Perinatal period comprehensive risk assessment system based on artificial intelligence
Through multi-dimensional feature extraction and multi-task joint learning network, the problems of feature instability and heavy computational burden in perinatal risk assessment are solved, achieving more efficient and accurate risk assessment.
Patent Information
- Application Number
- CN202510998379.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-21
- Publication Date
- 2025-10-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the existing technology, the medical image feature extraction method for perinatal risk assessment relies on a fixed color space and a single texture, and lacks shape analysis, resulting in unstable features and low information utilization. In addition, traditional risk assessment tasks are modeled in isolation, and each subtask cannot share the underlying feature information, resulting in heavy computational burden and unbalanced training, which affects the assessment quality.
A comprehensive feature vector is constructed by multi-dimensional feature extraction and weighted fusion of color, texture, and shape features. A multi-task joint learning network is constructed to integrate the subtasks under a unified framework for collaborative training. The weighted loss function is used to adjust the learning focus to avoid overfitting or underfitting.
It improves the ability to identify risk areas and information utilization, improves overall learning efficiency and classification accuracy, and achieves a more stable comprehensive risk assessment.
Smart Images
Figure CN120748735A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical risk assessment, and specifically to an artificial intelligence-based comprehensive perinatal risk assessment system. Background Art
[0002] The perinatal period is a critical stage in which both pregnant women and fetuses face high risks. Events such as premature birth, placental abruption, preeclampsia, and intrauterine infection carry high rates of disability and mortality. Conventional medical image feature extraction methods yield insufficient feature representation, impacting the accuracy of risk assessment. Their reliance on a fixed color space, single texture, and lack of shape analysis leads to low information utilization, unstable features, and insufficient ability to discern risk areas. Traditional risk assessment tasks are often modeled in isolation, with subtasks unable to share underlying feature information and lacking collaborative optimization. This leads to a heavy computational burden on the network and difficulty in globally balancing the training process across different tasks. This can easily lead to overfitting or underfitting of certain tasks, impacting the overall assessment quality of the system. Summary of the Invention
[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a perinatal comprehensive risk assessment system based on artificial intelligence. In view of the fact that the feature expression obtained by the general medical image feature extraction method is insufficient, which affects the accuracy of risk assessment, and relies on a fixed color space, a single texture and a lack of shape analysis, resulting in low information utilization, unstable features and insufficient ability to distinguish risk areas, this solution extracts multi-dimensional features from medical images, constructs a comprehensive feature vector by weighted fusion of color, texture and shape features, enhances the robustness of color features, improves the ability to distinguish risk areas, and improves the structural analysis ability, making the image feature expression more comprehensive; in view of the traditional risk assessment Tasks are often modeled in isolation, and the underlying feature information cannot be shared between subtasks. There is a lack of collaborative optimization, which leads to a heavy computational burden on the network and makes it difficult to globally balance the training process between different tasks. Some tasks are prone to overfitting or underfitting, affecting the overall evaluation quality of the system. This solution constructs a multi-task joint learning network, integrates the subtasks in a unified framework for collaborative training, enhances the utilization of data features, effectively improves the overall learning efficiency and classification accuracy, and dynamically adjusts the network learning focus by weighted summation of the loss values of each subtask, effectively avoiding the problem of over-training of a single task leading to performance degradation of other tasks, and the comprehensive risk assessment performance is more stable.
[0004] The present invention provides an artificial intelligence-based perinatal comprehensive risk assessment system, which includes a multimodal data acquisition module, an image feature extraction module, a structural feature extraction module, a feature fusion module, and a risk assessment module;
[0005] The multimodal data acquisition module acquires heterogeneous data of perinatal individuals, including clinical numerical indicators, medical images, and text data, sends the medical images to the image feature extraction module, stores the clinical numerical indicators and text data as structured data, and sends them to the structure feature extraction module;
[0006] The image feature extraction module pre-processes the medical image to obtain a processed image, extracts color features, texture features and shape features from the processed image, and generates an image feature vector that is sent to the feature fusion module;
[0007] The structural feature extraction module checks the logical consistency of the structured data, marks abnormal values, generates a structural feature vector and sends it to the feature fusion module;
[0008] The feature fusion module obtains the weight coefficient of each feature vector, performs weighted fusion on the image feature vector and the structural feature vector, generates a comprehensive feature vector, and sends the comprehensive feature vector to the risk assessment module;
[0009] The risk assessment module constructs a multi-task joint learning network, classifies and evaluates perinatal risk events, and outputs a comprehensive perinatal risk assessment report.
[0010] Furthermore, the image feature extraction module includes a color feature extraction unit, a texture feature extraction unit, a shape feature extraction unit and a feature fusion unit;
[0011] The color feature extraction unit extracts color features using the RGB color space and the HSV color space, calculates the mean and standard deviation of each channel from the processed image as color features, and constructs a color feature vector. The formula used is as follows: ;
[0012] Where, represents the color feature vector, The channels include R, G, B channels and H, S, V channels. represents the channel mean, represents the channel standard deviation;
[0013] The texture feature extraction unit extracts texture features from the processed image using GLCM (Gray Level Co-occurrence Matrix) and LBP (Local Binary Pattern) to construct a texture feature vector;
[0014] The shape feature extraction unit uses the Canny edge detection algorithm to extract edge contours from the processed image, calculates the Hu shape invariant moments of the edge contours through morphological analysis to describe the geometric shape, and constructs a shape feature vector. The formula used is as follows: ;
[0015] Where, represents the shape feature vector, Represents 7 Hu shape invariant moments;
[0016] The feature fusion unit combines the color feature vector, texture feature vector and shape feature vector into an image feature vector using the following formula: ;
[0017] Where, represents the image feature vector, represents the texture feature vector, Represents the weight coefficient.
[0018] Furthermore, the risk assessment module constructs a multi-task joint learning network to perform comprehensive risk assessment, including the following steps:
[0019] Step S1: Initialize the multi-task joint learning network and decompose the perinatal risk assessment task into multiple subtasks;
[0020] Step S2: Construct a shared feature layer to extract shared features between subtasks, including the following steps:
[0021] Step S21: constructing a convolutional feature extractor to extract local spatial features from the comprehensive feature vector based on a convolutional neural network;
[0022] Step S22: construct an attention mechanism module to enhance the key area response in the comprehensive feature vector through the self-attention mechanism;
[0023] Step S23: Construct a feature aggregation layer to fuse the outputs of the convolutional feature extractor and the attention mechanism module to generate shared features;
[0024] Step S24: construct a regularization module to prevent overfitting of network parameters through Dropout regularization;
[0025] Step S3: Design task-specific layers. Design independent task-specific layers for each perinatal risk subtask to capture the differences between subtasks.
[0026] Step S4: Construct a joint loss function to balance the loss values of each subtask by weighted summation. The formula used is as follows: ;
[0027] Where, represents the total loss value, Indicates the index of the subtask, Indicates the total number of subtasks, Indicates the The loss value of each subtask, Indicates the The weight coefficient of each subtask;
[0028] Step S5: Evaluation report generation, using the back propagation algorithm to optimize the network parameters until the loss value converges, and output the final perinatal risk assessment report.
[0029] The beneficial effects achieved by the present invention using the above scheme are as follows:
[0030] (1) In view of the problems that the feature expression obtained by general medical image feature extraction methods is insufficient, which affects the accuracy of risk assessment, and relies on a fixed color space, a single texture and lacks shape analysis, resulting in low information utilization, unstable features and insufficient ability to distinguish risk areas, this scheme performs multi-dimensional feature extraction on medical images, constructs a comprehensive feature vector by weighted fusion of color, texture and shape features, enhances the robustness of color features, improves the ability to distinguish risk areas, and improves the structural analysis ability, making the image feature expression more comprehensive.
[0031] (2) Traditional risk assessment tasks are often modeled in isolation. The underlying feature information cannot be shared between subtasks, and there is a lack of collaborative optimization, which leads to a heavy computational burden on the network and makes it difficult to globally balance the training process between different tasks. Some tasks are prone to overfitting or underfitting, which affects the overall assessment quality of the system. This solution constructs a multi-task joint learning network, integrates the subtasks in a unified framework for collaborative training, enhances the utilization of data features, effectively improves the overall learning efficiency and classification accuracy, and dynamically adjusts the network learning focus by weighted summation of the loss values of each subtask, effectively avoiding the problem of overtraining of a single task leading to a decline in the performance of other tasks. The comprehensive risk assessment performance is more stable. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 This is a schematic diagram of an artificial intelligence-based comprehensive perinatal risk assessment system proposed by the present invention;
[0033] Figure 2 Schematic diagram of the image feature extraction module;
[0034] Figure 3 Schematic diagram of a texture feature extraction unit;
[0035] Figure 4 Schematic diagram of the process of conducting comprehensive risk assessment for the risk assessment module.
[0036] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION
[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0038] Example 1, see Figure 1 , the present invention provides an artificial intelligence-based perinatal comprehensive risk assessment system, including a multimodal data acquisition module, an image feature extraction module, a structural feature extraction module, a feature fusion module, and a risk assessment module;
[0039] The multimodal data acquisition module acquires heterogeneous data of perinatal individuals, including clinical numerical indicators, medical images, and text data, sends the medical images to the image feature extraction module, stores the clinical numerical indicators and text data as structured data, and sends them to the structure feature extraction module;
[0040] The image feature extraction module preprocesses the medical image by using a non-local mean filtering algorithm to remove noise from the medical image, adjusting the differences between different staining methods of the medical image through normalization, and performing region segmentation using a threshold segmentation algorithm to obtain a processed image. The color features, texture features, and shape features of the processed image are extracted respectively, and an image feature vector is generated and sent to the feature fusion module;
[0041] The structural feature extraction module checks the logical consistency of the structured data and marks abnormal values; the clinical numerical indicators are normalized to the interval [0, 1], the text data is converted into word vector representation, and the relevant risk indicators are annotated with labels to generate structural feature vectors and send them to the feature fusion module;
[0042] The feature fusion module obtains the weight coefficient of each feature vector through expert scoring, performs weighted fusion on the image feature vector and the structural feature vector, generates a comprehensive feature vector, and sends the comprehensive feature vector to the risk assessment module;
[0043] The risk assessment module constructs a multi-task joint learning network, classifies and evaluates perinatal risk events, and outputs a comprehensive perinatal risk assessment report.
[0044] Example 2, see Figure 1This embodiment is based on the above embodiment. In the multimodal data acquisition module, clinical numerical indicators include blood pressure, blood sugar, BMI, fetal heart rate, and NT thickness; medical images include placenta and umbilical cord slice images, cervical smears or vaginal secretion micrographs, and post-processed ultrasound images; text data includes pregnancy examination records, doctor's evaluation opinions, and individual medical history and family history information.
[0045] Example 3, see Figure 1 This embodiment is based on the above embodiment. The image feature extraction module preprocesses the medical image, including denoising, normalization, and region segmentation. Specifically, the non-local mean filtering algorithm is used to calculate the similarity between pixels in the medical image to remove noise while retaining edge and texture information; the brightness and contrast of the medical image are adjusted to a uniform range to obtain a normalized image. The formula used is as follows: ;
[0046] Where, represents the normalized image, represents the pixel value of the medical image, and Represent the minimum and maximum pixel values in medical images respectively;
[0047] The normalized image is divided into foreground and background regions using the threshold segmentation algorithm to obtain the processed image.
[0048] Example 4, see Figure 1 and Figure 2 , this embodiment is based on the above embodiment, the image feature extraction module includes a color feature extraction unit, a texture feature extraction unit, a shape feature extraction unit and a feature fusion unit;
[0049] The color feature extraction unit extracts color features using the RGB color space and the HSV color space, calculates the mean and standard deviation of each channel from the processed image as color features, and constructs a color feature vector. The formula used is as follows: ; ; ;
[0050] Where, represents the color feature vector, Represents the index of the pixel point, Represents the total number of pixels in the processed image, Indicates The pixel value of the channel, The channels include R, G, B channels and H, S, V channels. represents the channel mean, represents the channel standard deviation;
[0051] The texture feature extraction unit extracts texture features from the processed image using GLCM and LBP. GLCM describes texture characteristics by statistically analyzing the grayscale value distribution of adjacent pixel pairs in the processed image. LBP compares the grayscale values of a central pixel with those of its neighboring pixels to generate a binary pattern and construct a texture feature vector.
[0052] The shape feature extraction unit uses the Canny edge detection algorithm to extract edge contours from the processed image, calculates the Hu shape invariant moments of the edge contours through morphological analysis to describe the geometric shape, and constructs a shape feature vector. The formula used is as follows: ; ;
[0053] Where, represents the normalized central moment, represents the central moment, represents the area of the shape, represents the order of the moment, represents the shape feature vector, Represents 7 Hu shape invariant moments;
[0054] The feature fusion unit combines the color feature vector, texture feature vector and shape feature vector into an image feature vector using the following formula: ;
[0055] Where, represents the image feature vector, represents the texture feature vector, Represents the weight coefficient.
[0056] By performing the above operations, the present invention solves the problems that the feature expression obtained by general medical image feature extraction methods is insufficient, which affects the accuracy of risk assessment, and relies on a fixed color space, a single texture and a lack of shape analysis, resulting in low information utilization, unstable features and insufficient ability to distinguish risk areas. In this solution, multi-dimensional feature extraction is performed on medical images, and a comprehensive feature vector is constructed by weighted fusion of color, texture and shape features. This enhances the robustness of color features, improves the ability to distinguish risk areas, and improves structural analysis capabilities, making the image feature expression more comprehensive.
[0057] Example 5, see Figure 1 、 Figure 2 and Figure 3 This embodiment is based on the above embodiment, and the texture feature extraction unit includes a GLCM calculation subunit, an LBP calculation subunit and a vector construction subunit:
[0058] The GLCM calculation subunit extracts texture characteristics by calculating GLCM in different directions and pixel spacings, and constructs GLCM feature vectors. The directions are selected as 0°, 45°, 90°, and 135°. The texture characteristics include contrast, energy, and entropy. The formula used is as follows: ; ; ; ;
[0059] Where, represents the GLCM eigenvector, 、 、 represent contrast, energy, and entropy respectively, and represents the grayscale value of adjacent pixels, represents the probability of symbiosis;
[0060] The LBP calculation subunit uses circular LBP to extract local texture structure, calculates the LBP value of each pixel and counts the LBP value distribution of all pixels to form an LBP histogram, and normalizes the LBP histogram to obtain the LBP feature vector. The formula used is as follows: ; ;
[0061] Where, represents a local binary pattern, represents the LBP feature vector, represents the binarization function, represents the center pixel, represents the neighborhood pixels, represents the number of neighborhood pixels, Represents the distance between neighboring pixels, Indicates the frequency of the LBP histogram;
[0062] The vector construction subunit combines the GLCM feature vector and the LBP feature vector into a texture feature vector as shown below: ;
[0063] Where, Represents the texture feature vector.
[0064] Example 6, see Figure 1 and Figure 4This embodiment is based on the above embodiment. The risk assessment module constructs a multi-task joint learning network to perform comprehensive risk assessment, including the following steps:
[0065] Step S1: Initialize a multi-task joint learning network and decompose the perinatal risk assessment task into multiple subtasks. The subtasks include: locating high-risk areas in medical images, including the placenta and cervix; classifying perinatal complications, including premature birth, placental abruption, and preeclampsia; and predicting pregnancy trends, including changes in uterine contraction frequency, weight gain curve, and fetal heart rate fluctuations.
[0066] Step S2: Construct a shared feature layer to extract shared features between subtasks, including the following steps:
[0067] Step S21: constructing a convolutional feature extractor to extract local spatial features from the comprehensive feature vector based on a convolutional neural network;
[0068] Step S22: Construct an attention mechanism module to enhance the key region response in the comprehensive feature vector through the self-attention mechanism, including the structural regions of cervical length and placental stratification boundaries related to the risk of premature birth in medical images, and the regions with dramatic fluctuations in clinical indicators such as blood pressure, weight gain, and urine protein in structured data;
[0069] Step S23: Construct a feature aggregation layer to fuse the outputs of the convolutional feature extractor and the attention mechanism module to generate shared features;
[0070] Step S24: construct a regularization module to prevent overfitting of network parameters through Dropout regularization;
[0071] Step S3: Design task-specific layers. For each perinatal risk subtask, design an independent task-specific layer to capture the differences between subtasks: For the high-risk area localization subtask, use the U-Net architecture to generate pixel-level segmentation results for placental slices and cervical ultrasounds; for the complication type classification subtask, use the Softmax function to output the predicted probability of each complication category; for the pregnancy trend prediction subtask, use a recurrent neural network to capture the temporal evolution of physical signs such as uterine contraction frequency changes, weight gain curves, and fetal heart rate fluctuations;
[0072] Step S4: Construct a joint loss function to balance the loss values of each subtask by weighted summation. The formula used is as follows: ;
[0073] Where, represents the total loss value, Indicates the index of the subtask, Indicates the total number of subtasks, Indicates the The loss value of each subtask, Indicates the The weight coefficient of each subtask;
[0074] Step S5: Generate an assessment report. Use the backpropagation algorithm to optimize the network parameters until the loss value converges, and output the final perinatal risk assessment report. The report content includes a visual annotation map of the image risk area, complication classification results, and a pregnancy trend prediction map.
[0075] By performing the above operations, traditional risk assessment tasks are often modeled in isolation, and the underlying feature information cannot be shared between subtasks, resulting in a lack of collaborative optimization, which leads to a heavy computational burden on the network and makes it difficult to globally balance the training process between different tasks. Some tasks are prone to overfitting or underfitting, affecting the overall assessment quality of the system. This solution constructs a multi-task joint learning network, integrates the subtasks in a unified framework for collaborative training, enhances the utilization of data features, effectively improves the overall learning efficiency and classification accuracy, and dynamically adjusts the network learning focus by weighted summation of the loss values of each subtask, effectively avoiding the problem of over-training of a single task leading to performance degradation of other tasks, and the comprehensive risk assessment performance is more stable.
[0076] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.
[0077] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.
[0078] The present invention and its embodiments are described above. Such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by the above, and does not deviate from the purpose of the present invention, without inventive design, a structure and embodiment similar to the technical solution should fall within the scope of protection of the present invention.
Claims
1. An artificial intelligence-based comprehensive perinatal risk assessment system, characterized by: It includes multimodal data acquisition module, image feature extraction module, structural feature extraction module, feature fusion module and risk assessment module; The multimodal data acquisition module acquires heterogeneous data of perinatal individuals, including clinical numerical indicators, medical images, and text data, sends the medical images to the image feature extraction module, stores the clinical numerical indicators and text data as structured data, and sends them to the structure feature extraction module; The image feature extraction module pre-processes the medical image to obtain a processed image, extracts color features, texture features and shape features from the processed image, and generates an image feature vector that is sent to the feature fusion module; The structural feature extraction module checks the logical consistency of the structured data, marks abnormal values, generates a structural feature vector and sends it to the feature fusion module; The feature fusion module obtains the weight coefficient of each feature vector, performs weighted fusion on the image feature vector and the structural feature vector, generates a comprehensive feature vector, and sends the comprehensive feature vector to the risk assessment module; The risk assessment module constructs a multi-task joint learning network, classifies and evaluates perinatal risk events, and outputs a comprehensive perinatal risk assessment report.
2. The artificial intelligence-based perinatal comprehensive risk assessment system according to claim 1, characterized in that: The image feature extraction module includes a color feature extraction unit, a texture feature extraction unit, a shape feature extraction unit and a feature fusion unit; The color feature extraction unit extracts color features using the RGB color space and the HSV color space, calculates the mean and standard deviation of each channel from the processed image as color features, and constructs a color feature vector; The texture feature extraction unit extracts texture features from the processed image using GLCM and LBP to construct a texture feature vector; The shape feature extraction unit uses the Canny edge detection algorithm to extract edge contours from the processed image, calculates the Hu shape invariant moments of the edge contours through morphological analysis to describe the geometric shape, and constructs a shape feature vector; The feature fusion unit combines the color feature vector, texture feature vector and shape feature vector into an image feature vector using the following formula: ； Where, represents the image feature vector, represents the color feature vector, Represents texture feature vector represents the shape feature vector, Represents the weight coefficient.
3. The artificial intelligence-based perinatal comprehensive risk assessment system according to claim 1, characterized in that: The risk assessment module constructs a multi-task joint learning network to perform comprehensive risk assessment, including the following steps: Step S1: Initialize the multi-task joint learning network and decompose the perinatal risk assessment task into multiple subtasks; Step S2: Construct a shared feature layer to extract shared features among subtasks; Step S3: Design task-specific layers. Design independent task-specific layers for each perinatal risk subtask to capture the differences between subtasks. Step S4: Construct a joint loss function and balance the loss values of each subtask through weighted summation; Step S5: Evaluation report generation, using the back propagation algorithm to optimize the network parameters until the loss value converges, and output the final perinatal risk assessment report.
4. The artificial intelligence-based perinatal comprehensive risk assessment system according to claim 3, characterized in that: Step S2, constructing a shared feature layer, includes the following steps: Step S21: constructing a convolutional feature extractor to extract local spatial features from the comprehensive feature vector based on a convolutional neural network; Step S22: construct an attention mechanism module to enhance the key area response in the comprehensive feature vector through the self-attention mechanism; Step S23: Construct a feature aggregation layer to fuse the outputs of the convolutional feature extractor and the attention mechanism module to generate shared features; Step S24: Construct a regularization module and use Dropout regularization to prevent overfitting of network parameters.