Fetal growth limitation prediction method and system based on deep learning and ultrasonic image
By fusing multimodal information from fetal ultrasound images and maternal clinical data using deep learning, a high-precision FGR prediction method is constructed. This solves the problem of reliance on human experience in diagnosis in existing technologies, and achieves highly accurate and interpretable diagnosis of fetal growth restriction, which is applicable to obstetric medical scenarios.
Patent Information
- Application Number
- CN202511287450.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-10
- Publication Date
- 2025-12-16
AI Technical Summary
Existing technologies rely on human experience in the diagnosis of fetal growth restriction (FGR), which is highly subjective and lacks cross-modal information fusion, resulting in insufficient diagnostic accuracy and a lack of interpretability, making it difficult to meet the needs of precision medicine.
By employing deep learning methods and combining fetal ultrasound images with pregnant women's clinical data, an image feature extraction network was constructed that integrates local attention mechanisms, global attention mechanisms, dynamic convolution, and SE channel attention modules. The Grad-CAM algorithm was used to generate heatmaps of the model's areas of interest, enabling interpretable diagnosis.
It improves the accuracy and interpretability of FGR diagnosis, enhances the identification ability of clinicians, and realizes the transformation from experience-based judgment to intelligent assistance. It is suitable for auxiliary screening of fetal developmental abnormalities in primary and professional obstetric medical scenarios.
Smart Images
Figure CN121148698A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of graphic data carrier reading and processing technology, and relates to an artificial intelligence and medical image processing method, particularly to a method and system for predicting fetal growth restriction based on deep learning and ultrasound images. Background Technology
[0002] Fetal growth restriction (FGR) is a common and clinically significant high-risk pregnancy condition in obstetrics, characterized by a fetal weight that fails to reach its expected growth potential during pregnancy. FGR is closely associated with adverse perinatal outcomes, including stillbirth, preterm birth, and neurodevelopmental delays, and significantly increases risks for both the newborn and the mother. Therefore, early and accurate identification of FGR is of significant clinical value.
[0003] Currently, the diagnosis of fetal growth restriction (FGR) mainly relies on two-dimensional ultrasound measurements (such as head circumference, abdominal circumference, and femur length) combined with gestational age to estimate fetal weight, which is then compared with a standard gestational age-weight curve. However, this process is highly dependent on the physician's subjective judgment and experience, and there are certain operational variability and risks of misdiagnosis. In addition, traditional manual assessment methods have low sensitivity when dealing with complex or borderline cases, making it difficult to meet the needs of precision medicine.
[0004] In recent years, artificial intelligence, especially deep learning technology, has shown great potential in the field of medical image analysis, enabling the automatic extraction of fine-grained features and pattern recognition from images. Some studies have attempted to use convolutional neural networks (CNNs) to process fetal ultrasound images to assist in the diagnosis of specific pathological conditions. However, existing methods generally suffer from the following problems:
[0005] 1) Most models only focus on the ultrasound images themselves, ignoring the clinical data of the mother and fetus, making it difficult to achieve multi-dimensional comprehensive analysis;
[0006] 2) The lack of visual explanations of key areas in the image affects the usability of the model and doctors' clinical trust.
[0007] 3) There is currently no structural design that combines local and global attention mechanisms, which is insufficient in capturing multi-scale features;
[0008] 4) It lacks a systematic deployment approach and has failed to be transformed into a truly clinically usable auxiliary diagnostic platform.
[0009] Therefore, there is an urgent need to propose an FGR prediction method and system that integrates deep learning and multimodal information, has high-precision prediction capabilities, is interpretable, and is practical, in order to meet the urgent need for intelligent and accurate assessment in obstetric clinical practice. Summary of the Invention
[0010] The purpose of this invention is to provide a method and system for predicting fetal growth restriction based on deep learning and ultrasound images, aiming to solve the problems of current FGR diagnosis, which relies too much on human experience, is highly subjective, lacks cross-modal information fusion and intelligent auxiliary tools, resulting in insufficient diagnostic accuracy and weak clinical interpretability.
[0011] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting fetal growth restriction based on deep learning and ultrasound images, comprising the following steps:
[0012] We collected fetal ultrasound images and structured clinical data from multiple centers, labeled the corresponding FGR status, and constructed the original training dataset.
[0013] The raw data undergoes uniform preprocessing, including image size standardization, image enhancement, clinical data normalization, and missing value imputation.
[0014] The preprocessed ultrasound images are input into an image feature extraction network that integrates local attention mechanism, global attention mechanism, dynamic convolution and SE channel attention module to extract multi-scale deep image features.
[0015] Structured clinical data is encoded and converted into fixed-dimensional feature vectors;
[0016] Image features and clinical features are concatenated and fused at the feature level to form a unified feature vector;
[0017] A deep classification network was constructed, and an FGR classification prediction model was trained using supervised learning. The model was then internally validated and externally evaluated to obtain the final model.
[0018] The prediction model is deployed to an interactive auxiliary diagnostic system for doctors, supporting image uploads and output of prediction results. At the same time, the Grad-CAM algorithm is used to generate a heat map of the model's region of interest, thereby achieving model interpretability.
[0019] As a preferred technical solution of the present invention, the acquisition process includes: acquiring two-dimensional ultrasound images from multiple centers, covering standard sections such as fetal head circumference, abdominal circumference, and femur length; and simultaneously acquiring clinical data such as gestational age, maternal age, weight, blood pressure, history of diabetes, and fetal movement, with the doctor indicating whether there is fetal growth restriction.
[0020] As a preferred technical solution of the present invention, the preprocessing process includes: uniformly cropping the image to a specified resolution, performing image enhancement operations such as random rotation, flipping, and translation to simulate real prenatal examination differences; performing mean filling or regression imputation on missing fields in clinical data, and performing standardization processing.
[0021] As a preferred technical solution of the present invention, the image feature extraction network adopts an improved ResNet structure. Its backbone includes multiple stacked Basic Block modules, embeds local attention modules at different levels to capture fine-grained features, combines global attention to enhance the overall image context awareness capability, and enhances the robustness of the model to complex morphological changes through dynamic convolution and SE modules.
[0022] As a preferred technical solution of the present invention, the fusion process adopts a feature stitching method, which merges image features and clinical features according to the channel dimension and sends them to a multilayer perceptron (MLP) for classification and prediction; at the same time, the cross-entropy loss function is used for optimization and training is carried out in conjunction with a learning rate scheduling strategy.
[0023] As a preferred technical solution of the present invention, a 7:1:2 training set, validation set, and test set partitioning strategy is adopted during model training. The effectiveness of the proposed model structure design is verified by comparing it with mainstream models such as ResNet50, CNN, ViT, and DenseNet and by stepwise ablation experiments.
[0024] A fetal growth restriction prediction system based on the above method includes:
[0025] The acquisition module is used to acquire fetal ultrasound images and structured clinical data of pregnant women to form a raw dataset;
[0026] The preprocessing module is used to normalize and enhance the images and clinical data to form a unified input format;
[0027] The image feature extraction module extracts high-quality image features based on local attention, global attention, dynamic convolution, and the SE module.
[0028] The fusion module is used to concatenate images with clinical feature vectors;
[0029] The prediction module outputs the probability value of whether the fetus has FGR;
[0030] The interpretation module is used to provide clinically interpretable evidence for regions of interest based on the Grad-CAM visualization model;
[0031] The user interaction module allows doctors to upload images and view prediction results and explanations.
[0032] As a preferred technical solution of the present invention, the system provides services through a web front-end deployment. After doctors upload standard ultrasound images and clinical information through a webpage, they can obtain prediction results within seconds.
[0033] An electronic device for predicting fetal growth restriction based on deep learning and ultrasound images includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the FGR prediction method based on deep learning and ultrasound images as described in any of the above embodiments.
[0034] A computer-readable storage medium for a method of predicting fetal growth restriction based on deep learning and ultrasound images, wherein the computer-readable storage medium stores a computer program that, when executed by a processor, implements the FGR prediction method based on deep learning and ultrasound images as described in any of the above embodiments.
[0035] Compared with the prior art, the beneficial effects of the present invention are:
[0036] The FGR prediction method and system provided in this application combine ultrasound images with the pregnant woman's clinical information to construct a deep learning network that integrates multi-attention mechanisms and dynamic modeling capabilities. This results in advantages such as high prediction accuracy, strong model interpretability, convenient deployment, and strong adaptability. Deployed through an online platform, it can significantly improve clinicians' ability to identify fetal developmental abnormalities, shifting from "experience-based judgment" to "intelligent assistance," and has promising clinical application prospects.
[0037] This patent collects two-dimensional ultrasound images of the fetus and relevant maternal / fetal clinical information; performs standardized preprocessing and data augmentation on the ultrasound images; constructs a deep learning feature extraction network integrating local attention mechanisms, global attention mechanisms, dynamic convolutional structures, and SE channel attention modules to achieve efficient perception of fetal structures in the images; performs feature-level concatenation of image features and structured clinical data, and completes FGR classification prediction through a multimodal fusion network; the system supports doctors uploading ultrasound images and instantly obtaining prediction results and visualized heatmaps of the model's areas of interest, improving the interpretability and auxiliary value of clinical judgment. Experimental results show that this method exhibits superior performance and robustness in multiple comparative models and ablation experiments. The system features high accuracy, strong scalability, and ease of use, and is suitable for auxiliary screening of fetal developmental abnormalities in primary and professional obstetric medical scenarios.
[0038] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description
[0039] Figure 1 This is a flowchart of a method for predicting fetal growth restriction based on deep learning and ultrasound images, according to an embodiment of this application.
[0040] Figure 2This is a schematic diagram of the working interface of an intelligent assisted diagnostic system after embedding a fetal growth restriction prediction model according to an embodiment of this application.
[0041] Figure 3 This is a schematic diagram of the structural composition of the fetal growth restriction prediction system provided according to an embodiment of this application;
[0042] Figure 4 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application, used to implement the deployment and execution of the method described in this application. Detailed Implementation
[0043] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.
[0044] The following describes, with reference to the accompanying drawings, a method and system for predicting fetal growth restriction based on deep learning and ultrasound images, according to embodiments of this application.
[0045] To address the issues mentioned in the background section regarding the current reliance on experience in identifying FGR (fetal growth restriction), the high degree of subjectivity in diagnosis, and the lack of interpretable AI-assisted systems, this application provides a deep learning-based decision-making method that integrates ultrasound images and clinical data. By constructing a deployable intelligent prediction model system, it achieves high-precision and interpretable identification of fetal growth status, thereby enhancing the consistency and scientific rigor of obstetric diagnosis.
[0046] Figure 1 This is a flowchart illustrating a method for predicting fetal growth restriction based on deep learning and ultrasound images, provided in an embodiment of this application. The method includes the following steps:
[0047] Step S101: Collect fetal ultrasound images and maternal clinical structured data from multiple centers, label the corresponding FGR status, and generate raw training data.
[0048] In this embodiment of the application, the data collection steps specifically include the following:
[0049] Two-dimensional grayscale ultrasound images of the fetus were acquired, covering standard sections commonly used in clinical practice to assess fetal development, such as head circumference (HC), abdominal circumference (AC), and femur length (FL). Image acquisition methods conformed to prenatal ultrasound examination technical specifications, and images were sourced from multiple medical centers, including provincial tertiary hospitals and primary care obstetric units. All images were derived from actual prenatal examination workflows, and the image format was uniformly PNG or DICOM, with image resolution controlled between 480×640 and 1024×768.
[0050] Collect structured clinical data corresponding to pregnant women, including but not limited to gestational age, pre-pregnancy BMI, fetal movement, amniotic fluid index (AFI), blood pressure, blood sugar, history of hypertensive disorders of pregnancy, history of gestational diabetes, history of previous adverse pregnancy outcomes, placental location, fetal heart rate monitoring score, etc. The above structured data is sourced from the prenatal examination electronic medical record system. After field mapping and format standardization, it is converted into a unified format input field. All fields are set with variable types and ranges according to the national clinical data standard set to ensure the requirements for subsequent modeling.
[0051] Collect the evaluation results after the fetus is born as the label signal in supervised learning. This label is determined by clinicians according to the standard that the birth weight is lower than the 10th percentile (<P10) of the same gestational age, and excluding non-FGR causes such as intrauterine infection, fetal structural abnormalities, and chromosomal abnormalities, and each sample is labeled as FGR or non-FGR.
[0052] Finally, complete the matching of image data, structured data, and labels to construct the original dataset. To ensure data quality, a quality control process is established to剔除 the following samples: cases with blurred images, cross-sections deviating from the standard plane, missing labels, or a missing rate of clinical data fields higher than 30%.
[0053] Step S102: Preprocess the initial data for images and clinical data to form standardized available training inputs.
[0054] Specifically, in an embodiment of the present application, the data preprocessing includes two parts: image preprocessing and structured clinical data preprocessing:
[0055] On the one hand, perform the following operations on the two-dimensional fetal ultrasound image data:
[0056] (1) Image size normalization: All input images are uniformly adjusted to a size of 512×512 pixels to ensure the consistency of different source images in the spatial scale and avoid the impact of size differences on the feature learning process. The image scaling uses bilinear interpolation to retain key structural details.
[0057] (2) Image grayscale normalization: Perform pixel value normalization on the original grayscale image to unify the grayscale value range to the interval from 0 to 1, enhance the robustness of the model to image contrast changes, and accelerate the convergence speed during the gradient descent process.
[0058] (3) Image enhancement: To improve the model's generalization ability and robustness to minor image perturbations, image enhancement techniques, including random rotation (±15 degrees), horizontal flipping, brightness perturbation, random cropping, and scaling, were used to expand the training set samples. All of the above enhancement operations were performed while keeping the image semantics unchanged, and the probability parameters of each enhancement operation were ensured to be within the control range (e.g., the execution probability of random rotation was 0.3, and the brightness adjustment range was controlled within ±15%).
[0059] On the other hand, the following processing steps are performed on structured clinical data:
[0060] (1) Missing value handling: For variables with a missing field ratio not exceeding 20%, use the mean to fill (continuous variables) or the mode to fill (categorical variables); remove fields with a missing rate exceeding the threshold or with unclear clinical significance.
[0061] (2) Standardization of numerical variables: For fields such as pregnant woman's age, weight, gestational age, amniotic fluid index, etc., the Min-Max normalization method is used to linearly map all values to the [0,1] interval to avoid the unbalanced impact of different units on the model weight update.
[0062] (3) Vectorization of categorical variables: variables such as history of gestational diabetes (yes / no), placental location (anterior / posterior / bottom), and history of cesarean section (yes / no) are vectorized using one-hot encoding so that categorical variables can participate in the model learning process.
[0063] After preprocessing the images and structured data, the process proceeds to the feature extraction stage. Image features are automatically learned end-to-end by a deep convolutional neural network, while structured clinical data is mapped into vectors of the same length as the image features through a multilayer fully connected network (MLP).
[0064] In the feature fusion process, a feature-level concatenation method is adopted to concatenate the high-dimensional semantic features of the image with the structured information vector in the channel dimension to form the final fused feature representation, which serves as the input to the classification model.
[0065] Step S103: Construct a multi-channel deep feature extraction network to extract image features and fuse clinical features to complete feature-level fusion.
[0066] In a preferred embodiment, the prediction model adopts a modular structure of "image coding network + clinical coding network + feature fusion network + classification output module". The overall framework has high scalability and strong feature expression capability, and is adaptable to multimodal input forms.
[0067] Specifically, the model structure is designed as follows:
[0068] (1) Image coding network: An improved ResNet architecture based on residual connection mechanism is adopted as the backbone for image feature extraction. This module contains multiple convolutional units composed of BasicBlocks, and the following structural enhancement modules are introduced in key layers:
[0069] Local attention mechanism module: Independently models local regions of the image by dividing the image into windows, captures fine-grained differences in fetal anatomical structures, and improves the model's ability to distinguish the edge regions of abdominal circumference and head circumference.
[0070] Global attention mechanism module: Employs a multi-head self-attention mechanism to model long-distance dependencies across the entire image, enhancing the model's comprehensive understanding of the overall image structure;
[0071] Dynamic convolution module: Adaptively generates convolution kernel parameters based on input samples, enabling the model to dynamically adapt to image differences (such as image brightness, background noise, device resolution, etc.).
[0072] The Squeeze-and-Excitation module introduces a channel-level attention mechanism. By explicitly modeling the dependencies between channels, it enhances important semantic feature channels, effectively improving the model's response strength to high-level semantics of images.
[0073] (2) Clinical information encoding network: Structured data is compressed and nonlinearly mapped through a two-layer fully connected neural network (FC). The activation function is ReLU, and the output dimension is consistent with the length of the image features to facilitate subsequent stitching.
[0074] (3) Feature fusion module: The feature vectors encoded from the image end and the structured data end are concatenated along the channel dimension and fed into the fusion classification network. The fusion network contains one or two fully connected layers and a batch normalization layer to prevent gradient vanishing and overfitting during training.
[0075] (4) Classification output module: After the fused features are regularized by Dropout, they are connected to the Softmax classification layer and output the probability distribution of belonging to FGR or not. The final predicted label is determined by the threshold (default is 0.5).
[0076] Step S104: Train, validate, and evaluate the model based on supervised learning to generate the final fetal growth restriction prediction model.
[0077] In this step, the system uses the fused input features as input to the deep classification model, trains the binary classification prediction model through supervised learning, and outputs the predicted probability of whether FGR exists.
[0078] During the model training phase, the embodiments of this application employ the following optimization and evaluation strategies:
[0079] Loss function: Cross-entropy loss is used as the model optimization objective, which is suitable for the current binary classification task;
[0080] Optimizer: Stochastic gradient descent optimizer with momentum (SGD with Momentum) is selected, where momentum is set to 0.9 and the initial value of the base learning rate is 1e-3;
[0081] Learning rate scheduling strategy: The Cosine Annealing Scheduler is used to control the learning rate to decay slowly from high to low, thereby improving training stability;
[0082] Regularization strategies: Dropout mechanism is added to the fully connected layer, with the Dropout ratio set to 0.3 to suppress overfitting; Batch Normalization is introduced in the Batch layer to accelerate convergence;
[0083] Training epochs: The total training epochs are set to 100 epochs. Each epoch performs an iterative update on the complete training set. The batch size is set to 32.
[0084] Step S105: A systematic evaluation process and comparative experiments are conducted to ensure the reliability, stability, and superiority of the proposed model.
[0085] After completing the structure construction and training process of the deep fusion model, in order to comprehensively evaluate the performance of the model in the fetal growth restriction prediction task, this application further designed and implemented a systematic evaluation process and comparative experiments to ensure the reliability, stability and superiority of the proposed model.
[0086] In one embodiment, model evaluation is divided into two stages: internal evaluation on the training set and external validation on an independent test set. Multiple evaluation metrics are used to comprehensively measure the model's classification performance. The specific process is as follows:
[0087] On the one hand, during the training phase, this application divides the original samples into a training set, a validation set, and a test set, with a ratio of 70%:10%:20%. Both training and validation samples are derived from actual prenatal checkup data from different centers to maintain diversity and mitigate distributional bias between samples. To further enhance the statistical robustness of the evaluation, a 5-fold cross-validation method is used to repeat the training and evaluation process, and the average result of the five experiments is taken as the final performance indicator.
[0088] On the other hand, during the external validation phase, data from centers not involved in the training were selected as an independent test set to examine the model's generalization ability in real-world scenarios. This type of external validation mechanism helps identify the risk of overfitting the model to data from specific centers and assesses its adaptability under different ultrasound equipment, physician procedures, and prenatal examination conditions.
[0089] To verify the effectiveness of the model structure, this application conducted the following two types of verification experiments:
[0090] Model comparison experiment: Performance comparison with mainstream image classification models such as ResNet50, CNN, ViT, and DenseNet121;
[0091] Ablation experiments: Local attention, global attention, dynamic convolution and SE modules were introduced step by step, and the contribution of each module was verified separately.
[0092] Performance evaluation metrics include AUC (area under the curve), accuracy, recall, and F1-score, ensuring that the final model achieves a reasonable balance between sensitivity and specificity.
[0093] Step S106: Model deployment and intelligent assisted diagnostic system design.
[0094] After completing model training and validation, to realize the clinical application of the fetal growth restriction prediction model, this application further constructed a complete intelligent auxiliary diagnostic system. This system can be used by obstetricians during routine prenatal checkups; after inputting standardized ultrasound images and the pregnant woman's clinical data, it quickly outputs a risk prediction result regarding the presence of fetal growth restriction.
[0095] In one embodiment, the system is developed based on a web service architecture, adopts a modular design, and includes the following functional components:
[0096] Image Upload and Format Verification Module: This module is used to receive fetal ultrasound images uploaded by users. It supports common medical image formats such as JPG, PNG, and DICOM. The system automatically verifies technical parameters such as image size, number of grayscale channels, and compression rate. If the image does not conform to the standard format, a prompt will be given and the image will be rejected.
[0097] Clinical Information Entry Module: Doctors can enter structured clinical data of pregnant women in the form format on the system's front-end interface, including but not limited to gestational age, weight, blood pressure, pregnancy history, fetal movement, and amniotic fluid volume. The system has automatic field recognition and unit verification functions to ensure the accuracy and completeness of data entry.
[0098] Data preprocessing module: The backend service automatically performs standardization processing on image data, including grayscale normalization, size adjustment, and image enhancement; and performs preprocessing steps such as normalization, encoding, and tensor transformation on structured data. All processing flows are consistent with the model training phase to ensure inference consistency.
[0099] Model inference module: The preprocessed image and structured data are fed as joint input into the trained deep fusion prediction model. The model returns the FGR prediction probability value and presents it as a binary classification result to determine whether the fetus is at risk of growth restriction.
[0100] Results Display and Interpretation Module: The system will feed back the prediction results to the doctor's interface as either output probabilities or output results.
[0101] Historical record management module: The system supports the automatic saving of image input, structured data and model output results for each prediction operation. Doctors can access historical results at any time for follow-up evaluation or academic research.
[0102] In terms of technical implementation, the system's backend is written in Python, with the core deep learning module based on the PyTorch framework. The inference service is deployed via ONNX to improve computational efficiency. The frontend is built using the lightweight web framework Streamlit, featuring a simple and intuitive interface for easy user operation. The entire system can be deployed on a local server or remotely in the cloud, adapting to the network and hardware environments of different medical institutions.
[0103] Step S107: Implementation of the electronic device and computer-readable storage medium.
[0104] In one embodiment, this application also provides an electronic device for implementing a method for predicting fetal growth restriction. The electronic device can be deployed in a hospital's clinical information system, cloud server, or medical workstation to load the deep learning prediction model and auxiliary diagnostic system provided in this application and perform functions such as image recognition, data processing, and result presentation.
[0105] The electronic device includes, but is not limited to, servers, terminals, laptops, tablets, portable ultrasound devices, or embedded processing platforms with certain computing capabilities, and includes at least the following components:
[0106] The memory is used to store computer program instructions and model parameter files. The computer program is used to implement the FGR prediction method as described in the foregoing embodiments, including steps such as data acquisition, preprocessing, feature fusion, model inference and result interpretation.
[0107] The processor is used to execute the computer program, perform calculations on the input images and clinical data, execute the model inference process, and output a prediction result on whether the fetus is at risk of growth restriction.
[0108] The communication interface is used to connect with the hospital's HIS, PACS, and RIS systems to enable the retrieval of image and medical record data, as well as the uploading and sharing of processing results.
[0109] The display device is used to show users prediction results and historical prediction records, and the interface supports graphic and text interaction and operation feedback.
[0110] In a preferred embodiment, the processor may be a general-purpose CPU, GPU, AI acceleration chip, or FPGA, used to meet the real-time and computational resource requirements of model inference. The memory may be an SSD or RAM, with sufficient capacity to store high-resolution image data and intermediate model cache.
[0111] In another embodiment, this application also provides a computer-readable storage medium storing a computer program for implementing the above-described method, which, when executed by a processor of an electronic device, implements the fetal growth restriction prediction method as defined in the claims of this application.
[0112] The computer-readable storage media include, but are not limited to: ROM (Read-Only Memory), RAM (Random Access Memory), magnetic disk, optical disk, flash memory, EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), USB flash drive, or other media capable of storing information electronically.
[0113] Through the design of the aforementioned electronic devices and computer-readable storage media, the FGR prediction method proposed in this invention not only has clinical applicability but also good cross-platform deployability. It can be widely applied in prenatal examination departments of large hospitals, primary healthcare institutions, and remote fetal health monitoring systems, providing important auxiliary support for the early identification of fetal abnormal development risks.
[0114] It should be understood that the terms "embodiments," "exemplary implementations," "certain implementations," or "specific implementation methods" used in this specification are only for illustrating the technical solutions in this application and are intended to present multiple alternative implementation approaches, and should not be construed as limiting the technical solution. Related features, structures, processes, or parameters can be combined and applied in various ways in different implementations, and unless the context explicitly limits them, these terms should not be construed as referring to only one specific solution.
[0115] Similarly, the terms "first," "second," and "third" used in this specification are merely used to distinguish different technical objects or steps, and do not represent any priority, execution order, or technical importance, nor should they be construed as meaning that the object or step must be unique. For example, defining something as "first module" does not preclude the existence of structural units with similar functions, such as "second module" or "third module."
[0116] Furthermore, the flowcharts, module divisions, or functional steps described in this application are not limited to being executed strictly in the listed order. Unless the process dependencies are clear, those skilled in the art can adjust the order of the steps as needed, or execute multiple steps in parallel or in an integrated manner while achieving equivalent technical effects. Similarly, the distribution of various functional modules in the system can be physically independent, or they can be integrated into a unified logical unit through a programmatic approach.
[0117] The term "computer-readable medium" as used in this application refers to various media that can store program code for controlling electronic devices to perform method steps. Such media include tangible carriers (such as ROM, RAM, disks, optical discs, USB flash drives, and flash memory), data stream files transmitted over a network in the form of electrical signals, and even printed paper or other visual symbol media that can be recognized and transcribed into electronic instructions.
[0118] Without affecting the overall technical solution of this application, the function can be implemented by hardware logic circuits, such as gate circuits, application-specific integrated circuits (ASICs), programmable logic arrays (FPGAs, CPLDs), etc.; it can also be implemented by software modules or firmware programs running on general computing platforms, or by a hybrid hardware and software approach.
[0119] Furthermore, the functional modules, process units, and operating steps described in this application can be tailored, merged, modified, or replaced according to specific scenarios during actual product deployment. Those skilled in the art can adapt them without inventive effort; therefore, such changes should be considered to fall within the scope of protection of this application.
Claims
1. A method for predicting fetal growth restriction based on deep learning and ultrasound images. The feature includes the following steps: We collected ultrasound images and clinical structured data from multiple centers of pregnant women, labeled them with whether they were fetal growth restricted (FGR), and generated initial training data. The initial data is preprocessed, including image normalization, resizing, and enhancement transformation, and the clinical data is standardized and missing value is handled. The processed image is input into a deep neural network model that integrates local attention, global attention, dynamic convolution, and SE modules to extract image features; The processed clinical data is converted into feature vectors and then fused with image features at the feature level. A deep fusion model is trained using supervised learning to output a prediction of whether fetal growth is restricted. The trained model is deployed to an online assisted diagnostic system that doctors can operate, and a visual heatmap is generated to explain the areas of interest of the model.
2. The method for predicting fetal growth restriction based on deep learning and ultrasound images according to claim 1, characterized in that: The ultrasound images are standard ultrasound cross-sectional images including the following areas: two-dimensional grayscale fetal abdominal circumference, head circumference, and femur length. The ultrasound images are derived from routine prenatal examination data from multiple hospitals.
3. The method for predicting fetal growth restriction based on deep learning and ultrasound images according to claim 1, characterized in that: The image feature extraction process includes: inputting the image into a backbone with a stacked structure of 3×3 convolutional layers, BatchNorm, ReLU and BasicBlock, and performing multi-scale feature encoding by combining local attention and global attention mechanisms.
4. The method for predicting fetal growth restriction based on deep learning and ultrasound images according to claim 1, characterized in that: The clinical structured data includes the following indicators: gestational age, maternal age, weight, fetal movement, blood pressure, and history of diabetes. These data are standardized and transformed into fixed-dimensional vector representations.
5. The method for predicting fetal growth restriction based on deep learning and ultrasound images according to claim 1, characterized in that: Feature-level fusion employs a feature concat operation, which connects image features and clinical features along the channel dimension before inputting them into the classification module.
6. The method for predicting fetal growth restriction based on deep learning and ultrasound images according to claim 1, characterized in that: Deep neural networks are trained using the cross-entropy loss function and optimized for model performance by combining a stochastic gradient descent optimizer with a dynamic learning rate scheduler.
7. The method for predicting fetal growth restriction based on deep learning and ultrasound images according to claim 1, characterized in that: Image data augmentation strategies, including rotation, translation, mirroring, and contrast adjustment, are introduced during training to improve the robustness of the model under different shooting conditions.
8. The method for predicting fetal growth restriction based on deep learning and ultrasound images according to claim 1, characterized in that: Grad-CAM heatmaps are generated based on backpropagation of convolutional feature maps and output class gradients, marking image regions where the model contributes to the diagnostic results.
9. The method for predicting fetal growth restriction based on deep learning and ultrasound images according to claim 1, characterized in that: The training set, validation set, and test set were divided in a ratio of 7:1:
2. Multiple mainstream model comparison experiments and ablation experiments were conducted on the test set to verify the effectiveness of the model.
10. A system for predicting fetal growth restriction based on deep learning and ultrasound images, characterized in that: include: The acquisition module is used to acquire two-dimensional ultrasound images of pregnant women and corresponding clinical structured data, and to label the FGR status to generate a training dataset. The preprocessing module is used to perform scale normalization and data augmentation on images, and to perform vector encoding on clinical data; The feature extraction module is used to extract image features that combine local attention, global attention, dynamic convolution, and SE modules; The fusion module is used to stitch and fuse image features with clinical vectors; The training module is used to train the FGR prediction model based on fused features; The reasoning and explanation module is used to output the prediction results; The interface module is used for doctors to upload images, display results, and provide interactive feedback.
11. An electronic device for predicting fetal growth restriction based on deep learning and ultrasound images, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the FGR prediction method as described in any one of claims 1-9.
12. A computer-readable storage medium for a method of predicting fetal growth restriction based on deep learning and ultrasound images, characterized in that: The program is executed by the processor to implement the FGR prediction method as described in any one of claims 1-9.