A breast tumor multi-modal evidence chain fusion discrimination method and system

Abstract

The application discloses a breast tumor multi-modal evidence chain fusion discrimination method and system, and relates to the fields of medical data intelligent processing and multi-modal information fusion. A computer acquires breast images, breast ultrasound, laboratory test and clinical text data collected in advance, generates uniform semantic features through feature extraction, standardization and semantic embedding, constructs evidence nodes and directed evidence chains according to relevant consistency, generates consistency potential according to in-chain support, in-chain dominance and node confidence, forms a reasoning output sequence according to evidence chain identification and weighted fusion, outputs a breast tumor malignancy tendency probability, and records the evidence chain structure, potential and reasoning output sequence. The method can reduce the influence of low-consistency links on the fusion result and realize traceable verification of the discrimination process.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical data processing and multimodal information fusion technology, specifically to a method and system for multimodal evidence chain fusion and discrimination of breast tumors. Background Technology

[0002] Current methods for assisting in the diagnosis of breast tumors often require the integration of multi-source information, including medical imaging, laboratory tests, pathological data, and clinical records. Existing clinical information and intelligent analysis technologies typically extract features from different modalities separately, and then fuse them through vector concatenation, weighted summarization, or alignment to output the diagnosis result. Because different modalities differ significantly in terms of acquisition conditions, dimensional scale, noise distribution, and semantic hierarchy, existing methods have shortcomings in terms of unified representation and cross-modal consistency constraints. Especially when modality is missing, acquisition quality fluctuates, or distribution drift is caused by equipment differences, the comparability of the fused input is difficult to guarantee, and the accumulation of bias in a particular modality can easily occur, leading to limited diagnostic stability.

[0003] Some existing technologies incorporate rule bases, knowledge graphs, or interpretable models to organize multi-source information into evidence items and conduct reasoning. However, the dependency order, support relationships, and conflict relationships between evidence items are mostly implicit in model parameters or rule sets, lacking a structured expression of the reasoning connections between evidence, making it difficult to form verifiable reasoning paths. In scenarios where multiple reasoning paths coexist, common fusion strategies still use fixed weights, empirical weights, or simple confidence levels as weight sources, making it difficult to distinguish consistency differences within different evidence chains, unable to quantify the degree of structural consistency within the chain, and making it difficult to explicitly suppress contradictions in the output of the links. The fusion results are highly sensitive to local anomalous evidence.

[0004] Furthermore, existing systems mostly write the discrimination results and some input information to logs or databases, primarily recording result fields, timestamps, and input summaries. They lack verifiable binding of key intermediate data such as the evidence chain structure, chain-level outputs, and fusion weights, making it difficult to guarantee a stable correspondence between the discrimination results and the generation process. It also makes it difficult to verify the consistency of structural fragments, chain-level outputs, and weight information of specific links in review scenarios. Therefore, in the multimodal discrimination scenario for breast tumors, a data processing method is still needed to provide a unified semantic space input basis for multimodal features, enable evidence nodes to form organizational relationships based on cross-modal correlation consistency, generate evidence chain structures through inference associations, allow chain-level consistency to participate in fusion inference as parameters, and simultaneously record the association with the evidence chain structure, chain-level weight parameters, and inference output sequences while outputting auxiliary discrimination results in the form of breast tumor malignancy probability. Summary of the Invention

[0005] Based on the shortcomings of the prior art described above, the purpose of this invention is to provide a method and system for multimodal evidence chain fusion and discrimination of breast tumors to solve the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for multimodal evidence chain fusion and discrimination of breast tumors, comprising: The computer acquires pre-collected breast imaging data, breast ultrasound data, laboratory test data, and clinical text data as multimodal raw data; Based on the original multimodal data, feature extraction, distribution normalization, and semantic embedding mapping are performed to generate multimodal feature data in a unified semantic space; Evidence nodes are constructed based on the correlation and consistency among multimodal feature data, and confidence parameters corresponding to the evidence nodes are generated. Construct an evidence chain based on the inferential relationships between evidence nodes, and generate evidence chain structure data; Based on the intra-chain edge weights and candidate external edges of nodes in the evidence chain structure data, the intra-chain support sequence and intra-chain dominant sequence are generated in combination with the confidence parameters of the evidence nodes. The evidence chain consistency potential parameters are generated based on the intra-chain support sequence, intra-chain dominant sequence and the confidence parameters of the intra-chain nodes. Based on the evidence chain structure data and confidence parameters, the evidence chain reasoning output for each evidence chain is generated. The reasoning output sequence is formed according to the evidence chain identifier. The reasoning output sequence is weighted and fused according to the consistency potential parameter to generate a fusion discrimination result in the form of the probability of malignancy of breast tumor. Based on the fusion discrimination results, the evidence chain structure data, evidence chain consistency potential parameters, and inference output sequence are associated with the records.

[0007] The present invention is further configured such that generating multimodal feature data in a unified semantic space includes: Preprocessing is performed on the raw data for each modality to generate preprocessed data and quality weights; Modality coding mapping is performed on the preprocessed data to generate deep feature sequences; Weighted convergence is performed on the deep feature sequence to generate fixed-length modal features. Intensity modulation is then performed on the fixed-length modal features according to the quality weights to generate modulated modal features. Perform distribution normalization on the modulation mode features to generate aligned mode features; Based on preset embedding mapping parameters, the aligned modal features are mapped to a shared semantic space. Based on preset amplitude constraint rules, the mapping results are constrained to generate multimodal feature data in a unified semantic space.

[0008] The present invention is further configured such that constructing evidence nodes based on the correlation consistency between multimodal feature data includes: Based on the multimodal feature data and quality weights in the unified semantic space, the computational domain mapping is performed by calling the preset modality projection parameters to generate computational domain features and effective quality weights; Based on the preset consistency measurement rules, cross-modal correlation consistency processing is performed on the features of the computational domain, and relevant consistency data is generated by combining effective quality weights. Generate consistent center data based on relevant consistency data and computational domain characteristics; Modal allocation weights are generated based on preset node query parameters and relevant consistency data; Evidence node features are generated by performing aggregation based on modality weights, consensus center data, and computational domain features, and evidence node sets are encapsulated based on evidence node identifiers.

[0009] The present invention is further configured such that the generation of confidence parameters corresponding to the evidence nodes includes: Node quality aggregation data is generated based on quality weights and modality-assigned weights. Node consistency support data is generated based on relevant consistency data and modal weighting. Node deviation data is generated based on consensus center data and evidence node characteristics; Based on the pre-set confidence mapping rules, the node quality aggregation data, node consistency support data, and node deviation data are converted into confidence parameters, and constraints are applied to the confidence parameters according to the pre-set boundary rules.

[0010] The present invention is further configured such that the generated evidence chain structure data includes: Generate node pair inference association scores based on evidence node characteristics and confidence parameters; Directional processing is performed on the inference association scores based on the nodes to generate directed inference association strength data; Based on the directed inference association strength data, sparsity selection and acyclic constraint filtering are performed to generate directed adjacency structure data; Based on the directed adjacency structure data, the execution path is extracted to generate the evidence chain node sequence and the in-chain edge weight sequence, and the evidence chain set is generated. Evidence chain structure data is generated based on the evidence chain set and the directed adjacency structure data. The evidence chain structure data includes a node identifier set, an edge identifier set, an edge weight set, a chain identifier set, and a chain order index.

[0011] The present invention is further configured such that the consistency potential parameter for generating the chain of evidence includes: Based on the evidence chain structure data parsing, the evidence chain node sequence, the chain inner edge weight sequence, and the node external connection candidate set are obtained, and the chain inner sequence data is generated. Generate an in-chain support sequence based on the in-chain edge weight sequence and the confidence parameters of adjacent nodes in the chain; Generate the dominant sequence within the chain based on the candidate set of external connections of nodes and the sequence of edge weights within the chain; Generate a chain consistency kernel based on the in-chain support sequence, the in-chain dominant sequence, and the in-chain node confidence parameters; Evidence chain consistency potential parameters are generated based on the chain consistency kernel.

[0012] The present invention is further configured such that the inference output for each chain of evidence generated based on the evidence chain structure data and confidence parameters includes: Based on the evidence chain node sequence and the intra-chain edge weight sequence in the evidence chain structure data, the evidence node features and confidence parameters are called according to the node identifier index to generate the intra-chain inference input sequence; Confidence gating encoding is performed on the first node of the chain based on the in-chain inference input sequence to generate the initial state of the chain. The in-chain state is updated by gating recursion based on the in-chain edge weight sequence and the confidence parameters of adjacent nodes, generating an in-chain state sequence. Generate chain-level loss factors based on the in-chain state sequence and the chain's built-in confidence parameters; Evidence chain inference output is generated based on the chain end state and chain-level loss factor, forming an inference output sequence.

[0013] The present invention is further configured such that the weighted fusion of the inference output sequence based on the consistency potential parameter to generate the fusion discrimination result includes: Map the inference output sequence to the fusion domain output sequence; The fusion domain output sequence is weighted and fused according to the consistency potential energy parameter to generate the fusion domain output. The fusion discrimination result is generated based on the output of the fusion domain.

[0014] The present invention is further configured such that the step of associating the evidence chain structure data, evidence chain consistency potential parameters, and inference output sequence based on the fusion discrimination result includes: Based on the fusion discrimination results, evidence chain structure data, evidence chain consistency potential parameters and inference output sequence, normalized encoding is performed to generate sample keys, result keys and evidence chain keys; Based on the evidence chain key, the evidence chain structure data is segmented into chain-level segments, and a chain-level summary sequence is generated by combining the corresponding evidence chain consistency potential parameters and the corresponding evidence chain reasoning output. Generate the evidence commitment root and the evidence chain key containing the proof sequence based on the chain-level digest sequence; Based on the fusion discrimination result and the evidence commitment root, a result binding code is generated, and the result binding code corresponds to the result key; Generate append-only archived records based on the result binding code, evidence commitment root, sample key, and result key.

[0015] This invention also provides a multimodal evidence chain fusion and discrimination system for breast tumors, the system comprising: Feature embedding module: The computer acquires pre-collected breast imaging data, breast ultrasound data, laboratory test data, and clinical text data as multimodal raw data. Based on the multimodal raw data, feature extraction, distribution normalization, and semantic embedding mapping are performed to generate multimodal feature data in a unified semantic space. Evidence construction module: Constructs evidence nodes based on the correlation and consistency among multimodal feature data, and generates confidence parameters corresponding to the evidence nodes; Evidence Chain Construction Module: Constructs an evidence chain based on the reasoning relationships between evidence nodes and generates evidence chain structure data; Potential energy generation module: Based on the intra-chain edge weights and candidate external edges of nodes in the evidence chain structure data, and combined with the confidence parameters of the evidence nodes, generate intra-chain support sequences and intra-chain dominant sequences, and generate evidence chain consistency potential energy parameters based on the intra-chain support sequences, intra-chain dominant sequences and the confidence parameters of intra-chain nodes. The reasoning fusion module generates evidence chain reasoning output for each evidence chain based on the evidence chain structure data and confidence parameters. It forms a reasoning output sequence according to the evidence chain identifier and performs weighted fusion of the reasoning output sequence based on the consistency potential parameter to generate a fusion discrimination result in the form of the probability of malignancy of breast tumor. The associated archiving module associates and records evidence chain structure data, evidence chain consistency potential parameters, and inference output sequences based on the fusion discrimination results.

[0016] This invention provides a method and system for multimodal evidence chain fusion and discrimination in breast tumors. The method uses pre-collected breast imaging data, breast ultrasound data, laboratory test data, and clinical text data acquired by a computer as multimodal raw data. Based on the multimodal raw data, feature extraction, distribution normalization, and semantic embedding mapping are performed to generate multimodal feature data in a unified semantic space. Evidence nodes are constructed based on the correlation and consistency among the multimodal feature data, generating confidence parameters corresponding to each evidence node. An evidence chain is constructed based on the inference relationships between evidence nodes, generating evidence chain structure data. The method then determines the evidence chain structure based on the intra-chain edge weights within the evidence chain structure data. Candidate edges are connected to external nodes. In-chain support sequences and dominant sequences are generated based on the confidence parameters of the evidence nodes. Evidence chain consistency potential parameters are then generated based on the in-chain support sequences, dominant sequences, and confidence parameters of the in-chain nodes. Evidence chain inference outputs are generated for each evidence chain based on the evidence chain structure data and confidence parameters. Inference output sequences are formed according to the evidence chain identifiers. These inference output sequences are then weighted and fused based on the consistency potential parameters to generate a fusion discrimination result in the form of a probability of malignancy in breast tumors. The fusion discrimination result is used to associate and record the evidence chain structure data, evidence chain consistency potential parameters, and inference output sequences. The resulting beneficial effects include: 1. Unified semantic input: Based on the preprocessing, quality weighting, coding convergence, distribution normalization and shared semantic mapping of multimodal raw data, a unified semantic space feature input is formed to reduce the impact of modality scale differences and distribution drift on subsequent evidence organization; 2. Potential Constraint Fusion: Based on cross-modal correlation consistency, evidence nodes are constructed and confidence parameters are generated. Then, a directed evidence chain is constructed and structured data is formed. The evidence chain consistency potential parameter is introduced to weight and fuse the chain-level inference output, suppressing the amplification of inconsistencies and local biases within the chain on the fusion result, and reducing the impact of low-consistency links and local biases on the fusion result. 3. Linked and Traceable Archiving: Standardized coding is performed on the fusion judgment results, evidence chain structure data, consistency potential parameters and chain-level inference output to generate a key-value system and construct chain-level summaries, evidence commitment roots and binding codes to form append-based archive records, realizing verifiable association between the results and the data of the entire evidence chain, providing a consistent basis for review, auditing and traceability.

[0017] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart illustrating a method for multimodal evidence chain fusion and discrimination of breast tumors is shown as an exemplary embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the structure of a multimodal evidence chain fusion discrimination system for breast tumors, as shown in an exemplary embodiment of the present invention. Detailed Implementation

[0019] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0020] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0021] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0022] Example 1:

[0023] A multimodal evidence chain fusion method for breast tumor discrimination, such as Figure 1 As shown, it includes: The computer acquires pre-collected breast imaging data, breast ultrasound data, laboratory test data, and clinical text data as multimodal raw data; Based on the original multimodal data, feature extraction, distribution normalization, and semantic embedding mapping are performed to generate multimodal feature data in a unified semantic space; Evidence nodes are constructed based on the correlation and consistency among multimodal feature data, and confidence parameters corresponding to the evidence nodes are generated. Construct an evidence chain based on the inferential relationships between evidence nodes, and generate evidence chain structure data; Based on the intra-chain edge weights and candidate external edges of nodes in the evidence chain structure data, the intra-chain support sequence and intra-chain dominant sequence are generated in combination with the confidence parameters of the evidence nodes. The evidence chain consistency potential parameters are generated based on the intra-chain support sequence, intra-chain dominant sequence and the confidence parameters of the intra-chain nodes. Based on the evidence chain structure data and confidence parameters, the evidence chain reasoning output for each evidence chain is generated. The reasoning output sequence is formed according to the evidence chain identifier. The reasoning output sequence is weighted and fused according to the consistency potential parameter to generate a fusion discrimination result in the form of the probability of malignancy of breast tumor. Based on the fusion discrimination results, the evidence chain structure data, evidence chain consistency potential parameters, and inference output sequence are associated with the records.

[0024] This embodiment performs a binary classification aid for determining the benign or malignant tendency of breast tumors, and uses a set of exemplary input data to illustrate the data transfer relationship between the above steps. The example values ​​are only used to illustrate the data processing procedure of this application and should not be considered as clinical statistical conclusions or independent diagnostic conclusions. In the offline stage, the model uses training data labeled with the discrimination category to generate encoding parameters, fragment query parameters, modality projection parameters, node query parameters, and output mapping parameters. The training data and the validation data used to determine the parameters are independent of each other.

[0025] In this embodiment, candidate values ​​are enumerated item by item on the verification data. The first selection condition is that the number of correctly fused samples is greater. When the number of correctly fused samples corresponding to the candidate values ​​is the same, the second selection condition is that the total number of generated evidence chains is smaller. If they are still the same, the candidate value with the smaller value is selected. After the parameters are determined, the parameters, encoding rules, and model parameters are written into the same configuration version, and the specific parameter values ​​are fixed in the corresponding processing steps.

[0026] As a feasible training method, training data is divided into training, validation, and test sets according to the subjects being examined, with data from the same subject only entering one set; category labels are taken from pathological results or verified follow-up conclusions. The encoding network, relational metric operator, and gated recursive parameters are jointly trained. The loss function consists of auxiliary discriminant cross-entropy, cross-modal consistency constraints, and directed relational binary cross-entropy, with weights set to 1, 0.2, and 0.2 respectively. Adaptive moment estimation optimization with a batch size of 32 and an initial learning rate of 0.001 is used, with a maximum training duration of 100 epochs. Training stops when the validation loss fails to decrease for 10 consecutive epochs. After training, the model parameters are fixed, and the training parameters, thresholds, and model version number are written to the same configuration version. In this embodiment, the inference phase only processes pre-collected data and does not include the data collection steps performed on the human body. The output is the probability of malignancy of breast tumors and auxiliary discriminant labels.

[0027] The present invention is further configured such that generating multimodal feature data in a unified semantic space includes: Preprocessing is performed on the raw data of each modality to generate preprocessed data and quality weights. Specifically, uniform preprocessing is performed on the raw data of each modality: image data undergoes grayscale range correction, spatial resolution resampling, and region of interest cropping according to the calibration information of the acquisition equipment; ultrasound sequences undergo jitter suppression and keyframe screening according to inter-frame stability; inspection indicators undergo outlier truncation and missing value imputation according to dimensional rules; text data undergoes synonym normalization and negative semantic annotation according to the terminology dictionary. Then, quality weights are generated: the usability indicators applicable to the target modality are converted into quality scores ranging from zero to one; the missing proportion score is one minus the missing proportion; the noise level score and artifact severity score are one minus the noise anomaly severity and artifact anomaly severity output by the quality detection branch, respectively; timestamp consistency is one when the acquisition interval between adjacent modalities does not exceed 24 hours, and when it exceeds 24 hours but does not exceed 72 hours, it is the result of 72 hours minus the actual interval and then divided by 48 hours. The value is set to zero after 72 hours; the cropped region coverage is the ratio of the number of pixels in the effective region of interest to the number of pixels in the target region. The minimum quality score applicable to the target modality is taken as the quality weight, so that any deficiency in any key quality item can limit the contribution of that modality; when the quality weight is below zero, it is truncated to zero; when it is above one, it is truncated to one; when it is below 0.25, it is marked as low quality. This step outputs preprocessed data, quality weights, and low quality labels, providing a common entry point for subsequent encoding, modulation, and effectiveness screening; in the specific operation of this embodiment, the input data includes four modalities: breast imaging data is the head-and-tail image and internal-external oblique image of the target breast, the original format is DICOM, and after grayscale correction and resampling, it is cropped into a lesion region of interest of 512×512 pixels; ultrasound data is a lesion scanning sequence containing 24 frames, after inter-frame stability screening, 8 key frames are retained, and it is cropped into a lesion region of 256×256 pixels; the test data includes cancer antigen 15-3 of 18.6. U / mL, carcinoembryonic antigen was 2.4 ng / mL and platelet count was Numerical coding was completed according to the corresponding dimensional range; text data included age, symptoms, family history, and examination description. An example examination description was "a hypoechoic nodule was seen in the upper outer quadrant of the right breast, with an irregular shape and unclear borders." After terminology normalization and negative semantic annotation, a sequence of text fragments was formed. According to the above quality weight generation rules, the quality weights of the imaging, ultrasound, laboratory, and text modalities were 0.88, 0.82, 0.94, and 0.90, respectively, all of which were not less than 0.25. Therefore, all four modalities participated in the subsequent relevant consistency calculation. Modality coding mapping is performed on the preprocessed data to generate deep feature sequences. Specifically, modality coding mapping is called on the preprocessed data to generate deep feature sequences: the image modality uses a coding network that combines multi-scale convolution and self-attention to decompose the region of interest into hierarchical semantic segments and output segment sequences; the ultrasound modality uses a temporal self-attention network to encode keyframe sequences into temporal semantic sequences; the testing modality uses a tabular coding network with interactive attention to encode the indicator set into indicator semantic sequences; and the text modality uses a domain-adaptive semantic coding network to encode sentence segments into semantic sequences. The coding network parameters are learned through cross-modality consistency constraints during offline training. During training, the proximity relationships of different modalities of the same tested object in the semantic space and the separability relationships of different objects are constrained simultaneously. The deep feature sequences of each modality are output, preserving local evidence clues rather than compressing them into a single low-order statistic. Each coding network also generates segment confidence clues in the range of zero to one for each semantic segment through a confidence cue branch that shares the previous-level features with the deep feature sequences, and establishes a one-to-one correspondence according to the segment order. Weighted convergence is performed on the deep feature sequence to generate fixed-length modal features. Intensity modulation is then applied to these fixed-length modal features based on quality weights to generate modulated modal features. Specifically, for each modality's deep feature sequence, content relevance, segment confidence hints, and contextual consistency are generated segment by segment. Content relevance is determined by performing same-dimensional normalization on the semantic segment features and the corresponding modal segment query parameters obtained from offline training. The normalization results are multiplied dimension by dimension and summed. The summation result is then linearly mapped from a range of -1 to 1 to a range of 0 to 1 to obtain the content relevance. Segment confidence hints are generated by the confidence hints from the modality coding network. The output of the branch is shown. The method for determining context consistency is as follows: The normalized features of the current semantic segment and its adjacent semantic segments are multiplied dimension-wise and summed, then mapped to zero to one. When the current semantic segment has two adjacent segments, the smaller of the two mapping results is taken; when it is located at the sequence boundary, the mapping result of the only adjacent segment is taken; when the sequence contains only one segment, one is taken. The content relevance, segment confidence cue, and context consistency of the same semantic segment are multiplied item by item to obtain the importance score of the semantic segment. This score is divided by the sum of all importance scores in the same sequence to obtain the segment weight. When the sum of all importance scores is zero, the weight is determined by multiplying the sum of all importance scores by one. At time zero, the semantic segment with the highest content relevance is weighted to one, and the weights of the remaining segments are weighted to zero. When there are ties, the segment with the smallest sequence number is selected. Deep feature sequences are then converged based on segment weights to obtain fixed-length modal features. The amplitude of these fixed-length modal features is then adjusted using quality weights, and the minimum feature retention value is set to 0.05 to preserve the archived representation of low-quality modalities. Modalities marked as low-quality have their contribution set to zero during subsequent effective quality weight screening. This process ensures that importance scores have a clear data source, calculation order, normalization rules, and anomaly handling methods. For the four modalities mentioned above, corresponding encoding networks are used. Deep feature sequences are formed separately. Segment weights are generated according to the multiplicative coupling rules of content relevance, segment confidence cues, and contextual consistency. The segment weights for the image modality are 0.08, 0.52, 0.27, and 0.13; for the ultrasound modality, they are 0.11, 0.48, 0.29, and 0.12; for the test modality, they are 0.46, 0.31, and 0.23; and for the text modality, they are 0.18, 0.44, 0.26, and 0.12. The sum of each set of segment weights is one. Convergence and mass intensity modulation are performed based on the corresponding segment weights to obtain four fixed-length modal features. Distribution normalization is performed on the modulation mode features to generate aligned mode features. Specifically, distribution normalization is performed on the modulation mode features to generate aligned mode features: a reference center and a second-order statistical reference are preset for each mode, and the reference is generated from the training data or the runtime sliding reference library; the input features are first centered and then scaled according to the second-order statistical reference; to avoid numerical instability caused by second-order statistical degradation, a shrinkage coefficient is introduced to stabilize the second-order statistics. The shrinkage coefficient is selected from 0.05, 0.10, 0.15 and 0.20 according to the validation data. In this embodiment, 0.10 is used, and it is written into the configuration version together with the corresponding second-order statistical reference; the aligned mode features are output, so that different modes are comparable in scale and related structure, providing stable input for shared semantic mapping and reducing the impact of device differences and batch drift. Based on preset embedding mapping parameters, a shared semantic space mapping is performed on the aligned modal features. Based on preset amplitude constraint rules, amplitude constraints are applied to the mapping results to generate multimodal feature data in a unified semantic space. Specifically, the shared semantic space mapping is performed on the aligned modal features by calling the preset embedding mapping parameters: each modality is configured with an independent mapping parameter group, and the aligned features are projected onto the same dimension of the unified semantic space; the mapping parameters are obtained by offline cross-modal consistency training, and the version number is kept consistent during version changes to support traceability; amplitude constraints are applied after mapping: each mapped feature is first normalized according to the L2 norm, and then each component is restricted to a fixed amplitude domain from negative one to one to ensure that the outputs of different modalities are in a unified numerical domain and to avoid extreme amplitudes that could lead to bias in subsequent related consistency calculations.

[0028] The present invention is further configured such that constructing evidence nodes based on the correlation consistency between multimodal feature data includes: Based on multimodal feature data and quality weights in a unified semantic space, a computational domain mapping is performed using preset modality projection parameters to generate computational domain features and effective quality weights. Specifically, the computational domain mapping uses linear projection with bias correction to transform different modality features to the same computational dimension while maintaining the separability of semantic directions within a modality. The projection parameters are obtained through offline training, which constrains the output dimension to be consistent and the numerical domain to be stable. Subsequently, the quality weights are validated: when a modality has missing labels or its quality weight is below 0.25, the effective quality weight of that modality is reset to zero; when the quality weight exceeds one, it is truncated to one; the rest remain at their original values. The computational domain features and effective quality weights are then output, ensuring that subsequent consistency measurements only contribute to available modalities, reducing the perturbation of evidence organization by missing and low-quality inputs. Cross-modal relevance consistency processing is performed on computational domain features according to preset consistency measurement rules, and relevant consistency data is generated by combining effective quality weights. Specifically, the relevance consistency processing adopts a combination of normalized bilinear similarity with a metric matrix and quality gating: first, the semantic direction consistency between any two modalities is calculated according to the preset consistency measurement rules, and then the consistency degree is gating and attenuated by the effective quality weights of the two modalities, so that the high-quality modality contributes more significantly to the consistency and the low-quality modality contribution is compressed. Then, the pairwise consistency degrees of all modalities are organized into a consistency data structure, which includes modality pair identifier, similarity strength, gating strength and statistical summary. Consistency center data is generated based on relevant consistency data and computational domain characteristics. Specifically, the generation of consistency centers adopts a robust consistency latent variable solution approach: the computational domain characteristics of each modality are treated as observations of the same potential consistency center. Influence factors are assigned to different modalities according to effective quality weights, and stronger suppression is applied to observations with large deviations, making the consistency centers closer to cross-modal common semantics rather than being driven by a single modality. This is achieved through iterative updates: the consistency center is initialized as the combined output of the two effective modalities with the highest consistency; when only one effective modality exists, its output is used as the initial center; subsequently, in each round of updates... The new algorithm adjusts the contribution ratio of each modality based on its deviation from the current center. When the maximum value of the absolute change of each dimension of the center in two adjacent rounds is not greater than 0.001, the update stops. If this condition is not met, the algorithm iterates a maximum of 30 times. In this embodiment, distribution normalization, shared semantic space mapping, and computational domain mapping are performed on the four fixed-length modal features to obtain comparable computational domain features. Effective quality weights are used to control the consistency of pairwise modal correlation. After 9 iterations, the maximum value of the absolute change of each dimension of the consistent center in two adjacent rounds is 0.0008, which meets the convergence condition of 0.001. The consistent center data is then output. Modality allocation weights are generated based on preset node query parameters and relevant consistency data. Specifically, the node query parameters are obtained through offline training and express the preference patterns of different evidence nodes for different modality information. The generation process includes: calculating the matching strength between the query parameters of each evidence node and the features of each modality computation domain; simultaneously introducing the overall consistency support strength of the modality in the relevant consistency data as an adjustment term, so that the modality in the region with high cross-modality consistency is more likely to be assigned to key nodes; and performing normalization processing on the matching strength to obtain the allocation weights, so that the weights of the same evidence node for all modalities meet the total constraint. Evidence node features are generated by aggregation based on modal allocation weights, consistency center data, and computational domain features. The evidence node set is then encapsulated based on the evidence node identifier. Specifically, the aggregation process employs center-gated aggregation: first, consistency center gating is applied to each modal feature, retaining components aligned with the consistency center direction and suppressing those deviating; then, the features after each modal gating are weighted and aggregated according to the modal allocation weights to obtain the evidence node features; finally, amplitude constraints and numerical stabilization are applied to the evidence node features to form node representations that can be directly used for subsequent inference and correlation scoring. Subsequently, the evidence node set is encapsulated based on the evidence node identifier. The evidence node set includes node identifiers, node features, modal allocation weight references, and consistency center references, enabling the evidence nodes to structurally reflect cross-modal consistency support and modal complementary allocation.

[0029] The present invention is further configured such that the generation of confidence parameters corresponding to the evidence nodes includes: Node quality aggregation data is generated based on quality weights and modality allocation weights. Specifically, for each target evidence node, the modality allocation weight corresponding to that node is taken one by one, and the validity screening is performed on the quality weight of each modality: when a modality is missing and marked as unusable or the quality weight is lower than a preset threshold, the contribution of that modality is set to zero; when the quality weight exceeds the upper limit, saturation truncation is performed to avoid abnormally high values ​​dominating aggregation. Then, the modality allocation weight is used as the contribution coefficient to aggregate the quality weights of the effective modalities to obtain the node quality aggregation value, and at the same time, quality coverage information is generated: the proportion of the number of effective modalities participating in the aggregation to the number of configured modalities is counted, and the cases where the proportion of effective modalities is not higher than 0.25, higher than 0.25 but not higher than 0.75, and higher than 0.75 are mapped to low, medium, and high coverage levels, respectively. This processing ensures at the data level that low-quality or missing modalities are unlikely to increase the credibility of nodes, while the coverage level constrains the upper limit of subsequent mapping, avoiding the inconsistency of a small number of modalities supporting high confidence. Node consistency support data is generated based on relevant consistency data and modality allocation weights. Specifically, in the cross-modal relevant consistency data structure, the consistency strength of the modality pair corresponding to the high-weight modality of the evidence node is selected to form a candidate consistency set. A coupling coefficient of the node allocation weight is introduced into the consistency strength of each modality pair: the coupling coefficient is the joint contribution of the two modality allocation weights, ensuring that the modality pair that the node mainly relies on has a higher proportion in the support evaluation. Then, robust aggregation is performed on the candidate consistency set: first, the consistency support strength is accumulated, and then the consistency support strength is subjected to discrete penalty processing. The discrete penalty is obtained by comparing the proportion of high consistency items and low consistency items in the candidate consistency set. Modality pairs with a consistency strength of not less than 0.70 are recorded as high consistency items, and modality pairs with a consistency strength of less than 0.40 are recorded as low consistency items. When the number of low consistency items exceeds half of the number of candidate consistency items, the consistency support level is reduced by one level. This processing makes the node consistency support not only reflect the overall consistency strength, but also explicitly characterize whether there are paired conflicts, thereby suppressing the misleading effect of cross-modal contradictions on the subsequent inference chain. Node deviation data is generated based on the consistency center data and evidence node features. Specifically, same-dimensional alignment is performed on the evidence node features and consistency center features. Both are normalized according to the L2 norm, and the sum of the dimension-wise products of the normalized features is calculated as the direction alignment value, which is between -1 and 1. The direction alignment value is increased by one and divided by two to obtain the direction consistency degree. The direction deviation is obtained by subtracting the direction consistency degree from one, and the direction deviation is limited to zero and one. The L2 norms of the evidence node features and consistency center features are calculated separately, and the absolute value of the difference between the two norms is divided by the L2 norm of the consistency center feature. The larger value between the number and 0.001 is used to obtain the amplitude difference ratio; when the amplitude difference ratio is not higher than 0.05, the amplitude deviation is zero; when the amplitude difference ratio exceeds 0.05 but is not higher than 0.50, the amplitude deviation is obtained by subtracting 0.05 from the amplitude difference ratio and dividing by 0.45; when the amplitude difference ratio exceeds 0.50, the amplitude deviation is one; when the directional deviation exceeds 0.60, the amplitude deviation exceeds 0.50, and the consistency support strength is lower than 0.50, a structural deviation marker is generated. Thus, the directional deviation, amplitude deviation, and structural deviation marker can all be determined by the input features. According to pre-set confidence mapping rules, node quality aggregation data, node consistency support data, and node deviation data are converted into confidence parameters, and constraints are applied to the confidence parameters according to pre-set boundary rules. Specifically, the node quality aggregation value is multiplied by the node consistency support strength to obtain the basic confidence value; the direction deviation is subtracted from one to obtain the direction reduction coefficient, and the amplitude deviation is subtracted from one to obtain the amplitude reduction coefficient; when no structural deviation marker is generated, the structural reduction coefficient is set to one; when a structural deviation marker is generated, the structural reduction coefficient is set to 0.50; the basic confidence value, direction reduction coefficient, amplitude reduction coefficient, and structural reduction coefficient are multiplied to obtain the initial confidence value. Initial confidence values ​​below 0.05 are set to 0.05, and initial confidence values ​​above 0.95 are set to 0.95, resulting in the confidence parameter. Values ​​below 0.40, between 0.40 and 0.70, and above 0.70 are set to... The reliability levels were categorized as low, medium, and high. In this embodiment, the node quality convergence values ​​of the four evidence nodes were 0.95, 0.91, 0.88, and 0.82, respectively; the node consistency support strengths were 0.92, 0.92, 0.90, and 0.84, respectively; the directional deviations were 0.02, 0.03, 0.04, and 0.06, respectively; and the amplitude deviations were 0.02, 0.03, 0.04, and 0.06, respectively. No structural deviation markers were generated. After calculation according to the multiplicative rule and retaining two decimal places, the confidence levels of the four evidence nodes were 0.84, 0.79, 0.73, and 0.61, respectively. The first evidence node represents evidence of irregular morphology, the second evidence node represents evidence of boundary and internal echo, the third evidence node represents evidence of the correspondence between local image highlighting and examination description, and the fourth evidence node represents evidence supporting the examination and clinical records.

[0030] The present invention is further configured such that the generated evidence chain structure data includes: A node pair inference association score is generated based on the evidence node features and confidence parameters. Specifically, for any two evidence nodes, two ordered node pairs are constructed in different directions. For each ordered node pair, the features of the starting node, the features of the ending node, the dimension-wise product of the two node features, and the difference between the starting node features and the ending node features are concatenated in a fixed order and input into the relation measurement operator. The relation measurement operator includes a first linear layer, a modified linear unit, and a second linear layer. The output of the second linear layer is mapped to zero to one by a logic function to obtain the basic association strength. Different parameter matrices are used for the starting node and the ending node to facilitate the exchange. By ordering the nodes, we can obtain the basic association strength in different directions. In the offline stage, the relationship measurement operator uses node pairs labeled with directed relationships as positive samples and node pairs without directed relationships as negative samples. It is trained using binary cross-entropy. The square root of the product of the confidence parameters of two nodes is used as the credible coupling coefficient. The basic association strength is multiplied by the credible coupling coefficient to obtain the node pair inference association score. Scores below 0.05 are set to 0.05, and scores above 0.95 are set to 0.95. Thus, the node pair inference association score can be jointly determined by the ordered node features, the relationship measurement parameters, and the node confidence. Directional processing is performed on the inference association scores of the nodes to generate directed inference association strength data. Specifically, the direction of bidirectional scores of the same node pair is determined to generate directed inference association strength data. The directional processing adopts a relative dominance mechanism: the difference between the two directional scores of the node pair is compared. When the difference between the two directional scores exceeds 0.10, the direction with the higher score is determined as the directed relationship direction and the other direction is suppressed. When the difference does not exceed 0.10, no directed edge is generated for the node pair to reduce the risk of loops caused by uncertain directions. At the same time, the directed strength is restricted to the range of 0.05 to 0.95 and arranged from high to low directed strength for subsequent sparsification selection. The benefit of this processing is that it transforms the inference relationship from undirected similarity into a directed association with traceable order constraints, thus establishing a structural foundation for chain inference. Based on the directed inference association strength data, sparsity selection and acyclic constraint filtering are performed to generate directed adjacency structure data. Specifically, for each starting node, sparsity selection is performed among all its candidate successor edges. The sparsity selection adopts a strategy that limits the number of outgoing edges and prioritizes edge strength: first, candidate edges are sorted by directed inference association strength, and then the set of outgoing edges is determined according to the node confidence level: high-confidence nodes retain the top 3 edges in the strength ranking, medium-confidence nodes retain the top 2 edges, and low-confidence nodes retain the top 1 edge. Subsequently, acyclic constraint filtering is performed: the selected edge set is checked for closure using a combination of path reachability checks and closure penalty rules. When a newly added edge causes a loop, the edge causing the closure is removed according to the minimum cost elimination principle. The cost is calculated by the edge strength and the confidence level of the starting node. The decision ensures that removal prioritizes low-intensity or low-confidence relevant edges. This process controls structural complexity while suppressing reasoning ambiguity caused by loops, making the generated adjacency structure more suitable for subsequent path extraction and chain-level reasoning. Node-based reasoning association scoring and directionality processing are performed on the four evidence nodes. The retained directed reasoning associations include: the first evidence node pointing to the second evidence node with an edge weight of 0.82; the second evidence node pointing to the third evidence node with an edge weight of 0.76; the first evidence node pointing to the fourth evidence node with an edge weight of 0.62; the fourth evidence node pointing to the third evidence node with an edge weight of 0.58; and the second evidence node pointing to the fourth evidence node with an edge weight of 0.55. All of these edges satisfy the direction threshold and maximum outgoing edge constraints, and no loops are formed after the acyclic filtering process. Path extraction is performed on the directed adjacency structure data to generate a sequence of evidence chain nodes and an in-chain edge weight sequence, thus creating an evidence chain set. Specifically, path extraction is performed on the directed adjacency structure to generate the evidence chain set. The path extraction adopts a mechanism of starting point identification-forward expansion-termination constraint: nodes with zero incoming edges or a maximum incoming edge strength lower than 0.40 are identified as candidate starting points, and those with higher confidence are prioritized among the candidate starting points; during forward expansion, the successor edge with the highest directed inference association strength is selected from the outgoing edge set of the current node, and when the outgoing edge set is exhausted, the successor edge is selected. When the strength difference between the first two candidate edges is no greater than 0.05, two branches are retained in parallel. The expansion is terminated when there is no available successor, the number of nodes in the chain reaches 5, or the product of the confidence of the nodes in the chain is less than 0.10. Then, the product of the strength of the nodes in the chain and the confidence of the nodes in the chain is used as the chain score. The first three candidate chains are retained in descending order of chain score to form an evidence chain set. This process can extract verifiable reasoning paths from the graph structure, avoid the excessive complexity caused by full graph enumeration, and reduce the noise chains introduced by weakly associated edges. Evidence chain structure data is generated based on the evidence chain set and directed adjacency structure data. This data includes a node identifier set, an edge identifier set, an edge weight set, a chain identifier set, and a chain order index. Specifically, the evidence chain set and directed adjacency structure are uniformly encapsulated to generate the evidence chain structure data. The encapsulation includes: a node identifier set to identify the range of nodes involved in the reasoning; an edge identifier set to identify directed reasoning-related edges; an edge weight set to record the reasoning association strength of each directed edge; a chain identifier set to identify the unique number of each evidence chain; and a chain order index to record the order of nodes within each chain and their corresponding edge weight references. Consistency checks are performed during encapsulation to ensure that adjacent nodes in the chain order index are in the edge identifier set. The presence of corresponding edges in the chain ensures a one-to-one correspondence between the edge weight set and the edge identifier, and a one-to-one correspondence between the chain identifier set and the chain sequence index. This processing enables the evidence chain structure data to have directly referenceable structural fields, providing stable input for the subsequent generation of consistency potential parameters and chain-level reasoning output, while improving the structured expression and traceability of the reasoning process. Based on the above directed adjacency structure, three evidence chains are extracted: the node order of the first evidence chain is the first evidence node, the second evidence node, and the third evidence node; the node order of the second evidence chain is the first evidence node, the fourth evidence node, and the third evidence node; and the node order of the third evidence chain is the first evidence node, the second evidence node, the fourth evidence node, and the third evidence node.

[0031] The present invention is further configured such that the consistency potential parameter for generating the chain of evidence includes: Based on the evidence chain structure data parsing, the evidence chain node sequence, the chain edge weight sequence, and the node external connection candidate set are obtained, generating the chain sequence data. Specifically, the evidence chain structure data is parsed chain by chain to obtain the evidence chain node sequence and the chain edge weight sequence, and at the same time, the node external connection candidate set is generated. The parsing process adopts the index consistency rule: the node identifier is read by the chain order index, and the chain edge weight reference is read by the adjacent node pair. For each node, the entire set of external connection candidate edges and their edge weights are read from the adjacency structure as the candidate domain for subsequent dominance calculation. During parsing, integrity verification is performed: ensuring that there are corresponding directed edges between adjacent nodes in the chain in the adjacency structure, ensuring that the chain edge weights and edge identifiers correspond one-to-one, and ensuring that the external connection candidate set does not contain missing or invalid edges. This step makes the chain information have a directly computable sequence form, providing a consistent data entry point for the generation of the supporting sequence and the dominant sequence. An intra-chain support sequence is generated based on the intra-chain edge weight sequence and the confidence parameters of adjacent nodes within the chain. Specifically, for each intra-chain edge, the square root of the product of the confidence of the starting node and the confidence of the ending node is calculated. This square root is then multiplied by the intra-chain edge weight to obtain the corresponding segment-level support value. Segment-level support values ​​below 0.05 are set to 0.05, and segment-level support values ​​above 0.95 are set to 0.95. The intra-chain support sequence is then formed according to the order of the intra-chain edges. This calculation simultaneously incorporates node confidence and intra-chain connection strength, so that when the confidence of any end node or the edge weight decreases, the segment-level support value decreases accordingly. An intra-chain dominant sequence is generated based on the candidate set of outgoing nodes and the sequence of intra-chain edge weights. Specifically, for each intra-chain edge, an intra-chain dominant sequence is generated based on the candidate set of outgoing nodes of the starting node of that edge. The dominance calculation adopts a relative advantage mechanism: the intra-chain edge weight is used as the target edge strength, and the strengths of all outgoing candidate edges of the starting node are taken to form a comparison set. The relative dominance of the target edge in the comparison set is calculated by dividing the target edge strength by the sum of the strengths of all outgoing candidate edges of the same starting node. When the dominance is not lower than 0.60, it is marked as high dominance; when it is lower than 0.40, it is marked as low dominance; otherwise, it is marked as medium dominance. When the dominance is lower than 0.05, it is set to 0.05; when it is higher than 0.95, it is set to 0.95. This avoids the incomparability of dominance caused by the difference in the outgoing edge scale of different nodes. This step makes the credibility of the intra-chain edge not only depend on the absolute edge weight, but also reflects whether the choice is the most reasonable locally, thereby suppressing the erroneous links caused by local structural ambiguity from being given excessive potential. A chain consistency kernel is generated based on the in-chain support sequence, the in-chain dominant sequence, and the in-chain node confidence parameters. Specifically, assuming the evidence chain contains m in-chain edges and n nodes, when m is greater than zero, the support convergence value is obtained by taking the m-th root of the product of the m segment-level support values, and the dominant convergence value is obtained by taking the m-th root of the product of the m dominant values. When m equals zero, both convergence values ​​are set to one. The proportion of low-confidence nodes is obtained by dividing the number of nodes with a confidence level below 0.40 by n, and the proportion of nodes with the longest consecutive low-confidence segment is obtained by dividing the number of nodes with the longest consecutive low-confidence segment by n. n yields the continuous gap ratio. The larger of the two values ​​is taken as the confidence gap ratio. The gap retention coefficient is obtained by subtracting the confidence gap ratio from one. The smaller of the result of dividing three by n and one is taken as the chain length retention coefficient. Evidence chains with more than three nodes are attenuated according to chain length. The support convergence value, dominant convergence value, gap retention coefficient, and chain length retention coefficient are multiplied to obtain the chain consistency kernel. Chain consistency kernels below 0.05 are set to 0.05, and chain consistency kernels above 0.95 are set to 0.95. Evidence chain consistency potential parameters are generated based on the chain consistency kernel. Specifically, the chain consistency kernel of each evidence chain is used as the initial potential value of that chain. The sum of all initial potential values ​​for the same sample is calculated, and the initial potential value of each evidence chain is divided by this sum to obtain the corresponding evidence chain consistency potential parameters. In this embodiment, segment-level support values ​​are calculated based on node confidence levels of 0.84, 0.79, 0.73, and 0.61, and intra-chain edge weights of 0.82, 0.76, 0.62, 0.58, and 0.55. Then, the degree of dominance is calculated based on the ratio of the intra-chain edge weight of each chain to the sum of all external candidate edge weights from the same starting point. The first evidence chain, the second evidence chain... The support convergence values ​​of the first and third evidence chains are 0.620909, 0.414453, and 0.462155, respectively, and the dominant convergence values ​​are 0.574774, 0.656167, and 0.620651, respectively. The gap retention coefficients are all one, and the chain length retention coefficients are one, one, and 0.75, respectively, resulting in chain consistency kernels of 0.356882, 0.271951, and 0.215128. The sum of the consistency kernels of the three chains is 0.843961. After cross-chain normalization, the consistency potential parameters of the three evidence chains are 0.423, 0.322, and 0.255, respectively, and the sum of the three is one.

[0032] The present invention is further configured such that the evidence chain reasoning output for generating each evidence chain based on the evidence chain structure data and confidence parameters includes: Based on the evidence chain node sequence and intra-chain edge weight sequence in the evidence chain structure data, the evidence node features and confidence parameters are retrieved by indexing the node identifier to generate the intra-chain inference input sequence. Specifically, the evidence chain structure data is retrieved chain by chain, and the evidence chain node sequence is read by indexing according to the chain order, while the intra-chain edge weight sequence corresponding to the adjacent nodes is read. Then, based on the node identifier, the node features and node confidence parameters are indexed in the evidence node set, and the intra-chain inference input sequence is formed according to the chain order. Consistency verification is performed when retrieving data: ensuring that adjacent nodes in the chain have corresponding items in the edge weight sequence, ensuring that the node identifier can be indexed to the node features and confidence parameters, and ensuring that there are no duplicate return visit marks in the chain nodes, so as to maintain the unidirectional propagation assumption of chain inference. This process aligns the intra-chain input in terms of time sequence, structural constraints, and credibility constraints, providing a stable data entry for subsequent gating recursion, and ensuring that the inference output can be reversed to locate specific links and nodes. The confidence-gated encoding of the first node in the chain is performed based on the in-chain inference input sequence to generate the initial state of the chain. Specifically, the features of the first node in the chain are input into the linear projection layer, the projection result is restricted to between -1 and 1 by the hyperbolic tangent function, and then multiplied by the larger value between the confidence parameter of the first node in the chain and 0.05 to obtain the initial state of the chain. The linear projection parameters are fixed after offline training and the version number is recorded. The in-chain state is recursively updated using gated operations based on the in-chain edge weight sequence and the confidence parameters of adjacent nodes, generating an in-chain state sequence. Specifically, the smaller values ​​among the current state, the next node's projected features, the current in-chain edge weights, and the confidence scores of the two adjacent nodes are linearly transformed and summed, then mapped using a logistic function to obtain a transitive gate. The current state and the next node's projected features are then subjected to another set of linear transformations and summed, then passed through a hyperbolic tangent function to obtain candidate states. The transitive gate is subtracted from 1 as the retention ratio of the current state, and the transitive gate is used as the injection ratio of the candidate states. The two parts are added together to obtain the updated state, and the components of the updated state are restricted to between -1 and 1. The above updates are repeated according to the in-chain node order to form an in-chain state sequence. A chain-level loss factor is generated based on the in-chain state sequence and the chain's built-in confidence parameters. Specifically, the number of nodes with a confidence level below 0.40 is divided by the total number of nodes in the chain to obtain the proportion of low-confidence nodes; the number of nodes with the longest consecutive low-confidence segments is divided by the total number of nodes in the chain to obtain the proportion of consecutive gaps; the number of updates with a transmission gating level below 0.30 is divided by the total number of in-chain updates to obtain the low-gating ratio, which is zero when there are no in-chain updates; the initial loss value is obtained by adding 0.50 times the low-confidence node ratio, 0.30 times the consecutive gap ratio, and 0.20 times the low-gating ratio; the initial loss value is set to 0.05 for values ​​below 0.05 and 0.95 for values ​​above 0.95 to obtain the chain-level loss factor. Evidence chain inference outputs are generated based on the chain end state and chain-level loss factor, forming an inference output sequence. Specifically, the chain end state is input into the linear output layer and mapped by a logic function to obtain the original probability of malign tendency. The difference between the original probability and 0.50 is multiplied by one minus the chain-level loss factor, and then 0.50 is added to obtain the reliability modulation probability. The reliability modulation probability not higher than 0.001 is set to 0.001, and the reliability modulation probability not lower than 0.999 is set to 0.999, obtaining the corresponding evidence chain inference output. In this embodiment, none of the three evidence chains have low confidence nodes or low-gated updates, and the chain-level loss factor is set to the minimum retention value of 0.05. The original probabilities of malign tendency for the three evidence chains are 0.8263, 0.6895, and 0.5737, respectively. After reliability modulation and retention to two decimal places, the evidence chain inference outputs are 0.81, 0.68, and 0.57, respectively, forming an inference output sequence according to the evidence chain identifier.

[0033] The present invention is further configured such that the weighted fusion of the inference output sequence based on the consistency potential parameter to generate the fusion discrimination result includes: The inference output sequence is mapped to the fusion domain output sequence. Specifically, fusion domain mapping is performed on the inference output of each evidence chain. The fusion domain mapping adopts a discriminant domain consistency strategy: when the inference output is in probabilistic form, probability pruning is first performed, setting probabilities no higher than 0.001 to 0.001 and probabilities no lower than 0.999 to 0.999 to avoid zero or one values ​​that would cause subsequent fusion numerical degradation. Then, logarithmic ratio transformation is performed to convert the probabilistic form into a fusion-compatible discriminant strength representation, so that positive and negative evidence are symmetrically expressed in the fusion domain. When the inference output is a multi-class probability vector, the same pruning and transformation are performed on each class component, while maintaining the consistency of the component order. During the fusion domain mapping process, the output is retained to 6 decimal places to complete fixed-point processing, ensuring the serialization consistency of different operating platforms. This processing makes the outputs of different evidence chains comparable in the same fusion domain, avoiding the excessive compression of small probabilities or the excessive amplification of large probabilities caused by direct superposition in the probability domain. The fusion domain output sequence is weighted and fused according to the consistency potential energy parameter to generate the fusion domain output. Specifically, the fusion domain output sequence is weighted and fused according to the consistency potential energy parameter. The weighted fusion adopts a potential energy normalization weighting mechanism: first, cross-chain normalization is performed on the potential energy parameter to ensure that the potential energy of each chain within the same sample meets the total constraint and maintains a relative proportion; when the potential energy parameter of a single evidence chain exceeds 0.90, it is first truncated to 0.90 before cross-chain normalization is performed; when the sum of the potential energy parameters of all evidence chains is less than 0.10, the potential energy parameter of each evidence chain is increased by 0. After 01, cross-chain normalization is re-executed; then, using the normalized potential energy as weight, the outputs of each chain's fusion domain are weighted and summed to obtain the fusion domain output; stabilization compression is performed on the fusion domain output, setting the discrimination strength below -8 to -8 and the discrimination strength above 8 to 8, to avoid nonlinear amplification of the single abnormal chain output by subsequent discrimination mapping; this process introduces the structural consistency of the evidence chain into the fusion weight, enabling the fusion process to automatically improve the contribution of structurally consistent and confidently continuous links, and suppress the impact of structural ambiguity or low-consistency links on the final result; Based on the output of the fusion domain, a fusion discrimination result is generated. Specifically, the output of the fusion domain undergoes an inverse transformation corresponding to the logarithmic ratio transformation to obtain the probability of malignancy of the breast tumor. When the probability of malignancy is not less than 0.50, a malignancy auxiliary discrimination label is generated; when it is less than 0.50, a benignity auxiliary discrimination label is generated. The results with the highest category probability below 0.60, between 0.60 and 0.80, and above 0.80 are respectively marked as low, medium, and high confidence levels. For this embodiment, the logarithmic ratios corresponding to the evidence chain inference outputs of 0.81, 0.68, and 0.57 are 1.450010, 0.753772, and 0.753772, respectively. The result is 0.281851; weighted with 0.423, 0.322 and 0.255 as consistency potential parameters for the corresponding evidence chain, the fusion domain output is 0.927941; performing the inverse transformation yields a malignancy probability of 0.716657, which is rounded to two decimal places as 0.72. Since 0.72 is not less than 0.50, the fusion discrimination result is a malignancy auxiliary discrimination label; since the highest category probability is between 0.60 and 0.80, the confidence level is medium. This output only represents auxiliary information obtained by the computer processing pre-collected multimodal medical data and needs to be reviewed by professionals in conjunction with clinical data.

[0034] The present invention is further configured such that the step of associating the evidence chain structure data, evidence chain consistency potential parameters, and inference output sequence based on the fusion discrimination result includes: Based on the fusion discrimination results, evidence chain structure data, evidence chain consistency potential energy parameters, and inference output sequence, normalized encoding is performed to generate sample keys, result keys, and evidence chain keys. Specifically, normalized encoding is performed on the fusion discrimination results, evidence chain structure data, evidence chain consistency potential energy parameter sequence, and inference output sequence. The normalized encoding adopts a deterministic serialization plus versioning rule approach: unified conventions are made for field order, numerical precision, null value representation, chain order index expansion method, and edge weight list arrangement method; the potential energy parameters and inference output are rounded to 6 decimal places according to the rounding rules, and pruning is performed according to the upper and lower limits determined in the generation steps of each field to ensure that the same input is used in the same way. The same encoding results are obtained under different operating environments; version numbers are assigned to all encoding rules and carried in the records; then sample keys, result keys, and evidence chain keys are generated. The sample key is obtained by using a deterministic digest of the inspected object identifier, collection batch identifier, and time count, and is used to uniquely point to the input sample; the result key is obtained by using a deterministic digest of the sample key and the discrimination result code, and is used to uniquely point to the discrimination output; the evidence chain key is obtained by using a deterministic digest of the evidence chain identifier, chain sequence index code, and chain edge weight code, and is used to uniquely point to a certain evidence chain. This process provides a stable index foundation for the three types of objects: sample, result, and evidence chain, which facilitates subsequent chain-level segmentation, proof generation, and archiving retrieval. The evidence chain structure data is segmented at the chain level based on the evidence chain key, and a chain-level summary sequence is generated by combining the corresponding evidence chain consistency potential parameters and the corresponding evidence chain inference output. Specifically, the evidence chain structure data is segmented at the chain level based on the evidence chain key to obtain the structural fragments corresponding to each evidence chain. The chain-level segmentation adopts a chain identifier-driven segmentation method: the chain sequence index, the edge weight reference within the chain, and the node identifier set are located according to the chain identifier to form a chain structure fragment; then, the consistency potential parameters and the inference output corresponding to the chain are synchronously indexed according to the chain identifier to form a chain-level record unit; subsequently, a chain-level summary sequence is generated. The generation of the chain-level summary adopts a concatenated summary method of structural fragment encoding, chain potential encoding, and chain output encoding: first, the chain structure fragment is serialized according to the normalization rules in the first step; then, the corresponding potential parameters and inference output are encoded with fixed precision and concatenated; finally, anti-collision summarization is performed on the concatenated result to generate a chain-level summary. This process enables each evidence chain to form an independent summary, which can represent the chain structure and bind the chain weights and chain outputs, providing the smallest verifiable unit for the subsequent commitment root construction. Evidence commitment roots and corresponding proof sequences are generated based on the chain-level summary sequence. Specifically, evidence commitment roots are constructed for the chain-level summary sequence using a Merkle tree commitment method: the chain-level summary sequence is sorted according to a preset sorting rule, which uses the lexicographical order of the evidence chain keys to ensure that a unique tree structure is generated for the same input; then, upper-level summaries are generated by combining them layer by layer until the root summary is obtained, which serves as the evidence commitment root; at the same time, a proof sequence is generated for each evidence chain, which records the set of path summaries from the chain-level summary to the evidence commitment root, used to verify that a certain evidence chain has indeed been included in the current judgment archive without exposing all chain-level summaries; this process compresses the archive information of multiple evidence chains into a single root commitment, while retaining verifiable proof materials for each chain, improving recording efficiency and enhancing verifiability; Based on the fusion discrimination result and the evidence commitment root, a result binding code is generated, which corresponds to the result key. Specifically, the generation of the result binding code for the fusion discrimination result and the evidence commitment root adopts an authentication code method that binds the result code and the evidence root: the normalized code of the fusion discrimination result and the evidence commitment root are concatenated in a fixed order, and the result key and rule version number are added as domain separation markers. Then, a preset authentication algorithm is called to generate the binding code. The authentication algorithm uses a key message authentication code based on SHA-256. The key or signature certificate is managed by the system security configuration and includes a version number. The result binding code corresponds one-to-one with the result key, ensuring that the binding verification fails when the result is tampered with or the evidence root is replaced. This ensures that the discrimination result and the chain-level evidence set remain inseparable, facilitating regulatory review and process auditing. Append-oriented archive records are generated based on the result binding code, evidence commitment root, sample key, and result key. Specifically, append-oriented archive records are generated based on the result binding code, evidence commitment root, sample key, and result key. Append-oriented archiving uses a method of referencing previous records plus a current record header summary: the header summary of the previous record is maintained in the archive storage; the current record encodes the header summary of the previous record, sample key, result key, evidence commitment root, and result binding code in a fixed order and generates the current record header summary, forming a chain connection between records; when an intermediate record is deleted or replaced, subsequent record header summaries are checked for mismatch; the archive record content includes the sample key, result key, evidence commitment root, result binding code, chain-level summary sequence reference information, and evidence sequence reference information, and, if necessary, includes a summary reference of the chain identifier set and chain sequence index to support evidence chain key-based archiving. Rapid location is achieved; this process enables the key intermediate information of a single judgment to be persistently stored in a verifiable manner, improving the traceability of chain structure, chain potential energy, and chain output during review, while reducing storage burden and maintaining the integrity and consistency of records; for the above fusion judgment results, the evidence chain structure data, the consistency potential energy parameters of the three evidence chains, and the inference output sequence are normalized and encoded to generate sample keys, result keys, and three evidence chain keys; chain-level summaries are generated chain by chain, evidence commitment roots and containing proof sequences are constructed according to the lexicographical order of the evidence chain keys, and then result binding codes are generated based on the fusion judgment results and evidence commitment roots, and written into an append-only archived record containing the header summary of the preceding record. Thus, the original multimodal inputs, evidence nodes, evidence chains, consistency potential energy, evidence chain inference outputs, and fusion judgment results of the same examined object form verifiable associated records.

[0035] Example 2:

[0036] Please see Figure 2 This exemplary multimodal evidence chain fusion discrimination system for breast tumors includes: Feature embedding module: The computer acquires pre-collected breast imaging data, breast ultrasound data, laboratory test data, and clinical text data as multimodal raw data. Based on the multimodal raw data, feature extraction, distribution normalization, and semantic embedding mapping are performed to generate multimodal feature data in a unified semantic space. Evidence construction module: Constructs evidence nodes based on the correlation and consistency among multimodal feature data, and generates confidence parameters corresponding to the evidence nodes; Evidence Chain Construction Module: Constructs an evidence chain based on the reasoning relationships between evidence nodes and generates evidence chain structure data; Potential energy generation module: Based on the intra-chain edge weights and candidate external edges of nodes in the evidence chain structure data, and combined with the confidence parameters of the evidence nodes, generate intra-chain support sequences and intra-chain dominant sequences, and generate evidence chain consistency potential energy parameters based on the intra-chain support sequences, intra-chain dominant sequences and the confidence parameters of intra-chain nodes. The reasoning fusion module generates evidence chain reasoning outputs for each evidence chain based on the evidence chain structure data and confidence parameters. It forms a reasoning output sequence according to the evidence chain identifiers and performs weighted fusion of the reasoning output sequences based on the consistency potential parameter to generate a fusion discrimination result in the form of the probability of malignancy of breast tumors. The associated archiving module associates and records evidence chain structure data, evidence chain consistency potential parameters, and inference output sequences based on the fusion discrimination results.

[0037] It should be noted that the breast tumor multimodal evidence chain fusion and discrimination system provided in the above embodiments and the breast tumor multimodal evidence chain fusion and discrimination method provided in the above embodiments belong to the same concept. The specific operation methods of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the breast tumor multimodal evidence chain fusion and discrimination system provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0038] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for multimodal evidence chain fusion and discrimination of breast tumors, characterized in that, include: The computer acquires pre-collected breast imaging data, breast ultrasound data, laboratory test data, and clinical text data as multimodal raw data; Based on the original multimodal data, feature extraction, distribution normalization, and semantic embedding mapping are performed to generate multimodal feature data in a unified semantic space; Evidence nodes are constructed based on the correlation and consistency among multimodal feature data, and confidence parameters corresponding to the evidence nodes are generated. Construct an evidence chain based on the inferential relationships between evidence nodes, and generate evidence chain structure data; Based on the intra-chain edge weights and candidate external edges of nodes in the evidence chain structure data, the intra-chain support sequence and intra-chain dominant sequence are generated in combination with the confidence parameters of the evidence nodes. The evidence chain consistency potential parameters are generated based on the intra-chain support sequence, intra-chain dominant sequence and the confidence parameters of the intra-chain nodes. Based on the evidence chain structure data and confidence parameters, the evidence chain reasoning output for each evidence chain is generated. The reasoning output sequence is formed according to the evidence chain identifier. The reasoning output sequence is weighted and fused according to the consistency potential parameter to generate a fusion discrimination result in the form of the probability of malignancy of breast tumor. Based on the fusion discrimination results, the evidence chain structure data, evidence chain consistency potential parameters, and inference output sequence are associated with the records.

2. The method for multimodal evidence chain fusion and discrimination of breast tumors according to claim 1, characterized in that, Generating multimodal feature data in a unified semantic space includes: Preprocessing is performed on the raw data for each modality to generate preprocessed data and quality weights; Modality coding mapping is performed on the preprocessed data to generate deep feature sequences; Weighted convergence is performed on the deep feature sequence to generate fixed-length modal features. Intensity modulation is then performed on the fixed-length modal features according to the quality weights to generate modulated modal features. Perform distribution normalization on the modulation mode features to generate aligned mode features; Based on preset embedding mapping parameters, the aligned modal features are mapped to a shared semantic space. Based on preset amplitude constraint rules, the mapping results are constrained to generate multimodal feature data in a unified semantic space.

3. The method for multimodal evidence chain fusion and discrimination of breast tumors according to claim 2, characterized in that, Evidence nodes are constructed based on the correlation and consistency among multimodal feature data, including: Based on the multimodal feature data and quality weights in the unified semantic space, the computational domain mapping is performed by calling the preset modality projection parameters to generate computational domain features and effective quality weights; Based on the preset consistency measurement rules, cross-modal correlation consistency processing is performed on the features of the computational domain, and relevant consistency data is generated by combining effective quality weights. Generate consistent center data based on relevant consistency data and computational domain characteristics; Modal allocation weights are generated based on preset node query parameters and relevant consistency data; Evidence node features are generated by performing aggregation based on modality weights, consensus center data, and computational domain features, and evidence node sets are encapsulated based on evidence node identifiers.

4. The method for multimodal evidence chain fusion and discrimination of breast tumors according to claim 3, characterized in that, The confidence parameters generated for the evidence nodes include: Node quality aggregation data is generated based on quality weights and modality-assigned weights. Node consistency support data is generated based on relevant consistency data and modal weighting. Node deviation data is generated based on consensus center data and evidence node characteristics; Based on the pre-set confidence mapping rules, the node quality aggregation data, node consistency support data, and node deviation data are converted into confidence parameters, and constraints are applied to the confidence parameters according to the pre-set boundary rules.

5. The method for multimodal evidence chain fusion and discrimination of breast tumors according to claim 4, characterized in that, The data for generating the chain of evidence structure includes: Generate node pair inference association scores based on evidence node characteristics and confidence parameters; Directional processing is performed on the inference association scores based on the nodes to generate directed inference association strength data; Based on the directed inference association strength data, sparsity selection and acyclic constraint filtering are performed to generate directed adjacency structure data; Based on the directed adjacency structure data, the execution path is extracted to generate the evidence chain node sequence and the in-chain edge weight sequence, and the evidence chain set is generated. Evidence chain structure data is generated based on the evidence chain set and the directed adjacency structure data. The evidence chain structure data includes a node identifier set, an edge identifier set, an edge weight set, a chain identifier set, and a chain order index.

6. The method for multimodal evidence chain fusion and discrimination of breast tumors according to claim 5, characterized in that, The parameters for generating a chain of evidence consistency potential include: Based on the evidence chain structure data parsing, the evidence chain node sequence, the chain inner edge weight sequence, and the node external connection candidate set are obtained, and the chain inner sequence data is generated. Generate an in-chain support sequence based on the in-chain edge weight sequence and the confidence parameters of adjacent nodes in the chain; Generate the dominant sequence within the chain based on the candidate set of external connections of nodes and the sequence of edge weights within the chain; Generate a chain consistency kernel based on the in-chain support sequence, the in-chain dominant sequence, and the in-chain node confidence parameters; Evidence chain consistency potential parameters are generated based on the chain consistency kernel.

7. The method for multimodal evidence chain fusion and discrimination of breast tumors according to claim 6, characterized in that, The evidence chain reasoning output generated for each evidence chain, based on the evidence chain structure data and confidence parameters, includes: Based on the evidence chain node sequence and the intra-chain edge weight sequence in the evidence chain structure data, the evidence node features and confidence parameters are called according to the node identifier index to generate the intra-chain inference input sequence; Confidence gating encoding is performed on the first node of the chain based on the in-chain inference input sequence to generate the initial state of the chain. The in-chain state is updated by gating recursion based on the in-chain edge weight sequence and the confidence parameters of adjacent nodes, generating an in-chain state sequence. Generate chain-level loss factors based on the in-chain state sequence and the chain's built-in confidence parameters; Evidence chain inference output is generated based on the chain end state and chain-level loss factor, forming an inference output sequence.

8. The method for multimodal evidence chain fusion and discrimination of breast tumors according to claim 7, characterized in that, The inference output sequence is weighted and fused based on the consistency potential parameter to generate the fused discrimination result, including: Map the inference output sequence to the fusion domain output sequence; The fusion domain output sequence is weighted and fused according to the consistency potential energy parameter to generate the fusion domain output. The fusion discrimination result is generated based on the output of the fusion domain.

9. The method for multimodal evidence chain fusion and discrimination of breast tumors according to claim 8, characterized in that, Based on the fusion discrimination results, the associated evidence chain structure data, evidence chain consistency potential parameters, and inference output sequence include: Based on the fusion discrimination results, evidence chain structure data, evidence chain consistency potential parameters and inference output sequence, normalized encoding is performed to generate sample keys, result keys and evidence chain keys; Based on the evidence chain key, the evidence chain structure data is segmented into chain-level segments, and a chain-level summary sequence is generated by combining the corresponding evidence chain consistency potential parameters and the corresponding evidence chain reasoning output. Generate the evidence commitment root and the evidence chain key containing the proof sequence based on the chain-level digest sequence; Based on the fusion discrimination result and the evidence commitment root, a result binding code is generated, and the result binding code corresponds to the result key; Generate append-only archived records based on the result binding code, evidence commitment root, sample key, and result key.

10. A multimodal evidence chain fusion and discrimination system for breast tumors, characterized in that, include: Feature embedding module: The computer acquires pre-collected breast imaging data, breast ultrasound data, laboratory test data, and clinical text data as multimodal raw data. Based on the multimodal raw data, feature extraction, distribution normalization, and semantic embedding mapping are performed to generate multimodal feature data in a unified semantic space. Evidence construction module: Constructs evidence nodes based on the correlation and consistency among multimodal feature data, and generates confidence parameters corresponding to the evidence nodes; Evidence Chain Construction Module: Constructs an evidence chain based on the reasoning relationships between evidence nodes and generates evidence chain structure data; Potential energy generation module: Based on the intra-chain edge weights and candidate external edges of nodes in the evidence chain structure data, and combined with the confidence parameters of the evidence nodes, generate intra-chain support sequences and intra-chain dominant sequences, and generate evidence chain consistency potential energy parameters based on the intra-chain support sequences, intra-chain dominant sequences and the confidence parameters of intra-chain nodes. The reasoning fusion module generates evidence chain reasoning output for each evidence chain based on the evidence chain structure data and confidence parameters. It forms a reasoning output sequence according to the evidence chain identifier and performs weighted fusion of the reasoning output sequence based on the consistency potential parameter to generate a fusion discrimination result in the form of the probability of malignancy of breast tumor. The associated archiving module associates and records evidence chain structure data, evidence chain consistency potential parameters, and inference output sequences based on the fusion discrimination results.

Images