Multi-view collective medical decision evaluation method and system based on dual-concept learning

Abstract

This invention provides a multi-view collective medical decision-making assessment method and system based on dual-concept learning, comprising: constructing a two-level decision interpretation mechanism to obtain decision concepts; modeling individual decisions of each view, inputting the decision concepts into an evidence-based neural network to obtain evidence quantity, and quantifying individual uncertainty; performing consensus judgment on individual decisions of each view, outputting a consistent collective opinion result, or obtaining prior importance and the importance of specific instances and modeling to obtain dynamic priorities, fusing each view according to the dynamic priorities to obtain a collective viewpoint and outputting it. This invention utilizes decision trajectory backtracking to achieve high transparency and interpretability in the decision-making process; it can proactively identify and warn of high-risk decisions, significantly improving the reliability of the system; thus, while retaining the differentiated information of the views, it achieves more logically reasonable and robust collective decision-making.

Description

Technical Field

[0001] This invention belongs to the field of medical intelligent decision-making technology, specifically, it relates to a multi-view collective medical decision-making evaluation method and system based on dual-concept learning, and in particular, a reliable and interpretable multi-view collective medical decision-making system based on dual-concept learning. Background Technology

[0002] Intelligent medical decision-making can provide reference suggestions, assisting medical professionals in developing treatment plans for patients more quickly and comprehensively. With the development of data acquisition technology, multi-view data has been widely used in fields such as medical diagnosis and autonomous driving. However, in actual decision-making processes, conflicts of opinion often arise among different perspectives due to differences in observation angles or professional backgrounds.

[0003] Existing multi-view decision and decision fusion technologies mainly suffer from the following shortcomings:

[0004] 1. The conflict resolution mechanism is too passive. Most existing methods treat differences between views as interference or negative factors, tending to seek consistency by minimizing or eliminating conflicts. This approach often obscures unique insights from different perspectives and fails to fully utilize the valuable information contained in the conflict.

[0005] 2. Lack of quantification of decision reliability. Traditional deep neural network models mostly focus only on classification accuracy, neglecting the reliability of the results. Although some evidence-based deep learning methods attempt to introduce uncertainty modeling, Evidential deep learning to quantify classification uncertainty (M. Sensoy, L. Kaplan, and M. Kandemir) mentions an evidence-based deep neural network, Evidential DNN, but when integrating conflicting information from multiple sources, it often cannot guarantee that the uncertainty of the integrated result will be reduced.

[0006] 3. Insufficient model interpretability. Most existing multi-view fusion models are "black box" models, capable of providing a final collective decision, but unable to clearly explain "why a certain view makes a specific prediction" or "why each view supports or opposes the final decision." Even existing conceptual bottleneck models often only focus on the reasons for the final decision, ignoring the intrinsic connection between the underlying features of the multi-view data and the decision concept.

[0007] The patent document "A Method and System for Constructing Reasoning Intelligent Agents for Intelligent Medical Guidance" (CN120525065A) discloses that by constructing a hierarchical modal completion network and a modal correlation knowledge graph, combined with contrastive learning and uncertainty quantification models, and dynamically adjusting modal weights, a high-quality diagnostic service for intelligent medical systems in areas with limited medical resources is achieved under incomplete data.

[0008] The patent document "Intelligent Decision System for Medical Diagnosis Based on Multimodal Data Fusion" (CN119495423A) discloses a method for data fusion using deep neural networks and Bayesian inference engines, and for optimizing feature weights using attention mechanisms to achieve personalized diagnostic analysis.

[0009] However, both mainly focus on weight optimization and dynamic adjustment of modal weights. Their core logic is still to seek a balance or consistency in multi-source data, and often regard differences and conflicts as interference.

[0010] The patent document "A Multi-Agent Collaborative Decision-Making System and Method Based on Neural Symbol Fusion" (CN120875036A) discloses a multi-agent collaborative decision-making system employing neural symbol fusion. It utilizes a neural symbol fusion engine to achieve two-layer representation and dynamic activation of domain knowledge, an agent coordinator optimizer to coordinate cognitive differences and verify candidate solutions, and an adaptive interpretation system to provide the decision evidence chain and learning feedback, forming a complete closed loop for decision optimization. It focuses on tracing the decision evidence chain, addressing the issue of transparency.

[0011] Meanwhile, existing technologies lack direct quantification of decision uncertainty. Bayesian uncertainty quantification is an indirect uncertainty assessment and is not suitable for multi-view fusion tasks.

[0012] Therefore, how to construct a system that can both actively handle view conflicts and provide reliability assessments and interpretability evidence for the medical decisions made is a technical problem that urgently needs to be solved in the field of intelligent medical decision-making. Summary of the Invention

[0013] To address the shortcomings of existing multi-view decision fusion methods, such as difficulty in quantifying reliability, lack of transparency in the decision-making process, and insufficient explanatory power for complex decision-making logic, the purpose of this invention is to provide a multi-view collective medical decision evaluation method and system based on dual-concept learning.

[0014] A multi-view collective medical decision assessment method based on dual-concept learning, provided by the present invention, includes: Step S1: Construct a two-level decision interpretation mechanism to obtain decision concepts. ; Step S2: Model the individual decisions for each view, and define the decision concepts. Input the evidence-based neural network to obtain the amount of evidence. And quantify individual uncertainty; Step S3: Make a consensus judgment on the individual decisions of each view and output a consistent collective opinion result or proceed to step S4; Step S4: Obtain prior importance and the importance of specific instances, and model to obtain dynamic priorities. According to dynamic priority Merge the various views to obtain a collective perspective and output it.

[0015] Preferably, step S1 includes: Step S1.1, for each view Using an encoder from the raw input Features are extracted and mapped to a feature concept space through an attention mechanism to obtain feature concepts. And guide the acquisition of enhanced features for each view. ; Step S1.2: Enhance the features Input capture network predicts individual decisions and their supportive attitudes toward collective decision-making, thus obtaining decision concepts. .

[0016] The decision-making concept There are two attitudes: one in favor and one against.

[0017] Where V represents the total number of views.

[0018] Preferably, in step S2, the non-negative excitation decision view output by the neural network is... For the Class of evidence :

[0019] in, This represents the parameters of the Dirichlet distribution.

[0020] view Individual uncertainty and its reliability for:

[0021]

[0022] in, Indicates the total number of categories; Indicates total intensity.

[0023] Preferably, the consensus judgment includes: For any view, if all individual decisions are the same and all attitudes are supportive, then the consensus is merged into a collective opinion result and output. If individual decisions differ or there is opposition, then step S4 is performed to merge the individual decisions.

[0024] Step S4 includes: Step S4.1: Obtain prior importance based on professional title. Based on the characteristics of different instances, through a dynamic perception layer The importance of obtaining a specific instance :

[0025] The superscript indicates the first. Parameters corresponding to each view; Represents a specific instance Dynamic perception; This indicates enhanced features.

[0026] Step S4.2: Use a normal distribution. Modeling is performed, based on evidence generation, using the Dirichlet distribution to generate instance-specific preferred distributions:

[0027] Step S4.3: Merge the viewpoints of each view according to dynamic priority to obtain a collective viewpoint:

[0028]

[0029]

[0030] in, This indicates the dynamic priority of the v-th view; This represents the amount of evidence for the v-th view; This represents the viewpoint of the v-th view; V represents the total number of views.

[0031] The present invention provides a multi-view collective medical decision evaluation system based on dual-concept learning, comprising: Module M1 obtains decision concepts based on a two-level decision interpretation mechanism. ; Module M2 models individual decisions for each view, and incorporates decision concepts. Input the evidence-based neural network to obtain the amount of evidence. And quantify individual uncertainty; Module M3 makes consensus judgments on individual decisions of each view and outputs the collective opinion result or triggers module M4 to perform collective decision fusion; Module M4 obtains prior importance and the importance of specific instances and models them to obtain dynamic priorities. According to dynamic priority By merging the various views, a collective perspective is obtained.

[0032] Preferably, the module M1 includes: Module M1.1, for each view Using an encoder from the raw input Features are extracted and mapped to a feature concept space through an attention mechanism to obtain feature concepts. And guide the acquisition of enhanced features for each view. ; Module M1.2 will enhance features Input capture network predicts individual decisions and their supportive attitudes toward collective decision-making, thus obtaining decision concepts. .

[0033] The decision-making concept There are two attitudes: one in favor and one against.

[0034] Where V represents the total number of views.

[0035] Preferably, in module M2, the non-negative excitation determination view output by the neural network... For the Class of evidence :

[0036] in, This represents the parameters of the Dirichlet distribution.

[0037] view Individual uncertainty and its reliability for:

[0038]

[0039] in, Indicates the total number of categories; Indicates total intensity.

[0040] Preferably, the consensus judgment includes: For any view, if all individual decisions are the same and all attitudes are supportive, then the consensus is merged into a collective opinion result and output. If individual decisions differ or there is opposition, the execution module M4 merges the individual decisions.

[0041] The module M4 includes: Module M4.1: Obtain prior importance based on professional title. Based on the characteristics of different instances, through a dynamic perception layer The importance of obtaining a specific instance :

[0042] The superscript indicates the first. Parameters corresponding to each view; Represents a specific instance Dynamic perception; This indicates enhanced features.

[0043] Module M4.2 adopts a normal distribution. Modeling is performed, based on evidence generation, using the Dirichlet distribution to generate instance-specific preferred distributions:

[0044] Module M4.3 merges the viewpoints of each view according to dynamic priority to obtain a collective viewpoint:

[0045]

[0046]

[0047] in, This indicates the dynamic priority of the v-th view; This represents the amount of evidence for the v-th view; This represents the viewpoint of the v-th view; V represents the total number of views.

[0048] A training method for a multi-view collective medical decision evaluation system based on dual-concept learning, provided by the present invention, includes: Step SA1: Parameterize the probability density function of the Dirichlet distribution. Obtain the second-order probability of the view and form multiple opinions. :

[0049] in, This represents the Beta function; express 1-dimensional simplex; This represents the class assignment probability of the nth type of simplicity in the vth view; This represents the probability of class assignment for the v-th view; The probability density function parameters represent the nth type Dirichlet distribution of the vth view; Let represent the probability density function parameters of the Dirichlet distribution for the v-th view.

[0050] Step SA2: Introduce the annealing coefficient Using the adjusted cross-entropy loss function KL divergence and binary cross-entropy Optimize the multi-view collective medical decision evaluation system based on dual-concept learning.

[0051] Preferably, step SA2 includes: Step SA2.1: Adjust the cross-entropy loss for a specific view:

[0052] in, Indicates the true predicted label for view v; Represents the double gamma function; Indicates by The integral of the defined cross-entropy loss function over the simplex; The total intensity is represented by k; k represents the category number, with a total of K categories.

[0053] Step SA2.2: Reduce the evidence for predicting mislabeled items to 0 using KL divergence:

[0054]

[0055] in, Indicates the removal of prediction parameters Dirichlet parameters after misleading evidence from China and Africa.

[0056] Introduce binary cross-entropy with logits loss:

[0057]

[0058] in, A binary value indicating whether the actual label and the collective label in the current view are consistent.

[0059] Step SA2.3: Obtain the loss. and its collective losses :

[0060]

[0061]

[0062] in, Indicates the annealing coefficient; Indicates the current training round; This indicates the preset number of annealing steps.

[0063] Compared with the prior art, the present invention has the following beneficial effects: 1. Compared with existing methods that only focus on prediction accuracy or simply eliminate view conflicts, this invention achieves high transparency and interpretability in the decision-making process by jointly mapping the concepts of feature level and decision level and utilizing decision trajectory backtracking.

[0064] 2. This invention introduces an evidence-based deep learning mechanism to achieve a quantitative assessment of the uncertainty of individual and collective decision-making. This enables the model to proactively identify and warn of high-risk decisions when processing conflict-based multi-view data, significantly improving the reliability of the system.

[0065] 3. This invention can flexibly adjust the voice of each view in collective decision-making based on the real-time reliability and attitude clarity of each view, thereby achieving a more logical and robust collective decision-making while preserving the differentiated information of each view. Attached Figure Description

[0066] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating a multi-view collective medical decision evaluation method based on dual-concept learning. Detailed Implementation

[0067] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0068] This invention provides a multi-view collective medical decision-making evaluation method based on dual-concept learning. It actively utilizes conflict and provides in-depth interpretation of the two-level system, offering an automated support tool that balances prediction accuracy, reliability, and decision transparency for complex decision-making scenarios such as multidisciplinary medical consultations. Specifically, it uses... Figure 1 For example, including: A two-level decision explanation mechanism is constructed, which reveals the reasoning logic of each doctor in conflict scenarios by constructing a "dual concept" mapping mechanism between feature-level concepts and decision-level concepts.

[0069] Step S1: Feature hierarchy concept extraction. For each view... Using an encoder from the raw input Features are extracted and mapped to a feature concept space through an attention mechanism to obtain feature concepts. Subsequently, the enhanced features of each view are obtained by being guided by the concept of features. .

[0070] In more preferred cases, multidisciplinary consultations are conducted for a single patient, integrating... Each view (corresponding to) (a famous doctor), and the patient information constitutes a multi-view feature set of the patient. .in, Indicates the first One characteristic.

[0071] Receive multi-view feature information The feature concepts are obtained by mapping them to the feature concept space through an attention mechanism. Guided by feature concepts, we can identify enhanced features from patient information that doctors are more interested in and that are more relevant to their own specialties. .

[0072] Step S2: Generation of decision-making hierarchy concepts. (This is for...) Predicting individual decisions and their attitudes toward collective decisions, and taking them as decision-making concepts. .

[0073] Specifically, after obtaining the enhanced features, they are input into a "supportive attitude" capture network to predict each doctor's attitude towards collective decision-making, i.e., the decision concept. There are both supportive and opposing attitudes, ensuring that their decision-making logic aligns with factual logic.

[0074] By establishing a mapping relationship between feature hierarchy and decision hierarchy concepts, the logical deduction process from original features to final decision is explained for each view. The feature hierarchy and decision hierarchy concepts respectively explain "which features each view focuses on (why a doctor would make this decision)" and "why they support or oppose collective decision (why a final collective decision is reached)." They also explain the view's "attitude" towards collective decision, achieving full backtracking of the decision trajectory and improving the model's transparency and user trust in critical tasks (such as medical diagnosis). This is a deep, closed-loop explanation from features to decision logic, going beyond a simple explanation of feature weights.

[0075] To achieve conflict-sensitive credible decision quantification, we utilize evidence-based deep learning to quantify the uncertainty and assess the reliability of individual physician decisions and final collective decisions in multi-view medical data.

[0076] Step S3: To quantify uncertainty, the classification problem is transformed into an evidence collection process. For the view... The output of follows a Dirichlet distribution. Wherein, the view For the Class of evidence The non-negative excitation output by the neural network determines the outcome, and the usual formula is:

[0077] in, These are the parameters of the Dirichlet distribution, total intensity. .

[0078] Furthermore, according to the theory of evidence, the view Individual uncertainty and its reliability The calculation formula is:

[0079] This reflects the "subjective uncertainty" arising from insufficient information in the view, among which... It represents the total number of categories. Indicates the reliability of collective decision-making. This represents the set of reliability of individual decisions.

[0080] Specifically, the concept of decision-making The data is fed into an evidence-based deep neural network to obtain the amount of evidence for each category of the doctor's decision-making process. We used the Dirichlet distribution to model individual physician opinions and quantified the uncertainty of individual decision-making based on the amount of evidence.

[0081] In more preferred examples, the features of Doctor 2 are fuzzy, and the amount of evidence output is... Lower, based on the subjective uncertainty calculated by the formula. It will significantly increase its reliability. The corresponding decrease is also expected.

[0082] Optimize consensus-building paths in complex scenarios and flexibly allocate resources based on fusion priority mechanisms. Addressing conflicting opinions arising from multi-view data input, achieve conflict-aware collective decision-making through consensus determination and dynamic priority allocation mechanisms.

[0083] Step S4: Consensus Determination and Conflict Branch Handling. Logical prediction is made based on the individual decisions and decision attitudes generated by each view: if... If all individuals make the same decision and all support it, then the output is a consistent decision, which is the final output, i.e., consensus fusion, and a highly credible collective viewpoint is obtained; otherwise, it indicates that there is a conflict among individuals, and a reliability-oriented collective decision fusion is initiated.

[0084] When conflicts arise, consensus is not forced. Instead, the conflicts are transformed into valuable assessment information. By incorporating dynamic importance specific to each case, a more suitable decision-making approach for the current situation is derived. This approach uses a subjective logic, requiring only a single forward input to capture uncertainty and directly outputting a reliable fusion result.

[0085] Specifically, if all doctors agree and are supportive, a highly reliable collective opinion can be obtained directly. However, if disagreements exist, it indicates a conflict of opinion among the doctors, requiring further reconciliation to arrive at a more reliable collective opinion.

[0086] By introducing a consensus determination phase, the decision-making process can be adaptively scheduled: when the view reaches a consensus, a high-confidence result is output, and when a conflict occurs, a reliability-oriented collective learning is initiated, thereby improving the fusion accuracy in complex conflict scenarios while ensuring decision-making efficiency.

[0087] Step S5: To obtain the dynamic priority of each view during merging. Two factors were considered to jointly determine this: (inherent) prior importance and the importance of the specific instance. .

[0088] The former does not change with instance variations and is typically related to the historical position of each view, expressed as a learnable average vector. This is represented by the initial values ​​of the decision proportions for each viewpoint obtained from the final decision scenario of a statistically normalized dataset. The latter requires that the viewpoint has a substantial influence on the judgment of a specific instance, not because of its position, but because it may provide crucial evidence. Represents a specific instance This dynamic awareness also indicates that it can be used as a measure of priority offset for instances, through the dynamic awareness layer. Get:

[0089] The superscript indicates the first. The parameters corresponding to each view. This represents a collection of all views.

[0090] Specifically, for the purpose of opinion fusion, the fusion priority of doctors' opinions is calculated. It mainly models the importance of priors and the dynamic importance of specific instances.

[0091] Obtain prior importance based on the doctor's professional title. In the preferred example, Doctor 1 is a chief physician and has a higher initial authority weight, while Doctor 6 is a newly appointed doctor and has a lower initial weight.

[0092] It considers not only the prior (static) importance based on professional titles, but also the dynamic importance of specific instances based on specific case characteristics to obtain the decision priorities of each doctor. Based on the characteristics of the patient's condition, a dynamic perception layer is used. To obtain the dynamic importance of a specific instance For example, if a patient's images show certain distinctive features, the importance of the radiologist will be increased, which is more in line with the logic of clinical consultation than optimizing the weight of a single feature.

[0093] By combining the inherent importance of views (such as doctor's title) with the dynamic changes in importance for different instance characteristics (such as the different conditions of different patients), a more flexible and reasonable acquisition of the fusion weights of each view can be achieved, which is more adaptable to more complex real-world situations and helps to focus on the opinions of a few but key doctors.

[0094] Step S6: After obtaining the two different importance levels, use a normal distribution. To model these two factors, based on evidence generation, a preferred distribution specific to instances is generated using the Dirichlet distribution:

[0095] Specifically, after obtaining the importance of two different dimensions, a doctor's opinion weight distribution for the case (dynamic priority during fusion) is generated through normal distribution and Dirichlet sampling. The set of physician decision priorities is represented as follows: By transforming the classification problem into an "evidence gathering" process, and using Dirichlet distribution modeling, the uncertainty of individuals and groups is quantified.

[0096] Use priority According to the rules, the evidence from each doctor is weighted and merged. Through this fusion method, even if a doctor with a higher professional title makes an incorrect decision, because the current uncertainty is high and the priority is dynamically lowered, the evidence will ultimately be reconciled to produce a more reliable and accurate collective decision. Specifically: Step S7: After obtaining the priority level of each view, for each view's perspective... To integrate.

[0097] In more preferred examples, viewpoints A and B are respectively , ,use Come is equivalent to The fusion process yields the fused viewpoints A and B. .

[0098] When the number of views is greater than 2, the rules for merging to obtain the collective viewpoint are as follows:

[0099]

[0100] in, This rule allows for the reconciliation of conflicting viewpoints based on priority.

[0101] Step S8: The probability density function of the Dirichlet distribution is parameterized as follows: The second-order probability of each view can be obtained, forming multiple opinions. :

[0102] in, It is a Beta function. yes One-dimensional simplex, It is a simple class assignment probability.

[0103] During the model training phase, the adjusted cross-entropy loss function is used. Ensure a high probability of correct labeling; utilize KL divergence. Compressing evidence generated by mislabeled data; utilizing binary cross-entropy Optimize the prediction accuracy of physician attitudes; introduce an annealing coefficient. This allows the system to fully explore the parameter space early in training, preventing the network output from exhibiting a flat and uniform distribution. Specifically, this includes: Step S9: Utilize the properties of the Beta and digamma functions to apply the properties of a specific view. The cross-entropy loss is adjusted, and the calculation formula is as follows:

[0104] in, These are the actual predicted labels for this view. It is a double gamma function. It is by The integral of the defined cross-entropy loss function over the simplex ensures that the correct label for each sample is higher than that for other classes. This indicates that the decision was made collectively by the doctors. This represents the set of individual decisions made by doctors.

[0105] Step S10: To ensure that fewer pieces of evidence are generated for mislabeled items, the evidence predicting mislabeled items is reduced to zero. The Kullback-Leibler (KL) divergence is used. This divergence minimizes the difference between the Dirichlet distribution and the uniform Dirichlet distribution after removing non-misleading evidence.

[0106] in, It is to remove prediction parameters Dirichlet parameters after misleading evidence from China and Africa.

[0107] Step S11: Simultaneously, guided by the decision concept, binary cross-entropy with logits loss is used to improve the accuracy of supporting attitude prediction. The calculation method is as follows:

[0108] in, This indicates whether the actual label in the current view matches the collective label.

[0109] Step S12: Obtain the loss for each view in each instance. and its collective losses :

[0110]

[0111] in, It is the annealing coefficient. This is the current training round. This is the preset number of annealing steps. To prevent insufficient parameter space exploration due to excessive focus on KL divergence in the early stages of network training, the number of steps can be gradually increased during training. The value.

[0112] By introducing an evidence-based deep learning mechanism, the system actively captures and utilizes conflicting information between multiple views, providing not only decision-making results but also quantitative reliability assessments for individual and collective decisions. This helps users judge the confidence level of decisions when there are differing opinions.

[0113] The present invention also provides a multi-view collective medical decision evaluation system based on dual-concept learning. The multi-view collective medical decision evaluation system based on dual-concept learning can be implemented by executing the process steps of the multi-view collective medical decision evaluation method based on dual-concept learning. That is, those skilled in the art can understand the multi-view collective medical decision evaluation method based on dual-concept learning as a preferred embodiment of the multi-view collective medical decision evaluation system based on dual-concept learning.

[0114] According to the present invention, a multi-view collective medical decision evaluation system based on dual-concept learning includes: Module M1 obtains decision concepts based on a two-level decision interpretation mechanism. ; Module M2 models individual decisions for each view, and incorporates decision concepts. Input the evidence-based neural network to obtain the amount of evidence. And quantify individual uncertainty; Module M3 makes consensus judgments on individual decisions of each view and outputs the collective opinion result or triggers module M4 to perform collective decision fusion; Module M4 obtains prior importance and the importance of specific instances and models them to obtain dynamic priorities. According to dynamic priority By merging the various views, a collective perspective is obtained.

[0115] In more preferred embodiments, module M1 includes: Module M1.1, for each view Using an encoder from the raw input Features are extracted and mapped to a feature concept space through an attention mechanism to obtain feature concepts. And guide the acquisition of enhanced features for each view. ; Module M1.2 will enhance features Input capture network predicts individual decisions and their supportive attitudes toward collective decision-making, thus obtaining decision concepts. .

[0116] The decision-making concept There are two attitudes: one in favor and one against.

[0117] Where V represents the total number of views.

[0118] In more preferred embodiments, in module M2, the non-negative excitation of the neural network output determines the view. For the Class of evidence :

[0119] in, This represents the parameters of the Dirichlet distribution.

[0120] view Individual uncertainty and its reliability for:

[0121]

[0122] in, Indicates the total number of categories; Indicates total intensity.

[0123] In more preferred embodiments, the consensus judgment includes: For any view, if all individual decisions are the same and all attitudes are supportive, then the consensus is merged into a collective opinion result and output. If individual decisions differ or there is opposition, the execution module M4 merges the individual decisions.

[0124] The module M4 includes: Module M4.1: Obtain prior importance based on professional title. Based on the characteristics of different instances, through a dynamic perception layer The importance of obtaining a specific instance :

[0125] The superscript indicates the first. Parameters corresponding to each view; Represents a specific instance Dynamic perception; This indicates enhanced features.

[0126] Module M4.2 adopts a normal distribution. Modeling is performed, based on evidence generation, using the Dirichlet distribution to generate instance-specific preferred distributions:

[0127] Module M4.3 merges the viewpoints of each view according to dynamic priority to obtain a collective viewpoint:

[0128]

[0129]

[0130] in, This indicates the dynamic priority of the v-th view; This represents the amount of evidence for the v-th view; This represents the viewpoint of the v-th view; V represents the total number of views.

[0131] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function as logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0132] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A multi-view collective medical decision-making evaluation method based on dual-concept learning, characterized in that, include: Step S1: Construct a two-level decision interpretation mechanism to obtain decision concepts. ; Step S2: Model the individual decisions for each view, and define the decision concepts. Input the evidence-based neural network to obtain the amount of evidence. And quantify individual uncertainty; Step S3: Make a consensus judgment on the individual decisions of each view and output a consistent collective opinion result or proceed to step S4; Step S4: Obtain prior importance and the importance of specific instances, and model to obtain dynamic priorities. According to dynamic priority Merge the various views to obtain a collective perspective and output it.

2. The multi-view collective medical decision evaluation method based on dual-concept learning according to claim 1, characterized in that, Step S1 includes: Step S1.1, for each view Using an encoder from the raw input Features are extracted and mapped to a feature concept space through an attention mechanism to obtain feature concepts. And guide the acquisition of enhanced features for each view. ; Step S1.2: Enhance the features Input capture network predicts individual decisions and their supportive attitudes toward collective decision-making, thus obtaining decision concepts. ; The decision-making concept There are two opposing attitudes: one in support and one in opposition. Where V represents the total number of views.

3. The multi-view collective medical decision evaluation method based on dual-concept learning according to claim 1, characterized in that, In step S2, the non-negative excitation determination view output by the neural network is... For the Class of evidence : in, The parameters representing the Dirichlet distribution; view Individual uncertainty and its reliability for: in, Indicates the total number of categories; Indicates total intensity.

4. The multi-view collective medical decision evaluation method based on dual-concept learning according to claim 1, characterized in that, The consensus judgment includes: For any view, if all individual decisions are the same and all attitudes are supportive, then the consensus is merged into a collective opinion result and output. If individual decisions differ or there is opposition, then step S4 is performed to merge the individual decisions; Step S4 includes: Step S4.1: Obtain prior importance based on professional title. Based on the characteristics of different instances, through a dynamic perception layer The importance of obtaining a specific instance : The superscript indicates the first. Parameters corresponding to each view; Represents a specific instance Dynamic perception; Indicates enhanced features; Step S4.2: Use a normal distribution. Modeling is performed, based on evidence generation, using the Dirichlet distribution to generate instance-specific preferred distributions: Step S4.3: Merge the viewpoints of each view according to dynamic priority to obtain a collective viewpoint: in, This indicates the dynamic priority of the v-th view; This represents the amount of evidence for the v-th view; This represents the viewpoint of the v-th view; V represents the total number of views.

5. A multi-view collective medical decision-making evaluation system based on dual-concept learning, characterized in that, include: Module M1 obtains decision concepts based on a two-level decision interpretation mechanism. ; Module M2 models individual decisions for each view, and incorporates decision concepts. Input the evidence-based neural network to obtain the amount of evidence. And quantify individual uncertainty; Module M3 makes consensus judgments on individual decisions of each view and outputs a consistent collective opinion result or triggers module M4; Module M4 obtains prior importance and the importance of specific instances and models them to obtain dynamic priorities. According to dynamic priority Merge the various views to obtain a collective perspective and output it.

6. The multi-view collective medical decision-making evaluation system based on dual-concept learning according to claim 5, characterized in that, The module M1 includes: Module M1.1, for each view Using an encoder from the raw input Features are extracted and mapped to a feature concept space through an attention mechanism to obtain feature concepts. And guide the acquisition of enhanced features for each view. ; Module M1.2 will enhance features Input capture network predicts individual decisions and their supportive attitudes toward collective decision-making, thus obtaining decision concepts. ; The decision-making concept There are two opposing attitudes: one in support and one in opposition. Where V represents the total number of views.

7. The multi-view collective medical decision-making evaluation system based on dual-concept learning according to claim 5, characterized in that, In module M2, the non-negative excitation determination view output by the neural network For the Class of evidence : in, The parameters representing the Dirichlet distribution; view Individual uncertainty and its reliability for: in, Indicates the total number of categories; Indicates total intensity.

8. The multi-view collective medical decision-making evaluation system based on dual-concept learning according to claim 5, characterized in that, The consensus judgment includes: For any view, if all individual decisions are the same and all attitudes are supportive, then the consensus is merged into a collective opinion result and output. If individual decisions differ or there is opposition, the execution module M4 merges the individual decisions; The module M4 includes: Module M4.1: Obtain prior importance based on professional title. Based on the characteristics of different instances, through a dynamic perception layer The importance of obtaining a specific instance : The superscript indicates the first. Parameters corresponding to each view; Represents a specific instance Dynamic perception; Indicates enhanced features; Module M4.2 adopts a normal distribution. Modeling is performed, based on evidence generation, using the Dirichlet distribution to generate instance-specific preferred distributions: Module M4.3 merges the viewpoints of each view according to dynamic priority to obtain a collective viewpoint: in, This indicates the dynamic priority of the v-th view; This represents the amount of evidence for the v-th view; This represents the viewpoint of the v-th view; V represents the total number of views.

9. A training method for a multi-view collective medical decision-making evaluation system based on dual-concept learning, wherein the multi-view collective medical decision-making evaluation system based on dual-concept learning as described in claims 5-8 is trained, characterized in that... include: Step SA1: Parameterize the probability density function of the Dirichlet distribution. Obtain the second-order probability of the view and form multiple opinions. : in, This represents the Beta function; express 1-dimensional simplex; This represents the class assignment probability of the nth type of simplicity in the vth view; This represents the probability of class assignment for the v-th view; The probability density function parameters represent the nth type Dirichlet distribution of the vth view; The parameters of the probability density function of the Dirichlet distribution for the v-th view are represented. Step SA2: Introduce the annealing coefficient Using the adjusted cross-entropy loss function KL divergence and binary cross-entropy Optimize the multi-view collective medical decision evaluation system based on dual-concept learning.

10. The training method for the multi-view collective medical decision-making evaluation system based on dual-concept learning according to claim 9, characterized in that, Step SA2 includes: Step SA2.1: Adjust the cross-entropy loss for a specific view: in, Indicates the true predicted label for view v; Represents the double gamma function; Indicates by The integral of the defined cross-entropy loss function over the simplex; Indicates total intensity; k represents the category ordinal number, and there are K categories in total; Step SA2.2: Reduce the evidence for predicting mislabeled items to 0 using KL divergence: in, Indicates the removal of prediction parameters Dirichlet parameters after misleading evidence from China and Africa; Introduce binary cross-entropy with logits loss: in, A binary value indicating whether the actual label and the collective label in the current view are consistent; Step SA2.3: Obtain the loss. and its collective losses : in, Indicates the annealing coefficient; Indicates the current training round; This indicates the preset number of annealing steps.

Images