A medical image feature conduction recognition method and system

Abstract

The application discloses a medical image feature conduction identification method and system, and relates to the technical field of medical image identification. By calling and analyzing the defect knowledge graph built in a large model, an image feature conduction benchmark is generated, then target medical image data is extracted and a hierarchical correspondence relationship with the image feature conduction benchmark is established, an adaptive feature set is output, a bidirectional knowledge conduction link is called for correlation operation to generate reinforced correlation features, the potential relationship between the features is further mined, the expressiveness and distinguishability of the features are enhanced, key areas are divided based on the reinforced correlation features, and a defect feature conduction link covering the whole area is constructed, so that the defect features in different areas and the conduction relationship can be accurately identified, and finally, a structured image identification report is integrated according to the medical image report specification, which is helpful to improve the accuracy, consistency and efficiency of image identification.

Description

Technical Field

[0001] This invention relates to the field of medical image recognition technology, and more specifically, to a method and system for medical image feature transmission and recognition. Background Technology

[0002] In the medical field, the early and accurate identification of birth defects in children is crucial for protecting children's health and developing timely and effective interventions. Medical image recognition is one of the important means of diagnosing birth defects; however, current traditional medical image recognition methods have many limitations.

[0003] On the one hand, current image recognition mainly relies on doctors' professional experience and subjective judgment. Different doctors may have different understandings and judgments of image features, making it difficult to guarantee the consistency and accuracy of diagnostic results. Especially when faced with complex and ever-changing images of children with birth defects, doctors may misdiagnose or miss diagnoses due to factors such as lack of experience or fatigue.

[0004] On the other hand, traditional image recognition methods lack a systematic integration and in-depth exploration of knowledge about birth defects in children. Birth defects involve various types and complex anatomical changes. Existing methods often only analyze features in images in isolation, without combining these features with the overall knowledge system of birth defects, making it difficult to comprehensively and accurately identify the type and severity of defects. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for medical image feature transmission and recognition, the method comprising:

[0006] The defect knowledge graph built into the large model is retrieved, and the association hierarchy and transmission logic of defect feature nodes and anatomical structure nodes in the defect knowledge graph are analyzed to generate image feature transmission benchmarks.

[0007] Extract the anatomical structure distribution data and pixel feature matrix of the target medical image, perform feature transfer adaptation operation based on the anatomical structure mapping rules, establish the hierarchical correspondence between the target medical image features and the image feature transfer benchmark, and output the adapted feature set;

[0008] The bidirectional knowledge transmission link of the large model is invoked, and each feature node in the adaptation feature set is connected to the starting end of the bidirectional knowledge transmission link. It is then associated with the upstream defect features, downstream defect features, and parallel defect features in the implementation defect knowledge graph to generate reinforced association features.

[0009] Based on the anatomical localization information of enhanced correlation features, key regions of the target medical image are divided, feature transmission sub-links are constructed within the key regions, and all feature transmission sub-links are connected by cross-regional fusion feature nodes to form a defect feature transmission link covering the entire domain.

[0010] Extract node association data, transmission path information, and defect matching results from the defect feature transmission chain, and integrate them into a structured image recognition report in accordance with medical image reporting standards. The structured image recognition report includes feature transmission atlas, defect association description, and anatomical location annotation.

[0011] Furthermore, embodiments of the present invention also provide a medical image feature transmission and recognition system, comprising:

[0012] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to perform the above-described medical image feature transduction and recognition method by executing the machine-executable instructions.

[0013] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, the processor of the medical image feature transmission and recognition system reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the medical image feature transmission and recognition system to perform the above-described medical image feature transmission and recognition method.

[0014] Based on the above, by retrieving and parsing the defect knowledge graph built into the large model, an image feature transmission benchmark is generated. Then, target medical image data is extracted and a hierarchical correspondence with the image feature transmission benchmark is established, outputting an adapted feature set. This captures key features related to birth defects in images. By invoking the bidirectional knowledge transmission link for association operations to generate enhanced association features, the potential connections between features are further explored, enhancing the expressiveness and discriminative power of the features. Based on the enhanced association features, key regions are divided and a full-domain defect feature transmission link is constructed, accurately identifying defect features and their transmission relationships in different regions. Finally, a structured image recognition report is generated according to medical image reporting standards, which helps improve the accuracy, consistency, and efficiency of image recognition for birth defects in children. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the execution flow of the medical image feature transmission and recognition method provided in the embodiments of the present invention;

[0016] Figure 2This is a flowchart illustrating the process of retrieving the built-in defect knowledge graph of a large model, analyzing the association hierarchy and transmission logic of defect feature nodes and anatomical structure nodes in the defect knowledge graph, and generating an image feature transmission benchmark, as provided in this embodiment of the invention.

[0017] Figure 3 This is a flowchart illustrating the bidirectional knowledge transmission link for calling a large model provided in this embodiment of the invention. Each feature node in the adaptation feature set is connected to the starting end of the bidirectional knowledge transmission link and associated with the upstream defect features, downstream defect features, and parallel defect features in the defect knowledge graph to generate reinforced associated features.

[0018] Figure 4 This is a schematic diagram of exemplary hardware and software components of the medical image feature transmission and recognition system provided in an embodiment of the present invention. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a medical image feature transmission and recognition method according to an embodiment of the present invention. The following is a detailed description of the medical image feature transmission and recognition method.

[0020] Step S110: Retrieve the defect knowledge graph built into the large model, analyze the association hierarchy and transmission logic of defect feature nodes and anatomical structure nodes in the defect knowledge graph, and generate image feature transmission benchmarks.

[0021] In this embodiment, ventricular septal defect, a type of congenital heart disease in children, will be used as the unified application scenario throughout the text. The large model is specifically a large model of birth defects.

[0022] For details, please refer to the following: Figure 2 Step S111: Retrieve the defect knowledge graph built into the large model. The defect knowledge graph stores the anatomical structural features, imaging features, and defect association features corresponding to various birth defects in children. All features exist in the form of independent nodes and the nodes are connected by the labeled association relationship.

[0023] For example, in the ventricular septal defect (VSD) scenario, the defect knowledge graph includes the anatomical features corresponding to the VSD, such as the normal thickness range and spatial relationship of the membranous and muscular portions of the VSD; imaging features, such as the size and shape of the echo loss area on echocardiography, and the direction and velocity of the shunt signal displayed by color Doppler; and defect-related features, such as increased pulmonary artery pressure, right ventricular wall thickening, and left ventricular enlargement. All features exist as independent nodes. For example, the "echo loss in the membranous portion of the VSD" node is connected to the "left-to-right shunt signal" node through a "direct display" association, and the "left-to-right shunt signal" node is connected to the "increased right ventricular volume overload" node through a "hemodynamic changes leading to" association. The associations between nodes are all labeled based on clinical pathological mechanisms.

[0024] Step S112: Extract the core image feature nodes corresponding to all defect types in the defect knowledge graph. Each core image feature node contains multiple information dimensions, including feature dimensions, manifestation form, anatomical location, and pathological correlation basis. Each information dimension is supported by specific clinical case data.

[0025] For ventricular septal defects (VSDs), the core imaging features extracted include "discontinuity of the ventricular septum," "transseptal blood flow signal," and "right ventricular enlargement." Taking the "discontinuity of the ventricular septum" feature as an example, its characteristic dimension is the structural morphological features in ultrasound images; its manifestation is that in multiple standard ultrasound sections (such as the parasternal long-axis section of the left ventricle, the apical four-chamber view, and the short-axis section of the great arteries), there is a local interruption of echo continuity in the ventricular septum, and the echo at the interruption ends may be enhanced or irregular; the anatomical location is divided into membranous part, muscular part (inflow tract, trabecular part, outflow tract), and atrioventricular canal type, according to the specific location of the interruption site in the ventricular septum; the pathological correlation is based on the incomplete fusion or excessive absorption of the corresponding part of the ventricular septum during embryonic development. Each information dimension corresponds to hundreds of clinical imaging cases of surgically confirmed ventricular septal defects, and the case data includes imaging manifestations of different defect sizes, types, and associated malformations.

[0026] Step S113: Analyze the transmission logic between core image feature nodes, and mark the order of appearance or the associated occurrence of different core image feature nodes in the defect evolution. The transmission logic is formed based on the pathological mechanism of defect occurrence and the time sequence of development process.

[0027] When analyzing the transmission logic between the core imaging features of ventricular septal defects (VSDs), we rely on their pathological development mechanisms. After a VSD develops, the first core imaging feature is "discontinuity of the ventricular septum," which is the morphological basis. Because the left ventricular pressure is higher than the right ventricular pressure, a "left-to-right shunt signal" appears, a direct manifestation of hemodynamic changes. Prolonged left-to-right shunt leads to increased right ventricular volume overload, resulting in "right ventricular enlargement." After right ventricular enlargement, pulmonary blood flow increases, gradually triggering "increased pulmonary artery pressure." When pulmonary artery pressure reaches a certain level, "right ventricular wall thickening" can occur. These features exhibit a clear sequential order of appearance in the defect's progression. Simultaneously, there are also accompanying correlations; for example, a significant "discontinuity of the ventricular septum" is often accompanied by "left ventricular enlargement," indicating a co-occurrence relationship, both caused by the increased volume overload resulting from the shunt.

[0028] Step S1131: Retrieve the clinical case database stored in the large model. The clinical case database contains confirmed cases of birth defects in children and corresponding diagnostic reports. Each case of birth defects in children records detailed data on the imaging manifestations of the birth defects and the timeline of their development.

[0029] In the case of ventricular septal defect (VSD), the retrieved clinical case database contains thousands of imaging cases of VSD diagnosed by echocardiography and confirmed by surgery or cardiac catheterization. The diagnostic report for each case records in detail the child's basic information (such as age and gender), ultrasound examination time, imaging data of each examination (such as the location, size, shunt velocity, size of each chamber of the heart, pulmonary artery pressure, etc.), and time points of defect development (such as the time of first defect discovery, the time of defect size change, and the time of complication onset).

[0030] Step S1132: Based on the matching parameters between the feature description of the core image feature node and the image performance data of the clinical case, select cases related to the core image feature node from the clinical case database. Each core image feature node corresponds to multiple clinical cases.

[0031] Taking the core imaging feature node "ventricular septal discontinuity interruption" as an example, its feature description (local echo discontinuity interruption of the ventricular septum) is matched with clinical case imaging data (echoic patterns of the ventricular septum region in ultrasound images). By calculating the similarity between key parameters in the feature description (such as the location and extent of the interruption) and corresponding parameters in the case imaging data, cases with matching parameters higher than a set threshold are selected. Each core imaging feature node corresponds to multiple clinical cases; for example, the "ventricular septal discontinuity interruption" node can filter out hundreds of clinical cases covering different defect locations and sizes.

[0032] Step S1133: Analyze the occurrence order of core imaging feature nodes in each case, record the timestamp or imaging layer identifier of the first occurrence of different core imaging feature nodes in each clinical case, form a list of node occurrence order, and use each possible combination of order as a statistical unit to count the frequency value of the occurrence order of core imaging feature nodes in all relevant clinical cases, forming a frequency table of order combinations.

[0033] Each selected clinical case of ventricular septal defect was analyzed individually. For example, in one case, the initial ultrasound examination (timestamp T1) revealed "discontinuity of the ventricular septum" in the parasternal long-axis view of the left ventricle (image plane label A). Color Doppler ultrasound in the same examination showed "left-to-right shunt signal" (image plane label A). A follow-up examination three months later (timestamp T2) revealed "right ventricular enlargement" (image plane label B). A follow-up examination six months later (timestamp T3) showed "elevated pulmonary artery pressure" (image plane label C). The timestamps and corresponding image plane labels of the first appearance of these core imaging features were recorded to form a list of the order of appearance of these features in the case: ventricular septal discontinuity (T1, A) → left-to-right shunt signal (T1, A) → right ventricular enlargement (T2, B) → elevated pulmonary artery pressure (T3, C). The list of node occurrence sequences for all cases is compiled, and each possible sequence combination (such as the above sequence combination, or other possible combinations such as interventricular septal discontinuity interruption → right ventricular enlargement → left-to-right shunt signal, etc.) is used as the statistical unit. The number of times each sequence combination appears in all relevant clinical cases is counted to form a sequence combination frequency table.

[0034] Step S1134: Retrieve pathological mechanism research data from the built-in knowledge base of the large model. The pathological mechanism research data includes the molecular biological mechanisms of defect occurrence and development, and the anatomical evolution law.

[0035] For ventricular septal defects, the retrieved pathological mechanism research data include the formation time of each part of the ventricular septum (membranous part and muscular part) during embryonic development, the cell populations involved, and the signal regulatory pathways; the molecular biological mechanisms of abnormal proliferation and differentiation of cardiomyocytes during ventricular septal defects; the pathophysiological process of left-to-right shunt leading to increased pulmonary blood flow and consequently pulmonary hypertension, including anatomical evolution patterns such as changes in pulmonary vascular endothelial cell function, vascular smooth muscle cell proliferation, and extracellular matrix remodeling; and the myocardial remodeling mechanism of long-term volume overload leading to ventricular wall hypertrophy and cardiac chamber enlargement.

[0036] Step S1135: Combining the frequency table of sequential combinations and pathological mechanism research data, verify the rationality of the occurrence order of the core imaging feature nodes, and determine the main transmission logic between the core imaging feature nodes. The main transmission logic is the sequential combination of nodes that have high-frequency occurrence records in the clinical case database and conform to the pathological mechanism.

[0037] The sequence combinations that appeared most frequently in the frequency table (e.g., ventricular septal discontinuity interruption → left-to-right shunt signal → right ventricular enlargement → increased pulmonary artery pressure) were compared and validated with pathological mechanism research data. Pathologically, ventricular septal discontinuity interruption is the structural basis and must precede the shunt signal; shunt leads to increased right ventricular volume load, and under long-term effects, right ventricular enlargement, consistent with hemodynamic principles; after right ventricular enlargement, pulmonary artery blood flow continues to increase, and pulmonary artery pressure gradually rises, consistent with the formation mechanism of pulmonary hypertension. This sequence combination appeared most frequently in the frequency table and fully conforms to the pathological mechanism; therefore, it was identified as the main transmission logic between core imaging feature nodes of ventricular septal defects.

[0038] Step S1136: Identify the secondary transmission logic between core image feature nodes, wherein the secondary transmission logic is a combination of nodes whose frequency value percentage does not meet the clinical statistical standard but is supported by a pathological mechanism or appears in specific labeled cases.

[0039] In the sequence combination frequency table, there may be some sequence combinations with a low frequency percentage (e.g., less than 20% of the primary conduction logic frequency) but supported by a pathological mechanism. For example, "interruption of ventricular septal continuity → right ventricular enlargement → left-to-right shunt signal"—this sequence combination may appear in cases where the defect is small and the shunt is not clearly shown by color Doppler in the early stages, but the cumulative effect of the shunt has already led to right ventricular enlargement. Although the frequency is low, it has its own specific pathological mechanism (e.g., instrument resolution limitations, low shunt velocity, etc.), and is therefore identified as a secondary conduction logic. In addition, for certain specially labeled cases (e.g., ventricular septal defects combined with other cardiac malformations), special node sequence combinations appearing, if there is a clear pathological mechanism explanation, are also identified as secondary conduction logics.

[0040] Step S1137: Label the trigger condition parameters for each conduction logic. The trigger condition parameters include factors that affect the selection of conduction paths, specifically including defect type coding, child age range, and anatomical structure status parameters.

[0041] Triggering parameters are labeled for primary and secondary conduction logics. For the primary conduction logic of ventricular septal defect (VSD), the triggering parameters include the defect type code (simple VSD, without other associated malformations), the child's age range (newborns to preschool children, during which pulmonary artery pressure gradually decreases and shunt is significant), and anatomical structure parameters (VSD diameter greater than a certain value, significant pressure gradient between the left and right ventricles). For secondary conduction logic, such as the above "interruption of VSD continuity → right ventricular enlargement → left-to-right shunt signal," the triggering parameters include the defect type code (VSD combined with mild pulmonary valve stenosis, reduced shunt velocity), the child's age range (early infancy, before pulmonary artery pressure has significantly decreased), and anatomical structure parameters (VSD location is special, situated in the trabecular region of the muscle, making ultrasound visualization difficult).

[0042] Step S1138: Mark the primary conduction logic as the priority conduction path and the secondary conduction logic as the alternative conduction path. In this way, organize all the marked conduction logics to form a list of conduction logics for core image feature nodes. The list of conduction logics includes the node combination of the conduction logic, the frequency value, the pathological basis summary, the trigger condition parameters, and the path type identifier.

[0043] Primary conduction logic is marked as the preferred conduction path, and its path type is identified as "P" in the conduction logic list; secondary conduction logic is marked as the alternative conduction path, and its path type is identified as "A". The specific content of the conduction logic list includes: for primary conduction logic, the node combination is "interruption of ventricular septal continuity → left-to-right shunt signal → right ventricular enlargement → increased pulmonary artery pressure", the frequency value is the number of specific cases obtained from statistics, the pathological basis summary is "the pathophysiological mechanism of left-to-right shunt caused by ventricular septal defect, resulting in increased right ventricular volume load and pulmonary hypertension", the trigger condition parameter is the trigger condition of the above primary conduction logic, and the path type is identified as "P". For each secondary conduction logic, the node combination, frequency value, pathological basis summary (a description of the specific pathological mechanism for this secondary logic), trigger condition parameter, and path type are recorded in the same format, thus forming a complete core imaging feature node conduction logic list.

[0044] Step S114: Based on the frequency of occurrence of this transmission logic recorded in the clinical case database, prioritize the transmission logic of each core imaging feature node, and determine the ranking result by combining the supporting conclusions in the literature on defect pathological mechanisms.

[0045] In the ventricular septal defect (VSD) scenario, for each core imaging feature node (e.g., "left-to-right shunt signal"), all related conduction logics (including those acting as the starting, intermediate, or terminating node) are collected. These conduction logics are initially ranked based on their frequency of occurrence recorded in the clinical case database, with higher-frequency logics having higher initial priority. For example, the conduction logic "left-to-right shunt signal → right ventricular enlargement" occurs much more frequently than "left-to-right shunt signal → left ventricular systolic dysfunction," so the former is initially prioritized. Then, literature on the pathological mechanism of the defect is searched to examine the supporting conclusions for each conduction logic. For "left-to-right shunt signal → right ventricular enlargement," numerous studies confirm that left-to-right shunts lead to increased right ventricular volume overload, thus causing right ventricular enlargement, providing strong support. However, "left-to-right shunt signal → left ventricular systolic dysfunction" is less common in the early stages of VSD, and the literature support is weaker. Combining the frequency ranking and the literature support, the final priority ranking of the conduction logics for each core imaging feature node is determined.

[0046] Step S115: Based on the sorted transmission logic, construct the transmission path template of the core image feature nodes. Each transmission path template corresponds to the feature transmission law of a type of defect. The transmission path template includes the starting node identifier, the intermediate node sequence, the ending node identifier, and the transmission direction parameters between nodes.

[0047] Based on the ordered conduction logic, conduction pathway templates are constructed for ventricular septal defects (VSDs). A primary template is built using the main conduction logic, while alternative templates are built using secondary conduction logic. The primary template's starting node is labeled "interruption of VSD continuity," the intermediate node sequence is "left-to-right shunt signal → right ventricular enlargement," and the ending node is labeled "increased pulmonary artery pressure." Conduction direction parameters between nodes are labeled as unidirectional or bidirectional. For example, the conduction from "interruption of VSD continuity" to "left-to-right shunt signal" is unidirectional (only the former can trigger the latter), and the conduction between "left-to-right shunt signal" and "right ventricular enlargement" is unidirectional (the shunt causes enlargement). Conduction direction parameters also include time-dependent characteristics, such as the "left-to-right shunt signal" triggering "right ventricular enlargement" only after a certain duration, labeled as time-delayed conduction. Each conduction pathway template clearly corresponds to the characteristic conduction patterns of VSDs.

[0048] Step S116: Extract the anatomical structure nodes that are directly associated with the core image feature nodes in the defect knowledge graph, establish a unique mapping relationship between the core image feature nodes and the anatomical structure nodes, and label the specific anatomical location coordinates and anatomical range boundaries corresponding to each core image feature node.

[0049] From the defect knowledge graph, extract anatomical structure nodes directly associated with the core imaging feature nodes of ventricular septal defect (VSD), such as "ventricular septum," "right ventricle," and "pulmonary artery." Establish a unique mapping relationship between the core imaging feature nodes and the anatomical structure nodes; for example, the node "discontinuity of ventricular septum" uniquely maps to the "ventricular septum" anatomical structure node, and the node "right ventricular enlargement" uniquely maps to the "right ventricle" anatomical structure node. Label the specific anatomical location coordinates corresponding to each core imaging feature node using a three-dimensional coordinate system (with the heart center as the origin, and the X, Y, and Z axes corresponding to the long axis, anterior-posterior axis, and left-right axis of the human body, respectively). For example, the anatomical location coordinates of "discontinuity of the membranous part of the ventricular septum" are the region from (X1, Y1, Z1) to (X2, Y2, Z2). Simultaneously label the anatomical boundaries; for example, the boundary of the membranous ventricular septum extends superiorly to the aortic valve annulus, inferiorly to the supraventricular crest, anteriorly to the anterior part of the ventricular septum, and posteriorly to the atrioventricular node region.

[0050] Step S117: Embed the three-dimensional spatial location information of the anatomical structure nodes into the transmission path template, add spatial dimension attributes to each transmission path template, and annotate the adjacency relationship, overlap range and distance parameters of the anatomical position corresponding to the core image feature node in three-dimensional space.

[0051] The three-dimensional spatial location information (such as the three-dimensional coordinate range and volume of each anatomical structure) of nodes such as "ventricular septum," "right ventricle," and "pulmonary artery" is embedded into the conduction path template. Spatial dimension attributes are added to the main conduction path template for ventricular septal defects. For example, the anatomical location corresponding to "interruption of ventricular septal continuity" (the membranous part of the ventricular septum) and the anatomical location of "right ventricle" are directly adjacent, with zero overlap (they are anatomically closely connected but do not overlap), and the distance parameter is the shortest distance from the inner surface of the membranous part of the ventricular septum to the inner wall of the right ventricle, which is zero. The adjacency relationship between "right ventricle" and "pulmonary artery" is that they are connected through the pulmonary valve, with the overlap being the pulmonary valve annulus region, and the distance parameter being the distance from the end of the right ventricular outflow tract to the beginning of the pulmonary artery. These spatial dimension attributes enable the conduction path template to include not only logical conduction relationships but also spatial location associations.

[0052] Step S118: Standardize the transmission path templates with embedded spatial attributes, integrate all standardized transmission path templates, and construct an initial image feature transmission benchmark framework. The initial image feature transmission benchmark framework includes feature transmission paths, spatial correlation data, and pathological mechanism descriptions for various defects.

[0053] The conduction pathway template for ventricular septal defects (VSDs) with embedded spatial attributes is standardized by unifying node identifier formats (e.g., using international medical terminology standard coding), spatial coordinate units (e.g., millimeters), conduction direction parameter representation methods (e.g., arrow symbols combined with textual descriptions), and correlation description standards (e.g., using unified medical relational terminology). The standardized main conduction pathway template and alternative conduction pathway templates are integrated to construct an initial imaging feature conduction baseline framework. This framework includes not only the characteristic conduction pathway of VSDs but also spatial correlation data for each core imaging feature node (e.g., adjacent structures, distance parameters), pathological mechanism descriptions (e.g., a detailed description of the hemodynamic mechanism of left-to-right shunt leading to pulmonary hypertension), while also reserving framework positions for other types of childhood birth defects, forming a baseline framework containing basic conduction information for multiple defect types.

[0054] Step S119: Prioritize and filter the conduction paths in the initial image feature conduction benchmark framework. Based on the data completeness and pathological mechanism clarity of the clinical cases associated with the conduction paths, distinguish between core conduction paths and reference conduction paths, and generate an image feature conduction benchmark containing hierarchical paths.

[0055] The conduction pathways (primary and alternative pathways) of ventricular septal defects (VSDs) in the initial imaging feature conduction benchmark framework were prioritized and screened. The data completeness of the clinical cases associated with each pathway was assessed, including the number of cases, the level of detail in the feature node records, and the completeness of follow-up data. The clarity of the pathological mechanism was also assessed, including whether there was sufficient literature support, whether the pathophysiological process was clearly defined, and whether there was any controversy. Pathways with high data completeness and high clarity of pathological mechanisms (such as the primary conduction pathway of VSDs) were identified as core pathways and assigned the highest priority. Pathways with moderate data completeness or moderate clarity of pathological mechanisms, but still possessing clinical reference value (such as some secondary pathways), were identified as reference pathways and assigned lower priority. Through the above differentiation and prioritization, a hierarchical imaging feature conduction benchmark including core and reference pathways was generated. The core pathways served as the primary basis for analysis, while the reference pathways served as supplementary auxiliary pathways.

[0056] Step S120: Extract the anatomical structure distribution data and pixel feature matrix of the target medical image, perform feature transfer adaptation operation based on the anatomical structure mapping rules, establish the hierarchical correspondence between the target medical image features and the image feature transfer benchmark, and output the adapted feature set.

[0057] In the context of ventricular septal defect, the target medical image is the pediatric echocardiogram to be analyzed. This step aims to extract key data from the target image and adapt it to the image feature transmission benchmark to establish a correspondence.

[0058] Step S121: Extract the overall anatomical structure distribution data of the target medical image through an image segmentation algorithm. The overall anatomical structure distribution data covers the names, three-dimensional spatial locations, morphological parameters, and spatial relationships of all identifiable anatomical structures in the image. It is obtained by scanning the image region by region and matching it with the anatomical structure feature library.

[0059] An image segmentation algorithm combining region growing and edge detection is employed for echocardiographic images of ventricular septal defects (VSDs), such as 3D echocardiographic volumetric data. First, initial seed points (e.g., the left ventricular cavity) are set, and region growing is performed based on pixel grayscale similarity and spatial continuity to initially segment the left ventricular cavity. Then, the image is scanned region by region, expanding outward from the left ventricular region to sequentially segment identifiable anatomical structures such as the interventricular septum, right ventricle, left atrium, right atrium, aorta, and pulmonary artery. Each segmented anatomical structure is matched against an anatomical structure feature database (containing typical grayscale features, morphological features, and texture feature templates for various cardiac structures) to determine the anatomical structure name. The three-dimensional spatial location is determined by calculating the three-dimensional coordinate range (minimum and maximum XYZ coordinate values) of the segmented region; morphological parameters are extracted, such as the major and minor diameters and volumes of each heart chamber, the thickness distribution of the interventricular septum, and the opening amplitude of the valves; the spatial relationships between various anatomical structures are analyzed, such as the positional relationship between the left ventricle and the interventricular septum (the left ventricle is located to the left of the interventricular septum), and the connection relationship between the aorta and the left ventricle (the root of the aorta is connected to the outflow tract of the left ventricle), and this information is integrated to form overall anatomical structure distribution data.

[0060] Step S122: Generate a pixel feature matrix of the target medical image using pixel matrix parsing technology. Decompose the pixel feature matrix into multiple regional feature sub-matrices according to the anatomical region division rules. The pixel feature matrix contains feature data of each pixel, including grayscale value, texture parameters, and edge gradient value.

[0061] Pixel matrix analysis is performed on each voxel of the target cardiac ultrasound image in two-dimensional sections or three-dimensional volume data. The image is converted into a digital matrix, where each element corresponds to a pixel (or voxel), and its grayscale value is recorded (reflecting the echo intensity of the tissue). By calculating the grayscale co-occurrence matrix of the pixels surrounding the pixel, texture parameters (such as energy, entropy, and contrast, reflecting the echo uniformity or complexity of the tissue) are extracted. The Sobel operator is used to calculate the gradient values ​​of the pixels in the horizontal and vertical directions, which are combined to form edge gradient values ​​(reflecting the intensity of grayscale changes at the pixel, used to identify tissue structure boundaries). According to the anatomical region division rules (based on the anatomical structure boundaries in the overall anatomical structure distribution data extracted in step S121), the pixel feature matrix is ​​decomposed into multiple regional feature sub-matrices, such as the left ventricular region feature sub-matrice, the interventricular septum region feature sub-matrice, and the right ventricular region feature sub-matrice. Each regional feature sub-matrice contains the grayscale values, texture parameters, and edge gradient values ​​of all pixels within the corresponding anatomical region.

[0062] Step S123: Compare the overall anatomical structure distribution data with the anatomical structure nodes in the image feature transmission benchmark, mark the specific anatomical regions in the target medical image that completely match the benchmark anatomical structure nodes, and record the three-dimensional coordinate range of each specific anatomical region in the image.

[0063] The overall anatomical structure distribution data obtained in step S121 (including the names and three-dimensional spatial locations of anatomical structures in the target image) is hierarchically compared with the anatomical structure nodes (such as "ventricular septum," "right ventricle," and "pulmonary artery" nodes, including their standard three-dimensional spatial locations and name codes) in the image feature transmission benchmark. First, the anatomical structure names are compared to filter out nodes with matching names; then, the three-dimensional spatial locations are compared to calculate the overlap between the three-dimensional coordinate range of the anatomical structure in the target image and the standard three-dimensional coordinate range of the corresponding anatomical structure node in the benchmark. If the overlap reaches a set threshold (e.g., more than 80%), it is considered a complete match. The specific anatomical regions in the target medical image that completely match the benchmark anatomical structure nodes are labeled, such as "membranous ventricular septum region" and "right ventricular free wall region," and the three-dimensional coordinate range of each specific anatomical region in the image is recorded (e.g., the three-dimensional coordinates of the "membranous ventricular septum region" range from (Xmin1, Ymin1, Zmin1) to (Xmax1, Ymax1, Zmax1).

[0064] Step S124: Extract features from the regional feature submatrix corresponding to each specific anatomical region. The extracted image features include the grayscale distribution range, grayscale concentration interval, texture direction parameters, edge morphology curves, and internal structure density distribution of the specific anatomical region.

[0065] Feature extraction is performed using the regional feature sub-matrix corresponding to the specific anatomical region "ventricular septum membranous region" as an example. The gray-level distribution range is determined by statistically analyzing the maximum and minimum gray-level values ​​of all pixels within the sub-matrix of this region; the gray-level concentration interval is identified by analyzing the frequency distribution histogram of gray-level values ​​to find the interval with the highest frequency gray-level values ​​(e.g., the number of pixels with gray-level values ​​in a certain range accounts for more than 60% of the total number of pixels in the region); the texture direction parameter is obtained by calculating the energy values ​​of the gray-level co-occurrence matrix in different directions (0 degrees, 45 degrees, 90 degrees, 135 degrees) within the region, with the direction with the highest energy value being the main texture direction, and at the same time, the consistency parameter of the texture direction (the degree of difference in energy values ​​in each direction) is calculated; the edge morphology curve is obtained by curve fitting of pixels with edge gradient values ​​higher than the threshold within the region to obtain the continuous curve shape of the edge (e.g., smooth curve, sawtooth curve); the internal structure density distribution is obtained by calculating the ratio of the number of pixels in different sub-regions (e.g., small regions divided by grids) to the area of ​​the sub-regions within the region to obtain the density distribution (e.g., uniform distribution, locally sparse or dense).

[0066] Step S1241: Identify edge lines with abrupt changes in grayscale values ​​in the image using an edge detection algorithm, and determine the boundary coordinates of each specific anatomical region based on the edge line coordinates. The boundary coordinates are used to delineate the range of the specific anatomical region.

[0067] The Canny edge detection algorithm is used to process the target medical image to identify edge lines (such as the outlines of tissues and organs) where grayscale values ​​change abruptly. For the specific anatomical region "membranous ventricular septum region", edge lines are searched within the three-dimensional coordinate range recorded in step S123. The coordinates of the detected edge line pixels are connected and fitted to form a continuous boundary contour. The set of coordinate points of this boundary contour is extracted and determined as the boundary coordinates of the "membranous ventricular septum region". These boundary coordinates accurately delineate the range of this specific anatomical region in the image, ensuring that subsequent feature extraction is performed only on pixels within this range.

[0068] Step S1242: Divide the specific anatomical region into multiple non-overlapping sub-regions according to the detailed distribution of anatomical structures within the specific anatomical region. Each sub-region contains a relatively independent anatomical detail or feature unit.

[0069] Based on the detailed distribution of anatomical structures within the "membranous septum region" (such as the connection points between the membranous septum and surrounding structures, and the area where membranous aneurysms form), it is divided into multiple non-overlapping sub-regions. For example, based on the boundary coordinates obtained in step S1241 and the grayscale distribution and texture changes within the region, the membranous region is divided into "anterior superior border sub-region," "central sub-region," and "posterior inferior border sub-region," etc. Each sub-region contains a relatively independent anatomical detail (such as the "central sub-region" which may contain potential sites of echo interruption, and the "anterior superior border sub-region" which is adjacent to the aortic valve annulus) or a characteristic unit (such as a local echo enhancement area).

[0070] Step S1243: Statistically analyze the grayscale value data of all pixels in each sub-region, determine the distribution range of grayscale values, the grayscale range in which they occur in a concentrated manner, and the morphological characteristics of the grayscale distribution, as the grayscale distribution characteristics of the sub-region.

[0071] The grayscale values ​​of all pixels within the "central sub-region of the membrane" are statistically analyzed. The maximum and minimum grayscale values ​​of pixels within this sub-region are calculated to determine the distribution range of grayscale values ​​(e.g., grayscale values ​​from Gmin to Gmax). By plotting a grayscale histogram, the number of pixels within each grayscale value range is counted, and the grayscale range with the most pixels is identified as the concentrated grayscale range (e.g., the range G1 to G2 has the highest proportion of pixels to the total number of pixels in the sub-region). The morphological characteristics of the grayscale histogram are analyzed to determine whether it is a normal, skewed, or multimodal distribution (e.g., the grayscale distribution of normal interventricular septum tissue may be close to a normal distribution, while peaks with low grayscale values ​​may appear if there is echo interruption). These constitute the grayscale distribution characteristics of the sub-region.

[0072] Step S1244: Use a multi-directional scanning algorithm to analyze the changing trend of pixel grayscale values ​​in the sub-region, and record the main extension direction angle of the texture, the consistency parameter of the direction, and the density change curve of the texture as the texture direction feature of the sub-region.

[0073] A multi-directional scanning algorithm (such as linear scanning in multiple directions like 0°, 45°, 90°, and 135°) is used for the "upper front edge sub-region of the membrane". The trend of gray value changes of pixels along the scan line in each direction is analyzed (such as the rise, fall, or fluctuation of gray value). By calculating the consistency of gray value changes in each direction (such as the average difference of gray value between adjacent pixels), the main extension direction angle of the texture (the scanning direction with the highest consistency) is determined. The consistency parameter of the direction is determined by calculating the standard deviation of the gray value change trend in the main direction. The smaller the standard deviation, the more consistent the direction. The density change curve of the texture is plotted by statistically analyzing the number of gray value change cycles (the number of gray value fluctuations per unit length) along the main direction at different positions, and the density curve changes with the position of the sub-region. These data together constitute the texture direction characteristics of the sub-region.

[0074] Step S1245: Extract line data with abrupt changes in grayscale values ​​within the sub-region, and analyze the continuity, curvature parameters, and thickness variation values ​​of the lines as edge morphological features of the sub-region.

[0075] Within the "lower rear edge sub-region of the membrane," line data (i.e., edge lines) with abrupt changes in grayscale value are extracted based on the edge gradient values ​​of pixels (extracted in step S122). By tracing the pixel sequence of the edge lines, the continuity of the lines is analyzed (e.g., the number of line breaks; fewer breaks indicate higher continuity). The curvature of each point on the line is calculated, and the average or maximum value is taken as the curvature parameter (reflecting the degree of curvature of the line). The width of the line at different positions (the number of pixels perpendicular to the line direction) is measured, and the difference between the maximum and minimum width values ​​is calculated as the thickness variation value. These parameters constitute the edge morphological features of the sub-region.

[0076] Step S1246: Analyze the aggregation state parameters, density difference values, and structural hierarchy distribution of pixels within the sub-region as internal structural features of the sub-region.

[0077] The aggregation parameters of pixels within the "central sub-region of the membrane" are analyzed. By calculating the spatial autocorrelation coefficient of pixels, it is determined whether pixels tend to cluster with similar gray values ​​(e.g., low gray value pixels cluster to form echo loss areas). The density difference value is calculated by comparing the number of pixels in different local small regions (e.g., a 5x5 pixel window) within the sub-region (considering the gray value threshold, counting the number of pixels above or below the threshold), and calculating the difference between the maximum and minimum density. The structural hierarchy distribution is determined by analyzing the changes in gray values ​​and texture parameters at different depths (along the direction of ultrasound beam propagation) within the sub-region to determine whether there are multi-layered structures (e.g., the three-layered structure of a normal ventricular septum: endocardium, myocardium, and epicardium). These constitute the internal structural characteristics of the sub-region.

[0078] Step S1247: Integrate the grayscale distribution features, texture direction features, edge morphology features, and internal structure features of each sub-region to form a comprehensive feature description of each sub-region. The comprehensive feature description of each sub-region contains all image feature data of that sub-region.

[0079] The grayscale distribution features (distribution range, concentration range, morphological features), texture direction features (main extension direction angle, direction consistency parameter, density change curve), edge morphological features (line continuity, curvature parameter, thickness variation value), and internal structural features (aggregation state parameter, density difference value, structural hierarchy distribution) of the "central sub-region of the membrane" are integrated and a comprehensive feature description of the sub-region is formed according to a unified data structure (such as feature name-feature value-feature description format). This ensures that all extracted image feature data of the sub-region are included, which is convenient for subsequent comparison with the benchmark features.

[0080] Step S1248: Based on the spatial positional relationship of each sub-region within the specific anatomical region, arrange the comprehensive feature descriptions of all sub-regions in spatial coordinate order to form a feature sequence of the specific anatomical region. The feature sequence of the specific anatomical region reflects the spatial relationship between the features of each sub-region.

[0081] Based on the spatial relationships (e.g., arranged from anterosuperior to posterior inferior along the long axis of the ventricular septum) of the subregions within the "membranous region of the ventricular septum," the comprehensive feature descriptions of each subregion are arranged according to the order of their spatial coordinates (e.g., from smallest to largest X-coordinate, or largest to smallest Y-coordinate). For example, the comprehensive feature descriptions of the "anterosuperior along the membrane" are arranged first, followed by the "central along the membrane," and then the "posterior inferior along the membrane," forming a feature sequence of the "membranous region of the ventricular septum." The order of the features of each subregion in this feature sequence reflects their spatial adjacency, front-back, and other related relationships.

[0082] Step S1249: Based on the spatial relationship of each sub-region, perform a fusion operation on the feature sequence of the specific anatomical region, and extract the core feature information after the fusion operation. The core feature information is a feature set representing the main image performance of the specific anatomical region, which includes features in multiple dimensions, including grayscale, texture, edge, and internal structure.

[0083] Based on the spatial relationships between the sub-regions of the "membranous ventricular septum region" (such as whether the grayscale transition between adjacent sub-regions is smooth and whether the texture direction is continuous), a fusion operation is performed on the feature sequence. For example, for the grayscale distribution features of adjacent sub-regions, the overlap of their grayscale concentration intervals is calculated. If the overlap is high and the distribution range is continuous, they are merged into a single grayscale distribution feature. For texture direction features, if the main extension direction angles of adjacent sub-regions are similar, they are merged into the main texture direction of the anatomical region. Through the above fusion operation, the core feature information representing the main imaging manifestations of the "membranous ventricular septum region" is extracted. This information includes features in multiple dimensions such as grayscale (overall grayscale distribution range, main concentration intervals), texture (overall texture direction, uniformity), edge (the continuity and morphology of the overall boundary of the region), and internal structure (whether there are abnormal density difference areas and whether the structural layers are clear).

[0084] Step S125: Compare the dimensions of the extracted image features of each specific anatomical region with the core image feature nodes associated with the corresponding anatomical structure nodes in the image feature transmission benchmark, align the feature dimensions one by one, and annotate the matching parameters and fit points of each feature dimension.

[0085] The core image features of the "membranous ventricular septum region" extracted in step S124 (including dimensions such as grayscale, texture, edge, and internal structure) are compared with the core image feature nodes associated with the "ventricular septum" anatomical structure nodes in the image feature transmission benchmark (such as the "ventricular septum continuity interruption" node, whose feature dimensions include grayscale distribution abnormal range, edge continuity features, and internal structure interruption features). Feature dimensions are aligned one by one, such as aligning the grayscale distribution range of the "membranous ventricular septum region" in the target image with the "grayscale distribution abnormal range" dimension of the benchmark node, and aligning the edge continuity of the target image with the "edge continuity feature" dimension of the benchmark node. Matching parameters for each aligned dimension are calculated, such as the overlap of grayscale distribution ranges and the similarity of edge continuity features; matching points are marked, such as a grayscale concentration range in the "membranous ventricular septum region" in the target image being completely consistent with the "grayscale distribution abnormal range" of the benchmark node, or the edge interruption position detected in the target image matching the typical interruption position described by the benchmark node.

[0086] Step S126: Filter image features that match the dimension of the reference core image feature nodes and mark them as candidate adaptation features. Each candidate adaptation feature includes the corresponding anatomical region name, three-dimensional coordinate range, feature dimension data, and matching parameter information with the reference node.

[0087] Based on the dimensional comparison results in step S125, image features that have a high degree of fit (matching parameters higher than the set threshold) with each feature dimension of the benchmark core image feature node (such as the "interventricular septal continuity interruption" node) are selected. For example, if the gray-scale distribution feature of the target image's "interventricular septal membranous region" contains a low gray-scale value range (significantly lower than the normal interventricular septal gray-scale), and the edge gradient value shows a sharp gray-scale abrupt change within this range (indicating a boundary), and it highly matches the "abnormal gray-scale distribution range" and "edge continuity interruption" feature dimensions of the "interventricular septal continuity interruption" node, then this image feature is marked as a candidate fitting feature. Each candidate fitting feature includes the corresponding anatomical region name (interventricular septal membranous region), three-dimensional coordinate range (the three-dimensional coordinates of the abnormal region), feature dimension data (the specific range of the low gray-scale value range, the length and location of the edge interruption), and matching parameter information with the benchmark node (the specific values ​​of the matching parameters for each dimension).

[0088] Step S127: Based on the three-dimensional spatial relationship of the anatomical region corresponding to the candidate fitting features, and combined with the transmission logic of the core image feature nodes in the image feature transmission benchmark, mark the spatial and logical associations between the candidate fitting features.

[0089] Analyze the three-dimensional spatial relationships of the anatomical regions corresponding to each candidate fitting feature, such as the positional relationship between the candidate fitting feature (echo interruption) in the "interventricular septum membranous region" and the candidate fitting feature (right ventricular enlargement) in the "right ventricular region" (the echo interruption region is located on the left side of the right ventricle, and the two are adjacent). Combine the transmission logic of the core image feature nodes in the image feature transmission benchmark (such as the "interventricular septum continuity interruption" node triggering the "left-to-right shunt signal" node, which in turn leads to the "right ventricular enlargement" node), and label the spatial relationships (such as adjacent, contained, and separated by a certain distance) and logical relationships (such as causal relationship, accompanying relationship, and inference based on the benchmark transmission logic that echo interruption may be the cause of right ventricular enlargement) between candidate fitting features.

[0090] Step S128: According to the transmission path template in the image feature transmission benchmark, adjust the arrangement order of the candidate adaptation features so that the sequence of candidate adaptation features is consistent with the node order in the corresponding transmission path template, forming a feature sequence corresponding to the benchmark transmission path.

[0091] Referring to the main conduction pathway template for ventricular septal defects in the reference image conduction feature benchmark (interruption of ventricular septal continuity → left-to-right shunt signal → right ventricular enlargement → increased pulmonary artery pressure), the order of candidate fitting features is adjusted. Assuming three candidate fitting features—"interruption of membranous ventricular septal echo," "left-to-right shunt signal," and "right ventricular enlargement"—are selected from the target image, they are arranged according to the order of node appearance in the benchmark conduction pathway template as the sequence "interruption of membranous ventricular septal echo → left-to-right shunt signal → right ventricular enlargement," ensuring this sequence matches the node order in the benchmark conduction pathway template, thus forming a feature sequence corresponding to the benchmark conduction pathway.

[0092] Step S129: Supplement detailed data for each candidate adaptation feature in the feature sequence. The supplementary data includes the specific performance parameters of the feature, its three-dimensional coordinates in the image, and the correlation parameters with adjacent candidate adaptation features. The supplementary data is generated based on the original pixel matrix of the target medical image and the feature description in the image feature transmission benchmark.

[0093] Supplementing detailed data for the candidate fitting feature "interruption of interventricular septal membranous echo" in the feature sequence: specific performance parameters include the shape of the interruption (e.g., circular, irregular shape), the maximum diameter length at the interruption point; the three-dimensional coordinates in the image are the minimum and maximum three-dimensional coordinate values ​​of the interruption area; the correlation parameters with the adjacent candidate fitting feature "left-to-right shunt signal" include the distance between the two in three-dimensional space (the distance from the center point of the interruption area to the starting point of the shunt signal), and the direction of the shunt signal relative to the interruption area (e.g., the shunt signal points from the left ventricle through the interruption area to the right ventricle). In the above supplementary data, the specific performance parameters and three-dimensional coordinates are calculated based on the original pixel matrix of the target medical image (e.g., converted to actual size by measuring pixel distance), and the correlation parameters are generated by combining the spatial correlation description of the "interventricular septal continuity interruption" and "left-to-right shunt signal" nodes in the image feature transmission benchmark.

[0094] Step S1210: Perform secondary hierarchical alignment between the supplemented feature sequence and the transmission path template in the image feature transmission benchmark, integrate all matched feature sequences, and output the adapted feature set.

[0095] The feature sequence (interruption of echo in the membranous part of the ventricular septum → left-to-right shunt signal → right ventricular enlargement) after supplementing detailed data is then subjected to secondary hierarchical alignment with the ventricular septal defect main conduction path template in the imaging feature conduction benchmark. The detailed data of each candidate fitting feature in the feature sequence are checked to ensure complete matching with the feature description of the corresponding node in the benchmark template (e.g., whether the maximum diameter of the interruption is within the typical range of the benchmark description, whether the direction of the shunt signal conforms to the hemodynamic direction in the benchmark), whether the conduction order is completely consistent with the node order in the template, and whether the spatial correlation parameters conform to the spatial dimension attributes in the benchmark). All compliant feature sequences are integrated. If there are candidate fitting feature sequences for other anatomical regions in the target image (e.g., feature sequences related to pulmonary hypertension), these are also integrated. The final output is a fitting feature set containing all candidate fitting features matching the imaging feature conduction benchmark and their associated information.

[0096] Step S130: Call the bidirectional knowledge transmission link of the large model, connect each feature node in the adaptation feature set to the starting end of the bidirectional knowledge transmission link, and perform association operations with the upstream defect features, downstream defect features, and parallel defect features in the defect knowledge graph to generate reinforced association features.

[0097] In the ventricular septal defect scenario, this step aims to leverage the bidirectional knowledge transfer capability of the large model to establish broader and deeper connections between isolated feature nodes in the fitting feature set and other relevant feature nodes in the knowledge graph, thereby enhancing the information content and relevance of the features and preparing for the subsequent construction of a complete defect feature transfer chain.

[0098] For details, please refer to the following: Figure 3 Step S131: Invoke the bidirectional knowledge transmission link of the large model. The bidirectional knowledge transmission link is a feature association transmission channel constructed based on the defect knowledge graph, which includes a vertical transmission path related to the pathological process and a horizontal transmission path related to the accompanying association.

[0099] The bidirectional knowledge transmission links are constructed based on the defect knowledge graph of the large model. For ventricular septal defects (VSDs), the longitudinal transmission path is related to the pathological process, transmitting along the timeline and causal relationship of disease development. For example, it transmits from the "ventricular septal defect" node to the "left-to-right shunt" node, then to the "right ventricular volume overload increase" node, and then to the "right ventricular enlargement" node, reflecting the pathological process from hemodynamic changes to cardiac structural changes. The lateral transmission path is related to the accompanying associations, connecting parallel feature nodes in the same pathological stage or those that influence each other. For example, the "ventricular septal defect" node transmits laterally to the "atrial septal defect" node (the two may coexist), and the "right ventricular enlargement" node transmits laterally to the "tricuspid regurgitation" node (right ventricular enlargement can lead to tricuspid annular enlargement, which in turn causes regurgitation). The lateral transmission path is established based on common clinical comorbidities, complications, or common pathological bases.

[0100] Step S132: Extract each feature node from the adaptation feature set, and connect each feature node as an independent input unit to the corresponding starting position of the bidirectional knowledge transmission link. Each feature node contains feature dimension data, anatomical positioning information, and initial transmission relationship parameters.

[0101] Each feature node is extracted from the adaptive feature set, such as the "interventricular septum membranous echo interruption" node, the "left-to-right shunt signal" node, and the "right ventricular enlargement" node. Each feature node is treated as an independent input unit, and based on its feature type (structural abnormality feature, hemodynamic feature, morphological change feature) and anatomical location information (interventricular septum membranous region, right ventricle), it is connected to the corresponding starting position in the bidirectional knowledge transmission link. For example, the "interventricular septum membranous echo interruption" node (structural abnormality feature, located in the interventricular septum membranous region) is connected to the starting end of the transmission path related to interventricular septal structural abnormalities, and the "left-to-right shunt signal" node (hemodynamic feature) is connected to the starting end of the transmission path related to shunts. The feature dimension data contained in each input node includes the specific performance parameters supplemented in step S129, the anatomical location information is the three-dimensional coordinate range, and the initial transmission relationship parameters are the initial correlation strength with other nodes set based on the imaging feature transmission benchmark (e.g., lower correlation strength with upstream potential causal nodes, higher correlation strength with downstream direct consequence nodes).

[0102] Step S133: Based on the feature type, anatomical location information and matching parameters of the starting node of the bidirectional knowledge transmission link of the input node, connect each input node to the corresponding transmission path so that the feature dimensions of the input node and the starting node of the link are fully aligned.

[0103] The matching parameters between the input node (e.g., "interruption of echo in the membranous part of the ventricular septum") and each starting node in the bidirectional knowledge transfer link are calculated. The starting node has preset feature type requirements and anatomical location ranges. For example, the "starting node for congenital ventricular septal defect" requires a feature type of structural abnormality and an anatomical location in the membranous or muscular part of the ventricular septum. The feature type matching degree (1.0 for a perfect match) and anatomical location overlap (the proportion of overlap between the three-dimensional coordinates of the input node and the preset anatomical location range of the starting node) are calculated and weighted to obtain the overall matching parameters. The input node is connected to the transfer path with the highest overall matching parameters. By adjusting the feature dimension data format of the input node (e.g., unifying parameter names and units), the feature dimensions of the input node and the starting node of the link are fully aligned (e.g., if the starting node requires a "maximum interruption diameter" parameter, the input node supplements this parameter and ensures the unit is millimeters).

[0104] Step S134: Through the vertical transmission path of the bidirectional knowledge transmission link, trace the upstream defect feature node corresponding to the input node. The upstream defect feature node is a feature node that appears before the input node in the pathological process. The tracing basis is the preset pathological logic association parameter in the bidirectional knowledge transmission link.

[0105] Taking the input node "interruption of membranous echo in the ventricular septum" as an example in the longitudinal conduction path, we trace its upstream defect feature nodes. The pre-defined pathological logic association parameters in the bidirectional knowledge transmission link define the causal relationship and sequence between nodes. For example, the association parameter between the node "incomplete embryonic ventricular septal development" and the node "interruption of membranous echo in the ventricular septum" is "direct cause, chronological order first." Based on these parameters, we search upstream in the longitudinal conduction path to find all nodes that have a "causal relationship and chronological order first" with the node "interruption of membranous echo in the ventricular septum," such as the node "impaired formation of the membranous ventricular septum" and the node "abnormal development of the endocardial cushion" (which may affect the membranous ventricular septum). These nodes are the upstream defect feature nodes corresponding to the input node, and they appear before "interruption of membranous echo in the ventricular septum" in the pathological process.

[0106] Step S1341: Based on the feature description and anatomical location information of the input node and the matching parameters of the input node in the bidirectional knowledge transmission link, determine the specific coordinate position of the input node in the bidirectional knowledge transmission link.

[0107] The input node "interruption of echo in the membranous ventricular septum" is characterized as "an interruption of echo continuity occurs in the membranous region of the ventricular septum, with a clear fracture end," and its anatomical location information is three-dimensional coordinates (X1, Y1, Z1) to (X2, Y2, Z2). Each node in the bidirectional knowledge transfer link has its coordinate position within the link topology, determined by attributes such as the node's feature description keywords and the anatomical location standard range. By calculating the semantic similarity between the input node's feature description and the feature descriptions of other nodes in the link (e.g., the semantic matching degree between "interruption of echo continuity" and the link node "interruption of structural continuity"), and the spatial matching degree between the anatomical location information and the anatomical location standard range of the link node, a comprehensive matching parameter is obtained. The link node with the highest comprehensive matching parameter is then determined as the specific coordinate position of the input node in the bidirectional knowledge transfer link.

[0108] Step S1342: Retrieve the vertical transmission path rule corresponding to the coordinate position in the bidirectional knowledge transmission link. The vertical transmission path rule is constructed based on the pathological development process of the defect and includes the possible upstream node identifier and transmission condition parameters corresponding to each node.

[0109] In the bidirectional knowledge transmission link, each coordinate position is associated with a specific longitudinal transmission path rule. For the coordinate position of the "interruption of echo in the membranous ventricular septum" node, its corresponding longitudinal transmission path rule is retrieved. This rule is constructed based on the pathological development process of ventricular septal defect and clarifies the possible upstream node identifiers, such as the "failure of membranous ventricular septal fusion during embryonic period" node identifier and the "abnormal proliferation of myocardial cells in the ventricular septum" node identifier. At the same time, the rule also includes transmission condition parameters, such as the condition parameter for the transmission from the "failure of membranous ventricular septal fusion during embryonic period" node to the current node being "fusion failure occurred after a specific week of pregnancy and no spontaneous closure occurred," and the transmission condition parameter for the "abnormal proliferation of myocardial cells in the ventricular septum" node being "the degree of abnormal proliferation reaches a threshold that leads to the disruption of structural continuity."

[0110] Step S1343: Based on the longitudinal transmission path rules, screen potential upstream nodes that are directly associated with the input node. The potential upstream node is the preceding feature node that directly leads to the appearance of the input node in the pathological process. Each input node corresponds to multiple potential upstream nodes.

[0111] Based on the longitudinal conduction path rules retrieved in step S1342, potential upstream nodes directly associated with the input node "interruption of echo in the membranous ventricular septum" are screened. The rule defines "direct association" as a direct pathological causal relationship between the upstream node and the input node, with no other intermediate nodes. For example, the node "failure of membranous ventricular septum fusion during embryonic period" directly leads to the failure of the membranous ventricular septum to close, resulting in echo interruption, and is therefore a directly associated potential upstream node; the node "excessive apoptosis of myocardial cells in the ventricular septum" directly leads to the loss of membranous myocardial tissue, and is also a directly associated potential upstream node. Each input node corresponds to multiple potential upstream nodes, all of which act as preceding characteristic nodes in the pathological process, directly leading to the appearance of the input node.

[0112] Step S1344: Extract the feature description, anatomical location coordinates, and transmission condition parameters of each potential upstream node from the defect knowledge graph.

[0113] From the defect knowledge graph of the large model, we extract the feature description of the potential upstream node "failure of fusion of the membranous part of the ventricular septum during embryonic development" (the left and right ventricular bulbar crests, the upper edge of the ventricular septum muscle, and the endocardial cushion tissue failed to fuse normally during embryonic development), anatomical coordinates (the standard three-dimensional coordinate range of the region where the membranous part of the ventricular septum forms in the embryonic heart), and conduction condition parameters (the embryonic developmental stage at which the fusion failure occurred and the abnormal expression of related regulatory genes). We also extract the feature description of the node "excessive apoptosis of cardiomyocytes in the ventricular septum" (the number of apoptotic cardiomyocytes in the membranous part of the ventricular septum exceeds the normal range, resulting in insufficient cell number), anatomical coordinates (highly overlapping with the anatomical coordinates of the input node "interruption of echo in the membranous part of the ventricular septum"), and conduction condition parameters (the concentration threshold of apoptosis-related factors and the degree of reduction in the activity of apoptosis inhibitory factors).

[0114] Step S1345: Compare the transmission condition parameters of each potential upstream node with the actual parameters of the input node. The actual parameters of the input node include its corresponding anatomical region state parameters and feature performance values.

[0115] The actual parameters for the input node "Interruption of echo in the membranous part of the ventricular septum" include anatomical region status parameters (such as the thickness of the membranous part of the ventricular septum and the echo intensity of the surrounding tissue) and characteristic values ​​(such as the maximum diameter of the interruption and the degree of echo enhancement at the interruption ends). The conduction condition parameters for the node "Failure of fusion in the membranous part of the ventricular septum during embryonic period" (such as the embryonic stage of the failed fusion) are compared with the actual parameters for the input node (such as the embryonic development status inferred from the child's age, combined with the interruption morphology to determine whether it is an embryonic fusion problem). The conduction condition parameters for the node "Excessive apoptosis of cardiomyocytes in the ventricular septum" (such as the apoptosis factor concentration threshold) are compared with the characteristic values ​​for the input node (such as the extent of tissue loss in the interrupted area, which can indirectly reflect the degree of apoptosis) to determine whether the conduction conditions of potential upstream nodes meet the actual conditions of the input node.

[0116] Step S1346: Filter potential upstream nodes whose conduction condition parameters match the actual parameters of the input node, and exclude potential upstream nodes whose conduction condition parameters do not meet the requirements. The potential upstream nodes whose conduction condition parameters match the actual parameters of the input node are the actual upstream nodes of the input node in the pathological process.

[0117] After comparison in step S1345, potential upstream nodes whose conduction condition parameters match the actual parameters of the input node are screened out. For example, if the interruption morphology of the input node "interventricular septum membranous echo interruption" is a smooth circle, and the patient is a newborn, and these actual parameters match the conduction condition parameters of the node "embryonic period interventricular septum membranous fusion failure" (interruptions caused by fusion failure are mostly congenital, with relatively regular morphology, and are common in newborns), then it is identified as a true upstream node; if the conduction condition parameters of the node "excessive apoptosis of interventricular septum cardiomyocytes" (such as the imaging manifestation of a large number of apoptotic cells in the tissue surrounding the area requiring interruption, but this manifestation is not present in the actual situation of the input node) do not match the actual situation of the input node, then the potential upstream node is excluded.

[0118] Step S1347: Based on the pathological mechanism data analysis, analyze the transmission logic between the upstream nodes and input nodes after screening, and mark the causal relationship chain that the upstream nodes cause or lead to the appearance of the input nodes.

[0119] We retrieved pathological mechanism data on ventricular septal defects from a large model and analyzed the transmission logic between the selected upstream node "embryonic failure of ventricular septal membranous fusion" and the input node "interruption of ventricular septal membranous echo." Pathological mechanism data shows that the ventricular septal membranous portion during embryonic development is formed by the fusion of endocardial cushion tissue, bulbar ridge tissue, and the upper margin of the ventricular septal muscular portion. If any part of the tissue maldevelops or fusion is impaired during the fusion process, it will lead to an interruption of membranous continuity. Based on this, the causal chain is labeled as: embryonic failure of ventricular septal membranous fusion (upstream node) → failure to form continuity of ventricular septal membranous tissue (intermediate pathological process) → postnatal ultrasound examination showing interruption of ventricular septal membranous echo (input node), clearly describing how the upstream node triggers the appearance of the input node through specific pathophysiological processes.

[0120] Step S1348: Trace back the upstream nodes that precede the selected upstream nodes, repeat the operations of potential node screening, transmission condition parameter comparison, and logic verification until the starting node of the vertical transmission path is traced back. Record all upstream nodes in the entire process in the tracing order to form an upstream node sequence. The upstream node sequence contains information about each upstream node, the transmission logic with adjacent nodes, and the transmission condition parameters.

[0121] For the screened true upstream node "embryonic ventricular septal membranous fusion failure," further upstream nodes are traced. The processes of potential node screening (finding potential upstream nodes related to "embryonic ventricular septal membranous fusion failure" from the defect knowledge graph, such as the "endocardial cushion developmental abnormality" node and the "glomerular ridge growth delay" node), conduction condition parameter comparison (comparing the conduction condition parameters of the aforementioned potential upstream nodes with the actual parameters of the "embryonic ventricular septal membranous fusion failure" node), and logical verification (verifying causal relationships based on pathological mechanism data) are repeated. If the tracing leads to the "embryonic gene expression abnormality" node, which is the initial cause of endocardial cushion developmental abnormalities and thus the starting node of the longitudinal conduction pathway, the tracing is stopped. Record all upstream nodes in tracing order to form an upstream node sequence, such as "abnormal gene expression during embryonic period → abnormal development of endocardial cushion → failure of fusion of membranous ventricular septum during embryonic period → interruption of membranous echo of ventricular septum". The sequence contains the characteristic description, anatomical location and other information of each upstream node, as well as the conduction logic between adjacent nodes (such as how abnormal gene expression leads to abnormal development of endocardial cushion) and conduction condition parameters (such as the duration and scope of influence of abnormal gene expression).

[0122] Step S1349: Integrate the correlation data between upstream node sequences and input nodes, label the transmission relationship type and pathological evidence identifier between input nodes and each upstream node, and form the vertical upstream node tracing result of input nodes.

[0123] The upstream node sequences are integrated with the associated data of the input node "interruption of echogenicity in the membranous part of the ventricular septum". The type of conduction relationship between the input node and each upstream node is labeled, such as "failure of membranous fusion in the embryonic period" having a "direct causal" relationship with the input node, and "abnormal development of the endocardial cushion" having an "indirect causal (through intermediate nodes)" relationship with the input node. Each conduction relationship is labeled with a pathological evidence identifier, which corresponds to the literature number of the pathological mechanism research or clinical guideline entry number in the defect knowledge graph that supports the conduction relationship. For example, the pathological evidence identifier for "failure of membranous fusion in the embryonic period leading to echogenicity interruption" corresponds to the literature number of a classic study on embryonic development of the ventricular septum. This information is integrated to form a complete longitudinal upstream node tracing result for the input node.

[0124] Step S135: Expand the parallel defect feature nodes corresponding to the input node through the horizontal transmission path of the bidirectional knowledge transmission link. The parallel defect feature nodes are feature nodes that appear simultaneously with the input node in the clinical case. The expansion is based on the pre-set accompanying association parameters in the bidirectional knowledge transmission link.

[0125] Taking the input node "interruption of membranous ventricular septal echo" as an example, parallel defect feature nodes are expanded through the lateral transmission path of the bidirectional knowledge transmission link. Preset associated parameters in the lateral transmission path define the probability of simultaneous occurrence, correlation strength, and common pathological basis between nodes. For example, the associated correlation parameters between the nodes "interruption of membranous ventricular septal echo" and "atrial septal defect" are "the probability of simultaneous occurrence in cases of congenital heart disease with malformations is 15%, and the common pathological basis is overall abnormal development of the cardiac septum during the embryonic period," with a moderate correlation strength. Based on these associated correlation parameters, nodes with associated correlation parameters higher than a set threshold are searched in the lateral transmission path and identified as parallel defect feature nodes, such as "atrial septal defect" and "patent ductus arteriosus," which often occur simultaneously with ventricular septal defect in clinical cases.

[0126] Step S136: Extract the complete information of the upstream defect feature nodes obtained by tracing and the parallel defect feature nodes obtained by expansion, including feature description, anatomical location coordinates, association strength parameters, and transmission priority parameters. The complete information comes from the stored data of the defect knowledge graph.

[0127] The complete information of upstream defect feature nodes (such as "embryonic ventricular septal membranous fusion failure" and "endocardial cushion developmental abnormalities") and parallel defect feature nodes (such as "atrial septal defect" and "patent ductus arteriosus") obtained from the defect knowledge graph is extracted. Feature descriptions include detailed clinical definitions and imaging manifestations of the defect feature represented by the node; anatomical coordinates are the three-dimensional coordinate range of the feature in a standard human anatomical model; the association strength parameter is the numerical value of the closeness of the association between the node and the input node "interruption of echogenicity in the membranous ventricular septum" (calculated based on clinical co-occurrence frequency and pathological mechanism correlation); and the transmission priority parameter is the level at which the node's information is prioritized during knowledge transmission (e.g., upstream nodes have higher transmission priority than parallel nodes). All this information is stored in the defect knowledge graph and retrieved through node identifiers.

[0128] Step S137: Perform fusion calculation on the upstream defect feature node information, parallel defect feature node information and corresponding input node information, retain the core feature data of the input node, add the correlation parameters of the upstream node and parallel node to form a fused feature node.

[0129] The information from upstream defect feature nodes (such as the feature description and correlation strength parameters of "embryonic ventricular septal membranous fusion failure"), parallel defect feature nodes (such as the anatomical location coordinates and conduction priority parameters of "atrial septal defect"), and the information from the input node "ventricular septal membranous echo interruption" are fused together. During the fusion process, the core feature data of the input node (such as the specific location, size, and morphology of the echo interruption) are retained as the main information of the fused feature node. Correlation parameters from upstream nodes (such as the correlation strength parameters and conduction relationship type with the "embryonic ventricular septal membranous fusion failure" node) and correlation parameters from parallel nodes (such as the associated probability and common pathological basis identifiers with the "atrial septal defect" node) are added. Through structured data integration, the above information is organized together to form a fused feature node containing multi-dimensional correlation information.

[0130] Step S138: Compare the transmission relationship between the fusion feature nodes with the preset transmission logic of the bidirectional knowledge transmission link, strengthen the association parameters of the fusion feature nodes based on the rules summarized from pathological mechanisms and clinical cases, and arrange all the strengthened fusion feature nodes according to the transmission order and association parameters of the bidirectional knowledge transmission link, and integrate them to form a strengthened association feature that includes input nodes, upstream defect feature nodes, downstream defect feature nodes, parallel defect feature nodes and all association parameters.

[0131] Compare the transmission relationships between various fusion feature nodes (such as the fusion node containing upstream information, the fusion node for "interruption of ventricular septal echo," the fusion node for "left-to-right shunt signal," and the fusion node for "atrial septal defect") with the preset transmission logic of the bidirectional knowledge transmission link (such as whether the preset logic for "ventricular septal defect → left-to-right shunt" is consistent with the actual association between the fusion nodes). Based on pathological mechanisms (such as hemodynamic principles) and rules summarized from clinical cases (such as numerous cases showing a positive correlation between the size of the ventricular septal defect and the shunt volume), strengthen the association parameters of the fusion feature nodes. For example, when the interruption area of ​​the "interruption of ventricular septal echo" fusion node is large, increase its association strength parameter with the "left-to-right shunt signal" fusion node. All enhanced fusion feature nodes are arranged according to the transmission order of the two-way knowledge transmission link (vertically from upstream to downstream, and horizontally sorted by correlation strength) and correlation parameters (such as correlation strength and transmission priority), and integrated to form an enhanced correlation feature. This feature includes input nodes, all upstream defect feature nodes (such as gene abnormality and developmental abnormality nodes), downstream defect feature nodes (such as right ventricular enlargement and pulmonary hypertension nodes), parallel defect feature nodes (such as atrial septal defect nodes), and all correlation parameters between them (correlation strength, transmission conditions, pathological basis, etc.).

[0132] Step S140: Based on the anatomical localization information of enhanced correlation features, the key regions of the target medical image are divided, and feature transmission sub-links within the key regions are constructed. All feature transmission sub-links are connected by cross-regional fusion feature nodes to form a defect feature transmission link covering the entire domain.

[0133] In the ventricular septal defect scenario, the enhanced correlation features include multiple defect-related fusion feature nodes, each with clear anatomical location information. This step aims to identify key regions in the target medical image based on this anatomical location information, construct sub-links within these key regions, and connect these sub-links through cross-regional nodes to form a complete transmission link covering the defect-related regions throughout the entire image.

[0134] Step S141: Extract the enhanced correlation features, which include multiple feature nodes generated by correlation operations and the correlation relationships between nodes; based on the anatomical location information of each feature node in the enhanced correlation features, divide the key regions of the target medical image, which are the collection of anatomical regions corresponding to the feature nodes, covering all anatomical sites that are associated with the defect features.

[0135] Extract all fused feature nodes from the enhanced correlation features, such as nodes like "interruption of echo in the membranous part of the ventricular septum," "left-to-right shunt signal," "right ventricular enlargement," "pulmonary artery widening," and "atrial septal defect," as well as the relationships between nodes (e.g., causal relationships, accompanying relationships). Based on the anatomical location information of each feature node (e.g., "interruption of echo in the membranous part of the ventricular septum" is located in the three-dimensional coordinate region of the membranous part of the ventricular septum, and "right ventricular enlargement" corresponds to the three-dimensional coordinate region of the right ventricle), merge these anatomical regions in the target medical image to delineate key regions. For example, merge the membranous part of the ventricular septum region, the right ventricle region, and the pulmonary artery region into the "Ventricular Septal Defect-Right Heart System-Pulmonary Artery Key Region"; and delineate the atrial septum region into the "Atrial Septal Parallel Defect Key Region." The key region is the collection of all anatomical sites that are directly or indirectly related to the defect features, ensuring that it covers the anatomical regions corresponding to all feature nodes in the enhanced correlation features.

[0136] Step S142: Based on the anatomical location information of each fused feature node, the key regions of the target medical image are divided. The key regions are the set of anatomical regions corresponding to the fused feature nodes, covering all anatomical sites that are associated with the defect features.

[0137] For each fusion feature node in the enhanced correlation features, such as the anatomical location information of the "interventricular septum membranous echo interruption" node (X1, Y1, Z1) to (X2, Y2, Z2), the anatomical location information of the "right ventricular enlargement" node (X3, Y3, Z3) to (X4, Y4, Z4), and the anatomical location information of the "pulmonary hypertension" node (X5, Y5, Z5) to (X6, Y6, Z6), the imaging regions corresponding to the above anatomical location information are analyzed. If the anatomical regions of multiple fusion feature nodes are spatially adjacent or have an inclusion relationship, they are merged into a key region. For example, the anatomical regions of "interventricular septum membranous echo interruption" and "right ventricular enlargement" are adjacent and have a clear causal relationship. Therefore, these two regions and their directly related surrounding small regions (such as the right ventricular outflow tract) are merged to form the "interventricular septum-right ventricular key region." The pulmonary artery anatomical region corresponding to "pulmonary hypertension" is connected to the right ventricular outflow tract and is merged into the "pulmonary artery key region." The division of key regions ensures that all anatomical sites associated with defect features are covered, without omitting any anatomical region corresponding to any fused feature node.

[0138] Step S143: Calculate the anatomical region coordinates corresponding to each fused feature node within the key region, and analyze the spatial association parameters between the fused feature nodes. The spatial association parameters include the distance value of the anatomical location, the overlap ratio, and the adjacent relationship identifier.

[0139] For the "critical region of interventricular septum-right ventricle," the anatomical coordinates (minimum and maximum 3D coordinate values) corresponding to each fusion feature node (such as "interventricular septum membranous echo interruption," "starting point of left-to-right shunt signal," and "thickening of the right ventricular free wall") within this region are calculated. Spatial correlation parameters between nodes are analyzed: the straight-line distance between the center point of the "interventricular septum membranous echo interruption" region and the "starting point of left-to-right shunt signal" is calculated, yielding the distance value; the proportion of overlap between the region traversed by the "left-to-right shunt signal" and the right ventricular cavity region is calculated, yielding the overlap ratio; by determining whether the anatomical boundaries of two nodes share common pixels, adjacency markers (such as "directly adjacent," "indirectly adjacent," and "not adjacent") are indicated. For example, the "interventricular septum membranous echo interruption" region and the "starting point of left-to-right shunt signal" region are directly adjacent and marked as "yes."

[0140] Step S144: Combine the spatial correlation parameters and the transmission direction parameters in the enhanced correlation features to determine the transmission order of the fused feature nodes in the key area. The transmission order simultaneously satisfies the spatial continuity parameter requirements and the logical pathological rationality requirements.

[0141] By combining spatial correlation parameters (such as small distance values ​​and direct adjacency) and conduction direction parameters in enhanced correlation features (such as unidirectional conduction from "interventricular septum membranous echo interruption" to "left-to-right shunt signal," the conduction sequence of fusion feature nodes within the critical region is determined. For example, in the "interventricular septum-right ventricle critical region," the "interventricular septum membranous echo interruption" node and the "left-to-right shunt signal" node are spatially close and directly adjacent, and the conduction direction parameter indicates that the former is the cause and the latter is the effect; therefore, the sequence is "interventricular septum membranous echo interruption → left-to-right shunt signal." The "left-to-right shunt signal" and the "right ventricular enlargement" node have causal conduction direction parameters, and the shunt signal region overlaps significantly with the right ventricular region; therefore, the sequence is "left-to-right shunt signal → right ventricular enlargement." This conduction sequence must meet the requirements of spatial continuity parameters (gradually increasing distance between nodes or distribution along a specific anatomical path) and logical pathological rationality (consistent with the pathological process of defect development, such as shunt precedes volume overload leading to enlargement).

[0142] Step S145: Connect the fusion feature nodes in the key area in sequence according to the determined transmission order to construct the initial defect feature transmission sub-link. Each initial defect feature transmission sub-link corresponds to the defect feature transmission logic in a key area.

[0143] Following the conduction sequence determined in step S144, the fusion feature nodes within the "interventricular septum-right ventricle key region" are connected sequentially, such as "interventricular septum membranous echo interruption → left-to-right shunt signal → right ventricular enlargement → right ventricular free wall thickening," thus constructing an initial defect feature conduction sub-link. Each node in this sub-link is arranged according to the conduction sequence, and the nodes are connected by directed arrows to indicate the conduction direction. The arrows are labeled with the correlation strength parameter and the conduction relationship type (such as "causal" or "accompanied"). Each key region corresponds to an initial defect feature conduction sub-link. For example, the sub-link for the "pulmonary artery key region" might be "left-to-right shunt signal → increased pulmonary artery blood flow → pulmonary artery widening → increased pulmonary artery pressure." Each sub-link reflects the conduction logic of the defect features within the corresponding key region.

[0144] Step S146: Filter the fusion feature nodes that belong to multiple key regions in the enhanced correlation features, and mark the fusion feature nodes that belong to multiple key regions as cross-key region fusion feature nodes. The cross-key region fusion feature nodes establish feature transmission channels between different key regions.

[0145] Examine all fusion feature nodes in the enhanced correlation features to determine whether their anatomical location information simultaneously belongs to multiple critical regions. For example, the anatomical location information of the "left-to-right shunt signal" node starts at the membranous portion of the interventricular septum (IVS-Right Ventricle) and terminates in the right ventricle. Simultaneously, its hemodynamic effects influence blood flow in the "pulmonary artery critical region." Therefore, part of the node's anatomical area or area of ​​influence belongs to the "pulmonary artery critical region," meaning it simultaneously belongs to two critical regions. Mark these nodes as cross-critical region fusion feature nodes. These nodes can serve as a conduction channel connecting the "IVS-Right Ventricle" and "pulmonary artery critical region," enabling the transfer of feature information between different critical regions.

[0146] Step S147: Using the cross-critical region fusion feature node as the connection hub, the initial defect feature transmission sub-links corresponding to different critical regions are connected in series to form a preliminary cross-region transmission link. The preliminary cross-region transmission link includes multiple initial defect feature transmission sub-links and connection nodes between the initial defect feature transmission sub-links.

[0147] Using the cross-critical region fusion feature node "left-to-right shunt signal" as the connecting hub, the initial sub-links of the "interventricular septum-right ventricle critical region" (interventricular septum membranous echo interruption → left-to-right shunt signal → right ventricular enlargement → right ventricular free wall thickening) and the initial sub-links of the "pulmonary artery critical region" (left-to-right shunt signal → increased pulmonary blood flow → pulmonary artery widening → increased pulmonary artery pressure) are connected in series. The common "left-to-right shunt signal" node in both sub-links is used as the connection point to merge them into a preliminary cross-regional conduction link: interventricular septum membranous echo interruption → left-to-right shunt signal → (branch 1: right ventricular enlargement → right ventricular free wall thickening; branch 2: increased pulmonary blood flow → pulmonary artery widening → increased pulmonary artery pressure). This preliminary cross-regional conduction link includes two initial sub-links and the "left-to-right shunt signal" node as the connecting node.

[0148] Step S1471: Extract all fusion feature nodes that simultaneously possess anatomical location information of multiple key regions from the enhanced correlation features. The fusion feature nodes that simultaneously possess anatomical location information of multiple key regions are marked as cross-key region fusion feature nodes. Cross-key region fusion feature nodes establish feature transmission channels between different key regions.

[0149] Each fusion feature node in the enhanced correlation features is traversed, and its anatomical location information is checked to see if it includes the three-dimensional coordinate range of multiple key regions. For example, the anatomical location information of the "left-to-right shunt signal" node extends from the membranous part of the interventricular septum (belonging to the "interventricular septum-right ventricle key region") to the right ventricular cavity and affects the pulmonary artery through the right ventricular outflow tract (belonging to the "pulmonary artery key region"). Its three-dimensional coordinate range simultaneously covers parts of both of these key regions. Such fusion feature nodes that simultaneously possess anatomical location information of multiple key regions are marked as cross-key region fusion feature nodes. Because these nodes spatially span multiple key regions, they can establish connections between the initial defect feature transmission sub-links in different key regions, forming feature transmission channels.

[0150] Step S1472: Label all key region names corresponding to each cross-key region fusion feature node, record the feature performance parameters, associated node identifiers, and propagation direction parameters of the cross-key region fusion feature node in each key region, and obtain complete information about the node in different regions.

[0151] For the marked cross-critical region fusion feature node "left-to-right shunt signal," its corresponding critical region names are labeled as "interventricular septum-right ventricle critical region" and "pulmonary artery critical region." In the "interventricular septum-right ventricle critical region," the node's characteristic parameters include the location of the shunt initiation point (interventricular septum membranous interruption) and the shunt velocity; associated nodes are identified as "interventricular septum membranous echo interruption" (upstream) and "right ventricular enlargement" (downstream); the conduction direction parameter is from the interventricular septum to the right ventricle. In the "pulmonary artery critical region," its characteristic parameters include the percentage increase in pulmonary artery blood flow caused by the shunt; associated nodes are identified as "increased pulmonary artery blood flow" (downstream); the conduction direction parameter is from the right ventricular outflow tract to the pulmonary artery. Recording this information allows us to understand the complete characteristic performance and correlations of this node in different critical regions.

[0152] Step S1473: Retrieve the initial defect feature transmission sub-link corresponding to each cross-critical region fusion feature node. Each initial defect feature transmission sub-link corresponds to a critical region. The initial defect feature transmission sub-link contains the transmission logic and associated node data of the cross-critical region fusion feature node in the corresponding region.

[0153] For the cross-critical region fusion feature node "left-to-right shunt signal", its corresponding initial defect feature conduction sub-link is retrieved. In the initial sub-link corresponding to the "interventricular septum-right ventricle critical region", it includes the node's conduction logic (receiving "interventricular septum membranous echo interruption" and triggering "right ventricular enlargement") and associated node data (association strength parameters with the upstream "interventricular septum membranous echo interruption" and conduction condition parameters with the downstream "right ventricular enlargement"). In the initial sub-link corresponding to the "pulmonary artery critical region", it includes its conduction logic (triggering "increased pulmonary blood flow") and associated node data (association strength and conduction direction with the "increased pulmonary blood flow" node). Each initial sub-link corresponds to a critical region and contains the specific conduction information of the cross-region node within that region.

[0154] Step S1474: Analyze the transmission role of cross-critical region fusion feature nodes in different initial defect feature transmission sub-links. The transmission role is divided into starting node, intermediate node, and ending node. The positioning of the transmission role is determined based on the position and transmission direction parameters of the node in the initial defect feature transmission sub-link.

[0155] The conduction role of the "left-to-right shunt signal" node in conduction sub-links with different initial defect characteristics was analyzed. In the initial sub-link of the "interventricular septum-right ventricular critical region" (interventricular septum membranous echo interruption → left-to-right shunt signal → right ventricular enlargement), this node is located between two nodes, with both an upstream node (interventricular septum membranous echo interruption) and a downstream node (right ventricular enlargement). Its conduction direction parameter is bidirectional (receiving upstream conduction and conducting downstream), therefore it is positioned as an intermediate node. In the initial sub-link of the "pulmonary artery critical region" (left-to-right shunt signal → increased pulmonary blood flow → pulmonary artery widening), this node is the first node in the link, with only a downstream node. Its conduction direction parameter is unidirectional output, therefore it is positioned as the starting node. Its conduction role is determined by the node's position (first, middle, or last) in the sub-link and its conduction direction parameters (input, output, bidirectional).

[0156] Step S1475: Determine the method of connecting different initial defect feature transmission sub-links of cross-critical area fusion feature nodes based on the transmission role. If the cross-critical area fusion feature node is an intermediate node in multiple initial defect feature transmission sub-links, then a direct serial connection method is used; if the roles are different, the connection order is adjusted according to the transmission logic.

[0157] Based on the conduction role of the "left-to-right shunt signal" node in the two initial sub-links ("interventricular septum-right ventricular critical region" as the intermediate node and "pulmonary artery critical region" as the starting node), the connection method is determined. Since this node has a downstream node in the "interventricular septum-right ventricular critical region" sub-link and is the starting node in the "pulmonary artery critical region" sub-link, the "pulmonary artery critical region" sub-link is connected downstream of this node in the "interventricular septum-right ventricular critical region" sub-link, i.e., "interventricular septum membranous echo interruption → left-to-right shunt signal → (right ventricular enlargement → right ventricular free wall thickening) and (increased pulmonary artery blood flow → pulmonary artery widening)," forming a branch connection method. If a cross-regional node is an intermediate node in multiple sub-links (e.g., node 2→node 3→node 4 in link A, and node 5→node 3→node 6 in link B), then a direct serial connection method is used to merge node 3 of link A and node 3 of link B, forming a serial structure with shared intermediate nodes of node 2→node 3→node 4 and node 5→node 3→node 6.

[0158] Step S1476: Using the cross-critical region fusion feature node as a connection hub, sequentially connect all the corresponding initial defect feature transmission sub-links according to the determined connection method to form a preliminary cross-region transmission link.

[0159] Using the "left-to-right shunt signal" node as the connection hub, and following the branch connection method determined in step S1475, the initial sub-links of the "interventricular septum-right ventricle key region" and the "pulmonary artery key region" are connected in series. Specifically, the sequence "interventricular septum membranous echo interruption → left-to-right shunt signal → right ventricular enlargement → right ventricular free wall thickening" is retained as the main branch. Another branch, "left-to-right shunt signal → increased pulmonary blood flow → pulmonary artery widening → increased pulmonary artery pressure," is drawn from the "left-to-right shunt signal" node, forming a preliminary cross-regional conduction link containing a main branch and sub-branches. The same operation is performed sequentially on other cross-critical region fusion feature nodes (if any), connecting their corresponding initial sub-links in series, continuously expanding the preliminary cross-regional conduction link.

[0160] Step S1477: Based on the pathological mechanism and clinical pattern analysis of the defects, analyze the consistency of the transmission logic in the preliminary cross-regional transmission links, and detect whether there are conflicts in the correlation of transmission directions after the transmission sub-links with different initial defect characteristics are connected.

[0161] Based on the pathological mechanism of ventricular septal defect (left-to-right shunt leading to increased right ventricular volume load, resulting in pulmonary hypertension) and clinical patterns (the degree of shunt is positively correlated with the timing of pulmonary hypertension), the consistency of conduction logic in preliminary cross-regional conduction pathways was analyzed. The conduction direction conflict was examined after connecting different sub-pathways. For example, if the "right ventricular enlargement" node is marked as the upstream node of "increased pulmonary artery pressure" in a certain sub-pathway, but according to the pathological mechanism, increased pulmonary artery pressure is usually a late complication after right ventricular enlargement, the conduction direction should be "right ventricular enlargement → increased pulmonary artery pressure." If the reverse occurs, it is considered a conflict of association. The conduction direction between all nodes in the pathway was examined to ensure it conforms to the pathological mechanism and clinical patterns.

[0162] Step S1478: If a conflict exists, adjust the conduction direction parameters according to the pathological mechanism and clinical patterns; adjust the parameters of the conduction links with logical conflicts, refer to the defect knowledge graph, select the conduction logic that conforms to the pathological mechanism to replace the conflicting links, and add cross-regional conduction association description data. The association description data is used to describe the pathological basis and clinical case support of the conduction sub-links with different initial defect features connected by cross-key region fusion feature nodes.

[0163] If a conflict is found during the analysis, such as a conflict between the conduction direction "increased pulmonary artery pressure → right ventricular enlargement" in a sub-link and the pathological mechanism, the conduction direction parameter should be adjusted to "right ventricular enlargement → increased pulmonary artery pressure" according to the pathological mechanism. For conduction links with logical conflicts (such as inconsistent association strength parameters of two sub-links for the same node), refer to the standard association strength range of the node in the defect knowledge graph and select parameter values ​​that conform to the pathological mechanism for adjustment (such as taking the medium association strength value recommended in the knowledge graph). Add association description data for cross-regional conduction, such as the pathological basis for the "left-to-right shunt signal" node connecting "ventricular septum-right ventricular critical area" and "pulmonary artery critical area" being "left-to-right shunt increases right ventricular volume load, thereby increasing pulmonary artery blood flow, which can lead to pulmonary hypertension in the long term," supported by clinical cases stating that "according to case statistics in the defect knowledge graph, 80% of large ventricular septal defect cases have this conduction path."

[0164] Step S1479: Integrate all connected cross-regional transmission links to form a cross-regional defect feature transmission link framework covering all key areas.

[0165] All sub-links connected and adjusted through cross-critical region fusion feature nodes are integrated. For example, sub-links such as "interventricular septum-right ventricle critical region," "pulmonary artery critical region," and "atrial septum parallel defect critical region" are connected through their respective cross-regional nodes (such as "left-to-right shunt signal" and "atrial septal defect"). If the "atrial septal defect" node and the "interruption of membranous echo in the interventricular septum" node are associated, the "atrial septum parallel defect critical region" sub-link is associated with the main link through this node, forming a cross-regional defect feature conduction link framework covering all critical regions in the target medical image. This framework includes all fused feature nodes, the conduction relationships between nodes, and cross-regional connection descriptions.

[0166] Step S148: Supplement the transmission relationship data in the preliminary cross-regional transmission link. Add corresponding supporting information to each transmission link. The supporting information includes pathological evidence literature identifiers and clinical case numbers. All supporting information comes from the defect knowledge graph of the large model.

[0167] Supporting information was added to each conduction link in the initial cross-regional conduction pathway (e.g., "interruption of echo in the membranous part of the ventricular septum → left-to-right shunt signal" and "left-to-right shunt signal → increased pulmonary blood flow"). For the "interruption of echo in the membranous part of the ventricular septum → left-to-right shunt signal" link, the added pathological literature identifiers corresponded to the literature numbers of hemodynamic studies on shunts caused by ventricular septal defects in the defect knowledge graph; the clinical case numbers corresponded to the confirmed case numbers with the same conduction relationship in the defect knowledge graph (e.g., Case ID: VSD2023001). All supporting information was retrieved from the defect knowledge graph of the large model to ensure that the conduction relationship had a reliable pathological basis and clinical examples to support it.

[0168] Step S149: Bind the original pixel data of the target medical image to the preliminary cross-regional conduction link, so that each fusion feature node in the preliminary cross-regional conduction link corresponds to a specific pixel region in the image; check whether all fusion feature nodes in the enhancement association features have been included in the bound preliminary cross-regional conduction link, and add the fusion feature nodes that have not been included to the corresponding positions of the link according to their anatomical location information and conduction logic, forming a full-domain coverage defect feature conduction link that covers all key areas related to defects in the target medical image.

[0169] The raw pixel data of the target medical image (including the coordinates and grayscale information of each pixel) is bound to a preliminary cross-regional conduction link, assigning a specific pixel region in the image to each fusion feature node in the link. For example, the "interruption of echo in the membranous ventricular septum" node is bound to the pixel region in the image where the echo in the membranous ventricular septum is interrupted (achieved through coordinate matching), and the "left-to-right shunt signal" node is bound to the pixel region in the color Doppler image showing the shunt. It is checked whether all fusion feature nodes in the enhancement-associated features have been included in the bound link. If the "right ventricular outflow tract stenosis" node is found to be missing, it is added to the corresponding upstream or downstream position of the "elevated pulmonary artery pressure" node based on its anatomical location information (three-dimensional coordinates of the right ventricular outflow tract) and conduction logic (possibly related to "elevated pulmonary artery pressure"). After supplementation, a defect feature conduction link covering the entire region is formed, ensuring that all critical regions and their feature nodes related to defects in the target medical image are included.

[0170] Step S150: Extract the node association data, transmission path information, and defect matching results in the defect feature transmission link, and integrate them into a structured image recognition report in accordance with the medical image reporting specifications. The structured image recognition report includes feature transmission atlas, defect association description, and anatomical location annotation.

[0171] In the case of ventricular septal defect, this step aims to extract key information from the established full-coverage defect feature transmission link and integrate it into a structured report according to the standard format of medical imaging reports, clearly presenting the image recognition results.

[0172] For example, step S151: Extract all fused feature node information in the defect feature transmission link. Each fused feature node information includes feature description, anatomical location coordinates, associated node identifier, transmission direction parameters, and pathological basis summary.

[0173] Information on all fusion feature nodes was extracted from the defect feature conduction link covering the entire domain. For example, the feature description of the node "interruption of echo in the membranous part of the ventricular septum" was "a visible interruption of echo continuity in the membranous part of the ventricular septum with a diameter of approximately D, and enhanced echo at the cut end"; the anatomical location coordinates were (Xmin, Ymin, Zmin) to (Xmax, Ymax, Zmax); the associated node was identified as "failure of membranous fusion in the embryonic period" (upstream) and "left-to-right shunt signal" (downstream); the conduction direction parameter was bidirectional (receiving upstream conduction and conducting downstream); the pathological basis summary was "fusion failure of the membranous part of the ventricular septum in the embryonic period led to interruption of structural continuity after birth, consistent with the typical manifestation of membranous ventricular septal defect". The above information was extracted for each node in the link to form a node information list.

[0174] Step S152: Organize the transmission relationships between the fusion feature nodes according to the transmission order to form a transmission relationship list. The transmission relationship list includes the starting node identifier, ending node identifier, transmission type, and association strength parameter of each transmission link.

[0175] Following the conduction sequence in the defect feature transmission chain (e.g., "abnormal gene expression during embryonic period → interruption of echo in the membranous ventricular septum → left-to-right shunt signal → right ventricular enlargement → increased pulmonary artery pressure"), organize the conduction relationships between nodes. Create a list of conduction relationships, where each entry corresponds to a conduction link: the starting node is identified by the name of the upstream node (e.g., "interruption of echo in the membranous ventricular septum"); the ending node is identified by the name of the downstream node (e.g., "left-to-right shunt signal"); the conduction type is indicated by "causal relationship," "accompanying relationship," etc. (e.g., causal relationship); the association strength parameter is a quantified value of the degree of association (e.g., 0.8, based on the standard setting in the defect knowledge graph). List all conduction links sequentially to form a complete list of conduction relationships.

[0176] Step S153: Based on the fusion of feature node information and the list of transmission relationships, a feature transmission relationship chart is generated. The feature transmission relationship chart presents all nodes and the transmission relationships between nodes in a visual form. The feature transmission relationship chart includes node identifiers, anatomical location labels, and transmission direction arrows.

[0177] Based on the fusion feature node information (including node identifiers and anatomical location coordinates) and the conduction relationship list (including conduction direction, start and end nodes), a feature conduction relationship chart is generated. In the chart, each fusion feature node is represented by a circular or square icon, with the node identifier (e.g., "interruption of echo in the membranous part of the ventricular septum") marked inside the icon; the anatomical location (e.g., "membranous part of the ventricular septum") is marked next to or inside the icon; the start and end nodes are connected by directed lines with arrows, the arrow direction indicating the conduction direction (e.g., from "interruption of echo in the membranous part of the ventricular septum" to "left-to-right shunt signal"); associated intensity parameters or conduction types can be marked next to the line segments. The overall layout of the chart is arranged according to anatomical location relationships and conduction order, making the spatial and logical relationships between nodes clearly visible.

[0178] Step S154: Extract the defect association information corresponding to each fusion feature node from the defect knowledge graph of the large model. The defect association information includes the defect type name, defect performance description, and associated defect name corresponding to the fusion feature node.

[0179] Defect association information was extracted from each fusion feature node in the defect knowledge graph. For the node "Interruption of echo in the membranous part of the ventricular septum," the corresponding defect type name was "Congenital ventricular septal defect (membranous type)"; the defect manifestation description was "Local discontinuity interruption in the membranous part of the ventricular septum, which may be accompanied by left-to-right shunt"; the associated defect names were "Aortic valve prolapse" and "Pulmonary hypertension" (common combined or secondary defects). For the node "Right ventricular enlargement," the defect type name was "Increased right ventricular volume overload"; the defect manifestation description was "Increased right ventricular diameter and enhanced wall motion amplitude"; the associated defect names were "Tricuspid regurgitation" and "Pulmonary valve insufficiency." This information from all nodes was extracted to form a list of defect association information.

[0180] Step S155: Classify and organize the defect association information according to the defect type name, and generate a defect association description based on the classified defect association information. The defect association description is used to describe the performance of the fusion feature node, the feature transmission process, and the association with other defects corresponding to each defect type. Each category includes the corresponding fusion feature node information, transmission relationship data, and pathological evidence summary.

[0181] Defect association information is categorized and organized according to defect type name (e.g., "Congenital Ventricular Septal Defect (Membranous Type)," "Pulmonary Hypertension," "Atrial Septal Defect"). Taking the "Congenital Ventricular Septal Defect (Membranous Type)" category as an example, the defect association description is generated as follows: the fusion feature nodes corresponding to this defect type are "interruption of membranous ventricular septal echo" (diameter D) and "left-to-right shunt signal" (velocity V); the characteristic conduction process is "interruption of membranous ventricular septal echo → left-to-right shunt signal → right ventricular enlargement → increased pulmonary artery blood flow → increased pulmonary artery pressure"; the association with other defects is "often accompanied by right ventricular enlargement and pulmonary hypertension, and a few may be associated with atrial septal defects." Each category includes information on all fusion feature nodes corresponding to this defect type (e.g., interruption diameter, shunt velocity), conduction relationship data (e.g., conduction order and association strength between nodes), and a summary of pathological evidence (e.g., the mechanism by which shunt leads to pulmonary hypertension).

[0182] Step S156: Determine the fixed framework of the structured image recognition report, which includes the report title, basic image information, feature transmission relationship analysis, defect association description, and link integrity.

[0183] The fixed framework for the structured image recognition report is determined, and the contents of each part are as follows: The report title is "Pediatric Echocardiography Image Recognition Report (Congenital Heart Disease Special Project)"; the basic image information includes the model of the examination equipment, the examination date, the patient ID (anonymized), the image modality (e.g., 2D ultrasound + color Doppler), and the list of imaging sections; the feature transmission relationship analysis section includes feature transmission relationship charts, a list of transmission relationships, and textual descriptions; the defect association description section presents the defect association descriptions generated in step S155 according to the defect type; the link integrity section describes the number of key areas covered by the defect feature transmission link, the number of fused feature nodes included, and the description of uncovered areas (if any).

[0184] Step S157: Extract the basic content of the target medical image from the raw image data, and fill the basic content of the target medical image into the position corresponding to the fixed frame. The basic content includes modality type, imaging range, and acquisition time.

[0185] Extract the basic information from the raw data of the target medical image: modality type "2D echocardiography + color Doppler flow imaging"; imaging range "heart and major blood vessels (standard sections such as parasternal, apical, and subxiphoid)"; acquisition time is the specific date and time of the examination (accurate to the minute). Fill the above basic information into the corresponding position of "basic image information" in the fixed framework of the structured image recognition report to ensure the accuracy of the information.

[0186] Step S158: Embed the feature transmission relationship chart and transmission relationship list into the feature transmission relationship analysis section of the report, and fill in the classified defect association information and the written defect association description into the defect association description section of the report.

[0187] Embed the feature transmission relationship diagram (visual node relationship diagram) generated in step S153 and the transmission relationship list (text table format) formed in step S152 into the "Feature Transmission Relationship Analysis" section of the report. The diagram and list should be arranged vertically, with the diagram on top and the list below. Add text descriptions to briefly summarize the main feature transmission paths and key nodes in the link. Fill the "Defect Association Description" section of the report with the defect association information classified in step S155 (such as information lists classified by "Congenital Ventricular Septal Defect (Membranous Type)" and "Pulmonary Hypertension") and the written defect association descriptions (text paragraphs). Each defect type should be a subheading, followed by the corresponding association information and description.

[0188] Step S159: Compile a list of key areas covered by the defect feature transmission link and the number of fused feature nodes that have been connected, as a description of the link integrity, and generate a structured image recognition report.

[0189] The report includes a list of key regions covered by the defect feature transmission pathways, such as the "interventricular septum-right ventricle key region," "pulmonary artery key region," and "atrial septum parallel defect key region," totaling N key regions. It also includes the number of fusion feature nodes with established connections, such as M nodes (including all defect-related nodes). These statistical results are added to the "Link Integrity" section of the report as a description of the link integrity, stating that "the feature transmission pathways in this report cover all defect-related key regions in the target image, totaling N, including M fusion feature nodes. The transmission relationships between nodes are clear and consistent with the pathological mechanism." All parts of the report are then integrated to form a complete structured image recognition report.

[0190] In one exemplary embodiment, a medical image feature transduction and recognition system is provided. This system can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 4As shown, the medical image feature transduction and recognition system includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements the medical image feature transduction and recognition method. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the shell of the medical image feature transmission and recognition system, or an external keyboard, touchpad, or mouse, etc.

[0191] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for medical image feature transmission and recognition, characterized in that, include: The defect knowledge graph built into the large model is retrieved, and the association hierarchy and transmission logic of defect feature nodes and anatomical structure nodes in the defect knowledge graph are analyzed to generate image feature transmission benchmarks. Extract the anatomical structure distribution data and pixel feature matrix of the target medical image, perform feature transfer adaptation operation based on the anatomical structure mapping rules, establish the hierarchical correspondence between the target medical image features and the image feature transfer benchmark, and output the adapted feature set. The bidirectional knowledge transmission link of the large model is invoked, and each feature node in the adaptive feature set is connected to the starting end of the bidirectional knowledge transmission link. It is then associated with the upstream defect features, downstream defect features, and parallel defect features in the defect knowledge graph to generate reinforced association features. The bidirectional knowledge transmission link is a feature association transmission channel built based on the defect knowledge graph, which includes a vertical transmission path related to the pathological process and a horizontal transmission path related to the accompanying association. Based on the anatomical location information of enhanced correlation features, key regions of the target medical image are delineated, and feature transmission sub-links are constructed within the key regions. All feature transmission sub-links are connected by cross-regional fusion feature nodes to form a defect feature transmission link covering the entire domain. The fusion feature node refers to the node generated by fusing the input node with the upstream defect feature node information obtained through bidirectional knowledge transmission link tracing and the parallel defect feature node information obtained through expansion. This fusion operation retains the core feature data of the input node and adds the correlation parameters of the upstream node and the parallel node, forming a comprehensive feature node that includes the input node, upstream defect feature node, downstream defect feature node, parallel defect feature node and all correlation parameters. The input node refers to each feature node in the adaptive feature set. Each feature node is connected as an independent input unit to the corresponding starting position of the bidirectional knowledge transmission link. Each feature node includes feature dimension data, anatomical location information and initial transmission relationship parameters. Extract node association data, transmission path information, and defect matching results from the defect feature transmission link, and integrate them into a structured image recognition report according to medical image reporting standards. The structured image recognition report includes feature transmission atlas, defect association description, and anatomical location annotation. The process involves retrieving the built-in defect knowledge graph of the large model, analyzing the association hierarchy and transmission logic of defect feature nodes and anatomical structure nodes in the defect knowledge graph, and generating an image feature transmission benchmark, including: The defect knowledge graph built into the large model is retrieved. The defect knowledge graph stores the anatomical structural features, imaging features, and defect association features corresponding to various birth defects in children. All features exist in the form of independent nodes and the nodes are connected by labeled association relationships. Extract the core image feature nodes corresponding to all defect types in the defect knowledge graph. Each core image feature node contains multiple information dimensions, including feature dimensions, manifestation form, anatomical location, and pathological correlation basis. Each information dimension is supported by specific clinical case data. The transmission logic between core image feature nodes is analyzed, and the order of appearance or associated occurrence of different core image feature nodes in the defect evolution is marked. The transmission logic is formed based on the pathological mechanism of defect occurrence and the time sequence of development process. Based on the frequency of occurrence of this transmission logic recorded in the clinical case database, the transmission logic of each core imaging feature node is prioritized and ranked, and the ranking result is determined in combination with the supporting conclusions in the literature on defect pathological mechanisms. Based on the sorted transmission logic, a transmission path template for core image feature nodes is constructed. Each transmission path template corresponds to the feature transmission law of a type of defect. The transmission path template includes a start node identifier, an intermediate node sequence, an end node identifier, and transmission direction parameters between nodes. Extract the anatomical structure nodes that are directly associated with the core image feature nodes from the defect knowledge graph, establish a unique mapping relationship between the core image feature nodes and the anatomical structure nodes, and label the specific anatomical location coordinates and anatomical range boundaries corresponding to each core image feature node. The three-dimensional spatial location information of the anatomical structure nodes is embedded into the transmission path template. A spatial dimension attribute is added to each transmission path template. The spatial dimension attribute marks the adjacency relationship, overlap range and distance parameters of the anatomical position corresponding to the core image feature node in three-dimensional space. The transmission path templates with embedded spatial attributes are standardized, and all standardized transmission path templates are integrated to construct an initial image feature transmission benchmark framework. The initial image feature transmission benchmark framework includes feature transmission paths, spatial correlation data, and pathological mechanism descriptions corresponding to various defects. Prioritize and filter the transmission paths in the initial image feature transmission benchmark framework. Based on the data completeness and pathological mechanism clarity of the clinical cases associated with the transmission paths, distinguish between core transmission paths and reference transmission paths, and generate an image feature transmission benchmark containing hierarchical paths. Based on anatomical structure mapping rules, feature transfer and adaptation operations are performed to establish a hierarchical correspondence between target medical image features and image feature transfer benchmarks, outputting a set of adapted features, including: The overall anatomical structure distribution data of the target medical image is extracted by an image segmentation algorithm. The overall anatomical structure distribution data covers the names, three-dimensional spatial locations, morphological parameters and spatial relationships of all identifiable anatomical structures in the image. It is obtained by scanning the image region by region and matching it with an anatomical structure feature library. A pixel feature matrix of the target medical image is generated using pixel matrix parsing technology. The pixel feature matrix is ​​then decomposed into multiple regional feature sub-matrices according to the anatomical region division rules. The pixel feature matrix contains feature data for each pixel, including grayscale value, texture parameters, and edge gradient value. The overall anatomical structure distribution data is hierarchically compared with the anatomical structure nodes in the image feature transmission benchmark. The specific anatomical regions in the target medical image that completely match the benchmark anatomical structure nodes are marked, and the three-dimensional coordinate range of each specific anatomical region in the image is recorded. Feature extraction is performed on the regional feature submatrix corresponding to each specific anatomical region. The extracted image features include the gray-level distribution range, gray-level concentration interval, texture direction parameters, edge morphology curves, and internal structure density distribution of the specific anatomical region. The image features of each extracted specific anatomical region are compared with the core image feature nodes associated with the corresponding anatomical structure nodes in the image feature transmission benchmark. The feature dimensions are aligned one by one, and the matching parameters and fitting points of each feature dimension are marked. Image features that match the dimensions of the baseline core image feature nodes are selected and marked as candidate fitting features. Each candidate fitting feature includes the corresponding anatomical region name, three-dimensional coordinate range, feature dimension data, and matching parameter information with the baseline node. Based on the three-dimensional spatial relationship of the anatomical region corresponding to the candidate adaptation features, and combined with the transmission logic of the core image feature nodes in the image feature transmission benchmark, the spatial and logical associations between the candidate adaptation features are marked. According to the transmission path template in the image feature transmission benchmark, adjust the arrangement order of candidate adaptation features so that the sequence of candidate adaptation features is consistent with the node order in the corresponding transmission path template, forming a feature sequence corresponding to the benchmark transmission path. For each candidate fitting feature in the feature sequence, supplementary detailed data is provided. The supplementary content includes the specific performance parameters of the feature, its three-dimensional coordinates in the image, and the correlation parameters with adjacent candidate fitting features. The supplementary data is generated based on the original pixel matrix of the target medical image and the feature description in the image feature transmission benchmark. The supplemented feature sequences are then aligned hierarchically with the transmission path template in the image feature transmission benchmark. All matched feature sequences are then integrated to output the adapted feature set.

2. The medical image feature transmission and recognition method according to claim 1, characterized in that, The bidirectional knowledge transmission link of the large model is invoked, connecting each feature node in the adaptive feature set to the starting end of the bidirectional knowledge transmission link, and performing association operations with the upstream defect features, downstream defect features, and parallel defect features in the defect knowledge graph to generate reinforced association features, including: Invoke the bidirectional knowledge transfer link of the large model; Extract each feature node from the adaptation feature set, and connect each feature node as an independent input unit to the corresponding starting position of the bidirectional knowledge transmission link. Each feature node contains feature dimension data, anatomical positioning information, and initial transmission relationship parameters. Based on the feature type of the input node, the anatomical location information and the matching parameters of the starting node of the bidirectional knowledge transmission link, each input node is connected to the corresponding transmission path, so that the feature dimensions of the input node and the starting node of the link are fully aligned. Through the vertical transmission path of the bidirectional knowledge transmission link, the upstream defect feature node corresponding to the input node is traced. The upstream defect feature node is a feature node that appears before the input node in the pathological process. The tracing basis is the preset pathological logic association parameter in the bidirectional knowledge transmission link. By expanding the parallel defect feature nodes corresponding to the input node through the horizontal transmission path of the bidirectional knowledge transmission link, the parallel defect feature nodes are feature nodes that appear simultaneously with the input node in clinical cases. The expansion is based on the pre-set accompanying association parameters in the bidirectional knowledge transmission link. Extract the complete information of upstream defect feature nodes obtained from tracing and parallel defect feature nodes obtained from expansion, including feature description, anatomical location coordinates, association strength parameters, and transmission priority parameters. All of the complete information comes from the stored data of the defect knowledge graph. The upstream defect feature node information, parallel defect feature node information and corresponding input node information are fused and calculated. The core feature data of the input node is retained, and the correlation parameters of the upstream node and parallel node are added to form the fused feature node. By comparing the transmission relationship between fusion feature nodes with the preset transmission logic of the bidirectional knowledge transmission link, the association parameters of fusion feature nodes are strengthened based on the rules summarized from pathological mechanisms and clinical cases. All strengthened fusion feature nodes are arranged according to the transmission order and association parameters of the bidirectional knowledge transmission link, and integrated to form a strengthened association feature that includes input nodes, upstream defect feature nodes, downstream defect feature nodes, parallel defect feature nodes and all association parameters.

3. The medical image feature transmission and recognition method according to claim 1, characterized in that, The anatomical localization information based on enhanced correlation features is used to delineate key regions of the target medical image, construct feature transmission sub-links within the key regions, and connect all feature transmission sub-links by fusing feature nodes across regions to form a defect feature transmission link covering the entire domain, including: The enhanced correlation features are extracted, which include multiple feature nodes generated by correlation operations and the correlation relationships between nodes; based on the anatomical location information of each feature node in the enhanced correlation features, the key regions of the target medical image are divided, which are the collection of anatomical regions corresponding to the feature nodes, covering all anatomical sites that are associated with the defect features; Based on the anatomical location information of each fused feature node, the key regions of the target medical image are divided. The key regions are the set of anatomical regions corresponding to the fused feature nodes, covering all anatomical sites that are associated with the defect features. Calculate the anatomical region coordinates corresponding to each fused feature node within the key region, and analyze the spatial association parameters between the fused feature nodes. The spatial association parameters include the distance value of the anatomical location, the proportion of the overlapping range, and the adjacent relationship identifier. By combining spatial correlation parameters and transmission direction parameters in enhanced correlation features, the transmission order of fused feature nodes in key areas is determined. The transmission order simultaneously satisfies the requirements of spatial continuity parameters and logical pathological rationality. According to the determined transmission order, the fusion feature nodes in the key area are connected sequentially to construct the initial defect feature transmission sub-link. Each initial defect feature transmission sub-link corresponds to the defect feature transmission logic in a key area. Filter and strengthen the correlation features. The fusion feature nodes that belong to multiple key regions are marked as cross-key region fusion feature nodes. The cross-key region fusion feature nodes establish feature transmission channels between different key regions. Using the cross-critical region fusion feature node as a connection hub, the initial defect feature transmission sub-links corresponding to different critical regions are connected in series to form a preliminary cross-region transmission link. The preliminary cross-region transmission link includes multiple initial defect feature transmission sub-links and connection nodes between the initial defect feature transmission sub-links. Supplement the preliminary cross-regional transmission link data with corresponding supporting information for each transmission link. The supporting information includes pathological evidence literature identifiers and clinical case numbers. All supporting information comes from the defective knowledge graph of the large model. The raw pixel data of the target medical image is bound to the preliminary cross-regional conduction link, so that each fusion feature node in the preliminary cross-regional conduction link corresponds to a specific pixel region in the image. It is then checked whether all fusion feature nodes in the enhancement correlation features have been included in the bound preliminary cross-regional conduction link. Unincluded fusion feature nodes are added to the corresponding positions in the link according to their anatomical location information and conduction logic, forming a full-domain defect feature conduction link that covers all key areas related to defects in the target medical image.

4. The medical image feature transmission and recognition method according to claim 1, characterized in that, The analysis of the transmission logic between core image feature nodes, and the annotation of the order of appearance or associated occurrence of different core image feature nodes in defect evolution, include: The clinical case database stored in the large model is retrieved. The clinical case database contains confirmed cases of birth defects in children and corresponding diagnostic reports. Each case of birth defect in children records detailed data on the imaging manifestations of the defect and the timeline of its development. Based on the matching parameters between the feature descriptions of the core image feature nodes and the image performance data of clinical cases, cases related to the core image feature nodes are selected from the clinical case database, with each core image feature node corresponding to multiple clinical cases. The occurrence order of core imaging feature nodes is analyzed case by case. The timestamp or imaging layer identifier of the first occurrence of different core imaging feature nodes in each clinical case is recorded to form a list of node occurrence order. Each possible combination is used as a statistical unit. The frequency values ​​of the occurrence order of core imaging feature nodes in all relevant clinical cases are counted to form a frequency table of sequence combinations. The pathological mechanism research data in the built-in knowledge base of the large model are retrieved. The pathological mechanism research data includes the molecular biological mechanisms of defect occurrence and development, and the anatomical evolution law. By combining the frequency table of sequential combinations and pathological mechanism research data, the rationality of the occurrence order of the core imaging feature nodes is verified, and the main transmission logic between the core imaging feature nodes is determined. The main transmission logic is the sequential combination of nodes that have high-frequency occurrence records in the clinical case database and conform to the pathological mechanism. Identify the secondary transmission logic between core image feature nodes, wherein the secondary transmission logic is a combination of nodes whose frequency value percentage does not meet the clinical statistical standard but is supported by a pathological mechanism or appears in specific labeled cases; The trigger condition parameters for each conduction logic are labeled. The trigger condition parameters include factors that affect the selection of conduction paths, specifically including defect type coding, child age range, and anatomical structure status parameters. Primary conduction logic is marked as a priority conduction path, and secondary conduction logic is marked as an alternative conduction path. All marked conduction logics are then organized to form a list of conduction logics for core image feature nodes. The list of conduction logics includes node combinations of conduction logics, frequency values, pathological evidence summaries, trigger condition parameters, and path type identifiers.

5. The medical image feature transmission and recognition method according to claim 1, characterized in that, The feature extraction process for each specific anatomical region's corresponding regional feature submatrix includes the extracted image features such as the grayscale distribution range, grayscale concentration intervals, texture orientation parameters, edge morphology curves, and internal structure density distribution of that specific anatomical region. An edge detection algorithm is used to identify edge lines in the image where gray values ​​change abruptly, and the boundary coordinates of each specific anatomical region are determined based on the edge line coordinates. The boundary coordinates are used to delineate the range of the specific anatomical region. Based on the detailed distribution of anatomical structures within a specific anatomical region, the specific anatomical region is divided into multiple non-overlapping sub-regions, each containing a relatively independent anatomical detail or feature unit. Collect grayscale data of all pixels in each sub-region, determine the distribution range of grayscale values, the concentrated grayscale range, and the morphological characteristics of grayscale distribution, and use these as the grayscale distribution characteristics of the sub-region. A multi-directional scanning algorithm is used to analyze the changing trend of pixel grayscale values ​​within a sub-region, and the main extension direction angle of the texture, the consistency parameter of the direction, and the density change curve of the texture are recorded as the texture direction features of the sub-region. Extract line data with abrupt changes in grayscale values ​​within a sub-region, and analyze the continuity, curvature parameters, and thickness variation values ​​of the lines as edge morphological features of the sub-region. The aggregation state parameters, density difference values, and structural hierarchy distribution of pixels within a sub-region are analyzed as internal structural features of the sub-region. The grayscale distribution features, texture direction features, edge morphology features, and internal structure features of each sub-region are integrated to form a comprehensive feature description of each sub-region. The comprehensive feature description of each sub-region contains all image feature data of that sub-region. Based on the spatial relationship between the sub-regions within a specific anatomical region, the comprehensive feature descriptions of all sub-regions are arranged in spatial coordinate order to form a feature sequence of the specific anatomical region. The feature sequence of the specific anatomical region reflects the spatial relationship between the features of each sub-region. Based on the spatial relationship between each sub-region, the feature sequence of the specific anatomical region is fused to extract the core feature information after fusion. The core feature information is a feature set representing the main image performance of the specific anatomical region, which includes features in multiple dimensions, including grayscale, texture, edge, and internal structure.

6. The medical image feature transmission and recognition method according to claim 2, characterized in that, The vertical transmission path through the bidirectional knowledge transmission link traces the upstream defect feature node corresponding to the input node. The upstream defect feature node is a feature node that appears before the input node in the pathological process. The tracing basis is the preset pathological logical association parameters in the bidirectional knowledge transmission link, including: Based on the feature description and anatomical location information of the input node and the matching parameters of the input node in the bidirectional knowledge transmission link, the specific coordinate position of the input node in the bidirectional knowledge transmission link is determined. The vertical transmission path rule corresponding to the coordinate position in the bidirectional knowledge transmission link is retrieved. The vertical transmission path rule is constructed based on the pathological development process of the defect and includes the possible upstream node identifier and transmission condition parameters corresponding to each node. Based on the vertical transmission path rules, potential upstream nodes that are directly associated with the input node are screened. The potential upstream node is the preceding feature node that directly leads to the appearance of the input node in the pathological process. Each input node corresponds to multiple potential upstream nodes. Extract the feature description, anatomical location coordinates, and transmission condition parameters of each potential upstream node from the defect knowledge graph; The transmission condition parameters of each potential upstream node are compared with the actual parameters of the input node. The actual parameters of the input node include its corresponding anatomical region state parameters and feature performance values. Potential upstream nodes whose conduction condition parameters match the actual parameters of the input node are screened, while potential upstream nodes whose conduction condition parameters do not meet the requirements are excluded. The potential upstream nodes whose conduction condition parameters match the actual parameters of the input node are the actual upstream nodes of the input node in the pathological process. Based on the transmission logic between upstream and input nodes after analysis and screening of pathological mechanism data, the causal relationship chain that causes or leads to the appearance of input nodes is marked. The upstream nodes that precede the selected upstream nodes are traced back to the previous upstream nodes. The operations of potential node screening, transmission condition parameter comparison, and logic verification are repeated until the starting node of the vertical transmission path is traced back. All upstream nodes in the entire process are recorded in the tracing order to form an upstream node sequence. The upstream node sequence contains information about each upstream node, the transmission logic with adjacent nodes, and the transmission condition parameters. By integrating the correlation data between upstream node sequences and input nodes, the transmission relationship type and pathological evidence identifier between the input node and each upstream node are labeled to form the vertical upstream node tracing result of the input node.

7. The medical image feature transmission and recognition method according to claim 3, characterized in that, The step of using the cross-critical region fusion feature node as a connection hub to connect the initial defect feature transmission sub-links corresponding to different critical regions to form a preliminary cross-region transmission link includes: Extract all fusion feature nodes that simultaneously possess anatomical location information of multiple key regions from the enhanced correlation features. These fusion feature nodes are marked as cross-key region fusion feature nodes. Cross-key region fusion feature nodes establish feature transmission channels between different key regions. Label all key region names corresponding to each cross-key region fusion feature node, and record the feature performance parameters, associated node identifiers, and propagation direction parameters of the cross-key region fusion feature node in each key region to obtain complete information about the node in different regions. Retrieve the initial defect feature transmission sub-link corresponding to each cross-critical region fusion feature node. Each initial defect feature transmission sub-link corresponds to a critical region. The initial defect feature transmission sub-link contains the transmission logic and associated node data of the cross-critical region fusion feature node in the corresponding region. The transmission roles of cross-critical region fusion feature nodes in different initial defect feature transmission sub-links are analyzed. The transmission roles are divided into starting nodes, intermediate nodes, and ending nodes. The positioning of the transmission role is determined based on the position and transmission direction parameters of the node in the initial defect feature transmission sub-link. Based on the transmission role, the method of connecting different initial defect feature transmission sub-links to the cross-critical area fusion feature node is determined. If the cross-critical area fusion feature node is an intermediate node in multiple initial defect feature transmission sub-links, then a direct serial connection is used; if the roles are different, the connection order is adjusted according to the transmission logic. Using cross-critical region fusion feature nodes as connection hubs, all their corresponding initial defect feature transmission sub-links are sequentially connected in a determined connection manner to form a preliminary cross-region transmission link. Based on the pathological mechanisms and clinical patterns of defects, the consistency of conduction logic in the preliminary cross-regional conduction links is analyzed. It is then used to detect whether there are conflicts in the correlation of conduction direction after connecting conduction sub-links with different initial defect characteristics. If conflicts exist, the conduction direction parameters are adjusted according to the pathological mechanisms and clinical patterns. For transmission links with logical conflicts, the parameters are adjusted. Referring to the defect knowledge graph, the transmission logic that conforms to the pathological mechanism is selected to replace the conflicting links, and cross-regional transmission association data is added. The association data is used to describe the pathological basis and clinical case support of the transmission sub-links with different initial defect features connected by cross-key region fusion feature nodes. By integrating all connected cross-regional transmission links, a cross-regional defect feature transmission link framework covering all key areas is formed.

8. A medical image feature transmission and recognition system, comprising: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the medical image feature transmission and recognition method according to any one of claims 1-7 by executing the machine-executable instructions.

Images