An intelligent processing system for urology diagnosis and treatment data based on artificial intelligence

Abstract

The present invention discloses an intelligent processing system for urology diagnosis and treatment data based on artificial intelligence, comprising the following modules: an original diagnosis and treatment data acquisition and preprocessing module for acquiring data from urology patients; a stage annotation text construction module for constructing a stage annotation text set; an image description structuring module for constructing a structured image description sequence; a cross-modal semantic fusion module for extracting a cross-modal joint semantic representation vector sequence; a named entity recognition module for outputting a structured medical entity set; an entity semantic graph construction module for constructing an entity semantic graph structure; a diagnosis and treatment relationship extraction module for outputting a diagnosis and treatment relationship set of a triple structure; a structured encapsulation module for outputting a structured diagnosis and treatment data unit; and an intelligent diagnosis and treatment decision module for performing disease identification, treatment path generation, and reasoning suggestion output. The present invention integrates multimodal semantic modeling and graph neural reasoning to construct an intelligent diagnosis and treatment relationship recognition system.

Description

Technical Field

[0001] The present invention relates to the field of medical artificial intelligence and clinical information structured processing technology, and in particular to an intelligent processing system for urology diagnosis and treatment data based on artificial intelligence. Background Art

[0002] With the continued development of electronic medical record systems, picture archiving systems (PACS), and clinical information systems (CIS), hospitals have accumulated a vast amount of complex medical data from diverse sources. In the field of urology, in particular, patient data often consists of heterogeneous content, including structured basic information fields, unstructured electronic medical record text, and image description reports. How to fully exploit the implicit semantics within this multi-source medical data and implement structured, standardized modeling and intelligent processing of the medical process is a key topic in the current research and application of medical artificial intelligence.

[0003] Existing medical data structuring technologies mostly focus on processing single-modal content, such as entity recognition and stage annotation methods for electronic medical record texts, and lexical normalization and phrase extraction methods for image reports. These methods generally suffer from problems such as information fragmentation between modalities, missing stage semantics, and insufficient context association capabilities, making it difficult to support semantic modeling of complete diagnosis and treatment pathways. At the same time, some studies have attempted to introduce deep learning models to semantically represent or extract relationships from medical texts. However, when processing urology diagnosis and treatment data with complex structures such as time evolution, stage jumps, and logical breaks, they lack the embedding mechanism for stage information and the global reasoning ability for path semantics, making it difficult to restore the continuity, logic, and individual differences inherent in the actual diagnosis and treatment process.

[0004] Furthermore, current knowledge graph construction technologies mostly rely on manual rules or general graph neural networks to model relationships between medical entities. These technologies suffer from weak inter-entity path representation capabilities, low granularity in identifying relationship types, and difficulty quantifying inference confidence. These issues make them unable to meet the practical needs of urology for extracting high-precision diagnostic and treatment logic. Regarding intelligent diagnosis and treatment recommendations, existing systems often employ static rules or historical case matching methods, lacking a semantic consistency verification mechanism driven by structured data units, making it difficult to generate personalized, interpretable, and well-structured intelligent diagnosis and treatment recommendations.

[0005] The artificial intelligence-based intelligent processing system for urology diagnosis and treatment data provided by the present invention is designed to solve the above problems. By introducing structured basic information fields and diagnosis and treatment stage labels in the multi-source data fusion stage, designing an embedded encoder and a cross-modal semantic fusion mechanism in the integration stage, the context alignment and semantic injection of text and image information are achieved; combining graph neural networks with dynamic multi-hop path aggregation algorithms, entity semantic graphs are constructed and structured continuous diagnosis and treatment triple relationships are extracted; further, stage information is bound to individual attributes and encapsulated into standardized structured diagnosis and treatment data units to drive diagnosis, recommendation, and auxiliary reasoning modules, ultimately improving the structuralization, semantic expression ability, and intelligent application level of urology diagnosis and treatment data.

[0006] Therefore, how to provide an intelligent processing system for urology diagnosis and treatment data based on artificial intelligence is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0007] One purpose of the present invention is to propose an intelligent processing system for urology diagnosis and treatment data based on artificial intelligence. The present invention integrates the cross-modal semantic information of electronic medical records and image descriptions, uses graph neural networks and semantic path reasoning technology to accurately identify urology diagnosis and treatment relationships, construct structured data units, realize intelligent diagnosis, personalized recommendations and clinical decision assistance, and improve the processing efficiency and intelligence level of diagnosis and treatment data.

[0008] An intelligent processing system for urology diagnosis and treatment data based on artificial intelligence according to an embodiment of the present invention includes the following modules:

[0009] The original diagnosis and treatment data collection and preprocessing module is used to collect structured basic information fields, electronic medical record texts, and description texts corresponding to medical images of urology patients;

[0010] The stage-annotated text construction module is used to identify main paragraphs, divide semantic blocks, normalize medical terms, and label diagnosis and treatment stages in electronic medical record texts, and build a stage-annotated text collection based on structured basic information fields;

[0011] The image description structuring module is used to perform normalization, vocabulary standardization, syntactic structure analysis, semantic entity recognition and sequence encapsulation on medical image description text to construct a structured image description sequence;

[0012] The cross-modal semantic fusion module is used to input the stage annotation text collection and the structured image description sequence into the semantic fusion network composed of the integrated stage embedding encoder and the cross-modal cross-attention mechanism to extract the cross-modal joint semantic representation vector sequence;

[0013] A named entity recognition module is used to input the cross-modal joint semantic representation vector sequence into the named entity recognition structure, identify medical entities in the diagnosis and treatment data, and output a structured medical entity set;

[0014] The entity semantic graph construction module is used to input the structured medical entity set into the graph neural network model, combine it with the urology-specific knowledge graph, and build the entity semantic graph structure based on graph convolution propagation and node semantic attention mechanism;

[0015] The diagnosis and treatment relationship extraction module is used to perform dynamic semantic graph pruning and multi-scale path integration algorithms based on the entity semantic graph structure, construct a set of high-confidence candidate paths, perform semantic aggregation and relationship type identification, and output a set of diagnosis and treatment relationships in a triple structure;

[0016] The structured encapsulation module is used to bind the diagnosis and treatment relationship set with the stage annotation information, and at the same time integrate the structured basic information fields to construct the patient's individual attribute vector, uniformly map it to the medical terminology standard system, and encapsulate it into a structured diagnosis and treatment data unit;

[0017] The intelligent diagnosis and treatment decision module is used to input structured diagnosis and treatment data units into the intelligent diagnosis module, personalized treatment recommendation module, and clinical decision support module, respectively performing disease identification, treatment path generation, and reasoning recommendation output, and generating intelligent diagnosis and treatment recommendation reports that can be used on system terminals or remote platforms;

[0018] The modules are implemented as follows:

[0019] S1. Collecting original diagnosis and treatment data of urology patients, wherein the original diagnosis and treatment data includes structured basic information fields, electronic medical record text, and description text corresponding to medical images;

[0020] S2. Perform sentence segmentation, segmentation, and medical terminology normalization on the electronic medical record text, combine the structured basic information fields with the text context, annotate the diagnosis and treatment stage labels, and construct a stage annotated text collection;

[0021] S3. Perform semantic cleaning, vocabulary standardization, and syntactic structure extraction on the medical image description text to extract the anatomical parts, lesion types, and diagnostic conclusions corresponding to the images, and construct a structured image description sequence;

[0022] S4. Input the stage annotation text collection and structured image description sequence into the semantic fusion network composed of the integration stage embedding encoder and the cross-modal cross attention mechanism to extract the cross-modal joint semantic representation vector sequence;

[0023] S5. Input the cross-modal joint semantic representation vector sequence into the named entity recognition module to identify medical entities in the diagnosis and treatment data;

[0024] S6. Input medical entities into the graph neural network and construct the entity semantic graph structure based on graph convolution propagation and node semantic attention mechanism;

[0025] S7. Based on the entity semantic graph structure, execute the dynamic semantic graph pruning and multi-scale path integration algorithm to extract the continuous diagnosis and treatment logical relationship between entities and output the diagnosis and treatment relationship set;

[0026] S8. Bind the diagnosis and treatment relationship set with the stage annotation information, associate the structured basic information fields, and encapsulate them into structured diagnosis and treatment data units according to medical terminology standards;

[0027] S9, inputting the structured diagnosis and treatment data unit into the intelligent diagnosis module, the personalized treatment recommendation module and the clinical decision support module;

[0028] The S6 specifically includes:

[0029] S61, inputting the structured medical entity set as node information into the graph neural network initialization module, and constructing an edge connection matrix between the entities based on the original text co-occurrence relationship, paragraph structure relationship and consistency of diagnosis and treatment stages between the medical entities;

[0030] S62. Extract standard entity nodes, term definition nodes, concept affiliation nodes, and clinical logic relationship edges from the pre-built urology-specific medical knowledge graph, and embed them into the current graph structure to form a knowledge-enhanced medical entity graph;

[0031] S63, inputting the medical entity graph into the graph convolution propagation unit of the graph neural network, adopting a layered propagation strategy, performing graph convolution calculation according to the node topology structure and initial node features, and updating the local neighborhood semantic feature representation of each node;

[0032] S64. Introducing a node semantic attention mechanism in the graph convolution propagation process, weighting adjacent edges by calculating the semantic similarity and context consistency between nodes to guide the direction of feature aggregation;

[0033] S65. Perform residual connection and feature fusion on the node embedding vectors output by each layer of graph convolution to generate a multi-level semantic representation, and perform node vector regularization on the entire graph.

[0034] S66. Combine the entity node semantic representation and edge relationship semantic representation finally output by the graph neural network and encapsulate them to construct an entity semantic graph structure.

[0035] Optionally, the S2 specifically includes:

[0036] S21. Perform main paragraph recognition on the electronic medical record text and divide the text into several diagnosis and treatment semantic blocks, including the chief complaint segment, medical history segment, examination segment, diagnosis segment, treatment segment, and follow-up segment. Use segmentation rules based on position priors and keyword semantic matching to perform text segmentation;

[0037] S22. Perform medical terminology normalization on each diagnosis and treatment semantic block, map non-standard vocabulary, abbreviations, and semantically ambiguous expressions within the diagnosis and treatment semantic block to standard medical terms, use a term mapping dictionary and context-aware rules to jointly determine standard terms, and output a normalized set of diagnosis and treatment text sub-blocks;

[0038] S23. Extract paragraph-level semantic features for each sub-block of medical text, construct a temporal semantic vector sequence based on the context window, and embed structured basic information fields into the vector representation. The basic information fields include the patient's age, gender, duration of the main complaint, admission method, initial diagnosis time, surgical history, and previous diagnosis labels.

[0039] S24. Inputting the semantic vector sequence of the fused structured field into a diagnosis and treatment stage classification model, the diagnosis and treatment stage classification model adopts a multi-channel parallel encoding structure, wherein the first channel processes the text semantic backbone, the second channel processes the structured label features, and a gated fusion mechanism is used to achieve feature interaction, and outputs the diagnosis and treatment stage label corresponding to each text sub-block;

[0040] S25. Performing stage boundary calibration on the predicted diagnosis and treatment stage label sequence, modifying the critical paragraph labels using a reordering mechanism based on local confidence score and context consistency, and generating a diagnosis and treatment stage boundary sequence;

[0041] S26. Bind and encapsulate the diagnosis and treatment text sub-blocks with the corresponding diagnosis and treatment stage labels to construct a stage annotated text set, which serves as the input data set for the subsequent semantic fusion network and named entity recognition module.

[0042] Optionally, the S3 specifically includes:

[0043] S31. Normalize the text of medical image descriptions, remove non-diagnosis-related tags, noise symbols, and non-structural terms without actual clinical significance, restructure sentence boundaries with disordered formatting, and construct sentence element sequences with standard word order;

[0044] S32, performing a lexical standardization operation on the normalized sentence sequence, using a static vocabulary based on a domain term library and context-related word expansion rules to normalize non-standard words, synonyms, and spelling variations contained in the medical image description into standard medical vocabulary, and replacing the original expressions;

[0045] S33. Use a multi-level syntactic analysis module to perform dependency structure analysis and part-of-speech tagging on the standardized sentence sequence, extract the subject-verb-object structure, modification structure and spatial positioning expression in the sentence, and construct a multi-level syntactic dependency graph;

[0046] S34. Identify anatomical part entities involved in the image description based on the semantic verbs, spatial prepositions, and anatomical lexemes in the multi-level syntactic dependency graph, select noun phrases with anatomical attributes, and perform entity classification;

[0047] S35. Utilize the attention-based lesion word recognition model to analyze the descriptive phrases and determinant word structures within the context of the anatomical site and extract the lesion type entities contained in the image report;

[0048] S36, combining verb structure, sentence-ending assertion pattern and diagnostic term matching rules, locate formative judgment sentences in image descriptions, extract corresponding diagnostic conclusion terms and map them to standard term nodes in the medical diagnostic vocabulary;

[0049] S37. Encapsulate the anatomical part entities, lesion type entities, and diagnostic conclusion terms identified in each image description text into a structured entity sequence in the order of appearance, construct a structured image description sequence, and use it as input content for the subsequent semantic fusion network.

[0050] Optionally, the S4 specifically includes:

[0051] S41. Grouping the stage annotation text set by diagnosis and treatment stage type to establish a first diagnosis stage set, an examination stage set, a diagnosis stage set, a treatment stage set, and a follow-up stage set, thereby forming a stage partition input structure;

[0052] S42. Each stage set is fed into a corresponding stage embedding encoder. The stage embedding encoder includes a diagnosis and treatment stage label embedding layer, a stage position embedding layer, and a stage semantic encoding module. The stage label embedding vector guides sentence element expression and combines relative position information to construct the internal context semantic structure of the stage.

[0053] S43. Perform stage feature extraction on the vector output from the stage semantic encoding module, using a multi-scale convolutional structure and attention filtering mechanism to extract local diagnosis and treatment expressions and global decision intentions, respectively, to form a stage-level semantic representation tensor;

[0054] S44, stacking all stage-level semantic representation tensors in a unified representation space, adjusting the semantic amplitude through a stage-level normalization mechanism, and generating a stage-aggregated text semantic vector sequence;

[0055] S45, inputting the structured image description sequence into the image semantic encoding module, performing entity-level embedding mapping, syntactic position information binding and feature tensor construction, and outputting an image description vector sequence;

[0056] S46. Input the stage-aggregated text semantic vector sequence and the image description vector sequence into the cross-modal cross attention module, and construct an interaction weight matrix using a dual-channel mechanism of text-to-image attention and image-to-text attention;

[0057] S47, based on the calculation result of the interaction weight matrix, perform semantic alignment and spatial mapping, inject the image semantic vector into the stage text semantic space, and generate a fused joint semantic representation;

[0058] S48. Connect the residual connection of the joint semantic representation input and the gate integration structure to complete cross-modal semantic compression and sequence encoding update, and output the final cross-modal joint semantic representation vector sequence for subsequent entity recognition task input.

[0059] Optionally, the S5 specifically includes:

[0060] S51, sending the cross-modal joint semantic representation vector sequence as input to a named entity recognition module, wherein the named entity recognition module includes a context feature extraction submodule, an entity boundary determination submodule, and a label classification submodule;

[0061] S52. In the context feature extraction submodule, a bidirectional gated recurrent unit network is used to perform bidirectional modeling on the input vector sequence, obtain a dynamic representation of each semantic unit in the current context, and form a context semantic feature sequence;

[0062] S53, inputting the contextual semantic feature sequence into the entity boundary determination submodule, constructing an entity candidate window based on the position annotation strategy, detecting the potential entity boundary position through the sliding window mechanism, and generating entity boundary candidate segments;

[0063] S54, jointly inputting the entity boundary candidate segment and the contextual semantic feature sequence into the label classification submodule, using a multi-task parallel structure to jointly discriminate the entity category and label position type, and outputting an entity label sequence, wherein the entity label types include a start label, an intermediate label, and a non-entity label;

[0064] S55. Reverse mapping the cross-modal joint semantic representation vector based on the entity label sequence, extracting the corresponding entity text fragments in the diagnosis and treatment text, and performing vocabulary reconstruction and semantic consistency verification to eliminate pseudo-entity areas that do not conform to the standard vocabulary;

[0065] S56. The finally identified medical entities are grouped by category, and classified into disease name, symptom description, surgical method, anatomical location, examination index and medication items, to construct a structured medical entity set, which serves as the basic input for subsequent knowledge graph modeling and entity relationship extraction.

[0066] Optionally, the S7 specifically includes:

[0067] S71. Path sampling initialization is performed based on the semantic representation of the entity nodes and the edge relationship types in the entity semantic graph structure. A depth-first search algorithm is used to generate a set of candidate multi-hop paths from the start node to any target node, each with no more than a preset number of hops. The candidate multi-hop path set consists of a sequence of entity nodes and a sequence of edge relationships, and is accompanied by path indexes and position markers between nodes.

[0068] S72. For each path in the candidate path set, calculate the semantic similarity score and the treatment stage label consistency score between the entity nodes, where the semantic similarity score is based on the weighted accumulation of the context feature similarity between all consecutive nodes in the path, and the treatment stage consistency score is based on the monotonically increasing and semantically consistent score of the treatment stage label sequence bound to the entities in the path;

[0069] S73. Establish a joint path confidence scoring mechanism, fuse the semantic similarity score and the stage consistency score according to the weighted coefficient, generate a path global confidence score vector, set a path pruning threshold interval, and perform dynamic pruning operations on paths below the threshold to eliminate semantically scattered, logically jumping, and disordered diagnosis and treatment stage paths, and obtain a high-confidence path subset;

[0070] S74. The pruned path subsets are layered according to path length to construct a hop-level structure. Multi-path aggregation is performed on all paths in each hop-level layer. Entity representation aggregation, edge relationship aggregation, and position weight adjustment strategies are used to generate a hop-level path representation tensor. Cross-layer attention weighted fusion is performed on path vectors between different hop levels to obtain a multi-scale path semantic feature matrix.

[0071] S75. Input the multi-scale path semantic feature matrix into a relationship type recognition module. The relationship type recognition module uses a multi-task structure including a fully connected classification layer and a graph structure constraint filter to predict the entity relationship type of each path and output the corresponding relationship label and path confidence level.

[0072] S76. Based on the path confidence level and structure matching rules, screen out the diagnosis and treatment paths with closed structure, clearly defined relationships and semantic continuity of entity pairs, extract the starting entity node, intermediate edge relationship and ending entity node in each path, organize and remove duplicates in triple format, and output a structured diagnosis and treatment relationship set as the input entity pair set for the subsequent semantic reasoning and knowledge fusion module.

[0073] Optionally, the S8 specifically includes:

[0074] S81, performing index-level alignment on the triple structure diagnosis and treatment relationship set output in S7 and the stage annotated text set constructed in S2, mapping and matching the diagnosis and treatment stage labels based on the starting position of the entity in the original text, and establishing a binding mapping table between the diagnosis and treatment relationship and the stage label;

[0075] S82. Perform a stage attribution check on the starting entity and the ending entity in each diagnosis and treatment relationship. If the two entities belong to different diagnosis and treatment stages, establish a sequence constraint relationship based on the time sequence of the stage labels and mark the diagnosis and treatment process direction attribute of the relationship path;

[0076] S83. Extract structured basic information fields, including patient age, gender, initial diagnosis time, duration of main complaint, department visited, and family history label, construct the patient's individual attribute vector, and fuse and map it with each entity node in the diagnosis and treatment relationship set as the entity context feature completion input;

[0077] S84. Perform semantic consistency determination on the set of diagnosis and treatment relationships that have been bound to stage information and individual attributes. Calculate confidence weights based on stage continuity, entity consistency, and context matching. Eliminate diagnosis and treatment triples that are logically contradictory or contextually disconnected, and retain high-confidence relationship sets that conform to clinical semantic rules.

[0078] S85. Combine the retained diagnosis and treatment triples with their corresponding stage labels, entity identifiers, individual attribute labels, and relationship direction markers, and uniformly map them to the medical terminology standard system, including ICD, SNOMED CT, UMLS, and the generic drug name coding system. Build a structured diagnosis and treatment field set based on the mapping results.

[0079] S86. Aggregate and encapsulate the triple relationship fields, stage labels and individual attribute codes under each set of standard term codes to construct a structured diagnosis and treatment data unit, and output it in a unified data format, including diagnosis and treatment stage fields, standard entity fields, standard relationship fields, individual feature fields and relationship confidence fields, for downstream diagnosis, recommendation or analysis module calls.

[0080] Optionally, the S9 specifically includes:

[0081] S91. Modularize and distribute the structured diagnosis and treatment data units, dividing them into a diagnosis input set, a treatment recommendation input set, and a decision reasoning input set according to the data field type, and send them to the intelligent diagnosis module, the personalized treatment recommendation module, and the clinical decision support module respectively;

[0082] S92. In the intelligent diagnosis module, standardized disease entities, symptom manifestations, examination indicators, and stage labels in the diagnosis input set are received and input into a diagnosis classification model built based on a multi-layer feature fusion and semantic alignment mechanism, and the corresponding disease identification label and diagnosis confidence score are output;

[0083] S93. In the personalized treatment recommendation module, the surgical method, drug entity, anatomical location, patient individual attribute vector, and diagnosis and treatment stage information in the treatment recommendation input set are received and input into a treatment matching model based on graph structure path retrieval and treatment effect history inversion mechanism, and the recommended treatment plan sequence and efficacy estimation score are output;

[0084] S94. In the clinical decision support module, the triple structure diagnosis and treatment relationship, stage label sequence, and patient full feature code in the decision reasoning input set are received and input into the reasoning engine including the rule reasoning layer and the neural symbol fusion layer. The module performs diagnosis and treatment pathway consistency verification, risk event simulation, and candidate pathway scoring, and outputs a decision support recommendation set.

[0085] S95. The diagnostic result labels, recommended treatment paths and auxiliary decision-making suggestions output by the three modules are fused and consistency checked to form a final intelligent diagnosis and treatment recommendation report. The report structure includes disease identification results, recommended intervention plans, risk prediction information and phased optional paths, which serves as the output content of the system terminal or remote consultation platform.

[0086] The beneficial effects of the present invention are:

[0087] First, this invention connects the previously isolated semantic dimensions of urology diagnosis and treatment data by constructing a multimodal input system covering electronic medical record text, image description text, and structured basic information fields. Through the synergistic effect of the stage annotation text construction module and the image description structuring module, a standardized expression of semantic segments and image features at different stages of the diagnosis and treatment process is achieved, effectively improving the accuracy of downstream semantic fusion and relationship extraction, and providing a solid data foundation for subsequent diagnosis and treatment logic modeling.

[0088] Secondly, the present invention introduces a cross-modal semantic fusion network, which embeds an encoder and a cross-attention mechanism in the fusion stage to fully integrate the semantic information of text and image, generating a context-consistent joint representation vector. This mechanism can significantly enhance the model's perception of stage context, semantic co-occurrence, and modality alignment, solving the problems of modality fragmentation and representation distortion in existing technologies, and effectively improving the accuracy and robustness of medical entity recognition and diagnosis and treatment relationship extraction.

[0089] Furthermore, after medical entity recognition, the present invention introduces a graph neural network and a urology knowledge graph to construct an entity semantic graph structure. Combining a semantic attention mechanism with a multi-scale path aggregation algorithm, this method identifies logically continuous diagnosis and treatment relationship paths. This path extraction method not only enhances the ability to model the contextual structure between entities but also improves the reliability of relationship type identification through path confidence assessment and structural constraint rules, addressing the lack of abstraction of diagnosis and treatment paths in existing methods.

[0090] Finally, the present invention deeply binds the diagnosis and treatment relationship set with stage labels and individual attributes, encapsulating it into a structured diagnosis and treatment data unit driven by standard terminology, and using it as a unified input interface for intelligent diagnosis, personalized recommendation, and auxiliary reasoning modules, completing a closed-loop processing flow from data understanding to intelligent decision-making. This mechanism not only achieves deep semantic modeling and reasoning calls for urology diagnosis and treatment data, but also effectively improves the interpretability, accuracy, and application feasibility of intelligent diagnosis and treatment recommendations, significantly promoting the systematic application of artificial intelligence technology in specialized clinical scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0092] Figure 1 This is a schematic diagram of the structure of an intelligent processing system for urology diagnosis and treatment data based on artificial intelligence proposed by the present invention;

[0093] Figure 2 This is an overall flow chart of the intelligent processing method for urology diagnosis and treatment data based on artificial intelligence proposed by the present invention. DETAILED DESCRIPTION

[0094] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0095] refer to Figure 1 , an intelligent processing system for urology diagnosis and treatment data based on artificial intelligence, including the following modules:

[0096] The original diagnosis and treatment data collection and preprocessing module is used to collect the structured basic information fields, electronic medical record texts and description texts corresponding to medical images of urology patients, and perform unified coding and format standardization operations on the data;

[0097] The stage-annotated text construction module is used to identify main paragraphs, divide semantic blocks, normalize medical terms, and label diagnosis and treatment stages in electronic medical record texts, and build a stage-annotated text collection based on structured basic information fields;

[0098] The image description structuring module is used to perform normalization, vocabulary standardization, syntactic structure analysis, semantic entity recognition and sequence encapsulation on medical image description text to construct a structured image description sequence;

[0099] The cross-modal semantic fusion module is used to input the stage annotation text collection and the structured image description sequence into the semantic fusion network composed of the integrated stage embedding encoder and the cross-modal cross-attention mechanism to extract the cross-modal joint semantic representation vector sequence;

[0100] The named entity recognition module is used to input the cross-modal joint semantic representation vector sequence into the named entity recognition structure to identify medical entities in the diagnosis and treatment data, including disease names, symptom descriptions, surgical methods, anatomical locations, examination indicators and medication items, and output a structured medical entity set;

[0101] The entity semantic graph construction module is used to input the structured medical entity set into the graph neural network model, combine it with the urology-specific knowledge graph, build the entity semantic graph structure based on graph convolution propagation and node semantic attention mechanism, and generate contextual semantic representations between entities;

[0102] The diagnosis and treatment relationship extraction module is used to perform dynamic semantic graph pruning and multi-scale path integration algorithms based on the entity semantic graph structure, construct a set of high-confidence candidate paths, perform semantic aggregation and relationship type identification, and output a set of diagnosis and treatment relationships in a triple structure;

[0103] The structured encapsulation module is used to bind the diagnosis and treatment relationship set with the stage annotation information, and at the same time integrate the structured basic information fields to construct the patient's individual attribute vector, uniformly map it to the medical terminology standard system, and encapsulate it into a structured diagnosis and treatment data unit;

[0104] The intelligent diagnosis and treatment decision module is used to input structured diagnosis and treatment data units into the intelligent diagnosis module, personalized treatment recommendation module and clinical decision support module, which are used to perform disease identification, treatment path generation and reasoning suggestion output respectively, and generate intelligent diagnosis and treatment recommendation reports that can be used for system terminals or remote platforms.

[0105] This system, with its modular architecture, closes the entire process of diagnostic and treatment data from raw data collection, semantic fusion, knowledge modeling, to intelligent output, breaking through the bottlenecks of existing systems, which include fragmented processing processes, difficult modal integration, and low data utilization. Through the logical coordination of various functional modules, the system enables in-depth processing of multimodal, multi-granular, and multi-level diagnostic and treatment data, providing structured, reasonable, and interactive knowledge support for clinicians, significantly enhancing the data-driven capabilities of urology departments in disease understanding, pathway decision-making, and treatment planning.

[0106] refer to Figure 2 , an intelligent processing method for urology diagnosis and treatment data based on artificial intelligence, comprising the following steps:

[0107] S1. Collect original diagnosis and treatment data of urology patients, including structured basic information fields, electronic medical record text, and description text corresponding to medical images, and perform unified coding and format standardization on all data;

[0108] S2. Perform sentence segmentation, segmentation, and medical terminology normalization on the electronic medical record text. Combine the structured basic information fields with the text context, use semantic rules and the diagnosis and treatment event model to annotate the diagnosis and treatment stage labels, and construct a stage annotated text collection;

[0109] S3. Perform semantic cleaning, vocabulary standardization, and syntactic structure extraction on the medical image description text to extract the anatomical parts, lesion types, and diagnostic conclusions corresponding to the images, and construct a structured image description sequence;

[0110] S4. Input the stage annotation text set and image description sequence into the semantic fusion network composed of the integration stage embedding encoder and the cross-modal cross attention mechanism to extract the cross-modal joint semantic representation vector sequence;

[0111] S5. Inputting the cross-modal joint semantic representation vector sequence into a named entity recognition module to identify medical entities in the diagnosis and treatment data, wherein the medical entities include disease names, symptom descriptions, surgical methods, anatomical locations, examination indicators, and medication items;

[0112] S6. Input the identified medical entities into the graph neural network constructed based on the urology-specific knowledge graph, build the entity semantic graph structure based on graph convolution propagation and node semantic attention mechanism, and generate contextual semantic representations between entities;

[0113] S7. Based on the entity semantic graph structure, a dynamic semantic graph pruning and multi-scale path integration algorithm is executed, including: constructing a candidate multi-hop path set; dynamically pruning low-correlation paths based on the semantic similarity and stage consistency between entities; performing feature weighting on the remaining paths using a path length hierarchical aggregation mechanism; performing relationship type discrimination and structure matching based on the weighted features, extracting the continuous diagnosis and treatment logical relationship between entities, and outputting a diagnosis and treatment relationship set with a triple structure;

[0114] S8. Bind the diagnosis and treatment relationship set with the stage annotation information, and associate the structured basic information fields for individual feature modeling and context constraint completion, and encapsulate them into structured diagnosis and treatment data units according to medical terminology standards;

[0115] S9: Input the structured diagnosis and treatment data units into the intelligent diagnosis module, personalized treatment recommendation module, and clinical decision support module for disease identification, intervention path planning, and auxiliary reasoning output, respectively;

[0116] The S6 specifically includes:

[0117] S61, inputting the structured medical entity set as node information into the graph neural network initialization module, and constructing an edge connection matrix between the entities based on the original text co-occurrence relationship, paragraph structure relationship and consistency of diagnosis and treatment stages between the medical entities;

[0118] S62. Extract standard entity nodes, term definition nodes, concept affiliation nodes, and clinical logic relationship edges from the pre-built urology-specific medical knowledge graph, and embed them into the current graph structure to form a knowledge-enhanced medical entity graph;

[0119] S63, inputting the medical entity graph into the graph convolution propagation unit of the graph neural network, adopting a layered propagation strategy, performing graph convolution calculation according to the node topology structure and initial node features, and updating the local neighborhood semantic feature representation of each node;

[0120] S64. Introducing a node semantic attention mechanism in the graph convolution propagation process, weighting adjacent edges by calculating the semantic similarity and context consistency between nodes to guide the direction of feature aggregation;

[0121] S65. Perform residual connection and feature fusion on the node embedding vectors output by each layer of graph convolution to generate a multi-level semantic representation, and perform node vector regularization on the entire graph.

[0122] S66. Combine the entity node semantic representation and edge relationship semantic representation finally output by the graph neural network and encapsulate them to construct an entity semantic graph structure.

[0123] The method standardizes, serializes, and repeats the medical data processing process, covering the entire process of structural analysis, semantic modeling, graph construction, and knowledge reasoning from the initial data collection to the final intelligent reasoning suggestions, thereby improving the controllability of data processing and process consistency. In particular, in terms of data quality control and information restoration, through steps such as format specification, semantic segmentation, and stage labeling, a basic layer of medical data with aligned semantics, clear segmentation, and clear structure is constructed, providing stable input for subsequent entity extraction and logical reasoning. By inputting the identified structured entities into the graph neural network and combining it with the knowledge graph enhancement in the field of urology, the system is able to construct an entity semantic graph with continuous semantic logic and rich entity connections. Through graph convolution propagation and node attention mechanism, the module can deeply encode the contextual relationships between entities, breaking through the previous graph construction method that only relies on syntactic adjacency or regular paths, and constructing a highly expressive and robust semantic structure, providing a strong structural foundation for subsequent path extraction and triple construction.

[0124] In this embodiment, S2 specifically includes:

[0125] S21. Perform main paragraph recognition on the electronic medical record text and divide the text into several diagnosis and treatment semantic blocks, including the chief complaint segment, medical history segment, examination segment, diagnosis segment, treatment segment, and follow-up segment. Use segmentation rules based on position priors and keyword semantic matching to perform text segmentation;

[0126] S22. Perform medical terminology normalization on each diagnosis and treatment semantic block, map non-standard vocabulary, abbreviations, and semantically ambiguous expressions within the diagnosis and treatment semantic block to standard medical terms, use a term mapping dictionary and context-aware rules to jointly determine standard terms, and output a normalized set of diagnosis and treatment text sub-blocks;

[0127] S23. Extract paragraph-level semantic features for each sub-block of medical text, construct a temporal semantic vector sequence based on the context window, and embed structured basic information fields into the vector representation. The basic information fields include the patient's age, gender, duration of the main complaint, admission method, initial diagnosis time, surgical history, and previous diagnosis labels.

[0128] S24. Inputting the semantic vector sequence of the fused structured field into a diagnosis and treatment stage classification model, the diagnosis and treatment stage classification model adopts a multi-channel parallel encoding structure, wherein the first channel processes the text semantic backbone, the second channel processes the structured label features, and a gated fusion mechanism is used to achieve feature interaction, and outputs the diagnosis and treatment stage label corresponding to each text sub-block;

[0129] S25. Performing stage boundary calibration on the predicted diagnosis and treatment stage label sequence, modifying the critical paragraph labels using a reordering mechanism based on local confidence score and context consistency, and generating a diagnosis and treatment stage boundary sequence;

[0130] S26. Bind and encapsulate the diagnosis and treatment text sub-blocks with the corresponding diagnosis and treatment stage labels to construct a stage annotated text set, which serves as the input data set for the subsequent semantic fusion network and named entity recognition module.

[0131] This module effectively overcomes problems such as irregular electronic medical record text structure, significant semantic ambiguity, and information clutter by identifying main paragraphs and dividing them into semantic blocks. By utilizing medical terminology normalization and a phase label mapping mechanism, a clear correspondence is established between diagnostic and treatment statements and structured medical labels, significantly improving the clarity of text semantics and the machine readability of the information. Furthermore, by embedding structured basic information fields into semantic expressions, text entities are given greater contextual adaptability and the ability to express individual patient characteristics, providing a more accurate input foundation for the phase classification model.

[0132] In this embodiment, S3 specifically includes:

[0133] S31. Normalize the text of medical image descriptions, remove non-diagnosis-related tags, noise symbols, and non-structural terms without actual clinical significance, restructure sentence boundaries with disordered formatting, and construct sentence element sequences with standard word order;

[0134] S32, performing a lexical standardization operation on the normalized sentence sequence, using a static vocabulary based on a domain term library and context-related word expansion rules to normalize non-standard words, synonyms, and spelling variations contained in the medical image description into standard medical vocabulary, and replacing the original expressions;

[0135] S33. Use a multi-level syntactic analysis module to perform dependency structure analysis and part-of-speech tagging on the standardized sentence sequence, extract the subject-verb-object structure, modification structure and spatial positioning expression in the sentence, and construct a multi-level syntactic dependency graph;

[0136] S34. Identify anatomical part entities involved in the image description based on the semantic verbs, spatial prepositions, and anatomical lexemes in the multi-level syntactic dependency graph, select noun phrases with anatomical attributes, and perform entity classification;

[0137] S35. Utilize the attention-based lesion word recognition model to analyze the descriptive phrases and determinant word structures within the context of the anatomical site and extract the lesion type entities contained in the image report;

[0138] S36, combining verb structure, sentence-ending assertion pattern and diagnostic term matching rules, locate formative judgment sentences in image descriptions, extract corresponding diagnostic conclusion terms and map them to standard term nodes in the medical diagnostic vocabulary;

[0139] S37. Encapsulate the anatomical part entities, lesion type entities, and diagnostic conclusion terms identified in each image description text into a structured entity sequence in the order of appearance, construct a structured image description sequence, and use it as input content for the subsequent semantic fusion network.

[0140] The Image Description Structuring Module achieves high-fidelity conversion of medical image text from natural language to structured entity sequences through syntactic analysis and lexical standardization. Specifically, it extracts key information such as anatomical site, lesion type, and diagnostic conclusion. By integrating an attention mechanism with a semantic verb localization strategy, it significantly improves the ability to interpret implicit diagnostic information in clinical image descriptions, overcoming the limitations of existing rule-based processing methods, which struggle to balance semantic complexity and structural accuracy.

[0141] In this embodiment, the S4 specifically includes:

[0142] S41. Grouping the stage annotation text set by diagnosis and treatment stage type to establish a first diagnosis stage set, an examination stage set, a diagnosis stage set, a treatment stage set, and a follow-up stage set, thereby forming a stage partition input structure;

[0143] S42. Each stage set is fed into a corresponding stage embedding encoder. The stage embedding encoder includes a diagnosis and treatment stage label embedding layer, a stage position embedding layer, and a stage semantic encoding module. The stage label embedding vector guides sentence element expression and combines relative position information to construct the internal context semantic structure of the stage.

[0144] S43. Perform stage feature extraction on the vector output from the stage semantic encoding module, using a multi-scale convolutional structure and attention filtering mechanism to extract local diagnosis and treatment expressions and global decision intentions, respectively, to form a stage-level semantic representation tensor;

[0145] S44, stacking all stage-level semantic representation tensors in a unified representation space, adjusting the semantic amplitude through a stage-level normalization mechanism, and generating a stage-aggregated text semantic vector sequence;

[0146] S45, inputting the structured image description sequence into the image semantic encoding module, performing entity-level embedding mapping, syntactic position information binding and feature tensor construction, and outputting an image description vector sequence;

[0147] S46. Input the stage-aggregated text semantic vector sequence and the image description vector sequence into the cross-modal cross attention module, and construct an interaction weight matrix using a dual-channel mechanism of text-to-image attention and image-to-text attention;

[0148] S47, based on the calculation result of the interaction weight matrix, perform semantic alignment and spatial mapping, inject the image semantic vector into the stage text semantic space, and generate a fused joint semantic representation;

[0149] S48. Connect the residual connection of the joint semantic representation input and the gate integration structure to complete cross-modal semantic compression and sequence encoding update, and output the final cross-modal joint semantic representation vector sequence for subsequent entity recognition task input.

[0150] This module, through a dual-path fusion mechanism of "stage embedding + cross-attention," achieves the first deep coupling of the semantic structure of the diagnosis and treatment stage with the semantic structure of the image. The stage embedding encoder ensures a clear hierarchy of text semantics across time and stage dimensions, while image semantic encoding complements the entity support of unstructured imaging content. Driven by the cross-attention mechanism, inter-modal information can be complementary integrated and aligned, ultimately outputting a joint semantic representation vector, effectively enhancing the system's medical entity reasoning capabilities and the depth of modal collaborative processing in complex contexts.

[0151] In this embodiment, the S5 specifically includes:

[0152] S51, sending the cross-modal joint semantic representation vector sequence as input to a named entity recognition module, wherein the named entity recognition module includes a context feature extraction submodule, an entity boundary determination submodule, and a label classification submodule;

[0153] S52. In the context feature extraction submodule, a bidirectional gated recurrent unit network is used to perform bidirectional modeling on the input vector sequence, obtain a dynamic representation of each semantic unit in the current context, and form a context semantic feature sequence;

[0154] S53, inputting the contextual semantic feature sequence into the entity boundary determination submodule, constructing an entity candidate window based on the position annotation strategy, detecting the potential entity boundary position through the sliding window mechanism, and generating entity boundary candidate segments;

[0155] S54, jointly inputting the entity boundary candidate segment and the contextual semantic feature sequence into the label classification submodule, using a multi-task parallel structure to jointly discriminate the entity category and label position type, and outputting an entity label sequence, wherein the entity label types include a start label, an intermediate label, and a non-entity label;

[0156] S55. Reverse mapping the cross-modal joint semantic representation vector based on the entity label sequence, extracting the corresponding entity text fragments in the diagnosis and treatment text, and performing vocabulary reconstruction and semantic consistency verification to eliminate pseudo-entity areas that do not conform to the standard vocabulary;

[0157] S56. The finally identified medical entities are grouped by category, and classified into disease name, symptom description, surgical method, anatomical location, examination index and medication items, to construct a structured medical entity set, which serves as the basic input for subsequent knowledge graph modeling and entity relationship extraction.

[0158] By integrating bidirectional context modeling, entity boundary determination, and joint label classification, the named entity recognition module accurately extracts fine-grained and diverse entities from medical texts. The module supports entity boundary awareness and multi-label parallel classification within complex syntactic structures, addressing the limitations of traditional models whose recognition capabilities are limited to terminology dictionaries. Recognition results not only support medical concepts such as disease, examination, and treatment, but can also be broken down into multiple categories, including symptom manifestations, medication items, and examination indicators, effectively supporting the structural integrity of subsequent knowledge mapping and reasoning input.

[0159] In this embodiment, the S7 specifically includes:

[0160] S71. Path sampling initialization is performed based on the semantic representation of the entity nodes and the edge relationship types in the entity semantic graph structure. A depth-first search algorithm is used to generate a set of candidate multi-hop paths from the start node to any target node, each with no more than a preset number of hops. The candidate multi-hop path set consists of a sequence of entity nodes and a sequence of edge relationships, and is accompanied by path indexes and position markers between nodes.

[0161] S72. For each path in the candidate path set, calculate the semantic similarity score and the treatment stage label consistency score between the entity nodes, where the semantic similarity score is based on the weighted accumulation of the context feature similarity between all consecutive nodes in the path, and the treatment stage consistency score is based on the monotonically increasing and semantically consistent score of the treatment stage label sequence bound to the entities in the path;

[0162] S73. Establish a joint path confidence scoring mechanism, fuse the semantic similarity score and the stage consistency score according to the weighted coefficient, generate a path global confidence score vector, set a path pruning threshold interval, and perform dynamic pruning operations on paths below the threshold to eliminate semantically scattered, logically jumping, and disordered diagnosis and treatment stage paths, and obtain a high-confidence path subset;

[0163] S74. The pruned path subsets are layered according to path length to construct a hop-level structure. Multi-path aggregation is performed on all paths in each hop-level layer. Entity representation aggregation, edge relationship aggregation, and position weight adjustment strategies are used to generate a hop-level path representation tensor. Cross-layer attention weighted fusion is performed on path vectors between different hop levels to obtain a multi-scale path semantic feature matrix.

[0164] S75. Input the multi-scale path semantic feature matrix into a relationship type recognition module. The relationship type recognition module uses a multi-task structure including a fully connected classification layer and a graph structure constraint filter to predict the entity relationship type of each path and output the corresponding relationship label and path confidence level.

[0165] S76. Based on the path confidence level and structure matching rules, screen out the diagnosis and treatment paths with closed structure, clearly defined relationships and semantic continuity of entity pairs, extract the starting entity node, intermediate edge relationship and ending entity node in each path, organize and remove duplicates in triple format, and output a structured diagnosis and treatment relationship set as the input entity pair set for the subsequent semantic reasoning and knowledge fusion module.

[0166] The diagnosis and treatment relationship extraction module comprehensively identifies the underlying semantic logic between diagnosis and treatment entities through multi-scale path sampling, confidence fusion, and cross-layer aggregation strategies. Compared to existing triple extraction methods based on sentence-level co-occurrence or rules, this module can perceive the sequence of diagnosis and treatment stages, entity context consistency, and semantic focus, thereby accurately constructing diagnosis and treatment relationship triplets with closed structure and clear clinical logic, significantly improving the system's modeling accuracy and confidence in diagnosis and treatment logic in complex semantic scenarios.

[0167] In this embodiment, S8 specifically includes:

[0168] S81, performing index-level alignment on the triple structure diagnosis and treatment relationship set output in S7 and the stage annotated text set constructed in S2, mapping and matching the diagnosis and treatment stage labels based on the starting position of the entity in the original text, and establishing a binding mapping table between the diagnosis and treatment relationship and the stage label;

[0169] S82. Perform a stage attribution check on the starting entity and the ending entity in each diagnosis and treatment relationship. If the two entities belong to different diagnosis and treatment stages, establish a sequence constraint relationship based on the time sequence of the stage labels and mark the diagnosis and treatment process direction attribute of the relationship path;

[0170] S83. Extract structured basic information fields, including patient age, gender, initial diagnosis time, duration of main complaint, department visited, and family history label, construct the patient's individual attribute vector, and fuse and map it with each entity node in the diagnosis and treatment relationship set as the entity context feature completion input;

[0171] S84. Perform semantic consistency determination on the set of diagnosis and treatment relationships that have been bound to stage information and individual attributes. Calculate confidence weights based on stage continuity, entity consistency, and context matching. Eliminate diagnosis and treatment triples that are logically contradictory or contextually disconnected, and retain high-confidence relationship sets that conform to clinical semantic rules.

[0172] S85. Combine the retained diagnosis and treatment triples with their corresponding stage labels, entity identifiers, individual attribute labels, and relationship direction markers, and uniformly map them to the medical terminology standard system, including ICD, SNOMED CT, UMLS, and the generic drug name coding system. Build a structured diagnosis and treatment field set based on the mapping results.

[0173] S86. Aggregate and encapsulate the triple relationship fields, stage labels and individual attribute codes under each set of standard term codes to construct a structured diagnosis and treatment data unit, and output it in a unified data format, including diagnosis and treatment stage fields, standard entity fields, standard relationship fields, individual feature fields and relationship confidence fields, for downstream diagnosis, recommendation or analysis module calls.

[0174] This module deeply binds diagnosis and treatment relationship sets with stage labels and individual patient attributes, and uniformly maps them to the international medical terminology standard system, achieving a full-process transformation from free text to structured coding fields. Its advantage lies in eliminating medical expression redundancy, semantic ambiguity, and inconsistent terminology, making the system's output data not only structured but also compatible with a wide range of standards, enabling migration to multi-terminal systems, remote consultation platforms, and big data platforms.

[0175] In this embodiment, the S9 specifically includes:

[0176] S91. Modularize and distribute the structured diagnosis and treatment data units, dividing them into a diagnosis input set, a treatment recommendation input set, and a decision reasoning input set according to the data field type, and send them to the intelligent diagnosis module, the personalized treatment recommendation module, and the clinical decision support module respectively;

[0177] S92. In the intelligent diagnosis module, standardized disease entities, symptom manifestations, examination indicators, and stage labels in the diagnosis input set are received and input into a diagnosis classification model built based on a multi-layer feature fusion and semantic alignment mechanism, and the corresponding disease identification label and diagnosis confidence score are output;

[0178] S93. In the personalized treatment recommendation module, the surgical method, drug entity, anatomical location, patient individual attribute vector, and diagnosis and treatment stage information in the treatment recommendation input set are received and input into a treatment matching model based on graph structure path retrieval and treatment effect history inversion mechanism, and the recommended treatment plan sequence and efficacy estimation score are output;

[0179] S94. In the clinical decision support module, the triple structure diagnosis and treatment relationship, stage label sequence, and patient full feature code in the decision reasoning input set are received and input into the reasoning engine including the rule reasoning layer and the neural symbol fusion layer. The module performs diagnosis and treatment pathway consistency verification, risk event simulation, and candidate pathway scoring, and outputs a decision support recommendation set.

[0180] S95. The diagnostic result labels, recommended treatment paths and auxiliary decision-making suggestions output by the three modules are fused and consistency checked to form a final intelligent diagnosis and treatment recommendation report. The report structure includes disease identification results, recommended intervention plans, risk prediction information and phased optional paths, which serves as the output content of the system terminal or remote consultation platform.

[0181] The intelligent diagnosis and treatment decision module integrates structured diagnosis and treatment data into three submodules: diagnosis, recommendation, and reasoning. Through deep model fusion and reasoning path consistency analysis, it outputs clinically interpretable diagnostic labels, personalized treatment pathways, and auxiliary strategy recommendations. Through a systematic module division of labor and a unified output mechanism, it ultimately generates a standardized, semantically clear, and clearly defined diagnosis and treatment recommendation report, fully supporting clinicians in improving decision-making efficiency and intelligence in multiple scenarios, including complex case management, remote consultations, and dynamic monitoring. Example 1:

[0182] To verify the feasibility of this invention, we applied it to the clinical diagnosis and treatment of urology patients. Faced with a vast amount of heterogeneous data resources (such as electronic medical records, structured examination indicators, and image description text), doctors often face difficulties in quickly establishing a complete diagnosis and treatment cognitive chain due to time constraints and data fragmentation. This leads to low diagnostic efficiency, insufficiently personalized treatment plans, and even misdiagnosis or disconnected plans. To address this issue, we proposed an AI-based intelligent processing system for urology diagnosis and treatment data, and verified its effectiveness through clinical simulation pilots conducted in various medical institutions.

[0183] In the actual application scenario, the research team selected real urology data from three comprehensive hospitals of different regions and sizes as experimental samples. Each institution provided the original data of nearly a thousand patients, including structured basic information (such as age, gender, and duration of the chief complaint), original electronic medical records (including the chief complaint, diagnosis, and treatment sections), imaging report texts, and pre- and post-operative diagnosis and treatment records. After the data collection is completed, the system automatically pre-processes the data, unifies the coding format, annotates the text and structured image description sequences in the extraction stage, completes cross-modal semantic fusion and medical entity recognition, and then constructs the entity semantic graph and outputs the diagnosis and treatment relationship triples. Finally, the processing results are encapsulated into structured diagnosis and treatment data units, and respectively input into the diagnosis, recommendation, and reasoning modules for terminal output. The comprehensive performance comparison data of the system of the present invention and the traditional manual decision-making assistance model in the clinical simulation environment are shown in Table 1.

[0184] Table 1: Performance comparison between intelligent diagnosis and treatment system and traditional methods

[0185]

[0186] As shown in the table, the intelligent system outperforms traditional models in all dimensions, achieving significant improvements in key metrics such as diagnosis-treatment relationship modeling, condition identification, and recommendation agreement. Average condition identification accuracy increased from 81.2% with the traditional approach to 93.5%, and treatment agreement increased to 91.1%. Processing time was particularly significant. Traditionally, doctors typically spend over 20 seconds manually reviewing and judging information, while the intelligent system delivers complete diagnosis and treatment recommendations in less than 7 seconds, significantly improving response efficiency. Patient entity recognition accuracy also jumped from 84.6% to 96.3%, providing a solid foundation for the accuracy of subsequent decision-making logic. Anonymous patient satisfaction surveys were also conducted throughout the system implementation period. The results showed that the structured reports generated by the intelligent system are clearer and more readable, facilitating interaction between doctors and patients. The final satisfaction score reached 4.7, significantly higher than the 3.9 with the traditional approach.

[0187] In summary, this embodiment not only verifies that the diagnosis and treatment accuracy and processing efficiency of the system of the present invention are significantly better than traditional methods, but also shows that the system has good practical application potential and scalability, and can effectively improve prominent problems such as the dispersion of multi-source data in urology, difficulty in semantic alignment, and unclear diagnosis and treatment pathways, and has extremely high practical application value.

[0188] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An intelligent processing system for urology diagnosis and treatment data based on artificial intelligence, characterized by: Includes the following modules: The original diagnosis and treatment data collection and preprocessing module is used to collect structured basic information fields, electronic medical record texts, and description texts corresponding to medical images of urology patients; A stage annotation text construction module is used to construct a stage annotation text set based on the electronic medical record text and structured basic information fields; The image description structuring module is used to perform normalization processing on the medical image description text and construct a structured image description sequence; The cross-modal semantic fusion module is used to input the stage annotation text set and the structured image description sequence into the semantic fusion network to extract the cross-modal joint semantic representation vector sequence; The named entity recognition module is used to input the cross-modal joint semantic representation vector sequence into the named entity recognition structure, identify medical entities in the diagnosis and treatment data, and output a structured medical entity set; The entity semantic graph construction module is used to input the structured medical entity set into the graph neural network model, combine it with the urology-specific knowledge graph, and build the entity semantic graph structure based on graph convolution propagation and node semantic attention mechanism; The diagnosis and treatment relationship extraction module is used to construct a high-confidence candidate path set based on the entity semantic graph structure and output a diagnosis and treatment relationship set with a triple structure; The structured encapsulation module is used to bind the diagnosis and treatment relationship set with the stage annotation information, and at the same time integrate the structured basic information fields to construct the patient's individual attribute vector and encapsulate it into a structured diagnosis and treatment data unit; The intelligent diagnosis and treatment decision module is used to input structured diagnosis and treatment data units into the intelligent diagnosis module, personalized treatment recommendation module, and clinical decision support module, respectively performing disease identification, treatment path generation, and reasoning recommendation output, and generating intelligent diagnosis and treatment recommendation reports that can be used on system terminals or remote platforms; The modules are implemented as follows: S1. Collecting original diagnosis and treatment data of urology patients, wherein the original diagnosis and treatment data includes structured basic information fields, electronic medical record text, and description text corresponding to medical images; S2. Perform sentence segmentation, segmentation, and medical terminology normalization on the electronic medical record text, combine the structured basic information fields with the text context, annotate the diagnosis and treatment stage labels, and construct a stage annotated text collection; S3. Perform semantic cleaning, vocabulary standardization, and syntactic structure extraction on the medical image description text to extract the anatomical parts, lesion types, and diagnostic conclusions corresponding to the images, and construct a structured image description sequence; S4. Input the stage annotation text collection and structured image description sequence into the semantic fusion network composed of the integration stage embedding encoder and the cross-modal cross attention mechanism to extract the cross-modal joint semantic representation vector sequence; S5. Input the cross-modal joint semantic representation vector sequence into the named entity recognition module to identify medical entities in the diagnosis and treatment data; S6. Input medical entities into the graph neural network and construct the entity semantic graph structure based on graph convolution propagation and node semantic attention mechanism; S7. Based on the entity semantic graph structure, execute the dynamic semantic graph pruning and multi-scale path integration algorithm to extract the continuous diagnosis and treatment logical relationship between entities and output the diagnosis and treatment relationship set; S8. Bind the diagnosis and treatment relationship set with the stage annotation information, associate the structured basic information fields, and encapsulate them into structured diagnosis and treatment data units according to medical terminology standards; S9, inputting the structured diagnosis and treatment data unit into the intelligent diagnosis module, the personalized treatment recommendation module and the clinical decision support module; The S6 specifically includes: S61, inputting the structured medical entity set as node information into the graph neural network initialization module, and constructing an edge connection matrix between the entities based on the original text co-occurrence relationship, paragraph structure relationship and consistency of diagnosis and treatment stages between the medical entities; S62. Extract standard entity nodes, term definition nodes, concept affiliation nodes, and clinical logic relationship edges from the pre-built urology-specific medical knowledge graph, and embed them into the current graph structure to form a knowledge-enhanced medical entity graph; S63, inputting the medical entity graph into the graph convolution propagation unit of the graph neural network, adopting a layered propagation strategy, performing graph convolution calculation according to the node topology structure and initial node features, and updating the local neighborhood semantic feature representation of each node; S64. Introducing a node semantic attention mechanism in the graph convolution propagation process, weighting adjacent edges by calculating the semantic similarity and context consistency between nodes to guide the direction of feature aggregation; S65. Perform residual connection and feature fusion on the node embedding vectors output by each layer of graph convolution to generate a multi-level semantic representation, and perform node vector regularization on the entire graph. S66. Combine the entity node semantic representation and edge relationship semantic representation finally output by the graph neural network and encapsulate them to construct an entity semantic graph structure.

2. The artificial intelligence-based intelligent processing system for urology diagnosis and treatment data according to claim 1, characterized in that: The S2 specifically includes: S21. Perform main paragraph recognition on the electronic medical record text and divide the text into several diagnosis and treatment semantic blocks, including the chief complaint segment, medical history segment, examination segment, diagnosis segment, treatment segment, and follow-up segment. Use segmentation rules based on position priors and keyword semantic matching to perform text segmentation; S22. Perform medical terminology normalization on each diagnosis and treatment semantic block, map non-standard vocabulary, abbreviations, and semantically ambiguous expressions within the diagnosis and treatment semantic block to standard medical terms, use a term mapping dictionary and context-aware rules to jointly determine standard terms, and output a normalized set of diagnosis and treatment text sub-blocks; S23. Extract paragraph-level semantic features for each sub-block of medical text, construct a temporal semantic vector sequence based on the context window, and embed structured basic information fields into the vector representation. The basic information fields include the patient's age, gender, duration of the main complaint, admission method, initial diagnosis time, surgical history, and previous diagnosis labels. S24. Inputting the semantic vector sequence of the fused structured field into a diagnosis and treatment stage classification model, the diagnosis and treatment stage classification model adopts a multi-channel parallel encoding structure, wherein the first channel processes the text semantic backbone, the second channel processes the structured label features, and a gated fusion mechanism is used to achieve feature interaction, and outputs the diagnosis and treatment stage label corresponding to each text sub-block; S25. Performing stage boundary calibration on the predicted diagnosis and treatment stage label sequence, modifying the critical paragraph labels using a reordering mechanism based on local confidence score and context consistency, and generating a diagnosis and treatment stage boundary sequence; S26. Bind and encapsulate the diagnosis and treatment text sub-blocks with the corresponding diagnosis and treatment stage labels to construct a stage annotated text set.

3. The artificial intelligence-based intelligent processing system for urology diagnosis and treatment data according to claim 1, characterized in that: The S3 specifically includes: S31. Normalize the text of medical image descriptions, remove non-diagnosis-related tags, noise symbols, and non-structural terms without actual clinical significance, restructure sentence boundaries with disordered formatting, and construct sentence element sequences with standard word order; S32, performing a lexical standardization operation on the normalized sentence sequence, using a static vocabulary based on a domain term library and context-related word expansion rules to normalize non-standard words, synonyms, and spelling variations contained in the medical image description into standard medical vocabulary, and replacing the original expressions; S33. Use a multi-level syntactic analysis module to perform dependency structure analysis and part-of-speech tagging on the standardized sentence sequence, extract the subject-verb-object structure, modification structure and spatial positioning expression in the sentence, and construct a multi-level syntactic dependency graph; S34. Identify anatomical part entities involved in the image description based on the semantic verbs, spatial prepositions, and anatomical lexemes in the multi-level syntactic dependency graph, select noun phrases with anatomical attributes, and perform entity classification; S35. Utilize the attention-based lesion word recognition model to analyze the descriptive phrases and determinant word structures within the context of the anatomical site and extract the lesion type entities contained in the image report; S36, combining verb structure, sentence-ending assertion pattern and diagnostic term matching rules, locate formative judgment sentences in image descriptions, extract corresponding diagnostic conclusion terms and map them to standard term nodes in the medical diagnostic vocabulary; S37. Encapsulate the anatomical part entities, lesion type entities, and diagnostic conclusion terms identified in each image description text into a structured entity sequence in the order of appearance, and construct a structured image description sequence.

4. The artificial intelligence-based intelligent processing system for urology diagnosis and treatment data according to claim 1, characterized in that: The S4 specifically includes: S41. Grouping the stage annotation text set by diagnosis and treatment stage type to establish a first diagnosis stage set, an examination stage set, a diagnosis stage set, a treatment stage set, and a follow-up stage set, thereby forming a stage partition input structure; S42. Each stage set is fed into a corresponding stage embedding encoder. The stage embedding encoder includes a diagnosis and treatment stage label embedding layer, a stage position embedding layer, and a stage semantic encoding module. The stage label embedding vector guides sentence element expression and combines relative position information to construct the internal context semantic structure of the stage. S43. Perform stage feature extraction on the vector output from the stage semantic encoding module, using a multi-scale convolutional structure and attention filtering mechanism to extract local diagnosis and treatment expressions and global decision intentions, respectively, to form a stage-level semantic representation tensor; S44, stacking all stage-level semantic representation tensors in a unified representation space, adjusting the semantic amplitude through a stage-level normalization mechanism, and generating a stage-aggregated text semantic vector sequence; S45, inputting the structured image description sequence into the image semantic encoding module, performing entity-level embedding mapping, syntactic position information binding and feature tensor construction, and outputting an image description vector sequence; S46. Input the stage-aggregated text semantic vector sequence and the image description vector sequence into the cross-modal cross attention module, and construct an interaction weight matrix using a dual-channel mechanism of text-to-image attention and image-to-text attention; S47, based on the calculation result of the interaction weight matrix, perform semantic alignment and spatial mapping, inject the image semantic vector into the stage text semantic space, and generate a fused joint semantic representation; S48. The joint semantic representation is input into the residual connection and gate integration structure to complete cross-modal semantic compression and sequence encoding update, and output the final cross-modal joint semantic representation vector sequence.

5. The intelligent processing system for urology diagnosis and treatment data based on artificial intelligence according to claim 1 is characterized in that: The S5 specifically includes: S51, sending the cross-modal joint semantic representation vector sequence as input to a named entity recognition module, wherein the named entity recognition module includes a context feature extraction submodule, an entity boundary determination submodule, and a label classification submodule; S52. In the context feature extraction submodule, a bidirectional gated recurrent unit network is used to perform bidirectional modeling on the input vector sequence, obtain a dynamic representation of each semantic unit in the current context, and form a context semantic feature sequence; S53, inputting the contextual semantic feature sequence into the entity boundary determination submodule, constructing an entity candidate window based on the position annotation strategy, detecting the potential entity boundary position through the sliding window mechanism, and generating entity boundary candidate segments; S54, jointly inputting the entity boundary candidate segment and the contextual semantic feature sequence into the label classification submodule, using a multi-task parallel structure to jointly discriminate the entity category and label position type, and outputting an entity label sequence, wherein the entity label types include a start label, an intermediate label, and a non-entity label; S55. Reverse mapping the cross-modal joint semantic representation vector based on the entity label sequence, extracting the corresponding entity text fragments in the diagnosis and treatment text, and performing vocabulary reconstruction and semantic consistency verification to eliminate pseudo-entity areas that do not conform to the standard vocabulary; S56. The finally identified medical entities are grouped by category, and classified into disease name, symptom description, surgical method, anatomical location, examination index and medication items to construct a structured medical entity set.

6. The artificial intelligence-based intelligent processing system for urology diagnosis and treatment data according to claim 1, characterized in that: The S7 specifically includes: S71. Path sampling initialization is performed based on the semantic representation of the entity nodes and the edge relationship types in the entity semantic graph structure. A depth-first search algorithm is used to generate a set of candidate multi-hop paths from the start node to any target node, each with no more than a preset number of hops. The candidate multi-hop path set consists of a sequence of entity nodes and a sequence of edge relationships, and is accompanied by path indexes and position markers between nodes. S72. For each path in the candidate path set, calculate the semantic similarity score and the treatment stage label consistency score between the entity nodes, where the semantic similarity score is based on the weighted accumulation of the context feature similarity between all consecutive nodes in the path, and the treatment stage consistency score is based on the monotonically increasing and semantically consistent score of the treatment stage label sequence bound to the entities in the path; S73. Establish a joint path confidence scoring mechanism, fuse the semantic similarity score and the stage consistency score according to the weighted coefficient, generate a path global confidence score vector, set a path pruning threshold interval, and perform dynamic pruning operations on paths below the threshold to eliminate semantically scattered, logically jumping, and disordered diagnosis and treatment stage paths, and obtain a high-confidence path subset; S74. The pruned path subsets are layered according to path length to construct a hop-level structure. Multi-path aggregation is performed on all paths in each hop-level layer. Entity representation aggregation, edge relationship aggregation, and position weight adjustment strategies are used to generate a hop-level path representation tensor. Cross-layer attention weighted fusion is performed on path vectors between different hop levels to obtain a multi-scale path semantic feature matrix. S75. Input the multi-scale path semantic feature matrix into a relationship type recognition module. The relationship type recognition module uses a multi-task structure including a fully connected classification layer and a graph structure constraint filter to predict the entity relationship type of each path and output the corresponding relationship label and path confidence level. S76. Based on the path confidence level and structure matching rules, screen out the diagnosis and treatment paths with closed structure, clearly defined relationships and semantic continuity of entities, extract the starting entity node, intermediate edge relationship and ending entity node in each path, organize and remove duplicates in triple format, and output a structured diagnosis and treatment relationship set.

7. The artificial intelligence-based intelligent processing system for urology diagnosis and treatment data according to claim 1, characterized in that: The S8 specifically includes: S81, performing index-level alignment on the triple structure diagnosis and treatment relationship set output in S7 and the stage annotated text set constructed in S2, mapping and matching the diagnosis and treatment stage labels based on the starting position of the entity in the original text, and establishing a binding mapping table between the diagnosis and treatment relationship and the stage label; S82. Perform a stage attribution check on the starting entity and the ending entity in each diagnosis and treatment relationship. If the two entities belong to different diagnosis and treatment stages, establish a sequence constraint relationship based on the time sequence of the stage labels and mark the diagnosis and treatment process direction attribute of the relationship path; S83. Extract structured basic information fields, including patient age, gender, initial diagnosis time, duration of main complaint, department visited, and family history label, construct the patient's individual attribute vector, and fuse and map it with each entity node in the diagnosis and treatment relationship set; S84. Perform semantic consistency determination on the set of diagnosis and treatment relationships that have been bound to stage information and individual attributes. Calculate confidence weights based on stage continuity, entity consistency, and context matching. Eliminate diagnosis and treatment triples that are logically contradictory or contextually disconnected, and retain high-confidence relationship sets that conform to clinical semantic rules. S85. Combine the retained diagnosis and treatment triples with their corresponding stage labels, entity identifiers, individual attribute labels, and relationship direction markers, and uniformly map them to the medical terminology standard system, including ICD, SNOMED CT, UMLS, and the generic drug name coding system. Build a structured diagnosis and treatment field set based on the mapping results. S86. Aggregate and encapsulate the triple relationship fields, stage labels and individual attribute codes under each set of standard term codes to construct a structured diagnosis and treatment data unit, and output it in a unified data format, including diagnosis and treatment stage fields, standard entity fields, standard relationship fields, individual feature fields and relationship confidence fields.

8. The artificial intelligence-based intelligent processing system for urology diagnosis and treatment data according to claim 1, characterized in that: The S9 specifically includes: S91. Modularize and distribute the structured diagnosis and treatment data units, dividing them into a diagnosis input set, a treatment recommendation input set, and a decision reasoning input set according to the data field type, and send them to the intelligent diagnosis module, the personalized treatment recommendation module, and the clinical decision support module respectively; S92. In the intelligent diagnosis module, standardized disease entities, symptom manifestations, examination indicators, and stage labels in the diagnosis input set are received and input into a diagnosis classification model built based on a multi-layer feature fusion and semantic alignment mechanism, and the corresponding disease identification label and diagnosis confidence score are output; S93. In the personalized treatment recommendation module, the surgical method, drug entity, anatomical location, patient individual attribute vector, and diagnosis and treatment stage information in the treatment recommendation input set are received and input into a treatment matching model based on graph structure path retrieval and treatment effect history inversion mechanism, and the recommended treatment plan sequence and efficacy estimation score are output; S94. In the clinical decision support module, the triple structure diagnosis and treatment relationship, stage label sequence, and patient full feature code in the decision reasoning input set are received and input into the reasoning engine including the rule reasoning layer and the neural symbol fusion layer. The module performs diagnosis and treatment pathway consistency verification, risk event simulation, and candidate pathway scoring, and outputs a decision support recommendation set. S95. The diagnostic result labels, recommended treatment paths and auxiliary decision-making suggestions output by the three modules are fused and checked for consistency to form a final intelligent diagnosis and treatment recommendation report. The report structure includes disease identification results, recommended intervention plans, risk prediction information and phased optional paths.

Images