Intelligent thyroid interrogation system

Abstract

The invention relates to the technical field of thyroid treatment, in particular to an intelligent thyroid inquiry system which comprises an inquiry system and is used for intelligent auxiliary diagnosis of thyroid diseases, and the inquiry system comprises a heterogeneous data integration unit which is used for synchronously accessing electronic medical record data, ultrasonic DICOM images, patient chief complaint voice and wearable equipment monitoring data; the multi-dimensional feature extraction unit is used for processing the ultrasonic image by adopting a deep convolutional network to extract nodule features, and analyzing a medical history text through a BiLSTM network; the dynamic decision center unit integrates a GNN-based knowledge graph inference engine and an interpretability algorithm, and outputs a differential diagnosis list; the diagnosis accuracy is improved through multi-modal data fusion, and comprehensive analysis is realized by integrating image, examination and medical history data; the diagnosis and treatment process is optimized by adopting an intelligent inquiry path, the problem sequence is dynamically adjusted, critical cases are preferentially processed, and the inquiry efficiency is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of thyroid treatment, and in particular to an intelligent thyroid diagnosis system. Background Art

[0002] With the development of society, people are paying more and more attention to their health, and thus medical and health services have gradually entered people's field of vision. Medical and health services provide patients with all-round, three-dimensional and one-stop health management services throughout the entire treatment process. The content is very extensive, covering health education before, during and after diagnosis, involving medical facilities, real-time patient consultation, continuous follow-up visits, family doctors, outpatient management, as well as diagnosis and treatment of common and frequently occurring diseases, on-site emergency care, home visits, home care, referral services, rehabilitation medicine and other convenient services.

[0003] After checking the publication number: CN118969240A, a thyroid consultation system is disclosed. This technology discloses "a thyroid consultation system, including: an intelligent medical module, nine thyroid diagnosis modules; the intelligent medical module is used to provide data analysis support for the nine thyroid diagnosis modules; the intelligent medical module includes: a user question preprocessing module, a module vector library and a large language model; the user question preprocessing module uses the NLP algorithm to achieve semantic understanding of user questions and named body recognition, including using keywords in the form of external dictionaries to analyze user questions, extract user question keywords, standardize user questions, and vectorize user questions; the function of the module vector library is to vectorize knowledge and store it, and match it in the vector library according to the standardized user question vector; the large language model is used to understand the user's questions and summarize the knowledge in combination with the questions and answers matched by the vector library and answer the user in an appropriate tone" and other technical solutions, with technical effects such as "capable of realizing intelligent consultation for thyroid diseases";

[0004] The current thyroid disease diagnosis and treatment system has two key flaws: First, the traditional diagnostic model is limited by data integration capabilities and cannot achieve the systematic integration of multi-source heterogeneous medical data such as ultrasound images, biochemical indicators, electronic medical records and patient complaints, resulting in the fragmentation of key diagnostic information; second, the existing consultation process uses a standardized template, which lacks an adaptive adjustment mechanism for individual patient differences (such as age, medical history, and symptom characteristics), and lacks the intelligent identification and priority treatment functions for acute and severe diseases such as hyperthyroid crisis, making it difficult to improve the efficiency and quality of diagnosis and treatment. Summary of the Invention

[0005] In response to the shortcomings of existing technologies, the present invention provides an intelligent thyroid consultation system, which improves diagnostic accuracy through multimodal data fusion and integrates imaging, test and medical history data to achieve comprehensive analysis; it uses intelligent consultation paths to optimize the diagnosis and treatment process, dynamically adjusts the question sequence and prioritizes critical cases, significantly improving consultation efficiency.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a thyroid intelligent consultation system, including a consultation system and used for intelligent auxiliary diagnosis of thyroid diseases, the consultation system comprising:

[0007] Heterogeneous data integration unit for synchronous access to electronic medical record data, ultrasound DICOM images, patient voice complaints, and wearable device monitoring data;

[0008] A multi-dimensional feature extraction unit uses a deep convolutional network to process ultrasound images to extract nodule features and uses a BiLSTM network to analyze medical history text;

[0009] The dynamic decision-making central unit integrates the GNN-based knowledge graph reasoning engine and explainability algorithm to output a differential diagnosis list.

[0010] Preferably, the multi-dimensional feature extraction unit includes:

[0011] The three-stage segmentation module sequentially performs the first stage of locating the thyroid contour, the second stage of segmenting the nodule area, and the third stage of annotating key features;

[0012] A real-time enhancement processing module uses a generative adversarial network to improve the resolution of low-quality images on mobile devices and automatically annotates key TI-RADS classification features;

[0013] The time series comparison module performs non-rigid registration on serial ultrasound images of follow-up patients and calculates the nodule volume change rate and elasticity score evolution trend.

[0014] Preferably, the dynamic decision-making central unit includes:

[0015] The graph neural network reasoning module generates a preliminary diagnostic hypothesis set based on graph attention calculation of the thyroid disease knowledge graph;

[0016] The interpretability analysis module quantifies the contribution of each clinical feature through SHAP value analysis and marks key decision factors;

[0017] The multimodal decision fusion module uses an adaptive weighted algorithm to integrate data and output a final diagnosis list with confidence scores.

[0018] Preferably, the medical consultation system further includes a medical consultation path generation unit for dynamically generating and optimizing the medical consultation process according to patient characteristics, and the medical consultation path generation unit includes:

[0019] The initial classifier module loads differentiated questionnaire templates based on patient demographic characteristics and basic medical history;

[0020] Dynamic adjustment module, using the DQN algorithm with reinforcement learning strategy to optimize the problem sequence;

[0021] The emergency triage mechanism module automatically jumps to the hyperthyroidism crisis assessment branch when TSH < 0.01mIU / L with rapid nodule growth characteristics is detected.

[0022] Preferably, the medical consultation system further includes a learning and training unit and is used for multi-center collaborative modeling under medical data privacy protection, and the learning and training unit includes:

[0023] Distributed training architecture module: each medical node stores the original data locally and only uploads the model gradient parameters;

[0024] The differentiated contribution evaluation module calculates the contribution of each institution's data to the global model through Shapley value and assigns weights accordingly;

[0025] The concept drift detection module verifies the consistency of model output in each region and triggers retraining when the Kappa value is less than 0.6.

[0026] Preferably, the contribution calculation formula in (152) is:

[0027]

[0028] Where N is the set of all medical institutions participating in the learning, S is the subset excluding i, and v(·) is the model performance evaluation function.

[0029] Preferably, the calculation formula of the Kappa value in (153) is:

[0030]

[0031] Among them, p o is the observed consistency ratio, p e is the expected consistency ratio by chance; the value range is [-1,1], >0.6 indicates acceptable consistency, and <0.4 indicates poor consistency.

[0032] Preferably, the medical consultation system further includes a decision interaction unit for three-dimensional dynamic display of thyroid disease risk maps, treatment plan comparisons, and AR-assisted patient education. The decision interaction unit includes:

[0033] The 3D risk heat map module integrates the probability of nodule malignancy, the degree of thyroid function abnormality, and the risk of complications into a 3D visualization model;

[0034] Treatment plan comparison module, showing the 5-year survival rate prediction of drug therapy, ablation therapy and surgical resection;

[0035] Patient education AR module displays a 3D holographic projection of the thyroid anatomy and lesion location.

[0036] The present invention provides an intelligent thyroid diagnosis system. Compared with the existing technology, it has the following advantages:

[0037] 1. Synchronous access and fusion of multi-source medical data are achieved through the heterogeneous data integration unit; the multi-dimensional feature extraction unit uses a deep convolutional network and a BiLSTM network to process image and text data respectively, ensuring the comprehensive extraction of nodule characteristics and medical history information; the dynamic decision-making center unit integrates multimodal features based on graph neural networks and interpretable algorithms to generate a high-confidence differential diagnosis list; this multimodal fusion technology significantly improves the comprehensiveness and accuracy of diagnosis, and is particularly suitable for complex diseases such as thyroid diseases that require a combination of imaging, laboratory indicators and clinical symptoms. It effectively avoids the limitations of a single data source and provides more reliable decision support for clinicians.

[0038] 2. The consultation path generation unit achieves real-time optimization of the personalized consultation process through an initial classifier module and a dynamic adjustment module. It loads differentiated questionnaires based on patient demographics and medical history, and utilizes a reinforcement learning algorithm to dynamically reduce the scale of questions, significantly improving consultation efficiency while ensuring diagnostic accuracy. The emergency triage mechanism module automatically identifies critical values ​​and immediately switches to a specialized assessment branch, ensuring that high-risk patients receive priority treatment. This not only shortens consultation time but also reduces redundant examinations through intelligent path adjustment, optimizing medical resource allocation. It is particularly suitable for thyroid specialists with high outpatient volume and complex disease types.

[0039] 3. The learning and training unit utilizes a federated learning architecture, where medical institutions share only encrypted model gradient parameters, while raw data is always stored locally, strictly adhering to medical data privacy requirements. The differentiated contribution assessment module quantifies the contribution of each institution's data to the global model using Shapley values, equitably distributing weights. The concept drift detection module regularly verifies regional model consistency and triggers adaptive retraining to address changes in data distribution. This design not only resolves the "data silo" problem between medical institutions but also enhances the model's generalization capabilities through multi-center collaborative training. It is particularly suitable for scenarios where the spectrum of thyroid diseases varies significantly across regions, ensuring both data security and model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a block diagram of the medical inquiry system of the present invention;

[0041] Figure 2is a block diagram of a multi-dimensional feature extraction unit in the present invention;

[0042] Figure 3 A block diagram of the dynamic decision-making central unit in the present invention;

[0043] Figure 4 This is a block diagram of a medical inquiry path generating unit in the present invention;

[0044] Figure 5 A block diagram of a learning and training unit in the present invention;

[0045] Figure 6 It is a block diagram of the decision interaction unit in the present invention.

[0046] In the figure: 1. Medical consultation system; 11. Heterogeneous data integration unit; 12. Multi-dimensional feature extraction unit; 121. Three-tiered segmentation module; 122. Real-time enhanced processing module; 123. Time series comparison module; 13. Dynamic decision-making center unit; 131. Graph neural network reasoning module; 132. Explainability analysis module; 133. Multimodal decision fusion module; 14. Medical consultation path generation unit; 141. Initial classifier module; 142. Dynamic adjustment module; 143. Emergency diversion mechanism module; 15. Learning and training unit; 151. Distributed training architecture module; 152. Differentiation contribution evaluation module; 153. Concept drift detection module; 16. Decision interaction unit; 161. Three-dimensional risk heat map module; 162. Treatment plan comparison module; 163. Patient education AR module. DETAILED DESCRIPTION

[0047] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0048] See also Figure 1 - Figure 6 The present invention provides a technical solution: a thyroid intelligent consultation system, including a consultation system 1 and used for intelligent auxiliary diagnosis of thyroid diseases, the consultation system 1 includes:

[0049] Heterogeneous data integration unit 11, used for synchronous access to electronic medical record data, ultrasound DICOM images, patient complaints and wearable device monitoring data;

[0050] Multi-dimensional feature extraction unit 12 uses a deep convolutional network to process ultrasound images to extract nodule features and uses a BiLSTM network to analyze medical history text;

[0051] The dynamic decision-making central unit 13 integrates the GNN-based knowledge graph reasoning engine and interpretability algorithm to output a differential diagnosis list.

[0052] In this embodiment, the heterogeneous data integration unit 11 receives multi-source medical data through a standardized interface protocol. The electronic medical record data is structured and converted into a unified format. The ultrasound DICOM image is decompressed and metadata is extracted. The patient's voice complaints are converted into text through automatic speech recognition technology. The wearable device data is filtered and feature extracted and then stored in a temporary cache. Secondly, the multi-dimensional feature extraction unit 12 starts the image processing and text analysis processes in parallel. The image processing channel uses a deep convolutional network to extract the morphological features and blood flow distribution features of nodules in ultrasound images. The text analysis channel uses a BiLSTM network to identify key clinical symptoms and time series patterns in medical history documents. Finally, The dynamic decision-making central unit 13 inputs the extracted multimodal features into the knowledge graph reasoning engine built based on the graph neural network, generates preliminary diagnostic hypotheses through node embedding and relational reasoning, and then analyzes the decision contribution of each feature through an interpretable algorithm, and outputs a differential diagnosis list with confidence assessment; the system establishes a full-process data quality control mechanism, sets up anomaly detection and automatic correction functions in key links such as feature extraction and decision reasoning, and provides a manual review interface to ensure the reliability of the diagnosis results; asynchronous communication is achieved between units through message queues to ensure system stability in high-concurrency scenarios, and fully records key intermediate results in the data processing and decision-making process to support subsequent traceability analysis and model optimization.

[0053] Specifically, the multi-dimensional feature extraction unit 12 includes:

[0054] The three-stage cascade segmentation module 121 sequentially performs the first stage of locating the thyroid contour, the second stage of segmenting the nodule area, and the third stage of labeling key features;

[0055] A real-time enhancement processing module 122 uses a generative adversarial network to improve the resolution of low-quality images on mobile devices and automatically annotate key features of TI-RADS classification;

[0056] The time series comparison module 123 performs non-rigid registration on the serial ultrasound images of the follow-up patients and calculates the nodule volume change rate and elasticity score evolution trend.

[0057] In this embodiment, the three-stage segmentation module 121 utilizes a cascaded deep neural network architecture. In the first stage, an encoder-decoder structure precisely locates the anatomical boundaries of the thyroid organ. In the second stage, an attention mechanism is used to focus on the nodule region for pixel-level segmentation. In the third stage, a morphological analysis and feature extraction algorithm are combined to annotate key imaging features such as blood flow distribution and calcification characteristics of the nodule. The real-time enhancement processing module 122 deploys a generative adversarial network model to enhance the quality of low-resolution ultrasound images acquired by mobile devices. Computer vision algorithms are also integrated to automatically identify and annotate imaging features relevant to TI-RADS classification. The temporal comparison module 123 uses non-rigid image registration technology to align ultrasound images from multiple follow-up visits of a patient. Differential analysis is used to calculate the dynamic trends of nodule morphology and elasticity characteristics, and to construct a three-dimensional spatiotemporal evolution model. Each module works collaboratively using a pipeline architecture. A reprocessing mechanism is automatically triggered when the output quality of the previous module does not meet standards. A quality control log for the entire feature extraction process is established to ensure the reliability and consistency of the feature data used for subsequent diagnostic decisions. The system supports physicians in manually verifying and correcting the automatically extracted features, and the correction results are fed back to each module for iterative model optimization.

[0058] Specifically, the dynamic decision-making central unit 13 includes:

[0059] The graph neural network reasoning module 131 generates a preliminary diagnostic hypothesis set based on graph attention calculation of the thyroid disease knowledge graph;

[0060] The interpretability analysis module 132 quantifies the contribution of each clinical feature through SHAP value analysis and marks key decision factors;

[0061] The multimodal decision fusion module 133 integrates the data using an adaptive weighting algorithm and outputs a final diagnosis list with confidence scores.

[0062] In this embodiment, the graph neural network reasoning module 131 constructs a knowledge graph containing thyroid disease entities and their associations, dynamically calculates the importance weights of different clinical feature nodes through the graph attention mechanism, and generates a preliminary diagnostic hypothesis set based on graph reasoning; the interpretability analysis module 132 applies the SHAP value analysis method to quantitatively evaluate the contribution of each clinical feature to the diagnostic result, identify and mark key decision-making factors, and generate a visual explanation report; the multimodal decision fusion module 133 receives multi-source feature data from ultrasound image analysis, laboratory tests, and medical history texts, uses an adaptive weighted fusion algorithm to integrate the intermediate results of different modalities, and generates a final diagnosis list with confidence scores through probability calibration; a two-way data channel is established between the modules. When the interpretability analysis module identifies that a key feature is missing, it automatically triggers the supplementary reasoning process of the graph neural network module, and at the same time, the multimodal fusion results will be fed back to the knowledge graph for dynamically updating the node association weights; during the diagnostic reasoning process, the system records the intermediate calculation results and decision paths of each module in real time to support physicians in reviewing and verifying the entire process.

[0063] Specifically, the medical consultation system 1 further includes a medical consultation path generation unit 14 and is used to dynamically generate and optimize the medical consultation process according to the patient's characteristics. The medical consultation path generation unit 14 includes:

[0064] The initial classifier module 141 loads a differentiated questionnaire template based on the patient’s demographic characteristics and basic medical history;

[0065] The dynamic adjustment module 142 uses the DQN algorithm with reinforcement learning strategy to optimize the problem sequence;

[0066] The emergency triage mechanism module 143 automatically jumps to the hyperthyroidism crisis assessment branch when TSH < 0.01mIU / L with rapid nodule growth characteristics is detected.

[0067] In this embodiment, the initial classifier module 141 analyzes structured data such as the patient's age, gender, and medical history to match the most suitable basic consultation process from a preset questionnaire template library and generate an initial question sequence. The dynamic adjustment module 142 uses a deep reinforcement learning algorithm to monitor the patient's answers to each question and their correlation with the collected clinical characteristics in real time, dynamically optimize the sorting and content selection of subsequent questions, and gradually reduce the number of questions while ensuring diagnostic accuracy. At the same time, the emergency triage mechanism module 143 continuously scans the input laboratory test results and imaging features. When an abnormal indicator combination that meets the hyperthyroidism crisis warning conditions is identified, the regular consultation process is immediately interrupted and automatically jumps to a dedicated assessment interface for critical cases that includes vital signs monitoring, emergency treatment recommendations, and specialist physician liaison functions. The system automatically saves the current consultation progress when executing path switching and provides an option to resume regular consultation after the critical situation is handled. All consultation path adjustment decisions are recorded in the audit log to support subsequent quality assessment and algorithm optimization.

[0068] Specifically, the medical consultation system 1 further includes a learning and training unit 15 and is used for multi-center collaborative modeling under medical data privacy protection. The learning and training unit 15 includes:

[0069] In the distributed training architecture module 151, each medical node stores the original data locally and only uploads the model gradient parameters;

[0070] The differentiated contribution evaluation module 152 calculates the contribution of each institution's data to the global model through Shapley value and assigns weights accordingly;

[0071] The concept drift detection module 153 verifies the consistency of the model outputs of each region and triggers retraining when the Kappa value is less than 0.6.

[0072] In this embodiment, the distributed training architecture module 151 deploys model training nodes on the local servers of each medical institution and establishes a federated learning network through a secure communication protocol. After each node completes forward propagation and backpropagation calculations on the local dataset, it only uploads the encrypted model gradient parameters to the central aggregation server. Secondly, the differentiated contribution assessment module 152 uses the Shapley value algorithm based on secure multi-party computation to evaluate the marginal contribution of the gradient updates provided by each medical node to the improvement of global model performance in an encrypted space, and dynamically adjusts the aggregation weight of each node in the next round of training based on the contribution ratio. At the same time, the concept drift detection module 153 regularly distributes standardized test case sets to each node, detects potential medical data distribution shifts by statistically analyzing the Kappa consistency coefficient of the output results of each node, and automatically triggers the model retraining process when significant differences are detected. This process includes steps such as gradient reweighting, local model fine-tuning, and global parameter fusion. The entire training process uses blockchain technology to record operation logs to ensure that the data usage and model update behavior of each medical node are auditable and traceable.

[0073] Specifically, the contribution calculation formula in the differentiated contribution evaluation module (152) is:

[0074]

[0075] Where N is the set of all medical institutions participating in the learning, S is the subset excluding i, and v(·) is the model performance evaluation function.

[0076] In this embodiment, the differentiated contribution evaluation module 152 receives encrypted model parameter update data from each participating medical institution; secondly, the Shapley value is approximately calculated through the Monte Carlo sampling method, and cooperative subset combinations of different sizes are constructed for each medical institution to evaluate the degree of change in the model performance indicators before and after the institution is added; then, the contribution ranking of each institution is generated based on the marginal contribution calculation results, and the data sampling ratio or gradient update weight in the next round of training is adjusted according to the preset weight distribution strategy; at the same time, the module will record the historical contribution change trend, and when it is detected that the contribution of an institution continues to decline, it will automatically trigger the data quality assessment process and generate an improvement suggestion report; in addition, the differentiated contribution evaluation module 152 establishes a data linkage mechanism with the concept drift detection unit 153 to dynamically optimize the weight distribution scheme of participating nodes during the model retraining process.

[0077] Specifically, the calculation formula of the Kappa value in the concept drift detection module (153) is:

[0078]

[0079] Among them, p o is the observed consistency ratio, pe is the expected consistency ratio by chance; the value range is [-1,1], >0.6 indicates acceptable consistency, and <0.4 indicates poor consistency.

[0080] In this embodiment, the concept drift detection module 153 constructs a cross-comparison matrix by regularly collecting the diagnostic results of the same batch of standardized test cases from each participating medical institution; secondly, by comparing the consistency of the diagnostic results between different institutions, the difference between the actually observed consistency ratio and the randomly expected consistency ratio is calculated, and the Kappa coefficient formula is applied for quantitative evaluation; when the consistency level is detected to be lower than the preset threshold, the module automatically triggers the following processing flow: generates an inter-institutional diagnostic difference report and marks the main difference categories, starts the incremental training mechanism of the federated learning model, and pushes model update notifications to each participating node; at the same time, the system records the occurrence time, impact scope and corrective measures of each concept drift event to form a traceable quality control log; and the concept drift detection module 153 works in conjunction with the differentiated contribution evaluation module 151 to ensure that the parameter update weights are dynamically adjusted according to the contribution of each institution during the model retraining process.

[0081] Specifically, the medical consultation system 1 further includes a decision interaction unit 16 for dynamically displaying a thyroid disease risk map, comparing treatment plans, and AR-assisted patient education in three dimensions. The decision interaction unit 16 includes:

[0082] The 3D risk heat map module 161 integrates the probability of nodule malignancy, the degree of thyroid function abnormality, and the risk of complications into a 3D visualization model;

[0083] Treatment plan comparison module 162, showing the 5-year survival predictions for drug therapy, ablation therapy, and surgical resection;

[0084] Patient education AR module 163 displays a three-dimensional holographic projection of the thyroid anatomy and lesion location.

[0085] In this embodiment, the three-dimensional risk heat map module 161 uses a multi-layer data fusion algorithm to weightedly integrate the patient's ultrasound imaging features, laboratory test indicators and clinical history data to generate a risk distribution model with deep information. The model supports physicians to rotate, slice and adjust transparency through touch gestures; the treatment plan comparison module 162 is based on the knowledge graph reasoning engine to extract the indications, expected efficacy and potential adverse reaction characteristics of different treatment plans, and construct an interactive parallel decision tree. When the physician chooses a specific treatment path, the system automatically highlights the key clinical evidence and typical case references corresponding to the plan; the patient education AR module 163 uses SLAM spatial positioning technology to capture the real environment plane through the smart terminal camera, and renders a three-dimensional holographic image of the thyroid gland that matches the patient's actual anatomy in real time. Physicians can use gestures to separate tissue layers, mark the range of lesions, and dynamically simulate the impact of different treatment plans on the anatomical structure. The system simultaneously generates a diagnosis and treatment record animation containing instructions for key steps.

[0086] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0087] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A thyroid intelligent diagnosis system, characterized by: The invention comprises a medical inquiry system (1) and is used for intelligent auxiliary diagnosis of thyroid diseases, wherein the medical inquiry system (1) comprises: Heterogeneous data integration unit (11), used for synchronous access to electronic medical record data, ultrasound DICOM images, patient complaints and wearable device monitoring data; A multi-dimensional feature extraction unit (12) uses a deep convolutional network to process ultrasound images to extract nodule features, and analyzes medical history text through a BiLSTM network; The dynamic decision-making central unit (13) integrates the GNN-based knowledge graph reasoning engine and interpretability algorithm to output a differential diagnosis list.

2. The intelligent thyroid diagnosis system according to claim 1, characterized in that: The multi-dimensional feature extraction unit (12) comprises: A three-stage cascade segmentation module (121) sequentially performs the first stage of locating the thyroid contour, the second stage of segmenting the nodule region, and the third stage of labeling key features; A real-time enhancement processing module (122) uses a generative adversarial network to improve the resolution of low-quality images on mobile devices and automatically annotates key features of TI-RADS classification; The time series comparison module (123) performs non-rigid registration on the serial ultrasound images of the follow-up patients and calculates the nodule volume change rate and elasticity score evolution trend.

3. The intelligent thyroid diagnosis system according to claim 1, characterized in that: The dynamic decision-making central unit (13) comprises: The graph neural network reasoning module (131) generates a preliminary diagnostic hypothesis set based on graph attention calculation of the thyroid disease knowledge graph; The interpretability analysis module (132) quantifies the contribution of each clinical feature through SHAP value analysis and marks the key decision factors; The multimodal decision fusion module (133) uses an adaptive weighting algorithm to integrate data and output a final diagnosis list with confidence scores.

4. The intelligent thyroid diagnosis system according to claim 1, characterized in that: The medical consultation system (1) further includes a medical consultation path generation unit (14) for dynamically generating and optimizing a medical consultation process according to patient characteristics. The medical consultation path generation unit (14) includes: The initial classifier module (141) loads differentiated questionnaire templates based on patient demographic characteristics and basic medical history; The dynamic adjustment module (142) uses the DQN algorithm with reinforcement learning strategy to optimize the problem sequence; The emergency triage mechanism module (143) automatically jumps to the hyperthyroidism crisis assessment branch when TSH < 0.01mIU / L with rapid nodule growth characteristics is detected.

5. The intelligent thyroid diagnosis system according to claim 1, characterized in that: The medical consultation system (1) further includes a learning and training unit (15) and is used for multi-center collaborative modeling under medical data privacy protection. The learning and training unit (15) includes: Distributed training architecture module (151), each medical node stores the original data locally and only uploads the model gradient parameters; The differentiated contribution evaluation module (152) calculates the contribution of each institution's data to the global model through the Shapley value and assigns weights accordingly; The concept drift detection module (153) verifies the consistency of the model outputs in each region and triggers retraining when the Kappa value is less than 0.

6.

6. The intelligent thyroid diagnosis system according to claim 5, characterized in that: The contribution calculation formula in the differentiated contribution evaluation module (152) is: Where N is the set of all medical institutions participating in the learning, S is the subset excluding i, and v(·) is the model performance evaluation function.

7. The intelligent thyroid diagnosis system according to claim 5, characterized in that: The calculation formula of the Kappa value in the concept drift detection module (153) is: Among them, p o is the observed consistency ratio, p e is the expected consistency ratio by chance; the value range is [-1,1], >0.6 indicates acceptable consistency, and <0.4 indicates poor consistency.

8. The intelligent thyroid diagnosis system according to claim 1, characterized in that: The medical consultation system (1) further includes a decision interaction unit (16) for three-dimensional dynamic display of thyroid disease risk maps, treatment plan comparisons, and AR-assisted patient education. The decision interaction unit (16) includes: The three-dimensional risk heat map module (161) integrates the probability of nodule malignancy, the degree of thyroid function abnormality, and the risk of complications into a three-dimensional visualization model; Treatment plan comparison module (162), which displays the 5-year survival predictions for drug therapy, ablation therapy, and surgical resection; Patient education AR module (163) displays a three-dimensional holographic projection of the thyroid anatomy and lesion location.

Images