Medical report quality control method and system based on knowledge anchoring and confidence decision

Abstract

This invention discloses a medical report quality control method and system based on knowledge anchoring and confidence-based decision-making, belonging to the field of medical data processing. The method includes: parsing medical image reports to generate metadata vectors; dynamically activating matching quality control item processors based on the metadata; bidirectionally anchoring the report text based on a three-layer structured medical knowledge base; inputting information into a multi-task medical model trained in two stages, outputting preliminary judgment results and confidence scores through thought chain reasoning; determining whether the confidence score is below a threshold, and if so, triggering a multi-perspective adversarial dispute resolution mechanism, generating a final judgment through independent reasoning using three strategies and majority voting; and aggregating all judgment results to generate a structured quality control report. This invention suppresses model illusions through knowledge anchoring and improves the accuracy of marginal sample judgment through confidence-driven multi-stage review, achieving high-precision, interpretable, and evolvable intelligent quality control of medical reports.

Description

Technical Field

[0001] This invention relates to the fields of medical information processing and artificial intelligence technology, and in particular to a medical report quality control method and system based on knowledge anchoring and confidence decision-making. Background Technology

[0002] Medical imaging diagnostic reports serve as a crucial basis for clinical decision-making and prognostic assessment; their standardization and accuracy are directly related to medical quality and patient safety.

[0003] Current medical report quality control methods have significant limitations. Manual quality control is inefficient, lacks coverage, and has difficulty in standardizing evaluation criteria. Rule-based quality control methods have poor semantic understanding and generalization capabilities, making it difficult to handle synonymous expressions and implicit logic in medical texts. Existing deep learning quality control models are mostly coarse-grained judgments and lack interpretable reasoning, making them prone to high-confidence misjudgments in complex and marginal cases. They also suffer from fragmented architecture, high maintenance costs, and an inability to adaptively control quality based on examination sites and modalities. Consequently, they cannot simultaneously achieve high-precision semantic error recognition, interpretable decision output, and flexible adaptation to multiple scenarios, failing to meet the actual clinical needs for intelligent and highly reliable report quality control. Summary of the Invention

[0004] To address the technical problems of low coverage, poor semantic understanding, and lack of interpretability in medical report quality control, this invention provides a medical report quality control method and system based on knowledge anchoring and confidence decision-making.

[0005] On the one hand, it provides a medical report quality control method based on knowledge anchoring and confidence-based decision-making, including: Obtain the original medical image diagnostic report text, parse the diagnostic report text, extract metadata fields, and generate a structured metadata vector; Based on the metadata vector, dynamically activate one or more quality control item processors that match the report from a predefined quality control processor registry; Based on a pre-built structured medical quality control knowledge base, the report text is bidirectionally anchored, and the anchored semantic information is output. The metadata vector, the activated matching quality control item processor, and the anchored semantic information are input into a multi-task medical big model trained in two stages. The model generates preliminary judgment results and their corresponding confidence scores based on the thinking chain reasoning strategy. If the confidence score is lower than a preset threshold, an adversarial dispute resolution mechanism is triggered, and a final judgment result is generated through multi-perspective independent reasoning and majority ruling; otherwise, the preliminary judgment result is directly adopted as the final judgment result. Aggregate the final judgment results of all activated and matched quality control item processors, generate a structured quality control report, and output it.

[0006] According to the medical report quality control method based on knowledge anchoring and confidence decision-making provided by the present invention, the metadata vector is:

[0007] in, For report ID; To examine what was seen; For the inspection conclusion; For the area to be inspected; To check the modality; The patient's gender.

[0008] According to the medical report quality control method based on knowledge anchoring and confidence decision provided by the present invention, each processor in the quality control processor registration library corresponds to a type of quality control rule, and all processors follow a unified input / output interface standard; the dynamic activation matching adopts an adaptive task routing mechanism.

[0009] According to the medical report quality control method based on knowledge anchoring and confidence decision provided by the present invention, the structured medical quality control knowledge base adopts a three-layer architecture design, including an entity layer, a semantic relationship layer, and a rule layer.

[0010] According to the medical report quality control method based on knowledge anchoring and confidence decision provided by the present invention, the bidirectional anchoring includes positive anchoring and negative filtering; wherein, positive anchoring converts the description in the report text into a high-dimensional embedding vector and performs semantic similarity calculation with the standard entity vector of the knowledge base entity layer; negative filtering systematically excludes interfering descriptions based on the rule layer template system.

[0011] According to the medical report quality control method based on knowledge anchoring and confidence decision-making provided by the present invention, the two-stage training of the multi-task medical big data model includes: In the first stage, adaptive pre-training in the medical field was conducted, using clinical desensitized medical image reports for incremental training to optimize the model's adaptation to the semantic distribution of radiology. The loss function was modeled using causal language. In the second stage, multi-task instruction fine-tuning and task-aware parameter adaptation are carried out by introducing a low-rank matrix adaptation structure to configure independent adaptation parameter groups for different quality control tasks.

[0012] According to the medical report quality control method based on knowledge anchoring and confidence decision provided by the present invention, dynamic loss weighting and gradient conflict suppression are introduced in the second stage of training.

[0013] According to the medical report quality control method based on knowledge anchoring and confidence decision provided by the present invention, the multi-task medical big model further introduces a confidence regularization strategy to generate boundary samples and adds a confidence regularization term on the basis of standard cross-entropy loss.

[0014] According to the medical report quality control method based on knowledge anchoring and confidence decision-making provided by the present invention, the adversarial dispute resolution mechanism initiates three independent inferences in parallel in an isolated inference environment, each employing one of the following three different strategies: Strategy A: Change the wording of the prompts; Strategy B: Increase the decoding temperature; Strategy C: Switch the evaluation perspective. By integrating the results of the three rulings through majority voting logic, if the final ruling is inconsistent with the preliminary ruling, a dispute resolution record is generated.

[0015] On the other hand, it provides a medical report quality control system based on knowledge anchoring and confidence-based decision-making, including: The report parsing and metadata extraction module obtains the original medical image diagnosis report text, parses the diagnosis report text, extracts metadata fields, and generates a structured metadata vector. The quality control item scheduling and registration module dynamically activates one or more quality control item processors that match the report from the predefined quality control processor registration library based on the metadata vector. The structured medical quality control knowledge base module, based on a pre-built structured medical quality control knowledge base, performs bidirectional anchoring on the report text and outputs the anchored semantic information; The reasoning and adversarial error correction module inputs the metadata vector, the activated matching quality control item processor, and the anchored semantic information into a multi-task medical big data model trained in two stages. The model generates a preliminary judgment result and its corresponding confidence score based on a thought chain reasoning strategy. It then determines whether the confidence score is lower than a preset threshold. If so, an adversarial dispute resolution mechanism is triggered, and a final judgment result is generated through multi-perspective independent reasoning and majority decision-making. If not, the preliminary judgment result is directly adopted as the final judgment result. The structured quality control report generation module aggregates the final judgment results of all activated and matched quality control item processors, generates and outputs a structured quality control report containing a metadata layer, a scoring layer, an evidence layer, and a recommendation layer.

[0016] The above technical solution has the following advantages or beneficial effects: This invention achieves a deep integration of high-precision semantic understanding and clinical interpretability in the field of medical report quality control by constructing a full-link technical architecture of "structured knowledge anchoring - multi-task large-scale model reasoning - confidence-driven verification - interpretable output". The solution uses a structured medical knowledge base as factual constraints and suppresses semantic illusions in large models through a two-way anchoring mechanism; it uses a two-stage trained multi-task large-scale model as a unified reasoning kernel, replacing traditional fragmented model deployment, and achieving soft isolation and collaborative optimization of multiple quality control tasks based on a shared semantic space; it innovatively introduces a multi-perspective adversarial dispute resolution mechanism using confidence scores as decision-making boundaries, enabling the system to self-correct marginal samples; and finally, it generates a structured quality control report containing a complete chain of evidence based on thought chain reasoning. Through the organic synergy of the above technical features, a high-precision, interpretable, and easily scalable intelligent quality control closed loop for medical reports is formed, effectively solving the technical problems of insufficient semantic understanding, high misjudgment rate of marginal samples, lack of interpretability, and high maintenance costs in existing technologies. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] Figure 1 This is the report quality control flowchart for Example 1; Figure 2 This is a structural diagram of the report quality control system in Example 2. Detailed Implementation

[0019] Example 1 like Figure 1 As shown, this embodiment provides a medical report quality control method based on knowledge anchoring and confidence decision-making. The method specifically includes: S1. Obtain the original medical image diagnostic report text, parse the diagnostic report text, extract metadata fields, and generate a structured metadata vector.

[0020] The parsing of the diagnostic report text includes: using the report's layout features and chapter title patterns to segment the original report text into different semantic regions, and further extracting metadata fields.

[0021] The metadata vector is:

[0022] in, For report ID; To examine what was seen; For the inspection conclusion; For the area to be inspected; To check the modality; The patient's gender.

[0023] For example, id: COR00001; finding: clear lung fields, normal lung markings, no solid lesions; conclusion: no abnormalities seen in both lungs; bodypart: chest CT scan; modality: CT; gender: male.

[0024] S2. Based on the metadata vector, dynamically activate one or more quality control item processors that match the report from the predefined quality control processor registry.

[0025] Specifically, the quality control processor registration library is based on a unified quality control item interface standard predefined by the system. Each processor corresponds to a type of quality control rule, and all processors follow a unified input / output interface standard. The quality control logic is completely decoupled from the main process. The quality control item processor covers ten major categories of quality control scenarios: completeness of conclusion basis, omission of inspection items, conflict between the location information of observation and conclusion, gender organ conflict, completeness of observed positive features, typos, completeness of mammography BI-RADS classification, completeness of prostate PI-RADS classification, unit description errors, and language logic and description standardization. Dynamic activation matching employs an adaptive task routing mechanism, and the set of activated processors is formally represented as follows:

[0026] in, For metadata vectors, For the i-th processor, match() is the matching function used to determine the processor. Does it satisfy the activation condition corresponding to the metadata vector m?

[0027] This mechanism allows new quality control items to take effect simply by implementing the standard interface and placing them in the specified directory, without requiring modifications to the main program framework. This achieves a high degree of modularity and "hot-swappable" deployment of quality control items.

[0028] This invention adopts an architecture of "routing allocation + pluggable quality control items" to decouple the quality control logic from the model base. This design allows medical institutions to flexibly configure personalized quality control rules for different quality control needs and different examination modalities (CT, MRI, DR, etc.), and supports the "hot-swappable" deployment of new rules, with excellent engineering versatility and low-cost iteration advantages.

[0029] S3. Based on a pre-built structured medical quality control knowledge base, the report text is bidirectionally anchored, and the anchored semantic information is output.

[0030] Specifically, the structured medical quality control knowledge base adopts a three-layer architecture design, including an entity layer, a semantic relation layer, and a rule layer, which provides verified factual background for subsequent reasoning and constitutes a domain-specific retrieval enhancement generation front link. Among them, the entity layer stores anatomical locations, disease names, imaging features, grading terms, and their synonym sets; the semantic relation layer defines semantic associations between entities such as "belongs to", "manifests as", and "excludes from"; and the rule layer stores judgment templates based on entities and logical relationships.

[0031] Based on a pre-built structured medical quality control knowledge base, report texts are bidirectionally anchored, including positive anchoring and negative filtering.

[0032] In this process, positive anchoring converts the descriptions in the report text into high-dimensional embedding vectors, which are then used to calculate semantic similarity with the standard entity vectors in the knowledge base entity layer. The semantic similarity calculation employs cosine similarity.

[0033] in, For embedding vectors, It is a text vector; For knowledge base entity vectors. When At that time, positive entity anchoring is completed, which accurately maps non-standard descriptions to standard entities and associates the relevant differential diagnosis and the features that must be described with the entity, storing them as external auxiliary clues in the context of the prompt words.

[0034] Negative filtering, based on rule-layer templates, systematically identifies and eliminates interfering descriptions to reduce over-quality control of normal physiological states, post-treatment states, and non-pathological changes. This process employs a Boolean logic decision mechanism: when a text fragment meets preset filtering rules, it is marked as a filter item and removed from subsequent quality control processes. For example, if a text fragment contains negative normal descriptions such as "no obvious abnormalities observed" or "no clear abnormal signals observed," it is marked as a filter item and excluded; similarly, if a text fragment contains descriptions of post-intervention states such as "post-operative changes" or "post-treatment changes," it is marked as a filter item and excluded.

[0035] For example, when a lung CT report is input, the system first searches the medical knowledge base. If the report mentions "cavitation in the upper lobe of the left lung," the knowledge base automatically associates common differential diagnoses of "cavitation" (tuberculosis, lung cancer, etc.) and essential characteristics (wall thickness, inner wall smoothness). This retrieved structured information is stored as "external auxiliary clues" in the prompt word container, serving as the anchored semantic information output. Through this two-way anchoring mechanism, the problem of misjudging normal reports due to semantic coupling is fundamentally solved, providing accurate factual constraints for subsequent large-scale model inference.

[0036] Aiming at the defect that general large models are prone to generate misleading outputs (hallucinations) when processing professional medical texts, the present invention constructs a retrieval-augmented generation (RAG) pre-check link. By performing bidirectional semantic anchoring on the report content through a structured medical knowledge base, it provides mandatory factual constraints for model reasoning and improves the rigor of the system in complex clinical scenarios.

[0037] S4. Input the metadata vector, activate the matching quality control item processor, and the anchored semantic information into a multi-task medical large model trained in two stages. The model generates a preliminary judgment result and its corresponding confidence score based on the chain of thought reasoning strategy.

[0038] Specifically, the multi-task medical large model is pre-constructed through the following two-stage training strategy: The first stage is medical domain adaptive pre-training, which uses clinically desensitized medical image reports for incremental training to optimize the model's adaptation to the semantic distribution of the radiology department. The loss function uses causal language modeling (CLM):

[0039] Among them, represents the causal language modeling loss; T represents the total length of the input sequence; represents the t-th token; x<t represents the context sequence before the t-th token; represents the model parameters; represents the conditional probability that the model generates the current token under given conditions. This stage aims to bridge the semantic gap between general domain corpora and radiology vertical domain corpora, thereby improving the model's understanding ability of professional terms, report expression methods, and domain knowledge.

[0040] The second stage is multi-task instruction fine-tuning and task-aware parameter adaptation. To achieve unified modeling of multiple quality control tasks, the present invention introduces a low-rank matrix adaptation (LoRA) structure. For the i-th quality control task, low-rank matrix adaptation is introduced:

[0041] Among them, represents the adapted weight matrix corresponding to the i-th quality control task; is the original weight matrix; and respectively represent the low-rank adaptation matrices corresponding to the i-th task. Usually, if , then , , among which, is the low-rank dimension. <By introducing low-rank incremental parameters, lightweight adaptation to different quality control tasks can be achieved while maintaining the core weights frozen. In this way, different quality control tasks are configured with independent adaptation parameter sets, thereby achieving soft parameter separation on the basis of a shared semantic space. This reduces the negative transfer effect in multi-task learning and improves the model's ability to specifically model various quality control sub-tasks.

[0043] The training organization adopts a multi-task alternating sampling strategy, with each training batch containing only single-task samples and different tasks rotating between batches.

[0044] Suppose there are N quality control tasks in total, then the loss of a single quality control task is:

[0045] in, Let represent the expected loss for the i-th quality control task; This represents the data distribution or dataset corresponding to the i-th quality control task; Indicates input sample and its annotation Dataset from task i ; Indicates the input sample; This represents the true label corresponding to the input sample; This represents the output result of the model on the i-th quality control task; This represents the single-sample loss function, used to measure the difference between the model's predicted results and the true labels; This represents the expectation of the sample distribution for task i.

[0046] The total loss is:

[0047] Where N represents the total number of quality control tasks. Task weights. To ensure stable convergence of the multi-task model in the high-dimensional parameter space, dynamic loss weighting and gradient conflict suppression are introduced. Task weights Dynamically adjust using gradient norm normalization:

[0048] in This is a very small constant to prevent division by zero; For gradient operators; This represents the specific model parameters corresponding to the i quality control tasks. When the gradient directions of two tasks conflict (inner product < 0), the parameters for each task are... The gradient is corrected by orthogonal projection:

[0049] in For the task The gradient in the shared parameter space, For the task The gradient in the shared parameter space, It is its L2 norm squared. This mechanism effectively reduces the risk of negative migration and ensures stable convergence across multiple tasks.

[0050] The multi-task medical model, after completing the two-stage process, further incorporates a confidence regularization strategy for optimization. This confidence regularization strategy generates approximately 8,000 boundary cases (synonyms, minor word order inversions, minor typos with semantic preservation, ambiguous "possible" expressions, etc.) based on rules and artificial perturbations. A confidence regularization term is added to the standard cross-entropy loss.

[0051] in, This represents the total loss after adding confidence regularization; This represents the standard cross-entropy loss, used to measure the difference between the model's predictions and the true labels; This represents the confidence regularization coefficient, used to control the strength of the influence of the regularization term on the total loss, and its value ranges from 0.08 to 0.15. This represents the predicted probability of the model's output for the i-th category; The information entropy related term represents the probability distribution predicted by the model and is used to constrain the output confidence distribution of the model on boundary samples.

[0052] The confidence regularization strategy is integrated into the second-stage fine-tuning process. A regularization term is applied simultaneously when calculating the multi-task loss. The probability distribution of the output of such fuzzy samples located in the cognitive fuzzy interval is applied, enabling the model to learn to identify its own capability boundaries and learn confidence calibration while optimizing task accuracy.

[0053] After completing the above pre-construction, in the quality control inference stage, the metadata vector, the activated matching quality control item processor, and the anchored semantic information are combined into prompt words and input into the model. The model follows a thought chain inference strategy, requiring the model's quality control to have complete intermediate inference steps. The inference process is broken down into an ordered sequence containing fields such as the original input text, step number, inference content, quality control summary, and confidence score, and based on this, a preliminary judgment result and its confidence score are output. The confidence score is calculated as follows:

[0054] Where z is the model output logits, and T is the temperature coefficient. For initial quality control without dispute, the output is in structured form.

[0055] This invention abandons the traditional fragmented deployment architecture of "one model for each quality control item" and adopts a technical approach of learning a single large model for all tasks. By injecting learning signals for specific tasks into a base model with shared weights, it achieves unified modeling of multiple quality control tasks.

[0056] This invention effectively solves the "negative transfer" problem in the joint training of a single model by introducing task-aware low-rank adaptation (LoRA) and orthogonal gradient projection (PCGrad) techniques. This mechanism ensures the coordination and consistency of gradients between different tasks, enabling the system to achieve a depth of medical semantic understanding and task focus that surpasses general-purpose large models while maintaining extremely low computational overhead.

[0057] Furthermore, unlike the "black box" output of traditional AI quality control, this invention constructs a structured reasoning engine based on the Chain of Reasoning (CoT) technology. Each quality control judgment is accompanied by a complete evidence attribution chain (original text citation - logical reasoning - standard recommendations), realizing the traceability and auditability of the quality control process. This not only reduces the cognitive load of manual review but also provides physicians with highly valuable modification suggestions, truly realizing the use of AI to assist in the continuous improvement of medical quality.

[0058] S5. Determine whether the confidence score is lower than a preset threshold. If so, trigger the adversarial dispute resolution mechanism and generate a final judgment result through multi-perspective independent reasoning and majority ruling. If not, directly adopt the preliminary judgment result as the final judgment result.

[0059] Specifically, preset reliability thresholds (Preferred in this embodiment) For each quality control item output in step S4, the preliminary judgment result and confidence score are... ,like If it is determined to be of high confidence, the preliminary judgment result is directly adopted as the final judgment result without further verification; if If the model is found to have entered a cognitive ambiguity zone, a multi-perspective adversarial dispute resolution mechanism will be automatically triggered.

[0060] The adversarial dispute resolution mechanism initiates three independent secondary inferences in parallel within an isolated reasoning environment. These three inferences do not interfere with each other and employ the following three different strategies to increase the diversity and robustness of the judgments: Strategy A: Change the wording of the prompts, replacing the original positive instruction prompts with rhetorical questions to guide the model to examine the problem from different perspectives; Strategy B increases the decoding temperature by adjusting the temperature parameter T to a higher value, thereby increasing the randomness and exploratory nature of the model output. Strategy C involves switching the evaluation perspective, changing the original "deduction-oriented" judgment logic to "evidence-finding-oriented," which requires the model to prioritize finding evidence supporting the report's correctness rather than errors.

[0061] Each of the three inference steps outputs a judgment result for the same quality control item. , , (Values ​​can be "deduct points" or "no points deducted"). The results of the three judgments are combined using majority rule logic, and the majority result is taken as the final judgment, expressed as:

[0062] Here, argmax represents selecting the decision item with the highest cumulative score. If the final decision differs from the initial decision, a dispute resolution record is generated, documenting the dispute process and the strategy combination used. Furthermore, this sample can be labeled as a "false positive difficult example" and fed back into the training set. Specifically, samples confirmed as false positives after dispute resolution are labeled as difficult examples, and their corresponding original inputs, decision labels, dispute records, and strategy information are structured and stored. During subsequent model training, these samples are proportionally fed back into the training set as high-value difficult examples to participate in incremental fine-tuning and confidence optimization, continuously improving the model's ability to discriminate complex boundary samples and reducing the false positive rate.

[0063] The model outputs a binary label of "deduct points / no points" and its confidence score for each judgment result of each quality control item. This confidence-driven decision-making process achieves an organic balance between rapid approval of high-confidence results and in-depth review of low-confidence results. While ensuring processing efficiency, it significantly improves the accuracy of judging marginal samples and eliminates the random errors of single inference to the greatest extent.

[0064] This invention addresses the core shortcomings of large language models, such as overconfidence and high false positive rates on marginal samples, by introducing an adversarial secondary quality control and dispute resolution mechanism under cognitive boundary perception, enabling the system to have self-correction capabilities.

[0065] Through an adversarial resolution mechanism, this invention can perceive its own "cognitive ambiguity" in judging specific reports in real time. For marginal samples with low confidence, it automatically triggers multi-perspective instruction game and decoding perturbation review. This decision-making process of "fast passage with high confidence and deep adversarial process with low confidence" simulates the reasoning process of senior physician consultation, greatly eliminating the randomness error of single reasoning and ensuring the reliability of quality control results.

[0066] S6. Aggregate the final judgment results of all activated and matched quality control item processors, generate and output a structured quality control report containing metadata layer, scoring layer, evidence layer and suggestion layer.

[0067] Specifically, the final judgment results of all activated and matched quality control item processors are aggregated, including the sub-scores, judgment labels, confidence scores, complete inference chains, and standard correction suggestions for each quality control item. The heterogeneous results output by each processor are then formatted into a standard format. The final generated quality control report contains the following levels: The metadata layer contains patient information and examination information; The scoring layer includes sub-item scores and total scores for ten major quality control items; The evidence layer includes the original text location, model inference chain, and confidence score for each deduction item; The suggestion layer contains standard correction suggestions for erroneous content (which can be invoked from the knowledge base rule layer or generated by the model).

[0068] This structured report is persistently stored in JSON format and outputs a summary of scores in CSV format, facilitating integration with higher-level quality control platforms for big data statistical analysis. The generated report ensures that each deduction decision is accompanied by a complete chain of evidence, meeting the rigid requirements of interpretability and audit trail in the medical field. It also provides a structured data foundation for subsequent data analysis and quality improvement, fundamentally changing the shortcomings of existing technologies that only output abstract scores, and providing a logically consistent chain of evidence at the output end.

[0069] Example 2 like Figure 2 As shown, this embodiment provides a medical report quality control system based on knowledge anchoring and confidence-based decision-making; The system mainly consists of a report parsing and metadata extraction module, a quality control item scheduling and registration module, a structured medical quality control knowledge base module, a reasoning and judgment module, and a structured quality control report generation module.

[0070] The report parsing and metadata extraction module obtains the original medical image diagnostic report text, parses the diagnostic report text, extracts metadata fields, and generates a structured metadata vector.

[0071] The quality control item scheduling and registration module dynamically activates one or more quality control item processors that match the report from the predefined quality control processor registration library based on the metadata vector.

[0072] As the system's scheduling hub, a unified interface standard for quality control items is predefined, encapsulating each type of quality control rule into an independent processor. Based on the metadata output by the report parsing module, this module dynamically filters and activates a set of quality control item processors matching the report from the quality control item registry, enabling differentiated loading and hot-swappable expansion of quality control tasks.

[0073] The Structured Medical Quality Control Knowledge Base module, based on a pre-built structured medical quality control knowledge base, performs bidirectional anchoring on report texts and outputs anchored semantic information.

[0074] This module adopts a three-layer architecture, including an entity layer that stores standard terms such as anatomical locations, disease names, and imaging features, as well as their synonyms; a relation layer that defines semantic relationships between entities; and a rule layer that stores judgment templates. The module incorporates semantic anchoring and filtering functions. On the one hand, positive anchoring accurately maps abnormal descriptions in reports to standard entities; on the other hand, negative filtering systematically excludes interfering descriptions such as "no abnormalities found," providing high-quality input for subsequent reasoning.

[0075] The reasoning and adversarial error correction module inputs the metadata vector, the activated matching quality control item processor, and the anchored semantic information into a multi-task medical big data model trained in two stages. The model generates a preliminary judgment result and its corresponding confidence score based on the thought chain reasoning strategy. It then determines whether the confidence score is lower than a preset threshold. If so, it triggers an adversarial dispute resolution mechanism to generate a final judgment result through multi-perspective independent reasoning and majority decision-making. If not, it directly adopts the preliminary judgment result as the final judgment result.

[0076] The structured quality control report generation module aggregates the final judgment results of all activated and matched quality control item processors, generates and outputs a structured quality control report containing a metadata layer, a scoring layer, an evidence layer, and a recommendation layer.

[0077] This module, located at the system output, aggregates the final judgment results of all quality control items, generating a fully structured report containing metadata, scoring, evidence, and recommendation layers. The report is persistently stored in JSON format and outputs a summary of scores in CSV format for easy subsequent statistical analysis.

[0078] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0079] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

Claims

1. A medical report quality control method based on knowledge anchoring and confidence-based decision-making, characterized in that, include: Obtain the original medical image diagnostic report text, parse the diagnostic report text, extract metadata fields, and generate a structured metadata vector; Based on the metadata vector, dynamically activate one or more quality control item processors that match the report from a predefined quality control processor registry; Based on a pre-built structured medical quality control knowledge base, the report text is bidirectionally anchored, and the anchored semantic information is output. The metadata vector, the activated matching quality control item processor, and the anchored semantic information are input into a multi-task medical big model trained in two stages. The model generates preliminary judgment results and their corresponding confidence scores based on the thinking chain reasoning strategy. If the confidence score is lower than a preset threshold, an adversarial dispute resolution mechanism is triggered, and a final judgment result is generated through multi-perspective independent reasoning and majority ruling; otherwise, the preliminary judgment result is directly adopted as the final judgment result. Aggregate the final judgment results of all activated and matched quality control item processors, generate a structured quality control report, and output it.

2. The medical report quality control method based on knowledge anchoring and confidence decision-making according to claim 1, characterized in that, The metadata vector is: in, For report ID; To examine what was seen; For the inspection conclusion; For the area to be inspected; To check the modality; The patient's gender.

3. The medical report quality control method based on knowledge anchoring and confidence decision-making according to claim 1, characterized in that, Each processor in the quality control processor registration library corresponds to a type of quality control rule, and all processors follow a unified input / output interface standard; the dynamic activation matching adopts an adaptive task routing mechanism.

4. The medical report quality control method based on knowledge anchoring and confidence decision-making according to claim 1, characterized in that, The structured medical quality control knowledge base adopts a three-layer architecture design, including an entity layer, a semantic relationship layer, and a rule layer.

5. The medical report quality control method based on knowledge anchoring and confidence decision-making according to claim 1, characterized in that, The bidirectional anchoring includes positive anchoring and negative filtering; wherein, positive anchoring converts the description in the report text into a high-dimensional embedding vector and performs semantic similarity calculation with the standard entity vector of the knowledge base entity layer; negative filtering systematically excludes interfering descriptions based on the rule layer template system.

6. The medical report quality control method based on knowledge anchoring and confidence decision-making according to claim 1, characterized in that, The two-stage training of the multi-task medical model includes: In the first stage, adaptive pre-training in the medical field was conducted, using clinical desensitized medical image reports for incremental training to optimize the model's adaptation to the semantic distribution of radiology. The loss function was modeled using causal language. In the second stage, multi-task instruction fine-tuning and task-aware parameter adaptation are carried out by introducing a low-rank matrix adaptation structure to configure independent adaptation parameter groups for different quality control tasks.

7. The medical report quality control method based on knowledge anchoring and confidence decision-making according to claim 6, characterized in that, The second phase of training introduces dynamic loss weighting and gradient conflict suppression.

8. The medical report quality control method based on knowledge anchoring and confidence decision-making according to claim 1, characterized in that, The multi-task medical big model further introduces a confidence regularization strategy to generate boundary samples, adding a confidence regularization term on the basis of standard cross-entropy loss.

9. The medical report quality control method based on knowledge anchoring and confidence decision-making according to claim 1, characterized in that, The adversarial dispute resolution mechanism initiates three independent inferences in parallel within an isolated inference environment, employing the following three different strategies: Strategy A: Change the wording of the prompts; Strategy B: Increase the decoding temperature; Strategy C: Switch the evaluation perspective. By integrating the results of the three rulings through majority voting logic, if the final ruling is inconsistent with the preliminary ruling, a dispute resolution record is generated.

10. A medical report quality control system based on knowledge anchoring and confidence-based decision-making, characterized in that, include: The report parsing and metadata extraction module obtains the original medical image diagnosis report text, parses the diagnosis report text, extracts metadata fields, and generates a structured metadata vector. The quality control item scheduling and registration module dynamically activates one or more quality control item processors that match the report from the predefined quality control processor registration library based on the metadata vector. The structured medical quality control knowledge base module, based on a pre-built structured medical quality control knowledge base, performs bidirectional anchoring on the report text and outputs the anchored semantic information; The reasoning and adversarial error correction module inputs the metadata vector, the activated matching quality control item processor, and the anchored semantic information into a multi-task medical big data model trained in two stages. The model generates a preliminary judgment result and its corresponding confidence score based on a thought chain reasoning strategy. It then determines whether the confidence score is lower than a preset threshold. If so, an adversarial dispute resolution mechanism is triggered, and a final judgment result is generated through multi-perspective independent reasoning and majority decision-making. If not, the preliminary judgment result is directly adopted as the final judgment result. The structured quality control report generation module aggregates the final judgment results of all activated and matched quality control item processors, generates and outputs a structured quality control report containing a metadata layer, a scoring layer, an evidence layer, and a recommendation layer.

Images