Disease screening system based on large model

Abstract

The invention provides a disease screening system based on a large model. The disease screening system is used for solving the technical problem that an existing disease screening system is intelligently used for single special disease screening. The system comprises a scheduling model and a plurality of AI auxiliary diagnosis models, the scheduling model is connected with the plurality of AI auxiliary diagnosis models, and the scheduling model is connected with a big data disease library. According to the method, a high-level scheduling model is utilized, multiple single AI auxiliary diagnosis models are managed in a centralized mode, automatic calling of a multi-disease AI system is achieved, automatic structured report generation is achieved through a large language model, historical medical history is combined, progress is predicted, treatment suggestions and reference cases are automatically given, an existing clinician film reading workflow is fitted, and the efficiency is improved. The whole process of actual diagnosis decision making of doctors is greatly fitted, and the working efficiency of the doctors can be improved to the maximum extent.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of AI-assisted diagnosis, and particularly relates to a disease screening system based on a large model. BACKGROUND

[0002] Generally, the existing image examination process is as follows: a radiologist analyzes information such as a patient's case, medical history and related symptoms, determines the part (head, chest or abdomen, etc.) and the examination method (CT, MRI or X-ray, etc.) that need to be examined, and then uses the corresponding picture reading software to read the corresponding image results for detailed analysis. In this process, the doctor needs to check and examine each DICOM image and constantly adjust the most suitable window width, window level and other parameters according to the image results of different parts, adjust the best imaging according to different organs or structures, and use various measurement tools to observe whether there are abnormalities and suspicious lesions in different structures and organs, and whether there are signs that may cause adverse consequences, so as to achieve comprehensive diagnosis of diseases, and finally give the possible diseases and supporting evidence involved in the image results to discover suspicious lesions as soon as possible.

[0003] At present, the working logic of most AI-assisted diagnosis software is single and based on specific part images to screen specific diseases, which is quite different from the conventional clinical workflow. Without the active selection of the doctor, all possible diseases on the image cannot be screened. The existing AI-assisted diagnosis system is usually set for specific diseases, for example, a lung nodule detection system can only detect lung nodules from chest X-rays. It cannot meet the existing clinical workflow of suspicious multiple disease screening at the same time, and can only be used in single disease screening scenarios.

[0004] Patent application number 202410396731.6 relates to the field of medical health information technology, specifically a disease prediction and risk assessment method based on a medical big model. The method comprises the following steps: S1: Collecting multi-dimensional health data from patients; S2: Cleaning, outlier processing, and feature engineering the data collected in S1; S3: Constructing a deep learning-based medical big model; S4: Optimizing the medical big model constructed in S3; S5: Applying the optimized medical big model to disease prediction and risk assessment; S6: Utilizing model interpretation techniques to clarify the decision-making basis for model predictions; and S7: Providing targeted health management and disease prevention recommendations based on the prediction and assessment results of S5. This invention, by integrating multi-dimensional health data to construct a deep learning medical big model, not only significantly improves the accuracy and interpretability of disease predictions, promotes personalized health management and disease prevention, but also provides effective support for reducing medical costs and improving public health quality. The method also examines whether the disease targets include all possible problems; whether the examination sites cover the entire body; whether the judgment principles are consistent with evidence-based medicine; and whether the patient-doctor relationship is patient-friendly. The information obtained by the patient and the doctor should be inconsistent, and whether information isolation is implemented. The aforementioned patent focuses on macroscopic, multidimensional health data integration and disease risk assessment, but fails to address specific, practical, and feasible technical designs or innovations for physicians' real-world workflows, efficiency improvements, multi-disease screening, automated structured report generation, progress tracking, and intelligent collaboration within medical imaging scenarios. The aforementioned patent fails to align with the actual workflow of imaging departments. Traditional radiologists rely heavily on reviewing DICOM images layer by layer, adjusting window widths and window positions in real time, flexibly switching measurement tools, and gradually narrowing the diagnostic scope for different anatomical structures. While emphasizing the integration of "multidimensional health data" to build a large deep learning model, the patent focuses on disease risk prediction and health management, but fails to optimize the actual image analysis process for physicians (such as layered review, tool switching, and dynamic adjustment of image parameters). There are no specific designs tailored to clinical needs for imaging software's human-computer interaction, actual review experience, and report generation. The aforementioned patent is unable to achieve automated screening for multiple diseases and scheduling of AI models. Specifically, the current AI software's "single disease, single model" is its biggest drawback, meaning it can only provide judgments for specific diseases. If doctors want to screen for multiple diseases concurrently, they must manually switch or repeat the operation, which is inefficient and out of touch with actual clinical work patterns. The large model described in the aforementioned patent "predicts the overall risk of disease," but does not mention how to automatically call multiple disease AI models during image analysis, nor does it have a high-level scheduling strategy to coordinate the detection of various diseases. The lack of a mechanism to achieve "automatic screening for all possible concurrent diseases in images" does not fundamentally address the actual need of clinicians to "detect all suspicious lesions." The aforementioned patent does not support structured reporting and disease progression analysis based on LLM.The modern clinical requirements for AI-assisted reporting are not only to make disease predictions, but also to automatically generate comprehensive structured diagnostic reports, including positive / negative evidence, important measurement values, findings conclusions, etc., and to link with previous image and medical record data to analyze the current disease trend. The above patent only vaguely mentions "health management suggestions" in the last stage, lacking details of image report automation and dynamic disease management. The above patent lacks sufficient explainability to meet the actual needs of medical treatment, although "model explanation" is added in the steps, but it does not specifically explain whether it can clearly mark the image lesion location, quantify the abnormal area, and correlate the evidence, etc., which cannot meet the high requirements of doctors for traceability and explainability. The visualization basis that cannot clearly support image diagnosis decision-making is not easy to integrate into the doctor's process. The above patent lacks innovative solutions for actual auxiliary operations in the imaging department, and is more inclined to "medical health big data risk control" or "comprehensive health management", which does not make customized design for the actual work characteristics of the medical imaging department, which are multiple, complex, and variable. It does not solve the many "pain points" of doctors in daily software review, such as automatic adaptation to different anatomical structures, batch measurement and comparison, existing examination / historical medical data linkage, etc. SUMMARY

[0005] In order to solve the technical problem that the existing disease screening system is intelligent for single disease screening, the application provides a disease screening system based on a large model, which uses a high-level scheduling model to centrally manage a plurality of single AI-assisted diagnosis models, thereby breaking the status quo that the existing AI system can only perform disease screening and diagnosis for specific functional images.

[0006] In order to achieve the above-mentioned purpose, the technical scheme of the application is as follows: a disease screening system based on a large model, comprising a scheduling model and a plurality of AI-assisted diagnosis models, the scheduling model being connected with the plurality of AI-assisted diagnosis models, and the scheduling model being connected with a big data disease library.

[0007] Preferably, the scheduling model adopts a three-layer scheduling architecture, comprising an image distribution layer, a model selection layer, and a result integration layer, the image distribution layer being connected with the model selection layer, and the model selection layer and the result integration layer both being connected with the AI-assisted diagnosis models; the result integration layer being connected with the big data disease library to read the previous medical history of the current patient and the case with higher similarity.

[0008] Preferably, the image distribution layer receives input medical image data, pre-processes the medical image by using DICOM analysis and image enhancement, and judges whether the image quality meets the diagnosis requirements; decomposes the medical image sequence into single slices, and identifies the anatomical region by using a human anatomical positioning algorithm; pre-processes the identified anatomical region by using corresponding window width and window level adjustment, image enhancement, and size normalization; decides to distribute the corresponding image data to the AI-assisted diagnosis model according to the image content; The image distribution layer utilizes an intelligent decision tree scheduling algorithm to automatically construct an optimal calling sequence according to the parameters of the examination site, sequence, and window width and window level in the DICOM metadata in the image; the intelligent decision tree scheduling algorithm performs serial / parallel screening according to the examination for common possible diseases, adjusts the execution order of the AI-assisted diagnosis model according to the diagnosis result, and automatically retrieves other examination results for joint determination if other lesions are suspected and there are other images or examination data.

[0009] Preferably, the model selection layer intelligently selects the AI-assisted diagnosis model that needs to be called according to the anatomical region identified by the image distribution layer; coordinates the parallel / serial execution order of multiple AI-assisted diagnosis models; sets appropriate detection parameters and thresholds for different AI-assisted diagnosis models; and manages the allocation of computing resources.

[0010] Preferably, the result integration layer collects the detection results output by each AI-assisted diagnosis model; processes the conflicts of different judgments on the same region by multiple AI-assisted diagnosis models; combines a big data disease library to make comprehensive judgments and risk assessments and develop treatment reference schemes based on the detection results; automatically generates a structured comprehensive diagnosis report; and utilizes a large language model for natural language processing and report optimization.

[0011] Preferably, each type of disease screening model in the AI-assisted diagnosis model includes but is not limited to a lung nodule detection model, a benign and malignant analysis model, a cerebral apoplexy detection and lesion segmentation and analysis model, a fracture detection model, and an organ-specific disease model for tumor screening. According to the principles of anatomy and pathology, the AI-assisted diagnosis model is divided into an organ positioning model, a lesion detection model, a quantitative analysis model, and a benign and malignant prediction model. The organ positioning model is connected to the lesion detection model, the lesion detection model is connected to the quantitative analysis model, and the quantitative analysis model is connected to the benign and malignant prediction model. The organ positioning model is realized by an All-in-One model of uAI. The lesion detection model includes more than 20 special disease detection models for lung nodule detection, liver lesion detection, and brain abnormality detection. The quantitative analysis model measures the quantitative parameters of the volume, density, and boundary of the lesion. The benign and malignant prediction model predicts the benign and malignant of the detected lesion.

[0012] Preferably, the intelligent decision tree scheduling algorithm includes but is not limited to at least one of FCFS, SPT, LPT, EDD, or CR. Adopting a dynamic resource allocation strategy, GPU / CPU resources are allocated according to model complexity and priority to ensure that critical models are executed first; the dynamic resource allocation strategy adopts a hybrid scheduling strategy, EDD for emergency queues and SPT for regular queues; CR is used for complex checks; batch processing is used; according to the current system load and task type, dynamically select the algorithm, set the priority weight of different disease types, analyze the patient's disease severity combined with the big prophecy model, and intelligently adjust the hybrid scheduling strategy; The large language model is trained based on 10 million radiology reports and 5,000 medical textbooks; the large language model receives the complete information of the patient, including the current imaging examination results, medical history, and laboratory examination data; through the deep analysis of the large language model enhanced by medical knowledge, the patient information is matched with the "disease-symptom-treatment-effect" four-element relationship network in the knowledge graph to identify the disease pattern and characteristics of the patient; combined with the evidence strength of different diagnoses and treatment schemes according to the evidence-based medicine grading standard, a comprehensive report containing diagnosis analysis, disease assessment, treatment recommendations, and prognosis prediction is generated; The large data disease library is a hierarchical treatment knowledge base, and the construction method is: a) bottom layer data source: integrate 30 million de-identified electronic medical records, 5 million follow-up records, and 100,000 clinical trial data; b) middle layer knowledge representation: construct a "disease-symptom-treatment-effect" four-element relationship network using knowledge graph technology to obtain an expert knowledge graph; c) top-level decision support: grade annotation of treatment schemes based on evidence-based medicine evidence levels.

[0013] Preferably, when multiple models give different conclusions for the same region, a combination of confidence weighting and expert rules is used to solve the conflict; the confidence weighting is that each AI-assisted diagnosis model outputs a confidence score when outputting the diagnosis result, indicating the certainty of the AI-assisted diagnosis model to the diagnosis result, and according to the algorithm performance, different weights are given to the model in different scenarios, including anatomical site weight, disease type weight, and image quality weight; the expert rule is a logical rule based on medical knowledge and clinical experience, including diagnosis standards in medical literature, clinical expert experience, and clinical pathways.

[0014] The implementation method of the combination of confidence weighting and expert rules is: input: multiple AI-assisted diagnosis model diagnosis results → first layer: expert rule pre-screening, removing medically impossible results and marking high-risk features → second layer: confidence weighting calculation, weighted voting on the remaining results to generate preliminary diagnosis probability → third layer: expert rule post-processing, adjusting the final probability according to the rules, adding diagnosis basis and suggestions → output: final diagnosis result + confidence + basis.

[0015] Preferably, the medical knowledge enhanced large language model is set in the result integration layer, based on a general large language model architecture, using more than 10 million radiology reports and 5,000 medical textbooks for fine-tuning in different fields; the fine-tuning is a phased fine-tuning, and the implementation method is: first phase: medical knowledge injection, using medical textbooks for knowledge pre-training; second phase: report generation fine-tuning, using radiology reports for supervised fine-tuning; third phase: instruction fine-tuning optimization, instruction fine-tuning for different tasks; The large language model includes at least one of GPT-3, GPT3.5, GPT4, XLNet, RoBERTa, and ALBERT.

[0016] Preferably, the result integration layer is provided with a multi-modal fusion processing algorithm, and a "image-text-history" three-modal fusion algorithm is designed to fuse the detection results of the AI assisted diagnosis model, the key features of the image original data, and the patient's past medical history through a self-attention mechanism to generate a comprehensive understanding. The image features are deep feature vectors extracted by CNN, the text data are feature vectors converted from structured JSON of AI detection results, and the medical history data are embedding vectors converted by a medical record encoder; the features of the three modalities are projected into the same feature space, and the projected feature vectors are spliced; a multi-head attention module is used for weighted fusion of multi-modal data features.

[0017] The result integration layer is provided with a disease evolution evaluation algorithm based on time series analysis, which compares and analyzes the results of continuous multiple examinations, combines the current multi-modal fusion results, uses an LSTM network structure with a medical knowledge module gate switch to model the disease progression trend, and combines an expert knowledge graph to constrain the disease development law. The modeling includes: feature alignment, aligning the same anatomical structures of the previous examinations; change detection: calculating the change rate of key indicators; LSTM modeling: inputting the feature sequence of historical time points.

[0018] A medical dedicated uncertainty quantification mechanism is introduced in the result integration layer as a final evaluation module for all prediction results; according to the evidence strength, the number of similar cases, and the prediction confidence, the certainty degree of diagnosis and prediction is expressed in a standardized language. A patient feature vector-based similar case retrieval system is developed, which comprehensively considers more than 30 dimensions of patient feature vectors including demographic characteristics, pathological indicators, genotypes, and image features; a high-efficiency indexing method combining local sensitive hashing and vector database is used to realize millisecond-level similar case retrieval; the first layer uses local sensitive hashing for rapid pre-screening, and 32 hash tables and 128 hash functions are used to map similar patients into the same or similar hash buckets; the second layer uses a vector database for accurate similarity calculation; a similarity weight adjustment mechanism verified by medical experts is introduced to strengthen the matching importance of key features, and an expert-verified feature importance weight library is established, and the basic weight of each feature is determined through multiple rounds of expert questionnaires and clinical verification; The result integration layer develops a treatment decision tree generation algorithm to generate the potential effect and risk probability of each treatment plan according to the treatment effect statistics of similar cases; based on the treatment path of similar cases, a decision tree node is generated, and each node is attached with effect statistics; a multi-objective optimization algorithm is used to balance in multiple dimensions of treatment effect, risk control, quality of life, and economic cost; a comprehensive medical ethics rule library is built in, including the principles of non-maleficence, beneficence, autonomy, and justice, and an ethical review is conducted on each treatment recommendation; The result integration layer designs a three-level evidence display mechanism, the first level displays the final treatment recommendation and expected effect; the second level displays the list of similar cases supporting the recommendation after clicking to expand, including the number of cases, success rate statistics, and evidence level; the third level displays the details of the specific original case, including patient characteristics, treatment process, and follow-up results; an interactive body adjustment interface of the treatment plan is developed; a learning feedback loop is established to continuously optimize the treatment recommendation algorithm according to the adoption of doctors.

[0019] Compared with the prior art, the beneficial effects of the present application are: by using a high-level scheduling model to realize automatic calling of a multi-disease AI system, introducing a large language model (LLM) to complete a comprehensive structured examination and diagnosis report, and automatically analyzing the report results combined with the previous medical history, giving the current progress of the related disease course, the future prediction of the progress (if there is a related disease), and giving customized treatment suggestions and related reference cases (visible to doctors) according to the existing related disease database, so that patients can understand their current physical condition and doctors can make the best treatment plan; it is consistent with the existing clinical doctor's reading work flow, which can maximize the doctor's work efficiency. The present application realizes automatic structured report generation through a large language model (LLM), combines with the history of the disease, predicts the progress, and automatically gives treatment suggestions and reference cases, which greatly fits the whole process of the doctor's actual diagnosis and decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1 This is a principle block diagram of the present invention.

[0022] In the figure, 1 is the scheduling model, 2 is the AI-assisted diagnosis model, and 3 is the big data disease library. DETAILED DESCRIPTION

[0023] The following will clearly and completely describe 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 creative work are within the scope of protection of the present invention.

[0024] like Figure 1 As shown, a disease screening system based on a large model includes a scheduling model 1 and multiple AI-assisted diagnosis models 2. The scheduling model 1 is connected to the multiple AI-assisted diagnosis models respectively, and the scheduling model 1 is connected to the big data disease library 3. The high-level scheduling model 1 is used to centrally manage many single AI-assisted diagnosis models, breaking the current situation where the existing AI system can only perform specialized disease screening and diagnosis based on specific functional images. Specifically, the scheduling model 1 is a layered scheduling architecture: a three-layer scheduling architecture is adopted, including an image distribution layer, a model selection layer, and a result integration layer, to achieve collaborative work between various AI-assisted diagnosis models. The image distribution layer is connected to the model selection layer, and the model selection layer and the result integration layer are both connected to the AI-assisted diagnosis model. The result integration layer is connected to the big data disease library 3 to read the current patient's past medical history and matching cases with high similarity in the big data disease library 3.

[0025] The image distribution layer receives input medical image data (such as DICOM format and pathology PNG files), performs image preprocessing, including DICOM parsing and image enhancement, and determines whether the image quality meets diagnostic requirements. It decomposes the medical image sequence into individual slices and uses human anatomical positioning algorithms to identify anatomical regions (such as the lungs, liver, and brain). It also performs preprocessing on the identified anatomical regions, including window width and window level adjustment, image enhancement, and size normalization. Distribution strategy development: Based on the image content, it determines which AI-assisted diagnosis models to distribute the corresponding image data to.

[0026] The functions of the model selection layer include: 1. Intelligent model matching: intelligently selecting the AI-assisted diagnosis model to be called according to the anatomical region identified by the image distribution layer; 2. Task scheduling management: coordinating the parallel / serial execution order of multiple AI-assisted diagnosis models; 3. Parameter configuration: setting appropriate detection parameters and thresholds for different AI-assisted diagnosis models; 4. Load balancing: managing the allocation of computing resources to avoid system overload.

[0027] The functions of the result integration layer include: 1. Result collection and summary: collecting the detection results output by each AI-assisted diagnosis model; 2. Conflict resolution: handling conflicts between different judgments of the same region by multiple AI-assisted diagnosis models; 3. Comprehensive analysis: making comprehensive judgments and risk assessments based on the detection results, and formulating treatment reference schemes, in combination with the big data disease library; 4. Report generation: automatically generating a structured comprehensive diagnosis report. Large language models (LLM) are used for natural language processing and report optimization.

[0028] The various disease screening models in the AI-assisted diagnosis model include, but are not limited to, lung nodule detection, benign and malignant analysis models, stroke detection lesion segmentation and analysis models, fracture detection models, and organ-specific disease models such as tumor screening models.

[0029] Model classification and dynamic calling: According to the principles of anatomy and pathology, the AI-assisted diagnosis model is divided into the following categories: organ positioning model, lesion detection model, quantitative analysis model, and benign and malignant prediction model. The organ positioning model is connected to the lesion detection model, the lesion detection model is connected to the quantitative analysis model, and the quantitative analysis model is connected to the benign and malignant prediction model. The organ positioning model is used to quickly locate the key organs and tissues in the image, and is used to locate large organs through the All-in-One model of uAI. The lesion detection model includes more than 20 special disease detection models such as lung nodule detection, liver lesion detection, and brain abnormality detection. The quantitative analysis model measures the volume, density, and boundary of the lesion. On the segmentation result, the general formula algorithm based on mathematical and physical priori is used for measurement. The benign and malignant prediction model predicts the benign and malignant of the detected lesions. Each model has its own benign and malignant discrimination model. The disease screening system proposed in the present application can be used for AI-assisted diagnosis models for disease screening, including but not limited to lung nodule detection software, coronary artery segmentation software, brain tumor segmentation and classification software, and fracture detection software.

[0030] Intelligent decision tree scheduling algorithm. The image distribution layer automatically constructs the optimal calling sequence according to the parameters such as examination site, sequence and window width and window level in the DICOM metadata in the image. For example, for chest CT, after processing by the image distribution layer, the model selection layer will preferentially call the sequence process of lung organ segmentation model → lung nodule detection model → emphysema evaluation model → mediastinal lymph node detection model. Similar to the description of lung organ segmentation, it is a further partitioned segmentation model for each organ, such as the division of the lung into left lung, right lung, lung segment and lung lobe, which belongs to the model selection layer. For a certain disease, it is an integrated integration of lesion detection, quantitative analysis and benign and malignant discrimination. The intelligent decision tree scheduling algorithm is to perform serial / parallel screening according to the examination for common possible diseases. For the lung, only the lung-related models are preferentially scheduled, but according to the diagnosis result, there will be some adjustment of the execution order of the models (for example, if pneumonia is diagnosed, then the next step is to judge whether there is pleural effusion; if there is no pneumonia, then the next step is to judge whether there is pulmonary fibrosis); and if other lesions are suspected and there are other images or examination data, the other examination results will be automatically called to make a joint diagnosis.

[0031] The scheduling algorithms that can be used by the scheduling model 1 include, but are not limited to, FCFS (First-Come, First-Served), SPT (Shortest Processing Time), LPT (Longest Processing Time), EDD (Earliest Due Date), and CR (Critical Ratio), etc. The advantages of FCFS are simple implementation, intuitive logic, absolute fairness, no starvation phenomenon, and small system overhead. The disadvantages are that the average waiting time can be long, it is not friendly to short tasks (short tasks can be blocked by long tasks), and the urgency of the task cannot be considered. In the medical scene, emergency patients can be delayed due to queuing. Medical application scenarios: suitable for routine physical examination and non-urgent screening. The advantages of SPT are minimizing the average waiting time, improving the overall throughput of the system, high efficiency of computing resource utilization, and quickly processing a large number of lightweight tasks. The disadvantages are that long tasks can be indefinitely postponed (starvation phenomenon), the processing time needs to be estimated in advance, important but time-consuming examinations can be delayed, and the medical importance of the task is not considered. Medical application scenarios: suitable for batch processing of simple screening (such as X-ray rapid screening). The advantages of LPT are avoiding long tasks being indefinitely postponed, contributing to load balancing in parallel processing environments, and suitable for complex scenarios mixed with short tasks. The disadvantages are that the waiting time of short tasks is increased, the overall response speed of the system is reduced, the condition of light symptoms but urgency can be delayed in medical, and the overall throughput is relatively low. Medical application scenarios: suitable for complex image analysis (such as comprehensive analysis of whole body PET-CT). The advantages of EDD are that it can meet the needs of time-sensitive tasks, suitable for medical scenarios with clear time limits, ensure that urgent tasks are processed first, and meet the needs of medical priority management. The disadvantages are that the deadline of the medical task is difficult to define accurately, which can lead to unbalanced resource utilization, and tasks without a deadline can be indefinitely postponed. Medical application scenarios: emergency department image diagnosis, preoperative necessary examination. The dynamic calculation priority of CR = (deadline - current time) / remaining processing time. The advantages are dynamic priority, considering both urgency and workload, balancing time requirements and resource consumption, strong adaptability, and adjusting the priority with time changes. It is theoretically the optimal comprehensive scheduling scheme. The disadvantages are high computational complexity, accurate time estimation and deadline setting, and frequent priority changes can cause system instability. Medical application scenarios: multi-department coordination in comprehensive hospitals, multi-modal image analysis of complex cases, and MDT scenarios.

[0032] Parallel computing resource allocation: Dynamic resource allocation strategy is adopted to allocate GPU / CPU resources according to model complexity and priority, ensuring that critical models are executed first. Generally, the dynamic resource allocation strategy adopts a hybrid scheduling strategy. Emergency queue: EDD is used to ensure that critical conditions are prioritized. Regular queue: SPT is used to improve overall efficiency. Complex examination: CR is used to balance urgency and resource consumption. Batch processing: FCFS is used to ensure fairness. When dynamic resource allocation is needed, the algorithm is dynamically selected based on the current system load and task type, and the priority weight of different disease types is set. Combined with LLM analysis of patient condition urgency, the scheduling strategy is intelligently adjusted. When GPU resources are limited, SPT is used first. When multi-modal AI models need to be coordinated, CR is used. In emergency medical situations, EDD mode is switched to.

[0033] Model conflict resolution mechanism. When multiple models give different conclusions for the same region, a combination of confidence weighting and expert rules is used to solve the conflict and improve diagnostic accuracy. Confidence weighting is that each AI model outputs a confidence score (usually between 0 and 1) when outputting the diagnostic result, indicating the model's certainty about the result. At the same time, different weights are given to models in different scenarios according to algorithm performance, where anatomical site weight is that some models perform better in certain parts, disease type weight is that specialized disease detection models get higher weight, and image quality weight is that all model weights are reduced when image quality is poor. Expert rules are logical rules based on medical knowledge and clinical experience. Diagnostic criteria in medical literature (such as Lung-RADS grading standard, BI-RADS breast diagnostic standard), clinical expert experience (diagnostic rules summarized by experienced physicians, such as certain image feature combinations usually indicate a specific disease), clinical pathway (mutually exclusive or symbiotic relationship between diseases, as well as statistical rules based on large sample research, etc.). Confidence weighting and expert rules combination: input: multi-model diagnostic results → first layer: expert rule pre-screening, remove medically impossible results, mark high-risk features → second layer: confidence weighting calculation, weighted voting on remaining results to generate preliminary diagnostic probability → third layer: expert rule post-processing, adjust final probability according to rules, add diagnostic basis and recommendations → output: final diagnostic result + confidence + basis.

[0034] The combination of confidence weighting and expert rules can bring the following effects: 1) accuracy improvement, false diagnosis rate reduction, through multi-model voting and medical rule double verification; false positive reduction: expert rule filtering results that do not conform to medical logic; complex case diagnosis improvement: comprehensive judgment combined with multiple evidence chains; 2) reliability enhancement, consistency guarantee: ensure that the results conform to medical common sense and clinical logic; robustness improvement: single model anomaly will not seriously affect the final result; confidence calibration: provide more accurate uncertainty estimation; 3) improved interpretability, transparent decision-making process: clearly show the voting process and rule application; sufficient medical basis: provide explanations based on medical knowledge; good traceability: can track the basis of each decision step; 4) clinical applicability, consistent with clinical thinking: simulate the process of comprehensive analysis by doctors, reduce learning cost: doctors are easy to understand and accept system decision, improve trust: transparent decision-making process enhances doctor trust; 5) system optimization, continuous learning: collect conflict cases to continuously optimize the rule base, personalized tuning: adjust parameters according to different hospital, department characteristics, quality monitoring: find weak links in the system through conflict analysis. This conflict resolution mechanism truly realizes the organic combination of "technical intelligence" and "medical wisdom", and provides more reliable and credible AI-assisted diagnosis support for clinical practice.

[0035] The disease screening system provided by the present application, the resources that need to be adjusted by the scheduling model 1 include but are not limited to the reasonable use of a plurality of AI auxiliary diagnosis models 2 under limited resources, the specificity of the current image, the reasonable use time of the LLM, etc.

[0036] Compared with existing AI disease detection software, the disease screening system provided by the present application is more in line with the routine workflow of radiologists, and is more easy to conduct comprehensive screening on concurrent diseases, and to achieve no missed detection of serious diseases and early detection of minor diseases.

[0037] The disease screening system provided by the present application can not only give the results of this image examination and issue a structured examination report by introducing a large language model, but also can conduct comprehensive analysis combined with the past medical history and the existing examination results, give the current disease progression, and estimate the future development. The specific implementation is as follows: The medical knowledge enhanced large language model is arranged in the result integration layer, which is based on a general large language model architecture and fine-tuned using more than 10 million radiology reports and 5,000 medical textbooks in a specific field. The result integration layer receives the detection results of multiple AI models, integrates multi-modal information, generates a final structured report, and conducts comprehensive analysis and prediction.

[0038] Staged fine-tuning method: first stage: medical knowledge injection, pre-training using medical textbooks (training target = masked language modeling + medical entity recognition + medical relation extraction / input: ["patient presents", "[MASK]", "signs, may be malignant lesions"] / target: ["patient presents", "hair spur", "signs, may be malignant lesions"]); second stage: report generation fine-tuning, supervised fine-tuning using radiology reports (input: image features + patient information / target: standardized radiology report / prompt = """image features: lung nodule, diameter 12mm, edge spurs; patient information: 65-year-old male, 30-year smoking history; past examination: 6 months ago, nodule diameter 8mm; please generate a standardized report: """ ); third stage: instruction fine-tuning optimization, instruction fine-tuning for specific tasks (instruction types = ["generate diagnosis report", "analyze disease progression", "predict disease development", "provide treatment recommendations"]).

[0039] Construct an enhanced LLM specifically for medical image report generation. The large language model used in this invention includes but is not limited to GPT-3 (Generative Pre-trained Transformer 3), GPT3.5, GPT4, XLNet (eXtended Language understanding Network), RoBERTa (Robustly optimized BERT approach), ALBERT (A LiteBERT); GPT3.5 is an improved version of GPT-3, enhancing instruction following ability; GPT4 is the fourth generation of GPT, with multi-modal capabilities, supporting text and image input.

[0040] Multi-modal fusion processing algorithm is set in the result integration layer, which collects key nodes from all information sources, including original features from the image distribution layer, results from multiple AI models called by the model selection layer, and external medical history data. A "image-text-history" three-modal fusion algorithm is designed to integrate AI-assisted diagnosis model detection results (in structured JSON format), key features of image original data, and patient medical history through self-attention mechanism to generate comprehensive understanding.

[0041] Image features are deep feature vectors extracted by CNN, text data is structured JSON converted from AI detection results to feature vectors, and medical history data is converted to embedding vectors by medical record encoder; project the features of the three modalities into the same feature space, and concatenate the projected feature vectors; then use the multi-head attention module for weighted fusion of multi-modal data features. The effects are: improved diagnostic accuracy: complementary multi-modal information; reduce false negative rate: medical history information makes up for early lesions with non-obvious images; enhance explainability: attention weight shows the contribution of each modality; personalized diagnosis: combined with patient-specific information.

[0042] Time series analysis and disease progression prediction: a "disease evolution evaluation algorithm" based on time series analysis is developed, which is set in the result integration layer as a special module, which needs to integrate historical multiple examination results, combined with the current multi-modal fusion results, to provide comprehensive progress evaluation. By comparing the results of continuous multiple examinations, a modified LSTM network structure is used to model the disease progression trend, and finally combined with the expert knowledge graph to constrain the disease development rule to improve the prediction accuracy. The modified LSTM network is a MedLSTM network with professional constraints by adding a medical knowledge module gate switch, including sequentially set time sequence main module, knowledge constraint gate module, disease-specific constraint module, final time step feature module and output module. The modeling includes: 1. Feature alignment, aligning the same anatomical structures of the previous examinations; 2. Change detection: calculate the change rate of key indicators (such as nodule volume change); 3. LSTM modeling: input the feature sequence of historical time points.

[0043] The process of disease development rule constraint is as follows: taking lung cancer and pneumonia as an example, load the knowledge graph, for lung cancer: the monthly growth upper limit is 30%, the diameter > 20mm has a higher risk of metastasis, and the volume doubling time is 30-400 days; for pneumonia: improvement within 14 days, double lung progression probability less than 0.2; constraint 1-growth rate limit, monthly growth exceeding 30% reduces the confidence of the model. Constraint 2-TMN staging logic constraint, rare combination of T4N0 reduces the confidence of the model. Constraint 3-recovery time constraint; more than 30 days still have progression, consider other diseases.

[0044] Uncertainty quantification and expression: A medical-specific uncertainty quantification mechanism is introduced in the result integration layer as the final evaluation module of all prediction results. According to the evidence strength, the number of similar cases and the prediction confidence, the certainty degree of diagnosis and prediction is expressed in standardized language. Evidence strength calculation: image feature clarity: 0.9 (clear) ~ 0.3 (fuzzy), model feature specificity: 0.9 (high) ~ 0.2 (low), multi-model consistency: standard deviation of multiple AI model outputs 0~1, historical result consistency: 0~1, knowledge base support degree: number of matching cases in knowledge base (log standardization). The prediction confidence is the variance of the probability distribution of LSTM prediction, and the variance of multiple inferences (MC Dropout). The expression in standardized language: when the certainty level is very high, the probability range is >95%, and the standard expression is "definite diagnosis"; when the certainty level is high, the probability range is 80-95%, and the standard expression is "highly suggestive"; when the certainty level is medium, the probability range is 60-80%, and the standard expression is "considered as a tendency"; when the certainty level is low, the probability range is 40-60%, and the standard expression is "cannot be ruled out"; when the certainty level is very low, the probability range is <40%, and the standard expression is "evidence is insufficient, follow-up is recommended".

[0045] Medical standard term mapping: An automatic mapping algorithm from natural language to standard medical terms (such as ICD-10, SNOMED CT) is designed to ensure the standardization and interoperability of reports. Post-processing before report generation, two-stage mapping architecture: A [natural language description]-->B (entity recognition model), B -->C {ICD-10 encoder}, C -->D [standard term library check], D -->E [SNOMED CT mapping]; Conflict resolution mechanism: specialty priority, high-frequency term priority, knowledge graph reasoning confidence priority; Handle special cases: New term learning: update the mapping table through online learning, dialect processing: build a local hospital terminology library.

[0046] The disease screening system provided by the present application has a structured human disease treatment database, which can not only use LLM to analyze the current patient, but also can give the corresponding treatment scheme and result by matching similar cases, so that the doctor can refer to it and give the best treatment scheme in the shortest time. The specific implementation is: LLM receives the complete information of the patient, including the current imaging results, medical history, laboratory tests, and other data. Then it conducts a deep analysis through a large language model enhanced by medical knowledge. The large language model is trained based on 10 million radiology reports and 5,000 medical textbooks, and has professional medical reasoning capabilities. During the analysis process, LLM matches the patient's information with the "disease-symptom-treatment-effect" four-element relationship network in the knowledge graph, identifies the patient's disease pattern and characteristics, and evaluates the evidence strength of different diagnoses and treatment plans based on evidence-based medicine grading standards. Finally, it generates a comprehensive report containing diagnosis analysis, disease assessment, treatment recommendations, and prognosis prediction.

[0047] The big data disease library 3 is a hierarchical treatment knowledge base, and the construction method is: a) bottom layer data source: 300 million de-identified electronic medical records, 5 million follow-up records, and 100,000 clinical trial data are integrated. b) middle layer knowledge representation: a four-element relationship network of "disease-symptom-treatment-effect" is constructed using knowledge graph technology to obtain an expert knowledge graph. c) top-level decision support: treatment plans are graded based on evidence-based medicine evidence levels (I-IV). The diseases included in the big data disease library 3 used in the invention include but are not limited to lung cancer, brain tumors, and heart diseases. Level I: High-quality randomized controlled trials and systematic reviews, with sample sizes typically exceeding 1000 cases; Level II: Small-scale randomized controlled trials and cohort studies, with sample sizes of 100-1000 cases; Level III: Case-control studies, with sample sizes of 50-100 cases; Level IV: Case series reports and expert opinions, with sample sizes less than 50 cases. The system automatically analyzes the supporting evidence behind each treatment recommendation, evaluates the research type, sample size, bias risk, and result consistency, and then automatically labels the corresponding evidence level. Each treatment plan clearly displays its evidence level, such as "I-level recommendation: based on 3 large-scale RCT studies" or "III-level suggestion: based on retrospective analysis". This hierarchical labeling significantly improves the doctor's trust in AI recommendations, allowing doctors to quickly identify treatment plans supported by high-quality evidence and avoid relying on low-quality evidence. Clinical applications show that doctors' acceptance of I-level and II-level evidence-supported recommendations exceeds 85%, and overall decision-making efficiency improves by 50%, reducing treatment delays due to insufficient evidence.

[0048] Multi-dimensional similarity matching algorithm: a) A patient feature vector-based similar case retrieval system was developed, considering more than 30 dimensions such as demographic characteristics, pathological indicators, genotypes, and imaging features. The patient feature vector includes 30+ dimensions, such as demographic characteristics (age group, gender, BMI classification, smoking status, drinking history, occupation type, geographic region), pathological indicators (tumor size, histological type, differentiation degree, TNM stage, biomarker status, gene mutation spectrum, inflammation marker, tumor marker, metastasis mode, vascular invasion, lymphatic invasion, neural invasion), genotypes (EGFR mutation, KRAS mutation, ALK fusion, ROS1 fusion, PD-L1 expression, microsatellite instability, tumor mutation burden, genetic syndromes), and imaging features (lesion morphology, enhancement pattern, location distribution, adjacent organ invasion, lymph node involvement, distant metastasis). Comprehensive feature representation enables precise patient matching, with a 35% improvement in matching accuracy compared to traditional single-feature-based matching methods. Multi-dimensional features ensure increased personalization, with each patient finding truly similar cases for reference. Meanwhile, feature selection based on medical knowledge ensures clinical relevance, avoiding irrelevant factors and making matching results more valuable. b) A high-efficiency indexing method combining local sensitive hashing (LSH) and vector databases was adopted to achieve millisecond-level similar case retrieval. The system uses a two-layer retrieval architecture to achieve millisecond-level search. The first layer uses local sensitive hashing (LSH) for fast pre-screening, mapping similar patients to the same or similar hash buckets through 32 hash tables and 128 hash functions. Each patient's 30-dimensional feature vector is converted into multiple hash signatures, and similar patients' signatures fall into the same bucket. The second layer uses a vector database (such as Faiss) for accurate similarity calculation. After LSH quickly narrows down the candidate range, the system only needs to perform accurate cosine similarity or Euclidean distance calculations on the candidate set, rather than a brute-force search of the entire database. This hierarchical strategy reduces search time from seconds to milliseconds. The system also implements Hamming distance proximity search, searching for patients with a Hamming distance of 1 even if the query vector's hash signature does not completely match a bucket, ensuring that high-similarity patients are not missed. c) A similarity weight adjustment mechanism with medical expert verification was introduced to strengthen the importance of key features in matching. The system established an expert-verified feature importance weight library, determining the base weight of each feature through multiple rounds of expert questionnaires and clinical verification. For example, in lung cancer patient matching, TNM stage and tumor size received a high weight of 0.95, EGFR mutation status received a weight of 0.85, and geographic region received a low weight of 0.20. The system dynamically adjusts weights based on disease type, increasing the weight of smoking history for lung cancer patients and increasing the weight of hormone receptor status for breast cancer patients.Meanwhile, the system analyzes the distribution variance of feature values. If a certain feature has little variation among candidate patients, its weight will be automatically reduced to avoid ineffective matching. The system also establishes a learning feedback mechanism to collect doctors' evaluations of matching results and continuously optimizes weight configurations through machine learning algorithms. When doctors believe that some recommended cases are not similar enough, the system will automatically adjust the weight of the relevant features to improve the accuracy of subsequent matching.

[0049] Personalized treatment plan generation system: a) Developed "treatment decision tree generation algorithm" set in the result integration layer in parallel with the big data disease library, according to the treatment effect statistics of similar cases, to generate the potential effect and risk probability of each treatment plan. Based on the treatment path of similar cases, generate decision tree nodes (such as "first-line treatment plan: A / B"), each node attached with effect statistics (such as 5-year survival rate of plan A). b) Adopt multi-objective optimization algorithm to balance in multiple dimensions such as treatment effect, risk control, quality of life, economic cost, etc. Use NSGA-II algorithm to balance multiple objectives to generate a set of Pareto optimal solutions, finally provide 3-5 optimized plans with different focuses for doctors to choose from. c) Introduce medical ethics constraint module to ensure that the generated treatment recommendations comply with medical ethics. The system has a comprehensive medical ethics rule library, including: non-maleficence principle (avoid recommending harmful treatments for patients), beneficence principle (prefer to recommend clear benefit plans), autonomy principle (respect for patient choice), justice principle (ensure fair distribution of resources). The system will conduct ethical review on each treatment recommendation, for example: for patients with advanced disease, do not recommend highly toxic but limited benefit treatment; for young patients, prefer to consider treatment plans that protect reproductive function; for economically disadvantaged patients, recommend treatment options with higher cost-effectiveness. The system will also check whether the treatment plan meets the requirements of relevant medical guidelines and regulations to ensure that all recommendations are within the scope of compliance. When potential ethical conflicts are found, the system will mark a warning and prompt the doctor to pay attention.

[0050] Doctor-AI collaborative decision support mechanism: a) A three-level evidence display mechanism is designed, and doctors can trace the treatment recommendations back to supporting evidence and original cases. The result integration layer is set in the user interface as the key interface for doctor-AI collaboration. The first level displays the final treatment recommendation and expected effect; the second level displays the list of similar cases supporting the recommendation after clicking to expand, including case quantity, success rate statistics and evidence level; the third level can view the details of specific original cases, including patient characteristics, treatment process and follow-up results. This hierarchical display ensures the simplicity of the interface while meeting the needs of doctors to view detailed evidence, enhancing the transparency and traceability of AI decision-making. b) Develop an interactive treatment plan adjustment interface, doctors can modify AI recommendations based on their own experience. c) Establish a learning feedback loop, the system will continuously optimize the treatment recommendation algorithm based on the doctor's adoption. The system establishes a complete learning feedback loop, including data collection, analysis and processing, and model updating. At the data collection level, the system records the doctor's adoption of AI recommendations, modification content and final treatment results for updating the similar case library and optimizing the decision tree, forming a closed-loop learning. The system uses reinforcement learning algorithm to continuously optimize the treatment recommendation. When doctors frequently modify certain recommendations, the system will identify patterns and adjust the corresponding decision rules. When the actual effect of some treatment plans does not match the prediction, the system will update the parameters of the prediction model. The system also regularly analyzes successful and failed cases, identifies patterns of accurate and inaccurate predictions, and continuously improves the similarity matching algorithm and treatment effect prediction model. This self-learning ability enables the system to continuously adapt to changes in clinical practice, providing more and more accurate treatment recommendations.

[0051] Take CT chest radiograph disease screening as an example: A patient with unexplained chest pain seeks medical treatment, and the attending physician issues a CT chest radiograph examination order for him.

[0052] After the radiologist obtains the CT image results of the patient's skull base to chest segment, the dispatch model 1 is enabled to analyze the CT image.

[0053] The dispatch model selects all AI software for any part of the skull base to the chest segment as the target object by screening the AI auxiliary diagnosis models 2 under it, and according to the input required by each software, it performs specific image splitting to split the CT image into lung image, trachea image, esophagus image, etc. It calls the target diagnosis software for each split image, such as pneumonia diagnosis, lung nodule detection software for the lung, esophageal abnormality detection software for the esophagus, to obtain the screening results for corresponding diseases, such as small cell lung cancer, esophageal wall thickening with mass.

[0054] The scheduling model selects an in-built large language model (LLM) to summarize and analyze the screening results of the three diseases, generating a structured report (including image findings and diagnostic opinions). It also labels the severity of the disease according to its impact and gives a warning prompt for the most dangerous disease.

[0055] The LLM first receives the structured detection results output by each AI software: { "Pulmonary Nodule Detection": { "Finding": "Right upper lobe nodule", "Size": "1.2cm × 0.8cm", "Location": "Right upper lobe lateral segment", "Morphology": "Irregular, with spiculated edges", "Density": "Solid", "Malignancy Probability": 0.87, "Confidence": 0.92 }, "Esophageal Abnormality Detection": { "Finding": "Esophageal mid-segment wall thickening", "Thickness": "8mm", "Range": "About 4cm in length", "Features": "Irregular thickening, lumen stenosis", "Suspected Nature": "Possibly malignant", "Confidence": 0.83 }, "Tracheal Examination": { "Finding": "Normal", "Confidence": 0.96 } } The LLM uses a medical knowledge base for analysis: · Correlation analysis: Identify possible correlations between pulmonary nodules and esophageal lesions (such as lung metastasis or primary double cancer) · Priority ranking: Sort findings according to malignancy probability and clinical urgency · Medical logic verification: Check for any medical logic contradictions between findings · Comprehensive diagnostic reasoning: Analyze the clinical significance of each finding in combination with patient symptoms (chest pain) The LLM generates a report in the standard radiology report format:

Image Findings

Diagnosis Opinion

[0056] Read the patient's past medical history in Big Data Disease Library 3, and analyze the current examination results to obtain the current disease development (lung cancer detection time, cancer doubling time, etc.), and future progression (survival period of 3 months, if the esophageal cancer stage enters the next stage, the survival period is halved, etc.).

[0057] Time series comparative analysis: Assume the patient's medical history shows: 6 months ago, chest CT: right lung upper lobe nodule 0.6 cm; 1 year ago, gastroscopy: mild esophagitis.

[0058] Analysis process: 1) Lesion growth rate calculation: nodule volume change: 0.6 cm → 1.2 cm (6 months); doubling time calculation: about 3 months (consistent with malignant characteristics); growth pattern: rapid growth, highly suggestive of malignancy; 2) Disease progression evaluation: esophagus from inflammation to occupying lesion; time span: significant change within 1 year; progression rate: suggests possible invasive lesion.

[0059] Prognosis prediction analysis (based on time series evolution and medical knowledge base): 1) Lung cancer prognosis: according to doubling time and morphological characteristics, it is predicted to be small cell lung cancer; combined with TNM staging prediction: if it is T1N0M0 stage, 5-year survival rate is about 60%; if metastasis is found, median survival is about 8-12 months.

[0060] 2) Esophageal cancer prognosis: tube wall thickness 8mm suggests T3 stage possible; if diagnosed as esophageal cancer T3 stage, 5-year survival rate is about 30%; if it progresses to T4 stage, survival period will be significantly shortened to 6-9 months.

[0061] The scheduling model matches similar cases in the built-in Big Data Disease Library 3, and gives the corresponding treatment plan (radiotherapy, chemotherapy, and medication, patient's clinical response, etc.) (only visible to doctors).

[0062] 1. Patient feature matching: a. Basic information matching: age (+5 years), gender (same gender preferred), smoking history (number of years and number of cigarettes per day) b. Disease feature matching: Lung nodule size (+2mm range), nodule location (same lobe or adjacent lobe), morphological features (spiculation, pleural indentation, etc.), esophageal lesion location (upper / middle / lower segment matching), lesion extent (+1 length range) c. Pathological feature matching: Histological type (adenocarcinoma, squamous cell carcinoma, small cell carcinoma, etc.), differentiation degree (high, medium, low differentiation), molecular markers (EGFR, ALK, PD-L1 expression status) 2. Matching algorithm process: a. Similarity calculation = (basic information similarity × 0.2 + image feature similarity × 0.3 + pathological feature similarity × 0.3 + treatment response similarity × 0.2) b. Select cases with similarity>0.8 as reference 3. Treatment recommendation results (based on matched similar cases): a. Recommended treatment plan A (best match 85% similarity): i. Treatment strategy: concurrent chemoradiotherapy ii. Chemotherapy regimen: carboplatin + etoposide × 4 cycles iii. Radiotherapy dose: 60Gy / 30 times iv. Patient response: complete remission rate 65%, partial remission rate 25% v. Adverse reactions: 3-4 grade hematological toxicity 30% vi. Median survival time: 14.2 months b. Recommended treatment plan B (suboptimal match 82% similarity): i. Treatment strategy: surgery + adjuvant chemotherapy ii. Surgical method: right upper lobe resection + lymph node dissection iii. Adjuvant chemotherapy: cisplatin + vinorelbine × 4 cycles iv. Patient response: 5-year survival rate 68% v. Postoperative complications: lung leak incidence 8% c. Risk prompt: i. If treatment is delayed for more than 4 weeks, the risk of disease progression increases by 40% ii. The possibility of double primary cancer needs to be verified by gene detection iii. It is recommended to have a multidisciplinary team consultation to develop an individualized plan.

[0063] The disease screening system proposed in the present application can be applied to all types of medical images, including but not limited to MR, CT, etc.

[0064] The above merely provides the preferred embodiment of the present application, and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A disease screening system based on a large model, characterized in that: The system includes a scheduling model (1) and multiple AI-assisted diagnosis models (2). The scheduling model (1) is connected to the multiple AI-assisted diagnosis models (2) respectively, and the scheduling model (1) is connected to the big data disease library (3).

2. The disease screening system based on a large model according to claim 1, characterized in that: The scheduling model (1) adopts a three-layer scheduling architecture, including an image distribution layer, a model selection layer, and a result integration layer. The image distribution layer is connected to the model selection layer, and the model selection layer and the result integration layer are both connected to the AI-assisted diagnosis model (2); the result integration layer is connected to the big data disease library (3) to read the current patient's past medical history and matching cases with high similarity in the big data disease library (3).

3. The disease screening system based on a large model according to claim 2, characterized in that: The image distribution layer receives input medical image data, pre-processes the medical images using DICOM parsing and image enhancement, and determines whether the image quality meets the diagnostic requirements; decomposes the medical image sequence into single slices, and identifies the anatomical regions using the human anatomy positioning algorithm; performs corresponding window width and window position adjustment, image enhancement, and size normalization pre-processing on the identified anatomical regions; and decides to distribute the corresponding image data to the AI-assisted diagnosis model based on the image content; The image distribution layer uses an intelligent decision tree scheduling algorithm to automatically construct the optimal calling sequence based on the examination site, sequence, and window width and window position parameters in the DICOM metadata in the image; the intelligent decision tree scheduling algorithm performs serial / parallel screening for common possible diseases based on the examination, and adjusts the execution order of the AI-assisted diagnosis model based on the diagnosis results. If other lesions are suspected and there are other images or examination data, other examination results are automatically retrieved for joint judgment.

4. The large model-based disease screening system according to claim 2 or 3, characterized in that: The model selection layer intelligently selects the AI-assisted diagnosis model (2) to be called based on the anatomical region identified by the image distribution layer; coordinates the parallel / serial execution order of multiple AI-assisted diagnosis models (2); sets appropriate detection parameters and thresholds for different AI-assisted diagnosis models (2); and manages the allocation of computing resources.

5. The disease screening system based on a large model according to claim 4, characterized in that: The result integration layer collects the test results output by each AI-assisted diagnosis model (2); processes conflicts among different judgments of the same region by multiple AI-assisted diagnosis models (2); combines with the big data disease database (3) to make comprehensive judgments and risk assessments based on the test results and formulate treatment reference plans; automatically generates a structured comprehensive diagnosis report; and uses a large language model to perform natural language processing and report optimization.

6. The large model-based disease screening system according to any one of claims 1, 2, 3, and 5, characterized in that: The various disease screening models in the AI-assisted diagnosis model include but are not limited to lung nodule detection and benign and malignant analysis models, stroke detection lesion segmentation and analysis models, fracture detection models, and organ-specific disease models of tumor screening models; Based on the principles of anatomy and pathology, the AI-assisted diagnosis model is divided into: organ localization model, lesion detection model, quantitative analysis model and benign and malignant prediction model. The organ localization model is connected to the lesion detection model, the lesion detection model is connected to the quantitative analysis model, and the quantitative analysis model is connected to the benign and malignant prediction model; the organ localization model is implemented through the uAI All-in-One model; the lesion detection model includes more than 20 special disease detection models for lung nodule detection, liver lesion detection, and brain abnormality detection; the quantitative analysis model measures the quantitative parameters of the volume, density, and boundary of the lesions; the benign and malignant prediction model predicts the benign and malignant nature of the detected lesions.

7. The disease screening system based on a large model according to claim 6, characterized in that: The intelligent decision tree scheduling algorithm includes but is not limited to at least one of FCFS, SPT, LPT, EDD or CR; A dynamic resource allocation strategy is used to allocate GPU / CPU resources based on model complexity and priority, ensuring that key models are executed first. This dynamic resource allocation strategy uses a hybrid scheduling strategy, with EDD used for emergency queues and SPT used for routine queues; CR used for complex examinations; and batch processing used CR. Based on the current system load and task type, algorithms are dynamically selected, priority weights are set for different disease types, and the urgency of the patient's condition is analyzed in conjunction with a large prediction model to intelligently adjust the hybrid scheduling strategy. The large language model is trained on 10 million radiology reports and 5,000 medical textbooks. It receives complete patient information, including current imaging results, medical history, and laboratory test data. Through the large language model enhanced by medical knowledge, it conducts in-depth analysis, matching patient information with the "disease-symptom-treatment-effect" four-element relationship network in the knowledge graph to identify the patient's disease pattern and characteristics. Combined with the evidence-based medicine grading standards, it evaluates the strength of evidence for different diagnostic and treatment options and generates a comprehensive report that includes diagnostic analysis, condition assessment, treatment recommendations, and prognosis prediction. The big data disease database (3) is a hierarchical treatment knowledge base, and its construction method is as follows: a) Bottom-level data source: integrating 30 million de-identified electronic medical records, 5 million follow-up records and 100,000 clinical trial data; b) Middle-level knowledge representation: using knowledge graph technology to construct a "disease-symptom-treatment-effect" four-element relationship network to obtain an expert knowledge graph; c) Top-level decision support: hierarchical labeling of treatment plans based on the level of evidence in evidence-based medicine.

8. The large model-based disease screening system according to claim 1 or 7, characterized in that: When multiple models reach different conclusions about the same region, a combination of confidence weighting and expert rules is used to resolve conflicts. Confidence weighting means that each AI-assisted diagnosis model outputs a confidence score when outputting a diagnosis, indicating the degree of certainty of the AI-assisted diagnosis model in the diagnosis. Based on the algorithm's performance, different weights are assigned to the model in different scenarios, including weights for anatomical location, disease type, and image quality. The expert rules are logical rules formulated based on medical knowledge and clinical experience, including diagnostic criteria in medical literature, clinical expert experience, and clinical pathways; The implementation method of the combination of confidence weighting and expert rules is as follows: input: diagnosis results of multiple AI-assisted diagnosis models → first layer: expert rule pre-screening, removing medically impossible results, marking high-risk features → second layer: confidence weighted calculation, weighted voting on the remaining results, generating preliminary diagnosis probability → third layer: expert rule post-processing, adjusting the final probability according to the rules, adding diagnostic basis and suggestions → output: final diagnosis result + confidence + basis.

9. The disease screening system based on a large model according to claim 5, characterized in that: The large language model enhanced with medical knowledge is set in the result integration layer. Based on the general large language model architecture, it is fine-tuned in different fields using more than 10 million radiology reports and 5,000 medical textbooks; Fine-tuning is a phased fine-tuning process, implemented as follows: Phase 1: medical knowledge injection, using medical textbooks for knowledge pre-training; Phase 2: report generation fine-tuning, using radiology reports for supervised fine-tuning; Phase 3: instruction fine-tuning optimization, fine-tuning instructions for different tasks; The large language model includes but is not limited to at least one of GPT-3, GPT3.5, GPT4, XLNet, RoBERTa, and ALBERT.

10. The large model-based disease screening system according to claim 7 or 9, characterized in that: The result integration layer is equipped with a multimodal fusion processing algorithm and a "image-text-medical history" trimodal fusion algorithm. It integrates the detection results of the AI-assisted diagnosis model, the key features of the original image data, and the patient's medical history through a self-attention mechanism to generate a comprehensive understanding. The image features are deep feature vectors extracted by CNN, the text data is the structured JSON of the AI ​​detection results converted into feature vectors, and the medical history data is converted into embedding vectors by a medical record encoder. The features of the three modalities are projected into the same feature space, and the projected feature vectors are spliced. A multi-head attention module is used to perform weighted fusion of multimodal data features. The result integration layer is equipped with a disease progression assessment algorithm based on time series analysis. It compares and analyzes the results of multiple consecutive examinations. Combined with the current multimodal fusion results, it uses an LSTM network structure with a medical knowledge module gate switch to model the disease progression trend. In combination with the expert knowledge graph, it constrains the disease development pattern. The modeling includes: feature alignment, aligning the same anatomical structures of previous examinations; change detection: calculating the rate of change of key indicators; LSTM modeling: inputting the feature sequence of historical time points; A medical-specific uncertainty quantification mechanism is introduced at the result integration layer as the final evaluation module for all prediction results. The degree of certainty of diagnosis and prediction is expressed in standardized language based on the strength of evidence, the number of similar cases, and the confidence level of the prediction. A similar case retrieval system based on patient feature vectors has been developed, comprehensively considering more than 30 dimensions of patient feature vectors, including demographic characteristics, pathological indicators, genotypes, and imaging features. A highly efficient indexing method combining local sensitive hashing and a vector database is used to achieve millisecond-level similar case retrieval. The first layer uses local sensitive hashing for rapid pre-screening, mapping similar patients to the same or similar hash buckets via 32 hash tables and 128 hash functions. The second layer uses a vector database for precise similarity calculations. A similarity weight adjustment mechanism validated by medical experts has been introduced to strengthen the matching importance of key features. A feature importance weight library validated by experts has been established, and the basic weight of each feature has been determined through multiple rounds of expert questionnaires and clinical validation. The results integration layer develops a treatment decision tree generation algorithm that generates the potential effects and risk probabilities of each treatment plan based on the treatment effect statistics of similar cases. Based on the treatment pathways of similar cases, decision tree nodes are generated, each with effect statistics attached. A multi-objective optimization algorithm is used to balance the multiple dimensions of treatment effect, risk control, quality of life, and economic cost. A comprehensive medical ethics rule library is built in, including the principles of non-maleficence, beneficence, autonomy, and fairness, to conduct ethical review of each treatment recommendation. The result integration layer is designed with a three-level evidence display mechanism. The first level displays the final treatment recommendation and expected effect; the second level, when clicked, displays a list of similar cases supporting the recommendation, including the number of cases, success rate statistics and evidence level; the third level views the specific original case details, including patient characteristics, treatment process and follow-up results; develops an interactive adjustment interface for the treatment plan; and establishes a learning feedback loop to continuously optimize the treatment recommendation algorithm based on the doctor's adoption.

Images