Data processing method, device, medium and program product

Abstract

This application provides a data processing method, device, medium, and program product, relating to the field of artificial intelligence technology. The method includes: acquiring medical text containing operational statements for medical image data; generating input features for a semantic analysis model based on the medical text; generating an ordered sequence of target operation instructions based on the input features using the semantic analysis model, wherein the target operation instruction sequence includes instructions for processing the medical image data; and invoking an execution module to perform corresponding processing tasks on the medical image data based on the target operation instruction sequence. Based on this method, users can complete complex lung watershed analysis tasks by inputting medical text with natural language instructions as needed, without manually switching software tools or system interfaces, reducing operational complexity, minimizing accuracy reduction due to human error, and improving analysis efficiency.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a data processing method, device, medium, and program product. Background Technology

[0002] In clinical practice, doctors or researchers need to perform pulmonary airflow analysis based on a patient's lung medical images to determine the specific location of lung lesions, the extent of lobectomy, or the surgical approach. For example, before lobectomy, doctors need to use image processing techniques to segment pulmonary vessels, bronchi, and lung segment boundaries, identify key anatomical structures (such as the ligation point of the S6 segment of the artery), and calculate relevant parameters (such as the area or volume of the lesion region).

[0003] In existing analysis and processing workflows, doctors need to manually perform multiple steps in medical image processing software, including image annotation, region segmentation, and feature measurement. Each step may require switching tools or interfaces, making the process cumbersome and time-consuming. Furthermore, due to the complex anatomy of the lungs and the extremely high precision requirements of image processing, doctors need extensive professional knowledge and experience to complete high-quality analyses, posing a significant technical barrier for primary healthcare institutions or non-specialist physicians.

[0004] Therefore, there is an urgent need for a method that can process medical image data with greater automation and intelligence to reduce the complexity of the operation during analysis and improve the efficiency and accuracy of analysis, thereby helping users to complete lung basin analysis tasks more efficiently. Summary of the Invention

[0005] This application provides a data processing method, device, medium, and program product for reducing operational complexity and improving analysis efficiency and accuracy in lung watershed analysis.

[0006] In a first aspect, embodiments of this application provide a data processing method, the method comprising: acquiring medical text containing operation statements on medical image data; generating input features for a semantic analysis model based on the medical text; generating an ordered sequence of target operation instructions based on the input features through the semantic analysis model, wherein the sequence of target operation instructions includes instructions for processing the medical image data; and invoking an execution module to perform corresponding processing tasks on the medical image data based on the sequence of target operation instructions.

[0007] In one possible implementation, generating an ordered sequence of target operation instructions based on input features using a semantic analysis model includes: performing semantic analysis on the input features based on a preset medical knowledge base, which stores structured knowledge data related to lung medical scenarios; and generating an ordered sequence of target operation instructions based on the results of the semantic analysis.

[0008] In one possible implementation, the medical knowledge base includes multiple knowledge data, each of which includes vectorized structured knowledge; semantic analysis is performed on the input features based on the preset medical knowledge base, including: retrieving target knowledge data that matches the input features from the multiple knowledge data in the medical knowledge base according to the input features; and performing semantic analysis based on the input features and the target knowledge data to obtain the result of the semantic analysis.

[0009] In one possible implementation, after the execution module is invoked to perform corresponding processing tasks on medical image data based on the target operation instruction sequence, the method further includes: feeding back the execution result of the execution module to the semantic analysis model; and performing execution effect analysis on the execution result through the semantic analysis model, so as to adjust the subsequent instruction generation logic based on the execution effect analysis of the execution result.

[0010] In one possible implementation, the execution module is invoked to perform corresponding processing tasks on the medical image data based on the target operation instruction sequence, including: analyzing whether there are sub-tasks that can be processed independently in the target operation instruction sequence, wherein the sub-tasks that can be processed independently include segmenting different anatomical structures in the medical image data, and / or measuring different lesion regions in the medical image data; if the analysis determines that any sub-task that can be processed independently exists in the target operation instruction sequence, the execution module is invoked to perform the corresponding processing tasks on the medical image data in a parallel execution manner.

[0011] In one possible implementation, the execution module is invoked to perform corresponding processing tasks on the medical image data in a parallel execution manner, including: executing at least one subtask that can be processed independently and obtaining at least one execution result in parallel; determining the execution result identifier of at least one execution result through a version control algorithm, wherein the execution result identifier is used to uniquely identify the execution result; determining the target execution result for result aggregation among the at least one execution result through the result selection rules of the version control algorithm; and performing result aggregation operation according to the execution result identifier of the target execution result to obtain the overall execution result after performing the corresponding processing tasks on the medical image data.

[0012] In one possible implementation, after the execution module is invoked to perform corresponding processing tasks on medical image data based on the target operation instruction sequence, the method further includes: receiving natural language instructions input by the user regarding the execution results of the execution module; adjusting the target operation instruction sequence based on the natural language instructions and generating an adjusted target operation instruction sequence; and / or, based on the natural language instructions, performing analysis on the execution results obtained after the execution module performs the corresponding processing tasks to obtain analysis information.

[0013] Secondly, embodiments of this application provide a data processing apparatus, comprising: an acquisition module for acquiring medical text containing operational statements on medical image data; a generation module for generating input features for a semantic analysis model based on the medical text; the generation module further for generating an ordered sequence of target operation instructions based on the input features through the semantic analysis model, wherein the sequence of target operation instructions includes instructions for processing the medical image data; and an execution module for invoking the execution module to perform corresponding processing tasks on the medical image data based on the sequence of target operation instructions.

[0014] In one possible implementation, the generation module is specifically used to: perform semantic analysis on the input features based on a preset medical knowledge base using a semantic analysis model, wherein the medical knowledge base stores structured knowledge data related to lung medical scenarios; and generate an ordered sequence of target operation instructions based on the results of the semantic analysis.

[0015] In one possible implementation, the medical knowledge base includes multiple knowledge data, each of which includes vectorized structured knowledge; the generation module is specifically used to: retrieve target knowledge data that matches the input features from the multiple knowledge data in the medical knowledge base based on the input features; and perform semantic analysis based on the input features and the target knowledge data to obtain the results of the semantic analysis.

[0016] In one possible implementation, the apparatus further includes a feedback module, which is used to: feed back the execution result of the execution module to the semantic analysis model; and perform execution effect analysis on the execution result through the semantic analysis model, so as to adjust the subsequent instruction generation logic based on the execution effect analysis of the execution result.

[0017] In one possible implementation, the execution module is specifically used to: analyze the target operation instruction sequence to see if there are subtasks that can be processed independently, the subtasks that can be processed independently include segmenting different anatomical structures in the medical image data, and / or measuring different lesion regions in the medical image data; if the analysis determines that there is any subtask that can be processed independently in the target operation instruction sequence, the execution module is called to perform the corresponding processing task on the medical image data in a parallel execution manner.

[0018] In one possible implementation, the execution module is specifically used to: execute at least one subtask that can be processed independently in parallel and obtain at least one execution result; determine the execution result identifier of at least one execution result through a version control algorithm, wherein the execution result identifier is used to uniquely identify the execution result; determine the target execution result for result aggregation among at least one execution result through the result selection rules of the version control algorithm; and perform result aggregation operation according to the execution result identifier of the target execution result to obtain the overall execution result after performing the corresponding processing task on the medical image data.

[0019] In one possible implementation, the execution module is further configured to: receive natural language instructions input by the user regarding the execution result of the execution module; adjust the target operation instruction sequence based on the natural language instructions and generate an adjusted target operation instruction sequence; and / or, based on the natural language instructions, perform analysis on the execution result obtained after the execution module performs the corresponding processing task to obtain analysis information.

[0020] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0021] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0022] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0023] The data processing method, device, medium, and program product provided in this application embodiment parse medical text into a structured sequence of target operation instructions through a semantic analysis model, and drive the execution module to complete the processing task of medical image data, which can efficiently complete the lung watershed analysis task. Based on this method, users can complete complex lung watershed analysis tasks by inputting medical text with natural language instructions as needed, without having to manually switch software tools or system interfaces, significantly reducing the complexity of operation, reducing the decrease in accuracy caused by human operation errors, and improving analysis efficiency by eliminating the need to operate multiple tools or interfaces. Attached Figure Description

[0024] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0025] Figure 1 A flowchart illustrating the data processing method provided in an embodiment of this application;

[0026] Figure 2 A flowchart illustrating the medical data processing method provided in this application embodiment;

[0027] Figure 3 This is a schematic diagram of the architecture of a medical data processing system provided in an embodiment of this application;

[0028] Figure 4 This is a schematic diagram of the execution flow of medical data processing provided in an embodiment of this application;

[0029] Figure 5 A schematic diagram illustrating the process of generating the final report provided in this application embodiment;

[0030] Figure 6 This is a schematic diagram of the interactive interface of the medical data processing system provided in the embodiments of this application;

[0031] Figure 7 This is a schematic diagram of the structure of the data processing apparatus provided in the embodiments of this application;

[0032] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0033] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0034] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0035] It should be noted that the users involved in this application can be understood as users of the methods, systems, and / or devices provided in the embodiments of this application, such as doctors, patients, researchers, etc. The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, and necessary confidentiality measures have been taken. These measures do not violate public order and good morals, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0036] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on personal rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0037] In this application, the terms "exemplarily" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplarily" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the terms "exemplarily" or "for example" is intended to present the relevant concepts in a specific manner.

[0038] In the embodiments of this application, the use of terms such as "first" and "second" is to distinguish between identical or similar items that have essentially the same function and effect. For example, "first electronic device" and "second electronic device" are merely used to distinguish different electronic devices and do not limit their order of execution. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that "first" and "second" do not necessarily imply that they are different.

[0039] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following associated objects have an "or" relationship.

[0040] In existing technologies, lung watershed analysis primarily relies on medical image processing software. Its main workflow includes image uploading and preprocessing, manual annotation and segmentation, feature measurement and lesion localization, and result output and analysis. The following section uses an application scenario where a physician performs lung watershed analysis as an example to introduce the existing technologies.

[0041] For example, in the image upload and preprocessing step, doctors need to upload the patient's lung computed tomography (CT) images to specialized software and perform preprocessing operations such as denoising and enhancement to improve the accuracy of subsequent analysis. In the manual annotation and segmentation step, doctors need to manually annotate key anatomical structures of the lungs (such as bronchi and vascular branches) through the software interface and use segmentation tools to divide lung lobes or segments into regions. This process requires repeated parameter adjustments, making it complex and time-consuming.

[0042] In the feature measurement and lesion localization steps, doctors need to manually measure the size, area, or volume of the lesion region and determine its location by considering anatomical relationships. This step requires a high level of professional knowledge from the doctor and is prone to errors due to subjective judgment. In the results output and analysis steps, the software generates an analysis report based on the above operations, and doctors need to further interpret the results to formulate a treatment plan.

[0043] Analysis of the existing technologies reveals that the current data processing methods suffer from at least the following technical problems.

[0044] Firstly, existing methods suffer from cumbersome operational procedures. For example, doctors need to perform annotation, segmentation, and measurement steps one by one in the software, frequently switching between tools and interfaces, and performing multiple manual operations, which leads to low overall efficiency.

[0045] Secondly, existing methods rely on user experience. For example, key steps such as image segmentation and lesion localization require doctors to manually adjust parameters, which demands a high level of professional skill from the operator, depends on personal experience, is highly subjective, and the results are greatly affected by subjective factors.

[0046] Thirdly, existing methods suffer from low consistency in analytical results. For example, manual operation makes it difficult to guarantee the repeatability and consistency of analytical results, especially when dealing with complex lung structures, where deviations in operation can easily occur in repeated analyses, leading to low consistency in analytical results.

[0047] Fourthly, existing methods suffer from insufficient data processing accuracy. For example, due to the lack of operational stability of hand movements, operational errors may be introduced, especially when dealing with areas with complex anatomical structures (such as lung segment boundaries), resulting in insufficient accuracy in segmentation, measurement, and other processing operations.

[0048] Fifthly, existing methods suffer from low levels of intelligence in data processing. For example, existing processing software or systems cannot directly parse natural language instructions; doctors must input operation parameters and perform operations item by item, lacking an end-to-end (direct output of results after data input) intelligent workflow.

[0049] For example, in the task of "locating the ligation point of the S6 segment of the right lower pulmonary artery," doctors need to manually segment the right lower lung, identify arterial branches, calculate the location of the S6 segment, and then mark the ligation point using a labeling tool. The entire process takes at least tens of minutes, and the results are prone to fluctuation due to differences in operators. It is evident that existing software lacks the ability to automatically parse natural language commands and cannot automatically generate operational instructions and execute tasks based on the doctor's text input (such as "find the ligation point of the S6 segment of the right lower pulmonary artery"). Therefore, the level of intelligent assistance is low.

[0050] Starting with problems identified in existing technologies, the inventors first proposed introducing Natural Language Processing (NLP) into medical image data analysis. First, they identified the processing needs of lung-related medical data. Doctors expect to automate and intelligently perform operations in clinical practice by inputting natural language commands (such as "segment lung lobes" or "measure lesion volume"), but existing systems cannot directly parse such natural language commands.

[0051] Secondly, for natural language instruction processing tasks, a Large Language Model (LLM) can be chosen as an important tool. Its powerful semantic understanding capabilities can transform natural language instructions into structured operation sequences. Furthermore, a modular execution flow is considered, decoupling the LLM from execution modules (such as segmentation, measurement, and annotation procedures). Execution modules are dynamically invoked through an instruction matching mechanism, improving the flexibility of the software system.

[0052] Furthermore, to better adapt to various application scenarios, LLM can be integrated with medical knowledge. For example, for pulmonary medicine scenarios, the LLM can be fine-tuned to understand technical terms (such as "S6 segment artery"), and the semantic analysis capabilities of the model can be enhanced through a knowledge base. Moreover, to achieve end-to-end workflows and reduce manual intervention in the operation process, a fully automated workflow can be built from text input to task execution and result output.

[0053] Based on the above technical concept, a method and system for processing medical image data driven by natural language commands can be developed, which can reduce operational complexity and improve analysis accuracy.

[0054] In view of this, embodiments of this application provide a data processing method that deeply integrates natural language processing technology with a medical image processing module to achieve an automated medical image data analysis process based on natural language commands. Specifically, the method parses user-input medical text (such as operation statements including natural language descriptions) using a large language model, transforms it into a structured sequence of target operation commands, and drives execution modules (such as segmentation, measurement, and annotation program code) to automatically complete medical image processing tasks, ultimately generating analysis results that meet clinical needs. This method combines natural language understanding capabilities with the medical image processing workflow, reducing operational complexity and improving analysis efficiency and accuracy through end-to-end automation.

[0055] The data processing method provided in this application is applicable to scenarios such as lung disease analysis, lung lesion localization, lung surgery planning, and postoperative evaluation. For example, before lobectomy, doctors need to analyze the lung anatomy (such as bronchi and vascular branches) and lesion distribution based on the patient's CT images to determine the surgical scope and key operation points (such as ligation locations). If the existing process is used, doctors need to manually perform multiple steps in the image processing software. However, using the method in this application, the natural language instructions input by the doctor (such as "segment the right lower lung S6 artery and mark the ligation point") can be parsed through a large language model, automatically calling the segmentation, marking, and other modules to complete the task, and finally outputting the processed images and analysis results.

[0056] When implementing the data processing method of the embodiments of this application, it can be implemented through a corresponding data processing system. The data processing system can be deployed on medical platforms such as surgical navigation systems, hospital information systems (HIS), laboratory information management systems (LIS), and picture archiving and communication systems (PACS). It can support linkage with multi-source information such as electronic medical records and laboratory data to form a complete intelligent medical analysis ecosystem.

[0057] The technical solutions of this application will be described in detail below with reference to specific embodiments. The specific embodiments described below can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0058] Figure 1This is a flowchart illustrating the data processing method provided in an embodiment of this application. The executing entity of this method can be an electronic device with corresponding data storage and computing capabilities, such as a computer, server, or server cluster. Figure 1 As shown, the method includes:

[0059] S101, Obtain medical text containing operation statements for medical image data.

[0060] For example, medical text can be understood as natural language text used to describe operations performed on medical imaging data, such as "segmenting the S6 artery and marking the ligation point in a CT image of the right lower lung." Medical imaging data can be understood as image data of body parts acquired through medical imaging technology, including at least lung imaging data. Examples of medical imaging data include CT medical imaging data, magnetic resonance imaging (MRI) medical imaging data, etc.

[0061] S102, based on medical text, generates input features for a semantic analysis model.

[0062] For example, a semantic analysis model can be understood as an algorithmic model capable of parsing natural language and generating structured instructions, such as large language models based on a Transformer architecture. Input features can be understood as encoding medical text into feature tensors that the model can process, such as high-dimensional vectors generated through word embeddings or one-hot encoding.

[0063] Taking the semantic analysis model as an example of a large language model including the Transformer architecture, the core principle of the large language model to realize natural language semantic understanding and analysis is based on pre-training learning and context-dependent modeling of massive text data. For example, it can be summarized as the attention mechanism of the Transformer architecture, unsupervised pre-training of massive text data, and fine-tuning and prompt learning.

[0064] The Transformer architecture's attention mechanism can capture long-distance dependencies between words and sentences in input text (such as medical texts) through self-attention, accurately identifying semantic associations (such as pronoun reference, modification, and logical relationships). Through unsupervised pre-training on massive amounts of text, it can learn grammatical rules, semantic knowledge, common-sense logic, and contextual patterns, forming a distributed representation of natural language (word embeddings / sentence embeddings), transforming text into computable vectors (such as input features). Due to fine-tuning and cue-based learning, after pre-training, the model can be fine-tuned using specific task data, or directly guided by cue words to align with specific semantic analysis needs. This allows the semantic analysis model to perform tasks such as semantic similarity calculation and intent recognition based on learned knowledge.

[0065] For example, a large language model can include several stacked Transformer modules, such as a Transformer encoder-decoder architecture or a decoder-only architecture based on Transformer. Its semantic understanding capabilities rely on a hierarchical module design. For instance, a token embedding layer can break down the input natural language text into the smallest unit, tokens, and then map each token to a fixed-dimensional vector (word embedding vector). Simultaneously, positional encoding vectors are superimposed, allowing the large language model to perceive the sequential position of tokens in the input text, providing a foundation for subsequent semantic modeling.

[0066] The Transformer module can also include a Multi-Head Self-Attention Layer, which can be used to compute in parallel with multiple attention heads. Each attention head assigns different weights to different tokens in the text, captures the semantic relationships between tokens (such as subject-verb-object relationships, referential relationships, and long-distance dependencies), and outputs a feature vector that fuses global semantics.

[0067] The Transformer module can also include a Feed-Forward Network (FFN) layer, which typically consists of two fully connected layers and activation functions (such as ReLU and GELU). It can perform position-by-position nonlinear transformations on the feature vector output by the self-attention layer, further extracting and enhancing the semantic features of the token, and improving the model's ability to fit complex semantics.

[0068] For multiple stacked Transformer modules, the outputs of the multi-head self-attention layer and feedforward neural network layer of the previous Transformer module can be combined and input into the next Transformer module for processing. After the feature vectors are processed by dozens or even more stacked Transformer modules, the basic grammar and lexical semantics of the text can be learned through the low-level network, while the higher-level network captures more abstract contextual logic and common-sense connections, thereby achieving semantic understanding from shallow to deep.

[0069] Large language models can also include layer normalization and residual connections. Layer normalization is added after each attention layer and feedforward layer to stabilize the parameter distribution during model training. Residual connections directly add the input and output of the module to solve the gradient vanishing problem in deep network training and ensure that the model can still effectively learn semantic information in deep structures.

[0070] Large language models can also include an output layer, which can map the feature vectors output by the last Transformer module back to the token space and calculate the probability distribution of the next token through functions such as the normalized exponential function (Softmax). In semantic analysis tasks, the output can also be mapped to task labels (such as text semantics and text intent) through fine-tuning.

[0071] S103, an ordered sequence of target operation instructions is generated based on input features using a semantic analysis model, wherein the sequence of target operation instructions includes instructions for processing medical image data.

[0072] For example, the target operation instruction sequence can be understood as an ordered set of instructions generated by a semantic analysis model to drive the medical image processing program. The target operation instruction sequence may include one or more operation instructions. For example, the target operation instruction sequence is "right lower lung segmentation → artery segmentation → S6 segment identification → generation of ligation point location", where "right lower lung segmentation" and "artery segmentation" are operation instructions.

[0073] In one possible implementation, generating an ordered sequence of target operation instructions based on input features using a semantic analysis model includes: performing semantic analysis on the input features based on a preset medical knowledge base, which stores structured knowledge data related to lung medical scenarios; and generating an ordered sequence of target operation instructions based on the results of the semantic analysis.

[0074] For example, a medical knowledge base can be understood as a dataset storing structured knowledge related to pulmonary medicine scenarios. This structured knowledge may include experiential data on pulmonary medical image processing, subject knowledge data related to pulmonary medicine scenarios, and medical case data related to pulmonary medicine, such as hierarchical relationships of pulmonary anatomical structures and mapping rules between operational instructions and anatomical structures. For example, the medical knowledge base may contain rules such as "the S6 segment artery belongs to a substructure of the right lower lung" and "arterial segmentation should take precedence over ligation point identification."

[0075] After the semantic analysis model parses the medical text to generate input features, the input features can be semantically analyzed by calling a preset medical knowledge base.

[0076] In one possible implementation, the medical knowledge base includes multiple knowledge data, each of which includes vectorized structured knowledge; semantic analysis is performed on the input features based on the preset medical knowledge base, including: retrieving target knowledge data that matches the input features from the multiple knowledge data in the medical knowledge base according to the input features; and performing semantic analysis based on the input features and the target knowledge data to obtain the result of the semantic analysis.

[0077] For example, knowledge data can be understood as data obtained by vectorizing structured knowledge. For instance, vectorizing a sentence or a passage of empirical data on lung medical image processing yields a semantic feature vector, which can be considered a piece of knowledge data. Similarly, vectorizing a document related to a lung medical scenario can yield several pieces of knowledge data. Vectorization can be achieved through feature vector extraction or vector space mapping.

[0078] The process of retrieving and matching input features and knowledge data to determine target knowledge data can be achieved, for example, through similarity calculation and combining the calculated similarity with preset filtering rules.

[0079] For example, structured knowledge related to lung-related medical scenarios in a medical knowledge base can be vectorized and stored as semantic feature vectors. After generating input features, the similarity between the input features and the semantic feature vectors in the medical knowledge base can be calculated. Based on this similarity, several target knowledge data related to lung-related medical scenarios with relatively high semantic matching degree with the current input features can be retrieved from the medical knowledge base.

[0080] For example, by calculating the cosine similarity between the input features and the semantic feature vectors of each structured knowledge, several structured knowledge points with a cosine similarity greater than a preset similarity threshold are identified as target knowledge data, or several structured knowledge points with the highest cosine similarity are identified as target knowledge data. Then, the semantic feature vectors of the target knowledge data related to lung medical scenarios retrieved from the medical knowledge base are concatenated with the input features, thus incorporating the structured knowledge related to lung medical scenarios stored in the medical knowledge base as context for the medical text into the large language model.

[0081] It should be understood that other similarity calculation methods, such as Euclidean distance and Manhattan distance, can also be used when calculating the similarity between the semantic feature vector of knowledge data and the input feature. This application embodiment does not limit this. The number of target knowledge data to be matched can also be determined according to the needs of the actual application scenario, and is not limited to matching only one target knowledge data for one input feature.

[0082] The method based on the embodiments of this application can quickly and accurately match target knowledge data for input features in a medical knowledge base, so that when generating an ordered sequence of target operation instructions through a semantic analysis model, it can combine richer contextual information rather than one-sided or isolated input information, thereby obtaining a sequence of target operation instructions that is more in line with the actual intent of the medical text and improving the accuracy of the target operation instruction sequence.

[0083] For example, when medical text contains "S6 segment artery," the semantic analysis model queries a medical knowledge base to confirm its anatomical hierarchy. Combining this with the rule that "artery segmentation takes precedence over ligation point identification," it generates a target operation instruction sequence: "Right lower lung segmentation → Artery segmentation → S6 segment identification → Generation of ligation point location." This sequence includes the order of operations, forming an ordered sequence. Therefore, the introduction of a medical knowledge base can improve the consistency between semantic analysis results and pulmonary medical logic, avoiding ambiguity in terminology analysis due to "unprofessional" semantic analysis, and thus preventing the generation of incorrect target operation instruction sequences such as sequence errors or instruction errors.

[0084] In this embodiment, structured knowledge data from a medical knowledge base can be used to expand input features, thereby enhancing the accuracy and professionalism of the semantic analysis model in analyzing input features. For example, when a medical text describes "locating the ligation point of the S6 artery," the semantic analysis model can not only parse the term "S6 artery" in the medical text, but also add structured knowledge data related to "S6 artery" through the medical knowledge base. This adds contextual information to aid understanding of "S6 artery," enabling the semantic analysis model to correctly understand that "S6 artery" belongs to a substructure of the right lower lung, and thus accurately match the corresponding segmentation and annotation instructions. This semantic enhancement mechanism based on a medical knowledge base can improve the matching accuracy of complex terms and operational logic, ensuring that the generated target operation instruction sequence conforms to the anatomical and operational norms of pulmonary medical scenarios.

[0085] S104, based on the target operation instruction sequence, calls the execution module to perform corresponding processing tasks on the medical image data.

[0086] For example, an execution module can be understood as a set of program code for performing medical image processing tasks. An execution module includes at least modules for performing watershed analysis on lung image data, such as a segmentation module, a labeling module, a measurement module, etc.

[0087] The method in this application embodiment inputs medical text into a semantic analysis model, generates input features for the model to parse, and then generates a target operation instruction sequence.

[0088] For example, when the input medical text is "Segment the S6 artery in the right lower lung CT image and mark the ligation point," the semantic analysis model will parse the operation instructions in the text and generate a target operation instruction sequence of "right lower lung segmentation → artery segmentation → S6 segment identification → generation of ligation point location." Subsequently, the execution module will sequentially call the program code of the segmentation module, annotation module, etc., according to the target operation instruction sequence to process the medical image data.

[0089] In the process of the method in this application embodiment, medical text is used as the input source, input features are used as the processing object of the semantic analysis model, and the target operation instruction sequence is used as the scheduling basis of the execution module. In the end, intelligent and automated end-to-end processing of medical image data is realized, and the degree of human intervention in the overall execution process is effectively reduced.

[0090] The data processing method provided in this application uses a semantic analysis model to parse medical text into a structured sequence of target operation instructions, and drives the execution module to complete the processing task of medical image data, thereby solving the problem in the prior art that multiple operation steps need to be manually executed to complete data processing.

[0091] For example, the semantic analysis model generates input features based on medical text and combines them with lung-related medical knowledge stored in a medical knowledge base. The model analyzes the operational instructions required to complete the operation statements in the medical text (such as "segmenting the S6 artery") and constructs a target operation instruction sequence, ensuring that the generated sequence matches the order of the actual medical image data analysis operations. For instance, when inputting the medical text "annotate the ligation point of the S6 artery," the semantic analysis model matches the target operation instruction sequence "right lower lung segmentation → artery segmentation → S6 segment identification → generation of ligation point location," and calls the segmentation and annotation modules to sequentially perform the segmentation and annotation processing tasks.

[0092] Therefore, using the method of this application embodiment, users can complete complex lung watershed analysis tasks by inputting medical text in natural language as needed, without manually switching software tools or system interfaces. This significantly reduces operational complexity, minimizes accuracy loss due to human error, and improves analysis efficiency by eliminating the need to operate multiple tools or interfaces. Furthermore, by fine-tuning the semantic analysis model, it can adapt to terminology in different medical scenarios (such as "S6 segment artery"), and can be extended to other medical analysis scenarios beyond lung watershed analysis, further improving the accuracy, efficiency, and adaptability of data processing.

[0093] In practical applications, there may be ambiguities in the mapping relationship between medical terminology and anatomical structures in medical images. To address this issue, the method provided in this application can also resolve the problem by querying a medical knowledge graph in a medical knowledge base to check for associations and / or eliminate ambiguities.

[0094] In one possible implementation, the medical knowledge base includes a medical knowledge graph. Semantic analysis of input features is performed based on the preset medical knowledge base, including: during the semantic analysis of input features based on the preset medical knowledge base, checking the correlation between the anatomical structures described in the medical text and the target operation instruction sequence by querying the medical knowledge graph, and / or, performing polysemy elimination operation on the terms in the medical text based on the medical knowledge graph.

[0095] For example, a medical knowledge graph can be constructed in a medical knowledge base. Structured knowledge such as lung anatomy (e.g., the right lower lung and the S6 artery) and its hierarchical relationships and operational logic (e.g., "the S6 artery is a substructure of the right lower lung") can be built into a highly professional knowledge network to form a medical knowledge graph.

[0096] Ambiguity elimination can be understood as using a medical knowledge base to determine the precise meaning of potentially ambiguous terms in medical texts. For example, "S6 segment" in a medical text may refer to different anatomical structures, but the knowledge base confirms that it specifically refers to "S6 segment of the right lower pulmonary artery".

[0097] In the process of semantic analysis of input features based on a pre-set medical knowledge base, the semantic analysis model can process at least one of two aspects through the medical knowledge base.

[0098] The first aspect of these two aspects involves verifying the correlation between the anatomical structures described in the medical text and the target operation instruction sequence. For example, verifying that "S6 segment artery" matches the instruction "artery segmentation → S6 segment identification". The second aspect involves performing semantic ambiguity analysis using knowledge stored in a medical knowledge base and performing operations to eliminate terminological ambiguity, such as confirming that "S6 segment" specifically refers to "right lower pulmonary artery S6 segment". For example, when inputting "locate S6 segment", the semantic analysis model uses the medical knowledge base to determine that it may point to different anatomical structures and selects the most matching instruction sequence based on the context.

[0099] During the instruction matching process, the semantic analysis model can use a medical knowledge graph to verify the medical rationality of any instruction in the target operation instruction sequence. For example, when parsing "annotate the S6 segment artery", the semantic analysis model can query the medical knowledge graph to confirm that "S6 segment artery" belongs to a substructure of the right lower lung and match the corresponding segmentation and annotation instructions, thus achieving semantically enhanced matching.

[0100] Furthermore, logical reasoning can be supported when verifying the relationship between anatomical structures described in medical texts and the sequence of target operation instructions. The semantic analysis model can perform logical reasoning based on medical knowledge graphs. For example, when parsing "calculate the volume ratio of the S6 artery segment to the surrounding tissues", it can automatically call the volume calculation module and associate the segmentation results of adjacent anatomical structures.

[0101] The introduction of medical knowledge graphs can enhance the semantic understanding of technical terms and operational logic in semantic analysis models. For example, when a doctor inputs "locate the ligation point of the S6 artery," the semantic analysis model can use the medical knowledge graph to confirm that the "S6 artery" belongs to a substructure of the right lower lung and match the target operation instruction sequence of "right lower lung segmentation → artery segmentation → S6 segment identification → generation of ligation point location." This semantically enhanced matching mechanism based on medical knowledge graphs effectively avoids instruction matching errors caused by terminological ambiguity (such as "S6 segment" potentially referring to different anatomical structures). Furthermore, supporting logical reasoning allows the method to handle more complex combined instructions (such as analysis tasks involving multiple anatomical structures), further enhancing the method's adaptability and intelligence.

[0102] In the embodiments of this application, medical knowledge graphs can be used to verify relationships and / or eliminate ambiguity, improve the accuracy of terminology parsing, reduce the probability of instruction matching errors caused by terminology ambiguity, and improve the consistency between the generated target operation instruction sequence and actual medical needs.

[0103] In practical applications, there may be a problem where the generated target operation instruction sequence needs to dynamically adapt to the quality differences in medical image data (such as low-resolution slices and blurred areas) during execution. To address this issue, the method provided in this application embodiment can also effectively solve this problem through an execution result feedback mechanism, a dynamic instruction adjustment strategy, and an adaptive execution process.

[0104] The execution result feedback mechanism can be understood as the real-time feedback of the operation result to the semantic analysis model after the execution module completes the operation. The dynamic instruction adjustment strategy can be understood as adjusting the matching strategy of subsequent instructions based on the feedback results. The adaptive execution process can be understood as the model dynamically adjusting the priority and execution order of the instruction sequence based on the feedback data of historical execution results (such as success rate and error rate) through a reinforcement learning mechanism, prioritizing the processing of high-confidence operation instructions.

[0105] In one possible implementation, after the execution module is invoked to perform corresponding processing tasks on medical image data based on the target operation instruction sequence, the method further includes: feeding back the execution result of the execution module to the semantic analysis model; and performing execution effect analysis on the execution result through the semantic analysis model, so as to adjust the subsequent instruction generation logic based on the execution effect analysis of the execution result.

[0106] For example, the execution result can be understood as the processing output of the execution module after completing the operation instructions, such as a segmentation mask, measurement values, and annotation positions. For example, the segmentation mask of the lower right lung output by the segmentation module and the coordinates of the ligation point position generated by the annotation module.

[0107] After the execution module completes any instruction in the target operation instruction sequence, it can feed back the corresponding execution result (such as a segmentation mask or measurement value) to the semantic analysis model in real time, or it can record key features (such as the integrity of the segmented region and the clarity of its boundaries). For example, if the segmentation result contains regions with ambiguous boundaries, the semantic analysis model can adjust the matching strategy for subsequent unexecuted instructions based on this feedback, such as inserting an "optimize segmentation parameters" instruction or switching to a higher-precision segmentation algorithm. This closed-loop feedback mechanism can improve the dynamic optimization of the target operation instruction sequence and adapt to the quality differences in medical image data.

[0108] The dynamic instruction adjustment strategy monitors execution results in real time and feeds them back to the model, enabling it to dynamically optimize the instruction sequence based on the quality of medical image data. For example, in low-resolution CT images, if the boundaries of the initial segmentation result are blurred, the model can automatically insert instructions to optimize segmentation parameters and call a higher-precision segmentation algorithm for reprocessing. This adaptive mechanism improves the robustness of the method in complex image scenarios and prevents the entire analysis process from being interrupted due to a single operation failure.

[0109] When adjusting the subsequent instruction generation logic based on the analysis of the execution results, an adaptive execution process can be combined with reinforcement learning to optimize the instruction execution order, so that key operations can be completed first when processing multi-step tasks, thereby improving the overall analysis efficiency and result consistency.

[0110] In this embodiment, feedback of execution results can improve the robustness of data processing. For example, in low-resolution CT images, if the initial segmentation result is unsatisfactory, the semantic analysis model can automatically adjust subsequent instructions, prioritizing the execution of operations that optimize segmentation parameters. This dynamic adjustment strategy can reduce instruction matching errors caused by differences in image quality, and improve the completeness of the analysis process in complex scenarios.

[0111] In practical applications, there may be a need to shorten the total processing time and make full use of computing resources when handling multi-step lung watershed analysis tasks. The method provided in this application embodiment can also meet this need through subtask analysis and parallel execution.

[0112] In one possible implementation, the execution module is invoked to perform corresponding processing tasks on the medical image data based on the target operation instruction sequence, including: analyzing whether there are sub-tasks that can be processed independently in the target operation instruction sequence, wherein the sub-tasks that can be processed independently include segmenting different anatomical structures in the medical image data, and / or measuring different lesion regions in the medical image data; if the analysis determines that any sub-task that can be processed independently exists in the target operation instruction sequence, the execution module is invoked to perform the corresponding processing tasks on the medical image data in a parallel execution manner.

[0113] For example, a subtask that can be processed independently can be understood as a medical image data processing unit operation that can be executed independently. For instance, the task of segmenting the right lower lung and the left upper lung can be understood as two subtasks that can be processed independently; similarly, the task of measuring the volume of two different lesion regions can also be understood as two subtasks that can be processed independently.

[0114] When invoking the execution module, the task dependencies between multiple tasks in the target operation instruction sequence can be analyzed first. If there are at least two subtasks that can be processed independently, the subtasks can be executed in parallel. Task dependency analysis can be understood as resolving the task dependencies in the target operation instruction sequence through a semantic analysis model (e.g., "segmenting lung lobes" is a prerequisite task for "calculating lesion volume"). In addition, a task map can be generated based on each subtask.

[0115] When executing subtasks that can be processed independently, a parallel scheduling strategy can be employed. A parallel scheduling strategy can be understood as allocating tasks that can be executed in parallel (such as segmenting different lung lobes or calculating the volume of multiple lesions) to independent execution threads or computing units, with a scheduler coordinating resource allocation. Computing units can be, for example, a Central Processing Unit (CPU) or a Graphics Processing Unit (GPU), or a cluster of both.

[0116] For example, if the target operation instruction sequence contains "right lower lung segmentation → left upper lung segmentation → lesion A measurement → lesion B measurement", then the semantic analysis model can determine that there are multiple subtasks that can be processed independently. The segmentation tasks of the right lower lung and the left upper lung can be executed in parallel, as can the measurement tasks of lesion A and lesion B. This parallel processing strategy can make full use of computing resources and shorten the overall analysis time.

[0117] Parallel execution architecture, through task dependency analysis and parallel scheduling strategies, can significantly reduce the total time required for complex lung watershed analysis processes. For example, when analyzing multiple lung lobes or lesions, the system can perform operations such as annotation, segmentation, and measurement in parallel, rather than waiting to execute them sequentially. This approach implements a resource-optimized scheduling mechanism, which can fully utilize computing resources and avoid resource idleness caused by sequential task execution.

[0118] After executing subtasks in parallel, a result aggregation mechanism can be used to combine the intermediate results (such as segmentation masks and measurements) from each subtask into a final overall execution result for unified display. Version control of the software system can also be used to avoid conflicts during aggregation. By using a result aggregation mechanism and version control, data consistency can be ensured, conflicts caused by parallel processing can be avoided, thereby improving efficiency while guaranteeing the accuracy of the analysis results.

[0119] In this embodiment of the application, by analyzing the target operation instruction sequence to see if there are subtasks that can be processed independently, and by adopting a parallel execution strategy to schedule and execute multiple subtasks in parallel in a timely manner, the total time consumption of the complex analysis process is shortened, resource idleness caused by serial execution is avoided, and analysis efficiency is improved.

[0120] For example, in some scenarios, parallel execution can allow for the independent processing of subtasks, or it can involve executing the same subtask multiple times in parallel. For instance, executing the operation instruction "measure the area of ​​lesion A" three times in parallel can yield three execution results.

[0121] In one possible implementation, the execution module is invoked in a parallel execution manner to perform corresponding processing tasks on the medical image data, including:

[0122] At least one subtask that can be processed independently is executed in parallel and at least one execution result is obtained. An execution result identifier is determined for at least one execution result through a version control algorithm. The execution result identifier is used to uniquely identify the execution result. The result selection rules of the version control algorithm are used to determine the target execution result for result aggregation from at least one execution result. The result aggregation operation is performed according to the execution result identifier of the target execution result to obtain the overall execution result after performing the corresponding processing tasks on the medical image data.

[0123] For example, a version control algorithm can be understood as an algorithm that performs version management and result selection on the execution results of subtasks that are allowed to be processed independently in parallel processing.

[0124] The execution result identifier includes identification information used to distinguish each execution result. This identification information can be, for example, a globally unique timestamp, a unique identifier, or a unique graphic identifier. The execution result identifier distinguishes each execution result; in other words, it identifies which execution version each result was obtained from.

[0125] Result selection rules can be understood as rules for filtering and determining multiple execution results. For example, preset result selection rules can be used to determine the target execution result from multiple parallel execution results. These rules might include selecting the latest-generated execution result as the target result based on the order of its generation; or selecting the execution result with the highest quality evaluation score as the target result; or determining the execution result as the target result when a subtask that allows independent processing has only one execution result.

[0126] When executing one or more subtasks that can be processed independently in parallel, a version control algorithm can assign a globally unique execution result identifier to each execution result obtained in parallel. This identifier allows for accurate identification of the target execution result. For example, when updating the segmentation mask of the same anatomical structure simultaneously in parallel, multiple segmentation mask results can be obtained. A version control algorithm can assign a globally unique execution result identifier to each segmentation mask result.

[0127] When aggregating the execution results of all subtasks, the execution result of each subtask can be selected as the target execution result based on the result selection rules of the version control algorithm. Aggregating these target execution results yields the overall execution result, which represents a comprehensive result for multiple parallel subtasks that are allowed to be processed independently. This approach provides a highly consistent and comprehensive overall execution result, avoids data conflicts caused by multiple versions processing tasks in parallel, and ensures the accuracy of the analysis results.

[0128] In this embodiment, a version control algorithm can improve the consistency of execution results in parallel processing. For example, when multiple subtasks simultaneously modify the segmentation results of the same anatomical structure, the version control algorithm can select the latest version of the execution result as the target execution result, avoiding data conflicts. This improves execution efficiency while ensuring the integrity and reliability of the execution results.

[0129] In practical applications, if users are not satisfied with the results of intermediate operations, it is necessary to quickly correct the instructions and adjust the subsequent task execution process. In addition, if users expect to conduct a series of related analyses on the execution results, it is necessary to receive and understand the natural language instructions input by the users in order to conduct the corresponding analyses and obtain analytical information.

[0130] In one possible implementation, after the execution module is invoked to perform corresponding processing tasks on medical image data based on the target operation instruction sequence, the method further includes: receiving natural language instructions input by the user regarding the execution results of the execution module; adjusting the target operation instruction sequence based on the natural language instructions and generating an adjusted target operation instruction sequence; and / or, based on the natural language instructions, performing analysis on the execution results obtained after the execution module performs the corresponding processing tasks to obtain analysis information.

[0131] For example, natural language instructions can be understood as natural language content provided by the user through text input. Based on the natural language instructions, the sequence of target operation instructions can be adjusted, and related content analysis can be performed based on the execution results. For example, the user inputs "move the boundary of the S6 artery segment 10 pixels to the left" to adjust the segmentation result.

[0132] For example, a user inputs a natural language command: "If this lung segment is removed, how much of the remaining lung function is expected to be lost?" By understanding this natural language command, it can be clear that the user expects to conduct relevant analysis based on the execution result. Then, the corresponding analytical information can be obtained through semantic understanding and medical analysis. For example, the obtained analytical information could be: "Based on watershed analysis calculations, the removal of segment S3 is expected to result in a loss of approximately 12% of lung function. Risk assessment: low risk."

[0133] After the execution module executes a portion of the operation instructions in the target operation instruction sequence, it obtains the corresponding execution results. The user can then input natural language instructions (such as "adjust boundaries") based on these results. The electronic device receives the user's input natural language instructions and can generate an adjusted target operation instruction sequence based on the instructions.

[0134] For example, after completing any operation command, the execution module can provide intermediate results (such as segmented regions and labeled positions) in a visual format to the user interface, enabling real-time display of the execution results. Visualization formats include highlighting and visualization of 3D (three-dimensional) models.

[0135] After viewing the displayed execution results, users can provide natural language commands by entering text or directly interacting with the displayed execution results. The semantic analysis model parses the text entered by the user and the user's interaction with the execution results, and matches the corresponding operation commands based on the parsing results. The target operation command sequence is updated by adding the matched operation commands to the end of the target operation command sequence.

[0136] For example, after a user inputs a natural language command such as "move the boundary of the S6 artery segment 10 pixels to the left," the semantic analysis model can generate a "call the boundary optimization module" command and update the subsequent operation flow in the target operation command sequence. As another example, a user can zoom in or out on the image in the displayed execution result by scrolling the mouse or typing a shortcut key; the semantic analysis model can directly parse the user's interactive operation and adjust the target operation command sequence accordingly. Furthermore, after detecting a user's interactive operation on the displayed execution result, the model can automatically generate corresponding operation logic text as input to the semantic analysis model and adjust the target operation command sequence accordingly. This interactive mechanism allows users to quickly correct the analysis results without re-executing the entire process.

[0137] During the adjustment process, a context memory mechanism can be used. The semantic analysis model records historical operations and natural language instructions, and the context information is retained through the memory module (such as the caching mechanism of Transformer), which can improve the coherence of subsequent instruction matching.

[0138] The dynamic interaction mechanism significantly enhances doctors' control over the analysis process through real-time visual feedback and natural language commands. For example, when segmentation results are unsatisfactory, doctors can quickly correct them using simple text commands (such as "adjust boundaries") without having to re-enter the entire procedure. Contextual memory ensures seamless integration of corrected commands with historical operations, preventing process interruptions due to corrections. This interactive adjustment mechanism not only improves operational flexibility but also reduces errors caused by model misjudgments, thereby improving accuracy while maintaining analytical efficiency.

[0139] In this embodiment, natural language commands enhance operational flexibility, enrich system functionality, and improve intelligence. For example, when segmentation results are unsatisfactory, users can quickly correct them using simple natural language commands, automatically generating an adjusted sequence of target operation commands and updating the analysis process. This reduces erroneous operations caused by misjudgments in the semantic analysis model and improves the user's control over the analysis process. Furthermore, when further medical analysis is required, natural language commands can be input to conduct corresponding extended analyses based on the execution results, yielding richer and more comprehensive information.

[0140] In one possible implementation, when generating the adjusted target operation instruction sequence, the association between the adjusted target operation instruction sequence and the original target operation instruction sequence can also be preserved based on the context information of the historical operation records.

[0141] For example, contextual information can be understood as the associated data in the historical operation record related to the current adjustment instruction. For example, the relationship between the target operation instruction sequence before adjustment "right lower lung segmentation → artery segmentation → S6 segment identification" and the target operation instruction sequence after adjustment "right lower lung segmentation → artery segmentation → optimization of segmentation parameters → S6 segment identification".

[0142] For example, after a user corrects the "boundary of the S6 artery segment," the system records the association between this adjustment command and the "S6 segment identification" step in the original command. This allows subsequent operations (such as ligation point generation) to be performed based on the corrected boundary data. This mechanism ensures the consistency between the adjusted command and the original process.

[0143] In this embodiment, a context information retention mechanism ensures the continuity between the adjusted target operation instruction sequence and the original target operation instruction sequence. For example, after a user corrects the segmentation boundary, the input data for subsequent operations (such as ligation point generation) can be updated, avoiding interruption of the execution process due to the adjustment, improving the efficiency and accuracy of interactive correction, and forming a complete intelligent medical analysis ecosystem.

[0144] In one possible implementation, the method provided in this application can also achieve context-aware instruction parsing based on multimodal input. For example, based on the aforementioned medical text input, medical image data can be used as auxiliary input, and the image features of text and image can be jointly analyzed through a multimodal fusion model (such as a Transformer architecture model with text-image alignment) to improve the accuracy of instruction parsing.

[0145] For example, when processing multimodal inputs, medical text and corresponding medical image data (such as CT slices) can be simultaneously input into the image-text analysis model. Image features are extracted by an image encoder, such as a residual neural network (ResNet), and fused with the feature tensor generated by the text encoder.

[0146] When performing context-aware semantic analysis, image features can be introduced as contextual information into the attention mechanism of large language models. For example, when parsing "locating the ligation point of the right lower pulmonary artery S6 segment", the segmented right lower lung region in the image can be semantically aligned with the "S6 segment" in the text description to ensure more accurate matching of operation instructions.

[0147] Furthermore, dynamic instruction correction can be performed. For example, if there are blurred areas in the image (such as low-resolution slices), the semantic analysis model can also dynamically adjust the instruction matching strategy based on contextual information, such as prioritizing segmentation algorithms based on image features.

[0148] Based on this, through multimodal fusion, the model can more accurately understand the contextual information in medical texts (such as the correspondence between "S6 segment" and actual anatomical structures in images), improving the parsing accuracy of complex instructions (such as combined operations involving multiple anatomical structures). Furthermore, the dynamic correction mechanism can reduce instruction matching errors caused by differences in image quality, ensuring the reliability of operational results.

[0149] This application also provides a data processing method, which includes: invoking a semantic analysis module integrated in a medical data processing system to generate a target operation instruction sequence based on input medical text, wherein the semantic analysis module includes a large language model fine-tuned by a medical knowledge base; invoking an execution engine module integrated in the medical data processing system to perform processing tasks on medical image data based on the target operation instruction sequence, wherein the execution engine module includes multiple independently callable image processing sub-modules; and feeding back the execution results to the semantic analysis module through a communication interface module in the medical data processing system to adjust the subsequent instruction generation logic.

[0150] For example, a medical data processing system can be understood as a hardware or software architecture comprising a semantic analysis module, an execution engine module, and a communication interface module. The semantic analysis module can be understood as a software component integrating a large language model and a medical knowledge base, used to parse medical text and generate instruction sequences. The execution engine module can be understood as a collection of programs containing multiple image processing sub-modules (such as segmentation and annotation modules), used to execute target operation instruction sequences. The communication interface module can be understood as a hardware or software component that transmits data within the system, such as an application programming interface (API) or a data bus.

[0151] The medical data processing system receives medical text input through a semantic analysis module and uses a fine-tuned large language model combined with a medical knowledge base to generate a sequence of target operation instructions. The execution engine module then calls the corresponding image processing sub-modules (such as segmentation and annotation modules) to process the medical image data based on the instruction sequence. The communication interface module is responsible for transmitting execution results between the execution engine module and the semantic analysis module to dynamically adjust the subsequent instruction generation logic.

[0152] In this embodiment, the modular design of the medical data processing system improves overall processing efficiency and flexibility. For example, the integrated design of the semantic analysis module and the execution engine module ensures seamless connection between instruction generation and execution, and the introduction of the communication interface module enables real-time feedback of execution results, thereby dynamically optimizing the instruction sequence. This structured system design significantly reduces communication latency between modules and improves the processing consistency of complex tasks.

[0153] The data processing method provided in this embodiment is similar to the data processing methods provided in the above embodiments in terms of implementation principle and technical effect. For details, please refer to the descriptions in the above embodiments. This embodiment will not repeat them here.

[0154] This application also provides a data processing system, including: a semantic analysis module for generating a target operation instruction sequence based on input medical text, wherein the semantic analysis module includes a large language model fine-tuned by a medical knowledge base; an execution engine module for performing processing tasks on medical image data based on the target operation instruction sequence, wherein the execution engine module includes multiple independently callable image processing sub-modules; and a communication interface module for transmitting data between the semantic analysis module and the execution engine module to feed back the execution results to the semantic analysis module to adjust the subsequent instruction generation logic.

[0155] For example, the semantic analysis module can be understood as a hardware or software component integrating a large language model and a medical knowledge base, used to parse medical text and generate instruction sequences. The execution engine module can be understood as a collection of programs containing multiple image processing sub-modules (such as segmentation modules and annotation modules), used to execute target operation instruction sequences. The communication interface module can be understood as a hardware or software component that transmits data within the system, such as an API interface or a data bus.

[0156] The data processing system receives medical text input through the semantic analysis module and uses a fine-tuned large language model combined with a medical knowledge base to generate a sequence of target operation instructions. The execution engine module then calls the corresponding image processing sub-modules (such as segmentation and annotation modules) to process the medical image data based on the instruction sequence. The communication interface module is responsible for transmitting execution results between the execution engine module and the semantic analysis module to dynamically adjust the subsequent instruction generation logic.

[0157] In this embodiment, functional decoupling and efficient collaboration are achieved through a modular architecture of the data processing system. For example, the independent design of the semantic analysis module and the execution engine module ensures the parallelism of instruction generation and execution, while the introduction of the communication interface module enables real-time data feedback, thus forming a closed-loop optimization process. This structured design significantly improves the system's scalability and robustness, adapting to the processing needs of different medical scenarios.

[0158] The data processing system provided in this application embodiment is similar to the data processing methods provided in the above embodiments in terms of implementation principle and technical effect. For details, please refer to the descriptions in the above embodiments. This embodiment will not repeat them here.

[0159] Based on the methods or systems provided in this application, the operational complexity of lung watershed analysis can be reduced. By introducing a large language model, doctors' natural language instructions are automatically parsed into an ordered sequence of target operation instructions, eliminating the need for doctors to manually switch tools or interfaces, thereby simplifying the operation process and improving analysis efficiency. It can also improve the accuracy and consistency of analysis results. Semantic analysis using a large language model combined with a medical knowledge base ensures accurate matching of operation instructions, and the execution module automates task execution, reducing human error and improving the reliability of results.

[0160] Based on the methods or systems provided in the embodiments of this application, the intelligence and adaptability of the system can be enhanced. By fine-tuning the training of the large language model, it can understand specific terminology and operational requirements in pulmonary medicine scenarios, thereby supporting the parsing of more complex instructions and adapting to the needs of different clinical scenarios. End-to-end automation of medical image data processing can also be achieved. The entire process, from medical text input to operation instruction generation, task execution, and result output, is automated, reducing repetitive work for doctors and allowing them to focus more on diagnosis and decision-making.

[0161] Through the above technical solutions, the methods or systems of this application embodiment can significantly optimize the process of lung watershed analysis, reduce the technical threshold, and improve analysis efficiency and accuracy, thereby providing more efficient and reliable intelligent support for lung disease diagnosis and surgical planning.

[0162] This application also provides a medical data processing method. Figure 2 This is a flowchart illustrating the medical data processing method provided in the embodiments of this application, as shown below. Figure 2 As shown, the method includes:

[0163] S201, Obtain medical text and determine the first input features of the large language model based on the medical text;

[0164] S202, using a large language model to perform semantic analysis based on the first input features, and based on the results of the semantic analysis, determining at least one ordered target operation instruction from a set of multiple candidate operation instructions;

[0165] S203, based on at least one ordered target operation instruction, the execution module is sequentially called to execute the data processing task matching the target operation instruction, and the operation execution result is obtained;

[0166] S204 generates the first output of the large language model based on the operation execution result.

[0167] In step S201, the medical text includes at least operational text. The operational text can be a piece of natural language input by the first object at the application interface of the large language model, describing the operation that needs to be performed based on the medical image data. The target operation instruction can be understood as the operation instruction in the target operation instruction sequence in the above embodiments, and one or more ordered target operation instructions can constitute a target operation instruction sequence.

[0168] In this embodiment, the medical imaging data includes at least lung imaging data, such as medical imaging data obtained by scanning the second object using computed tomography (CT) scan. Of course, in other possible embodiments, the medical imaging data may also be images obtained using other medical imaging techniques, and this embodiment does not limit the scope of the medical imaging data.

[0169] It should be noted that, in the embodiments of this application, the first object can be understood as the object that performs lung watershed analysis using the medical data processing method proposed in the embodiments of this application, and the second object can be understood as the object that provides medical imaging data used for lung watershed analysis.

[0170] In another possible embodiment, the medical text further includes medical information text, which is text used to characterize the medical information of the second object corresponding to the medical image data. In one possible embodiment, the medical information text can be obtained based on the medical report of the second object, which may include the diagnostic report, medical image analysis report, electronic medical record, etc. of the second object. Specifically, the medical information text can be obtained by performing optical character recognition on the image of the medical report of the second object, or by extracting structured text from the digital medical report of the second object.

[0171] In addition, in one possible embodiment, the input end of the large language model can also be equipped with an additional interface to access the medical information management system. For example, it can access the hospital information system, laboratory information management system, image archiving and communication system corresponding to the first object, and obtain the medical information associated with the second object from these medical system management systems to obtain medical information text. Then, these medical information texts and operation texts are input into the large language model together.

[0172] The first input feature is a feature tensor derived from the medical text. In this embodiment, the medical text can be encoded into a corresponding feature tensor using a trained text encoder to obtain the first input feature. For example, the text encoder can encode the medical text using word embedding or one-hot encoding to obtain the first input feature; this is not limited in this embodiment.

[0173] In step S202, after obtaining the first input feature, the large language model can perform semantic analysis based on the first input feature. Specifically, the large language model includes N Transformer modules cascaded sequentially. Each Transformer module is equipped with an attention mechanism. After obtaining the first input feature, the first input feature is used as the input to the attention mechanism in the first Transformer module for attention operation. Subsequently, the output of each Transformer module is used as the input to the next cascaded Transformer module for attention operation again. Through the attention mechanism, outputs with strong correlation to the input medical text can be captured at the semantic level, thereby performing semantic analysis on the input medical text and obtaining the corresponding results.

[0174] It should be noted that, in this embodiment, the large language model is configured to determine at least one ordered target operation instruction from a set of pre-defined candidate operation instructions based on the results of semantic analysis. The candidate operation instructions are pre-set instructions used to invoke program code to perform specific operations on medical image data. The target operation instruction includes one or more instructions from the candidate operation instructions required to perform the operation corresponding to the text on the medical image data. For example, the target operation instruction may include instructions to divide lung image data into different regions according to a pre-defined algorithm, instructions to highlight a specific region in the lung image data, etc.

[0175] For example, if the medical text input into the large language model includes "find the ligation point of the S6 segment of the right lower pulmonary artery and show it with a red arrow", after semantic analysis, the large language model first determines that the S6 segment is located in the dorsal segment of the right lower lobe of the lung when the lung is divided based on the segmental bronchus. Then, it analyzes that it is necessary to find the ligation point of the dorsal segment of the right lower lobe of the lung artery. Based on this, it matches the target operation instructions required to complete this operation in the medical image data as follows: right lower lung segmentation, artery segmentation, S6 segment identification, determining the starting position of the S6 segment, and generating the ligation point position, thereby obtaining at least one ordered target operation instruction corresponding to the operation text.

[0176] In related technologies, the attention mechanism in each Transformer module of a large language model is configured with three weight matrices, namely the query weight matrix. Key weight matrix Sum weight matrix Each Transformer module's input is first multiplied by the three weight matrices to obtain the corresponding query matrix, key matrix, and value matrix. Then, attention operations are performed based on the query matrix, key matrix, and value matrix to obtain the output of the Transformer module.

[0177] In this embodiment, the large language model can also be fine-tuned to learn how to perform semantic analysis based on the operational text and match the corresponding target operation instructions. Specifically, the pre-trained large language model can be fine-tuned using knowledge data related to the lung medical scenario. By using knowledge data from the lung medical scenario to fine-tune the large language model, it can be made to perform semantic analysis on medical texts more accurately.

[0178] In this embodiment, an additional query weight fine-tuning matrix group and a value weight fine-tuning matrix group are set in the attention mechanism of each Transformer module. The two query weight fine-tuning matrices in the query weight fine-tuning matrix group are denoted as follows: , The two value weight fine-tuning matrices in the value weight fine-tuning matrix group are denoted as follows: , The sizes of these two sets of fine-tuning parameter matrices are determined based on the sizes of the query weight matrix and the value weight matrix, respectively.

[0179] For example, if the size of the query weight matrix is ​​denoted as M*N, then... and The matrix sizes are M*r and r*N respectively; similarly, the size of the value weight matrix is ​​denoted as M*N. and The matrix sizes are M*r and r*N, respectively. During the fine-tuning training of the large language model, the input to each Transformer module is not only related to the query weight matrix... Key weight matrix Sum weight matrix Performing matrix multiplication, and also fine-tuning the query weight matrix. , as well as , Perform matrix multiplication and accumulate the results into the input and query weight matrices of the Transformer module, respectively. and the key weight matrix The matrix multiplication results in the corresponding query matrix and value matrix. Specifically, in the attention mechanism of each Transformer module of the large language model, the process of determining the corresponding query matrix, key matrix, and value matrix based on the input features of the Transformer module can be referred to the following formula:

[0180] ;

[0181] ;

[0182] ;

[0183] in, Represents the query matrix; Represents the key matrix; Represents a value matrix; This represents the input to the Transformer module; , and These represent the query weight matrix, key weight matrix, and value weight matrix in the attention mechanism, respectively. These represent the two query weight fine-tuning matrices set in this attention mechanism; These represent the two value weight fine-tuning matrices set in this attention mechanism.

[0184] In each Transformer module, after calculating the query matrix, key matrix, and value matrix, matrix multiplication is performed on the query matrix and key matrix, and the attention weight matrix is ​​determined based on the result. Then, the value matrix is ​​weighted based on the attention weight matrix to obtain the output result of the attention mechanism in that Transformer module.

[0185] When fine-tuning a large language model using training samples from a lung medicine scenario, the neural network parameters of the large language model itself, such as the query weight matrix, key weight matrix, and value weight matrix in each Transformer module, can be frozen, and only the additional query weight fine-tuning matrix and value weight fine-tuning matrix can be optimized.

[0186] Specifically, each training sample includes sample medical text and at least one ordered label operation instruction, where the label operation instruction is used to complete the operation corresponding to the sample medical text. During the fine-tuning training of the large language model, all neural network parameters in the large language model except for the query fine-tuning weight matrices and value fine-tuning weight matrices are frozen. The sample medical text is input into the large language model, which performs semantic analysis and instruction matching based on the sample medical text to obtain the sample operation instructions output by the large language model.

[0187] Next, a loss function is calculated based on the sample operation instructions and the corresponding label operation instructions. This loss function characterizes the difference between the sample operation instructions generated by the large language model and the standard label operation instructions. Then, backpropagation is performed based on the loss function to optimize the query weight fine-tuning matrix and value weight fine-tuning matrix in each Transformer module until the optimized large language model meets the preset conditions, such as optimizing until the loss function value is less than a set threshold, or optimizing the relevant performance indicators of the optimized large language model on the preset test dataset to be higher than the set threshold.

[0188] This allows for fine-tuning of the large language model using training sample data from pulmonary medical scenarios, enabling the large language model to more accurately analyze and match the corresponding target operation instructions based on the input medical text.

[0189] In step S203, after determining the required target operation instructions, the large language model sends these instructions to the execution engine. The execution engine is a processing engine capable of processing medical images; it can perform various processing operations on medical images by calling program code. It is easy to understand that since the medical images need to be processed through the execution engine, the medical image data of the second object needs to be input into the execution engine. After receiving the corresponding target operation instructions, the execution module will call the corresponding program code to process the medical image data according to the target operation instructions.

[0190] For example, it can identify specific regions in medical images, measure the distance between specific points in medical images, segment medical images, and calculate the area of ​​a certain region in a medical image. Thus, it can perform operations corresponding to target operation instructions on the medical image and obtain the corresponding operation execution results.

[0191] In step S204, after obtaining the operation execution result, the execution engine returns the operation execution result to the large language model. The large language model generates a corresponding first output based on the operation execution result returned by the execution engine and displays the first output in the application interface of the large language model. The first output is used to characterize the result obtained by performing the operation corresponding to the operation text on the medical image data. The first output can include output results in various forms such as text and images. For example, the first output can include medical image data after being processed by the execution engine, such as medical images annotated or stained according to the operation text.

[0192] Furthermore, in one possible embodiment, when the operation text representation requires calculating the area of ​​a specific region in the medical image data, the corresponding first output may include a text describing the area of ​​the corresponding region. Of course, it is understood that, since medical image data needs to be processed, in one possible embodiment, the large language model may also have a medical knowledge base, which stores knowledge data related to lung medical scenarios. This knowledge data is encoded into feature vectors and stored in the knowledge base, and semantic analysis is performed based on the first input features using the large language model.

[0193] The medical data processing method provided in this application involves acquiring medical text input by a user, which contains operational text related to medical image data. This medical text is then input into a large language model, where the large language model performs semantic analysis on the input medical text to identify at least one ordered target operation instruction required to complete lung basin analysis of the medical image data. Subsequently, based on the target operation instruction, the corresponding execution modules are sequentially called to perform the corresponding operations on the medical image data, thereby completing the corresponding lung basin analysis based on the medical image data.

[0194] During the lung watershed analysis process, users only need to input medical text, and the large language model and various execution modules can automatically process the medical image data to obtain the lung watershed analysis results. This eliminates the need for users to frequently manually process medical images, simplifying the lung watershed analysis process and reducing its difficulty.

[0195] The medical data processing method provided in this application embodiment is similar to the data processing methods provided in the above embodiments in terms of implementation principle and technical effect. For details, please refer to the descriptions in the above embodiments. This embodiment will not repeat them here.

[0196] Figure 3 This is a schematic diagram of the architecture of a medical data processing system provided in an embodiment of this application. This medical data processing system can be used to execute the aforementioned medical data processing method. Figure 3 As shown, the medical data processing system may include a user interaction layer, an intelligent processing layer, a core algorithm layer, and a data support layer.

[0197] The user interaction layer may include units such as a natural language interface, a graphical user interface, and multimodal input processing. This layer is used to acquire medical text input by the user and detect user interactions through other methods. The intelligent processing layer may include units such as a large language model engine, an instruction engine, and a context manager. These units analyze the medical text input by the user and match instructions based on the analysis results. The core algorithm layer may include units such as a watershed analysis engine, a graphical visualization engine, a report generation engine, and an instruction function library. These units execute instructions by calling the corresponding instruction execution engine based on the instructions matched by the intelligent processing layer and visualize the execution results.

[0198] The data support layer may include units such as medical image databases, professional knowledge bases, patient databases, and surgical case databases. These units are used to access and retrieve data from the databases or knowledge bases in response to the intelligent processing layer or the core algorithm layer. For example, in response to the large language model engine of the intelligent processing layer, medical texts are analyzed, and matching structured knowledge data is retrieved from the professional knowledge base. Or, in response to the instruction engine, target operation instructions are executed on medical image data, and the medical image data to be processed is read from the medical image database.

[0199] Figure 4 This is a schematic diagram of the execution flow of medical data processing provided in the embodiments of this application, such as... Figure 4As shown, after medical data processing begins, the doctor inputs text instructions (medical text). The LLM (Limited Language Management) analyzes and identifies the intent behind these instructions. Based on the workflow logic, it generates instruction sets (target operation instruction sequences). Then, it performs workflow execution and standard instruction invocation steps to call the execution module to perform data processing tasks matching the target operation instructions. The interface can then be updated. During workflow execution and standard instruction invocation, adjustments can also be made to the doctor's input text instructions through text feedback.

[0200] Figure 5 This is a schematic diagram of the process for generating the final report provided in an embodiment of this application, such as... Figure 5 As shown, after the report generation is triggered, the template analysis step can be performed. This step first obtains the preset surgical planning report template, and then performs processing such as content to be filled analysis, page logical symbol analysis, chart generation analysis, and text filling analysis on the surgical planning report template, thereby determining the type of corresponding filling content for each area to be filled in the surgical planning report template.

[0201] After template analysis, content population can be performed. For example, patient information can be filled into the patient information population area of ​​the surgical planning report template. The large language model, combined with a knowledge base, analyzes the watershed analysis results obtained after the instruction execution engine executes the target operation instruction sequence. Content to be filled into the surgical planning report template is extracted from the watershed analysis results and filled into the corresponding areas. Afterward, the surgical planning report is automatically generated. Thus, users can automatically generate corresponding surgical planning reports based on the watershed analysis results using the large language model, eliminating the need for manual filling and reducing user workload.

[0202] Figure 6 This is a schematic diagram of the interactive interface of the medical data processing system provided in the embodiments of this application, such as... Figure 6 As shown, the left side of the interface displays a natural language interaction interface, where an intelligent surgical planning assistant can be displayed. Users can input natural language text, such as medical text, into the input boxes of the intelligent surgical planning assistant. The right side of the interface displays a 3D visualization and watershed analysis interface, which can include lung medical images and real-time analysis results.

[0203] Figure 7 This is a schematic diagram of the structure of the data processing apparatus provided in the embodiments of this application, such as... Figure 7As shown in the figure, this application embodiment provides a data processing apparatus, which includes: an acquisition module 701, used to acquire medical text containing operation statements on medical image data; a generation module 702, used to generate input features for a semantic analysis model based on the medical text; the generation module 702 is further used to generate an ordered sequence of target operation instructions based on the input features through the semantic analysis model, wherein the sequence of target operation instructions includes instructions for processing medical image data; and an execution module 703, used to call the execution module to perform corresponding processing tasks on the medical image data based on the sequence of target operation instructions.

[0204] In one possible implementation, the generation module 702 is specifically used to: perform semantic analysis on the input features based on a preset medical knowledge base using a semantic analysis model, wherein the medical knowledge base stores structured knowledge data related to lung medical scenarios; and generate an ordered sequence of target operation instructions based on the results of the semantic analysis.

[0205] In one possible implementation, the medical knowledge base includes multiple knowledge data, each of which includes vectorized structured knowledge; the generation module 702 is specifically used to: retrieve target knowledge data that matches the input features from the multiple knowledge data in the medical knowledge base based on the input features; and perform semantic analysis based on the input features and the target knowledge data to obtain the result of the semantic analysis.

[0206] In one possible implementation, the apparatus further includes a feedback module, which is used to: feed back the execution result of the execution module to the semantic analysis model; and perform execution effect analysis on the execution result through the semantic analysis model, so as to adjust the subsequent instruction generation logic based on the execution effect analysis of the execution result.

[0207] In one possible implementation, the execution module 703 is specifically used to: analyze the target operation instruction sequence to see if there are subtasks that can be processed independently, the subtasks that can be processed independently include segmenting different anatomical structures in the medical image data, and / or measuring different lesion regions in the medical image data; if the analysis determines that there is any subtask that can be processed independently in the target operation instruction sequence, the execution module is called to perform the corresponding processing task on the medical image data in a parallel execution manner.

[0208] In one possible implementation, the execution module 703 is specifically configured to: execute at least one subtask that can be processed independently in parallel and obtain at least one execution result; determine the execution result identifier of at least one execution result through a version control algorithm, wherein the execution result identifier is used to uniquely identify the execution result; determine the target execution result for result aggregation among at least one execution result through the result selection rules of the version control algorithm; and perform result aggregation operation according to the execution result identifier of the target execution result to obtain the overall execution result after performing the corresponding processing task on the medical image data.

[0209] In one possible implementation, the execution module is further configured to: receive natural language instructions input by the user regarding the execution result of the execution module; adjust the target operation instruction sequence based on the natural language instructions and generate an adjusted target operation instruction sequence; and / or, based on the natural language instructions, perform analysis on the execution result obtained after the execution module performs the corresponding processing task to obtain analysis information.

[0210] The data processing apparatus provided in this application embodiment can be used to execute the technical solution of the data processing method in any of the above embodiments of this application. Its implementation principle and technical effect are similar, and will not be described again here.

[0211] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, such as... Figure 8 As shown, the electronic device of this embodiment may include: at least one processor 801; and a memory 802 communicatively connected to the at least one processor; wherein the memory 802 stores instructions that can be executed by the at least one processor 801, and the instructions are executed by the at least one processor 801 to cause the electronic device to perform the method as described in any of the above embodiments.

[0212] Optionally, the memory 802 can be either standalone or integrated with the processor 801.

[0213] The implementation principle and technical effects of the electronic device provided in this embodiment can be found in the foregoing embodiments, and will not be repeated here.

[0214] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the method of any of the foregoing embodiments.

[0215] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the method of any of the foregoing embodiments.

[0216] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0217] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0218] It should be understood that the aforementioned processor can be a central processing unit (CPU) or other general-purpose processors. The processor can also be a digital signal processor (DSP) or an application-specific integrated circuit (ASIC), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0219] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device, and may also be various media that can store program code, such as USB flash drives, portable hard drives, read-only memory (ROM), disks or optical discs.

[0220] The aforementioned storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof. Examples of storage media include Static Random-Access Memory (SRAM) or Electrically Erasable Programmable Read Only Memory (EEPROM).

[0221] Storage media can be, for example, erasable programmable read-only memory (EPROM) or programmable read-only memory (PROM). Storage media can also be read-only memory (ROM), magnetic storage, flash memory, magnetic disks, or optical disks. Storage media can be any available medium accessible to general-purpose or special-purpose computers.

[0222] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. The processor and storage medium can reside within an application-specific integrated circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components within an electronic device or host device.

[0223] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0224] The sequence numbers of the embodiments in this application are merely for description and do not represent the superiority or inferiority of the embodiments. Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0225] Based on this understanding, the technical solution of this application, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0226] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0227] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0228] It should be further noted that although the steps in the flowchart are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise explicitly stated in this document, there is no strict order requirement for the execution of these steps, and they can be executed in other orders.

[0229] Furthermore, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but may be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0230] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0231] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0232] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A data processing method, characterized in that, The method includes: Retrieve medical text containing statements that manipulate medical image data; Input features for semantic analysis models are generated based on the aforementioned medical text; The semantic analysis model generates an ordered sequence of target operation instructions based on the input features, wherein the sequence of target operation instructions includes instructions for processing the medical image data; Based on the target operation instruction sequence, the execution module is invoked to perform corresponding processing tasks on the medical image data.

2. The method according to claim 1, characterized in that, The step of generating an ordered sequence of target operation instructions based on the input features using the semantic analysis model includes: The semantic analysis model performs semantic analysis on the input features based on a preset medical knowledge base, which stores structured knowledge data related to lung medical scenarios. An ordered sequence of target operation instructions is generated based on the results of the semantic analysis.

3. The method according to claim 2, characterized in that, The medical knowledge base includes multiple knowledge data, each of which includes vectorized structured knowledge; the semantic analysis of the input features based on the preset medical knowledge base includes: Based on the input features, target knowledge data matching the input features is retrieved from multiple knowledge data in the medical knowledge base; Semantic analysis is performed based on the input features and the target knowledge data to obtain the semantic analysis results.

4. The method according to any one of claims 1-3, characterized in that, After the method involves calling the execution module to perform corresponding processing tasks on the medical image data based on the target operation instruction sequence, the method further includes: The execution results of the execution module are fed back to the semantic analysis model; The execution results are analyzed using the semantic analysis model to adjust the subsequent instruction generation logic based on the analysis of the execution results.

5. The method according to any one of claims 1-3, characterized in that, The step of invoking the execution module to perform corresponding processing tasks on the medical image data based on the target operation instruction sequence includes: The target operation instruction sequence is analyzed to determine if there are subtasks that can be processed independently. These subtasks include segmenting different anatomical structures in the medical image data and / or measuring different lesion regions in the medical image data. If the analysis determines that there is any subtask in the target operation instruction sequence that can be processed independently, the execution module is invoked to perform the corresponding processing task on the medical image data in a parallel execution manner.

6. The method according to claim 5, characterized in that, The parallel execution method is used to call the execution module to perform corresponding processing tasks on the medical image data, including: At least one subtask that can be processed independently is executed in parallel and at least one execution result is obtained. An execution result identifier of the at least one execution result is determined by a version control algorithm. The execution result identifier is used to uniquely identify the execution result. Based on the result selection rules of the version control algorithm, a target execution result for result aggregation is determined from the at least one execution result; Based on the execution result identifier of the target execution result, the execution result aggregation operation is performed to obtain the overall execution result after performing the corresponding processing task on the medical image data.

7. The method according to any one of claims 1-3, characterized in that, After the method involves calling the execution module to perform corresponding processing tasks on the medical image data based on the target operation instruction sequence, the method further includes: Receive natural language instructions from the user regarding the execution result of the execution module; Based on the natural language instructions, the target operation instruction sequence is adjusted, and an adjusted target operation instruction sequence is generated. And / or, based on the natural language instructions, analyze the execution results obtained after the execution module performs the corresponding processing tasks to obtain analysis information.

8. An electronic device, characterized in that, include: Memory and processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method according to any one of claims 1-7.

Images