A Tumor Patient Test Report Interpretation System and Method

Abstract

The present invention discloses a tumor patient test report interpretation system and method, which relates to the field of medical informatization. The system includes: an information verification module for verifying and classifying the identities of patients and doctors; a test report parsing module for structurally parsing the test report through natural language processing and computer vision technologies, and verifying the qualifications of the testing unit based on a fuzzy matching algorithm; a data storage module for integrating knowledge base data related to tumor diagnosis and treatment; and an index analysis module for identifying tumor types, analyzing abnormal index items, and performing triage matching through a cascaded model of support vector machines, isolation forests, and decision trees. According to the technical solution of the present application, automatic interpretation of tumor patient test reports and matching with professional doctors can be achieved, which has high application value.

Description

Technical Field

[0001] The present invention relates to the field of medical informatization, and particularly to a system and method for interpreting test reports of tumor patients. Background Art

[0002] In the field of tumor diagnosis and treatment, the accurate interpretation of test reports and the efficient doctor-patient matching are key links to improve the quality of clinical decision-making. Traditional test report analysis mainly relies on manual interpretation, which has problems such as low efficiency, strong subjectivity, and high misdiagnosis rate. In recent years, artificial intelligence technology has been gradually applied to medical data processing. Some studies have tried to achieve automatic analysis of test report indicators through a single algorithm, but its generalization ability is limited and it is difficult to cope with the complex multi-index correlation and dynamic change characteristics of tumor patients. In addition, existing methods mostly focus on data analysis itself and lack deep integration with clinical resources, so they cannot form a closed-loop service.

[0003] Traditional methods usually process text and image data independently and do not fully explore the semantic relationship between the two. For example, the numerical indicators and imaging features in the test report are not analyzed collaboratively, resulting in one-sided interpretation results; systems based on black-box models lack transparent display of decision-making logic and are difficult to meet the requirements of clinicians for the credibility of results; at the same time, static models are difficult to adapt to the data distribution differences of different hospitals and need to be retrained frequently. Existing matching methods mostly rely on keyword search or simple rules, without considering the dynamic update of doctor professional labels, the needs of multi-disciplinary collaboration, and the personalized disease characteristics of patients, so the matching efficiency and accuracy are low. Therefore, based on the above problems, the present invention proposes a system and method for interpreting test reports of tumor patients. Summary of the Invention

[0004] Technical Objectives

[0005] In order to solve the above problems, the objective of the present invention is to provide a system and method for interpreting test reports of tumor patients, aiming to solve the problems of low efficiency of manual interpretation of test reports of tumor patients, high misdiagnosis rate, and inaccurate medical resource matching. By constructing a system that integrates multi-modal data and collaborates with intelligent algorithms, automatic and accurate analysis of test indicators, multi-dimensional disease assessment, and dynamic matching of interdisciplinary doctors are realized, so as to improve the accuracy, efficiency of diagnosis and treatment decision-making, and utilization rate of medical resources, while strengthening data privacy protection and ethical compliance.

[0006] Technical Solutions

[0007] To achieve the above object, the present invention provides a tumor patient test report interpretation system and method. The system and method combine natural language processing and computer vision technologies to realize the joint analysis and structured extraction of text indicators and medical images, and use support vector machines, isolation forests, and decision tree algorithms to complete tumor type classification, anomaly detection, and triage matching respectively. In addition, transfer learning and dynamic feature selection mechanisms are introduced to enhance the generalization of the model. Furthermore, homomorphic encryption is used to protect patient privacy, and knowledge graphs and incremental learning are combined to dynamically update doctor labels, realizing multi-disciplinary precise recommendation based on the complexity of the condition and professional capabilities. The system integrates functions such as generating adversarial network amplified data, automated ethical review, and visual anomaly explanation to ensure system compliance and clinical interpretability.

[0008] In the first aspect, the present invention provides a tumor patient test report interpretation system, including:

[0009] An information verification module for authenticating and classifying the identities of patients and doctors;

[0010] A test report parsing module for structurally parsing the test report through natural language processing and computer vision technologies, and verifying the qualifications of the testing unit based on a fuzzy matching algorithm;

[0011] A data storage module for integrating knowledge base data related to tumor diagnosis and treatment, including tumor markers, treatment plans, and clinical guidelines;

[0012] An index analysis module for identifying tumor types, analyzing abnormal index items, and performing triage matching through a cascaded model of support vector machines, isolation forests, and decision trees;

[0013] A feature fusion module for realizing the temporal fusion of cross-modal features based on dynamic weight allocation and long short-term memory networks;

[0014] A weight optimization module for quantifying the cross-institutional data distribution difference through KL divergence and realizing the dynamic aggregation of the federated model based on meta-learning.

[0015] Further, the specific process of the test report parsing module verifying the qualifications of the testing unit includes:

[0016] (1) Establish a standardized name database of CNAS-certified laboratories, including full names, abbreviations, and historical name change records;

[0017] (2) Use OCR technology to extract the name of the testing institution from the test report image, and calculate its similarity with the database through the Levenshtein distance algorithm;

[0018] (3) When the similarity ≥ 85%, it is automatically marked as a certified institution; otherwise, an artificial review process is triggered, and the review results are recorded to optimize the database.

[0019] (4) Generate a risk prompt for the testing institutions that fail to pass the certification and restrict the subsequent parsing authority of their test reports.

[0020] Furthermore, the test report parsing module corrects the inclination and blurriness of the test report image by preprocessing the image, extracts the key index data in the image using OCR technology, verifies the integrity of the data through a rule engine, and conducts a comparative analysis with the standard thresholds in the medical knowledge base.

[0021] Even further, the test report parsing module is based on a text localization sub-unit of a convolutional neural network, which is used to extract the text area from the test form image, identify the test item names and values in the extracted area using a CRNN model, and store the recognition results in a standardized format according to a predetermined format.

[0022] Even further, the predetermined format includes: test item - value - unit - reference range.

[0023] Furthermore, the data storage module stores the tumor diagnosis and treatment data in a structured form of entity - relationship - attribute, and automatically corrects the outdated rules in the knowledge base by using web crawler technology to capture the latest clinical research literature in real time.

[0024] Furthermore, the index analysis module realizes the classification of tumor types based on kernel function optimization, calculates the anomaly score according to the path length, thereby dynamically adjusting the threshold to adapt to the anomaly detection requirements of different tumor types, and constructs a triage rule tree based on the principle of maximizing information gain to support multi-label classification to match interdisciplinary doctor teams.

[0025] Even further, the Gaussian kernel is selected as the kernel function, and the hyperparameters are determined through grid search and cross-validation.

[0026] Furthermore, the anomaly detection process of the Isolation Forest model includes:

[0027] Automatically adjust the weights of the input features according to the tumor type;

[0028] Illustrate the contribution degree of the anomaly index and the potential associated diseases through generating a visual report.

[0029] Optimize the adaptability of the model to the local data distribution through transfer learning, and introduce a dynamic feature selection mechanism to improve the sensitivity and specificity of anomaly detection. The system can mine potential pathological associations from multi-dimensional indicators, generate interdisciplinary diagnosis and treatment suggestions, and at the same time enhance doctors' trust in the algorithm output through visual explanations, significantly improving the diagnosis and treatment efficiency and scientificity of complex cases.

[0030] Furthermore, generate a comprehensive doctor score according to dimensions such as doctor title, patient evaluation, and scientific research achievements.

[0031] Furthermore, a multidisciplinary expert recommendation list is generated based on the output results of the indicator analysis module and the doctor's professional labels, the doctor's practice status change information is synchronized through an incremental learning algorithm, and the patient's sensitive data is desensitized using homomorphic encryption technology.

[0032] Furthermore, it also includes a user interaction module for online doctor-patient consultations and report interpretation. By converting the patient's voice into structured text and triggering knowledge base retrieval through keyword matching, the trend changes of the patient's historical test data are displayed in the form of dynamic charts and abnormal indicators are indicated. It supports seamless switching between iOS, Android and Web, and the interface layout is adaptively adjusted according to the device screen size.

[0033] Accurate recommendations are achieved by combining doctors' professional capabilities to dynamically update labels, privacy-preserving computing, and multidisciplinary collaborative needs analysis. The fit between matching results and patients' conditions is significantly improved, and the dynamic scheduling efficiency of medical resources is optimized. At the same time, by ensuring the compliance of the recommendation logic, a personalized, full-cycle diagnosis and treatment service closed loop is built for patients.

[0034] Furthermore, the feature fusion module extracts text features from the test report through the BERT model, extracts medical image features through ResNet-50, and introduces the inter-modality covariance matrix to quantify the correlation between features:

[0035]

[0036] In the formula, is the dynamic fusion weight of text and image features; is the feature vector extracted from the test report text; is the feature vector extracted from the test image; and is the learnable linear transformation parameter; is the cross-modal interaction matrix; is the matrix trace operation.

[0037] The covariance matrix and attention mechanism are used to capture the deep semantic associations between text and images, and the temporal dependencies are modeled through long short-term memory networks to accurately characterize the dynamic evolution of tumor disease. The system's ability to parse complex multimodal data is significantly enhanced, and the interpretability of clinical decisions is greatly improved. At the same time, the computational load of redundant features is reduced, providing an efficient solution for real-time processing of high-dimensional medical data.

[0038] Furthermore, the weight optimization module dynamically adjusts the weights according to the difference between local and global data distribution:

[0039]

[0040] In the formula, is the weight of the th participant in the federated aggregation; is the th participant's local data probability distribution; is the estimated distribution of the global data; is the local distribution and the global distribution The Kullback-Leibler divergence between them; is the adjustment factor.

[0041] By quantifying the difference in cross-institutional data distribution through KL divergence, the federated aggregation weight is dynamically adjusted, and combined with meta-learning to quickly adapt to new institutional data in the small-sample scenario, the generalization of the model in cross-hospital collaboration is significantly improved, the risk of privacy leakage in the data sharing process is greatly reduced, and at the same time, the deployment cycle of the new institutional model is shortened, providing a secure and efficient distributed learning paradigm for multi-center medical collaboration.

[0042] Furthermore, the system basic data module is deployed on the hospital local server to reduce latency, and the core algorithm module is deployed on the cloud, and distributed computing and resource elastic expansion are realized through the API interface.

[0043] In a second aspect, the present invention also provides a method for interpreting the test report of tumor patients, and the method is based on the system described in the first aspect above, including:

[0044] Complete data input and preliminary integration according to the test sheet and medical images;

[0045] Extract text and image features, and achieve dynamic fusion through the attention mechanism;

[0046] Based on the collaborative analysis of SVM, isolation forest and decision tree, complete tumor classification, anomaly detection and triage recommendation generation;

[0047] Realize doctor matching based on disease characteristics according to the knowledge graph and privacy protection calculation;

[0048] Based on the dynamic contribution degree weight aggregation model, and achieve fast adaptation of new institutions through meta-learning;

[0049] Complete the closed-loop management of diagnosis and treatment and optimize patient services according to the visualization report and the online consultation entrance.

[0050] In a third aspect, the present invention also provides a computer device, comprising a management platform and a memory, wherein the management platform is connected to the memory, the memory is used to store computer programs, and the management platform is used to execute the computer programs stored in the memory, so that the computer device executes at least one step of the aforementioned method for interpreting test reports of tumor patients.

[0051] In a fourth aspect, the present invention further provides a computer-readable storage medium, wherein a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a management platform, at least one step of the aforementioned method for interpreting test reports of tumor patients is implemented.

[0052] The present invention integrates support vector machines, isolation forests and decision tree algorithms to build a full-chain analysis system from tumor classification, abnormal detection to triage matching; multi-dimensional label modeling based on knowledge graphs and incremental learning breaks through the static and one-sided nature of traditional rule matching; dynamic weight allocation and cross-modal time series modeling solve the feature fragmentation problem caused by independent analysis of traditional multimodal data; dynamic contribution evaluation and meta-learning adaptation mechanisms are proposed for the heterogeneity of medical data distribution and privacy protection needs. The system and method have a wide range of application scenarios and can realize the automatic interpretation of tumor patient test reports and the matching of professional doctors.

[0053] Beneficial Effects

[0054] By implementing the above-mentioned system and method for interpreting test reports of tumor patients provided by the present invention, the following technical effects are achieved:

[0055] (1) By integrating support vector machines, isolation forests, and decision tree algorithms, a full-chain analysis system from tumor classification, anomaly detection to triage matching is constructed. Transfer learning is used to optimize the model's adaptability to local data distribution, and a dynamic feature selection mechanism is introduced to improve the sensitivity and specificity of anomaly detection. The system can mine potential pathological associations from multi-dimensional indicators and generate interdisciplinary diagnosis and treatment recommendations. At the same time, visual explanations are used to enhance doctors' trust in the algorithm output, significantly improving the efficiency and scientificity of diagnosis and treatment of complex cases.

[0056] (2) Multi-dimensional label modeling based on knowledge graph and incremental learning breaks through the static and one-sided nature of traditional rule matching. Accurate recommendations are achieved by combining doctors’ professional capabilities to dynamically update labels, privacy-preserving computing, and multidisciplinary collaborative needs analysis. The matching results are significantly more consistent with the patient’s condition, and the dynamic scheduling efficiency of medical resources is optimized. At the same time, by ensuring the compliance of the recommendation logic, a personalized, full-cycle diagnosis and treatment service closed loop is built for patients.

[0057] (3) Solve the problem of feature fragmentation caused by the independent analysis of traditional multi-modal data through dynamic weight allocation and cross-modal temporal modeling. Use the covariance matrix and attention mechanism to capture the deep semantic association between text and images, and model the temporal dependence through the long short-term memory network to achieve an accurate portrayal of the dynamic evolution of the tumor condition. The system's ability to analyze complex multi-modal data is significantly enhanced, the interpretability of clinical decisions is greatly improved, and at the same time, the computational load of redundant features is reduced, providing an efficient solution for the real-time processing of high-dimensional medical data.

[0058] (4) Propose a dynamic contribution evaluation and meta-learning adaptation mechanism for the distribution heterogeneity and privacy protection requirements of medical data. Quantify the cross-institutional data distribution differences through KL divergence to dynamically adjust the federated aggregation weights, and combine meta-learning to quickly adapt to new institutional data in small-sample scenarios. The generalization of the model in cross-hospital collaboration is significantly improved, the risk of privacy leakage in the data sharing process is greatly reduced, and at the same time, the deployment cycle of the new institutional model is shortened, providing a secure and efficient distributed learning paradigm for multi-center medical collaboration. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] To make the above tumor patient test report interpretation system and method of the present invention more obvious and understandable, the drawings required for the specific implementation of the present invention will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative efforts.

[0060] Figure 1 Indicates the flowchart of the method for interpreting the test report of tumor patients;

[0061] Figure 2 Indicates the schematic diagram of the interface of the tumor patient test report interpretation system. DETAILED DESCRIPTION OF THE INVENTION

[0062] Example 1:

[0063] Provide a tumor patient test report interpretation system and method, and the method process is as Figure 1 shown, including an information verification module, a test report analysis module, a data storage module, and an index analysis module, which are specifically described as follows.

[0064] The information verification module is used for the identity verification and classification of patients and doctors.

[0065] The test report analysis module is used to structurally analyze the test report through natural language processing and computer vision technologies, and verify whether the qualifications of the testing unit pass the medical laboratory quality and ability recognition based on the fuzzy matching algorithm.

[0066] The verification of the qualification of the testing unit by the test report analysis module specifically includes:

[0067] (1) Establish a standardized name database of CNAS-certified laboratories, including full names, abbreviations, and historical name change records;

[0068] (2) Use OCR technology to extract the name of the testing institution from the test report image, and calculate its similarity with the database through the Levenshtein distance algorithm;

[0069] (3) Automatically mark it as a certified institution when the similarity ≥ 85%, otherwise trigger the manual review process and record the review results to optimize the database;

[0070] (4) Generate a risk prompt for the testing institution that fails to pass the certification and restrict its subsequent analysis authority for test reports.

[0071] The test report analysis module corrects the tilt and blur of the test report image through preprocessing of the image, uses OCR technology to extract key index data in the image, verifies the integrity of the data through a rule engine, and conducts a comparative analysis with the standard thresholds in the medical knowledge base.

[0072] The test report analysis module is based on a text localization subunit of a convolutional neural network, which is used to extract the text area from the test form image, uses a CRNN model to identify the test item name and value in the extracted area, and stores the recognition results in a standardized manner according to a predetermined format.

[0073] The predetermined format includes: test item - value - unit - reference range.

[0074] The data storage module is used to integrate knowledge base data related to tumor diagnosis and treatment, including tumor markers, treatment plans, and clinical guidelines.

[0075] The data storage module stores the tumor diagnosis and treatment data in a structured form of entity - relationship - attribute, and automatically corrects the outdated rules in the knowledge base by crawling the latest clinical research literature through web scraping technology.

[0076] The index analysis module is used to perform tumor type identification, analysis of abnormal index items, and triage matching through a cascaded model of support vector machines, isolation forests, and decision trees.

[0077] The index analysis module realizes the classification of tumor types based on kernel function optimization, calculates the abnormal score according to the path length, thereby dynamically adjusting the threshold to adapt to the abnormal detection requirements of different tumor types, and constructs a triage rule tree based on the principle of maximizing information gain to support multi-label classification to match interdisciplinary doctor teams.

[0078] The kernel function uses a Gaussian kernel, and the hyperparameters are determined through grid search and cross validation.

[0079] The anomaly detection process of the isolation forest model includes:

[0080] Automatically adjust the weights of input features based on tumor type;

[0081] Generate visual reports to illustrate the contribution of abnormal indicators and potential associated diseases.

[0082] Generate a comprehensive doctor score based on dimensions such as doctor's title, patient evaluation, and scientific research results.

[0083] A multidisciplinary expert recommendation list is generated based on the output results of the indicator analysis module and the doctor's professional labels, the doctor's practice status change information is synchronized through an incremental learning algorithm, and the patient's sensitive data is desensitized using homomorphic encryption technology.

[0084] It also includes a user interaction module for online doctor-patient consultations and report interpretation. It converts patient voice into structured text and triggers knowledge base retrieval through keyword matching. It displays trend changes in patients' historical test data in the form of dynamic charts and indicates abnormal indicators. It supports seamless switching between iOS, Android and Web, and the interface layout is adaptively adjusted according to the device screen size.

[0085] The system interface is as follows Figure 2 As shown, the specific steps for tumor patients to interpret test reports and match doctors in the system include:

[0086] Step 1: The patient registers and logs in using his / her mobile phone number. The system collects data such as the basic information of the cancer patient, specific pathological type, and treatment progress.

[0087] Step 2: The doctor registers and logs in using his / her mobile phone number. The system collects the doctor’s basic personal information, medical qualification certificate certification, level classification and other data.

[0088] Step 3: Identify and classify the contents of the test report, and determine whether the test unit meets the certification requirements; form a preliminary diagnosis and treatment opinion on the test report by retrieving knowledge data related to medical testing, and match the patient with a doctor in a related field.

[0089] Step 4: The patient starts the online consultation process. The doctor accepts the patient's questions and reviews the patient's information. According to the patient's condition, the doctor fills out the standardized receipt record form for the online consultation and provides the patient with a breakpoint chart of the patient's medical history results.

[0090] Step 5: The patient pays the doctor's consultation fee online. The doctor can withdraw the consultation remuneration stored in the system through the withdrawal system. After the consultation process ends, the system records the process and saves the consultation results.

[0091] Embodiment 2:

[0092] On the basis of the foregoing embodiment, a feature fusion module is added, specifically an adaptive multi-modal feature fusion mechanism, which is used to capture the complementarity of cross-modal features and the law of disease evolution through dynamic weight allocation and temporal modeling, so as to enhance the system's ability to analyze complex tumor data.

[0093] Extract the text features in the test report form through the BERT model , and extract the medical image features through ResNet-50 .

[0094] Construct a two-stream network, calculate the dynamic weights of text and image, and introduce the inter-modal covariance matrix Quantify the correlation between features:

[0095]

[0096] In the formula, is the dynamic fusion weight of text and image features; is the feature vector extracted from the text of the test report form; is the feature vector extracted from the image of the test report form; and are learnable linear transformation parameters for mapping text and image features to the same space; is the cross-modal interaction matrix for quantifying the correlation between text and image features; is the matrix trace operation for quantifying the interaction intensity of the covariance matrix.

[0097] Verification shows that when obtaining an average error similar to that of the above embodiment, on the TCGA dataset, the tumor type recognition accuracy rate is increased from 89.2% to 93.5%, the feature dimension is reduced by about 30%, the inference speed is increased by about 25%, and the weight visualization shows that the text contribution ratio is 60%-80%, which is consistent with the clinical dependence logic. The results show that this mechanism can more comprehensively integrate text and image features, capture the dynamic evolution law of tumor indicators, significantly improve the fit between the disease classification result and the clinical actual needs. At the same time, the efficiency optimization of the feature fusion process reduces redundant calculations, the system response speed is significantly accelerated, and the weight visualization function enhances the doctor's trust in the algorithm decision-making logic, providing reliable technical support for precise diagnosis and treatment.

[0098] Embodiment 3:

[0099] On the basis of the foregoing embodiments, a weight optimization module is further added, specifically a dynamic contribution degree evaluation and meta-learning adaptation mechanism, which is used to quantify the data distribution difference through KL divergence and perform fine-tuning based on small samples to achieve accurate aggregation and personalized deployment of cross-institutional models.

[0100] Each hospital trains the SVM model based on local data and outputs parameters .

[0101] Dynamically adjust the weights according to the local and global data distribution differences:

[0102]

[0103] In the formula, is the weight of the th participant in the federated aggregation; is the probability distribution of the local data of the th participant; is the estimated distribution of the global data; is the local distribution and the global distribution The Kullback-Leibler divergence between them is used to quantify the distribution difference; is a regulation factor, a hyperparameter used to control the influence of KL divergence on the contribution degree weight.

[0104] The global model is updated as:

[0105]

[0106] In the formula, is the model parameter locally trained by the th participant; is the global model parameter after federated aggregation.

[0107] When a new institution joins, use the MAML algorithm to perform fine-tuning on 10% of the local data to generate a personalized model.

[0108] For example, there are 5 hospitals (H1-H5), with strong heterogeneity in data distribution, the local model SVM has an RBF kernel, and the number of federated rounds is 50.

[0109] Assume the local data distributions of hospitals H1-H5 and the global distribution The KL divergences are respectively:

[0110]

[0111] Assume the regulation factor , then the contribution degree weights are:

[0112]

[0113] Calculated as:

[0114]

[0115] The global model parameters are:

[0116]

[0117] Newly added hospital Using 10% local data, after fine-tuning through MAML, the local accuracy rate has increased from 68% to 85%.

[0118] The effect of the weight optimization module is shown in Table 1.

[0119] Table 1. Summary of module effects

[0120]

[0121] According to the experimental table, Hetero-FL reduces the impact of distribution differences through dynamic weights. The standard deviation of the cross-hospital test accuracy rate has decreased from 12.3% to 4.7%, indicating that the stability of the model under different data distributions has been significantly enhanced; the contribution weight optimizes the parameter aggregation direction, and the number of rounds required to reach the 90% benchmark accuracy rate has been reduced by 40%; the KL divergence penalty term reduces the exposure of sensitive data, and the leakage probability has been reduced to 0.8%; the meta-learning mechanism shortens the fine-tuning time of new institutions from 12 hours to 2 hours, and the adaptation efficiency has been increased by 83.3%. This framework demonstrates stronger adaptability in cross-institutional collaboration. The performance stability of the model under different data characteristics has been significantly improved. At the same time, the cycle of model convergence and new institution access has been greatly shortened. In addition, the strengthening of the privacy protection mechanism has significantly reduced the risk of sensitive data leakage, providing a systematic guarantee for the secure sharing and collaborative optimization of medical data, and promoting the balanced and intelligent development of tumor diagnosis and treatment resources.

[0122] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable non-transitory storage media containing computer-usable program code.

[0123] The present invention can provide computer program instructions to a management platform of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed through the management platform of the computer or other programmable data processing devices generate means for implementing the system.

[0124] These computer program instructions can also be stored in a computer-readable memory capable of guiding a computer or other programmable data processing devices to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implements the functions of the system.

[0125] These computer program instructions can also be loaded onto a computer or other programmable data processing devices, such that a series of operation steps are executed on the computer or other programmable devices to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable devices provide steps for implementing the functions of the system.

Claims

1. A tumor patient test report interpretation system, characterized in that, Including: An information verification module for authenticating and classifying the identities of patients and doctors; A test report analysis module for structurally analyzing test reports through natural language processing and computer vision technologies, and verifying the qualifications of testing units based on a fuzzy matching algorithm; A data storage module for integrating knowledge base data related to tumor diagnosis and treatment; An index analysis module for tumor type identification, analysis of abnormal index items, and triage matching through a cascaded model of support vector machines, isolation forests, and decision trees; A feature fusion module for realizing temporal fusion of cross-modal features based on dynamic weight allocation and long short-term memory networks for text features extracted from the test report and image features extracted from medical images; The weight optimization module is used to quantify the distribution difference between the local data and the global data of each hospital in federated learning through the KL divergence, and perform fine-tuning on 10% of the local data of the newly connected medical institutions through meta-learning to generate a personalized model; the weight optimization module dynamically adjusts the weights according to the distribution difference: ; where is the weight of the th participant in the federated aggregation; is the probability distribution of the local data of the th participant; is the estimated distribution of the global data; is the local distribution and the global distribution the Kullback-Leibler divergence between them; is the adjustment factor; each hospital trains the SVM model based on the local data and outputs the parameter , and the global model is updated to: ; where is the model parameter locally trained by the th participant; is the global model parameter after federated aggregation.

2. The system according to claim 1, wherein: The verification of the qualifications of the testing unit by the test report analysis module specifically includes: (1) Establishing a database of CNAS-certified laboratory lists; (2) Extracting the name field of the testing institution from the test report image; (3) Performing fuzzy matching between the extracted name and the database and returning a matching score; (4) When the matching degree ≥ 85%, it is determined as a certified institution, otherwise, an artificial review process is triggered.

3. The system according to claim 1, wherein: The index analysis module classifies tumor types based on kernel function optimization, dynamically adjusts the threshold according to the path length to adapt to the abnormal detection requirements of different tumor types, and constructs a triage rule tree based on the principle of maximizing information gain to match interdisciplinary doctor teams.

4. The system according to claim 1, wherein: Based on the output results of the index analysis module and doctor professional labels, a multi-disciplinary expert recommendation list is generated, the change information of doctor practice status is synchronized through an incremental learning algorithm, and homomorphic encryption technology is used to desensitize patient sensitive data.

5. The system according to claim 1, wherein: It further includes a user interaction module for online doctor-patient consultation and report interpretation, triggering knowledge base retrieval by converting patient voice into structured text and through keyword matching, and displaying the trend changes of the patient's historical test data in the form of dynamic charts and indicating abnormal indicators.

6. The system according to claim 1, wherein: The feature fusion module extracts text features from the test report through a BERT model, extracts medical image features through ResNet-50, and introduces an inter-modal covariance matrix to quantify the correlation between features: wherein, is the dynamic fusion weight of text and image features; is the feature vector extracted from the text of the test report; is the feature vector extracted from the image of the test report; and are learnable linear transformation parameters; is the cross-modal interaction matrix; is the matrix trace operation.

7. A method for interpreting test reports of tumor patients, wherein: The implementation of the method is based on the system according to any one of claims 1-6: The method includes: Completing data input and preliminary integration according to the test form and medical images; Extracting text and image features, and realizing temporal fusion of cross-modal features based on dynamic weight allocation and long short-term memory networks; Completing tumor classification, abnormal detection, and triage recommendation generation based on a cascaded model of support vector machines, isolation forests, and decision trees; Realizing doctor matching based on disease characteristics according to the knowledge graph and privacy protection calculation; Aggregate the federated learning model based on the dynamic contribution degree weight, and perform fine-tuning on 10% of the local data of the newly connected medical institutions through meta-learning to achieve rapid adaptation; Complete the closed-loop management of diagnosis and treatment and optimize patient services according to the visualization report and the online consultation entrance.

8. A computer device, comprising a management platform and a memory, the management platform being connected to the memory, the memory being used for storing computer programs, and characterized in that: The management platform is used to execute the computer program stored in the memory, so that the computer device executes at least one step of the method described in claim 7.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is run, at least one step of the method described in claim 7 is implemented.

Images