Accurate identification method for placenta position in abdominal antenatal examination image

Abstract

The invention relates to the technical field of fetal position image recognition, and discloses an abdominal antenatal examination image placenta position accurate recognition method, which comprises the following steps: constructing a sound velocity inversion growth dimension, extracting echo time sequence and displacement field data from a continuous frame video stream, generating an individualized sound velocity distribution field through sound ray tracing and parallax inversion, and realizing pixel-by-pixel geometric correction. The distance measurement between the lower edge of the placenta and the inner opening of the cervix is made to return to a real anatomical reference, the adhesion state is evaluated from breathing sliding displacement by constructing a mechanical fingerprint growth dimension, inverting the elastic modulus from probe pressurization strain response, and bimodal mutual verification is performed through a sound velocity-elastic modulus consistency index, so that single-modal multiplicity is eliminated. And finally, superposing the mechanical characteristic field to the corrected image, identifying an abnormal area, generating placenta position classification and implanting risk level identification, and realizing full-process automation of image correction, risk assessment and report generation.

Description

Methods for accurate identification of placental location in abdominal prenatal examination images Technical Field

[0001] This invention relates to the field of fetal position image recognition technology, and in particular to a method for accurate identification of placental position in abdominal prenatal examination images. Background Technology

[0002] Placental location identification refers to determining the attachment site of the placenta within the uterus and its spatial relationship with the internal cervical os using medical imaging technology. It is one of the core components of obstetric ultrasound examination. In a normal pregnancy, the placenta can attach to the anterior wall, posterior wall, fundus, or lateral wall of the uterus. When the placenta is abnormally attached to the lower segment of the uterus, or even covers the internal cervical os, it forms placenta previa—a significant cause of late pregnancy bleeding and premature birth, and closely related to perinatal maternal and fetal complications and mortality. According to the Chinese Medical Association guidelines, placenta previa can be classified into low-lying placenta (placental edge <20mm from the internal cervical os), marginal placenta previa, partial placenta previa, and complete placenta previa. Accurate identification of placental location is not only crucial for the diagnostic classification of placenta previa but also directly guides the choice of delivery method: for complete placenta previa, cesarean section is the recommended delivery method; while for low-lying placenta, vaginal delivery can be attempted, but the risk of postpartum hemorrhage must be carefully monitored. In addition, placental location is closely related to the occurrence of placenta accreta, especially for pregnant women with a history of cesarean section. Accurate assessment of the placental-uterine interface is a key basis for predicting serious complications such as intraoperative bleeding and bladder injury.

[0003] Currently, ultrasound examination has become the preferred imaging method for placental localization due to its advantages such as being non-invasive, real-time, and repeatable. The routine clinical procedure involves transabdominal ultrasound scanning, where doctors diagnose and classify the placenta by observing its echogenicity and measuring the distance between the lower edge of the placenta and the internal cervical os. With the development of deep learning technology, automated placental identification methods have made significant progress. For example, patent application CN120852284A discloses an MRI-based placental location detection system and method. This system uses a U-Net three-dimensional segmentation network incorporating an attention mechanism to segment MRI images, extracting three-dimensional models of the placenta, uterus, and cervix, and assessing the risk of placental implantation based on placental vascular density and the depth of myometrial invasion. The invention patent with announcement number CN120747113B proposes an automatic placental location identification method based on point cues. It adopts a two-stage training architecture: first, a medical image encoder is pre-trained through self-supervised learning, and then point cues from doctors are introduced to fine-tune the model. Finally, it achieves automatic classification of normal placenta, low-lying placenta, marginal placenta previa, partial placenta previa, and complete placenta previa in abdominal ultrasound images.

[0004] The aforementioned and existing related technologies often have the following drawbacks: First, conventional abdominal ultrasound examinations assume that sound waves propagate at a constant speed in tissues (1540 m / s). However, the actual sound speeds differ significantly between the abdominal wall fat layer, muscle layer, and uterine myometrium, leading to systematic geometric distortions in placental images—depth distortion and morphological distortion. This causes the measurement of the distance between the lower edge of the placenta and the internal cervical os based on the image to lose its true physical reference, directly affecting the accuracy of placenta previa classification. The core of this problem lies in the fact that traditional imaging models cannot adapt to individualized tissue sound speed distributions, resulting in a systematic deviation between the spatial position of the image and the actual anatomical structure.

[0005] Secondly, existing risk assessment methods rely solely on static morphological features (such as placental thickness and infiltration depth), completely ignoring the rich dynamic information sources during the examination process—the elastic deformation of tissues under probe pressure and the relative sliding of the placenta-uterus interface during maternal respiration. These "dynamic mechanical fingerprints," which could reflect the rigidity and adhesion of tissues at the placental implantation site, have not been effectively explored and utilized. As a result, the assessment of implantation risk remains at the morphological level, lacking a quantitative perception of the tissue's functional state. The essence of this problem is that single-modal static features cannot characterize the mechanical response characteristics of tissues under dynamic stress, and there is a lack of physical constraints and mutual verification mechanisms between multimodal features. Summary of the Invention

[0006] The technical problem to be solved by this invention is that the existing technology has the disadvantages of lacking individualized sound velocity correction and not integrating dynamic mechanical information. To address this, we propose a method for accurate identification of placental location in abdominal prenatal examination images.

[0007] To achieve the above objectives, this application adopts the following technical solution: a method for accurate placental location identification in abdominal prenatal examination images, comprising constructing a dual-modal growth network for placental location identification, determining the sound velocity inversion growth dimension and the mechanical fingerprint growth dimension of the dual-modal growth network, and setting initial feature nodes for each growth dimension; extracting echo timing features of the sound beam passing through different tissue layers and displacement field data of tissue feature points between adjacent frames from the continuous frame ultrasound video stream of abdominal scans, generating feature sub-nodes corresponding to each growth dimension, and constructing physical quantity inversion rules for each feature sub-node; and extracting tissue strain response data under probe pressure and maternal respiration data from the continuous frame ultrasound video stream. The motion-induced slippage displacement data of the placenta-uterus interface is used to construct the correlation and coupling rules between the response data and displacement data and the feature sub-nodes of each growth dimension. This is used to establish the physical constraint relationship between the sound velocity inversion and mechanical fingerprint dual modes. Based on the correlation and coupling rules, the physical quantity inversion rule parameters of the feature sub-nodes are iteratively optimized. The ultrasound video stream to be detected is input into the optimized dual-modal growth network to generate an individualized tissue sound velocity distribution field and a placenta-uterus interface mechanical feature field. Based on the sound velocity distribution field, the original ultrasound image is corrected pixel by pixel for geometric distortion. Based on the mechanical feature field, abnormal tissue areas are marked on the corrected image. The results are then fused to generate a precise identification result of the placental location.

[0008] Furthermore, the generation of feature sub-nodes corresponding to each growth dimension, and the construction of physical quantity inversion rules for each feature sub-node, include: for the sound velocity inversion growth dimension, extracting echo arrival time series data of the sound beam passing through each tissue layer from the continuous frame ultrasound video stream, generating sound ray refraction offset data based on the difference in echo arrival time between adjacent angles, generating a first-level feature sub-node based on the refraction offset data, and constructing sound ray tracing inversion rules for inferring the local sound velocity values ​​of each tissue layer; extracting displacement field data of tissue feature points between adjacent frames from the continuous frame ultrasound video stream, generating a second-level feature sub-node based on the displacement field data, and constructing rules for calculating based on parallax displacement. The parallax inversion rule for equivalent sound velocity; for the mechanical fingerprint growth dimension, image sequences of the probe pressurization and depressurization phases are extracted from continuous frame ultrasound video streams, tissue strain response data are generated based on tissue compression deformation before and after pressurization, and first-level feature sub-nodes are generated based on the tissue strain response data to construct an elastic inversion rule for back-calculating the local tissue elastic modulus value; image sequences within the maternal respiratory cycle are extracted from continuous frame ultrasound video streams, and the relative sliding trajectory of the placenta-uterine myometrium interface is tracked to generate sliding displacement data, and second-level feature sub-nodes are generated based on the sliding displacement data to construct an adhesion assessment rule for judging the adhesion status of the placenta-uterus interface.

[0009] Furthermore, the construction of the association and coupling rules between the response data and displacement data and the feature sub-nodes of each growth dimension includes: setting the acquisition trigger conditions for tissue strain response data; when the probe pressure value exceeds a preset pressure threshold, capturing the image sequence of the pressurization stage and extracting the compression deformation time series data of each tissue layer; setting the acquisition period for sliding displacement data; capturing the image sequence within the respiratory cycle according to the mother's respiratory frequency and tracking the displacement trajectory of interface feature points; constructing association and coupling rules between tissue strain response data and feature sub-nodes for the feature sub-nodes of the growth dimension of sound velocity inversion, specifying the relationship mapping between compression deformation and sound velocity changes corresponding to different tissue layer types, used to transform mechanical response features into constraints for sound velocity inversion; constructing association and coupling rules between sliding displacement data and feature sub-nodes for the feature sub-nodes of the mechanical fingerprint growth dimension, specifying the quantitative relationship between sliding displacement features and adhesion degree corresponding to different tissue strain rate intervals, used to transform sliding displacement data into verification basis for elastic inversion; establishing a mapping relationship table between the association and coupling rules and the feature sub-nodes of each growth dimension, and storing it in the collaborative optimization link configuration module of the dual-modal growth network.

[0010] Further, the physical quantity inversion rule parameters of the feature sub-nodes are iteratively optimized based on the aforementioned correlation coupling rules, including: receiving tissue strain response data and sliding displacement data through a collaborative optimization link; matching the corresponding feature sub-nodes and correlation coupling rules based on the mapping relationship table; for feature sub-nodes in the sound velocity inversion growth dimension, extracting the mechanical-sound velocity relationship mapping specified in the correlation coupling rules; inputting the compression deformation parameters in the tissue strain response data into the mapping to generate sound velocity inversion correction coefficients; and adjusting the acoustic ray tracing inversion rule parameters or disparity inversion rule parameters of the feature sub-node based on the correction coefficients; for feature sub-nodes in the mechanical fingerprint growth dimension, extracting the sliding-adhesion quantization relationship specified in the correlation coupling rules; and inputting the sliding amplitude parameters in the sliding displacement data into the physical quantity inversion rule table. The system generates elastic inversion verification coefficients based on the relationships between the coefficients and the elastic modulus values ​​output by the feature sub-nodes. If the comparison deviation exceeds a preset threshold, the elastic inversion rule parameters are adjusted. The correction frequency of each feature sub-node is counted. For feature sub-nodes whose correction frequency reaches a preset frequency threshold, the number of their associated sub-nodes is increased to expand the inversion dimension. The coupling relationships between feature sub-nodes are analyzed, and the sound velocity-elastic modulus consistency index corresponding to each pair of coupling relationships is calculated. For coupling relationships with consistency indices lower than a preset threshold, the connection weights of the feature sub-nodes are adjusted or intermediate coupling nodes are added to optimize the coupling logic path. The adjusted feature sub-node parameters, the expanded number of sub-nodes, and the optimized coupling relationships are integrated to generate an iteratively optimized placental location accuracy identification network.

[0011] Further, the step of extracting image sequences of the probe pressurization and depressurization phases from a continuous frame ultrasound video stream includes: identifying a complete pressure change cycle from the baseline pressure value to the peak pressure value and back to the baseline pressure value from the original ultrasound video stream; extracting the image sequence of the pressure rise phase within each cycle as the pressurization phase image sequence; and extracting the image sequence of the pressure fall phase as the depressurization phase image sequence; performing inter-frame registration processing on the pressurization phase image sequence and the depressurization phase image sequence respectively; and extracting the longitudinal images of each tissue layer from the registered pressurization phase image sequence. The compression displacement data is used to extract elastic recovery displacement data of each tissue layer from the registered decompression phase image sequence; the extraction of the image sequence within the maternal respiratory cycle from the continuous frame ultrasound video stream includes: identifying the start and end time points of the periodic undulation movement of the abdominal wall from the original ultrasound video stream, and extracting the complete image sequence from the inspiratory phase to the expiratory phase within each respiratory cycle as the respiratory cycle image sequence; performing inter-frame registration processing on the respiratory cycle image sequence, and tracing the displacement trajectories of characteristic points at the junction of the lower edge of the placenta and the internal cervical os, and at the junction of the posterior part of the placenta and the myometrium from the registered image sequence.

[0012] Further, for the feature sub-nodes of the sound velocity inversion growth dimension, the mechanical-sound velocity relationship mapping specified in the correlation coupling rules is extracted, and the compression deformation parameters in the tissue strain response data are input into the mapping to generate sound velocity inversion correction coefficients. This includes: extracting the compression deformation parameters of the tissue layer governed by the corresponding feature sub-node from the tissue strain response data; if it is the fat layer, the thickness compression percentage is extracted; if it is the myometrium, the shear wave propagation velocity change is extracted; if it is the placental tissue, the overall deformation displacement is extracted; the mechanical-sound velocity relationship mapping function stored in the correlation coupling rules of the feature sub-node is retrieved, and the mapping function is constructed based on experimental fitting curves, theoretical correlation models, or numerical simulation results of different tissue types; the compression deformation parameters are input into the mapping function to calculate the sound velocity correction coefficients of the corresponding tissue layer; the sound velocity correction coefficients corresponding to each tissue layer are stored according to the tissue layer type to form a sound velocity inversion correction coefficient vector for the feature sub-nodes of the sound velocity inversion growth dimension to call.

[0013] Further, the analysis of the coupling relationships between feature sub-nodes and the calculation of the sound velocity-elastic modulus consistency index corresponding to each pair of coupling relationships include: traversing all pairs of feature sub-nodes with coupling relationships in the dual-modal growth network, each pair containing a feature sub-node of the sound velocity inversion growth dimension and a feature sub-node of the mechanical fingerprint growth dimension, with the two sub-nodes governing the same spatial region's organizational layer; extracting the sound velocity distribution data of the organizational region from the inversion results of the feature sub-node of the sound velocity inversion growth dimension, and extracting the elastic modulus distribution data of the organizational region from the inversion results of the feature sub-node of the mechanical fingerprint growth dimension; and dividing the sound velocity distribution data into... The sound velocity distribution data and the elastic modulus distribution data are paired at the same coordinate points, and the correlation coefficient between the sound velocity value and the elastic modulus value is calculated. If the correlation coefficient is lower than a preset threshold, the consistency index is determined to be low consistency. If the correlation coefficient is higher than or equal to the preset threshold, the local difference map between the sound velocity distribution and the elastic modulus distribution is further calculated, and the number of pixels with difference values ​​exceeding the preset threshold is counted. The consistency index value is determined based on the ratio of this number to the total number of pixels in the tissue region. The coupling association identifier and consistency index value of each pair of feature sub-nodes are organized into a coupling association consistency statistics table and stored in the collaborative optimization link configuration module of the dual-modal growth network.

[0014] Further, the step of performing pixel-by-pixel geometric distortion correction on the original ultrasound image based on the sound velocity distribution field, identifying abnormal tissue regions on the corrected image based on the mechanical feature field, and fusing to generate accurate placental location identification results includes: extracting individualized tissue sound velocity distribution fields; for each pixel in the original ultrasound image, calculating the actual propagation time of the sound beam based on the local sound velocity value sequence of all tissue layers in the depth direction of that pixel, calculating the standard propagation time based on the standard sound velocity value, calculating the ratio of the two and correcting the depth coordinates of that pixel, generating a corrected coordinate mapping table for each pixel, and resampling and interpolating the original ultrasound image based on the coordinate mapping table to generate a corrected image of the true anatomical structure; The mechanical characteristic field of the placenta-uterus interface is captured, and the elastic modulus distribution data and interface sliding displacement distribution data are superimposed on the corrected real anatomical structure image to identify abnormal areas where the elastic modulus value exceeds the normal range and suspected adhesion areas where the sliding amplitude is below the normal threshold. Based on the spatial location of the abnormal areas and suspected adhesion areas, the minimum distance between the lower edge of the placenta and the internal cervical os in the corrected image is measured. The positional relationship between the abnormal areas and the internal cervical os is combined to generate placental location classification results and implantation risk level labels. The placental location classification results, implantation risk level labels, and abnormal area labels are superimposed on the corrected real anatomical structure image to generate a fused image of accurate placental location identification.

[0015] A system for accurate placental location identification in abdominal prenatal examination images, used to implement a method for accurate placental location identification in abdominal prenatal examination images, includes: a probe for emitting ultrasound waves into the pregnant woman's abdomen and receiving ultrasound echo signals; a transmission and reception circuit connected to the probe for controlling the transmission and reception of the probe; a processor connected to the transmission and reception circuit, configured to execute the method for accurate placental location identification in abdominal prenatal examination images; a memory connected to the processor for storing the continuous frame ultrasound video stream, the network parameters of the dual-modal growth network, the individualized tissue sound velocity distribution field, and the placenta-uterus interface mechanical characteristic field; and a display device connected to the processor for displaying the fused image of the accurate placental location identification result.

[0016] A computer program product, characterized in that the computer program product includes machine-executable instructions stored in a computer-readable storage medium, a processor of an abdominal prenatal examination image placental location accurate recognition system reads the machine-executable instructions from the computer-readable storage medium, and the processor executes the machine-executable instructions, causing the abdominal prenatal examination image placental location accurate recognition system to perform an abdominal prenatal examination image placental location accurate recognition method.

[0017] The technical effects and advantages of this invention are as follows: Traditional ultrasound examinations assume that sound waves propagate at a constant speed in tissues. However, the actual sound speeds differ significantly between the abdominal fat layer, muscle layer, and uterine myometrium of pregnant women, leading to depth distortion and morphological distortion in placental images. This causes the measurement of the distance between the lower edge of the placenta and the internal cervical os, based on the image, to lose its physical reference, directly affecting the accuracy of placenta previa classification. This solution constructs a sound speed inversion growth dimension, extracts echo temporal features and displacement field data from continuous frame ultrasound video streams, and generates an individualized tissue sound speed distribution field using ray tracing inversion rules and parallax inversion rules. Pixel-by-pixel geometric distortion correction is then applied to the original ultrasound image, restoring the morphology of deep tissues such as the posterior placenta and the acoustic shadowing zone to their true anatomical structure. The accuracy of the measurement of the distance between the lower edge of the placenta and the internal cervical os is improved to the sub-millimeter level, providing a reliable data foundation for accurate classification of placenta previa.

[0018] In this invention, existing methods rely solely on static morphological features (such as placental thickness and invasion depth), completely ignoring the rich dynamic information sources during the examination process. This approach constructs a mechanical fingerprint growth dimension, inverts the elastic modulus from the tissue strain response under probe pressure, and assesses the adhesion state from the interface sliding displacement of maternal respiratory motion, establishing a quantitative correlation between tissue mechanical features and pathological states. More importantly, this approach constructs a dual-modal correlation coupling rule between sound velocity inversion and mechanical fingerprinting. The inversion results are cross-validated by calculating a sound velocity-elastic modulus consistency index. When the consistency index falls below a preset threshold, the coupling logic path is automatically adjusted, ensuring that the final generated sound velocity distribution field and mechanical feature field have high intrinsic self-consistency. This significantly improves the accuracy and reliability of implantation risk assessment, making pathological changes that are difficult to identify using traditional methods, such as micro-implantation lesions and early adhesions, explicitly apparent.

[0019] In this invention, elastic modulus distribution data and interface sliding displacement distribution data are superimposed onto a corrected image of the actual anatomical structure to visually identify areas of abnormal elasticity and suspected adhesion. Based on this, the minimum distance between the lower edge of the placenta and the internal cervical os in the corrected image is measured to generate a placental location classification result and an implantation risk level label. Furthermore, the overall risk is quantitatively assessed by calculating a comprehensive placental implantation risk index P, and the correction depth is adaptively adjusted using a confidence factor Q. This ensures that the final fused image not only contains accurate anatomical location information but also embeds risk quantification indicators based on tissue mechanics, providing clinicians with "what you see is what you get" decision support. This method integrates the previously separate image correction, risk assessment, and report generation into a fully automated closed loop, requiring no additional hardware or changes to doctors' operating habits. It is particularly suitable for resource-limited primary hospitals, possessing strong clinical application value and broad prospects for promotion and application. Attached Figure Description

[0020] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals refer to the same components: Figure 1 is a main flowchart of the construction and iterative optimization of the dual-modal growth network of this invention; Figure 2 is a sub-flowchart of the iterative optimization based on the correlation coupling rule of this invention; Figure 3 is a flowchart of the generation of accurate placental location identification results of this invention; Figure 4 is a system flowchart of this invention. Detailed Implementation

[0021] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0022] This invention provides a technical solution: a method for accurate identification of placental location in abdominal prenatal examination images, comprising the following steps: Referring to Figure 1, step S110: constructing a dual-modal growth network for placental location identification, determining the sound velocity inversion growth dimension and the mechanical fingerprint growth dimension of the dual-modal growth network, and setting initial feature nodes for each growth dimension.

[0023] This embodiment uses an obstetric ultrasound examination scenario in a tertiary hospital as an example. Ms. Zhang, a 28-year-old pregnant woman at 28 weeks gestation, is clinically suspected of having placenta previa and requires an abdominal ultrasound examination to determine the placental location and assess the risk of implantation. First, a bimodal growth network for placental location identification needs to be constructed. This network, through its bimodal parallel growth architecture, aims to address the geometric distortion caused by uneven sound velocity in traditional ultrasound imaging and the inability of static morphological features to characterize tissue functional states. The network comprises two core growth dimensions: a sound velocity inversion growth dimension and a mechanical fingerprint growth dimension. For each growth dimension, corresponding initial feature nodes are set, carrying the basic attribute information for that dimension. For example, the initial feature nodes of the sound velocity inversion growth dimension carry basic attribute information including the basic assumptions of sound velocity inversion and tissue layer classification standards; the initial feature nodes of the mechanical fingerprint growth dimension carry basic attribute information including the reference range of elastic modulus and the normal threshold of interface sliding displacement.

[0024] Step S120: Extract the echo time-series features of the sound beam passing through different tissue layers and the displacement field data of tissue feature points between adjacent frames from the continuous frame ultrasound video stream of the abdominal scan. Generate feature sub-nodes corresponding to each growth dimension and construct physical quantity inversion rules for each feature sub-node.

[0025] In this embodiment, the doctor used an ultrasound probe with pressure sensing function to scan the abdomen of pregnant woman Zhang, acquiring a continuous frame ultrasound video stream. This video stream contains a multi-angle ultrasound echo sequence acquired as the probe oscillates along a standard scanning path. Specifically, the time-series data of echo arrivals through different tissue layers, such as the abdominal wall fat layer, muscle layer, uterine myometrium, and placental tissue, are extracted from this video stream, along with displacement field data of tissue feature points (such as fascial interfaces, blood vessel walls, and placental chorionic plate) between adjacent frames. Based on this data, feature sub-nodes are generated corresponding to each growth dimension, and physical quantity inversion rules are constructed for each sub-node.

[0026] Step S121: For the sound velocity inversion growth dimension, extract the echo arrival time series data of the sound beam passing through each tissue layer from the continuous frame ultrasound video stream, generate sound ray refraction offset data based on the difference in arrival time of adjacent angle echoes, generate first-level feature sub-nodes based on the refraction offset data, and construct sound ray tracing inversion rules for back-inferring the local sound velocity values ​​of each tissue layer; extract the displacement field data of tissue feature points between adjacent frames from the continuous frame ultrasound video stream, generate second-level feature sub-nodes based on the displacement field data, and construct disparity inversion rules for calculating the equivalent sound velocity based on disparity displacement.

[0027] Specifically, echo arrival time series data of the sound beam traversing each tissue layer are extracted from continuous frame ultrasound video streams. For example, for the abdominal wall fat layer, the time it takes for the sound beam to travel from the probe surface to the lower interface of the fat layer and back is recorded. By comparing the echo arrival time differences at adjacent angles (e.g., probe swing of 1 degree), the refraction offset of the sound ray at each tissue layer interface is calculated. Based on these refraction offset data, a first-level feature sub-node for the sound velocity inversion growth dimension is generated. This sub-node corresponds to different tissue layer types (e.g., fat layer, muscle layer, uterine myometrium, placental tissue), and a sound ray tracing inversion rule is constructed for each tissue layer. This rule uses the refraction offset to infer the local sound velocity value of that tissue layer.

[0028] Simultaneously, displacement field data of tissue feature points between adjacent frames are extracted from continuous frame ultrasound video streams. For example, the pixel displacement of feature points on the placental villi between two adjacent frames is tracked. Based on these displacement field data, a second-level feature sub-node for the sound velocity inversion growth dimension is generated. This sub-node corresponds to different feature point displacement patterns, and a disparity inversion rule is constructed for each pattern. This rule calculates the equivalent sound velocity of the sound beam traversing the path based on the disparity displacement of feature points between frames at different angles.

[0029] Step S122: For the mechanical fingerprint growth dimension, extract image sequences of the probe pressurization and depressurization phases from the continuous frame ultrasound video stream, generate tissue strain response data based on the tissue compression deformation before and after pressurization, generate first-level feature sub-nodes based on the tissue strain response data, and construct elastic inversion rules for back-calculating the local tissue elastic modulus value; extract image sequences within the maternal respiratory cycle from the continuous frame ultrasound video stream, track the relative sliding trajectory of the placenta and uterine wall interface to generate sliding displacement data, generate second-level feature sub-nodes based on the sliding displacement data, and construct adhesion assessment rules for judging the adhesion status of the placenta-uterus interface.

[0030] In this embodiment, the ultrasound probe has a pressure sensing function, enabling real-time recording of the contact pressure between the probe and the abdominal wall. During the scan, the doctor naturally applies and releases pressure, forming a pressure change cycle. The complete pressure change cycle, from baseline to peak and back to baseline, is identified from the continuous frame ultrasound video stream. The image sequence of the pressure increase phase is extracted as the pressurization phase image sequence, and the image sequence of the pressure decrease phase is extracted as the decompression phase image sequence. These two sets of image sequences are inter-frame registered to eliminate the influence of minor probe movements. Then, longitudinal compression displacement data of each tissue layer is extracted from the registered pressurization phase image sequence, and elastic recovery displacement data of each tissue layer is extracted from the registered decompression phase image sequence. These compression deformation data are the tissue strain response data. Based on this data, a first-level feature sub-node of the mechanical fingerprint growth dimension is generated. This sub-node corresponds to different tissue strain rate intervals, and an elastic inversion rule is constructed for each interval to infer the elastic modulus value of the local tissue.

[0031] Simultaneously, image sequences within the maternal respiratory cycle are extracted from continuous frame ultrasound video streams. By identifying the start and end points of the periodic undulations of the abdominal wall, complete image sequences from the inspiratory to expiratory phases within each respiratory cycle are extracted as respiratory cycle image sequences. Inter-frame registration is performed on these image sequences, and the displacement trajectories of characteristic points at the junction of the lower edge of the placenta and the internal cervical os, and at the junction of the posterior placenta and the myometrium, are tracked to generate sliding displacement data. Based on this data, a second-level feature sub-node corresponding to the mechanical fingerprint growth dimension is generated. This sub-node corresponds to different sliding displacement patterns, and adhesion assessment rules are constructed for each pattern to determine the adhesion status of the placenta-uterus interface. For example, the adhesion assessment rule can be set as follows: if the sliding displacement amplitude is lower than a normal threshold (e.g., 2 mm), it is determined to be a suspected adhesion area.

[0032] Step S130: Extract tissue strain response data under probe pressure and placental-uterine interface sliding displacement data caused by maternal respiratory motion from the continuous frame ultrasound video stream, construct the correlation and coupling rules between the response data and displacement data and feature sub-nodes of each growth dimension, and iteratively optimize the physical quantity inversion rule parameters of the feature sub-nodes based on the correlation and coupling rules.

[0033] In this embodiment, tissue strain response data under probe pressure (such as time-series data of compression deformation of each tissue layer) and placental-uterine interface sliding displacement data caused by maternal respiratory motion (such as displacement trajectory of interface feature points) are further extracted from the continuous frame ultrasound video stream. Then, the association coupling rules between these dynamic response data and feature sub-nodes of each growth dimension are constructed.

[0034] Step S131: Set the acquisition trigger conditions for tissue strain response data. When the probe pressure value exceeds the preset pressure threshold, capture the image sequence of the pressurization stage and extract the compression deformation time series data of each tissue layer. Set the acquisition period for sliding displacement data. Capture the image sequence within the respiratory cycle according to the mother's respiratory rate and track the displacement trajectory of interface feature points.

[0035] Specifically, the acquisition trigger conditions for tissue strain response data are set: when the probe pressure value exceeds a preset pressure threshold (e.g., 5N higher than the baseline pressure value), acquisition is automatically triggered, capturing image sequences during the pressurization phase, and extracting compression deformation time-series data for each tissue layer, including maximum compression, compression recovery time, and compression hysteresis. The acquisition cycle for sliding displacement data is set: according to the maternal respiratory rate (approximately 12-20 breaths per minute), image sequences within each respiratory cycle are automatically captured, tracking the displacement trajectory of interface feature points, and extracting data such as sliding amplitude, consistency of sliding direction, and the phase difference between the sliding start time and the respiratory cycle.

[0036] Step S132: For the feature sub-nodes of the sound velocity inversion growth dimension, construct the association and coupling rules between tissue strain response data and feature sub-nodes, and specify the relationship mapping between compression deformation and sound velocity change corresponding to different tissue layer types, which is used to transform mechanical response features into constraints for sound velocity inversion; For the feature sub-nodes of the mechanical fingerprint growth dimension, construct the association and coupling rules between sliding displacement data and feature sub-nodes, and specify the quantitative relationship between sliding displacement features and adhesion degree corresponding to different tissue strain rate ranges, which is used to transform sliding displacement data into verification basis for elastic inversion.

[0037] For the feature sub-nodes in the growth dimension of sound velocity inversion, association coupling rules are constructed between tissue strain response data and sub-nodes. For example, for the fat layer, a mapping relationship is defined between its compression deformation (thickness compression percentage) and sound velocity change. This mapping is constructed based on the experimental fitting curve of fat tissue compression rate and sound velocity change rate. In this way, mechanical response characteristics can be transformed into constraints for sound velocity inversion, i.e., the sound velocity inversion results are corrected by tissue compression deformation. For the feature sub-nodes in the growth dimension of mechanical fingerprint, association coupling rules are constructed between sliding displacement data and sub-nodes. For example, for different tissue strain rate ranges, a quantitative relationship between the corresponding sliding displacement characteristics and adhesion degree is defined. This relationship is established based on the prior knowledge that the smaller the sliding displacement amplitude, the greater the possibility of adhesion. In this way, sliding displacement data can be transformed into the verification basis for elasticity inversion, i.e., the reliability of the elastic modulus inversion results is verified by the interface sliding situation.

[0038] Step S133: Establish a mapping table between the association coupling rules and the feature sub-nodes of each growth dimension, and store it in the collaborative optimization link configuration module of the dual-modal growth network.

[0039] A mapping table is established between the constructed association coupling rules and the corresponding feature sub-nodes of each growth dimension. This table records the association coupling rule identifier and response data type for each sub-node. This mapping table is stored in the collaborative optimization link configuration module of the bimodal growth network for rapid location and use during subsequent optimization processes.

[0040] Step S140: Referring to Figure 2, iteratively optimize the physical quantity inversion rule parameters of the feature sub-nodes based on the association coupling rule.

[0041] The system receives real-time tissue strain response data and sliding displacement data through a collaborative optimization link, matches the corresponding feature sub-nodes and associated coupling rules based on the mapping relationship table, and then iteratively optimizes the physical quantity inversion rule parameters.

[0042] Step S141: Receive tissue strain response data and sliding displacement data through the collaborative optimization link, and match the corresponding feature sub-nodes and associated coupling rules based on the mapping relationship table.

[0043] In this embodiment, new dynamic data is continuously generated as the pregnant woman breathes and the doctor performs procedures. The system receives this data in real time through a collaborative optimization link and finds the corresponding feature sub-nodes and associated coupling rules based on the mapping table. For example, upon receiving compression deformation data of the fat layer, it matches the first-level feature sub-nodes and their associated coupling rules corresponding to the fat layer in the sound velocity inversion growth dimension.

[0044] Step S142: For the feature sub-nodes of the sound velocity inversion growth dimension, extract the mechanical-sound velocity relationship mapping specified in the correlation coupling rule, input the compression deformation parameters in the tissue strain response data into the mapping to generate sound velocity inversion correction coefficients, and adjust the acoustic ray tracing inversion rule parameters or disparity inversion rule parameters of the feature sub-node based on the correction coefficients.

[0045] For feature sub-nodes in the growth dimension of sound velocity inversion, the mechanics-sound velocity relationship mapping in their correlation coupling rules is extracted. For example, the thickness compression percentage of the fat layer is input into the mapping function to obtain the fat layer sound velocity correction coefficient. If the original sound tracing inversion rule assumes a fixed value for the fat layer sound velocity (e.g., 1450 m / s), it is adjusted according to the correction coefficient to obtain a local sound velocity value that better reflects the current tissue state. Similarly, for the myometrium or placental tissue, the corresponding compression deformation parameters (e.g., shear wave propagation velocity change, overall deformation displacement) are extracted to generate the corresponding sound velocity correction coefficient, and the inversion rule parameters of the corresponding sub-nodes are adjusted.

[0046] Step S143: For the feature sub-nodes in the mechanical fingerprint growth dimension, extract the sliding-adhesion quantization relationship specified in the association coupling rule, input the sliding amplitude parameter in the sliding displacement data into the quantization relationship to generate elastic inversion verification coefficients, and compare the verification coefficients with the elastic modulus value output by the feature sub-node. If the comparison deviation exceeds the preset threshold, adjust the elastic inversion rule parameters.

[0047] For feature sub-nodes in the mechanical fingerprint growth dimension, the sliding-adhesion quantization relationship in their association coupling rules is extracted. For example, the sliding amplitude parameter of the placenta-uterus interface is input into this quantization relationship to obtain the elasticity inversion verification coefficient. This coefficient is compared with the elastic modulus value currently output by the sub-node. If the deviation exceeds a preset threshold (e.g., 10%), it indicates that the elasticity inversion result may be inaccurate, and the elasticity inversion rule parameters need to be adjusted. For example, the elastic modulus interval threshold in the elasticity inversion rule can be adjusted to better reflect reality.

[0048] Step S144: Statistically count the correction frequency of each feature sub-node. For feature sub-nodes whose correction frequency reaches a preset frequency threshold, increase the number of their associated sub-nodes to expand the inversion dimension. Analyze the coupling relationship between feature sub-nodes, calculate the sound velocity-elastic modulus consistency index corresponding to each pair of coupling relationships, and for coupling relationships with consistency indices lower than a preset threshold, adjust the connection weights of feature sub-nodes or add intermediate coupling nodes to optimize the coupling logic path. Integrate the adjusted feature sub-node parameters, the expanded number of sub-nodes, and the optimized coupling relationship to generate an iteratively optimized placental location accurate identification network.

[0049] In this embodiment, the system counts the number of times each feature sub-node is corrected. For example, if a fat layer sound velocity inversion sub-node is frequently corrected (the correction frequency reaches a preset threshold), it indicates that the inversion dimension of this sub-node may be insufficient. The system automatically adds associated sub-nodes for it, such as further subdividing the fat layer into superficial and deep fat layers to more precisely invert sound velocity. Simultaneously, the system analyzes the coupling relationships between feature sub-nodes and calculates the sound velocity-elastic modulus consistency index corresponding to each pair of coupling relationships. For example, for sound velocity inversion sub-nodes and elastic modulus quantum nodes in the same spatial region, the system extracts the sound velocity distribution data and elastic modulus distribution data of that region from their inversion results and calculates their correlation coefficient. If the correlation coefficient is lower than a preset threshold (e.g., 0.7), the coupling relationship is considered to have poor consistency and needs optimization. Optimization methods include adjusting the connection weights between sub-nodes or adding intermediate coupling nodes (e.g., introducing a new feature layer) to improve the coupling logic. Finally, all adjusted sub-node parameters, the expanded number of sub-nodes, and the optimized coupling relationships are integrated to generate an iteratively optimized placental location accuracy identification network.

[0050] Referring to Figure 3, step S150: Input the ultrasound video stream to be detected into the optimized dual-modal growth network to generate an individualized tissue sound velocity distribution field and a placental-uterine interface mechanical characteristic field.

[0051] In this embodiment, for pregnant woman Zhang, her newly acquired abdominal ultrasound video stream is input into the optimized bimodal growth network. After iterative optimization, the network can more accurately invert the individualized tissue sound velocity distribution field (i.e., the local sound velocity value at each pixel location) and the placental-uterine interface mechanical characteristic field (including elastic modulus distribution and interface sliding displacement distribution).

[0052] Step S160: Perform pixel-by-pixel geometric distortion correction on the original ultrasound image based on the sound velocity distribution field, mark abnormal tissue areas on the corrected image based on the mechanical feature field, and fuse to generate accurate placental location identification results.

[0053] First, individualized tissue sound velocity distribution fields are extracted. For each pixel in the original ultrasound image, the actual propagation time of the sound beam from the probe surface to the pixel location is calculated based on the local sound velocity value sequence of all tissue layers along the depth direction of that pixel. Simultaneously, the standard propagation time is calculated based on the standard sound velocity value of 1540 m / s. The ratio of the actual propagation time to the standard propagation time is calculated, and this ratio is used to correct the depth coordinates of the pixel, generating a corrected coordinate mapping table for each pixel. Then, the original ultrasound image is resampled and interpolated based on this coordinate mapping table to generate a corrected image of the true anatomical structure. In this way, the geometric distortion caused by uneven sound velocity is eliminated, and the distance measurement between the lower edge of the placenta and the internal cervical os returns to the true anatomical benchmark.

[0054] Next, the mechanical feature field of the placenta-uterus interface was extracted, and the elastic modulus distribution data and the interface sliding displacement distribution data were superimposed on the corrected real anatomical structure image to identify abnormal areas where the elastic modulus value exceeded the normal range (e.g., 20% higher than the surrounding tissue) and suspected adhesion areas where the sliding amplitude was lower than the normal threshold (e.g., 2 mm).

[0055] Finally, based on the spatial location of abnormal areas and suspected adhesion areas, the minimum distance between the lower edge of the placenta and the internal cervical os in the corrected image is measured. Combined with the positional relationship between the abnormal areas and the internal cervical os, placental location classification results (e.g., complete placenta previa, partial placenta previa, etc.) and implantation risk level indicators (e.g., low risk, medium risk, high risk) are generated. The placental location classification results, implantation risk level indicators, and abnormal area annotations are then overlaid onto the corrected image of the actual anatomical structure to generate a fused image of accurate placental location identification for physician reference.

[0056] Step S170: Calculate the comprehensive risk index of placenta accreta.

[0057] Based on the generated fused image, a comprehensive placental implantation risk index P is further calculated to quantify the overall implantation risk. This index is calculated using the following formula:

[0058] in, The number of tissue blocks to be divided for the region of interest (e.g., dividing the placental region into 10×10×10 voxel blocks). For the first The local average value of the sound velocity distribution field of the layered block. The thickness of this layer is... For reference thickness (e.g., take the average thickness of a normal placenta as 30mm). This represents the amplitude of the interface sliding displacement of the layer during the breathing cycle. This is a reference value for the maximum sliding displacement (for example, take the maximum normal sliding value of 5mm). This is the elastic modulus value of the layer. This is a reference value for the elastic modulus of normal tissue (for example, the elastic modulus of a normal placenta is 5 kPa). This represents the contrast feature of the gray-level co-occurrence matrix in this layer (reflecting the complexity of image texture). For reference contrast (e.g., the average contrast of a normal placental area). , These are constant coefficients related to the organizational layer type, and , This is a pre-defined fixed value. The higher the index P, the greater the implantation risk. In this embodiment, the calculated P value for pregnant woman Zhang is 2.8, corresponding to a high-risk level.

[0059] Step S180: Calculate the confidence factor and perform adaptive depth correction.

[0060] In the pixel-by-pixel geometric distortion correction process, to further improve the correction accuracy, the correction depth coordinates of each pixel are adjusted using a confidence factor. Adaptive adjustment is performed. Confidence factor Calculated by the following formula:

[0061] in, This represents the number of tissue layers traversed along the sound beam path of the pixel. For the first Local sound velocity inversion values ​​of layered tissues The standard speed of sound is 1540 m / s. The echo arrival time variance of this layer (reflecting signal stability). This represents the maximum time variance (used for normalization). This represents the acoustic refraction offset of this layer of tissue. Use the reference refraction offset (e.g., take the average offset of a normal tissue interface). This represents the grayscale gradient magnitude of the layer (reflecting the clarity of the boundaries). For the maximum gradient magnitude, This is the thickness uniformity coefficient of the tissue layer (reflecting the consistency of tissue layer thickness). The average value of the uniformity coefficient. Confidence factor. The closer the value is to 1, the more reliable the sound velocity inversion and depth correction for that pixel. For For pixels with lower values, the depth correction range can be appropriately relaxed, or smoothing can be achieved through interpolation. In this embodiment, the system automatically calculates the depth correction value for each pixel. The values ​​are used for weighted fusion of the final correction depth to generate a higher-precision corrected image.

[0062] Step S190: Display and output the results.

[0063] Finally, the ultrasound equipment displays the generated images showing the precise location of the placenta, the placenta accreta risk index (P), and the confidence level distribution map, providing the doctor with this information. The doctor can then make a more accurate diagnosis and surgical plan based on this data.

[0064] Referring to Figure 4, in an exemplary embodiment, a system for accurate placental location identification in abdominal prenatal examination images is provided. This system includes: a probe, a transmitting and receiving circuit, a processor, a memory, and a display device. The probe is used to emit ultrasound waves into the pregnant woman's abdomen and receive the echoes; the transmitting and receiving circuit controls the probe's transmission and reception; the processor is configured to execute the steps described in the above method embodiment; the memory stores ultrasound video streams, intermediate data, optimized network parameters, etc.; and the display device displays the final results.

[0065] In one exemplary embodiment, a computer program product is provided, comprising machine-executable instructions stored in a computer-readable storage medium. A processor of an abdominal prenatal image placental location accuracy identification system reads the machine-executable instructions from the computer-readable storage medium, and the processor executes the machine-executable instructions, causing the system to perform an abdominal prenatal image placental location accuracy identification method.

[0066] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for accurate identification of placental location in abdominal prenatal examination images, characterized in that, This includes constructing a dual-modal growth network for placental location identification, determining the acoustic velocity inversion growth dimension and the mechanical fingerprint growth dimension of the dual-modal growth network, and setting initial feature nodes for each growth dimension; extracting echo time-series features of the sound beam passing through different tissue layers and displacement field data of tissue feature points between adjacent frames from continuous frame ultrasound video streams of abdominal scans, generating feature sub-nodes for each growth dimension, and constructing physical quantity inversion rules for each feature sub-node; generating feature sub-nodes for each growth dimension and constructing physical quantity inversion rules for each feature sub-node, including: for the acoustic velocity inversion growth dimension, extracting from continuous frame ultrasound video streams... The system extracts echo arrival time series data of the sound beam traversing each tissue layer. Based on the difference in arrival time between adjacent angle echoes, it generates ray refraction offset data. A first-level feature sub-node is generated based on this refraction offset data, and a ray tracing inversion rule for retrieving local sound velocity values ​​from each tissue layer is constructed. Displacement field data of tissue feature points between adjacent frames is extracted from the continuous frame ultrasound video stream. A second-level feature sub-node is generated based on this displacement field data, and a disparity inversion rule for calculating equivalent sound velocity based on disparity displacement is constructed. For the mechanical fingerprint growth dimension, image sequences of the probe pressurization and depressurization phases are extracted from the continuous frame ultrasound video stream. Tissue strain response data is generated based on tissue compression deformation before and after pressurization. First-level feature sub-nodes are generated based on this tissue strain response data, and elastic inversion rules for inferring local tissue elastic modulus values ​​are constructed. Image sequences within the maternal respiratory cycle are extracted from continuous frame ultrasound video streams, and the relative sliding trajectory of the placenta-uterine wall interface is tracked to generate sliding displacement data. Second-level feature sub-nodes are generated based on this sliding displacement data, and adhesion assessment rules for determining the adhesion status of the placenta-uterus interface are constructed. Tissue strain response data under probe pressurization and placenta-uterus interface sliding data are extracted from the continuous frame ultrasound video streams. Displacement data is used to construct the correlation and coupling rules between the response data and displacement data and the feature sub-nodes of each growth dimension. This is used to establish the physical constraint relationship between the sound velocity inversion and mechanical fingerprint dual modes. Based on the correlation and coupling rules, the physical quantity inversion rule parameters of the feature sub-nodes are iteratively optimized. The ultrasound video stream to be detected is input into the optimized dual-modal growth network to generate an individualized tissue sound velocity distribution field and a placenta-uterus interface mechanical feature field. Based on the sound velocity distribution field, the original ultrasound image is corrected pixel by pixel for geometric distortion. Based on the mechanical feature field, abnormal tissue regions are marked on the corrected image. The results are then fused to generate a precise placental location identification result.

2. The method for accurate identification of placental location in abdominal prenatal examination images according to claim 1, characterized in that: The construction of the association and coupling rules between the response data and displacement data and the feature sub-nodes of each growth dimension includes: setting the acquisition trigger conditions for tissue strain response data; when the probe pressure value exceeds a preset pressure threshold, capturing the image sequence of the pressurization stage and extracting the compression deformation time series data of each tissue layer; setting the acquisition period for sliding displacement data; capturing the image sequence within the respiratory cycle according to the maternal respiratory frequency and tracking the displacement trajectory of interface feature points; constructing association and coupling rules between tissue strain response data and feature sub-nodes for the feature sub-nodes of the growth dimension of sound velocity inversion, specifying the relationship mapping between compression deformation and sound velocity changes corresponding to different tissue layer types, used to transform mechanical response features into constraints for sound velocity inversion; constructing association and coupling rules between sliding displacement data and feature sub-nodes for the feature sub-nodes of the mechanical fingerprint growth dimension, specifying the quantitative relationship between sliding displacement features and adhesion degree corresponding to different tissue strain rate intervals, used to transform sliding displacement data into verification basis for elastic inversion; establishing a mapping relationship table between the association and coupling rules and the feature sub-nodes of each growth dimension, and storing it in the collaborative optimization link configuration module of the dual-modal growth network.

3. The method for accurate identification of placental location in abdominal prenatal examination images according to claim 2, characterized in that: The physical quantity inversion rule parameters of the feature sub-nodes are iteratively optimized based on the aforementioned correlation coupling rules, including: receiving tissue strain response data and sliding displacement data through a collaborative optimization link; matching the corresponding feature sub-nodes and correlation coupling rules based on a mapping relationship table; for feature sub-nodes in the sound velocity inversion growth dimension, extracting the mechanical-sound velocity relationship mapping specified in the correlation coupling rules; inputting the compression deformation parameters in the tissue strain response data into the mechanical-sound velocity relationship mapping to generate sound velocity inversion correction coefficients; and adjusting the ray tracing inversion rule parameters or disparity inversion rule parameters of the feature sub-node based on the correction coefficients; for feature sub-nodes in the mechanical fingerprint growth dimension, extracting the sliding-adhesion quantization relationship specified in the correlation coupling rules; and inputting the sliding amplitude parameters in the sliding displacement data into the mapping relationship table. The quantification relationship generates elastic inversion verification coefficients. These verification coefficients are compared with the elastic modulus value output by the feature sub-node. If the comparison deviation exceeds a preset threshold, the elastic inversion rule parameters are adjusted. The correction frequency of each feature sub-node is statistically analyzed. For feature sub-nodes with correction frequencies reaching a preset frequency threshold, the number of their associated sub-nodes is increased to expand the inversion dimension. The coupling relationships between feature sub-nodes are analyzed, and the sound velocity-elastic modulus consistency index corresponding to each pair of coupling relationships is calculated. For coupling relationships with consistency indices below a preset threshold, the connection weights of the feature sub-nodes are adjusted or intermediate coupling nodes are added to optimize the coupling logic path. The adjusted feature sub-node parameters, the expanded number of sub-nodes, and the optimized coupling relationships are integrated to generate an iteratively optimized placental location accuracy identification network.

4. The method for accurate identification of placental location in abdominal prenatal examination images according to claim 1, characterized in that: The extraction of image sequences from the continuous frame ultrasound video stream for the probe pressurization and depressurization phases includes: identifying a complete pressure change cycle from baseline pressure to peak pressure and back to baseline pressure in the original ultrasound video stream; extracting the image sequence of the pressure rise phase within each cycle as the pressurization phase image sequence; and extracting the image sequence of the pressure fall phase as the depressurization phase image sequence; performing inter-frame registration processing on the pressurization and depressurization phase image sequences respectively; extracting longitudinal compression displacement data of each tissue layer from the registered pressurization phase image sequence; and extracting elastic recovery displacement data of each tissue layer from the registered depressurization phase image sequence. The extraction of image sequences within the maternal respiratory cycle from the continuous frame ultrasound video stream includes: identifying the start and end times of the periodic undulation movement of the abdominal wall in the original ultrasound video stream; extracting a complete image sequence from the inspiratory to expiratory phase within each respiratory cycle as the respiratory cycle image sequence; performing inter-frame registration processing on the respiratory cycle image sequence; and tracing the displacement trajectories of characteristic points at the junction of the lower edge of the placenta and the internal cervical os, and at the junction of the posterior placenta and the myometrium, from the registered image sequence.

5. The method for accurate identification of placental location in abdominal prenatal examination images according to claim 3, characterized in that: The process for identifying feature sub-nodes in the sound velocity inversion growth dimension involves extracting the mechanical-sound velocity relationship mapping specified in the correlation coupling rules. Compression deformation parameters from tissue strain response data are input into this mechanical-sound velocity relationship mapping to generate sound velocity inversion correction coefficients. This includes: extracting compression deformation parameters from the tissue strain response data for the tissue layer governed by the corresponding feature sub-node; extracting the thickness compression percentage for the fat layer, the shear wave propagation velocity change for the myometrium, and the overall deformation displacement for the placental tissue; retrieving the mechanical-sound velocity relationship mapping function stored in the correlation coupling rules of the feature sub-node, where the mapping function is constructed based on experimental fitting curves, theoretical correlation models, or numerical simulation results for different tissue types; inputting the compression deformation parameters into the mapping function to calculate the sound velocity correction coefficients for the corresponding tissue layer; and storing the sound velocity correction coefficients for each tissue layer according to tissue layer type to form a sound velocity inversion correction coefficient vector for use by feature sub-nodes in the sound velocity inversion growth dimension.

6. The method for accurate identification of placental location in abdominal prenatal examination images according to claim 3, characterized in that: The analysis of the coupling relationships between feature sub-nodes and the calculation of the sound velocity-elastic modulus consistency index corresponding to each pair of coupling relationships include: traversing all pairs of feature sub-nodes with coupling relationships in the dual-modal growth network, each pair containing a feature sub-node of the sound velocity inversion growth dimension and a feature sub-node of the mechanical fingerprint growth dimension, with both sub-nodes governing the tissue layer of the same spatial region; extracting the sound velocity distribution data of the tissue layer of the same spatial region from the inversion results of the feature sub-nodes of the sound velocity inversion growth dimension, and extracting the elastic modulus distribution data of the tissue layer of the same spatial region from the inversion results of the feature sub-nodes of the mechanical fingerprint growth dimension. The sound velocity distribution data and the elastic modulus distribution data are paired at the same coordinate points, and the correlation coefficient between the sound velocity value and the elastic modulus value is calculated. If the correlation coefficient is lower than a preset threshold, the consistency index is judged to be low consistency. If the correlation coefficient is higher than or equal to the preset threshold, the local difference map between the sound velocity distribution and the elastic modulus distribution is further calculated, and the number of pixels with difference values ​​exceeding the preset threshold is counted. The consistency index value is determined based on the ratio of this number to the total number of pixels in the tissue area. The coupling association identifier and consistency index value of each pair of feature sub-nodes are organized into a coupling association consistency statistics table and stored in the collaborative optimization link configuration module of the dual-modal growth network.

7. The method for accurate identification of placental location in abdominal prenatal examination images according to claim 1, characterized in that: Based on the sound velocity distribution field, pixel-by-pixel geometric distortion correction is performed on the original ultrasound image. Based on the mechanical feature field, abnormal tissue regions are identified on the corrected image. The results are then fused to generate a precise placental location identification result. This includes: extracting an individualized tissue sound velocity distribution field; for each pixel in the original ultrasound image, calculating the actual propagation time of the sound beam based on the local sound velocity value sequence of all tissue layers along the depth direction of that pixel; calculating the standard propagation time based on the standard sound velocity value; calculating the ratio of the two and correcting the depth coordinates of that pixel; generating a corrected coordinate mapping table for each pixel; and resampling and interpolating the original ultrasound image based on the coordinate mapping table to generate a corrected image of the true anatomical structure. The placental-uterine interface mechanical feature field is used to superimpose elastic modulus distribution data and interface sliding displacement distribution data onto the corrected real anatomical structure image, identifying abnormal areas where the elastic modulus value exceeds the normal range and suspected adhesion areas where the sliding amplitude is below the normal threshold. Based on the spatial location of abnormal areas and suspected adhesion areas, the minimum distance between the lower edge of the placenta and the internal cervical os in the corrected image is measured. Combining the positional relationship between the abnormal areas and the internal cervical os, a placental location classification result and implantation risk level label are generated. The placental location classification result, implantation risk level label, and abnormal area label are superimposed onto the corrected real anatomical structure image to generate a fused image of accurate placental location identification.

8. A system for accurately identifying the placental location in abdominal prenatal examination images, used to implement the method for accurately identifying the placental location in abdominal prenatal examination images according to any one of claims 1-7, characterized in that, Includes: a probe used to emit ultrasound waves into the pregnant woman's abdomen and receive ultrasound echo signals; A transmitting and receiving circuit, connected to the probe, is used to control the transmission and reception of the probe; a processor, connected to the transmitting and receiving circuit, is configured to execute a method for accurate identification of placental location in abdominal prenatal examination images; a memory, connected to the processor, is used to store the continuous frame ultrasound video stream, the network parameters of the dual-modal growth network, the individualized tissue sound velocity distribution field, and the placental-uterine interface mechanical characteristic field. A display device, connected to the processor, is used to display the fused and generated accurate placental location identification results.

9. A computer program product, characterized in that, The computer program product includes machine-executable instructions stored in a computer-readable storage medium. The processor of the abdominal prenatal image placental location accurate recognition system reads the machine-executable instructions from the computer-readable storage medium and executes the machine-executable instructions, causing the abdominal prenatal image placental location accurate recognition system to perform the abdominal prenatal image placental location accurate recognition method.

Images