Ultrasound image-based placental perfusion and fetal development prediction method and system

Abstract

The application provides a placenta perfusion and fetal development prediction method and system based on ultrasound images, and relates to the technical field of medical image analysis and diagnosis, which comprises the following steps: obtaining a placenta ultrasound image and pre-processing and quality control; segmenting the placenta region through a deep neural network, extracting multi-dimensional features of gray scale, texture, shape and blood flow; inputting a machine learning model to predict the perfusion and fetal development state; visualizing the results and key factors, combining clinical verification and generating intervention suggestions. The system comprises image processing, segmentation, prediction, visualization and verification interpretation modules. The application realizes objective quantitative evaluation, solves AI problems, forms full-chain clinical support, improves early warning accuracy, helps precise intervention and reduces perinatal risks.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis and diagnostic technology, and in particular to a method and system for predicting placental perfusion and fetal development based on ultrasound images. Background Technology

[0002] Fetus growth restriction (FGR) is a common obstetric complication, referring to a fetus's failure to reach its genetic growth potential, typically defined as an estimated fetal weight below the 10th percentile for gestational age. The global incidence of FGR is approximately 5%-10% of all pregnancies, and it is a significant factor contributing to increased perinatal mortality and morbidity. FGR not only increases the risk of perinatal complications such as neonatal asphyxia and death but may also have adverse effects on long-term neurodevelopment and cardiovascular health.

[0003] The main pathophysiological basis of fetal growth restriction (FGR) is placental insufficiency. In a normal pregnancy, the placenta provides the fetus with sufficient oxygen and nutrients through the maternal-placental-fetal circulatory system, ensuring normal fetal growth and development. When placental perfusion is inadequate, it can lead to placental dysfunction, thereby affecting fetal development. Therefore, accurate assessment of placental perfusion status during pregnancy is crucial for the early prediction, diagnosis, and intervention of FGR.

[0004] Based on the above background, this invention proposes a method and system for predicting placental perfusion and fetal development based on ultrasound imaging. Summary of the Invention

[0005] This invention provides a method and system for predicting placental perfusion and fetal development based on ultrasound imaging, in order to solve the aforementioned technical problems.

[0006] This invention provides a method for predicting placental perfusion and fetal development based on ultrasound imaging, comprising: Step 1: Obtain ultrasound images of the placenta of the pregnant woman, and perform preprocessing and quality control processing on the placental ultrasound images; Step 2: Input the processed ultrasound image into a deep neural network segmentation model to obtain the segmented placental region. Perform quantitative feature analysis on the segmented placental region to extract multidimensional feature vectors, including grayscale features, texture features, morphological features, and blood flow features. Before performing quantitative feature analysis on the segmented placental region, perform boundary correction on the segmented placental region. The boundary correction includes multi-boundary line alignment, deformation sequence construction, dynamic expansion region determination, reference trajectory screening, and boundary adjustment. Step 3: Input the multidimensional feature vectors into a pre-trained machine learning model to predict whether the placental perfusion status is normal or poor, and whether the fetal development status is normal or developmentally restricted. Step 4: Perform visualization analysis on the prediction results to present the predicted results and key influencing factors of placental perfusion status and fetal development status. Step 5: Verify the accuracy of the prediction results in obstetric clinical practice and provide corresponding medical interpretations.

[0007] Preferably, acquiring ultrasound images of the placenta of a pregnant woman includes: acquiring ultrasound images of the placenta at different gestational weeks of pregnancy, including grayscale ultrasound images and color Doppler ultrasound images, wherein the acquired ultrasound images contain complete area and boundary information of the placenta.

[0008] Preferably, the preprocessing and quality control processing includes: resizing, standardizing brightness, enhancing contrast, and denoising the ultrasound images of the placenta using Gaussian filtering; and selecting images that meet quality standards based on quality scores for image sharpness, contrast, and signal-to-noise ratio.

[0009] Preferably, the deep neural network segmentation model is one or a combination of several of U-Net, Mask R-CNN, or DeepLab series networks, and the segmentation model is trained using a labeled placental ultrasound image dataset.

[0010] Preferably, before performing quantitative feature analysis on the segmented placental region, the following steps are also included: The first colored boundary line of the processed ultrasound image is obtained. At the same time, the processed ultrasound image is subjected to grayscale processing to obtain the first gray boundary line and the second gray boundary line of the segmented placental region is obtained. The first gray boundary line and the first colored boundary line are aligned in position, and the overlapping pixel group of each boundary point in the second gray boundary line is obtained in sequence. The overlapping pixel group is the combination of the overlapping state of the corresponding boundary point of the second gray boundary line with the first gray boundary line and the first colored boundary line. Simultaneously, extract the eight neighboring points of each boundary point in the second gray boundary line; Obtain the set of eight neighboring pixels for each current boundary point on the second gray boundary line, and select a set of sub-points from the set of eight neighboring pixels that are consistent with the local extension direction of the second gray boundary line; The representational deformation of each sub-point in the sub-point set relative to the current boundary point is determined, and an ordered deformation sequence is constructed for each current boundary point based on the order of the sub-point set on the trajectory of the second gray boundary line. The representational deformation is determined based on the changes in gray-level gradient, texture continuity, and distance between the sub-point and the first colored boundary line and the first gray boundary line. Along the second gray boundary line, the relative rate of change between the deformation sequence corresponding to each current boundary point and the adjacent upstream boundary point is calculated sequentially to form a set of relative rates of change for each boundary point, wherein the relative rate of change reflects the evolution rate of the local deformation mode along the boundary trajectory; Extract the maximum value from the set of relative rates of change of all boundary points, and construct the first rate of change curve in the order of the boundary points. At the same time, extract the minimum value from the set of relative rates of change of all boundary points, and construct the second rate of change curve in the same order. Based on the first rate of change curve and the second rate of change curve, as well as the position coordinates of the sub-points corresponding to each extreme point on the curve in the image, a dynamically expanding boundary region around the second gray boundary line is jointly determined. The width and shape of the dynamically expanding boundary region are adaptively adjusted by the spatial distribution characteristics implied by the amplitude difference and coordinates of the first rate of change curve and the second rate of change curve at the corresponding positions. For each preset control unit within the dynamically expanded boundary region, the change concentration rate of the corresponding boundary point in the relative change rate set is calculated, and the unit size of the control unit is adjusted in combination with the overlapping pixel group corresponding to the corresponding boundary point. The change concentration rate is used to quantify the internal dispersion of the relative change rate set of the corresponding boundary point. Based on all the adjusted reference units, one or more optimized boundary trajectories are generated as several reference trajectories for correcting the second gray boundary line; The second gray boundary line is sequentially combined with each reference trajectory and output to the trajectory comparison model to output the reference value of the corresponding reference trajectory. The reference value of all reference trajectories is statistically analyzed, and useful trajectories are selected by combining the concentrated distribution distance between each reference trajectory and the second gray boundary line. The second gray boundary line is adjusted based on the reference value of the useful trajectory to obtain the effective placental region, and quantitative feature analysis is performed on the effective placental region.

[0011] Preferably, the grayscale features include: average grayscale value, standard deviation, skewness, and kurtosis; the texture features include: grayscale co-occurrence matrix features, local binary pattern features, and Gabor filter features; the morphological features include: placental area, circumference, thickness, and uniformity index; and the blood flow features include: placental blood flow velocity, resistance index, pulsatility index, and blood flow distribution uniformity.

[0012] Preferably, the visualization analysis includes: displaying key areas in the placental region related to perfusion status in the form of a heatmap, generating a feature importance ranking chart, providing confidence scores for prediction results, and generating clinical recommendations for different prediction results.

[0013] Preferably, it also includes: generating interventional treatment recommendations based on the predicted poor placental perfusion and / or fetal developmental restriction, including drug treatment plans, nutritional adjustment recommendations, pregnancy management plans, and delivery timing recommendations.

[0014] This invention provides a system for predicting placental perfusion and fetal development based on ultrasound imaging, comprising: an image processing module for acquiring ultrasound images of the placenta of pregnant women, and performing preprocessing and quality control processing on the placental ultrasound images; The segmentation module inputs the processed ultrasound image into a deep neural network segmentation model to obtain segmented placental regions. It then performs quantitative feature analysis on these regions, extracting multidimensional feature vectors including grayscale, texture, morphology, and blood flow features. Before quantitative feature analysis, boundary correction is performed on the segmented placental regions, including multi-boundary line alignment, deformation sequence construction, dynamic expansion region determination, reference trajectory selection, and boundary adjustment. The prediction module inputs the multidimensional feature vectors into a pre-trained machine learning model to predict whether placental perfusion is normal or poor, and whether fetal development is normal or restricted. The visualization module visualizes the prediction results, presenting the predicted placental perfusion status and fetal development status, along with key influencing factors. The validation and interpretation module is used to validate the accuracy of prediction results in obstetric clinical practice and provide corresponding medical interpretations.

[0015] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements any of the steps of the method described above.

[0016] Compared with the prior art, the beneficial effects of this application are as follows: 1. Based on the actual needs of obstetric clinical practice, and addressing the difficulty in early diagnosis of intrauterine growth restriction in fetuses, this study innovatively establishes a predictive correlation between placental perfusion status and fetal development status. Through deep learning segmentation and feature extraction technology, it achieves accurate analysis of placental ultrasound images, transforming traditional subjective judgment into objective quantitative assessment. This significantly improves the consistency, repeatability, and early detection rate of diagnosis, providing a scientific basis for early clinical intervention and effectively solving the clinical challenges of diagnosing different placentas within the same placenta and diagnosing the same placentas outside the same placenta in traditional ultrasound assessment.

[0017] 2. A multi-dimensional feature comprehensive analysis system was creatively constructed, simultaneously extracting and analyzing the gray-scale statistical features, texture features, morphological features, and hemodynamic features of the placenta, achieving a comprehensive assessment of the placental structure and functional status. In particular, after accurately segmenting the placental region using a deep neural network, multi-scale feature extraction was performed on the segmented region. This not only captured minute changes and complex patterns that are difficult for the human eye to perceive, but also improved the sensitivity and specificity of early placental perfusion abnormalities by using feature optimization and selection algorithms, laying the foundation for accurate prediction of fetal development status.

[0018] 3. A unique comprehensive prediction and decision support system for placental perfusion and fetal development has been developed. This system not only predicts placental perfusion and fetal development status but also automatically generates clinical intervention recommendations and follow-up plans based on the prediction results. The system uses heatmap visualization technology to intuitively display the placental areas that contribute most to the prediction results. Feature importance analysis helps doctors understand the basis for the predictions, solving the "black box" problem of AI. Furthermore, the system has established a tiered early warning mechanism, providing timely warnings for high-risk cases and offering reasonable monitoring recommendations for medium- and low-risk cases. This achieves full-process support from prediction to clinical decision-making, significantly improving the precision of perinatal management.

[0019] 4. It possesses significant practical and promotional value, applicable to images acquired by routine clinical ultrasound equipment. No special equipment or technology is required, and the fully automated processing significantly reduces operational complexity and time costs. The system adopts a modular design and standardized interfaces, allowing for flexible deployment in the cloud or locally, and easy integration with existing hospital information systems. More importantly, the system has continuous learning capabilities, constantly optimizing model performance based on new clinical data. It can also serve as an auxiliary diagnostic tool for primary healthcare institutions, promoting hierarchical medical services, narrowing regional healthcare service gaps, and possessing significant social and economic benefits, providing innovative technological support for improving perinatal healthcare quality.

[0020] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for predicting placental perfusion and fetal development based on ultrasound imaging, as described in an embodiment of the present invention. Figure 2 This is a structural diagram of a system for predicting placental perfusion and fetal development based on ultrasound imaging, as described in an embodiment of the present invention. Detailed Implementation

[0023] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0024] This invention provides a method for predicting placental perfusion and fetal development based on ultrasound imaging, such as... Figure 1 As shown, it includes: Step 1: Obtain ultrasound images of the placenta of the pregnant woman, and perform preprocessing and quality control processing on the placental ultrasound images; Step 2: Input the processed ultrasound image into a deep neural network segmentation model to obtain the segmented placental region. Perform quantitative feature analysis on the segmented placental region to extract multidimensional feature vectors, including grayscale features, texture features, morphological features, and blood flow features. Before performing quantitative feature analysis on the segmented placental region, perform boundary correction on the segmented placental region. The boundary correction includes multi-boundary line alignment, deformation sequence construction, dynamic expansion region determination, reference trajectory screening, and boundary adjustment. Step 3: Input the multidimensional feature vectors into a pre-trained machine learning model to predict whether the placental perfusion status is normal or poor, and whether the fetal development status is normal or developmentally restricted. Step 4: Perform visualization analysis on the prediction results to present the predicted results and key influencing factors of placental perfusion status and fetal development status. Step 5: Verify the accuracy of the prediction results in obstetric clinical practice and provide corresponding medical interpretations.

[0025] Preferably, acquiring ultrasound images of the placenta of a pregnant woman includes: acquiring ultrasound images of the placenta at different gestational weeks of pregnancy, including grayscale ultrasound images and color Doppler ultrasound images, wherein the acquired ultrasound images contain complete area and boundary information of the placenta.

[0026] Preferably, the preprocessing and quality control processing includes: resizing, standardizing brightness, enhancing contrast, and denoising the ultrasound images of the placenta using Gaussian filtering; and selecting images that meet quality standards based on quality scores for image sharpness, contrast, and signal-to-noise ratio.

[0027] Preferably, the deep neural network segmentation model is one or a combination of several of U-Net, Mask R-CNN, or DeepLab series networks, and the segmentation model is trained using a labeled placental ultrasound image dataset.

[0028] Preferably, the grayscale features include: average grayscale value, standard deviation, skewness, and kurtosis; the texture features include: grayscale co-occurrence matrix features, local binary pattern features, and Gabor filter features; the morphological features include: placental area, circumference, thickness, and uniformity index; and the blood flow features include: placental blood flow velocity, resistance index, pulsatility index, and blood flow distribution uniformity.

[0029] Preferably, the visualization analysis includes: displaying key areas in the placental region related to perfusion status in the form of a heatmap, generating a feature importance ranking chart, providing confidence scores for prediction results, and generating clinical recommendations for different prediction results.

[0030] Preferably, it also includes: generating interventional treatment recommendations based on the predicted poor placental perfusion and / or fetal developmental restriction, including drug treatment plans, nutritional adjustment recommendations, pregnancy management plans, and delivery timing recommendations.

[0031] In this embodiment, placental ultrasound examination is performed on the pregnant woman to be examined. Two-dimensional grayscale ultrasound images and color Doppler blood flow images containing the placenta are acquired. The placental ultrasound images are acquired using a color Doppler ultrasound diagnostic instrument equipped with a 3.5-5.0MHz abdominal probe. The examination is performed on pregnant women between 18 and 40 weeks of gestation. 3-5 images are acquired along both the long and short axes of the placenta to ensure that the placental boundaries are clearly visible. The scanning parameters are as follows: the gain is set in the range of 55-65dB, the depth is adjusted to completely cover the placenta, and the focal position is placed in the middle of the placenta; for Doppler images, the pulse repetition frequency is set to 0.6-1.2kHz, the color gain is adjusted to the highest level to eliminate background noise, and the sampling volume covers the entire placental area.

[0032] In this embodiment, the preprocessing and quality control processes specifically include: first, converting the original DICOM format ultrasound image to PNG format; then, calculating the image sharpness score based on indicators such as image gradient and edge sharpness, calculating the signal-to-noise ratio to assess the noise level, calculating the image contrast, and assessing the coverage ratio of the placenta in the image; next, selecting images that meet the quality standards based on the quality assessment results; standardizing the size of the selected images (adjusting them to 512×512 pixels), applying histogram equalization for grayscale standardization, using Gaussian filtering to reduce noise, and applying contrast-limited adaptive histogram equalization (CLAHE) to enhance image details; finally, generating additional training samples through operations such as rotation, flipping, scaling, and translation, and the quality control adopts the following standards: sharpness score greater than 70 points (out of 100), signal-to-noise ratio greater than 5dB, placental area accounting for no less than 40% of the image area, and complete display of the placenta at the image edges without missing parts.

[0033] In this embodiment, the deep neural network segmentation model adopts the U-Net architecture, which consists of an encoder and a decoder. The encoder consists of four downsampling blocks, each containing two 3×3 convolutional layers and one 2×2 max-pooling layer. The decoder consists of four upsampling blocks, each containing one 2×2 deconvolutional layer and two 3×3 convolutional layers. Feature fusion is achieved between the encoder and decoder through skip connections. The network is trained using the Dice loss function and the Adam optimizer. The initial learning rate is set to 0.001, and the learning rate is halved every 30 epochs if the performance on the validation set does not improve. An early stopping strategy is used to avoid overfitting.

[0034] The training dataset for the deep neural network segmentation model was obtained by experienced obstetric ultrasound physicians manually annotating placental regions. The training dataset contains 500 ultrasound images of different gestational weeks, placental locations, and image qualities, divided into training, validation, and test sets in a 7:2:1 ratio. Segmentation performance was evaluated using metrics such as Dice coefficient, Jaccard index, sensitivity, and specificity. To further improve segmentation accuracy, morphological operations (such as opening and closing operations) were applied to optimize segmentation boundaries, remove small isolated regions, and interactive tools were provided to allow experts to fine-tune the segmentation results.

[0035] In this embodiment, the open-source package PyRadiomics (version 3.0.1; https: / / pyradiomics.readthedocs.io) is used in a Python environment to extract radiomics features. Specifically, the following types of features are extracted: Gray-scale statistical characteristics: Calculate the first-order statistical characteristics of the placental region, such as the average gray value, standard deviation, median, maximum value, minimum value, skewness, kurtosis, energy, and entropy.

[0036] Texture features: Extract gray-level co-occurrence matrix (GLCM) features, including energy, contrast, correlation, homogeneity, etc.; extract gray-level running length matrix (GLRLM) features; extract gray-level size region matrix (GLSZM) features; extract adjacent gray-level tone difference matrix (NGTDM) features; extract gray-level dependency matrix (GLDM) features; extract local binary pattern (LBP) features to capture the local texture structure of placental tissue.

[0037] Morphological characteristics: Calculate basic morphological parameters of the placenta such as area, circumference, maximum diameter, and thickness; extract shape descriptors such as roundness, elongation, and rectangularity; analyze the complexity and irregularity of the placental boundary.

[0038] Blood flow characteristics: The proportion of colored pixels in the placental region is calculated as blood flow density; the distribution characteristics of blood flow signals in different regions of the placenta are analyzed; parameters characterizing blood flow velocity are extracted; and the branching patterns and complexity of blood vessels in the placenta are analyzed.

[0039] In this embodiment, feature extraction employs a multi-scale analysis method, dividing the placental region into a 3×3 grid, extracting features from each sub-region, and analyzing the feature differences between the central and peripheral regions of the placenta. Furthermore, different filters (including Laplacian of Gaussian filters and wavelet transform) are applied to the original image to extract features, capturing multi-scale information. The extracted features are optimized and selected, the correlation between features is calculated, and highly correlated redundant features are removed. Recursive feature elimination and LASSO feature selection algorithms are applied to filter the most discriminative feature subset. The filtered features are then Z-score standardized to ensure comparability of features with different dimensions.

[0040] In this embodiment, the machine learning model employs a random forest algorithm, containing 100 decision trees, each with a maximum depth of 10 and a minimum number of splits of 5. Gini impurity is used as the splitting criterion. Model training utilizes 10-fold cross-validation, and evaluation metrics include accuracy, sensitivity, specificity, and AUC. The training dataset for the machine learning model comprises 300 cases with known placental perfusion status and fetal developmental status. Placental perfusion status is determined based on Doppler blood flow parameters and clinical outcomes, while fetal developmental status is determined based on estimated fetal weight percentiles and perinatal outcomes. To address class imbalance, SMOTE technology is used for oversampling.

[0041] In this embodiment, the method also constructs a multi-task learning model to simultaneously predict placental perfusion status and fetal development status, leveraging the inherent connection between the two tasks to improve overall prediction performance. Furthermore, for cases with multiple ultrasound examination records, a time-series prediction model is constructed to predict fetal development trends. The method also performs model interpretability analysis, identifying the features with the greatest impact on prediction results through feature importance analysis; evaluating the contribution of each feature to individual prediction results through SHAP value analysis; and conducting local interpretability analysis to analyze influencing factors for specific prediction results.

[0042] In this embodiment, the visualization analysis includes: generating a heatmap using Grad-CAM technology to visually display the area in the placental region that contributes the most to the prediction results; drawing a feature importance bar chart to show the top 10 features that have the greatest impact on the prediction results and their relative importance; generating a decision curve to show the clinical utility of the model at different thresholds; and visually comparing the current case with typical normal / abnormal cases.

[0043] In this embodiment, the method also develops interactive analysis tools that allow doctors to adjust regions of interest and view the distribution of different features; provides feature adjustment functions to analyze the impact of feature changes on prediction results; allows doctors to add clinical information to optimize prediction results; the method also automatically generates standardized analysis reports, including prediction results, key features, and suggested areas of focus; generates targeted clinical recommendations based on prediction results; compares the analysis results with clinical reference ranges to highlight abnormal findings; and compares and analyzes the examination results of the same pregnant woman at different times, generating trend charts to visually display changes.

[0044] Clinical validation employed a prospective study design to validate system performance in 100 new cases; multi-center validation was conducted in three hospitals of different levels using different brands of ultrasound equipment to verify the applicability and stability of the system; and the consistency and advantages of the system's predictive results were compared with traditional assessment methods (such as Doppler indices, biophysical scores, etc.).

[0045] This method also provides evidence-based clinical intervention recommendations based on different prediction results. For cases predicted as having poor placental perfusion, it is recommended to increase the frequency of prenatal checkups and appropriately supplement with drugs such as low molecular weight heparin or aspirin; for cases predicted as having a risk of fetal growth restriction, it is recommended to strengthen nutritional support, develop a personalized delivery plan, and hospitalize for observation if necessary; personalized recommendations are provided based on the individual patient's condition. In this embodiment, grayscale ultrasound images refer to images acquired through the grayscale mode of an ultrasound diagnostic instrument. Different tissues in the image are represented by different grayscale levels (from black to white) according to their ultrasound reflection intensity, which can clearly display the anatomical information such as the morphology, outline, and internal structure of the placenta. That is, the ultrasound diagnostic instrument is switched to grayscale mode, the scanning parameters are adjusted (gain 55-65dB, depth 10-15cm, focal position placed in the middle of the placenta), and the placenta is scanned to obtain grayscale ultrasound images. For example, in a certain grayscale ultrasound image, placental tissue appears as medium grayscale (grayscale value 100-150), with clear outlines and no obvious abnormal hyperechoic or hypoechoic areas inside; amniotic fluid appears as low grayscale (grayscale value 30-50); and the uterine wall appears as high grayscale (grayscale value 180-220).

[0046] Color Doppler ultrasound images refer to images acquired using the color Doppler mode of an ultrasound diagnostic instrument. This mode utilizes the Doppler effect to superimpose blood flow signals onto a grayscale image in the form of color codes. It can visually display the direction, velocity, and distribution of blood flow in the placental vessels. Red usually indicates blood flow towards the probe, and blue indicates blood flow away from the probe. The faster the blood flow velocity, the more vibrant the color. To obtain a color Doppler ultrasound image, switch the ultrasound diagnostic instrument to color Doppler mode, adjust the pulse repetition frequency (0.6-1.2kHz), color gain (adjusted to the highest level to eliminate background noise, typically 60-70dB), and ensure the sampling volume covers the entire placental area. In a color Doppler ultrasound image, the blood flow signals in the placental vessels are distributed in red and blue. The blood flow signals in the main vessels are continuous, uniform, and vibrant, indicating normal blood flow velocity and good perfusion.

[0047] The complete area and boundary information of the placenta refer to the ultrasound image containing all placental tissues (including placental parenchyma, placental lobules, placental edges, etc.), and the clear and distinguishable boundary between the placenta and surrounding tissues (such as amniotic fluid, uterine wall, and fetus), without any missing or blurred placental areas. This is achieved by adjusting the angle and position of the ultrasound probe during scanning, using multi-plane scanning (long-axis, short-axis, and coronal planes) to ensure that each plane fully displays different parts of the placenta. Simultaneously, image gain, contrast, and other parameters are adjusted to enhance the distinction between the placental boundary and surrounding tissues. For example, during long-axis scanning, the probe moves along the long axis of the placenta to ensure that the image includes both the near-uterine wall and near-fetal ends of the placenta with clear boundaries; during short-axis scanning, the probe moves perpendicular to the long axis to ensure that the image includes the left and right edges of the placenta without any omissions.

[0048] Size adjustment refers to uniformly adjusting ultrasound images of inconsistent size and resolution acquired under different scanning conditions to a preset fixed size, which facilitates the consistency of subsequent deep neural network segmentation model processing and feature extraction. In other words, all ultrasound images are uniformly adjusted to a size of 512×512 pixels.

[0049] Brightness standardization refers to eliminating the differences in ultrasound image brightness caused by different scanning conditions (such as probe gain, scanning depth, and the thickness of abdominal fat in pregnant women), so that the brightness distribution of all images is consistent. It uses a histogram equalization algorithm to adjust the grayscale histogram of the image, so that the grayscale values ​​are evenly distributed in the range of 0-255, thereby enhancing the overall brightness consistency of the image.

[0050] Contrast enhancement refers to using the contrast-limited adaptive histogram equalization (CLAHE) algorithm to improve the grayscale difference between placental tissue and surrounding tissues (such as amniotic fluid and uterine wall) in ultrasound images, making the outline and internal structure of the placenta clearer, which facilitates subsequent segmentation and feature extraction.

[0051] Build A two-dimensional Gaussian filter with a value of 1.0 and a kernel size of 5×5 is used to achieve noise reduction by convolving the filter with the ultrasound image.

[0052] In this embodiment, the quality of ultrasound images is comprehensively evaluated, including: Image sharpness: Calculated using the gradient magnitude mean method. Gradient magnitude reflects the rate of change of pixel grayscale values. The larger the average gradient magnitude, the sharper the edges and details of the image. The specific calculation process is as follows: First, Sobel edge detection is performed on the image, calculating the gradient in the x and y directions for each pixel. Then, the gradient magnitude of each pixel is calculated. Finally, the average gradient magnitude of all pixels is calculated as the sharpness index. This index is then normalized to a score of 0-100, i.e., sharpness score = (average gradient magnitude / maximum possible gradient magnitude) × 100.

[0053] Contrast Ratio: Calculated using the grayscale dynamic range method. The grayscale dynamic range is the difference between the maximum and minimum grayscale values ​​in an image. Contrast Ratio Score = (Grayscale Dynamic Range / 255) × 100, where 255 is the maximum grayscale value range of an 8-bit grayscale image. For example, if an image has a maximum grayscale value of 220, a minimum grayscale value of 30, and a grayscale dynamic range of 190, then the contrast ratio score = (190 / 255) × 100 ≈ 74.5 points.

[0054] Signal-to-noise ratio (SNR): Calculated as the ratio of the mean of the signal region to the standard deviation of the noise region. First, manually or automatically select the signal region (placental region) and the noise region (background region without tissue signal) in the image, calculate the mean gray level of the signal region and the standard deviation of the gray level of the noise region. SNR = mean gray level / standard deviation of gray level. Then, normalize the SNR to 0-100 points, i.e., SNR score = (SNR / maximum possible SNR) × 100. The maximum possible SNR is set to 20 (determined according to the actual situation of clinical ultrasound images).

[0055] Selecting images that meet the quality standards involves setting a threshold based on the comprehensive score calculated using the aforementioned quality scoring system, selecting images with a comprehensive score ≥ 70 as those meeting the quality standards, and discarding low-quality images with a comprehensive score < 70.

[0056] The beneficial effects of the above technical solution are as follows: through a complete process of ultrasound image acquisition, preprocessing and quality control, deep neural network segmentation, multidimensional feature extraction, machine learning prediction, visualization analysis and clinical validation, the traditional subjective ultrasound assessment is transformed into an objective quantitative analysis, realizing accurate prediction of placental perfusion status and fetal development status. This effectively solves the problems of poor consistency and low early diagnosis rate in traditional ultrasound assessment. At the same time, visualization analysis and medical interpretation solve the AI ​​model problem, providing clinicians with scientific and reliable diagnostic basis and decision support, which helps in the early detection and timely intervention of FGR, and reduces the incidence and mortality of perinatal complications.

[0057] In this invention, a single boundary line is easily affected by image noise and scanning angle (e.g., Doppler ultrasound boundaries may be affected by blood flow signals, and the segmentation model boundary may be offset). Multi-source boundary lines characterize the placental boundary from different dimensions, forming complementarity. Position alignment is a prerequisite for multi-source comparison to avoid comparison failure caused by coordinate system differences. Therefore, this invention provides a method for predicting placental perfusion and fetal development based on ultrasound images. Before performing quantitative feature analysis on the segmented placental region, it also includes: The first colored boundary line of the processed ultrasound image is obtained. At the same time, the processed ultrasound image is subjected to grayscale processing to obtain the first gray boundary line and the second gray boundary line of the segmented placental region is obtained. The first gray boundary line and the first colored boundary line are aligned in position, and the overlapping pixel group of each boundary point in the second gray boundary line is obtained in sequence. The overlapping pixel group is the combination of the overlapping state of the corresponding boundary point of the second gray boundary line with the first gray boundary line and the first colored boundary line. Simultaneously, extract the eight neighboring points of each boundary point in the second gray boundary line; Obtain the set of eight neighboring pixels for each current boundary point on the second gray boundary line, and select a set of sub-points from the set of eight neighboring pixels that are consistent with the local extension direction of the second gray boundary line; The representational deformation of each sub-point in the sub-point set relative to the current boundary point is determined, and an ordered deformation sequence is constructed for each current boundary point based on the order of the sub-point set on the trajectory of the second gray boundary line. The representational deformation is determined based on the changes in gray-level gradient, texture continuity, and distance between the sub-point and the first colored boundary line and the first gray boundary line. Along the second gray boundary line, the relative rate of change between the deformation sequence corresponding to each current boundary point and the adjacent upstream boundary point is calculated sequentially to form a set of relative rates of change for each boundary point, wherein the relative rate of change reflects the evolution rate of the local deformation mode along the boundary trajectory; Extract the maximum value from the set of relative rates of change of all boundary points, and construct the first rate of change curve in the order of the boundary points. At the same time, extract the minimum value from the set of relative rates of change of all boundary points, and construct the second rate of change curve in the same order. Based on the first rate of change curve and the second rate of change curve, as well as the position coordinates of the sub-points corresponding to each extreme point on the curve in the image, a dynamically expanding boundary region around the second gray boundary line is jointly determined. The width and shape of the dynamically expanding boundary region are adaptively adjusted by the spatial distribution characteristics implied by the amplitude difference and coordinates of the first rate of change curve and the second rate of change curve at the corresponding positions. For each preset control unit within the dynamically expanded boundary region, the change concentration rate of the corresponding boundary point in the relative change rate set is calculated, and the unit size of the control unit is adjusted in combination with the overlapping pixel group corresponding to the corresponding boundary point. The change concentration rate is used to quantify the internal dispersion of the relative change rate set of the corresponding boundary point. Based on all the adjusted reference units, one or more optimized boundary trajectories are generated as several reference trajectories for correcting the second gray boundary line; The second gray boundary line is sequentially combined with each reference trajectory and output to the trajectory comparison model to output the reference value of the corresponding reference trajectory. The reference value of all reference trajectories is statistically analyzed, and useful trajectories are selected by combining the concentrated distribution distance between each reference trajectory and the second gray boundary line. The second gray boundary line is adjusted based on the reference value of the useful trajectory to obtain the effective placental region, and quantitative feature analysis is performed on the effective placental region.

[0058] In this embodiment, adjusting the size of the control unit includes: ,in, This is the adjustment amount for the cell size of the control cell corresponding to the i-th boundary point; This is the preset baseline adjustment amount; The basic adjustment coefficient, with a value ranging from 0 to 1; The sensitivity coefficient is related to the concentration ratio of change, and its value is >0; It is the attenuation constant, with a value > 0; Let be the concentration ratio of change at the i-th boundary point; The overlapping pixel group of the i-th boundary point; Based on the overlapping pixel group Consistency weight function; ; And satisfy This function makes the adjustment amount positively correlated with the degree of consistency of the boundaries of the multi-source images; Adjusted control unit size ,in, Let be the cell size of the reference cell corresponding to the i-th boundary point.

[0059] In this embodiment, the rate of change in statistical characteristics is established. Overlapping pixel groups in logical judgment information The mathematical coupling relationship between them is expressed through the exponential decay term. and piecewise weight function The two physical quantities reflecting boundary reliability, namely the stability of local deformation patterns and the consistency of multi-source image boundaries, are unified and quantified into a coordinated adjustment command for a single geometric parameter.

[0060] Using exponential function Construct a stability adjustment factor when the deformation mode is highly stable ( When the factor is very small, it approaches 1, allowing for near-full amplitude (due to the coefficient). Adjustments to the control system are made to actively optimize the reliable region; however, when the deformation mode is unstable ( When the value increases, the factor decays rapidly, producing strong nonlinear suppression of the adjustment amount, thereby effectively avoiding blind and excessive adjustments in areas with blurred or contradictory boundaries, and greatly enhancing the robustness of the algorithm in complex image regions.

[0061] In this embodiment, the design of the consistency weight function directly addresses the specific clinical challenge of the possible separation between the color Doppler blood flow boundary and the grayscale tissue boundary in placental ultrasound. By defining three weight levels with a clear size relationship, it accurately encodes the confidence level represented by different overlap states, so that the adjustment strategy strictly corresponds to the consistency of multimodal image evidence.

[0062] In this embodiment, The default value is 0.5, which balances basic adjustment and dynamic optimization and adapts to the boundary features of most ultrasound images. The default value is 0.8. It has been verified through a large number of clinical ultrasound images (500 training sets) to ensure that the adjustment amount in the stable region is fully optimized without the adjustment amount overflowing due to excessive value (exceeding the pixel-level accuracy requirements of ultrasound images). The default value is 1.2. This value is designed based on the statistical characteristics of placental ultrasound boundary deformation (the concentration rate of change C(i) of the relative rate of change set is usually in the range of 0-0.5). When C(i) > 0.3 (deformation is relatively unstable), the attenuation term quickly drops below 0.7, effectively suppressing over-adjustment. , , The weights are 0.8, 0.5, and 0.3 respectively, with a difference of 0.2-0.3. This not only distinguishes the weight differences in different overlapping states but also prevents abrupt changes in the adjustment amount due to excessive differences. It adapts to the continuity characteristics of ultrasound boundaries. This weight level is verified by the boundary characteristics of multimodal ultrasound images (grayscale + color Doppler), solving the clinical problem of "separation of blood flow boundary and tissue boundary" and ensuring that the weight matches the actual reliability of the boundary.

[0063] In this embodiment, the first colored boundary line refers to the line extracted from the processed color Doppler ultrasound image that characterizes the boundary between the placental region and surrounding tissues (such as amniotic fluid and uterine wall). This boundary line is determined based on the grayscale difference or color coding difference between blood flow signals and non-blood flow signals in the color Doppler image. Specifically, the Canny edge detection algorithm is used to extract the first colored boundary line. The specific steps are: first, convert the color Doppler ultrasound image to a grayscale image (taking the average value of the RGB three channels); then apply Gaussian filtering (Gaussian filtering... Values ​​are adjusted according to the processing stage: Preprocessing stage =1.0 (primarily for noise reduction), boundary extraction stage The grayscale image is denoised using a threshold of 0.8 (preserving edge details); then the gradient magnitude and direction of the image are calculated; non-edge pixels are removed using non-maximum suppression (NMS); finally, a dual thresholding method (low threshold 50, high threshold 150) is used to determine edge points, which are then connected to form a continuous first colored boundary line. For example, after processing a color Doppler ultrasound image, the extracted first colored boundary line is distributed along the edge of the placenta, clearly distinguishing the placental region (containing blood flow signals) and the amniotic fluid region (without blood flow signals).

[0064] The first gray boundary line refers to the line representing the boundary between the placental region and surrounding tissues extracted after grayscale processing of the processed ultrasound image (which can be a grayscale image or a grayscale image converted from a color Doppler image). It is determined solely based on the difference in grayscale values ​​of the image. Specifically, an adaptive thresholding segmentation algorithm combined with edge detection is used to extract the first gray boundary line. The specific steps are as follows: First, the processed ultrasound image is converted to grayscale (if it is a color image); then, adaptive thresholding segmentation (block size 11, constant 2) is used to segment the image into foreground (placental region) and background (non-placental region); next, morphological opening (kernel size 3×3) is performed on the segmentation result to remove small noise; finally, the Sobel edge detection algorithm is used to extract the boundary of the foreground region, forming the first gray boundary line. For example, after adaptive thresholding segmentation of a grayscale ultrasound image, the placental region is accurately segmented as the foreground, and the first gray boundary line extracted by edge detection closely matches the actual edge of the placenta.

[0065] The second gray boundary line refers to the boundary line of the placental region extracted from the segmentation result (binary image) of the processed ultrasound image after the deep neural network segmentation model has segmented it. It is the initial placental boundary output by the segmentation model. Specifically, the boundary is extracted from the binary segmented image output by the segmentation model (placental region is 1, non-placental region is 0). A morphological gradient algorithm is used to obtain the boundary pixels of the placental region by calculating the difference between the dilated and eroded images. These boundary pixels are then connected sequentially to form a continuous second gray boundary line. For example, in the binary segmented image output by the U-Net model, the placental region is white (grayscale value 255), and the non-placental region is black (grayscale value 0). After processing by the morphological gradient algorithm, the edge pixels of the white region are extracted and connected to form the second gray boundary line.

[0066] Position alignment processing refers to matching and calibrating the spatial positions of the first gray boundary line and the first colored boundary line, aligning the two boundary lines in the same coordinate system. This eliminates potential positional shifts that may occur during image acquisition and processing, ensuring the accuracy of subsequent boundary point comparisons. A feature point matching-based alignment method is employed, with the following steps: First, feature points (such as corner points and inflection points) are extracted from both the first gray and first colored boundary lines. The Shi-Tomasi corner detection algorithm is used, with a maximum feature point count of 50, a quality level of 0.01, and a minimum Euclidean distance of 10. Then, a FLANN matcher is used to match the two sets of feature points, selecting feature point pairs with high matching degrees (matching distance < 2.0). Finally, based on the matched feature point pairs, an affine transformation is used to adjust the position of the first gray boundary line, aligning the two sets of feature points and thus achieving position alignment of the two boundary lines. For example, if the first gray boundary line and the first colored boundary line are slightly offset due to image rotation, after feature point matching and affine transformation, the corresponding feature points of the two boundary lines coincide, maintaining overall consistency in position.

[0067] The overlapping pixel group refers to each boundary point on the second gray boundary line, determining whether the point overlaps with the first gray boundary line and the first colored boundary line at the pixel level (i.e., whether the coordinates of the point belong to the pixel sets of the other two boundary lines simultaneously), forming a combination of three judgment results (overlapping / not overlapping with the first gray boundary line, overlapping / not overlapping with the first colored boundary line, and whether both overlap). First, the pixel coordinates of the three boundary lines are stored in three sets (S1: pixel set of the first gray boundary line, S2: pixel set of the first colored boundary line, S3: pixel set of the second gray boundary line); then, for each pixel P(x,y) in S3, it is determined whether P belongs to S1 (yes / no) and whether it belongs to S2 (yes / no); finally, the two judgment results are combined to form the overlapping pixel group, including three cases: (yes, yes) indicates overlap with both boundary lines, (yes, no) indicates overlap with only the first gray boundary line, (no, yes) indicates overlap with only the first colored boundary line, and (no, no) indicates no overlap with either boundary line. For example, pixel P1 (100, 200) on the second gray boundary line belongs to both S1 and S2, and its overlapping pixel group is (yes, yes); pixel P2 (150, 250) belongs to S1 but not to S2, and its overlapping pixel group is (yes, no).

[0068] The eight-neighborhood points refer to the eight adjacent pixels surrounding a given boundary point on the second gray boundary line, including horizontal, vertical, and diagonal neighbors. These are used to analyze the local neighborhood information of that boundary point. Specifically, for each boundary point P(x,y) on the second gray boundary line, the coordinates of its eight-neighborhood points are (x-1,y-1), (x-1,y), (x-1,y+1), (x,y-1), (x,y+1), (x+1,y-1), (x+1,y), and (x+1,y+1). By traversing these coordinates, the corresponding pixels are obtained as the eight-neighborhood points of that boundary point. For example, the eight neighboring points of the boundary point P(200,300) are (199,299), (199,300), (199,301), (200,299), (200,301), (201,299), (201,300), (201,301).

[0069] The eight-neighbor pixel set refers to the set of all eight neighboring pixels of a current boundary point on the second gray boundary line. It contains the coordinates of the eight pixels and the corresponding image information (such as gray value, texture features, etc.). That is, for each current boundary point P(x,y), the coordinates of the eight neighboring pixels are collected according to the above method of calculating the coordinates of the eight neighboring pixels. Then, the gray value, gray gradient, texture features, and other information of each neighboring pixel are extracted from the processed ultrasound image to form the eight-neighbor pixel set of the current boundary point. For example, in the eight-neighbor pixel set of the current boundary point P(200,300), each neighboring pixel contains coordinates (e.g., 199,299), gray value (e.g., 120), gray gradient (e.g., 15), and other information.

[0070] The local extension direction refers to the tangent direction of the second gray boundary line at the current boundary point, reflecting the extension trend of the boundary line at that point. It is used to filter out neighboring points that are consistent with the extension direction of the boundary line. For example, the local extension direction can be calculated using the current boundary point and its two adjacent boundary points. Let the current boundary point be... ( , The previous boundary point is ( , The next boundary point is ( , ), then the boundary line is The direction vector of the point is ( , The direction of this vector is the local extension direction. For example, the current boundary point. (200, 300), the previous point (198, 298), the next point (202,302), then the direction vector is (4,4), and the local extension direction is 45° (the angle between the vector and the positive x-axis).

[0071] The sub-point set refers to the set of neighboring points selected from the eight-neighbor pixel set whose direction is consistent with the local extension direction of the second gray boundary line. It is used to analyze the local deformation of the current boundary point. Specifically, first, the direction vector (dx, dy) of each eight-neighbor point relative to the current boundary point is calculated as (x-neighbor - x_current, y-neighbor - y_current); then, the angle θ between this direction vector and the local extension direction vector (dx0, dy0) is calculated using the dot product formula. ; Finally, the included angle threshold was set to 30°, when When the angle is ≤30°, the direction of the neighboring point is considered to be consistent with the local extension direction, and it is included in the sub-point set. For example, if the local extension direction vector is (4,4) and the direction vector of a neighboring point is (3,3), the calculation is as follows: Its actual range is from -1 to 1, therefore, here ≈0°, this neighboring point is included in the sub-point set; the direction vector of another neighboring point is (-1,1), cosθ=0, =90°, the neighboring point is excluded.

[0072] Deformation is characterized by the morphological changes of each sub-point in the sub-point set relative to the current boundary point. It is determined comprehensively based on the sub-point's grayscale gradient, texture continuity, and distance changes between the sub-point and the first colored boundary line and the first gray boundary line, and is used to quantify the degree of difference between the sub-point and the current boundary point. A quantitative index D representing deformation is constructed, and its calculation formula is as follows: ,in: The difference in grayscale gradient between the sub-point and the current boundary point. =| | / ,in, The gray-level gradient of the sub-point. The grayscale gradient of the current boundary point; To determine the texture continuity between sub-points and the current boundary point, the correlation feature of the gray-level co-occurrence matrix is ​​used for calculation. ,in, The correlation feature value of the region where the sub-point is located. L1 represents the correlation feature value of the region where the current boundary point is located, ranging from 0 to 1; L1 represents the distance change between the sub-point and the two reference boundary lines (the first colored boundary line and the first gray boundary line). Where d1 is the distance from the sub-point to the first colored boundary line, d2 is the distance from the sub-point to the first gray boundary line, d10 is the distance from the current boundary point to the first colored boundary line, and d20 is the distance from the current boundary point to the first gray boundary line. , , Here are the weighting coefficients, with values ​​of 0.3, 0.4, and 0.3 respectively, satisfying... For example, if a sub-point has G1=0.2, T1=0.8, and L1=1.1, then its characteristic deformation D=0.71.

[0073] An ordered deformation sequence refers to a sequence formed by arranging the characteristic deformation of each sub-point in order based on the order of the sub-point set on the trajectory of the second gray boundary line. This sequence reflects the changing pattern of the local deformation of the current boundary point. First, the positional order of each sub-point in the sub-point set on the trajectory of the second gray boundary line is determined. Using the current boundary point as the center, the sub-points are sorted from closest to furthest from the current boundary point (or clockwise if the distance is the same). Then, the characteristic deformation values ​​of each sub-point are arranged sequentially to form an ordered deformation sequence. For example, if the sub-point set of the current boundary point contains 4 sub-points, and their characteristic deformation values ​​after being sorted are 0.65, 0.71, 0.58, and 0.62, then the ordered deformation sequence is [0.65, 0.71, 0.58, 0.62].

[0074] The adjacent upstream boundary point refers to the boundary point on the trajectory of the second gray boundary line that is located before the current boundary point (in the opposite direction of the boundary line's extension) and directly adjacent to the current boundary point. It is used to calculate the deformation sequence change between the current boundary point and the upstream boundary point. That is, all boundary points of the second gray boundary line are numbered sequentially in a clockwise direction. , ,..., For the current boundary point (i≥2), its adjacent upstream boundary point is ; For the current boundary point Its adjacent upstream boundary point is (Connected end to end). For example, the boundary points of the second gray boundary line are numbered as follows: , , , , The current boundary point is Then its adjacent upstream boundary point is The current boundary point is Its adjacent upstream boundary point is .

[0075] The relative rate of change refers to the degree of change between the ordered deformation sequence at the current boundary point and the ordered deformation sequence at the adjacent upstream boundary point, reflecting the evolution rate of the local deformation mode along the boundary trajectory. That is, let the current boundary point... The ordered deformation sequence is =[s1,s2,...,sm], adjacent upstream boundary points The ordered deformation sequence is =[s1',s2',...,sm'], then the relative rate of change at each corresponding position. j=1,2,...,m, where, To avoid correction terms with a denominator of 0, if ,but .

[0076] The relative rate of change set refers to the set of relative rates of change at all corresponding positions of the current boundary point, containing information on the deformation rate of change between the current boundary point and the upstream boundary point at each sub-point location. That is, the relative rate of change at each corresponding position calculated above. Arrange the sub-points in order to form the current boundary points. set of relative rates of change .

[0077] In this embodiment, the first rate of change curve refers to the curve formed by extracting the maximum value of the relative rate of change set of all boundary points, arranging the boundary points in order along the second gray boundary line, and reflecting the maximum deformation rate of change along the boundary line. This process iterates through each boundary point of the second gray boundary line. From its relative rate of change set Extract the maximum value Then, using the boundary point numbers as the x-coordinates, the corresponding... Using the vertical axis as the ordinate, plot the curve; this is the curve representing the first rate of change. (Example: boundary points) , , , , The maximum values ​​of the relative rate of change are 0.08, 0.09, 0.079, 0.085, and 0.07, respectively. The coordinate points of the first rate of change curve are (1, 0.08), (2, 0.09), (3, 0.079), (4, 0.085), and (5, 0.07). Connecting these points forms a curve.

[0078] The second rate of change curve refers to the curve formed by extracting the minimum value of the relative rate of change set of all boundary points, arranging the boundary points in order along the second gray boundary line, and reflecting the minimum deformation rate of change along the boundary line. That is, it iterates through each boundary point of the second gray boundary line. From its relative rate of change set Extract the minimum value Then, using the boundary point numbers as the x-coordinates, the corresponding... Using the vertical axis as the ordinate, plot the curve; this is the second rate of change curve. (Example: boundary points) , , , , The minimum values ​​of the relative rate of change are 0.04, 0.035, 0.042, 0.038, and 0.045, respectively. The coordinate points of the second rate of change curve are (1, 0.04), (2, 0.035), (3, 0.042), (4, 0.038), and (5, 0.045). Connecting these points forms a curve.

[0079] Extreme points refer to the points on the first and second rate of change curves where the values ​​are extremely large or extremely small. These points correspond to boundary points and sub-points that are typically located in areas of drastic deformation, and are crucial for boundary correction. A sliding window method (window size 3) is used to identify extreme points. For a point Q on the curve, if its value is greater than the values ​​of its two adjacent points, then Q is a maximum point; if its value is less than the values ​​of its two adjacent points, then Q is a minimum point. For example, if the numerical sequence of the first rate of change curve is [0.08, 0.09, 0.079, 0.085, 0.07], then the maximum points are (2, 0.09) and (4, 0.085), and the minimum points are (3, 0.079) and (5, 0.07).

[0080] The dynamically expanded boundary region refers to the area surrounding the second gray boundary line, determined based on the coordinates of the first rate of change curve, the second rate of change curve, and the sub-points corresponding to the extreme points on the curves. The width and shape of this region can adaptively adjust according to the amplitude difference of the rate of change curves and the coordinate space distribution characteristics, for subsequent boundary optimization. The formula for calculating the width of the dynamically expanded boundary region is: Where k is the width increment corresponding to the unit rate of change amplitude difference (for an ultrasound image with a resolution of 512×512 pixels, the default value is 5 pixels; if the image resolution is 1024×1024 pixels, it can be adjusted to 8 pixels; if it is 256×256 pixels, it can be adjusted to 3 pixels to ensure the consistency of width adjustment under different resolutions). W0 represents the amplitude difference between the first and second rate of change curves at the current boundary point, and W0 is the base width (3 pixels). The shape of the region is adjusted according to the coordinates of the sub-points corresponding to the extreme points. At the extreme points, the region expands in the direction of increasing amplitude difference; at non-extreme points, the region maintains the base width. For example, if the amplitude difference at a boundary point is 0.05, the region width W = 5 × 0.05 + 3 = 3.25 pixels, rounded to 3 pixels; if the amplitude difference at an extreme point is 0.08, the width W = 5 × 0.08 + 3 = 3.4 pixels, rounded to 3 pixels, and the region shape at that location slightly expands in the direction of the sub-points.

[0081] The preset control unit refers to the smallest unit pre-divided within the dynamically expanding boundary area for analyzing the concentration rate of boundary point changes. It is usually a square or rectangular pixel block. For example, within the dynamically expanding boundary area, the control units are evenly divided into 3×3 pixel sizes.

[0082] The concentration rate of change is an indicator used to quantify the internal dispersion of the relative rate of change set at a corresponding boundary point. Its core function is to reflect the clustering of values ​​within the relative rate of change set. A smaller concentration rate indicates a more concentrated set of values ​​and a more stable local deformation pattern; conversely, a larger concentration rate indicates dispersed values ​​and poorer deformation pattern stability, providing crucial quantitative evidence for adjusting the size of the control unit. The concentration rate of change is calculated using the coefficient of variation (the ratio of standard deviation to mean) combined with the entropy method. The specific formula is as follows: ,in, For information entropy, ,in, It represents the probability of the j-th value in the set of relative change rates, which is the ratio of the frequency of that value in the set to the total number of values.

[0083] The control unit size adjustment refers to the adaptive adjustment of the initial size of the preset control unit within the dynamically expanded boundary area based on the change concentration rate and overlapping pixel groups of each boundary point. This makes the size of the control unit match the deformation stability and multi-source boundary consistency of the boundary points, thereby improving the accuracy of subsequent boundary trajectory optimization.

[0084] The optimized boundary trajectory refers to a trajectory line generated based on all adjusted control units. This trajectory line, generated by analyzing the grayscale features, texture continuity, and distance relationship of pixels within the unit to the reference boundary, more accurately represents the actual boundary of the placenta and is used to correct the second gray boundary line obtained from the initial segmentation. Specifically, for each adjusted control unit, the average grayscale gradient of pixels within the unit and the average distance to the first colored boundary line and the first gray boundary line are calculated. Pixels with an average gradient greater than a threshold (preset to 15) and an average distance less than a threshold (preset to 5 pixels) are selected as valid pixels. A polynomial fitting algorithm is used to fit all valid pixels into a continuous curve along the boundary line extension direction; each curve is an optimized boundary trajectory. For example, if the dynamically expanded boundary region contains 50 adjusted control units, 800 valid pixels are obtained after screening. A third-order polynomial fitting yields two continuous curves, corresponding to different possible trajectories of the placental boundary, which are the optimized boundary trajectories.

[0085] The trajectory comparison model is used to evaluate the degree of matching and clinical rationality of each reference trajectory with the second gray boundary line, and outputs the usefulness (reference value) of the reference trajectory for boundary correction. It employs a model based on similarity indices and clinical constraints. Similarity indices include the Dice coefficient of the trajectory (the degree of overlap with the second gray boundary line) and the Hausdorff distance (maximum distance deviation). Clinical constraints include whether the area enclosed by the trajectory falls within the normal placental area range (determined according to gestational age, e.g., a normal placental area of ​​150-300 mm at 28 weeks of gestation). ), and boundary smoothness (curvature change rate less than the threshold of 0.1). The model outputs a reference value score (0-10 points), score = (Dice coefficient × 4 + (1 - hausdorff distance / 50) × 3 + clinical constraint compliance × 3), where clinical constraint compliance is 1 for satisfying all constraints, 0.5 for violating one constraint, and 0 for violating two or more constraints. For example, if the Dice coefficient of a reference trajectory and the second gray boundary line is 0.85, the hausdorff distance is 8 pixels, and all clinical constraints are met, then the reference value score = 8.92 points.

[0086] Reference value refers to the score output by the trajectory comparison model, used to quantify the effectiveness of the reference trajectory in correcting the second gray boundary line. A higher score indicates that the reference trajectory is closer to the actual placental boundary, and the greater the correction value. The reference value score is calculated using multi-dimensional indicators of the trajectory comparison model, ranging from 0 to 10 points, with 8-10 points indicating high reference value, 5-7 points indicating medium reference value, and 0-4 points indicating low reference value. For example, a reference trajectory with a score of 8.92 has high reference value, indicating that the trajectory can effectively correct the second gray boundary line; a reference trajectory with a score of 4.2 has low reference value, indicating that its correction significance is limited.

[0087] The clustered distribution distance refers to the average Euclidean distance between all corresponding points on each reference trajectory and the second gray boundary line. It measures the overall deviation between the two trajectories; a smaller average value indicates a more concentrated trajectory distribution and a higher degree of matching. Specifically, sampling points are taken at equal intervals (preset to 5 pixels) between the second gray boundary line and the reference trajectory. The Euclidean distance between each pair of sampling points is calculated, and the arithmetic mean of all distances is the clustered distribution distance. For example, if 100 sampling points are taken from both the second gray boundary line and a certain reference trajectory, and the sum of the Euclidean distances between all pairs of sampling points is 300 pixels, then the clustered distribution distance = 300 / 100 = 3 pixels, indicating that the overall deviation between the two trajectories is relatively small.

[0088] Useful trajectories are those selected from all reference trajectories that simultaneously meet the criteria of a reference value score and a concentration distribution distance, effectively serving to correct the second gray boundary line. This involves setting a selection threshold: a reference value score ≥ 6 (medium-high reference value) and a concentration distribution distance ≤ 10 pixels (small deviation). Simultaneously, statistical analysis is performed on the reference value of all reference trajectories (calculating the mean and standard deviation), eliminating outlier trajectories with scores below the mean minus the standard deviation. For example, if 10 reference trajectories have reference value scores of 8.92, 7.5, 6.8, 5.2, 4.9, 7.2, 6.5, 3.8, 9.1, and 6.1, and concentration distribution distances of 3, 5, 4, 12, 15, 6, 8, 20, 2, and 7, the selected useful trajectories are 7 trajectories with a score ≥ 6 and a distance ≤ 10 (scores 8.92, 7.5, 6.8, 7.2, 6.5, 9.1, 6.1; distances 3, 5, 4, 6, 8, 2, 7).

[0089] The effective placental region refers to the area that accurately reflects the actual location, shape, and boundary of the placenta after the second gray boundary line is fused and adjusted with all useful trajectories. It is the final target area for subsequent quantitative feature analysis. Specifically, a weighted fusion algorithm is used to adjust the second gray boundary line, with the weight of each useful trajectory being the ratio of its reference value score to the sum of the reference value scores of all useful trajectories. For each boundary point of the second gray boundary line, the weighted average coordinates of its corresponding points on all useful trajectories are calculated. ,in, Let k be the weight of the useful trajectory. Let (the coordinates of the corresponding point on the k-th trajectory) be the boundary coordinates. Connect all the adjusted boundary points in sequence to form the final effective placental region boundary. The area enclosed by the boundary is the effective placental region. For example, the reference value scores of the three useful trajectories are 8.92, 9.1, and 7.5, respectively, with weights of 8.92 / (8.92+9.1+7.5)≈0.34, 9.1 / (25.52)≈0.357, and 7.5 / 25.52≈0.294, respectively. The coordinates of a point on the second gray boundary line are (200, 300), and the coordinates of the corresponding points on the three trajectories are (201, 301), (200, 300), and (202, 302), respectively. The adjusted coordinates are (200.99, 300.99), and the adjusted coordinates of this point are approximately (201, 301). The area formed by all the adjusted points is the effective placental region.

[0090] The beneficial effects of the above technical solution are as follows: By introducing comparative analysis of the first colored boundary line, the first gray boundary line, and the segmentation boundary line, and combining the deformation characteristics, rate of change curves, and dynamic expansion region optimization of the boundary points, accurate correction of the initially segmented placental region is achieved, effectively solving the problems of boundary offset and ambiguity that may exist in a single segmentation model, and significantly improving the segmentation accuracy and reliability of the placental region; through steps such as adaptive adjustment of the control unit, reference trajectory screening and fusion, the boundary trajectory is further optimized to ensure that the effective placental region can truly reflect the actual structure of the placenta, providing high-precision regional data for subsequent quantitative feature analysis, thereby improving the accuracy of placental perfusion and fetal development status prediction.

[0091] This invention provides a system for predicting placental perfusion and fetal development based on ultrasound imaging, such as... Figure 2 As shown, it includes: an image processing module, used to acquire ultrasound images of the placenta of pregnant women, and to perform preprocessing and quality control processing on the placental ultrasound images; The segmentation module inputs the processed ultrasound image into a deep neural network segmentation model to obtain segmented placental regions. It then performs quantitative feature analysis on these regions, extracting multidimensional feature vectors including grayscale, texture, morphology, and blood flow features. Before quantitative feature analysis, boundary correction is performed on the segmented placental regions, including multi-boundary line alignment, deformation sequence construction, dynamic expansion region determination, reference trajectory selection, and boundary adjustment. The prediction module inputs the multidimensional feature vectors into a pre-trained machine learning model to predict whether placental perfusion is normal or poor, and whether fetal development is normal or restricted. The visualization module visualizes the prediction results, presenting the predicted placental perfusion status and fetal development status, along with key influencing factors. The validation and interpretation module is used to validate the accuracy of prediction results in obstetric clinical practice and provide corresponding medical interpretations.

[0092] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements any of the steps of the method described above.

[0093] The beneficial effects of the above technical solution are as follows: through a complete process of ultrasound image acquisition, preprocessing and quality control, deep neural network segmentation, multidimensional feature extraction, machine learning prediction, visualization analysis and clinical validation, the traditional subjective ultrasound assessment is transformed into an objective quantitative analysis, realizing accurate prediction of placental perfusion status and fetal development status. This effectively solves the problems of poor consistency and low early diagnosis rate in traditional ultrasound assessment. At the same time, the visualization analysis and medical interpretation solve the AI ​​model problem, providing clinicians with scientific and reliable diagnostic basis and decision support, which helps in the early detection and timely intervention of FGR, and reduces the incidence and mortality of perinatal complications.

[0094] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting placental perfusion and fetal development based on ultrasound imaging, characterized in that, include: Step 1: Obtain ultrasound images of the placenta of the pregnant woman, and perform preprocessing and quality control processing on the placental ultrasound images; Step 2: Input the processed ultrasound image into a deep neural network segmentation model to obtain the segmented placental region. Perform quantitative feature analysis on the segmented placental region to extract multidimensional feature vectors, including grayscale features, texture features, morphological features, and blood flow features. Before performing quantitative feature analysis on the segmented placental region, perform boundary correction on the segmented placental region. The boundary correction includes multi-boundary line alignment, deformation sequence construction, dynamic expansion region determination, reference trajectory screening, and boundary adjustment. Step 3: Input the multidimensional feature vectors into a pre-trained machine learning model to predict whether the placental perfusion status is normal or poor, and whether the fetal development status is normal or developmentally restricted. Step 4: Perform visualization analysis on the prediction results to present the predicted results and key influencing factors of placental perfusion status and fetal development status. Step 5: Verify the accuracy of the prediction results in obstetric clinical practice and provide corresponding medical interpretations.

2. The method according to claim 1, characterized in that, Acquiring ultrasound images of the placenta in pregnant women includes: acquiring ultrasound images of the placenta at different gestational weeks, including grayscale ultrasound images and color Doppler ultrasound images. The acquired ultrasound images contain complete area and boundary information of the placenta.

3. The method according to claim 1, characterized in that, The preprocessing and quality control processes include: resizing, normalizing brightness, enhancing contrast, and denoising the ultrasound images of the placenta using Gaussian filtering; and selecting images that meet quality standards based on quality scores for image sharpness, contrast, and signal-to-noise ratio.

4. The method according to claim 1, characterized in that, The deep neural network segmentation model is one or a combination of several of the U-Net, MaskR-CNN, or DeepLab series networks, and the segmentation model is trained using a labeled placental ultrasound image dataset.

5. The method according to claim 1, characterized in that, Before performing quantitative feature analysis on the segmented placental regions, the following steps are also included: The first colored boundary line of the processed ultrasound image is obtained. At the same time, the processed ultrasound image is subjected to grayscale processing to obtain the first gray boundary line and the second gray boundary line of the segmented placental region is obtained. The first gray boundary line and the first colored boundary line are aligned in position, and the overlapping pixel group of each boundary point in the second gray boundary line is obtained in sequence. The overlapping pixel group is the combination of the overlapping state of the corresponding boundary point of the second gray boundary line with the first gray boundary line and the first colored boundary line. Simultaneously, extract the eight neighboring points of each boundary point in the second gray boundary line; Obtain the set of eight neighboring pixels for each current boundary point on the second gray boundary line, and select a set of sub-points from the set of eight neighboring pixels that are consistent with the local extension direction of the second gray boundary line; The representational deformation of each sub-point in the sub-point set relative to the current boundary point is determined, and an ordered deformation sequence is constructed for each current boundary point based on the order of the sub-point set on the trajectory of the second gray boundary line. The representational deformation is determined based on the changes in gray-level gradient, texture continuity, and distance between the sub-point and the first colored boundary line and the first gray boundary line. Along the second gray boundary line, the relative rate of change between the deformation sequence corresponding to each current boundary point and the adjacent upstream boundary point is calculated sequentially to form a set of relative rates of change for each boundary point, wherein the relative rate of change reflects the evolution rate of the local deformation mode along the boundary trajectory; Extract the maximum value from the set of relative rates of change of all boundary points, and construct the first rate of change curve in the order of the boundary points. At the same time, extract the minimum value from the set of relative rates of change of all boundary points, and construct the second rate of change curve in the same order. Based on the first rate of change curve and the second rate of change curve, as well as the position coordinates of the sub-points corresponding to each extreme point on the curve in the image, a dynamically expanding boundary region around the second gray boundary line is jointly determined. The width and shape of the dynamically expanding boundary region are adaptively adjusted by the spatial distribution characteristics implied by the amplitude difference and coordinates of the first rate of change curve and the second rate of change curve at the corresponding positions. For each preset control unit within the dynamically expanded boundary region, the change concentration rate of the corresponding boundary point in the relative change rate set is calculated, and the unit size of the control unit is adjusted in combination with the overlapping pixel group corresponding to the corresponding boundary point. The change concentration rate is used to quantify the internal dispersion of the relative change rate set of the corresponding boundary point. Based on all the adjusted reference units, one or more optimized boundary trajectories are generated as several reference trajectories for correcting the second gray boundary line; The second gray boundary line is sequentially combined with each reference trajectory and output to the trajectory comparison model to output the reference value of the corresponding reference trajectory. The reference value of all reference trajectories is statistically analyzed, and useful trajectories are selected by combining the concentrated distribution distance between each reference trajectory and the second gray boundary line. The second gray boundary line is adjusted based on the reference value of the useful trajectory to obtain the effective placental region, and quantitative feature analysis is performed on the effective placental region.

6. The method according to claim 1, characterized in that, The grayscale features include: average grayscale value, standard deviation, skewness, and kurtosis; the texture features include: grayscale co-occurrence matrix features, local binary pattern features, and Gabor filter features; the morphological features include: placental area, circumference, thickness, and uniformity index; the blood flow features include: placental blood flow velocity, resistance index, pulsatility index, and blood flow distribution uniformity.

7. The method according to claim 1, characterized in that, The visualization analysis includes: displaying key areas in the placental region related to perfusion status in the form of a heatmap, generating a feature importance ranking chart, providing confidence scores for prediction results, and generating clinical recommendations for different prediction results.

8. The method according to claim 1, characterized in that, Also includes: When the predicted results indicate poor placental perfusion and / or fetal growth restriction, intervention recommendations are generated, including drug treatment plans, nutritional adjustment recommendations, pregnancy management plans, and delivery timing recommendations.

9. A system for predicting placental perfusion and fetal development based on ultrasound imaging, characterized in that, include: An image processing module is used to acquire ultrasound images of the placenta of pregnant women and to perform preprocessing and quality control processing on the placental ultrasound images. The segmentation module inputs the processed ultrasound image into a deep neural network segmentation model to obtain segmented placental regions. It then performs quantitative feature analysis on these regions, extracting multidimensional feature vectors including grayscale, texture, morphology, and blood flow features. Before quantitative feature analysis, boundary correction is performed on the segmented placental regions, including multi-boundary line alignment, deformation sequence construction, dynamic expansion region determination, reference trajectory selection, and boundary adjustment. The prediction module inputs the multidimensional feature vectors into a pre-trained machine learning model to predict whether placental perfusion is normal or poor, and whether fetal development is normal or restricted. The visualization module visualizes the prediction results, presenting the predicted placental perfusion status and fetal development status, along with key influencing factors. The validation and interpretation module is used to validate the accuracy of prediction results in obstetric clinical practice and provide corresponding medical interpretations.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method described in any one of claims 1-8.

Images