A multi-part recognition method and intelligent agent system based on RAG optimization of large medical models

Abstract

The present invention discloses a method and intelligent system for multi-part identification of medical examinations based on RAG optimization of a large medical model, wherein the method comprises: receiving first medical examination data to be collected from a first medical institution; determining whether there is a first identification parameter corresponding to the first medical institution; if so, performing data preprocessing and part feature extraction on the first medical examination data based on the first identification parameter to obtain a part identification feature; if not, performing data preprocessing and part feature extraction on the first medical examination data based on a standard identification parameter to obtain a part identification feature; performing a fusion analysis on the part identification text feature and the part identification image feature to obtain a fusion feature; and outputting data-standardized part identification data corresponding to the first medical examination data based on the fusion feature. The present invention can accurately identify the input medical data and output standard and standardized identification data to achieve the purpose of standardizing medical data.

Description

Technical Field

[0001] The present invention relates to the field of multi-part identification during inspection, and in particular to a multi-part identification method and intelligent system based on RAG optimization of a large medical model. Background Art

[0002] Medical site examination has evolved from empirical judgment to scientific and precise methods, and has now become an indispensable diagnostic and treatment tool in clinical practice. Its development can be traced back to ancient intuitive diagnosis. Later, breakthroughs in microscopy, automated equipment, and molecular biology technologies have gradually led to the segmentation and intelligentization of testing items. The widespread adoption of technologies such as fully automated analyzers and genetic testing has driven the rise of personalized medicine and remote testing. Medical site examination is widely used for disease diagnosis (such as CT diagnosis of early lung cancer), disease assessment (cardiac ultrasound to observe cardiac structure), screening and prevention (low-dose CT lung cancer screening), treatment monitoring (real-time imaging of interventional treatments), and disease staging (CT / MRI tumor staging), providing comprehensive clinical support.

[0003] In existing technologies, each medical institution's information system uses its own standards for system development and design, resulting in uneven data quality. As a result, the imaging cloud platforms in various cities are unable to accurately identify the number and name of parts according to the examination site, causing difficulties in imaging data management, the inability to innovate project operation models, and the difficulty in carrying out cross-institutional imaging examination sharing and mutual recognition. Summary of the Invention

[0004] In view of some of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is to provide a multi-part inspection recognition method and intelligent system based on RAG optimization of the medical large model, which aims to accurately identify the input medical data and output standard and standardized recognition data to achieve the purpose of standardizing and normalizing medical data.

[0005] To achieve the above objectives, the present invention discloses, in a first aspect, a method for multi-site identification based on RAG optimization of a large medical model. The method is used for the collection and management of medical data from medical institutions across different imaging cloud platforms. The method comprises:

[0006] Step S1: receiving first medical examination data to be collected from a first medical institution; wherein the first medical examination data is multimodal data including at least corresponding text data and image data; the first medical institution is one of multiple medical institutions on different imaging cloud platforms to which data is to be collected;

[0007] Step S2: Determine whether there is a first identification parameter corresponding to the first medical institution. If so, perform data preprocessing and part feature extraction on the first medical examination data based on the first identification parameter to obtain the part identification feature of the first medical examination data; if not, perform data preprocessing and part feature extraction on the first medical examination data based on the standard identification parameter to obtain the part identification feature of the first medical examination data; wherein, the first identification parameter is obtained by pre-training the medical examination data set of the first medical institution, and the identification parameters obtained by training different medical institutions are different. The first identification parameter is used to improve the model recognition accuracy according to the differences in writing standards and inspection equipment of different medical institutions; the standard identification parameter is obtained by mixed training of a general medical examination data set; the part identification feature includes part identification text feature and part identification image feature;

[0008] Step S3: performing a fusion analysis on the part identification text feature and the part identification image feature based on the first recognition parameter or the standard recognition parameter to obtain a similarity between the part identification text feature and the part identification image feature, and further obtain a first confidence level between the part identification text feature and the part identification image feature;

[0009] Step S4: in response to the first confidence level being greater than a standard confidence level, obtaining a fusion feature, inputting the fusion feature into a RAG industry model analysis system, causing the RAG industry model analysis system to output data-standardized part identification data corresponding to the first medical examination data;

[0010] After step S1, the method further includes step S1A:

[0011] Obtaining first patient information corresponding to the first medical examination data; obtaining historical medical examination data of the first patient based on the first patient information;

[0012] In response to the existence of second medical examination data in the historical medical examination data, the second medical examination data is obtained in response to the existence of second medical examination data whose time difference with the first medical examination data does not exceed a first threshold and whose examination site is the same as that of the first medical examination data;

[0013] In response to the fact that the level of the second medical institution corresponding to the second medical examination data is greater than the level of the first medical institution, the first medical examination correction data corresponding to the first medical examination data and complying with the data standardization specifications are output according to the writing specifications of the second medical examination data and input into step S2 for processing as the first medical examination data.

[0014] Optionally, the step S1A further includes:

[0015] Acquire the first medical examination data and the second medical examination data having a time difference of no more than a first threshold and the same examination site as the first medical examination data, and construct a data pair set of the first medical examination data and the second medical examination data;

[0016] Based on the image data in the data pair set, an image data correction model of the correspondence between the first medical examination data and the second medical examination data is trained; the image data correction model is used to autonomously generate second mirror medical examination data corresponding to the first medical examination data of the first medical institution and the writing standards of the second medical institution, and is used for aggregation and as the first medical examination data to execute subsequent steps.

[0017] Optionally, in response to the existence of second medical examination data in the historical medical examination data, which has a time difference with the first medical examination data that does not exceed a first threshold and has the same examination site as the first medical examination data, before obtaining the second medical examination data, the method includes:

[0018] Extracting features corresponding to the first medical examination data and the historical medical examination data respectively, and obtaining examination parts corresponding to the first medical examination data and the historical medical examination data;

[0019] According to the examination parts corresponding to the first medical examination data and the historical medical examination data, second medical examination data having a time difference with the first medical examination data that does not exceed a first threshold and having the same examination part as the first medical examination data is obtained.

[0020] Optionally, in step S2, the method further includes:

[0021] In response to the absence of the first identification parameter corresponding to the first medical institution, extracting common features from the medical examination data set from the first medical institution; wherein the common features include at least text word usage features, text format features, examination equipment model features, and image clarity features;

[0022] Training is performed based on the common features to obtain the first recognition parameters.

[0023] Optionally, in step S1 , the data of multiple modalities of the first medical examination data are transmitted separately or in mixed collaborative transmission.

[0024] Optionally, the data preprocessing of the first medical examination data in step S2 includes:

[0025] Noise filtering is performed on the text data in the first medical examination data; and a corresponding processing method is selected according to the type of image data in the first medical examination data.

[0026] Optionally, step S3 includes:

[0027] According to the first recognition parameter or the standard recognition parameter, the recognition is performed through a cross-attention mechanism;

[0028] A confidence weighting model is constructed according to the DS evidence theory, and the first confidence between the part identification text feature and the part identification image feature is obtained through the confidence weighting model.

[0029] The second aspect of the present invention discloses a multi-site inspection intelligent agent system based on RAG optimization of a large medical model. The intelligent agent system is used for the collection and management of medical data of medical institutions across different imaging cloud platforms. The intelligent agent system includes: a multimodal input and output layer, an intelligent analysis engine layer, and a RAG industry model analysis system;

[0030] The multimodal input and output layer is used to receive first medical examination data to be collected from a first medical institution, or output part identification data in accordance with the data standardization specification of the RAG industry model analysis system; wherein the first medical examination data is multimodal data including at least corresponding text data and image data; the first medical institution is one of the medical institutions on multiple different imaging cloud platforms whose data is to be collected;

[0031] The intelligent analysis engine layer includes a feature extraction module and a multimodal decision fusion module;

[0032] The feature extraction module is used to determine whether there is a first identification parameter corresponding to the first medical institution. If so, data preprocessing and part feature extraction are performed on the first medical examination data based on the first identification parameter to obtain the part identification feature of the first medical examination data; if not, data preprocessing and part feature extraction are performed on the first medical examination data based on the standard identification parameter to obtain the part identification feature of the first medical examination data; wherein, the first identification parameter is obtained by pre-training the medical examination data set of the first medical institution, and the identification parameters obtained by training are different for different medical institutions. The first identification parameter is used to improve the model recognition accuracy according to the differences in writing standards and inspection equipment of different medical institutions; the standard identification parameter is obtained by mixed training of a general medical examination data set; the part identification feature includes part identification text feature and part identification image feature;

[0033] The multimodal decision fusion module is configured to perform a fusion analysis on the part identification text feature and the part identification image feature based on the first recognition parameter or the standard recognition parameter to obtain a similarity between the part identification text feature and the part identification image feature, and further obtain a first confidence level between the part identification text feature and the part identification image feature; in response to the first confidence level being greater than the standard confidence level, a fusion feature is obtained, and the fusion feature is input into the RAG industry model analysis system;

[0034] The RAG industry model analysis system is configured to output, to the multimodal input and output layer, data-standardized part identification data corresponding to the first medical examination data based on the fusion features;

[0035] The intelligent system further comprises: a historical data acquisition module, a same inspection data acquisition module, and a standard data rapid acquisition module;

[0036] The historical data acquisition module is configured to obtain first patient information corresponding to the first medical examination data; and obtain the historical medical examination data of the first patient based on the first patient information;

[0037] The same examination data acquisition module is configured to obtain the second medical examination data in response to the presence of second medical examination data in the historical medical examination data, the second medical examination data having a time difference with the first medical examination data that does not exceed a first threshold and has the same examination site;

[0038] The standard data rapid acquisition module is used to respond to the fact that the level of the second medical institution corresponding to the second medical examination data is greater than the level of the first medical institution. Then, according to the writing specifications of the second medical examination data, the first medical examination correction data corresponding to the first medical examination data that complies with the data standardization specifications is output and input into the feature extraction module for processing as the first medical examination data.

[0039] Optionally, the agent system further comprises: a data pair set construction module and a correction model training module;

[0040] The data pair set construction module is configured to obtain the first medical examination data and the second medical examination data having a time difference of no more than a first threshold and the same examination site as the first medical examination data, and to construct a data pair set of the first medical examination data and the second medical examination data;

[0041] The correction model training module is used to train an image data correction model of the correspondence between the first medical examination data and the second medical examination data based on the image data in the data pair set; the image data correction model is used to autonomously generate second mirror medical examination data corresponding to the first medical examination data of the first medical institution and the writing standards of the second medical institution, and is used for aggregation and as the first medical examination data to execute subsequent steps.

[0042] Optionally, the intelligent body system further comprises: an examination part obtaining module and a second medical examination data obtaining module;

[0043] The examination part acquisition module is used to extract features corresponding to the first medical examination data and the historical medical examination data, respectively, to obtain examination parts corresponding to the first medical examination data and the historical medical examination data;

[0044] The second medical examination data acquisition module is used to obtain second medical examination data with a time difference of no more than a first threshold and the same examination site as the first medical examination data based on the examination sites corresponding to the first medical examination data and the historical medical examination data.

[0045] Optionally, the agent system further includes a first recognition parameter obtaining module, wherein the first recognition parameter obtaining module is specifically configured to:

[0046] In response to the absence of the first identification parameter corresponding to the first medical institution, extracting common features from the medical examination data set from the first medical institution; wherein the common features include at least text word usage features, text format features, examination equipment model features, and image clarity features;

[0047] Training is performed based on the common features to obtain the first recognition parameters.

[0048] Optionally, the data of multiple modalities of the first medical examination data are transmitted separately or in mixed and coordinated manner.

[0049] The beneficial effects of the present invention are as follows: 1. The intelligent agent system constructed by artificial intelligence can accurately identify and examine multiple body part data from non-standardized body part description text, DICOM files, and image images, thereby solving the problem of "most medical institutions' information systems being poorly designed and data quality being uneven," and achieving the purpose of accurately identifying input medical data and outputting standardized and regulated body part identification data. 2. The present invention can perform targeted feature extraction on the medical examination data to be collected using corresponding recognition parameters. Because the recognition parameters are pre-trained using the medical examination data sets of the corresponding medical institutions, the recognition parameters can improve the model recognition accuracy based on the writing standards and examination equipment differences of the corresponding medical institutions, thereby improving the recognition accuracy of the medical examination data to be collected. 3. The present invention can also standardize and revise the medical examination data of lower-level medical institutions based on the medical examination data of higher-level medical institutions for the same patients in the same time period and at the same body part, thereby making the medical examination data of the lower-level medical institutions more standardized and regulated, thereby improving the accuracy of subsequent recognition of the medical examination data of the lower-level medical institutions.

[0050] In summary, the present invention can accurately identify input medical data and output standard and standardized identification data, so as to achieve the purpose of standardizing and normalizing medical data. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flowchart of a method for multi-site identification based on RAG optimization of a large medical model provided by a specific embodiment of the present invention;

[0052] Figure 2 This is a structural diagram of an intelligent system for multi-part identification based on RAG optimization of a large medical model provided by a specific embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of a processing flow of an intelligent system for multi-part identification based on RAG optimization of a large medical model provided by a specific embodiment of the present invention;

[0054] Figure 4 This is a client screenshot of an intelligent inspection system for multi-part recognition based on RAG optimization of a large medical model provided by a specific embodiment of the present invention. DETAILED DESCRIPTION

[0055] The present invention discloses a method and intelligent system for multi-part identification based on RAG optimization of a large medical model. Those skilled in the art can refer to the content of this article and appropriately improve the technical details. It should be pointed out in particular that all similar replacements and modifications are obvious to those skilled in the art and are considered to be included in the present invention. The method and application of the present invention have been described through preferred embodiments. Relevant personnel can obviously modify or appropriately change and combine the methods and applications described herein without departing from the content, spirit and scope of the present invention to implement and apply the technology of the present invention.

[0056] The imaging cloud platform is unable to accurately identify the number and name of each part examined, leading to difficulties in imaging data management, a lack of innovation in project operation models, and difficulty in implementing cross-institutional imaging sharing and mutual recognition. Artificial intelligence can be used to standardize chaotic data, but existing AI recognition technology cannot accurately identify specific data from individual medical institutions, resulting in low recognition accuracy.

[0057] Therefore, the present invention provides a multi-site identification method based on RAG optimization of a large medical model, which is used for the collection and management of medical data of medical institutions on different imaging cloud platforms, such as Figure 1 As shown, the method includes:

[0058] Step S1: Receive first medical examination data to be collected from a first medical institution.

[0059] The first medical examination data is multimodal data including at least corresponding text data and image data. The first medical institution is a medical institution to be data collected among multiple medical institutions on different image cloud platforms.

[0060] In this specific embodiment, in step S1 , data of multiple modalities of the first medical examination data are transmitted separately or in mixed and coordinated manner.

[0061] It should be noted that the input channels mainly include the following: text processing channel: supports free text input (including electronic medical record text, examination application form, etc.); image processing channel: compatible with DICOM / JPEG / PNG format image input; mixed input mode: allows collaborative text input of images to improve recognition confidence.

[0062] After step S1, the method further includes step S1A:

[0063] Obtaining first patient information corresponding to the first medical examination data; obtaining historical medical examination data of the first patient based on the first patient information;

[0064] In response to the existence of second medical examination data in the historical medical examination data, the time difference between the first medical examination data and the first medical examination data and the same examination site as the first medical examination data does not exceed a first threshold, obtaining the second medical examination data;

[0065] In response to the fact that the level of the second medical institution corresponding to the second medical examination data is greater than the level of the first medical institution, the first medical examination correction data corresponding to the first medical examination data and conforming to the data standardization specifications are output according to the writing specifications of the second medical examination data and input into step S2 for processing as the first medical examination data.

[0066] It should be noted that, through the processing of step S1A, the first medical examination data can be corrected according to the data specifications of a higher-level medical institution, so that the corrected first medical examination data is more standardized, thereby improving its recognition accuracy in subsequent steps.

[0067] In this specific embodiment, step S1A further includes:

[0068] Acquire first medical examination data and second medical examination data having a time difference of no more than a first threshold and the same examination site as the first medical examination data, and construct a data pair set of the first medical examination data and the second medical examination data;

[0069] Based on the image data in the data pair set, an image data correction model is trained to determine the correspondence between the first medical examination data and the second medical examination data; the image data correction model is used to autonomously generate second mirror medical examination data corresponding to the first medical examination data of the first medical institution and the writing standards of the second medical institution, and is used for aggregation and as the first medical examination data to execute subsequent steps.

[0070] It should be noted that through model training, a correspondence can be established between the medical examination data of low-level medical institutions and the medical examination data of high-level medical institutions, and then the medical examination data of low-level medical institutions can be corrected according to the relationship, so that the recognition accuracy of the medical examination data of low-level medical institutions in subsequent steps can be higher.

[0071] In this specific embodiment, in response to the presence of second medical examination data in the historical medical examination data, which has a time difference of no more than a first threshold with respect to the first medical examination data and has the same examination site as the first medical examination data, before obtaining the second medical examination data, the method includes:

[0072] Extracting features corresponding to the first medical examination data and the historical medical examination data respectively, and obtaining examination parts corresponding to the first medical examination data and the historical medical examination data;

[0073] According to the examination parts corresponding to the first medical examination data and the historical medical examination data, second medical examination data with a time difference of no more than a first threshold and the same examination part as the first medical examination data is obtained.

[0074] It should be noted that this embodiment is intended to ensure that the first medical examination data and the second medical examination data are examination data of the same part within the same time period, so as to avoid correction errors caused by mismatching, thereby affecting subsequent recognition.

[0075] Step S2: Determine whether there is a first identification parameter corresponding to the first medical institution. If so, perform data preprocessing and part feature extraction on the first medical examination data based on the first identification parameter to obtain the part identification feature of the first medical examination data; if not, perform data preprocessing and part feature extraction on the first medical examination data based on the standard identification parameter to obtain the part identification feature of the first medical examination data.

[0076] Among them, the first recognition parameter is obtained through pre-training of the medical examination data set of the first medical institution, and the recognition parameters obtained by training in different medical institutions are different. The first recognition parameter is used to improve the model recognition accuracy based on the writing standards and inspection equipment differences of different medical institutions; the standard recognition parameter is obtained by mixed training of the general medical examination data set; the part recognition features include part recognition text features and part recognition image features.

[0077] It should be noted that conventional medical examination data is typically identified based on standard recognition parameters, rarely specifically addressing differences in documentation and equipment between different medical institutions, resulting in low recognition accuracy. However, the present invention uses separate training to obtain recognition parameters corresponding to different medical institutions, and then identifies the corresponding medical data based on these recognition parameters. This improves model recognition accuracy by addressing differences in documentation standards and examination equipment across different medical institutions.

[0078] In this specific embodiment, in step S2, the method further includes:

[0079] In response to the absence of a first identification parameter corresponding to the first medical institution, extracting common features from the medical examination dataset from the first medical institution; wherein the common features include at least text word usage features, text format features, examination equipment model features, and image clarity features;

[0080] Training is performed based on the common features to obtain the first recognition parameters.

[0081] It should be noted that this embodiment can perform feature extraction training on medical data from medical institutions that do not have existing identification parameters, thereby obtaining corresponding identification parameters to improve the accuracy of subsequent medical data recognition. This embodiment obtains identification parameters through training on continuously accumulated medical data, thereby making the identification parameters more reliable.

[0082] In this specific embodiment, the data preprocessing of the first medical examination data in step S2 includes:

[0083] Noise filtering is performed on the text data in the first medical examination data; and a corresponding processing method is selected according to the type of image data in the first medical examination data.

[0084] The part recognition text feature extraction process includes the following:

[0085] Non-standardized text cleaning: A noise filtering model built using regular expressions + Bi-LSTM;

[0086] Part feature extraction: A domain-specific BERT model (MedBERT) built on the medical knowledge graph;

[0087] Contextual reasoning: Use the Transformer architecture to establish a mapping relationship between part descriptions and standard terminology.

[0088] Different parts of the image are identified using different methods to extract the following images:

[0089] Anatomical structure positioning: Improved U-Net network achieves three-dimensional spatial positioning (for CT / MRI);

[0090] Multi-scale feature fusion: Using the Inception-ResNet hybrid architecture to process images of different resolutions; Projection recognition subsystem: Integrating the CheXNet model to process X-ray two-dimensional projection recognition.

[0091] Step S3: Based on the first recognition parameter or the standard recognition parameter, the part recognition text feature and the part recognition image feature are fused and analyzed to obtain the similarity between the part recognition text feature and the part recognition image feature, and then obtain the first confidence between the part recognition text feature and the part recognition image feature.

[0092] In this specific embodiment, step S3 includes:

[0093] According to the first recognition parameter or the standard recognition parameter, it is realized by the cross attention mechanism;

[0094] A confidence weighted model is constructed based on the DS evidence theory, and the first confidence between the part recognition text features and the part recognition image features is obtained through the confidence weighted model.

[0095] It should be noted that feature fusion can further integrate feature information from different modalities to generate a more expressive and robust feature representation technology.

[0096] Step S4: In response to the first confidence being greater than the standard confidence, a fusion feature is obtained, and the fusion feature is input into the RAG industry model analysis system, so that the RAG industry model analysis system outputs data-standardized part identification data corresponding to the first medical examination data.

[0097] In this specific embodiment, the RAG industry model analysis system includes:

[0098] Standard terminology database: integrating domestic and international data standards and specifications;

[0099] Inspection item comparison table: Build a comparison table containing more than 100 medical institution inspection items and standard parts;

[0100] Incremental learning mechanism: continuously updating the recognition model through online learning modules;

[0101] Medical LLM model reasoning: Relying on a dynamic knowledge base to match body part features, the matching data is optimized for user and system prompts before entering the LLM model reasoning, and the reasoning results are finally output and returned to the user.

[0102] Based on the method of the above embodiment, the embodiment of the present invention also provides a multi-site inspection intelligent system based on the optimization of the medical large model RAG. The intelligent system is used for the collection and management of medical data of medical institutions between different imaging cloud platforms, such as Figure 2 The shown intelligent agent system includes: a multimodal input and output layer 201, an intelligent analysis engine layer 202, and a RAG industry model analysis system 203;

[0103] The multimodal input and output layer 201 is configured to receive first medical examination data to be aggregated from a first medical institution, or output part identification data in accordance with data standardization specifications of the RAG industry model analysis system 203; wherein the first medical examination data is multimodal data including at least corresponding text data and image data; and the first medical institution is one of multiple medical institutions on different imaging cloud platforms whose data is to be aggregated;

[0104] The intelligent analysis engine layer 202 includes a feature extraction module and a multimodal decision fusion module;

[0105] A feature extraction module is used to determine whether there is a first identification parameter corresponding to the first medical institution. If so, data preprocessing and part feature extraction are performed on the first medical examination data based on the first identification parameter to obtain the part identification feature of the first medical examination data; if not, data preprocessing and part feature extraction are performed on the first medical examination data based on the standard identification parameter to obtain the part identification feature of the first medical examination data; wherein, the first identification parameter is obtained by pre-training the medical examination data set of the first medical institution, and the identification parameters obtained by training are different for different medical institutions. The first identification parameter is used to improve the model recognition accuracy according to the writing standards and inspection equipment differences of different medical institutions; the standard identification parameter is obtained by mixed training of the general medical examination data set; the part identification feature includes part identification text feature and part identification image feature;

[0106] a multimodal decision fusion module configured to perform a fusion analysis on the part identification text features and the part identification image features based on the first recognition parameter or the standard recognition parameter, obtain a similarity between the part identification text features and the part identification image features, and further obtain a first confidence level between the part identification text features and the part identification image features; obtain a fusion feature in response to the first confidence level being greater than the standard confidence level, and input the fusion feature into the RAG industry model analysis system 203;

[0107] The RAG industry model analysis system 203 is configured to output, based on the fusion features, data-standardized part recognition data corresponding to the first medical examination data to the multimodal input and output layer 201;

[0108] The intelligent system also includes: historical data acquisition module, same inspection data acquisition module, standard data fast acquisition module;

[0109] A historical data acquisition module is used to obtain first patient information corresponding to the first medical examination data; and obtain historical medical examination data of the first patient based on the first patient information;

[0110] a same examination data acquisition module, configured to obtain second medical examination data in response to the presence of second medical examination data in the historical medical examination data, the second medical examination data having a time difference with the first medical examination data that does not exceed a first threshold and has the same examination site;

[0111] The standard data rapid acquisition module is used to respond to the fact that the level of the second medical institution corresponding to the second medical examination data is greater than the level of the first medical institution. Then, according to the writing specifications of the second medical examination data, the module outputs the first medical examination correction data that complies with the data standardization specifications corresponding to the first medical examination data and inputs it into the feature extraction module for processing as the first medical examination data.

[0112] Optionally, the agent system further includes: a data pair set construction module and a correction model training module;

[0113] a data pair set construction module, configured to obtain first medical examination data and second medical examination data having a time difference of no more than a first threshold and the same examination site as the first medical examination data, and to construct a data pair set of the first medical examination data and the second medical examination data;

[0114] The correction model training module is used to train an image data correction model of the correspondence between the first medical examination data and the second medical examination data based on the image data in the data pair set; the image data correction model is used to autonomously generate the first medical examination data of the first medical institution and the second mirror medical examination data corresponding to the writing standards of the second medical institution, and is used for aggregation and as the first medical examination data to execute subsequent steps.

[0115] Optionally, the intelligent body system further includes: an examination part obtaining module and a second medical examination data obtaining module;

[0116] An examination part acquisition module, configured to extract features corresponding to the first medical examination data and the historical medical examination data, respectively, and obtain examination parts corresponding to the first medical examination data and the historical medical examination data;

[0117] The second medical examination data acquisition module is used to obtain second medical examination data with a time difference of no more than a first threshold and the same examination site as the first medical examination data based on the examination sites corresponding to the first medical examination data and the historical medical examination data.

[0118] Optionally, the intelligent agent system further includes a first recognition parameter obtaining module, and the first recognition parameter obtaining module is specifically configured to:

[0119] In response to the absence of a first identification parameter corresponding to the first medical institution, extracting common features from the medical examination dataset from the first medical institution; wherein the common features include at least text word usage features, text format features, examination equipment model features, and image clarity features;

[0120] Training is performed based on the common features to obtain the first recognition parameters.

[0121] Optionally, data of multiple modalities of the first medical examination data are transmitted separately or in mixed and coordinated manner.

[0122] In this specific embodiment, the processing flow diagram of the intelligent agent system can be as follows: Figure 3 As shown, Figure 3In the process, the user end (i.e., the end of different medical institutions) transmits data (text data and impact data) to the multimodal input and output layer 201, and the multimodal input and output layer 201 transmits the data to the intelligent analysis engine layer 202. The intelligent analysis engine layer 202 performs feature extraction and feature fusion on the data to generate part features (fused features) and inputs them into the RAG industry model analysis system 203. The RAG industry model analysis system 203 enters the medical big voice model reasoning after optimizing the prompt words of the dynamic knowledge base, and finally outputs the reasoning results to the multimodal input and output layer 201 and returns them to the user.

[0123] In this specific embodiment, the user interface can be as follows Figure 4 As shown, by inputting text and images, standardized description text of the identified parts can be obtained.

[0124] The embodiment of the present invention uses an intelligent system composed of artificial intelligence to accurately identify and examine multiple body parts data from non-standardized body description texts, DICOM files and image pictures, thereby solving the problem of "improper design of information systems in most medical institutions and uneven data quality" and achieving the purpose of accurately identifying input medical data and outputting standard and standardized identification data.

[0125] The embodiment of the present invention can perform targeted feature extraction on the medical examination data to be collected through corresponding recognition parameters. Since the recognition parameters are obtained by pre-training the medical examination data set of the corresponding medical institution, the recognition parameters can improve the model recognition accuracy based on the writing standards and examination equipment differences of the corresponding medical institution, thereby improving the recognition accuracy of the medical examination data to be collected.

[0126] The embodiment of the present invention can also standardize the medical examination data of low-level medical institutions based on the medical examination data of high-level medical institutions for the same part of the patient within the same time period, so that the medical examination data of low-level medical institutions are more standardized, thereby improving the accuracy of the medical examination data of low-level medical institutions in subsequent identification.

[0127] In summary, the embodiments of the present invention can accurately identify input medical data and output standard and standardized identification data, so as to achieve the purpose of standardizing and normalizing medical data.

[0128] 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," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0129] Each embodiment in this specification is described in a related manner. Similar portions between the embodiments can be referenced to each other. Each embodiment focuses on the differences from other embodiments. In particular, the system embodiments are generally similar to the method embodiments, so their description is relatively simple. For related portions, reference can be made to the description of the method embodiments.

[0130] The above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. A multi-site identification method based on RAG optimization of a large medical model, characterized by: The method is used for the collection and management of medical data of medical institutions across different imaging cloud platforms, and the method includes: Step S1: receiving first medical examination data to be collected from a first medical institution; wherein the first medical examination data is multimodal data including at least corresponding text data and image data; the first medical institution is one of multiple medical institutions on different imaging cloud platforms to which data is to be collected; Step S2: Determine whether there is a first identification parameter corresponding to the first medical institution. If so, perform data preprocessing and part feature extraction on the first medical examination data based on the first identification parameter to obtain the part identification feature of the first medical examination data; if not, perform data preprocessing and part feature extraction on the first medical examination data based on the standard identification parameter to obtain the part identification feature of the first medical examination data; wherein, the first identification parameter is obtained by pre-training the medical examination data set of the first medical institution, and the identification parameters obtained by training different medical institutions are different. The first identification parameter is used to improve the model recognition accuracy according to the differences in writing standards and inspection equipment of different medical institutions; the standard identification parameter is obtained by mixed training of a general medical examination data set; the part identification feature includes part identification text feature and part identification image feature; Step S3: performing a fusion analysis on the part identification text feature and the part identification image feature based on the first recognition parameter or the standard recognition parameter to obtain a similarity between the part identification text feature and the part identification image feature, and further obtain a first confidence level between the part identification text feature and the part identification image feature; Step S4: in response to the first confidence level being greater than a standard confidence level, obtaining a fusion feature, inputting the fusion feature into a RAG industry model analysis system, causing the RAG industry model analysis system to output data-standardized part identification data corresponding to the first medical examination data; After step S1, the method further includes step S1A: Obtaining first patient information corresponding to the first medical examination data; obtaining historical medical examination data of the first patient based on the first patient information; In response to the existence of second medical examination data in the historical medical examination data, the second medical examination data is obtained in response to the existence of second medical examination data whose time difference with the first medical examination data does not exceed a first threshold and whose examination site is the same as that of the first medical examination data; In response to the second medical examination data corresponding to a level of a second medical institution being higher than a level of the first medical institution, outputting, according to a writing specification of the second medical examination data, first medical examination corrected data corresponding to the first medical examination data and conforming to a data standardization specification and inputting the corrected data as the first medical examination data into step S2 for processing; Wherein, the step S1A further includes: Acquire the first medical examination data and the second medical examination data having a time difference of no more than a first threshold and the same examination site as the first medical examination data, and construct a data pair set of the first medical examination data and the second medical examination data; Based on the image data in the data pair set, an image data correction model of the correspondence between the first medical examination data and the second medical examination data is trained; the image data correction model is used to autonomously generate second mirror medical examination data corresponding to the first medical examination data of the first medical institution and the writing standards of the second medical institution, and is used for aggregation and as the first medical examination data to execute subsequent steps.

2. The inspection multi-site identification method based on RAG optimization of a large medical model according to claim 1 is characterized in that: Before obtaining the second medical examination data in response to the existence of second medical examination data in the historical medical examination data, which has a time difference with the first medical examination data that does not exceed a first threshold and has the same examination site as the first medical examination data, the method includes: Extracting features corresponding to the first medical examination data and the historical medical examination data respectively, and obtaining examination parts corresponding to the first medical examination data and the historical medical examination data; According to the examination parts corresponding to the first medical examination data and the historical medical examination data, second medical examination data having a time difference with the first medical examination data that does not exceed a first threshold and having the same examination part as the first medical examination data is obtained.

3. The inspection multi-site identification method based on RAG optimization of a large medical model according to claim 1 is characterized in that: In step S2, the method further includes: In response to the absence of the first identification parameter corresponding to the first medical institution, extracting common features from the medical examination data set from the first medical institution; wherein the common features include at least text word usage features, text format features, examination equipment model features, and image clarity features; Training is performed based on the common features to obtain the first recognition parameters.

4. The inspection multi-site identification method based on RAG optimization of a large medical model according to claim 1 is characterized in that: In step S1 , the data of multiple modalities of the first medical examination data are transmitted separately or in mixed and coordinated manner.

5. The inspection multi-part identification method based on RAG optimization of a large medical model according to claim 1 is characterized in that: The data preprocessing of the first medical examination data in step S2 includes: Noise filtering is performed on the text data in the first medical examination data; and a corresponding processing method is selected according to the type of image data in the first medical examination data.

6. The method for multi-site identification based on RAG optimization of a large medical model according to claim 1, characterized in that: The step S3 comprises: According to the first recognition parameter or the standard recognition parameter, the recognition is performed through a cross-attention mechanism; A confidence weighting model is constructed according to the DS evidence theory, and the first confidence between the part identification text feature and the part identification image feature is obtained through the confidence weighting model.

7. A multi-part intelligent inspection system based on RAG optimization of a large medical model, characterized by: The intelligent system is used for the collection and management of medical data of medical institutions across different imaging cloud platforms. The intelligent system includes: a multimodal input and output layer, an intelligent analysis engine layer, and a RAG industry model analysis system; The multimodal input and output layer is used to receive first medical examination data to be collected from a first medical institution, or output part identification data in accordance with the data standardization specification of the RAG industry model analysis system; wherein the first medical examination data is multimodal data including at least corresponding text data and image data; the first medical institution is one of the medical institutions on multiple different imaging cloud platforms whose data is to be collected; The intelligent analysis engine layer includes a feature extraction module and a multimodal decision fusion module; The feature extraction module is used to determine whether there is a first identification parameter corresponding to the first medical institution. If so, data preprocessing and part feature extraction are performed on the first medical examination data based on the first identification parameter to obtain the part identification feature of the first medical examination data; if not, data preprocessing and part feature extraction are performed on the first medical examination data based on the standard identification parameter to obtain the part identification feature of the first medical examination data; wherein, the first identification parameter is obtained by pre-training the medical examination data set of the first medical institution, and the identification parameters obtained by training are different for different medical institutions. The first identification parameter is used to improve the model recognition accuracy according to the differences in writing standards and inspection equipment of different medical institutions; the standard identification parameter is obtained by mixed training of a general medical examination data set; the part identification feature includes part identification text feature and part identification image feature; The multimodal decision fusion module is configured to perform a fusion analysis on the part identification text feature and the part identification image feature based on the first recognition parameter or the standard recognition parameter to obtain a similarity between the part identification text feature and the part identification image feature, and further obtain a first confidence level between the part identification text feature and the part identification image feature; in response to the first confidence level being greater than the standard confidence level, a fusion feature is obtained, and the fusion feature is input into the RAG industry model analysis system; The RAG industry model analysis system is configured to output, to the multimodal input and output layer, data-standardized part identification data corresponding to the first medical examination data based on the fusion features; The intelligent system further comprises: a historical data acquisition module, a same inspection data acquisition module, and a standard data rapid acquisition module; The historical data acquisition module is configured to obtain first patient information corresponding to the first medical examination data; and obtain the historical medical examination data of the first patient based on the first patient information; The same examination data acquisition module is configured to obtain the second medical examination data in response to the presence of second medical examination data in the historical medical examination data, the second medical examination data having a time difference with the first medical examination data that does not exceed a first threshold and has the same examination site; The standard data rapid acquisition module is configured to, in response to the second medical examination data corresponding to the second medical examination data having a level greater than that of the first medical institution, output, according to the writing specifications of the second medical examination data, the first medical examination corrected data that conforms to the data standardization specification and corresponds to the first medical examination data, and input the data as the first medical examination data to the feature extraction module for processing; Wherein, the intelligent agent system further comprises: a data pair set construction module and a correction model training module; The data pair set construction module is configured to obtain the first medical examination data and the second medical examination data having a time difference of no more than a first threshold and the same examination site as the first medical examination data, and to construct a data pair set of the first medical examination data and the second medical examination data; The correction model training module is used to train an image data correction model of the correspondence between the first medical examination data and the second medical examination data based on the image data in the data pair set; the image data correction model is used to autonomously generate second mirror medical examination data corresponding to the first medical examination data of the first medical institution and the writing standards of the second medical institution, and is used for aggregation and as the first medical examination data to execute subsequent steps.

8. The multi-part intelligent inspection system based on RAG optimization of a large medical model according to claim 7 is characterized in that: The intelligent body system further comprises: an examination part obtaining module and a second medical examination data obtaining module; The examination part acquisition module is used to extract features corresponding to the first medical examination data and the historical medical examination data, respectively, to obtain examination parts corresponding to the first medical examination data and the historical medical examination data; The second medical examination data acquisition module is used to obtain second medical examination data with a time difference of no more than a first threshold and the same examination site as the first medical examination data based on the examination sites corresponding to the first medical examination data and the historical medical examination data.

Images