in the region to be recognized in the fetal image.
6. The prenatal ultrasound-assisted diagnosis and quality control method based on artificial intelligence according to claim 1, wherein S3 includes the following steps: Input the fetal ultrasound images recognized in S2, and use the level 0 object classifier to obtain an initial segmentation prediction probability map of the fetal region to be detected; Add the probability map predicted at level 0 to the original image in a certain proportion to obtain a fused image, and obtain a segmentation prediction probability map of the fetal region to be detected at level 1; Similarly, still based on the fetal ultrasound images recognized in S2, fuse them with the probability map predicted at level 1 to obtain a segmentation probability map of the level 2 object; Finally, use the prediction results optimized by 2 iterations to obtain a segmentation map of the contour of the region to be measured, and complete the automatic measurement of the region to be measured through the conversion between pixel length and physical length.
7. An artificial intelligence-based prenatal ultrasound assisted diagnosis and quality control system, using the method described in any one of claims 1-6, characterized in that, It includes the following modules: Image data acquisition module: used to obtain fetal ultrasound image data; Quality control AI module: used to establish a quality control AI model, and input the fetal ultrasound image data in S1 into the quality control AI model for training; Automatic measurement AI module: used to establish an automatic measurement AI model, and input the fetal ultrasound images output by S2 into the automatic measurement AI model for training; Recognition module: input the fetal ultrasound image to be measured into the quality control AI model that has completed training, and identify it through the automatic measurement AI model to complete the measurement of the fetal ultrasound image.