Gestational diabetes auxiliary analysis system and method based on placenta ultrasonic texture

Abstract

The invention discloses a gestational diabetes auxiliary analysis system and method based on placenta ultrasonic textures, and the system comprises a multi-modal database which is used for storing placenta two-dimensional ultrasonic image data and clinical comprehensive data; the preprocessing module is used for screening the clinical comprehensive data and carrying out standardization processing and labeling on the image data; the image feature extraction module is used for constructing an image segmentation model based on a CNN network, performing image processing on the input placenta two-dimensional ultrasonic image through the image segmentation model, and obtaining an image feature data set; the risk prediction module is used for constructing a GDM risk prediction model; and the judgment module is used for carrying out risk classification on the GDM risk of the current pregnant woman by utilizing the image segmentation model and the GDM risk prediction model. By introducing a deep learning technology, intelligent analysis is performed on placenta ultrasonic images, clinical comprehensive data and other multi-modal data, and a reliable tool is provided for auxiliary analysis of gestational diabetes mellitus.

Description

Technical Field

[0001] The present invention relates to the technical field of medical multimodal data analysis, and in particular to a gestational diabetes auxiliary analysis system and method based on placental ultrasound texture. Background Art

[0002] Gestational diabetes mellitus (GDM) is one of the most common pregnancy complications, potentially posing a threat to both maternal and fetal health. Short-term consequences include pregnancy-induced hypertension, premature birth, fetal macrosomia, and postpartum hemorrhage. Long-term, GDM increases the mother's risk of type 2 diabetes and cardiovascular disease, and also increases the risk of obesity, diabetes, and metabolic syndrome in the offspring. Ultrasound examinations are primarily used to monitor fetal growth, placental position, and amniotic fluid volume. However, current placental ultrasound analysis typically does not include in-depth analysis of the placental ultrasound texture; the imaging data is primarily used for macroscopic assessment.

[0003] Currently, clinicians typically subjectively assess the risk of gestational diabetes mellitus (GDM) based on factors such as the pregnant woman's personal history, weight, blood sugar levels, family history, and placental ultrasound imaging. This approach relies heavily on the physician's experience and intuition, and is subject to certain errors and limitations. Furthermore, most traditional methods rely on blood sugar monitoring or an oral glucose tolerance test (OGTT) to analyze and predict gestational diabetes. However, this approach focuses solely on blood sugar levels and ignores other potential health factors of pregnant women, such as placental function and individual differences among pregnant women. Relying solely on blood sugar levels may not fully reflect the overall health of the pregnant woman, leading to undetected early or mild GDM.

[0004] Therefore, it is necessary to provide a gestational diabetes auxiliary analysis system and method based on placental ultrasound texture. By introducing automated image segmentation and deep learning models, combined with multimodal databases and clinical data, the problems of relying on manual judgment and a single data source in traditional solutions can be optimized. By efficiently and accurately extracting placental detail features, a more personalized and intelligent GDM risk analysis tool can be provided, providing more powerful technical support for the comprehensive utilization and analysis of multimodal data during pregnancy. Summary of the Invention

[0005] In view of this, the present invention provides a gestational diabetes auxiliary analysis system and method based on placental ultrasound texture, which is used to solve the technical problems that current placental ultrasound image analysis is mostly subjective, the risk of GDM is usually predicted based on a single data source, there is a lot of manual intervention, lacks universal standards and correlation analysis, and it is difficult to fully utilize and analyze multimodal data.

[0006] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a gestational diabetes mellitus auxiliary analysis system based on placental ultrasound texture, comprising:

[0008] A multimodal database for storing historical placental two-dimensional ultrasound image data and corresponding comprehensive clinical data of different pregnant women;

[0009] The preprocessing module is used to screen the comprehensive clinical data and establish a correlation analysis data set corresponding to the comprehensive clinical data and image data; it is also used to standardize and annotate each placenta two-dimensional ultrasound image data to construct an image segmentation data set;

[0010] An image feature extraction module is used to construct an image segmentation model based on a CNN network and train the image segmentation model using an image segmentation dataset to obtain a fully trained image segmentation model. The trained image segmentation model is used to process the input two-dimensional ultrasound image of the placenta and obtain an image feature dataset based on the processed image.

[0011] The risk prediction module is used to build a GDM risk prediction model. The risk prediction dataset is established using the image feature dataset and the association analysis dataset. The GDM risk prediction model is trained using the risk prediction dataset to obtain a fully trained GDM risk prediction model.

[0012] The judgment module is used to obtain the clinical comprehensive data corresponding to the two-dimensional ultrasound image dataset of the placenta of the current pregnant woman, and input it into the well-trained image segmentation model and GDM risk prediction model to analyze and predict the GDM risk of the current pregnant woman.

[0013] Furthermore, an image segmentation model based on the CNN network is constructed, including:

[0014] The image segmentation model includes a downsampling unit and an upsampling unit, and the downsampling unit and the upsampling unit are symmetrical structures;

[0015] The downsampling unit is a dual-branch encoder, comprising a first input layer, a second input layer, a low-layer shared encoder, and a high-layer independent encoder;

[0016] The first input layer is used to receive multi-section ultrasound images of the placenta to capture the placenta morphology from different perspectives, and the second input layer is used to receive two-dimensional ultrasound images of the placenta to obtain global texture information. Both the low-level shared encoder and the high-level independent encoder include multiple convolutional blocks, each of which consists of multiple consecutive convolutional layers, and the last convolutional block is connected to a maximum pooling layer. Each convolutional layer is connected to a batch normalization layer and a Relu activation layer. The low-level shared encoder shares the same set of weight parameters, while the high-level independent encoders remain independent, extracting hierarchical features and encoding the information within the receptive field.

[0017] The upsampling unit includes multiple decoders, each decoder includes multiple depooling layers, each depooling layer is connected to multiple convolutional layers, and the last decoder is connected to an output layer, which uses Softmax to generate pixel-level image segmentation results.

[0018] Furthermore, the dual-branch encoder includes a multi-kernel inverted residual module, which is used to extract multi-scale placental texture features in parallel through convolution kernels of different sizes, capturing local details and global morphology respectively.

[0019] Furthermore, an image feature dataset is obtained based on the processed image, including:

[0020] Perform statistical analysis on the segmentation features of the image to determine the imaging quantitative features and radiomics features, and obtain the imaging feature dataset based on the imaging quantitative features and radiomics features;

[0021] The imaging quantitative features include at least placental thickness, gray-level co-occurrence matrix, local binary pattern, placental hemodynamic parameters, placental vascular distribution and density, and placental echo characteristics;

[0022] The imaging genomics features include at least a pixel grayscale histogram, a texture feature map, a shape feature map, and a wavelet characteristic map.

[0023] Furthermore, the GDM risk prediction model is constructed based on a fully connected network with a cross-attention fusion mechanism, including a feature input layer, a feature fusion layer, and an output layer;

[0024] The feature input layer is used to input image feature data and clinical comprehensive data corresponding to the image feature data, and map the clinical comprehensive data into a continuous vector;

[0025] The feature fusion layer is used to calculate the cross-modal correlation weight using the image feature vector as the query and the clinical comprehensive data vector as the key, and to concatenate the image features with the attention-weighted clinical comprehensive data to form a fusion feature matrix;

[0026] The output layer uses the Sigmoid function to output the GDM risk analysis results.

[0027] Furthermore, the feature fusion layer is used to use the image feature vector as the query and the clinical comprehensive feature vector as the key to calculate the cross-modal correlation weight, and to splice the image features with the attention-weighted clinical comprehensive data vector to form a fusion feature matrix, including:

[0028] The clinical comprehensive feature vector is dynamically weighted based on the attention weight to generate a contextual representation aligned with the image features, which can be expressed as:

[0029]

[0030] Among them, Q is the image feature, K is the clinical comprehensive feature, V is the weighted sum feature, d k is the scaling factor.

[0031] Furthermore, the comprehensive clinical data includes at least basic health information, postpartum placental pathological examination, blood biochemical index data and delivery outcome data.

[0032] Furthermore, each placenta two-dimensional ultrasound image data is standardized and annotated to construct an image segmentation dataset, including:

[0033] Converting each acquired two-dimensional ultrasound image data of the placenta into a single image in a unified format;

[0034] Perform denoising and image enhancement operations on the images in unified format;

[0035] Each image is labeled to indicate whether the image comes from a GDM patient or a normal pregnant woman, and an image segmentation dataset is obtained.

[0036] Furthermore, the clinical comprehensive data is screened to establish a correlation analysis data set corresponding to the clinical comprehensive data and the imaging image data, including:

[0037] Clinical comprehensive data with no corresponding placental ultrasound images or incomplete images were eliminated;

[0038] Align the comprehensive clinical data with the corresponding imaging data according to gestational age;

[0039] Establish a temporal correlation between placental ultrasound images and clinical comprehensive data of the same period.

[0040] In another aspect, the present invention further provides a method for auxiliary analysis of gestational diabetes based on placental ultrasound texture, comprising:

[0041] Obtain historical placental two-dimensional ultrasound image data and corresponding comprehensive clinical data of different pregnant women;

[0042] Screen the comprehensive clinical data, establish a correlation analysis dataset corresponding to the comprehensive clinical data and image data, standardize and annotate each placenta two-dimensional ultrasound image data, and construct an image segmentation dataset;

[0043] Construct an image segmentation model based on a CNN network and train it using an image segmentation dataset to obtain a fully trained image segmentation model. Use the trained image segmentation model to process the input two-dimensional ultrasound image of the placenta, and obtain an image feature dataset based on the processed image.

[0044] Construct a GDM risk prediction model, use the image feature dataset and the association analysis dataset to establish a risk prediction dataset, and train the GDM risk prediction model with the risk prediction dataset to obtain a fully trained GDM risk prediction model;

[0045] The clinical comprehensive data corresponding to the current pregnant woman's placental two-dimensional ultrasound image dataset is obtained and input into the well-trained image segmentation model and GDM risk prediction model to classify the current pregnant woman's GDM risk.

[0046] Compared with the existing technology, the gestational diabetes auxiliary analysis system and method based on placental ultrasound texture proposed in the present invention has the following advantages:

[0047] (1) The present invention breaks through the limitations of a single data type and solves the data island problem in traditional methods by constructing a correlation analysis model between clinical data and imaging data. It improves the ability to integrate multi-source heterogeneous data and realizes cross-modal feature fusion (such as the correlation between ultrasound texture features and physiological indicators), providing a new technical path for the analysis of complex physiological states.

[0048] (2) The image segmentation model based on the CNN network replaces manual labeling, realizing automatic recognition of key areas (such as chorionic structure and vascular distribution) in placental ultrasound images and quantitative feature extraction (such as texture complexity and grayscale distribution entropy), significantly reducing manual operation errors and time costs.

[0049] (3) It supports dynamic feature tracking of continuous imaging data (such as texture evolution trends) to achieve early identification of changes in physiological status, which is superior to traditional static analysis methods. Through risk stratification management, medical resources can be allocated preferentially to high-concern groups.

[0050] In summary, the present invention solves the core problems of low medical image processing efficiency, insufficient data utilization, and poor equipment compatibility through multimodal data fusion, automated image analysis, and dynamic modeling technology, and provides technical solutions for the comprehensive analysis of multimodal data during pregnancy. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a schematic diagram of the structure of the gestational diabetes mellitus auxiliary analysis system based on placental ultrasound texture provided by the present invention;

[0052] Figure 2 A schematic diagram of the construction process of the multi-source heterogeneous database of early pregnancy GDM placenta provided by the present invention;

[0053] Figure 3 A schematic diagram of the process of the auxiliary analysis method for gestational diabetes based on placental ultrasound texture provided by the present invention;

[0054] Figure 4 The present invention provides a schematic diagram of the multimodal data processing flow. DETAILED DESCRIPTION

[0055] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0056] See Figure 1 This embodiment provides a gestational diabetes mellitus auxiliary analysis system 100 based on placental ultrasound texture, comprising:

[0057] A multimodal database 101 is used to store historical placental two-dimensional ultrasound imaging data and corresponding comprehensive clinical data of different pregnant women;

[0058] The pre-processing module 102 is used to screen the comprehensive clinical data and establish a correlation analysis data set corresponding to the comprehensive clinical data and the image data; it is also used to perform standardization and annotation on each placenta two-dimensional ultrasound image data to construct an image segmentation data set;

[0059] An image feature extraction module 103 is configured to construct an image segmentation model based on a CNN network, train the image segmentation model using an image segmentation dataset, obtain a fully trained image segmentation model, perform image processing on an input two-dimensional ultrasound image of the placenta using the trained image segmentation model, and obtain an image feature dataset based on the processed image;

[0060] The risk prediction module 104 is used to construct a GDM risk prediction model, establish a risk prediction dataset using the image feature dataset and the association analysis dataset, and train the GDM risk prediction model using the risk prediction dataset to obtain a fully trained GDM risk prediction model;

[0061] The judgment module 105 is used to obtain the clinical comprehensive data corresponding to the two-dimensional ultrasound image data set of the placenta of the current pregnant woman, and input it into the well-trained image segmentation model and GDM risk prediction model to classify the GDM risk of the current pregnant woman.

[0062] The system of this embodiment utilizes joint modeling of placental ultrasound imaging features and clinical data, overcoming the limitations of a single data source. A dual-branch CNN-based image segmentation model achieves precise multi-scale placental texture segmentation. By building a GDM risk prediction model, the system combines imaging and clinical data to comprehensively assess the overall health of pregnant women and provide a more comprehensive and accurate risk assessment. This system improves decision-making efficiency and accuracy through automated analysis, thereby reducing human judgment errors.

[0063] As a specific example, the data in multimodal database 101 includes 1,000 cases each of GDM and healthy control groups. The GDM group is further subdivided into two categories: those diagnosed in early pregnancy and those diagnosed in late pregnancy. This includes multimodal data such as prenatal ultrasound examinations, postpartum placental pathology examinations, blood biochemical indicators, and birth outcomes. The healthy control group will also be matched with non-GDM pregnant women of the same gestational age to ensure the reliability of the control group data. Through the collection and structuring of multi-center, multimodal, heterogeneous data, a disease-specific database covering the functional and structural changes of the GDM placenta during early pregnancy will be established. Data collection and storage will adhere to unified standards to ensure data quality and consistency.

[0064] As a preferred embodiment, the comprehensive clinical data at least includes basic health information, postpartum placental pathological examination, blood biochemical index data and delivery outcome data.

[0065] As a preferred embodiment, screening the comprehensive clinical data and establishing a correlation analysis data set corresponding to the comprehensive clinical data and the image data include:

[0066] Clinical comprehensive data with no corresponding placental ultrasound images or incomplete images were eliminated;

[0067] Align the comprehensive clinical data with the corresponding imaging data according to gestational age;

[0068] Establish a temporal correlation between placental ultrasound images and clinical comprehensive data of the same period.

[0069] As a specific example, when screening the comprehensive clinical data, clinical data without postpartum placental pathology examination results, incomplete placental collection or substandard quality are first excluded. Each included data is double-checked by experts. If the results are inconsistent, it will not be included in the final analysis. Figure 2 As shown, Figure 2 The construction process of a multi-source heterogeneous database of early pregnancy GDM placenta is demonstrated.

[0070] As a preferred embodiment, each placenta two-dimensional ultrasound image data is standardized and annotated to construct an image segmentation dataset, including:

[0071] Converting each acquired two-dimensional ultrasound image data of the placenta into a single image in a unified format;

[0072] Perform denoising and image enhancement operations on the images in unified format;

[0073] Each image is labeled to indicate whether the image comes from a GDM patient or a normal pregnant woman, and an image segmentation dataset is obtained.

[0074] In actual operation, the collected imaging data were unified into a single image in JPG format without blood flow and annotation information, and the image was preprocessed (including denoising, normalization, and image enhancement steps) to ensure the quality and consistency of the input data.

[0075] During annotation, we developed a protocol for annotating early pregnancy GDM placental ultrasound images, referencing the ISUOG Guidelines for Obstetric Ultrasound Examination and related placental imaging standards. This protocol ensures standardization and consistency of ultrasound sections, and establishes annotation rules for placental microvascular pathology. Each image in the dataset was annotated to indicate whether it was from a GDM patient or a healthy pregnant woman.

[0076] As a preferred embodiment, an image segmentation model based on a CNN network is constructed, including:

[0077] The image segmentation model includes a downsampling unit and an upsampling unit, and the downsampling unit and the upsampling unit are symmetrical structures;

[0078] The downsampling unit is a dual-branch encoder, comprising a first input layer, a second input layer, a low-layer shared encoder, and a high-layer independent encoder;

[0079] The first input layer is used to receive multi-section placental ultrasound images (such as sagittal and coronal planes) to capture the placental morphology from different perspectives. The second input layer is used to receive two-dimensional placental ultrasound images to obtain global texture information. Both the low-level shared encoder and the high-level independent encoder include multiple convolution blocks, each of which includes multiple consecutive convolution layers, and the terminal convolution block is connected to a maximum pooling layer. Each convolution layer is connected to a batch normalization layer and a Relu activation layer. The low-level shared weight parameters (such as basic features such as edges and textures) are shared, while the high-level independent encoders remain independent, extracting hierarchical features separately and encoding the information within the receptive field. The high-level independent encoders extract specific features (such as hemodynamic parameters and calcification areas) for different modalities to enhance the model's ability to represent complex placental structures.

[0080] The upsampling unit includes multiple decoders, each decoder includes multiple depooling layers, each depooling layer is connected to multiple convolutional layers, and the last decoder is connected to an output layer, which uses Softmax to generate pixel-level image segmentation results.

[0081] The image segmentation model adopted in this embodiment can extract information from different levels respectively through the shared encoder and independent encoder structure in the downsampling unit, thereby enhancing the multi-scale learning ability of the model. During the downsampling process, the neural network can extract hierarchical features for encoding the information within the receptive field. During the upsampling process, the neural network can gradually restore the resolution of the prediction results. By combining multi-section placenta images with two-dimensional placenta images, it is ensured that the system can fully understand the morphological and texture characteristics of the placenta. In the upsampling unit, the combination of depooling and convolution operations can generate accurate pixel-level segmentation results, thereby improving the accuracy of image segmentation. By combining the downsampling and upsampling processes, accurate segmentation of multiple targets can be achieved simultaneously, making full use of the spatial information in the ultrasound image.

[0082] As a preferred embodiment, the dual-branch encoder includes a multi-core inverted residual module, which is used to extract multi-scale placental texture features in parallel through convolution kernels of different sizes, capturing local details and global morphology respectively. Specifically, the branches within the module use convolution kernels of different sizes (such as 3×3, 5×5, 7×7) to capture local details (such as chorionic plate microstructure) and global morphology (such as the overall calcification area of the placenta). In some embodiments, the outputs of the two branches can also be combined by channel splicing (Concatenation) or weighted addition (such as SENet's channel attention) to generate a multi-scale joint representation. By adopting the "expansion → convolution → compression" process, the number of channels is first expanded by 1×1 convolution, and then the amount of calculation is reduced by depthwise separable convolution (Depthwise Conv), and finally the channel is compressed to restore the dimension, which can alleviate the gradient vanishing problem, improve the training stability of the deep network, and ensure information integrity.

[0083] Assuming the input feature map is X, the output Y of the multi-core inverted residual module can be expressed as:

[0084] Y=X+∑ k∈K (DW k×k (PW ext (X)))

[0085] Among them, K = {3, 5, 7} represents the set of convolution kernel sizes; PW ext (X) is the 1×1 point convolution expansion channel; DW k×k It is a separable convolution with a depth of k×k.

[0086] Since the segmentation of soft boundaries of placental ultrasound (such as the junction between blood vessels and placental substance) is easily affected by unpooling errors, in order to improve the problem that relying solely on the unpooling layer to restore spatial information may lose details, in some embodiments, hollow convolution or jump connection is also introduced in the decoder to improve edge accuracy in combination with shallow features.

[0087] As a preferred embodiment, obtaining an image feature dataset based on the processed image includes:

[0088] Perform statistical analysis on the segmentation features of the image to determine the imaging quantitative features and radiomics features, and obtain the imaging feature dataset based on the imaging quantitative features and radiomics features;

[0089] The imaging quantitative features include at least placental thickness, gray-level co-occurrence matrix, local binary pattern, placental hemodynamic parameters, placental vascular distribution and density, and placental echo characteristics;

[0090] The imaging genomics features include at least a pixel grayscale histogram, a texture feature map, a shape feature map, and a wavelet characteristic map.

[0091] The meanings of imaging quantitative features and imaging omics features are explained below.

[0092] Placental thickness: The thickness of the placenta is directly measured through ultrasound images; Gray-level co-occurrence matrix (GLCM): Texture features are extracted by analyzing the grayscale value distribution and spatial relationship of pixels in ultrasound images; Local binary pattern (LBP): Local texture features are extracted by comparing the grayscale value of each pixel with its surrounding pixels; Placental hemodynamic parameters: Parameters such as placental blood flow velocity and resistance index are measured through Doppler ultrasound; Placental vascular distribution and density: The degree of placental vascularization is evaluated by analyzing the vascular distribution and density in ultrasound images; Placental echo characteristics: The tissue characteristics of the placenta are evaluated by analyzing the echo intensity and pattern of the placenta in ultrasound images.

[0093] Pixel grayscale histogram: extracts statistical features by analyzing the distribution of pixel grayscale values in ultrasound images; texture feature map: extracts texture features of images through methods such as GLCM and LBP; shape feature map: analyzes the shape features of the placenta, such as area, circumference, roundness, etc.; wavelet characteristic map: extracts multi-scale features of images through wavelet transform.

[0094] As a preferred embodiment, the GDM risk prediction model is constructed based on a fully connected network with a cross-attention fusion mechanism, including a feature input layer, a feature fusion layer and an output layer;

[0095] The feature input layer is used to input image feature data and clinical comprehensive data corresponding to the image feature data, and map the clinical comprehensive data into a continuous vector;

[0096] The feature fusion layer is used to calculate the cross-modal correlation weight using the image feature vector as the query and the clinical comprehensive data vector as the key, and to concatenate the image features with the attention-weighted clinical comprehensive data to form a fusion feature matrix;

[0097] The output layer uses the Sigmoid function to output the GDM risk analysis results.

[0098] As a preferred embodiment, the feature fusion layer is used to use the image feature vector as the query and the clinical comprehensive feature vector as the key to calculate the cross-modal correlation weight, and to splice the image features with the attention-weighted clinical comprehensive data vector to form a fusion feature matrix, including:

[0099] The clinical comprehensive feature vector is dynamically weighted based on the attention weight to generate a contextual representation aligned with the image features, which can be expressed as:

[0100]

[0101] Among them, Q is the image feature, K is the clinical comprehensive feature, V is the weighted sum feature, d k is the scaling factor.

[0102] As a specific example, when training a model, an independent dataset or cross-validation can be used to comprehensively evaluate model performance. Evaluation metrics may include accuracy, sensitivity, and specificity. Based on the validation results, the model is adjusted and optimized to improve its performance. This may include adjusting hyperparameters, increasing the dataset size, or employing other techniques.

[0103] The embodiment of the present invention further provides a method for auxiliary analysis of gestational diabetes based on placental ultrasound texture, comprising:

[0104] Step S301: Acquire historical placental two-dimensional ultrasound image data and corresponding comprehensive clinical data of different pregnant women;

[0105] Step S302: screening the comprehensive clinical data, establishing a correlation analysis data set corresponding to the comprehensive clinical data and the image data, performing standardization processing and image annotation on each placenta two-dimensional ultrasound image data, and constructing an image segmentation data set;

[0106] Step S303: constructing an image segmentation model based on a CNN network, and training the image segmentation model using an image segmentation dataset to obtain a fully trained image segmentation model, performing image processing on an input two-dimensional ultrasound image of the placenta using the trained image segmentation model, and obtaining an image feature dataset based on the processed image;

[0107] Step S304: constructing a GDM risk prediction model, using the image feature dataset and the association analysis dataset to establish a risk prediction dataset, and training the GDM risk prediction model using the risk prediction dataset to obtain a fully trained GDM risk prediction model;

[0108] Step S305: Obtain clinical comprehensive data corresponding to the two-dimensional ultrasound image dataset of the placenta of the current pregnant woman, and input it into the well-trained image segmentation model and GDM risk prediction model to classify the GDM risk of the current pregnant woman.

[0109] like Figure 4 As shown, Figure 4 This method demonstrates a schematic diagram of its multimodal data processing workflow. By incorporating deep learning technology, it automates and intelligently analyzes placental texture features in ultrasound images, improving the accuracy and consistency of imaging diagnoses. By building a large-scale ultrasound image database and training models, it enables accurate early prediction of GDM, demonstrating significant clinical application prospects.

[0110] The placental ultrasound texture-based gestational diabetes auxiliary analysis system and method provided by the present invention provide multimodal data support for the early analysis of GDM, which can improve the comprehensiveness and accuracy of the analysis; through automated image segmentation and feature extraction, manual intervention is reduced and efficiency is improved; comprehensive consideration of image features and clinical data is used to make accurate risk predictions; and deep learning technology is used to enhance the intelligence level of the system, which can effectively improve the risk prediction accuracy of gestational diabetes and provide doctors with reliable auxiliary decision support.

[0111] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A gestational diabetes mellitus auxiliary analysis system based on placental ultrasound texture, characterized in that: include: A multimodal database for storing historical placental two-dimensional ultrasound imaging data and corresponding comprehensive clinical data of different pregnant women; The preprocessing module is used to screen the comprehensive clinical data and establish a correlation analysis data set corresponding to the comprehensive clinical data and the image data; it is also used to standardize and annotate each placenta two-dimensional ultrasound image data to construct an image segmentation data set; An image feature extraction module is used to construct an image segmentation model based on a CNN network and train the image segmentation model using an image segmentation dataset to obtain a fully trained image segmentation model. The trained image segmentation model is used to process the input two-dimensional ultrasound image of the placenta and obtain an image feature dataset based on the processed image. The risk prediction module is used to build a GDM risk prediction model. The risk prediction dataset is established using the image feature dataset and the association analysis dataset. The GDM risk prediction model is trained using the risk prediction dataset to obtain a fully trained GDM risk prediction model. The judgment module is used to obtain the clinical comprehensive data corresponding to the two-dimensional ultrasound image dataset of the placenta of the current pregnant woman, and input it into the well-trained image segmentation model and GDM risk prediction model to classify the GDM risk of the current pregnant woman.

2. The gestational diabetes auxiliary analysis system based on placental ultrasound texture according to claim 1, characterized in that: Construct an image segmentation model based on the CNN network, including: The image segmentation model includes a downsampling unit and an upsampling unit, and the downsampling unit and the upsampling unit are symmetrical structures; The downsampling unit is a dual-branch encoder, comprising a first input layer, a second input layer, a low-layer shared encoder, and a high-layer independent encoder; The first input layer is used to receive multi-section ultrasound images of the placenta to capture the placenta morphology from different perspectives, and the second input layer is used to receive two-dimensional ultrasound images of the placenta to obtain global texture information. Both the low-level shared encoder and the high-level independent encoder include multiple convolutional blocks, each of which consists of multiple consecutive convolutional layers, and the last convolutional block is connected to a maximum pooling layer. Each convolutional layer is connected to a batch normalization layer and a Relu activation layer. The low-level shared encoder shares the same set of weight parameters, while the high-level independent encoders remain independent, extracting hierarchical features and encoding the information within the receptive field. The upsampling unit includes multiple decoders, each decoder includes multiple depooling layers, each depooling layer is connected to multiple convolutional layers, and the last decoder is connected to an output layer, which uses Softmax to generate pixel-level image segmentation results.

3. The gestational diabetes auxiliary analysis system based on placental ultrasound texture according to claim 2, characterized in that: The dual-branch encoder includes a multi-kernel inverted residual module, which is used to extract multi-scale placental texture features in parallel through convolution kernels of different sizes, capturing local details and global morphology respectively.

4. The gestational diabetes auxiliary analysis system based on placental ultrasound texture according to claim 1, characterized in that: The image feature dataset is obtained based on the processed image, including: Perform statistical analysis on the segmentation features of the image to determine the imaging quantitative features and radiomics features, and obtain the imaging feature dataset based on the imaging quantitative features and radiomics features; The imaging quantitative features include at least placental thickness, gray-level co-occurrence matrix, local binary pattern, placental hemodynamic parameters, placental vascular distribution and density, and placental echo characteristics; The imaging genomics features include at least a pixel grayscale histogram, a texture feature map, a shape feature map, and a wavelet characteristic map.

5. The gestational diabetes auxiliary analysis system based on placental ultrasound texture according to claim 1, characterized in that: The GDM risk prediction model is constructed based on a fully connected network with a cross-attention fusion mechanism, including a feature input layer, a feature fusion layer, and an output layer; The feature input layer is used to input image feature data and clinical comprehensive data corresponding to the image feature data, and map the clinical comprehensive data into a continuous vector; The feature fusion layer is used to calculate the cross-modal correlation weight using the image feature vector as the query and the clinical comprehensive data vector as the key, and to concatenate the image features with the attention-weighted clinical comprehensive data to form a fusion feature matrix; The output layer uses the Sigmoid function to output the GDM risk analysis results.

6. The gestational diabetes auxiliary analysis system based on placental ultrasound texture according to claim 5, characterized in that: The feature fusion layer is used to use the image feature vector as the query and the clinical comprehensive feature vector as the key to calculate the cross-modal correlation weight, and to splice the image features with the attention-weighted clinical comprehensive data vector to form a fusion feature matrix, including: The clinical comprehensive feature vector is dynamically weighted based on the attention weight to generate a contextual representation aligned with the image features, which can be expressed as: Among them, Q is the image feature, K is the clinical comprehensive feature, V is the weighted sum feature, d k is the scaling factor.

7. The gestational diabetes auxiliary analysis system based on placental ultrasound texture according to claim 1, characterized in that: The comprehensive clinical data includes at least basic health information, postpartum placental pathological examination, blood biochemical index data and delivery outcome data.

8. The gestational diabetes auxiliary analysis system based on placental ultrasound texture according to claim 1, characterized in that: Each placenta 2D ultrasound image data is standardized and annotated to construct an image segmentation dataset, including: Converting each acquired two-dimensional ultrasound image data of the placenta into a single image in a unified format; Perform denoising and image enhancement operations on the images in unified format; Each image is labeled to indicate whether the image comes from a GDM patient or a normal pregnant woman, and an image segmentation dataset is obtained.

9. The gestational diabetes auxiliary analysis system based on placental ultrasound texture according to claim 1, characterized in that: Screen the comprehensive clinical data and establish a correlation analysis data set corresponding to the comprehensive clinical data and imaging image data, including: Clinical comprehensive data with no corresponding placental ultrasound images or incomplete images were eliminated; Align the comprehensive clinical data with the corresponding imaging data according to gestational age; Establish a temporal correlation between placental ultrasound images and clinical comprehensive data of the same period.

10. A method for auxiliary analysis of gestational diabetes based on placental ultrasound texture, characterized in that: include: Obtain historical placental two-dimensional ultrasound image data and corresponding comprehensive clinical data of different pregnant women; Screen the comprehensive clinical data, establish a correlation analysis dataset corresponding to the comprehensive clinical data and image data, standardize and annotate each placenta two-dimensional ultrasound image data, and construct an image segmentation dataset; Construct an image segmentation model based on a CNN network and train it using an image segmentation dataset to obtain a fully trained image segmentation model. Use the trained image segmentation model to process the input two-dimensional ultrasound image of the placenta, and obtain an image feature dataset based on the processed image. Construct a GDM risk prediction model, use the image feature dataset and the association analysis dataset to establish a risk prediction dataset, and train the GDM risk prediction model with the risk prediction dataset to obtain a fully trained GDM risk prediction model; The clinical comprehensive data corresponding to the current pregnant woman's placental two-dimensional ultrasound image dataset is obtained and input into the well-trained image segmentation model and GDM risk prediction model to classify the current pregnant woman's GDM risk.

Images