Method and system for driving intelligent diagnosis, treatment, prevention and control of birth defects based on clinical detection

Abstract

The invention relates to the technical field of medical information processing, in particular to a birth defect intelligent diagnosis, treatment, prevention and control method and system based on clinical detection driving, and the method comprises the steps: obtaining a fetus ultrasonic image, gestational week information and clinical detection data; performing organ segmentation on the ultrasonic image, generating segmentation uncertainty information, extracting morphological features and topological features, and calculating a stability index set under a preset image transformation set; constructing an organ evidence graph taking organs as nodes on the basis of the information, and generating a residual evidence set in combination with a gestational week normal development benchmark; determining a candidate mechanism set according to clinical detection data and a mechanism knowledge base, and performing evidence and mechanism bidirectional message transmission by adopting an evidence weight rule to obtain mechanism probability distribution; and determining a minimum irreplaceable mechanism set through a minimum description length criterion and an anti-fact ablation test, and outputting a risk layering result and an active acquisition plan. According to the method, the image evidence utilization reliability and the mechanism inference interpretability are improved, and repeated inspection is reduced.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, specifically to a method and system for intelligent diagnosis, treatment and prevention of birth defects based on clinical testing. Background Technology

[0002] Early identification and stratified management of birth defects directly impact perinatal diagnostic and treatment efficiency, referral pathways, and the allocation of prenatal counseling and follow-up resources, playing a crucial role in reducing disability rates and optimizing regional maternal and child health service systems. Currently, clinical examinations heavily rely on ultrasound image recognition capabilities and operator experience. Significant differences exist among institutions in equipment configuration, completeness of slice acquisition, and noise and artifact control, leading to large fluctuations in the quality of imaging evidence at the same gestational age. Simultaneously, phenotypic nonspecificity and inconsistencies often exist between laboratory tests, gestational developmental patterns, and imaging findings, resulting in situations of "visible abnormalities but unclear mechanisms" or "abnormal indicators but unclear localization." Existing technologies often treat image recognition results as single-point feature inputs, lacking a unified expression of segmentation confidence, feature stability, and deviation from developmental benchmarks. They also lack a closed loop for interpretable inference and refutation verification at the candidate mechanism level, easily leading to problems such as unstable risk stratification, lack of targeted follow-up plans, and high costs of repeated examinations. Summary of the Invention

[0003] This invention provides a method and system for intelligent diagnosis and prevention of birth defects based on clinical testing, which addresses at least the problem of how to achieve interpretable mechanism inference, stable risk stratification, and targeted re-examination and collection planning under conditions of fluctuations in fetal ultrasound image recognition quality and inconsistencies between multi-source clinical testing.

[0004] In a first aspect, the present invention provides a method for intelligent diagnosis and prevention of birth defects based on clinical testing, the method comprising: To obtain fetal ultrasound images, gestational age information, and clinical test data; Fetal organ segmentation is performed based on fetal ultrasound images, generating segmentation uncertainty information, extracting morphological features and topological features based on continuous homology, and calculating a set of stability indices based on a preset set of image transformations. Based on morphological features, topological features based on continuous homology, segmentation of uncertainty information and stability index set, an organ evidence map with organs as nodes is constructed, and a residual evidence set is generated based on the normal fetal development benchmark corresponding to gestational age information. A set of candidate mechanisms is determined based on clinical test data and a mechanism knowledge base. A two-way message passing system based on evidence weighting rules using a stability index set is used to obtain the mechanism probability distribution. Based on the mechanism probability distribution and residual evidence set, the minimum description length criterion and counterfactual ablation test are used to determine the minimum set of irreplaceable mechanisms, and the risk stratification results and active collection plan are output based on the minimum set of irreplaceable mechanisms.

[0005] Secondly, this invention provides a clinical testing-driven intelligent diagnosis and treatment prevention system for birth defects, used to implement a clinical testing-driven intelligent diagnosis and treatment prevention method for birth defects. The system includes: The data acquisition module is used to acquire fetal ultrasound images, gestational age information, and clinical test data; The feature generation module is used to segment fetal organs based on fetal ultrasound images to generate segmentation uncertainty information, extract morphological features and topological features based on continuous homology, and calculate a set of stability indices based on a preset set of image transformations. The evidence mapping module is used to construct an organ evidence map with organs as nodes based on morphological features, topological features based on continuous homology, segmentation uncertainty information and stability index set, and generate a residual evidence set based on the normal fetal development benchmark corresponding to gestational age information. The mechanism inference module is used to determine the set of candidate mechanisms based on clinical test data and mechanism knowledge base, and to use evidence weight rules based on stability index set to perform two-way message passing between evidence and mechanism to obtain the mechanism probability distribution; The results output module is used to determine the minimum set of irreplaceable mechanisms based on the mechanism probability distribution and residual evidence set, using the minimum description length criterion and counterfactual ablation test, and output the risk stratification results and active collection plan based on the minimum set of irreplaceable mechanisms.

[0006] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: By segmenting organs and generating segmentation uncertainty information, the credibility of image recognition was quantitatively represented; by extracting morphological and topological features and calculating a set of stability indicators under preset image transformations, the robustness of features and evidence weighting were achieved; by constructing an organ evidence map with organs as nodes and generating a residual evidence set in conjunction with the normal development benchmark at gestational age, structured attribution clues for abnormal deviations were realized; by bidirectional message passing between evidence and mechanism based on evidence weight rules, interpretable inference of mechanism probability distribution was achieved; and by screening out the minimum set of irreplaceable mechanisms using the minimum description length criterion and counterfactual ablation test, a closed-loop generation of risk stratification and proactive collection plans was achieved, thereby reducing redundant review and improving the consistency of follow-up decisions. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the execution flow of the method of the present invention; Figure 2The following are example ultrasound images, in sequence, of the fetal lateral ventricle section, four-chamber view, bilateral kidney transverse section, and sagittal view of the spine, as described in a specific embodiment of the present invention. Figure 3 This is a stability score statistics chart in a specific embodiment of the present invention; Figure 4 This is a schematic diagram of organ evidence in a specific embodiment of the present invention; Figure 5 This is a statistical diagram of the mechanism probability distribution in a specific embodiment of the present invention; Figure 6 This is a structural block diagram of the system of the present invention. Detailed Implementation

[0008] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation.

[0009] Clinical testing typically encompasses two main sources of information: imaging examinations and laboratory tests. On the one hand, imaging examinations such as ultrasound can non-invasively obtain structural and developmental clues about key fetal organs. Image recognition technology can standardize the extraction of cross-sections, organ regions, and suspicious structures, forming calculable imaging evidence. On the other hand, laboratory tests such as serological, biochemical, and genetic indicators can reflect physiological states, risk signals, and potential etiologies, providing mechanistic clues that complement imaging. In actual clinical practice, these two types of tests are often obtained in batches over time, with significant differences in the strength of evidence. Furthermore, situations exist where "imaging suggests abnormalities but test indicators do not deviate significantly" or "tests indicate high risk but imaging manifestations are atypical." Therefore, a method is needed that can uniformly represent, characterize the credibility of, and infer mechanistic correlations between imaging evidence obtained through image recognition and clinical test indicators within the same framework. This method should further translate the inference results into actionable risk stratification and subsequent proactive data collection recommendations to support a more consistent and interpretable clinical decision-making process.

[0010] like Figure 1 As shown, a method for intelligent diagnosis and prevention of birth defects based on clinical testing is proposed, the method comprising: To obtain fetal ultrasound images, gestational age information, and clinical test data; During a prenatal examination, fetal ultrasound images, gestational age information, and clinical test data are collected and linked together. Fetal ultrasound images are derived from image files and frame sequences output by the ultrasound equipment; gestational age information comes from examination registration or electronic medical records; and clinical test data comes from serological screening, genetic testing, or routine biochemical tests. At the time of collection, an examination identifier is generated for each examination, recording the collection time, gestational age, and a list of examination items. The fetal ultrasound images and clinical test data are then bound according to the examination identifier, forming an input data package for subsequent processing, including organ segmentation, feature extraction, evidence construction, and mechanism inference.

[0011] Acquiring fetal ultrasound images involves acquiring ultrasound frame sequences corresponding to multiple preset section types, and performing section classification and frame-level quality control on the ultrasound frame sequences to remove unusable frames, resulting in an image evidence frame set; fetal organ segmentation is performed based on the image evidence frame set.

[0012] In one embodiment, acquiring fetal ultrasound images includes acquiring corresponding ultrasound frame sequences for multiple preset section types, and forming an image evidence frame set after acquisition. The image evidence frame set is used to support subsequent fetal organ segmentation. The preset section types are configured according to the standard prenatal ultrasound examination guidelines and include at least one or more of the following: four-chamber view of the heart, three-vessel trachea view, transverse view of the brain, and transverse view of the abdomen.

[0013] During the acquisition phase, the ultrasound equipment outputs a continuous playback frame sequence, retaining metadata such as timestamp, frame rate, imaging depth, and gain for each frame. Frame sequences of the same slice type are archived using inspection tags and slice type. In the slice classification phase, the slice type is determined frame by frame in the ultrasound frame sequence, forming a slice label sequence. Slice classification is implemented using an image classification model. The image classification model is trained using manually labeled slice image libraries, and grayscale and scale normalization is performed on samples from different equipment to reduce domain differences. Slice classification outputs a slice type confidence score. When the slice type confidence score is lower than a preset threshold, a review process is triggered. The review process is based on key anatomical structure visibility rules.

[0014] The frame-level quality control stage performs usability screening on candidate frames based on section classification to remove unusable frames. Usability evaluation includes sharpness evaluation, occlusion evaluation, and field-of-view coverage evaluation. Sharpness evaluation uses image gradient energy to obtain a sharpness score; occlusion evaluation uses the proportion of acoustic shadowing and saturation areas to obtain an occlusion score; and field-of-view coverage evaluation uses the overlap ratio between the key anatomical structure detection box and the imaging sector to obtain a coverage score. The sharpness score, occlusion score, and coverage score are combined to form a quality score. Frames with a quality score below a threshold are marked as unusable and removed. For each section type, a predetermined number of image evidence frames are selected from the remaining frames, sorted by quality score, and uniformly extracted over time to reduce redundancy. The image evidence frame set is stored indexed by inspection identifier, section type, frame index, and quality score for direct access during fetal organ segmentation. To ensure consistency in subsequent fetal organ segmentation, the image evidence frame set also records the section label and metadata summary for each frame. The fetal organ segmentation stage reads the image evidence frame set sequentially by section type.

[0015] Fetal organ segmentation is performed based on fetal ultrasound images, generating segmentation uncertainty information, extracting morphological features and topological features based on continuous homology, and calculating a set of stability indices based on a preset set of image transformations. Fetal organ segmentation is performed based on fetal ultrasound images. The input consists of a set of image evidence frames and corresponding section type information, and the output is a segmentation probability mask. Segmentation uncertainty information is generated simultaneously. Morphological features and topological features based on continuous homology are calculated on the segmentation probability mask to form a set of structured features oriented towards the organ level. A preset set of image transformations is applied to the set of image evidence frames, and segmentation and feature calculation are repeated to obtain a set of stability indices. Evidence weighting rules are generated based on the stability indices and segmentation uncertainty information to provide a unified confidence input for subsequent organ evidence image weighting and two-way message passing between evidence and mechanism.

[0016] Fetal organ segmentation generates a segmentation probability mask, and the segmentation uncertainty information includes a pixel-level uncertainty map and an organ-level uncertainty score; the organ-level uncertainty score is determined based on the variance of the segmentation probability mask obtained by performing multiple inferences on the image evidence frame set.

[0017] In one embodiment, fetal organ segmentation generates a segmentation probability mask, and segmentation uncertainty information is generated based on the segmentation probability mask. During processing, image evidence frames are read according to the section type, and grayscale normalization, scale unification, and fan-shaped imaging region cropping are performed on each frame. Regions outside the fan are masked to avoid background pixel interference; meanwhile, metadata such as pixel spacing, imaging depth, gain parameters, and frame-level quality score are retained.

[0018] The segmentation model outputs a multi-organ probability map, which provides the membership probability of each target organ at each pixel location. A segmentation probability mask is obtained by organ channel. The segmentation probability mask then undergoes connected component filtering and morphological closing operations. Connected component filtering incorporates anatomical location priors based on the section type, retaining only the connected components with the highest overlap with the expected organ region. Small holes are filled to obtain continuous boundaries.

[0019] Segmentation uncertainty information includes a pixel-level uncertainty map and an organ-level uncertainty score. The pixel-level uncertainty map characterizes the degree of hesitation in pixel classification. It is calculated based on a segmentation probability mask and can be measured using the complement of the pixel's maximum membership probability or an information entropy-based metric. Neighborhood averaging is used in the boundary neighborhood to reduce scatter points. The organ-level uncertainty score is determined based on the variance of the segmentation probability mask obtained from multiple inferences performed on the image evidence frame set. Multiple inferences can be achieved by enabling a random deactivation layer or using multiple independently trained segmentation models. The segmentation probability variance from multiple inferences is calculated at each pixel location for the same organ, and then weighted and aggregated within the organ region according to the frame-level quality score to obtain the organ-level uncertainty score.

[0020] To avoid bias caused by differences in organ size, organ-level uncertainty scores are normalized by organ region area after aggregation, and a minimum area threshold is set. When the organ region is smaller than the minimum area threshold, the organ-level uncertainty score is marked as high-risk uncertainty. For multi-frame results of the same cross-section type, the organ-level uncertainty score can be further weighted and averaged inter-frame according to the frame-level quality score to form a cross-section-level uncertainty score, which is used for subsequent evidence weighting. In the output stage, the segmentation probability mask, pixel-level uncertainty map, and organ-level uncertainty score are indexed and stored according to inspection identifier, cross-section type, organ name, and frame index, and the segmentation model version used during generation is recorded for subsequent recalculation and traceability. When the organ-level uncertainty score exceeds the preset threshold, the system adds the corresponding organ and cross-section type to the review list, providing candidates for subsequent active acquisition plans.

[0021] Morphological features are determined based on segmentation probability masks and include at least one of the following: area, perimeter, major axis length, minor axis length, axis ratio, symmetry index, and relative position index between organs.

[0022] In one embodiment, morphological features are determined based on a segmentation probability mask. These morphological features include at least one or more of the following: area, perimeter, major axis length, minor axis length, axis ratio, symmetry index, and relative position index between organs. Before computation, the segmentation probability mask is thresholded to obtain binary organ regions, and connected component filtering and boundary smoothing are performed. The connected component filtering incorporates the anatomical location prior of the section type to preserve the main connected components.

[0023] The area is calculated from the number of pixels in the binary organ region and the pixel spacing, and then weighted by averaging the frame-level quality score and organ-level uncertainty score across multiple frames of the same organ to form the organ-level area. The perimeter is obtained by extracting the boundary pixel chain through boundary tracking and converting it into physical length. The major and minor axis lengths are obtained by performing principal component analysis on the pixel coordinates of the binary organ region to obtain the principal axis direction, and then calculating the projection range in the direction orthogonal to the principal axis direction; the axis ratio is determined by the ratio of the major axis length to the minor axis length.

[0024] Symmetry indices are used to evaluate bilateral organ consistency. The area, axial ratio, and centroid position of the left and right organs are calculated separately under the same gestational age and the same section type. Normalized differences are then calculated to form the symmetry index. When the left and right organs do not appear simultaneously in the same frame, matching is performed among image evidence frames with the same section type and temporal proximity, prioritizing clear frames based on frame-level quality scores. The relative position index between organs is calculated based on the centroid of each organ. The centroid is obtained from the mean coordinates of the binary organ region. Relative position is expressed as the distance and direction between organ centroids and can be normalized according to imaging depth to reduce the influence of body size differences.

[0025] To avoid bias introduced by low-quality segmentation, when the organ-level uncertainty score of an organ exceeds a threshold, the morphological features of that organ are marked as low-confidence and their weights are reduced during inter-frame aggregation. After calculation, morphological feature records are output by organ name and stored together with frame index, section type, gestational age information, and organ-level uncertainty score for direct use in subsequent topological feature calculation, stability assessment, and evidence construction stages. Morphological consistency checks can be performed before output, such as checking whether the axial ratio falls within a preset range and whether the area monotonically changes within a reasonable range with gestational age information; if the check fails, the corresponding frame is added to the review queue and the original segmentation probability mask is retained for recalculation.

[0026] Topological features based on continuous cohomology include: generating a distance transformation map based on a segmentation probability mask, performing continuous cohomology calculation based on the distance transformation map, and obtaining a topological feature vector as the topological feature based on continuous cohomology.

[0027] In one embodiment, topological features based on continuous homology are used to characterize the connectivity and porosity of organ segmentation regions, supplementing shape differences that are difficult to express by segmentation boundary morphological features. During processing, a segmentation probability mask is used as input. First, thresholding is performed on the target organ channels to obtain binary organ regions, and then a distance transformation map is generated on the binary organ regions.

[0028] The pixel values ​​in the distance transformation map represent the shortest distance from that pixel to the organ boundary. The distance increases with depth within the boundary, while the distance outside the boundary is recorded as zero, thus forming a continuous scalar field. A threshold sequence is constructed based on the distance transformation map as the filtering basis. This threshold sequence can be sampled at equal intervals or by distance quantiles, generating a set of sub-regions according to the threshold from smallest to largest. For each sub-region set, the occurrence and disappearance of connected components and holes as the threshold changes are calculated, yielding a continuous cohomology result. This continuous cohomology result forms a set of event pairs, each containing an event occurrence threshold and an event disappearance threshold.

[0029] To facilitate fusion with morphological features, the continuous cohomology results are converted into fixed-dimensional topological feature vectors. The conversion involves binning the event duration intervals and summing the results for different event types to obtain the topological feature vectors. To suppress transient events caused by noise, a minimum duration threshold is set; events with duration intervals smaller than this threshold are not included in the topological feature vector. The minimum duration threshold is configured based on pixel spacing and organ scale.

[0030] For the topological feature vectors of the same organ across multiple frames, a weighted average is calculated based on frame-level quality scores and organ-level uncertainty scores to output an organ-level topological feature vector. This vector is then stored in conjunction with the frame index and section type to avoid mixing different projections. Continuous cohomology calculation can be performed using a topological data analysis library, with inputs being a distance transformation graph and a threshold sequence, and output being a set of event pairs. To ensure repeatability, the threshold sequence length, sampling method, and minimum persistence threshold are fixed configurations within the same project and written to the processing log.

[0031] To avoid instability in the distance transformation map caused by noise in the segmentation probability mask, a boundary smoothing is performed on the binary organ regions before the distance transformation, and the maximum distance value is limited to prevent extreme values ​​from dominating binning statistics. The dimension of the topological feature vector and the binning boundary are determined during system initialization. The binning boundary can be obtained statistically based on historical training data, making data from different gestational weeks and different devices comparable on the same scale. In the output stage, the topological feature vector, along with the corresponding organ name, section type, gestational week information, and organ-level uncertainty score, is encapsulated into a topological feature record for repeated generation and comparison under a preset image transformation set during subsequent stability index set calculations.

[0032] The preset image transformation set includes at least two of the following: brightness perturbation, gain perturbation, affine transformation, and local occlusion transformation; the stability index set includes a stability score, which is determined based on the relative fluctuations of morphological features and topological features based on continuous cohomology under the preset image transformation set; the evidence weight rule is determined based on the stability score and segmentation uncertainty information, and is used to weight the evidence in the organ evidence map to perform two-way message passing between evidence and mechanism.

[0033] In one embodiment, a preset image transformation set is used to simulate common imaging perturbations and evaluate the sensitivity of features to these perturbations. The preset image transformation set includes at least two of the following: brightness perturbation, gain perturbation, affine transformation, and local occlusion transformation. Brightness perturbation changes the overall brightness through linear grayscale mapping; gain perturbation simulates device gain changes through contrast scaling; affine transformation simulates probe attitude changes through small-angle rotation and micro-scaling; and local occlusion transformation simulates sound shadows or local occlusion through random occlusion blocks.

[0034] Image transformations are applied frame by frame to the image evidence frame set to obtain a transformed frame set. The transformed frame set is then subjected to the same process as the untransformed frames, repeating fetal organ segmentation, segmentation uncertainty calculation, morphological feature extraction, and topological feature vector calculation to form a feature sequence of the same organ under multiple transformation conditions. The stability index set includes stability scores, which characterize the relative fluctuation of the feature sequence and can be calculated as follows: ; in, For stability score, This represents the mean of the same feature under various image transformation conditions. The standard deviation of the same feature under various image transformation conditions. To prevent small positive numbers with zero denominators, stability scores are calculated for key components of area, axial ratio, symmetry index, and topological feature vector, and then aggregated into organ-level stability scores according to preset weights, forming a stability index set.

[0035] The evidence weighting rule is determined based on the stability score and segmentation uncertainty information. Specifically, higher stability scores and lower organ-level uncertainty scores are assigned higher weights; lower stability scores or higher organ-level uncertainty scores result in lower weights, and the corresponding organs are added to the supplementary sampling list. The evidence weighting rule is used to weight the evidence in the organ evidence map and control the magnitude of the evidence's update to the mechanism probability distribution during subsequent two-way message passing between evidence and mechanism, thereby suppressing false evidence introduced by imaging perturbations and segmentation instability.

[0036] The output phase stores the stability index set, evidence weighting rules, organ names, section types, and gestational age information for direct retrieval in subsequent phases. To ensure comparability between different section types, organ-level stability scores are normalized to a range of zero to one after aggregation, and a minimum sample size constraint is set. When the number of valid frames for a certain organ in the transform frame set is insufficient, a conservative value is adopted for the stability score, and the weight is reduced. The evidence weighting rules are truncated with upper and lower limits during application to prevent a single high-weighted piece of evidence from dominating the update. Furthermore, the weights and update amounts are recorded after each round of bidirectional message passing for easy auditing and playback.

[0037] Based on morphological features, topological features based on continuous homology, segmentation of uncertainty information and stability index set, an organ evidence map with organs as nodes is constructed, and a residual evidence set is generated based on the normal fetal development benchmark corresponding to gestational age information. Morphological features, topological features based on persistent homology, segmentation uncertainty information, and stability indices are aggregated into organ-level evidence units, and an organ-based evidence map is constructed accordingly. This organ-based evidence map simultaneously encodes developmental consistency relationships between organs, enabling the joint organization of multi-organ evidence within a single examination. Combining gestational age information, a reference interval for the corresponding gestational age is obtained from a normal fetal development benchmark. The morphological features of each organ are compared with the reference interval using topological features based on persistent homology, forming a residual evidence set. This residual evidence set provides the direction and magnitude of each deviation, while retaining uncertainty and stability information, serving as weighted evidence input for subsequent inference stages.

[0038] The organ evidence map includes organ nodes and developmental constraint edges, which are used to characterize bilateral symmetric constraint relationships or anatomical correlation constraint relationships. The normal fetal development benchmark provides a reference interval corresponding to the gestational age information. The residual evidence set includes the direction and magnitude of deviation of morphological features and topological features based on continuous coherence relative to the reference interval.

[0039] In one embodiment, the organ evidence graph consists of organ nodes and developmental constraint edges. Organ nodes are uniquely identified by the organ name, and their attributes include at least organ-level morphological feature records, organ-level topological feature vectors, organ-level uncertainty scores, and organ-level stability scores. The source of the cross-section type and the set of frame indexes are also retained for traceability.

[0040] Organ-level morphological feature records may include one or more of the following: area, perimeter, major axis length, minor axis length, axial ratio, symmetry index, and relative position index between organs; organ-level topological feature vectors are derived from the continuous cohomology calculation results of the distance transformation graph. To ensure the usable boundary of node evidence, organ nodes also record evidence confidence labels, which are jointly determined by the organ-level uncertainty score and the organ-level stability score. When the uncertainty score is high and the stability score is low, the node is marked as a low-confidence node, and the deviation magnitude of the node is conservatively processed subsequently.

[0041] Developmental constraint edges are used to express developmental consistency relationships between organs, and include at least bilateral symmetry constraint edges and anatomical association constraint edges. Bilateral symmetry constraint edges connect pairs of organs, and their attributes include the symmetry measurement type and allowable deviation range, used to constrain the consistency of the left and right organs in terms of area, axial ratio, and centroid position. Anatomical association constraint edges connect pairs of organs that have anatomical coupling or developmental linkage, and their attributes include the association feature type and reference relationship, such as the reasonable range of relative position indicators between organs, the morphological proportion relationship of a specific organ combination, or the cooperative change relationship of topological features.

[0042] The configuration of developmental constraint edges is derived from structural prior entries and clinical examination standards in the mechanism knowledge base. These edges can be fixed as edge templates during system initialization and automatically instantiated according to the actual available organ nodes during each examination. The normal fetal development benchmark uses gestational age as an index to provide reference intervals for various morphological and topological features of each organ. The reference intervals can be represented using quantile intervals, such as the low and high quantile boundaries obtained from normal sample statistics, and non-integer gestational age reference intervals are obtained by linear interpolation in the gestational age dimension.

[0043] The generation of the residual evidence set is centered on reference interval comparison. For each organ and each feature, the deviation direction and magnitude are calculated separately. The deviation direction indicates whether the feature value is below, within, or above the reference interval, while the deviation magnitude quantifies the degree of deviation. The deviation magnitude can be obtained by normalizing it according to the reference interval width. ; in, For deviation range, For eigenvalues, The midpoint of the reference interval, For reference interval width, To prevent small positive numbers with a denominator of zero, the deviation magnitude of low-confidence nodes is attenuated before being written into the residual evidence set. The attenuation coefficient is determined by both the stability score and the uncertainty score, ensuring that unstable or uncertain segmentation results do not form overly strong evidence in the residual evidence set.

[0044] The residual evidence set is indexed by organs and organized into multiple residual records. Each residual record includes at least the organ name, feature name, deviation direction, deviation magnitude, stability score, and uncertainty score. It can also include edge residual terms related to developmental constraints, such as the symmetry residual of bilateral symmetry constraints and the relative position residual of anatomically related constraints. Edge residual terms are obtained by performing the same interval comparison on the reference relationship defined by the edge attributes, thus achieving simultaneous quantification of node evidence and edge constraints. Through this construction method, the organ evidence map provides a structured multi-organ evidence organization, and the residual evidence set provides weighted deviation evidence based on gestational age, providing a consistent data entry point for subsequent mechanism probability distribution inference stages.

[0045] A set of candidate mechanisms is determined based on clinical test data and a mechanism knowledge base. A two-way message passing system based on evidence weighting rules using a stability index set is used to obtain the mechanism probability distribution. Candidate mechanism sets are determined based on clinical test data and a mechanism knowledge base, with organ evidence maps and residual evidence sets serving as inputs for imaging-based evidence. First, the mechanism knowledge base is initially screened using clinical test data to obtain a candidate mechanism set. Then, according to evidence weighting rules, the candidate mechanism set is associated with organ nodes in the organ evidence map, performing iterative updates through bidirectional message passing between evidence and mechanisms. This aligns the imaging-based residual evidence with the clinical screening results within the same probability space, ultimately outputting a mechanism probability distribution, which serves as input for subsequent screening of the minimum irreplaceable mechanism set and the generation of an active acquisition plan.

[0046] The mechanism knowledge base includes birth defect-related mechanism entries and image signature templates. The candidate mechanism set is determined based on the matching relationship between clinical test data and birth defect-related mechanism entries. The two-way message transmission between evidence and mechanism includes establishing the association relationship between candidate mechanisms and organ nodes between the candidate mechanism set and the organ evidence map based on the image signature template, and iteratively updating the association relationship based on the evidence weight rule to obtain the mechanism probability distribution.

[0047] In one embodiment, the mechanism knowledge base consists of birth defect-related mechanism entries and image signature templates, and supports the matching of clinical test data and the association of organ evidence maps using a unified indexing method. Birth defect-related mechanism entries use mechanism identifiers as the primary key and include at least the mechanism name, applicable gestational age range, a set of key clinical test items, reference intervals or judgment conditions for each clinical test item, a set of susceptible organs, and a set of signature entries related to imaging manifestations. The set of key clinical test items covers at least one of serological screening indicators, genetic testing conclusions, routine biochemical indicators, or infection screening conclusions, and for each item, it provides the value type, unit, effective time window, and missing value handling strategy.

[0048] The image signature template is indexed by the mechanism identifier and contains at least signature entries that can be aligned with the node attributes of the organ evidence map. Each signature entry consists of the organ name, feature name, expected deviation direction, expected deviation range, and a set of associated constraints related to the developmental constraint edge. The expected deviation direction is used to indicate the direction of deviation of the feature from the normal fetal development benchmark. The expected deviation range is used to constrain the reasonable range of the deviation. The set of associated constraints is used to indicate common mismatch patterns of bilateral symmetric constraints or anatomical associated constraints under this mechanism.

[0049] The candidate mechanism set is determined based on the matching relationship between clinical test data and birth defect-related mechanism entries. Specifically, a clinical matching score is calculated for each mechanism entry, and threshold screening and ranking truncation are performed. The calculation of the clinical matching score first standardizes the clinical test data, unifying test results from different sources into the value space defined by the mechanism knowledge base; for numerical test items, the landing point is determined according to the reference interval and the degree of deviation is calculated; for categorical test items, a matching score is given according to category consistency; for textual conclusions, a dictionary mapping is used to merge the conclusions into a preset category before calculating consistency.

[0050] A conservative strategy is adopted for missing detection items, meaning that missing items are neither awarded points nor directly rejected. However, when a missing item is a required detection item for a mechanism entry, the mechanism is marked as needing completion, and the upper limit of the clinical matching score is lowered. Based on the clinical matching scores, the top few mechanism entries are selected from high to low to form a candidate mechanism set, and the clinical matching interpretation record of each candidate mechanism is retained for the traceability of subsequent outputs.

[0051] The two-way message passing between evidence and mechanism is executed after establishing the association between candidate mechanisms and organ nodes between the candidate mechanism set and the organ evidence map. The association is established based on the image signature template: for each candidate mechanism, the set of signature entries in the image signature template is read, the organ name in the signature entry is mapped to the organ node in the organ evidence map, and the feature name in the signature entry is mapped to the corresponding residual record in the residual evidence set, thus obtaining the association list from mechanism to evidence; when the organ evidence map is missing an organ node or the corresponding feature is missing in the residual evidence set, the association is marked as unavailable and the interpretability upper limit of the candidate mechanism is lowered to avoid misjudging missing evidence as evidence of contradiction.

[0052] The iterative update employs an evidence weighting rule to weight associations: the update weight of each association is determined by both the stability score and segmentation uncertainty information. Associations with high stability scores and low organ-level uncertainty scores receive higher weights; associations with low stability scores or high uncertainty receive lower weights and are limited in their maximum impact during the update. The update process is divided into two directions: mechanism-to-evidence and evidence-to-mechanism.

[0053] The mechanism-to-evidence direction is used to generate the expected interpretation of the current residual evidence by the mechanism. Specifically, for each candidate mechanism, based on the expected deviation direction and expected deviation range given by the image signature template, the interpretation consistency of each association with the residual evidence is calculated, and the interpretation consistency is written as the interpretation message for the corresponding residual record. The calculation of interpretation consistency adopts a combination of interval consistency and direction consistency. If the direction is consistent, a point is added; if the magnitude falls within the range, a point is added; if the magnitude is far from the range, a point is deducted. The result is truncated to a preset range to avoid extreme values ​​dominating.

[0054] The evidence-to-mechanism approach is used to update the relative support of candidate mechanisms. Specifically, it summarizes the consistency of interpretation of the residual records associated with each candidate mechanism, performs a weighted summation using update weights, and then merges it with the clinical matching score. During fusion, the clinical matching score is maintained as a prior constraint, and the consistency of interpretation of residual evidence is used as a likelihood support, so that the mechanism update does not deviate from the basic boundaries given by the clinical test data.

[0055] After each update, the support of all candidate mechanisms is normalized to obtain the mechanism probability distribution, and the change in support before and after normalization is recorded. The iteration termination condition can be that the change in probability distribution is less than a threshold or the maximum number of iterations is reached. The threshold and the maximum number of iterations are preset in the system configuration. When the termination condition is met, the final mechanism probability distribution is output, along with a list of the main evidence contributions for each candidate mechanism. The list includes the organ node with the largest contribution, the corresponding feature residual record, the update weight, and the consistency, which are used for subsequent screening of irreplaceable mechanisms and evidence localization in the active collection plan.

[0056] Based on the mechanism probability distribution and residual evidence set, the minimum description length criterion and counterfactual ablation test are used to determine the minimum set of irreplaceable mechanisms, and the risk stratification results and active collection plan are output based on the minimum set of irreplaceable mechanisms.

[0057] Based on the mechanism probability distribution and residual evidence set, a combined screening of candidate mechanisms is performed. First, the minimum description length criterion is used to select interpretable and compact candidate mechanism sets. Then, counterfactual ablation tests are used to verify the irreplaceability of each mechanism in explaining the residual evidence set, thereby determining the minimum irreplaceable mechanism set. Subsequently, the minimum irreplaceable mechanism set is mapped to risk stratification results, and combined with the reference interval of the organ node with the largest deviation in the residual evidence set and the normal fetal development benchmark, an active data collection plan including section type and follow-up gestational age window is generated to guide subsequent follow-up and evidence supplementation.

[0058] The minimum description length criterion is used to select candidate mechanisms from the candidate mechanism set; the counterfactual ablation test includes removing candidate mechanisms one by one from the candidate mechanism set, and after each removal, re-performing bidirectional message passing between evidence and mechanism to obtain the probability distribution of the contrast mechanism. The minimum set of non-substitutable mechanisms is determined based on the difference between the probability distribution of the contrast mechanism and the probability distribution of the mechanism on the residual evidence set; the active acquisition plan includes the section type and the gestational age window for re-examination. The section type and the gestational age window for re-examination are determined based on the preset section type corresponding to the organ node with the largest deviation in the residual evidence set and the reference interval corresponding to the normal fetal development benchmark.

[0059] In one embodiment, the minimum description length criterion is used to select mechanism set candidates from the candidate mechanism set, and the counterfactual ablation test is used to determine the minimum set of non-substitutable mechanisms from the mechanism set candidates, and output the risk stratification results and active collection plan based on the set.

[0060] The generation of candidate mechanism sets is based on the prior ranking of mechanism probability distributions. First, mechanisms are sequenced from high to low probability distribution, and a candidate pool is formed by combining evidence coverage from clinically matched interpretation records and imaging signature templates. Then, a progressive expansion approach is used to construct multiple candidate mechanism sets: starting with a single mechanism set, the next candidate mechanism is added sequentially to form multi-mechanism combinations. For each combination, the description length cost and residual interpretation cost are calculated. The description length cost is used to constrain the size and complexity of the combinations. It can be obtained by summing the number of mechanisms and the complexity of mechanism entries. The complexity of mechanism entries can be characterized by the number of signature entries associated with that mechanism in the mechanism knowledge base and the number of key clinical tests, thus avoiding the selection of combinations containing a large number of redundant mechanisms.

[0061] The residual interpretation cost measures the extent to which the combination fails to explain the residual evidence set. It is calculated by summing the interpretation consistency of each residual record in the residual evidence set under the image signature template. The method for obtaining interpretation consistency is the same as the interval consistency and directional consistency used in the aforementioned two-way message passing between evidence and mechanisms, with a smaller weight applied to low-confidence residual records to reduce interference from unstable evidence. The total description cost is obtained by summing the description length cost and the residual interpretation cost. Several mechanism set candidates with the smallest total description cost are selected from all candidate mechanism sets for counterfactual ablation testing. To ensure feasibility and reproducibility, the calculation of the total description cost uses a fixed weight configuration and truncation rule. The weight configuration is given during system initialization and recorded in the check log during each execution. A maximum upper limit is set for the size of the mechanism set candidates; when the combination size exceeds the upper limit, it is no longer expanded to control computational load and avoid generating excessively large combinations that are difficult to interpret.

[0062] Counterfactual ablation testing is used to verify the irreplaceability of each candidate mechanism in a candidate mechanism set. For each candidate mechanism set, a removal process is performed on each candidate mechanism sequentially. Removal refers to removing the candidate mechanism from the current candidate mechanism set while keeping the remaining candidate mechanisms unchanged. After each removal process, bidirectional message passing between evidence and mechanism is re-executed to obtain the probability distribution of the contrast mechanism, and the contrastive interpretation results are calculated on the same residual evidence set. The calculation of the contrastive interpretation results includes two parts: first, the difference between the probability distribution of the contrast mechanism and the original mechanism probability distribution, which reflects how the overall probability quality is redistributed after removing a mechanism; second, the change in the consistency of interpretation of the residual evidence set, which reflects whether the interpretation of key residual records has significantly deteriorated after removing a mechanism.

[0063] The aforementioned differences are synthesized into an ablation impact score, with a higher ablation impact score indicating a less replaceable mechanism. To avoid being misled by low-confidence evidence during the ablation process, the calculation of the ablation impact score follows the evidence weighting rule, reducing the weight of residual records with low stability scores or high organ-level uncertainty scores, and setting an upper limit truncation for probability distribution differences to prevent misjudgments caused by extreme probability oscillations. Mechanisms within the candidate mechanism set are screened based on the ablation impact score, retaining those with ablation impact scores exceeding a threshold as irreplaceable mechanisms. When redundancy still exists in the retained mechanism set, the ablation test is repeated until the minimum requirement is met, meaning that removing any mechanism from the set will significantly worsen the interpretation of key deviations in the residual evidence set or cause a structural change in the mechanism probability distribution exceeding a threshold. Finally, a minimum set of irreplaceable mechanisms is obtained, and a list of key evidence contributions corresponding to this set is output. The list includes the organ node with the largest contribution, the corresponding residual record, evidence weight, and ablation impact score, facilitating the identification of key areas for supplementary collection when generating an active collection plan.

[0064] Risk stratification results are determined based on a minimum set of irreplaceable mechanisms and a mechanism probability distribution. Specifically, risk weights are extracted for each mechanism in the minimum set of irreplaceable mechanisms. These risk weights are derived from the clinical severity label and applicable gestational age range label of the mechanism in the mechanism knowledge base, and are then fused with the mechanism probability distribution to obtain a comprehensive risk score. When the comprehensive risk score falls within a preset stratification interval, the corresponding risk level is output, along with a list of organs requiring special attention and a list of recommended follow-up examinations.

[0065] The proactive data collection plan includes section type and follow-up gestational age window. Section type is determined by prioritizing the organ node with the largest deviation in the residual evidence set. The pre-defined section type corresponding to this organ node is read. If multiple section types correspond to the organ, the top few sections are selected based on their ability to cover the deviation characteristics. The follow-up gestational age window is determined based on the reference interval corresponding to the normal fetal development benchmark. Specifically, for the residual record with the largest deviation, its time sensitivity to potential regression or further deviation within the reference interval is calculated. The earliest and latest follow-up gestational ages are determined by combining gestational age information and routine clinical follow-up intervals. When the deviation direction exceeds the upper limit of the reference interval and the deviation magnitude is greater than the threshold, the follow-up gestational age window is narrowed proximally for faster verification. When the deviation magnitude is small and the evidence confidence label is moderate or higher, the follow-up gestational age window can be appropriately widened to ensure clinical feasibility. The final active data collection plan outputs a list of cross-sectional types and corresponding gestational week windows for follow-up examinations, along with the names of organ nodes and features that require key verification, ensuring that subsequent follow-up examinations can directly collect and verify the most critical residual evidence items.

[0066] In one specific embodiment, a set of reproducible sample data is used to illustrate the specific implementation process and calculation method of the present invention, with parameter ranges consistent with those of clinical fetal ultrasound examinations. The gestational age is 22 weeks and 4 days, and the clinical test data includes serological indicators and routine biochemical indicators: alpha-fetoprotein 2.6 MoM, free human chorionic gonadotropin 1.1 MoM, free estriol 0.9 MoM, and C-reactive protein 3.2 mg / L.

[0067] Ultrasound acquisition employed a multi-slice frame sequence approach. Preset slice types included four-chamber view, lateral ventricle view, bilateral renal transverse section, sagittal section of the spine, and abdominal transverse section, acquiring a total of 5 frame sequences, each approximately 120 frames. Each frame sequence was first classified by slice type, followed by frame-level quality control: frames with acoustic window obstruction, freeze artifacts, or significant defocus were marked as unusable and removed, resulting in a total of 318 image evidence frames, including 74 frames from the four-chamber view, 71 frames from the lateral ventricle view, 63 frames from the bilateral renal transverse section, 55 frames from the sagittal section of the spine, and 55 frames from the abdominal transverse section. To visually illustrate the effective evidence forms of different slice types, one frame from each slice was selected as an example from the image evidence frame set: [Example image would be inserted here]. Figure 2 As shown, Figure 2 Top left: Example of a fetal lateral ventricle cross-section ultrasound image. The lateral ventricle region is visible within the fan-shaped sound beam. The measurement direction and position are marked with an elliptical outline and caliper lines in the image. Figure 2 The upper right image is an example of a fetal four-chamber view ultrasound image. The outline outside the four-chamber area gives the approximate envelope of the heart structure, and the cross reference line indicates the location of the heart center. Figure 2 Example of a transverse ultrasound image of the fetus's kidneys in the lower left corner. The approximate area of ​​the kidneys is marked with an ellipse on both sides, and the midline reference line is used to indicate the symmetry between the left and right sides. Figure 2 Example of a sagittal section ultrasound image of the fetal spine in the lower right corner, with multiple arc-shaped markings indicating the continuous direction and visible range of the spinal segments. Subsequent organ segmentation and feature extraction are performed based on the image evidence frame set to ensure that the input evidence has a verifiable basis.

[0068] Fetal organ segmentation was performed on the image evidence frame set, covering structures such as the lateral ventricle, heart, left kidney, right kidney, and spine, and segmentation uncertainty information was generated simultaneously. The segmentation output uses a segmentation probability mask to express the probability that a pixel belongs to the target organ. The segmentation uncertainty information includes a pixel-level uncertainty map and an organ-level uncertainty score: the pixel-level uncertainty map shows band-like enhancement near the organ boundary, used to mark the fluctuation of the boundary position in different inference results; the organ-level uncertainty score is used to summarize the overall instability of the organ in the same frame or the same slice sequence. The organ-level uncertainty score is obtained by pooling the variance of the segmentation probability mask of multiple inference results; in this embodiment, five inferences were performed frame by frame on the image evidence frame set to obtain organ-level uncertainty scores: lateral ventricle 0.14, heart 0.19, and both kidneys 0.12. This score is subsequently used in evidence weighting to reduce the interference of high uncertainty evidence on mechanism judgment.

[0069] Morphological features are extracted from the segmentation probability mask and used to characterize the geometric shape and relative positional relationship of organs, including area, perimeter, major axis length, minor axis length, axis ratio, symmetry index, and relative positional index between organs. In this embodiment, the 10 frames with the highest quality scores were selected from the lateral ventricle section, and the lateral ventricle widths were obtained as follows: 11.2, 11.4, 11.6, 11.8, 11.7, 11.9, 11.5, 11.8, 12.0, and 11.7 mm, with an average of 11.76 mm. Five frames were taken from the four-chamber view, and the long axis lengths of the heart were obtained as follows: 18.6, 18.9, 19.1, 19.0, and 19.3 mm, and the short axis lengths as follows: 13.8, 13.9, 14.2, 14.0, and 14.1 mm, with an axis ratio of approximately 1.35. In the bilateral kidney cross-section, the average area of ​​the left kidney was 235 square millimeters, and that of the right kidney was 228 square millimeters, with a bilateral symmetry index of 0.97. To compensate for structural differences that may be missed when relying solely on geometric quantities, topological features based on persistent cohomology are further extracted: first, a distance transformation map is generated from a segmentation probability mask, and then persistent cohomology calculation is performed on the distance transformation map to obtain a topological feature vector, which is used to express the patterns of appearance and disappearance of connected structures and porous structures during the filtering process. In this embodiment, the length of the topological feature vectors for the lateral ventricle and cardiac regions is set to 32 dimensions, as supplementary evidence in addition to morphological features.

[0070] To suppress instability caused by common ultrasound acquisition issues such as brightness drift, gain variation, slight probe angle changes, and local occlusion, a pre-defined image transformation set was introduced, and a stability index was calculated. The pre-defined image transformation set selected four categories: brightness perturbation, gain perturbation, affine transformation, and local occlusion transformation. Key features of each organ were repeatedly calculated under each transformation condition to obtain a stability score. The stability score was characterized using a relative fluctuation form. ; in, Stability score; This represents the mean of the same feature under various transformation conditions. The standard deviation of the same feature under multiple transformation conditions; To prevent positive constants with a denominator of zero, let's take the lateral ventricle width as an example. The mean values ​​under the original and type IV transformations are 11.76, 11.70, 11.65, 11.48, and 11.31 mm, respectively, yielding a stability score of approximately 0.86; the stability score for the long axis length of the heart is approximately 0.81; and the stability score for the area of ​​both kidneys is approximately 0.77. For example... Figure 3 As shown, the stability scores of the lateral ventricles, heart, and kidneys decreased under different transformation conditions, with a more significant decrease under affine transformation and local occlusion conditions. This result was used to drive the adaptive adjustment of subsequent evidence weights: evidence with lower stability was automatically downweighted to reduce the amplification effect of "single imaging chance" on the overall inference.

[0071] Entering the evidence fusion stage, an organ evidence map is constructed based on morphological features, topological features based on persistent cohomology, segmentation uncertainty information, and stability scores, with organs as nodes. This is then combined with gestational age information to generate a normal fetal development benchmark and residual evidence set. The nodes of the organ evidence map include the lateral ventricle, heart, left kidney, right kidney, and spine, etc. Developmental constraint edges are used to express bilateral symmetry constraints and anatomical association constraints. In this embodiment, the left kidney-right kidney is set as a symmetry constraint edge, and the heart-spine and lateral ventricle-spine are set as anatomical association constraint edges. A weak association edge between the heart and lateral ventricle is also introduced to express co-occurrence associations in a clinically statistical sense. Figure 4 As shown, the node text labels correspond to the organ names, and the node size is used to express the deviation from the normal developmental benchmark. The lateral ventricle node is larger, indicating that its deviation is more significant. The left and right kidneys are connected by lines to express a symmetrical constraint relationship. The left link contains the relationship between the heart, lateral ventricle, and spine, which is used to constrain the over-expansion of the "single point anomaly explanation" in subsequent mechanism inference.

[0072] Reference intervals for normal fetal development are provided in segments according to gestational age. In this example, within the reference interval corresponding to 22 weeks and 4 days, the reference range for lateral ventricle width is 6.0 to 10.0 mm, with an observed average of 11.76 mm. The residual evidence set expresses the degree of abnormality through the direction and magnitude of deviation, with the magnitude of deviation expressed in a normalized form. ; in, For the deviation range; These are the observed values; The reference interval center value; The reference interval is half-width. Taking the width of the lateral ventricle as an example, with a reference center value of 8.0 mm and a half-width of 2.0 mm, the deviation is approximately 1.88. For the cardiac axial ratio, the reference center value is 1.25 and the half-width is 0.15; the observed value is 1.35, resulting in a deviation of approximately 0.67. For the bilateral renal symmetry index, the reference center value is 0.98 and the half-width is 0.05; the observed value is 0.97, resulting in a deviation of approximately 0.20. The deviation, along with the stability score and organ-level uncertainty score, forms the input for the evidence weight. The evidence weighting rule automatically reduces the weight of evidence with high uncertainty and low stability, making the residual evidence set more focused on "verifiable and repeatable" abnormal signals.

[0073] The candidate mechanism set is obtained by matching clinical test data with a mechanism knowledge base. The mechanism knowledge base contains birth defect-related mechanism entries and image signature templates. The image signature templates specify the interpretable association methods between mechanisms and organ evidence map nodes, such as the combination patterns of cerebrospinal fluid circulation abnormalities with lateral ventricle residuals and spinal morphology residuals, and the combination patterns of cardiac structural abnormalities with four-chamber heart structural morphology residuals and consistency of heart-spine association edges. In this embodiment, based on elevated alpha-fetoprotein and significant lateral ventricle residuals, four candidate mechanisms were screened: cerebrospinal fluid circulation abnormalities, cardiac structural abnormalities, delayed urinary tract development, and chromosomal abnormality risk. Subsequently, an association relationship is established between the candidate mechanisms and the organ evidence map, and bidirectional message passing between evidence and mechanism is executed according to the evidence weight rules, iteratively updating the mechanism probability distribution. Figure 5 As shown, the column groups present the changes in the probability of mechanisms before, after and after bidirectional transmission, and after ablation: the probability of abnormal cerebrospinal fluid circulation increases after transmission and becomes the dominant explanation; the urinary developmental delay is downweighted after transmission because the evidence of bilateral kidneys becomes less stable and more uncertain under local occlusion conditions, which is suppressed by the evidence weighting rule, thereby reducing the proportion of false positive explanations.

[0074] In the mechanism selection phase, candidate mechanisms are chosen from the candidate mechanism set according to the minimum description length criterion, and the minimum set of irreplaceable mechanisms is determined through counterfactual ablation test. In this embodiment, each candidate mechanism is removed one by one. After each removal, bidirectional message passing between evidence and mechanism is re-executed to obtain the probability distribution of the contrasting mechanisms, and the explanatory difference of the contrasting mechanism probability distribution on the residual evidence set is calculated. After removing cerebrospinal fluid circulation abnormalities, the explanatory gap for the lateral ventricle residuals is the largest, and the mechanism probability distribution shifts towards cardiac structural abnormalities and delayed urinary development. Simultaneously, the fitting error for lateral ventricle nodes increases significantly. After removing cardiac structural abnormalities, the explanatory gap for cardiac axis ratio deviation increases, but its overall impact is less than that of lateral ventricle-related mechanisms. The final determined minimum set of irreplaceable mechanisms consists of cerebrospinal fluid circulation abnormalities and cardiac structural abnormalities. The risk stratification results are: high risk for nervous system-related mechanisms, medium risk for cardiac-related mechanisms, and low risk for urinary-related mechanisms. The active acquisition plan consists of the section type and the follow-up gestational age window: it is recommended to add magnified lateral ventricle sections and cavum septum pellucidum sections at 23 to 24 weeks, and standard fetal echocardiography sections at 24 to 25 weeks; the follow-up window is determined based on the section type corresponding to the organ node with the largest deviation in the residual evidence set, and in combination with the rate of change of the reference interval of normal fetal development benchmark in this gestational age range, so as to maximize the differentiation of competitive mechanisms within a reasonable follow-up period.

[0075] like Figure 6 As shown, a clinical testing-driven intelligent diagnosis and treatment system for birth defects is used to implement a clinical testing-driven intelligent diagnosis and treatment method for birth defects. The system includes: The data acquisition module is used to acquire fetal ultrasound images, gestational age information, and clinical test data. Hardware-wise, the data acquisition module can consist of an ultrasound imaging device, a clinical information access gateway, and a time synchronization unit. The ultrasound end includes a probe, a transmitter / receiver front-end (high-voltage pulser, T / R switch, low-noise amplification and variable gain amplification), an analog-to-digital converter, and a local buffer, used to output frame sequences and necessary imaging parameters. Gestational age information and clinical test data are accessed through an HIS / LIS interface gateway or acquisition terminal. The gateway has wired Ethernet / Wi-Fi / 4G communication, data format parsing, and local encrypted storage capabilities. The time synchronization unit (RTC / PTP / NTP) is used to timestamp multi-source data and align data within the same examination session.

[0076] The feature generation module is used for fetal organ segmentation based on fetal ultrasound images to generate segmentation uncertainty information, extract morphological features and topological features based on persistent homology, and calculate a stability index set based on a preset set of image transformations. The feature generation module is typically deployed on an edge computing unit or server accelerator card, with its core consisting of a CPU + GPU / NPU / FPGA coprocessor, equipped with high-speed memory and video memory to support batch and multiple inferences for organ segmentation. The module takes an ultrasound frame sequence as input, and performs image preprocessing pipelines (denoising, normalization, scaling) and parallel inference on the hardware side. It outputs a segmentation probability mask through multiple inferences and performs uncertainty calculations such as variance / entropy on the accelerator. Simultaneously, morphological features are calculated on the CPU side or in the GPU kernel function. Distance transformations and topological vector calculations related to persistent homology can be performed in parallel on the GPU or accelerated by the FPGA to reduce latency. Batch processing of the image transformation set is executed by the accelerator to obtain the stability index set and write it back to shared memory.

[0077] The evidence graphing module is used to construct an organ evidence graph with organs as nodes based on morphological features, topological features based on continuous cohomology, segmentation uncertainty information, and stability index sets. It also generates a residual evidence set based on the normal fetal development benchmark corresponding to gestational age information. Hardware-wise, the evidence graphing module can consist of a CPU-driven graph computing unit and a memory subsystem, with the option to use a GPU for matrix operations acceleration when necessary. This module receives the morphological, topological, uncertainty, and stability indexes from the preceding output. The CPU constructs a graph structure (sparse adjacency list / CSR, etc.) of organ nodes and developmental constraint edges in memory, and generates residual evidence based on the developmental benchmark corresponding to gestational age (local database or read-only cache). To ensure throughput, an NVMe solid-state drive can be configured to store the developmental benchmark and historical statistical parameters, and random access efficiency can be improved through memory mapping or caching strategies. In multi-case parallel processing, a multi-core CPU thread pool can be used to complete node / edge updates and residual calculations.

[0078] The mechanism inference module is used to determine the set of candidate mechanisms based on clinical test data and a mechanism knowledge base. It then uses evidence weighting rules based on a stability index set to perform bidirectional message passing between evidence and mechanisms to obtain the mechanism probability distribution. The mechanism inference module can be implemented in hardware as a combination of a graph reasoning / probability inference acceleration unit and a mechanism knowledge base storage unit. Candidate mechanism matching can be performed on the CPU for retrieval and filtering (inverted index, vector index, or rule matching), and the association between candidate mechanisms and organ evidence graphs is loaded into GPU memory / RAM. Iterative updates of bidirectional message passing typically involve a large number of vector / matrix and graph neighborhood aggregation operations, which can be performed in parallel iteration by GPU / NPU or implemented with low-latency pipelined updates by FPGA. The mechanism knowledge base can be stored on a local SSD and kept hot-cache in memory to ensure fast reading of image signature templates and mechanism entries. Simultaneously, a security module (TPM / TEE) performs integrity verification and access control on the knowledge base and inference results.

[0079] The results output module is used to determine the minimum set of irreplaceable mechanisms based on the mechanism probability distribution and residual evidence set, using the minimum description length criterion and counterfactual ablation test. Based on this set, it outputs risk stratification results and an active data collection plan. The hardware of the results output module consists of a display and interactive terminal, a report generation and printing interface, and an external system communication interface. This module performs strategy control for the minimum description length criterion and counterfactual ablation test on the CPU (scheduling multiple inferences, recording and comparing probability distributions, and summarizing the set of irreplaceable mechanisms), and encodes the risk stratification results and active data collection plan into structured data. The output side can connect to a doctor's workstation monitor, touch input device, and voice input peripheral to support review and confirmation. It also provides DICOM / HL7 / FHIR or intra-hospital private network API interfaces to transmit results back to PACS / HIS. To meet audit and compliance requirements, the module is equipped with local encrypted storage for retaining versioned reports and key intermediate data, and an optional hardware encryption card can be added to accelerate transmission and storage encryption.

[0080] The above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for intelligent diagnosis and prevention of birth defects based on clinical testing, characterized in that, The method includes: To obtain fetal ultrasound images, gestational age information, and clinical test data; Fetal organ segmentation is performed based on the fetal ultrasound images to generate segmentation uncertainty information, extract morphological features and topological features based on continuous homology, and calculate a set of stability indices based on a preset set of image transformations. Based on the morphological features, the topological features based on continuous homology, the segmentation uncertainty information, and the stability index set, an organ evidence map with organs as nodes is constructed, and a residual evidence set is generated based on the normal fetal development benchmark corresponding to the gestational age information. Based on the clinical test data and mechanism knowledge base, a set of candidate mechanisms is determined, and evidence and mechanism bidirectional message passing is performed using evidence weight rules based on the stability index set to obtain the mechanism probability distribution. Based on the probability distribution of the mechanism and the residual evidence set, the minimum description length criterion and counterfactual ablation test are used to determine the minimum set of irreplaceable mechanisms, and the risk stratification results and active collection plan are output based on the minimum set of irreplaceable mechanisms.

2. The method according to claim 1, characterized in that, The process of acquiring fetal ultrasound images includes acquiring ultrasound frame sequences corresponding to multiple preset section types, and performing section classification and frame-level quality control on the ultrasound frame sequences to remove unusable frames, thereby obtaining an image evidence frame set. The fetal organ segmentation is performed based on the set of image evidence frames.

3. The method according to claim 2, characterized in that, The fetal organ segmentation generates a segmentation probability mask, and the segmentation uncertainty information includes a pixel-level uncertainty map and an organ-level uncertainty score. The organ-level uncertainty score is determined based on the variance of a segmentation probability mask obtained by performing multiple inferences on the image evidence frame set.

4. The method according to claim 3, characterized in that, The morphological features are determined based on the segmentation probability mask, and the morphological features include at least one of the following: area, perimeter, major axis length, minor axis length, axis ratio, symmetry index, and relative position index between organs.

5. The method according to claim 3, characterized in that, The topological features based on continuous homology include: A distance transformation map is generated based on the segmentation probability mask, and a continuous cohomology calculation is performed based on the distance transformation map to obtain a topological feature vector as the topological feature based on continuous cohomology.

6. The method according to claim 1, characterized in that, The preset image transformation set includes at least two of the following: brightness perturbation, gain perturbation, affine transformation, and local occlusion transformation; The stability index set includes a stability score, which is determined based on the relative fluctuations of the morphological features and the topological features based on continuous coherence under the preset image transformation set. The evidence weighting rule is determined based on the stability score and the segmentation uncertainty information, and is used to weight the evidence in the organ evidence map to perform the two-way message passing between the evidence and the mechanism.

7. The method according to claim 1, characterized in that, The organ evidence map includes organ nodes and developmental constraint edges, which are used to characterize bilateral symmetric constraint relationships or anatomical association constraint relationships. The normal fetal development benchmark provides a reference interval corresponding to the gestational age information based on the gestational age information; The residual evidence set includes the morphological features and the deviation direction and magnitude of the topological features based on continuous coherence relative to the reference interval.

8. The method according to claim 1, characterized in that, The mechanism knowledge base includes birth defect-related mechanism entries and image signature templates, and the candidate mechanism set is determined based on the matching relationship between the clinical test data and the birth defect-related mechanism entries; The two-way message transmission between evidence and mechanism includes establishing an association between candidate mechanisms and organ nodes between the candidate mechanism set and the organ evidence map based on the image signature template, and iteratively updating the association based on the evidence weight rules to obtain the probability distribution of the mechanism.

9. The method according to claim 1, characterized in that, The minimum description length criterion is used to select candidate mechanisms from the candidate mechanism set. The counterfactual ablation test includes performing a removal process on each candidate mechanism in the candidate mechanism set, and re-executing the evidence and mechanism bidirectional message passing after each removal process to obtain the comparison mechanism probability distribution, and determining the minimum set of irreplaceable mechanisms based on the difference between the comparison mechanism probability distribution and the mechanism probability distribution on the residual evidence set. The active data collection plan includes a section type and a follow-up gestational age window. The section type and the follow-up gestational age window are determined based on the preset section type corresponding to the organ node with the largest deviation in the residual evidence set and the reference interval corresponding to the normal fetal development benchmark.

10. A clinical testing-driven intelligent diagnosis and treatment system for birth defects, used to implement the clinical testing-driven intelligent diagnosis and treatment system for birth defects as described in any one of claims 1-9, characterized in that, The system includes: The data acquisition module is used to acquire fetal ultrasound images, gestational age information, and clinical test data; The feature generation module is used to perform fetal organ segmentation based on the fetal ultrasound image to generate segmentation uncertainty information, extract morphological features and topological features based on continuous homology, and calculate a set of stability indices based on a preset set of image transformations. The evidence mapping module is used to construct an organ evidence map with organs as nodes based on the morphological features, the topological features based on continuous homology, the segmentation uncertainty information and the stability index set, and to generate a residual evidence set based on the normal fetal development benchmark corresponding to the gestational age information. The mechanism inference module is used to determine a set of candidate mechanisms based on the clinical test data and the mechanism knowledge base, and to perform two-way message passing between evidence and mechanism using evidence weight rules based on the stability index set to obtain the mechanism probability distribution. The results output module is used to determine the minimum set of irreplaceable mechanisms based on the probability distribution of the mechanism and the residual evidence set, using the minimum description length criterion and counterfactual ablation test, and to output the risk stratification results and active collection plan based on the minimum set of irreplaceable mechanisms.

Images