Fetal congenital heart disease artificial intelligence multi-dimensional diagnosis system and diagnosis method

Abstract

The invention discloses a fetal congenital heart disease artificial intelligence multi-dimensional diagnosis system and a fetal congenital heart disease artificial intelligence multi-dimensional diagnosis method. The system comprises a multi-source heterogeneous special disease database module, a standard section intelligent identification module, a precise diagnosis and risk assessment module and an intelligent auxiliary decision making and cloud platform module. The method comprises the following steps: constructing a special disease database; automatically identifying a standard section in the ultrasonic video by using an AI model; carrying out segmentation quantization on the tangent plane; image features and clinical data are fused for multi-dimensional diagnosis; and generating a structured report. According to the method, the multi-modal data and the advanced AI algorithm are integrated, so that automatic and precise auxiliary diagnosis of the fetal congenital heart disease is realized, the missed diagnosis and misdiagnosis rate is effectively reduced, and the method is particularly suitable for improving the prenatal screening capability of primary medical institutions.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, and in particular to an artificial intelligence multidimensional diagnostic system and method for fetal congenital heart disease (CHD). Background Technology

[0002] Congenital heart disease is the leading cause of birth defects and mortality in my country. National maternal and child health surveillance data shows its incidence is increasing year by year, placing a heavy economic and emotional burden on affected children's families and posing a challenge to public health security. Currently, clinical screening and diagnosis of fetal congenital heart disease mainly rely on prenatal ultrasound examinations, such as fetal echocardiography, but this technology has significant limitations: High diagnostic dependence: There are many types of congenital heart disease in fetuses, such as atrial septal defect, tetralogy of Fallot, transposition of the great arteries, etc., and the quality of ultrasound images is greatly affected by fetal position, gestational age, amniotic fluid content and operator experience. Due to the lack of professional personnel, the rate of missed diagnosis and misdiagnosis in primary medical units remains high, and the level of diagnosis and treatment varies significantly in different regions.

[0003] Weak data integration capabilities: The diagnosis of congenital heart disease in fetuses requires the combination of ultrasound images, such as two-dimensional ultrasound BMUS, color Doppler CDFI, three-dimensional ultrasound 3D US, clinical indicators, such as gestational age, maternal history of diabetes / hypertension, epidemiological data, such as genetic history, exposure to high-risk environments, and postpartum follow-up information, such as surgical results. However, current technologies lack the ability to standardize the integration and collaborative analysis of multi-source heterogeneous data, making it difficult to form a diagnostic basis for the entire life cycle.

[0004] Limitations of AI technology applications: Existing research on artificial intelligence in the field of fetal echocardiography focuses on single sections, such as the identification of four-chamber views, without covering the entire chain of automatic section selection, image quantitative analysis, multimodal diagnosis, and clinical application. Moreover, the models mostly rely on single-center data, have poor generalization ability, cannot adapt to the complex clinical scenarios of primary hospitals, and lack integration with cloud computing and 5G technologies, making it difficult to achieve widespread promotion.

[0005] Difficulties in implementing medical resources at the grassroots level: uneven distribution of medical resources, difficulty for patients in remote areas to obtain guidance from experts in higher-level hospitals, existing diagnostic tools require professional operation, and there are security risks in data transmission and sharing, which prevents high-quality medical resources from reaching the grassroots level.

[0006] Therefore, there is an urgent need for a multi-dimensional diagnostic system and method for fetal congenital heart disease that can integrate multi-source data, achieve intelligent diagnosis across the entire process, and be adapted to grassroots application scenarios, in order to solve the aforementioned technical pain points. Summary of the Invention

[0007] The primary objective of this invention is to overcome the shortcomings of existing technologies and provide a fetal congenital heart disease artificial intelligence multidimensional diagnostic system and corresponding diagnostic methods that can improve diagnostic accuracy and reduce reliance on doctors' experience. This system enables fully automated or semi-automated analysis from standard section recognition to precise disease diagnosis and risk assessment, and can be easily deployed in medical institutions at different levels to promote the downward flow of high-quality medical resources.

[0008] To achieve the above objectives, the present invention provides an artificial intelligence multidimensional diagnostic system for fetal congenital heart disease, comprising: A multi-source heterogeneous disease-specific database module is used to store and manage standardized fetal cardiac multimodal clinical data; The standard section intelligent recognition module is connected to the disease-specific database module and is used to receive fetal cardiac ultrasound image data and automatically identify the standard cardiac sections in it using the first artificial intelligence model. The precision diagnosis and risk assessment module is connected to the standard section intelligent recognition module and the disease database module. It is used to analyze the standard cardiac section using a second artificial intelligence model, integrate clinical data, construct a multi-task diagnostic model based on graph neural network, and output the risk assessment and disease classification diagnosis results of congenital heart disease. The intelligent decision support and cloud platform module integrates the above modules and is used to provide user interaction, process requests, generate diagnostic reports, and support model updates.

[0009] Furthermore, the multi-source heterogeneous disease database module utilizes natural language processing technology and medical knowledge graphs to perform structuring and standardization on unstructured text data.

[0010] Furthermore, the first artificial intelligence model is a dynamic target recognition algorithm that integrates spatiotemporal context information, including a feature extraction unit based on residual networks and a video sequence analysis unit based on spatiotemporal convolutional neural networks.

[0011] Furthermore, the second artificial intelligence model includes: Image segmentation sub-model is used to perform pixel-level segmentation of standard cross-sectional images and extract quantitative parameters of the heart structure; The diagnostic sub-model employs a graph neural network to fuse the aforementioned quantitative parameters, radiomics features, and clinical data, performing multi-task learning to output diagnostic results. The graph neural network includes a joint loss function consisting of a risk assessment loss function and a disease classification loss function. The risk assessment loss function is a binary classification cross-entropy loss function, expressed as: ; in The total number of samples in the training batch. Let i be the true label of the i-th sample. The model is used to predict the disease probability of the i-th sample.

[0012] The disease classification loss function is a multi-class cross-entropy loss function, expressed as follows: ; in The total number of disease categories. Let c be the true label of the disease category corresponding to the i-th sample. To predict the disease category of the i-th sample using the model The probability of.

[0013] The joint loss function is expressed as follows: ; in For risk assessment task weighting coefficients, This is the weighting coefficient for the disease subtyping task, used to balance the influence of the two tasks.

[0014] Furthermore, the intelligent decision-making assistance and cloud platform module is built on cloud computing and 5G communication technologies and can be deployed in a private cloud or hybrid cloud mode. The structured diagnostic report includes identified standard profiles, quantitative indicators, diagnostic conclusions, and personalized suggestions.

[0015] On the other hand, the present invention also provides an artificial intelligence multi-dimensional diagnostic method for fetal congenital heart disease applied to the above-mentioned system, comprising the following steps: S1: Construct a standardized disease-specific database containing multimodal clinical data; S2: Utilize the first artificial intelligence model to automatically analyze fetal echocardiography videos, identify and extract standard cardiac cross-sectional images; S3: Use the second artificial intelligence model to segment the standard cross-sectional image and quantitatively extract the morphological parameters of key cardiac anatomical structures; S4: Integrate the morphological parameters, radiomics features and clinical information to construct a multi-task diagnostic model based on graph neural networks, and output the risk probability of congenital heart disease and the diagnosis of disease type. S5: Generate structured decision support reports.

[0016] Furthermore, in step S2, the first artificial intelligence model filters high-quality standard sections by analyzing the spatiotemporal context information of the ultrasound video.

[0017] Furthermore, in step S3, the standard cross-sectional image is segmented using the second artificial intelligence model, and the morphological parameters of key cardiac anatomical structures are quantitatively extracted as follows: S31, a deep fully convolutional segmentation neural network is used to perform pixel-level segmentation on the input multimodal ultrasound image, wherein: the neural network adopts an encoder-decoder architecture, the encoder gradually reduces the spatial dimension of the image through pooling layers and extracts feature descriptors from different scales; the decoder gradually restores the spatial dimension of the target details through deconvolution operations and uses skip connections to improve the coarseness of the upsampling process to achieve category judgment for each pixel; in the shallow representation of the segmentation process, blood flow information from Doppler ultrasound images is fused to improve the efficiency and robustness of the segmentation model; S32. The segmentation results obtained in S31 are post-processed and optimized based on Conditional Random Field (CRF), where: a fully connected CRF model is applied, feature representation is enhanced by double Gaussian kernels, the size and shape of homogeneous category regions are learned, local noise and artifacts are suppressed, and regularization constraints are introduced to obtain more accurate structured segmentation results; combined with prior knowledge of fetal cardiac anatomy, the segmented images are automatically identified and quantitatively analyzed to extract various signs including heart location, atrial location, atrioventricular connection, ventricular loop, aortic-ventricular connection, and venous-atrial connection; at the same time, based on the segmentation results, radiomics methods are used to extract multi-dimensional image features such as morphological, statistical, and texture features.

[0018] Furthermore, in step S4, the morphological parameters, radiomics features, and clinical information are integrated to construct a multi-task diagnostic model based on a graph neural network. Specifically, the radiomics features extracted in S3, the quantitative analysis results, and the patient-related clinical data are used as input. A graph neural network is used to model the association between different feature indicators and disease types. The graph neural network maps discrete features to nodes and edges in a graph structure, capturing the dependencies and constraints between features to form a network topology spectrum for disease diagnosis. The diagnostic task is decomposed into two sub-tasks: fetal congenital heart disease risk assessment and specific disease classification. Through a multi-task collaborative training mechanism, these two sub-tasks are optimized simultaneously to reduce the impact of redundant indicators and improve the robustness and prediction accuracy of the model.

[0019] Furthermore, in step S5, the report is presented through a cloud platform and supports remote consultation and data sharing.

[0020] The beneficial effects of this invention are: For the first time, multimodal data (images, clinical texts, follow-up information) are deeply integrated with multi-level AI models (section recognition, structural segmentation, and disease diagnosis). In particular, by introducing a risk assessment binary classification cross-entropy loss function and a disease subtyping multi-class cross-entropy loss function into the graph neural network diagnostic model, combined with a weighted collaborative joint loss function, the system achieves precise optimization of both the risk of congenital heart disease and specific disease subtyping. This allows the model to more efficiently capture the correlation between features and diseases, and the end-to-end, multi-dimensional intelligent analysis from image preprocessing to final diagnosis is more targeted, significantly improving diagnostic accuracy and comprehensiveness, and reducing diagnostic errors caused by single-task model bias.

[0021] The system employs a dynamic section recognition algorithm that integrates spatiotemporal information, a fine-grained segmentation network combined with conditional random fields, and a multi-task diagnostic model based on graph neural networks, forming a complete technology chain. Specifically, the graph neural network, through the mapping of discrete features to graph structures and the synergistic optimization of dual tasks using a customized loss function, effectively addresses complex challenges in clinical practice such as inconsistent image quality, fetal movement, and high data heterogeneity. Furthermore, the flexible adjustment of weight coefficients in the joint loss function can adapt to the diagnostic needs of different clinical scenarios, further improving the system's adaptability and generalization ability to complex clinical data, and reducing diagnostic bias caused by extreme cases or data imbalance.

[0022] By building an integrated cloud platform, the results of AI models, including those optimized with graph neural network loss functions, are transformed into easy-to-use clinical support tools that are seamlessly integrated into existing workflows. Structured diagnostic reports intuitively present risk probabilities and classification results calculated based on accurate loss functions, helping primary care physicians quickly grasp core diagnostic criteria and effectively reducing reliance on physician experience. Simultaneously, the cloud platform's private / hybrid cloud deployment mode, combined with 5G technology, ensures efficient data transmission, addressing the issues of insufficient professional personnel and difficulties in data sharing in primary healthcare units. This effectively enhances the diagnostic capabilities of primary hospitals and promotes the implementation of hierarchical medical services.

[0023] The modular design of the system not only facilitates the subsequent integration of new AI algorithms or the expansion of disease diagnosis types, but also reserves flexible optimization space for graph neural network diagnostic models. It can dynamically adjust the weight coefficients in the joint loss function based on feedback from multi-center clinical data, or supplement the dimensions of the loss function based on new disease types, so as to achieve continuous iteration of the model. The cloud platform architecture supports real-time updates and collaborative training of the model. As clinical data accumulates, the graph neural network model based on customized loss functions can continuously optimize its diagnostic performance to meet the development needs of medical research and clinical practice. Attached Figure Description

[0024] Figure 1 This is a standardized disease-specific database of multimodal clinical data in this embodiment of the invention.

[0025] Figure 2 This is a flowchart of the diagnostic method in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described in conjunction with specific embodiments thereof. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0027] This embodiment provides an artificial intelligence multi-dimensional diagnostic system for fetal congenital heart disease, including: Multi-source heterogeneous disease-specific database module: used to store and manage fetal cardiac multimodal data from different medical institutions. See [link to module]. Figure 1 The fetal CHD study involved a population and corresponding clinical data, including 5000 cases of normal fetal heart disease and 5000 cases of fetal CHD from outpatient or inpatient visits at no fewer than three hospitals. The fetal CHD group (comprising 10 groups of congenital heart diseases: fetal arrhythmia, atrial septal defect, ventricular septal defect, patent ductus arteriosus, double outlet right ventricle, tetralogy of Fallot, transposition of the great arteries, persistent truncus arteriosus, pulmonary stenosis, and aortic coarctation) included more than 500 cases in each group. A multi-center, multimodal, heterogeneous data collection and structuring scheme was feasible, and a specialized fetal CHD database was constructed. Exclusion criteria: no postpartum echocardiography results, no surgical, autopsy, or great vessel cast confirmation, or, if no such confirmation, inconsistent results after confirmation by two physicians with associate chief physician qualifications or above. Referring to the ISUOG Guidelines for Fetal Echocardiography, a standard section annotation scheme for fetal echocardiography was determined. Referring to the Expert Consensus on Detailed Risk Stratification Diagnostic Techniques for Fetal Echocardiography, a annotation scheme for fetal congenital heart disease risk and type was determined. Annotation rules and standards for precision diagnostic research on fetal congenital heart disease were also established. The data includes fetal echocardiography images (such as 2D ultrasound, color Doppler, and 3D ultrasound), clinical text data (such as gestational age and medical history), laboratory test data, postpartum follow-up information, and diagnostic gold standards (such as autopsy and surgical confirmation results). This module utilizes natural language processing technology and medical knowledge graphs to structure and standardize unstructured text data, constructing a high-quality disease-specific dataset.

[0028] Standard Section Intelligent Recognition Module: Connected to the disease-specific database module, this module receives fetal echocardiography video streams or image sequences. It incorporates a first artificial intelligence model based on deep learning. This model employs a dynamic target recognition algorithm that integrates spatiotemporal contextual information, such as combining residual networks with spatiotemporal convolutional neural networks, to automatically identify and filter standard cardiac section images that meet diagnostic requirements from the dynamic ultrasound video, such as four-chamber view, three-vessel view, and outflow tract view.

[0029] Fetal echocardiography has been widely used in clinical practice. Standard ultrasound sections are crucial for the diagnosis of fetal heart disease. However, in actual clinical settings, a large number of ultrasound images with different sections, depths, and angles are obtained during fetal ultrasound examinations. The standard sections used for disease diagnosis require a high level of professional knowledge and clinical experience from doctors, and there are also differences in gestational age, body size, and fetal movement.

[0030] In order to automatically identify standard sections during complex fetal cardiac dynamic ultrasound examinations, this embodiment proposes a multi-class dynamic target recognition algorithm, which integrates information from multiple dimensions to screen out high-quality standard sections from ultrasound examination videos. The algorithm mainly includes three parts: (1) Based on the residual network, an end-to-end category prediction model is constructed, which can perform hierarchical feature extraction on image information within the receptive field. The low-level features of ultrasound images extracted in the shallow part of the network express the common features of the image such as brightness and contour. By learning the weights of the large natural image dataset ImageNet, the learning of basic features of ultrasound images is effectively improved. In the deep part of the network, high-level features are extracted. The high-level features in different datasets have large heterogeneity. The stability of ultrasound image feature expression is enhanced by combining the max pooling layer and the fully connected layer. (2) Based on the spatiotemporal convolutional neural network, the spatiotemporal context information is utilized, and the long-term memory module effectively enhances the correlation information between frames before and after the ultrasound examination video, effectively eliminating non-standard sections such as motion blur and artifacts, and alleviating the problem of imbalance between standard and non-standard sections; through the correlation module, the global information of ultrasound video frames and the local information of independent images are learned simultaneously, improving the multi-dimensional discrimination ability of standard sections. (3) Through the feature attribute clustering module, category attribute learning is performed on the unevenly distributed ultrasound images, and the shared visual features between non-standard sections and ultrasound images of different standard sections are fully explored from the deep features, promoting the learning of different category attributes, and adding discriminable information and category relationship modules to enhance the discriminativeness of category embedding and semantic relevance, improve the fine-grained attributes for distinguishing different standard sections, and reduce the number of manually labeled samples.

[0031] Precision Diagnosis and Risk Assessment Module: Connected to the standard section intelligent recognition module and the disease-specific database module, this module analyzes the identified standard sections. It incorporates a second artificial intelligence model based on deep learning, configured as follows: a. Perform pixel-level segmentation on standard cross-sectional images, for example, using a fully convolutional network with an encoder-decoder structure to extract quantitative parameters of the heart structure, such as the size of the heart chambers, the thickness of the ventricular walls, and the diameter of the great vessels.

[0032] b. Integrate the quantitative parameters, radiomics features, and relevant clinical multimodal data from the disease-specific database module, and perform multi-task learning using a neural network (GNN) to output the risk assessment results of the fetus having congenital heart disease and specific disease classification diagnoses, such as atrial septal defect and tetralogy of Fallot.

[0033] Once the standard section identification for fetal heart examination is completed, it can improve the diagnostic efficiency and accuracy of routine fetal CHD by ordinary ultrasound physicians. However, for the diagnosis of non-routine patients, it often requires senior ultrasound physicians to make a diagnosis and may require quantitative and qualitative identification of the disease.

[0034] In the precision diagnosis and risk assessment module of this invention, a multi-task learning framework based on graph neural networks (GNNs) is adopted. This framework can simultaneously handle two tasks: risk assessment of fetal congenital heart disease and diagnosis of specific disease types. This multi-task learning structure design allows for the full utilization of shared features between different tasks, improving the overall performance of the model, especially in the face of scarce or imbalanced data.

[0035] 1. Structural Design This multi-task learning model decomposes the task into two sub-tasks: fetal congenital heart disease risk assessment and disease classification diagnosis, and trains these two tasks simultaneously in the same network. By constructing synergistic effects between tasks in the graph neural network, the model can share learned features, thereby achieving knowledge transfer and enhancement between multiple tasks.

[0036] Specifically, Graph Neural Networks (GNNs) capture the dependencies between different features by mapping input data to nodes and edges in a graph structure. In each layer of graph convolution operations, the model continuously aggregates adjacency information of nodes and optimizes feature representation through the graph's topological structure. This structure is particularly suitable for handling the complex relationships between multimodal clinical and imaging data related to fetal congenital heart disease.

[0037] 2. Loss Function Design To effectively optimize the two sub-tasks in multi-task learning, this invention designs a joint loss function that comprehensively considers risk assessment and disease classification. This loss function consists of the following two parts: Risk Assessment Loss: For the risk assessment task of congenital heart disease in fetuses, we adopted a binary cross-entropy loss function, the goal of which is to maximize the accuracy of risk assessment, i.e., correctly predict the probability of the fetus having congenital heart disease. This loss function optimizes the model parameters by measuring the difference between the predicted value and the actual label.

[0038] This function is used for a binary classification task to determine whether a fetus has congenital heart disease. Parameter definition: : The loss value of the risk assessment task. The smaller the value, the more accurate the model's judgment of health / illness.

[0039] The total number of samples in a training batch (i.e., the number of cases of fetal echocardiography).

[0040] Sample index, .

[0041] : No. The true label of each sample. This is a binary value. = 1 indicates that the fetus has been diagnosed with congenital heart disease. = 0 indicates good health.

[0042] Model for the first The predicted probability of a sample having congenital heart disease. This is a continuous value between 0 and 1, generated by the model's Sigmoid output layer.

[0043] Disease Subtyping Loss: For disease typology diagnostic tasks, we employ a multi-class cross-entropy loss function, aiming to optimize the model's prediction accuracy across different disease subtypes. This loss function is calculated by comparing the output probabilities of the multi-class classification with the true labels.

[0044] This function is used to specifically determine the type of congenital heart disease the fetus has (such as atrial septal defect, tetralogy of Fallot, etc.): ; Parameter definition: : Loss value for the disease classification task. The smaller the value, the more accurate the model's judgment of the specific disease type.

[0045] The total number of samples in a training batch.

[0046] : Total number of disease categories (based on your disease database, C=10, corresponding to 10 groups of congenital heart disease).

[0047] Sample index, .

[0048] Disease category index .

[0049] : No. Each sample represents a disease category. The true label. This is a one-hot encoded vector, representing the true diagnosis of the sample as the class. hour, = 1, otherwise 0.

[0050] Model prediction of the first Each sample belongs to a disease category The predicted probability. This is a continuous value between 0 and 1, with the sum of the predicted probabilities of all C categories being 1, generated by the model's Softmax output layer.

[0051] 3. Comprehensive Loss Function To optimize the learning process for both tasks simultaneously, we weighted and summed the risk assessment loss and disease classification loss to form the final joint loss function. This function combines the losses from the two tasks through weighted summation, enabling the model to simultaneously learn "whether the patient has the disease" and "what specific disease it is."

[0052] Parameter definition: The overall joint loss of the model. The goal of training is to minimize this value.

[0053] Risk assessment loss as defined above.

[0054] The disease subtyping loss defined above.

[0055] : Weighting coefficient for the risk assessment task. Used to control the contribution of the loss from this task to the total loss.

[0056] : Weighting coefficients for the disease subtyping task.

[0057] These are weighting coefficients for the risk assessment and disease classification tasks, used to balance the influence of the two tasks. By adjusting these coefficients, the model can be flexibly adjusted to suit the performance of the two tasks in practical applications.

[0058] 4. Training Strategies During training, gradient descent (such as the Adam optimizer) is used to minimize the overall loss function. Through backpropagation, the model adjusts its parameters in each iteration based on the gradient of the current loss, thereby progressively optimizing the predictive ability of the multi-task learning framework. To avoid overfitting, we also incorporate an early stopping strategy to ensure that the model's performance on the validation set remains optimized.

[0059] To achieve comprehensive and accurate diagnosis of different heterogeneities in fetal CHD, this invention proposes a multi-stage fusion task prediction framework, which mainly includes the following three parts: (1) Constructing a multimodal ultrasound deep fully convolutional segmentation neural network. Without other auxiliary segmentation steps, the input image can be completely segmented through two steps: encoder and decoder. The encoder uses pooling layers to gradually reduce the spatial dimension of the input image and extracts image feature descriptors from different scales. The decoder gradually restores the spatial dimension of the target details through deconvolution and uses skip links to improve the coarsening of the upsampled spatial restoration, thereby realizing the category judgment of each pixel in the input image. Since Doppler ultrasound images contain blood flow information, fusing blood flow information in the shallow features of the segmentation can improve the segmentation efficiency and robustness of the model. (2) Post-processing of multi-target quantization analysis based on conditional random field (CRF). Since the fully connected convolutional neural network shares a lot of spatial context information between adjacent pixels, the segmentation result is often relatively smooth. Therefore, a fully connected conditional random field is added as a post-processing step. The feature expression of the average approximation field is improved by two Gaussian kernels. The size and shape of homogeneous category regions are learned in the feature space to reduce the influence of local and stray noise in the input image. Regularization constraint relationship is introduced to achieve more accurate structured segmentation results. Combined with the prior information of anatomical structure, more accurate automatic identification and quantitative analysis of various signs such as fetal heart position, atrial position, atrioventricular connection, ventricular loop, aortic ventricular connection, and venous atrial connection in ultrasound are achieved. At the same time, radiomics methods are combined to extract multi-dimensional features such as morphology, statistics, and texture phenotype. (3) A graph neural network that integrates multi-dimensional clinical information is used to carry out a multi-task diagnostic model. The graph neural network constructs the relationship between different indicators and diseases, and combines the actual clinical situation to decompose the task into the risk of congenital heart disease and specific disease type classification. The graph neural network can convert the auxiliary relationship in different discrete feature domains into graph connections, capture the dependency relationship and constraint relationship between different features, and while retaining the feature disease spectrum network topology, continuously aggregate node feature information to form a disease diagnosis network spectrum. Since there is an inherent link between the risk of congenital heart disease in fetuses and the fetal congenital heart disease classification model, simultaneous training by constructing synergistic effects between tasks can effectively reduce redundant indicators, improve the robustness of the model, and mutually promote the predictive accuracy of the tasks.

[0060] Intelligent Decision Support and Cloud Platform Module: Integrates the above modules and provides a user interface. This module is built on cloud computing and 5G communication technologies and is capable of: a. Receive inspection data uploaded by users.

[0061] b. Call the standard cross-section intelligent recognition module and the precise diagnosis and risk assessment module for processing and analysis.

[0062] c. Generate a structured diagnostic report, which includes identified standard cross-sectional images, quantitative indicators of cardiac structure, diagnostic conclusions, risk assessments, and personalized management recommendations.

[0063] d. Supports the updating and optimization of the parameters of the aforementioned artificial intelligence model. This module can be deployed in a private cloud (suitable for a single hospital) or hybrid cloud (suitable for regional medical consortia or multi-center applications) mode.

[0064] By employing an AI-powered fetal CHD diagnostic prediction model, and combining it with postpartum imaging, intraoperative results, anatomical castings, and clinical follow-up information, the accuracy of the CHD diagnostic prediction model was comprehensively evaluated and feedback was provided for correction. A diagnostic and treatment process / standards for fetal CHD were proposed, broadening the application scope and diagnostic efficacy of AI in fetal CHD diagnosis. The project's implementation reduces the limitations of ultrasound techniques and operator experience in prenatal diagnosis. Utilizing 5G, cloud computing, and other engineering technologies, without interfering with the actual diagnostic process, the project standardizes the use of the predictive model for fetal cardiac examinations at different gestational weeks by both patients and doctors. Structured reporting methods are used to provide diagnostic conclusions and personalized prognostic assessment reference plans, and training and promotion are conducted in hospitals at different levels and regional medical consortia. For in-hospital applications, a private cloud platform is built, directly acquiring data through the hospital information management system, reducing intermediate data processing and transmission time and clinical information leakage issues. For regional or multi-center applications and patient-side applications, a hybrid cloud platform is built, using a clinical electronic data acquisition system as data middleware to standardize clinical data from different sources, reducing the risk of clinical information leakage and improving the platform's application scope and value.

[0065] Secondly, this invention provides an artificial intelligence-based multidimensional diagnostic method for congenital heart disease in fetuses applied to the aforementioned system, see [link to relevant documentation]. Figure 2 This includes the following steps: S1: Data Collection and Construction: Gather multi-center, multi-modal fetal heart-related clinical data to construct a standardized, multi-source, heterogeneous disease database.

[0066] S2: Automatic recognition of standard sections: Using a trained first artificial intelligence model, the input fetal echocardiogram video is analyzed in real time to automatically identify and extract high-quality standardized cardiac section images.

[0067] S3: Heart Structure and Feature Extraction: Using the image segmentation sub-model in the trained second artificial intelligence model, the standard cross-sectional image obtained in step S2 is finely segmented to quantitatively extract the morphological parameters of key cardiac anatomical structures.

[0068] S4: Multidimensional Fusion Diagnosis: The quantitative parameters and radiomics features extracted in step S3 are fused with the patient's clinical information (such as gestational age, maternal factors, etc.), and a comprehensive analysis is performed using the diagnostic sub-model (as shown in the neural network) in the second artificial intelligence model to output diagnostic results including the probability of congenital heart disease risk and the specific disease type.

[0069] S5: Generate a decision support report: Integrate the analysis results from steps S2 to S4 to automatically generate a structured diagnostic report, which is then presented to doctors through a cloud platform to support clinical decision-making.

[0070] Finally, the method described in this application is merely a preferred embodiment and is not intended to limit the scope of protection of this invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A fetal congenital heart disease artificial intelligence multi-dimensional diagnostic system, characterized in that, The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence.

2. The system of claim 1, wherein, The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence.

3. The system of claim 1, wherein, The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence.

4. The system of claim 3, wherein, The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence.

5. The system of claim 1, wherein, The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence.

6. A fetal congenital heart disease artificial intelligence multi-dimensional diagnostic method applied to the system of any one of claims 1-5, characterized in that, The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence.

7. The method of claim 6, wherein, The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence.

8. The method of claim 6, wherein, The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease diagnosis system based on multi-source heterogeneous databases and artificial intelligence. The application relates to a fetal heart disease S31, performing pixel-level segmentation on the input multi-modal ultrasound image using a deep fully convolutional segmentation neural network, wherein: the neural network adopts an encoder-decoder architecture, the encoder gradually reduces the spatial dimensions of the image through a pooling layer, and extracts feature descriptors from different scales; the decoder gradually restores the spatial dimensions of the target details through deconvolution operations, and uses a skip connection to improve the roughness of the upsampling process, so as to realize the class judgment of each pixel; in the shallow representation of the segmentation process, the blood flow information in the Doppler ultrasound image is fused to improve the efficiency and robustness of the segmentation model; S32, performing post-processing optimization based on conditional random field (CRF) on the segmentation result obtained in S31, wherein: a fully connected conditional random field model is applied, a double Gaussian kernel is used to enhance feature expression, the size and shape of the homogeneous class region are learned, local noise and artifacts are suppressed, and regularization constraints are introduced to obtain more accurate structured segmentation results; combining the prior knowledge of the fetal heart anatomical structure, the segmented image is automatically identified and quantitatively analyzed, and multiple signs including heart position, atrium position, atrioventricular connection, ventricular loop, aortic ventricular connection, and venous atrial connection are extracted; at the same time, based on the segmentation result, imageomics method is used to extract multi-dimensional image features such as morphological, statistical and texture features.

9. The method of claim 6, wherein, In step S4, the morphological parameters, imageomics features and clinical information are fused to construct a multi-task diagnosis model based on a graph neural network, specifically, the imageomics features, quantitative analysis results and clinical data related to the patient extracted in S3 are input, and the association between different feature indicators and disease types is modeled through a graph neural network; the graph neural network maps discrete features to nodes and edges in the graph structure, captures the dependency relationship and constraint condition between features, and forms a network topology spectrum for disease diagnosis; the diagnosis task is divided into two sub-tasks of fetal congenital heart disease risk assessment and specific disease classification, and the two sub-tasks are simultaneously optimized through a multi-task collaborative training mechanism, so as to reduce the influence of redundant indicators, improve the robustness and prediction accuracy of the model.

10. The method of claim 6, wherein, In step S5, the report is presented through a cloud platform, and remote consultation and data sharing are supported.

Images