Obstetrical ultrasonic examination data intelligent analysis system and method

Abstract

The invention discloses an intelligent analysis system and method for ultrasonic examination data in the obstetrics department, and the method comprises the steps: firstly, solving the problems of data inconsistency and characteristic incompatibility caused by the variation of the body position and section of a fetus through standardization processing and attitude-anatomy alignment based on a spatial transformation network; further, a multi-scale feature descriptor containing rich anatomical structure information is generated from the aligned high-dimensional feature map to replace a sparse traditional biological measurement value; and finally, carrying out time sequence trajectory modeling on a sequence formed by the feature descriptors, and deeply mining a dynamic process and an individualized trend of fetal growth. Through the mode, accurate feature registration of ultrasonic images of different gestational weeks can be realized, a personalized dynamic growth model reflecting growth rate and mode change is constructed, and a quantitative growth abnormality risk score is finally output, so that the capability of accurately predicting fetal growth abnormality in an early stage is remarkably improved.

Description

Technical Field

[0001] This application relates to the field of intelligent analysis, and more specifically, to an intelligent analysis system and method for obstetric ultrasound examination data. Background Technology

[0002] Obstetric ultrasound examination is a core technology for monitoring fetal vital signs and assessing growth and development, and is crucial for the early identification and intervention of potential growth abnormalities. In clinical practice, the interpretation and analysis of massive amounts of dynamic ultrasound images not only increases the workload of physicians, but also makes diagnostic results susceptible to subjective experience, hindering complete standardization and precision. Therefore, developing an intelligent analysis system for obstetric ultrasound examination data to improve diagnostic efficiency, objectivity, and predictive accuracy has become an important development direction in the field of smart healthcare.

[0003] However, existing data analysis methods still have significant shortcomings in depth and accuracy. Current mainstream technologies mainly rely on extracting a few one-dimensional biological measurements from specific cross-sectional images and performing isolated static comparisons with standard growth curves based on large-scale statistics. This approach first has limitations in information dimensionality, compressing high-dimensional image information containing rich morphological, textural, and tissue features into sparse low-dimensional values, failing to fully reflect the complexity of fetal growth. Second, due to the variability of fetal posture and scanning cross-sections, it is difficult to achieve precise alignment of ultrasound images acquired at different stages at the feature level, hindering the construction of longitudinal analysis models that can track the dynamic evolution of anatomical structures. More importantly, this memoryless, single-point assessment method ignores the continuity and temporal dependence of the growth process, failing to effectively capture changes in growth rate and individualized trend characteristics, resulting in insufficient capabilities in deep growth trajectory modeling and early risk prediction.

[0004] Therefore, we look forward to an optimized intelligent analysis solution for obstetric ultrasound examination data. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide an intelligent analysis system and method for obstetric ultrasound examination data. First, it resolves data inconsistencies and feature incomparability caused by variations in fetal position and cross-section through standardization processing and pose-anatomical alignment based on a spatial transformation network. Next, it generates multi-scale feature descriptors containing rich anatomical structural information from the aligned high-dimensional feature map to replace sparse traditional biological measurements. Finally, it models the temporal trajectory of the sequence composed of these feature descriptors to deeply explore the dynamic process and individualized trends of fetal growth. In this way, it can achieve precise feature registration of ultrasound images at different gestational weeks, construct a personalized dynamic growth model reflecting changes in growth rate and pattern, and ultimately output a quantitative risk score for growth abnormalities, thereby significantly improving the ability to predict fetal growth abnormalities early and accurately.

[0006] According to one aspect of this application, an intelligent analysis method for obstetric ultrasound examination data is provided, comprising: The original ultrasound data sequence is standardized to obtain a standardized ultrasound image sequence; A pose-anatomical alignment based on a spatial transformation network is performed on standardized ultrasound image sequences to obtain aligned high-dimensional ultrasound feature map sequences. The aligned high-dimensional ultrasound feature map sequence is subjected to feature descriptor generation based on multi-scale anatomical structure to obtain an ultrasound feature descriptor sequence; Temporal trajectory modeling and risk prediction were performed on ultrasound feature descriptor sequences to obtain growth abnormality risk scores.

[0007] According to another aspect of this application, an intelligent analysis system for obstetric ultrasound examination data is provided, comprising: The standardization module is used to standardize the raw ultrasound data sequence to obtain a standardized ultrasound image sequence; The pose-anatomical alignment module is used to perform pose-anatomical alignment on standardized ultrasound image sequences based on a spatial transformation network to obtain aligned high-dimensional ultrasound feature map sequences. The feature descriptor generation module is used to generate feature descriptors based on multi-scale anatomical structures from the aligned high-dimensional ultrasound feature map sequence to obtain an ultrasound feature descriptor sequence. The growth abnormality analysis module is used to perform time-series trajectory modeling and risk prediction on ultrasound feature descriptor sequences to obtain growth abnormality risk scores.

[0008] Compared with existing technologies, the intelligent analysis system and method for obstetric ultrasound examination data provided in this application firstly solves the problems of data inconsistency and feature incomparability caused by variations in fetal position and cross-section through standardization processing and pose-anatomical alignment based on spatial transformation networks. Then, it generates multi-scale feature descriptors containing rich anatomical structural information from the aligned high-dimensional feature map to replace sparse traditional biological measurements. Finally, it models the temporal trajectory of the sequence composed of these feature descriptors to deeply explore the dynamic process and individualized trends of fetal growth. In this way, it can achieve accurate feature registration of ultrasound images at different gestational weeks, construct personalized dynamic growth models reflecting changes in growth rate and pattern, and ultimately output quantitative growth abnormality risk scores, thereby significantly improving the ability to predict fetal growth abnormalities early and accurately. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of an intelligent analysis method for obstetric ultrasound examination data according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow in the intelligent analysis method for obstetric ultrasound examination data according to an embodiment of this application; Figure 3 This is a block diagram of an intelligent analysis system for obstetric ultrasound examination data according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0016] The technical solution of this application proposes an intelligent analysis method for obstetric ultrasound examination data. Figure 1 This is a flowchart of an intelligent analysis method for obstetric ultrasound examination data according to an embodiment of this application. Figure 2 This is a system architecture diagram of the intelligent analysis method for obstetric ultrasound examination data according to an embodiment of this application. Figure 1 and Figure 2 As shown, the intelligent analysis method for obstetric ultrasound examination data according to an embodiment of this application includes the following steps: S1, standardizing the original ultrasound data sequence to obtain a standardized ultrasound image sequence; S2, performing pose-anatomical alignment on the standardized ultrasound image sequence based on a spatial transformation network to obtain an aligned high-dimensional ultrasound feature map sequence; S3, generating feature descriptors based on multi-scale anatomical structures on the aligned high-dimensional ultrasound feature map sequence to obtain an ultrasound feature descriptor sequence; S4, performing temporal trajectory modeling and risk prediction on the ultrasound feature descriptor sequence to obtain a growth abnormality risk score.

[0017] Specifically, S1 involves standardizing the original ultrasound data sequence to obtain a standardized ultrasound image sequence. It should be understood that due to differences in fetal position, amniotic fluid volume, gestational age, and operator technique, the acquired original ultrasound images exhibit significant inconsistencies in quality, size, grayscale distribution, and the anatomical sections they present. This inconsistency makes it difficult to directly and effectively compare features and perform longitudinal analysis on image data from different time points and different objects, severely limiting the accuracy and reliability of subsequent high-level intelligent analysis tasks. Therefore, in the technical solution of this application, a series of standardization processes are used to eliminate these non-substantial differences, transforming heterogeneous original data into homogeneous, comparable standardized ultrasound image sequences, providing a stable and standardized data foundation for subsequent steps such as pose-anatomical alignment, feature extraction, and temporal modeling.

[0018] In practice, the first step is to extract a candidate image frame pool from the original ultrasound data sequence. In this step, the original ultrasound data sequence usually refers to a dynamic video stream or a series of static image frames acquired during a single obstetric examination. Extracting the candidate image frame pool involves decomposing these dynamic or static original data into individual image frames and collecting them to form a dataset to be processed. Next, the candidate image frame pool is subjected to standard section identification and optimization based on a hierarchical classification model to obtain an optimal section mapping table. In this process, the system uses a pre-trained hierarchical classification model to analyze each frame in the pool. First, it determines whether the frame belongs to a pre-defined standard anatomical section with clinical diagnostic value (such as biparietal diameter, head circumference, abdominal circumference, femur length, etc.). Then, within a class of standard sections, an image quality assessment algorithm scores all frames that meet the criteria, selecting the frame with the highest quality as the optimal representative of that section. Finally, the identification information of all standard sections and their corresponding optimal image frames is recorded in an optimal section mapping table. This optimal section mapping table is a key data structure, stored in key-value pairs. The key is the name of the standard anatomical section, and the value is the index or identifier of the corresponding optimal image frame in the candidate frame pool, providing a direct index for subsequent precise processing. Furthermore, based on the preferred cross-section mapping table, image standardization is performed on each optimal image frame object in the candidate image frame pool to obtain the standardized ultrasound image sequence. Specifically, grayscale conversion, size normalization, and intensity normalization are performed on each optimal image frame object to obtain the standardized ultrasound image sequence. In this process, firstly, grayscale conversion is performed on each optimal image frame object, that is, the original image, which may contain pseudocolor or Doppler information, is uniformly converted into a single-channel grayscale image to focus on anatomical structure and tissue texture information; secondly, size normalization is performed, and image frames of different resolutions and sizes are uniformly scaled to a preset standard size (e.g., 256x256 pixels) through interpolation and other algorithms to meet the input requirements of subsequent deep learning models; finally, intensity normalization is performed, and the pixel intensity values ​​of the image are linearly mapped from the original range (e.g., 0-255) to a fixed range more suitable for numerical calculation (e.g., 0-1 or -1-1) to accelerate model convergence and improve stability.

[0019] Specifically, in step S2, a pose-anatomical alignment based on a spatial transformation network is performed on the standardized ultrasound image sequence to obtain an aligned high-dimensional ultrasound feature map sequence. It should be understood that even after preliminary standardization, key anatomical structures in ultrasound images (such as the fetal head and abdomen) will still exhibit different poses, positions, and scales due to the fetus's free movement within the uterus and subtle differences in the ultrasound probe's scanning angle. This pose inconsistency is a major obstacle to subsequent accurate longitudinal analysis and temporal modeling, rendering direct comparison of images from different gestational weeks meaningless at the pixel or feature level. Therefore, in the technical solution of this application, a spatial transformation network is introduced to actively and intelligently learn and execute a spatial transformation, aligning the anatomical structures within each frame of the sequence to a unified and standardized pose. This eliminates interference caused by pose differences, ensuring that subsequently extracted features have cross-temporal comparability and consistency, laying a solid foundation for constructing an accurate growth trajectory model.

[0020] In practice, the first step is to extract a first standardized ultrasound image from the standardized ultrasound image sequence. This step indicates that subsequent steps must perform the same processing on each frame of the standardized image in the sequence. Next, the first standardized ultrasound image is input into a localization network dedicated to parameter regression to obtain affine transformation parameters. It should be understood that although each frame of the standardized ultrasound image is uniform in size and grayscale, the pose (i.e., position, orientation, and size) of its key anatomical structures remains arbitrary. To align these structures to a unified standard template, a unique set of spatial transformation parameters must be calculated for each frame to transform it from its current pose to the target standard pose. The localization network for parameter regression is a core component of the Spatial Transformation Network (STN), which is typically a convolutional neural network. Its task is not classification or segmentation, but rather to regress a set of specific parameters that describe the desired spatial transformation. In the technical solution of this application, the first standardized ultrasound image is input into a localization network dedicated to parameter regression to deeply analyze the current pose of the anatomical structures in the first standardized ultrasound image and directly output precise mathematical parameters for correcting this pose through regression.

[0021] Specifically, in this process, the parameter regression process includes: first, inputting the first normalized ultrasound image into the encoder of the localization network to obtain a global feature vector of the anatomical structure pose. It should be understood that the original image matrix contains massive amounts of highly redundant low-level information (such as the brightness value of a single pixel), but does not explicitly express high-level, semantic features (such as contours, shapes, textures, and spatial relationships). Therefore, in the technical solution of this application, the encoder of the localization network effectively compresses and refines information from the input image. This encoder can map the image from pixel space to a more discriminative feature space, generating a compact global feature vector that discards irrelevant details and specifically captures and encodes essential information related to the pose of the anatomical structure, such as spatial location, rotation angle, and scale.

[0022] The specific execution process of this step is based on hierarchical feature extraction using a deep convolutional neural network (CNN). This structure typically consists of several stacked convolutional modules, each usually containing one or more convolutional layers, an activation function, and a pooling layer. In this process, firstly, in the shallower convolutional layers, the network uses small-sized convolutional kernels (filters) to perform sliding window operations on the image to detect and respond to simple features at the lower levels, such as edges, corners, gradients, and textures. The outputs of these operations form a series of feature maps; subsequently, these feature maps are passed through an activation function, introducing non-linearity, enabling the network to learn more complex patterns. Next, the pooling layer downsamples the feature map to reduce the spatial dimension of the data, thereby reducing the complexity and number of parameters in subsequent calculations. As the image data is processed deeper into the network, subsequent convolutional layers operate on the feature map output by the previous layer, learning and combining low-level features to form higher-order, more abstract semantic features, such as combining edges and textures to form the local contours or parts of specific organs of the fetus (such as the skull and abdominal cavity). After processing by the last convolutional module, a multi-channel 3D feature map is obtained, which is then flattened to obtain the global feature vector of the anatomical structure pose.

[0023] Next, the global feature vector of anatomical structure posture is fine-tuned at the feature level to obtain an optimized global feature vector of anatomical structure posture. It is understandable that due to differences in fetal movement within the mother's body and ultrasound scanning angles, even under the same standard section, images acquired at different times may still exhibit subtle differences in spatial structure and posture representation. This directly affects the quality of subsequent feature alignment and the stability of temporal modeling. Therefore, the global feature vector of anatomical structure posture extracted after initial posture alignment via a spatial transformation network is further optimized to improve the purity and task relevance of its feature representation. Specifically, in the technical solution of this application, a structured, adaptive feature refinement mechanism is used to eliminate redundant information, noise, or components irrelevant to the current analysis task that may be contained in the original posture features, thereby enhancing the discriminative power and robustness of the feature representation. Through refined control at the feature level, the overall model's ability to capture the risk of abnormal growth and development in complex ultrasound image sequences is enhanced, promoting the development of obstetric ultrasound diagnosis towards a higher degree of automation, standardization, and precision.

[0024] Specifically, firstly, a structured sparse transformation is performed on the global feature vector of the anatomical structure posture to obtain a global sparse feature vector of the anatomical structure posture. It should be understood that due to interference from fetal posture, scanning angle, and maternal tissue, even after preliminary processing by a spatial transformation network, the extracted global feature vector may still contain a large amount of redundant information or noise unrelated to anatomical alignment and growth assessment. This redundant information not only increases the difficulty of subsequent decoder calculation of affine transformation parameters but may also introduce biases, affecting the comparability of image features at different time points, thus hindering the construction of accurate temporal trajectories. Therefore, in the technical solution of this application, a structured sparse transformation is performed on the global feature vector of the anatomical structure posture to remove feature components that may represent image background, scanning artifacts, or unstable posture information, thereby generating a sparser and more compact candidate feature representation. This improves the information density and robustness of the features without losing key anatomical structure information, making them more focused on expressing the stable and essential anatomical posture characteristics of the fetus. This directly enhances the quality of subsequent multi-scale feature descriptor sequences and the reliability of temporal modeling, ultimately resulting in a more continuous and individualized characterization of fetal growth trajectory, as well as a more sensitive and accurate predictive ability of early growth abnormality risk, thereby improving the overall clinical practical value of the intelligent analysis system.

[0025] Specifically, the global feature vector of the anatomical structure pose is subjected to a structured sparse transformation using the following formula to obtain the global sparse feature vector of the anatomical structure pose: , in, This is the weight matrix. This represents the global feature vector of the anatomical structure's pose. Let be the bias vector, where This represents the global sparse feature vector of the anatomical structure's pose. The generation is parameterized by the Bernoulli distribution: ,in, These are learnable parameters; Next, the substantial gain of the global sparse feature vector of the anatomical structure pose relative to the global feature vector of the anatomical structure pose is calculated. It should be understood that simple sparse transformation is an exploratory operation; it may remove beneficial noise and redundancy, but it also carries the risk of accidentally deleting crucial anatomical pose information. Since the analysis of fetal ultrasound images is directly related to the accuracy of growth and development assessment, any feature-level optimization must have clear benefit justification and cannot rely on subjective assumptions or fixed rules. Therefore, in the technical solution of this application, a quantifiable mechanism associated with the final goal is used to verify its effectiveness. Specifically, by calculating the substantial gain, the system can accurately measure the performance change of the sparsified feature vector in driving the localization network decoder: whether it brings an improvement in alignment accuracy (positive gain) or a decrease in alignment quality due to the loss of key information (negative gain). This mechanism transforms the feature optimization process from blind to a data-driven, goal-oriented intelligent decision-making process.

[0026] Specifically, the feature sparse anatomical structure pose substantial gain relative to the global feature vector of the anatomical structure pose is calculated using the following formula: , in, Denotes KL divergence, For sparsity gain, The pose gain of the sparse anatomical structure is a feature-specific gain. Furthermore, in response to the substantial gain of the sparse anatomical structure pose satisfying a preset condition, the global sparse feature vector of the anatomical structure pose is set as the optimized global feature vector of the anatomical structure pose. Since fetal growth assessment is crucial for clinical decision-making, any feature optimization operation must be based on solid evidence of performance improvement, and cannot be chosen solely based on probability or experience. This conditional response mechanism is essentially an intelligent safety valve, ensuring that only feature transformations verified as effective are adopted, thereby avoiding the loss of critical anatomical information due to inappropriate feature sparsification, which could lead to subsequent pose alignment distortion. Therefore, in the technical solution of this application, a binary gating selection is performed through preset strict conditions (such as substantial gain satisfying a preset condition): the system will only adopt the optimized sparse feature vector when quantitative evaluation proves that the sparse transformation does indeed bring substantial performance improvement, i.e., it can help the localization network more accurately regress the affine transformation parameters, thereby improving the quality of anatomical structure alignment. Otherwise, the original feature vector is conservatively retained to ensure the robustness of the system. In this way, subsequent growth trajectory modeling and risk prediction can be based on more accurate and consistent anatomical structural representation, ultimately enabling the system to achieve truly meaningful early anomaly identification and individualized growth monitoring.

[0027] Specifically, in response to the sparse anatomical structure pose gain satisfying a preset condition, the global sparse feature vector of the anatomical structure pose is set as the global feature vector of the optimized anatomical structure pose using the following formula: , in, express variance For adjustable hyperparameters, This represents the global sparse feature vector of the anatomical structure's pose. This represents the global feature vector of the anatomical structure's pose. To optimize the global feature vector of anatomical structure pose.

[0028] Subsequently, the optimized anatomical structure pose global feature vector is input into the decoder of the localization network to obtain the affine transformation parameters. It should be understood that the optimized anatomical structure pose global feature vector is itself an intrinsic, high-dimensional feature representation. Although it contains all the key information about the pose, it cannot be directly applied to image pixels. Therefore, in the technical solution of this application, to achieve geometric correction of the image, the optimized anatomical structure pose global feature vector is input into the decoder of the localization network to decode it into a set of parameters with clear physical meaning and mathematical definition. It is worth mentioning that since the affine transformation is fully defined by 6 parameters in two-dimensional space, 6 independent output channels are required. Therefore, the output layer of the decoder is strictly designed to have 6 neurons, and typically no or no linear activation function is used. The output affine transformation parameters are a vector containing 6 real values. These 6 values ​​are arranged in a predetermined order and can be directly filled into a 2x3 matrix to form the affine transformation matrix, which is used for subsequent spatial transformation operations on the image.

[0029] Furthermore, a parameter-based differential image spatial transformation is performed on the first standardized ultrasound image based on the affine transformation parameters to obtain the first aligned ultrasound image. It should be understood that the affine transformation parameters output by the localization network are simply a set of values. To achieve image alignment, in this application's technical solution, a parameter-based differential image spatial transformation is performed on the first standardized ultrasound image based on the affine transformation parameters to perform corresponding spatial distortion and reconstruction of the original image. Specifically, differential image transformation means that the entire spatial transformation process is differentiable. This ensures that during the training of the neural network, the gradient generated by the loss function can smoothly pass through this transformation layer and propagate back to the localization network. This allows the localization network to learn how to generate better transformation parameters in an end-to-end manner to minimize the final task loss.

[0030] The specific execution process of this step mainly includes two closely linked sub-steps: sampling grid generation and differential sampling. First, the sampling grid is generated. This process calculates the corresponding coordinates in the source image (i.e., the normalized image) for each pixel coordinate of the target image (i.e., the aligned image) based on the affine transformation parameters (a vector containing 6 elements) output by the localization network. Specifically, the system first creates a regular coordinate grid with the exact same size as the target output image, where each point on the grid represents a pixel position in the output image. Then, using the transformation defined by the affine transformation parameters, each point on this regular grid is transformed to find its source coordinates in the source image space. By performing this operation on all pixel coordinates of the target image, a sampling grid is generated; next, differential sampling is performed from the source image using the generated sampling grid. Since the calculated source coordinates are usually floating-point numbers and cannot be directly mapped to a discrete pixel grid, an interpolation method is needed to estimate the pixel value of the non-integer coordinate point. To ensure the differentiability of the entire process, this application employs a bilinear interpolation mechanism. For each calculated source coordinate, bilinear interpolation finds its four nearest integer pixel coordinates. Then, based on the relative distances between the source coordinate and these four points, a weighted average of the pixel values ​​at these four points is calculated to obtain the interpolation result for that point. This sampling process is performed on every pixel of the output image to obtain the first aligned ultrasound image.

[0031] Finally, the first aligned ultrasound image is input into a pre-trained backbone convolutional neural network to obtain a high-dimensional feature map of the first aligned ultrasound image. It should be understood that although the first aligned ultrasound image has achieved uniformity in the macroscopic anatomical structure posture, it is essentially still a two-dimensional pixel matrix. To deeply analyze the fetal growth and development, it is necessary to extract features from the image that reflect deeper information such as anatomical structure morphology, texture, and tissue quality. Direct analysis at the pixel level is difficult and inefficient. Therefore, this application introduces a pre-trained backbone convolutional neural network as a feature extractor to effectively transform the complex visual patterns in the image into a structured, high-dimensional feature map, providing a more comprehensive and accurate decision-making basis for subsequent growth abnormality risk assessment.

[0032] In practice, the system first loads a pre-trained backbone convolutional neural network model, such as classic architectures like ResNet, VGGNet, or Inception trained on large-scale image datasets like ImageNet. Then, the first aligned ultrasound image generated in the previous step is used as input to the network. Since many pre-trained models default to accepting three-channel color images, while ultrasound images are typically single-channel grayscale images, the single-channel grayscale image is copied three times and stacked to form a three-channel image to match the network's input requirements. After the image is input into the network, it flows through a series of convolutional layers, activation function layers, pooling layers, and normalization layers from front to back, according to the network architecture design. In the shallow layers of the network, convolutional kernels are mainly used to capture basic features in the image, such as edges, corners, and textures. As the data passes through the network deeper, subsequent convolutional layers combine and abstract features extracted from the previous layer, thereby learning higher-level, more semantic features, such as the contours, shapes, and internal structures of specific organs. Specifically, this network does not perform the last few layers used for classification; instead, it truncates the forward propagation process at some intermediate point in the network. Typically, the output is extracted after the last convolutional block and before any global pooling layer or fully connected classification layer to obtain the first aligned high-dimensional ultrasound feature map.

[0033] Specifically, in step S3, feature descriptors based on multi-scale anatomical structures are generated from the aligned high-dimensional ultrasound feature map sequence to obtain an ultrasound feature descriptor sequence. It should be understood that although the aligned high-dimensional ultrasound feature map sequence contains rich information, the importance of different spatial locations and feature channels in the feature map is not equal; certain regions or features may contribute more to determining the growth status. Therefore, in the technical solution of this application, a more refined processing method is used to ensure that the final generated descriptor can retain the core information crucial for growth assessment to the greatest extent possible.

[0034] In practice, firstly, a first aligned high-dimensional ultrasound feature map is extracted from the aligned ultrasound high-dimensional feature map sequence. Next, the first aligned high-dimensional ultrasound feature map is input into an attention module to obtain a first weight mask. This attention module is typically a small convolutional neural network that receives the high-dimensional feature map as input and learns to output a weight mask, or attention map, with the same spatial dimension as the input feature map but only one channel. Each value on this mask is between 0 and 1, representing the importance of the corresponding spatial location in the input feature map. Through learning, this module can automatically identify the anatomical structure regions most important for the current analysis task. For example, when analyzing abdominal circumference, it might assign higher weights to the abdominal contour region and lower weights to the background region, thus achieving adaptive focusing on multi-scale anatomical structures. Then, the first weight mask is weighted with the first aligned high-dimensional ultrasound feature map to obtain a first weighted high-dimensional ultrasound feature map. Specifically, the first weight mask can be weighted with the first aligned high-dimensional ultrasound feature map element-wise to obtain the first weighted high-dimensional ultrasound feature map. This operation preserves or enhances the feature values ​​of important regions in the feature map, while weakening the feature values ​​of less important regions. Subsequently, global mean pooling and fully connected encoding are performed on the first weighted high-dimensional ultrasound feature map to obtain the first ultrasound feature descriptor. During this process, global mean pooling calculates the mean of all spatial locations in each channel of the weighted feature map, thus compressing a three-dimensional feature map into a one-dimensional vector. This vector is then fed into an encoder consisting of one or more fully connected layers for final nonlinear transformation and dimensionality adjustment, ultimately outputting a vector of a preset length, which is the ultrasound feature descriptor.

[0035] Specifically, in step S4, time-series trajectory modeling and risk prediction are performed on the ultrasound feature descriptor sequence to obtain a growth abnormality risk score. It should be understood that fetal growth and development is a continuous dynamic process, not a static snapshot of an isolated time point. Traditional analysis methods, which compare single measurements with standard curves, are essentially memoryless assessments and cannot effectively capture changes in growth rate, inflection points in growth patterns, and individualized developmental trends. Therefore, they are insufficient for early prediction of problems such as growth restriction or developmental delay. Therefore, in the technical solution of this application, a model capable of deeply understanding time-series data is constructed to perform longitudinal trajectory modeling on the entire ultrasound feature descriptor sequence. This learns the inherent laws and dynamic patterns in the fetal growth process, and based on this model, accurately predicts the likelihood of future growth abnormalities, ultimately outputting a quantitative risk score.

[0036] In practice, the ultrasound feature descriptor sequence is first input into a Transformer-based temporal encoder to obtain the ultrasound feature temporal context encoding vector. In this step, the sequence composed of ultrasound feature descriptors corresponding to different gestational weeks is fed into the model as a whole. The core of the Transformer architecture is its self-attention mechanism, which can simultaneously process all time points in the sequence and calculate the interdependencies and importance between any two time points. This means that when interpreting data from week 28, the model can dynamically and selectively associate and reference feature information from weeks 20 and 24, thereby accurately capturing key temporal patterns such as a change in growth rate from rapid to plateauing, or a continuous decline in growth percentile. After processing by the Transformer encoder, the information of the entire input sequence is condensed and encoded into a single, fixed-length high-dimensional vector, namely the ultrasound feature temporal context encoding vector. This vector not only contains information from all time points in the sequence but also implies the complex temporal dependencies and dynamic evolution patterns between these information points. Then, the ultrasonic feature temporal context encoding vector is input into a feedforward neural network to obtain the ultrasonic feature temporal context encoding vector. In this process, the feedforward neural network receives a context encoding vector containing information about the entire growth trajectory and maps it to a final prediction result through a series of nonlinear transformations. To ensure good interpretability of the output result, for example, by constraining it to a probability range of 0 to 1, the activation function of the last layer of the feedforward neural network is... Activation function. Through this function, the final output of the network is converted into a value between 0 and 1. This value is the growth abnormality risk score. The closer the value is to 1, the higher the predicted growth abnormality risk.

[0037] In summary, the intelligent analysis method for obstetric ultrasound examination data according to the embodiments of this application is explained. First, it addresses the issues of data inconsistency and feature incomparability caused by variations in fetal position and cross-section through standardization processing and pose-anatomical alignment based on a spatial transformation network. Then, it generates multi-scale feature descriptors containing rich anatomical structural information from the aligned high-dimensional feature map to replace sparse traditional biological measurements. Finally, it models the temporal trajectory of the sequence composed of these feature descriptors to deeply explore the dynamic process and individualized trends of fetal growth. In this way, it is possible to achieve accurate feature registration of ultrasound images at different gestational weeks, construct a personalized dynamic growth model reflecting changes in growth rate and pattern, and ultimately output a quantitative risk score for growth abnormalities, thereby significantly improving the ability to predict fetal growth abnormalities early and accurately.

[0038] Furthermore, an intelligent analysis system for obstetric ultrasound examination data is also provided.

[0039] Figure 3 This is a block diagram of an intelligent analysis system for obstetric ultrasound examination data according to an embodiment of this application. Figure 3 As shown, the obstetric ultrasound examination data intelligent analysis system 300 according to an embodiment of this application includes: a standardization module 310, used to standardize the original ultrasound data sequence to obtain a standardized ultrasound image sequence; a posture-anatomical alignment module 320, used to perform posture-anatomical alignment based on a spatial transformation network on the standardized ultrasound image sequence to obtain an aligned high-dimensional ultrasound feature map sequence; a feature descriptor generation module 330, used to generate feature descriptors based on multi-scale anatomical structures from the aligned high-dimensional ultrasound feature map sequence to obtain an ultrasound feature descriptor sequence; and a growth abnormality analysis module 340, used to perform temporal trajectory modeling and risk prediction on the ultrasound feature descriptor sequence to obtain a growth abnormality risk score.

[0040] As described above, the intelligent analysis system 300 for obstetric ultrasound examination data according to embodiments of this application can be implemented in various wireless terminals, such as servers with intelligent analysis algorithms for obstetric ultrasound examination data. In one possible implementation, the intelligent analysis system 300 for obstetric ultrasound examination data according to embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the intelligent analysis system 300 for obstetric ultrasound examination data can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the intelligent analysis system 300 for obstetric ultrasound examination data can also be one of many hardware modules of the wireless terminal.

[0041] Alternatively, in another example, the obstetric ultrasound examination data intelligent analysis system 300 and the wireless terminal can also be separate devices, and the obstetric ultrasound examination data intelligent analysis system 300 can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0042] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. An intelligent analysis method of obstetric ultrasound examination data, characterized in that, The method comprises the following steps: standardizing a raw ultrasound data sequence to obtain a standardized ultrasound image sequence; performing spatial transformation network-based pose-anatomy alignment on the standardized ultrasound image sequence to obtain an aligned ultrasound high-dimensional feature map sequence; generating a multi-scale anatomy-based feature descriptor based on the aligned ultrasound high-dimensional feature map sequence to obtain an ultrasound feature descriptor sequence; performing time series trajectory modeling and risk prediction on the ultrasound feature descriptor sequence to obtain a growth abnormality risk score.

2. The method of intelligent analysis of obstetric ultrasound examination data according to claim 1, characterized in that, The method for standardizing a raw ultrasound data sequence to obtain a standardized ultrasound image sequence comprises the following steps: extracting a candidate image frame pool from the raw ultrasound data sequence; performing standard section recognition and optimization based on a hierarchical classification model on the candidate image frame pool to obtain an optimal section mapping table; based on the optimal section mapping table, performing image standardization on each optimal image frame object in the candidate image frame pool to obtain the standardized ultrasound image sequence.

3. The method of intelligent analysis of obstetric ultrasound examination data according to claim 1, characterized in that, Based on the optimal section mapping table, performing image standardization on each optimal image frame object in the candidate image frame pool to obtain the standardized ultrasound image sequence comprises: performing grayscale conversion, size normalization and intensity normalization on each optimal image frame object to obtain the standardized ultrasound image sequence.

4. The method of intelligent analysis of obstetric ultrasound examination data according to claim 1, characterized in that, The method for performing spatial transformation network-based pose-anatomy alignment on the standardized ultrasound image sequence to obtain an aligned ultrasound high-dimensional feature map sequence comprises the following steps: extracting a first standardized ultrasound image from the standardized ultrasound image sequence; inputting the first standardized ultrasound image into a positioning network dedicated to parameter regression to obtain affine transformation parameters; performing parameter-based differential imaging spatial transformation on the first standardized ultrasound image based on the affine transformation parameters to obtain a first aligned ultrasound image; inputting the first aligned ultrasound image into a pre-trained backbone convolutional neural network to obtain a first aligned ultrasound high-dimensional feature map.

5. The method of intelligent analysis of obstetric ultrasound examination data according to claim 4, characterized in that, The method for inputting the first standardized ultrasound image into a positioning network dedicated to parameter regression to obtain affine transformation parameters comprises the following steps: inputting the first standardized ultrasound image into an encoder of the positioning network to obtain an anatomical structure pose global feature vector; performing feature-level fine-tuning on the anatomical structure pose global feature vector to obtain an optimized anatomical structure pose global feature vector; inputting the optimized anatomical structure pose global feature vector into a decoder of the positioning network to obtain the affine transformation parameters.

6. The method of intelligent analysis of obstetric ultrasound examination data according to claim 1, characterized in that, The method for generating a multi-scale anatomy-based feature descriptor based on the aligned ultrasound high-dimensional feature map sequence to obtain an ultrasound feature descriptor sequence comprises the following steps: extracting a first aligned ultrasound high-dimensional feature map from the aligned ultrasound high-dimensional feature map sequence; inputting the first aligned ultrasound high-dimensional feature map into an attention module to obtain a first weight mask; performing weighting on the first weight mask and the first aligned ultrasound high-dimensional feature map to obtain a first weighted ultrasound high-dimensional feature map; performing global mean pooling and fully connected encoding on the first weighted ultrasound high-dimensional feature map to obtain a first ultrasound feature descriptor.

7. The method of intelligent analysis of obstetric ultrasound examination data according to claim 1, characterized in that, The method for performing time series trajectory modeling and risk prediction on the ultrasound feature descriptor sequence to obtain a growth abnormality risk score comprises the following steps: inputting the ultrasound feature descriptor sequence into a Transformer-based time series encoder to obtain an ultrasound feature time series context encoding vector; inputting the ultrasound feature temporal context encoding vector into a feedforward neural network to obtain the ultrasound feature temporal context encoding vector.

8. The method of intelligent analysis of obstetric ultrasound examination data according to claim 7, characterized in that, The activation function of the last layer of the feedforward neural network is activation function.

9. The method of intelligent analysis of obstetric ultrasound examination data according to claim 5, characterized in that, performing feature-level fine-tuning on the anatomical pose global feature vector to obtain an optimized anatomical pose global feature vector, comprising: performing structured sparse transformation on the anatomical pose global feature vector to obtain an anatomical pose global sparse feature vector; calculating a feature sparse anatomical pose substantial gain of the anatomical pose global sparse feature vector relative to the anatomical pose global feature vector; in response to the feature sparse anatomical pose substantial gain satisfying a preset condition, setting the anatomical pose global sparse feature vector as the optimized anatomical pose global feature vector.

10. An obstetric ultrasound examination data intelligent analysis system, characterized in that, comprising: a normalization module configured to normalize the original ultrasound data sequence to obtain a normalized ultrasound image sequence; a pose-anatomy alignment module configured to perform pose-anatomy alignment on the normalized ultrasound image sequence based on a spatial transformation network to obtain an aligned ultrasound high-dimensional feature map sequence; a feature descriptor generation module configured to perform multi-scale anatomical structure-based feature descriptor generation on the aligned ultrasound high-dimensional feature map sequence to obtain an ultrasound feature descriptor sequence; a growth abnormality analysis module configured to perform temporal trajectory modeling and risk prediction on the ultrasound feature descriptor sequence to obtain a growth abnormality risk score.

Images