CN-122025183-A - Medical data labeling and training method, device, equipment, medium and product

CN122025183ACN 122025183 ACN122025183 ACN 122025183ACN-122025183-A

Abstract

The application discloses a method, a device, equipment, a medium and a product for marking and training medical data, and relates to the technical field of medical data processing. The inference labeling model instance comprises an initial node, an intermediate node and a terminal node, execution nodes meeting preset execution conditions are used for acquiring execution results according to preset execution steps, the execution results corresponding to the inference model are used for labeling input data, the next-stage node of the execution nodes is used as the execution node, and the preset execution conditions are repeatedly judged until the execution node does not meet the preset execution conditions or the execution node is the terminal node. The method can effectively integrate medical data and complete data labeling based on multi-source data parameters.

Inventors

Yang Tianshuai
LIU WEIPING
ZHANG GUANGYUN
ZHAO HUI
LI XIANG
GUO HONGWEI
HU JUNYI
ZHAO KEJIA
HAN FENG
WANG CHENYANG
SUN YINGLI
JIN JING
LI XIN
SU ZHEN
LI JING

Assignees

浙江泉林智能医疗科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260205

Claims (10)

1. A method for labeling and training medical data, the method comprising: Acquiring patient instance information, and acquiring an inference annotation model instance and a corresponding initial node as executing nodes according to the patient instance information, wherein the inference annotation model instance comprises an initial node, an intermediate node and a terminal node, the initial node is the preset basic item node, the intermediate node is one or more of the preset feature extraction node, the preset basic task node and the preset multi-stage task node, and the terminal node is the preset basic task node and/or the preset multi-stage task node; If the execution node is a preset basic task node/a preset multi-level task node, the corresponding preset execution step is to process input data according to a preset reasoning model to obtain an execution result, and the execution result is used as a label of the input data; And taking the next level node of the execution node as the execution node, and repeating the step of judging whether the execution node meets the corresponding preset execution condition or not until the execution node does not meet the preset execution condition or the execution node is the terminal node.
2. The method for labeling and training medical data according to claim 1, wherein the example of the inference labeling model is a directed acyclic graph, and the initial node, the intermediate node and the end node are preset nodes in the directed acyclic graph; If the preset node is the preset basic item node, the preset basic item is a patient medical record or a patient examination item, the preset execution condition is that the preset execution step is directly executed, and the preset execution step is that the result of the patient medical record or the patient examination item is obtained; If the preset node is the preset basic task node, the preset execution condition is that the execution result of the previous level node is obtained, and the preset execution step is that the execution results of all the previous level nodes are input into a corresponding preset reasoning model as input data to obtain the execution result; if the preset node is the preset multi-stage task node, the preset execution condition is that the execution results of all the previous-stage nodes are obtained in a preset time; And if the preset node is the preset feature extraction node, the preset execution condition is that the execution results of all the previous-stage nodes are obtained, and the preset execution step is that the execution results of all the previous-stage nodes are used as input data to be processed through corresponding preset feature extraction modules to obtain the execution results.
3. The method for labeling and training medical data according to claim 2, further comprising, after the obtaining of the execution result according to the corresponding preset execution step: Aiming at each execution node, acquiring a node type according to the node attribute of the execution node, wherein each preset node comprises a corresponding node attribute, and the node attribute is used for indicating that the type of the preset node is one of the preset basic item node, the preset basic task node, the preset multi-stage task node and the preset feature extraction node; if the node type is a preset basic task node/the preset multi-stage task node, acquiring the confidence coefficient of the execution result; if the confidence coefficient is not greater than a first preset value, adding the execution result and related information into an expert auditing queue; If the confidence is larger than the first preset value and not larger than the second preset value, adding the execution result and related information into the queue to be checked; If the confidence coefficient is larger than the second preset value, adding the execution result and related information into a spot check auditing queue; Transmitting the expert auditing queue, the necessary auditing queue and the spot inspection auditing queue to a manual auditing terminal through corresponding interfaces; and acquiring an indication of the completion of the audit, taking the execution result as a label of the input data as a response of the indication, and storing the input data and the label as sub-data into a database of the preset reasoning model.
4. The method of labeling and training medical data according to claim 3, wherein said obtaining a confidence level of the execution result comprises: The task type is obtained according to the model attribute identifier corresponding to the preset reasoning model, wherein the preset reasoning model comprises the corresponding model attribute identifier, and the model attribute identifier is used for indicating that the task type of the preset reasoning model belongs to one of a classification task, a target detection task, a semantic segmentation task and a regression task; if the task type is a classification task, taking the classification probability corresponding to the execution result as the confidence coefficient; If the task type is a target detection task, taking the classification probability and the object score corresponding to the execution result as the confidence; If the task type is a semantic segmentation task, taking the region average confidence coefficient of the execution result as the confidence coefficient; And if the task type is a regression task, taking the reciprocal of the variance estimation of the execution result as the confidence coefficient.
5. The method of labeling and training medical data according to claim 2, wherein the patient instance information includes a patient identification and at least one underlying item information, the initial node corresponding to the patient instance information is obtained as an execution node, comprising: Traversing the executing reasoning annotation model instance according to the patient identification, and judging whether the executing reasoning annotation model instance exists or not; if the executing reasoning annotation model instance exists, determining an initial node according to basic item information on the existing reasoning annotation model instance; and if the executing reasoning annotation model instance does not exist, creating the reasoning annotation model instance, and determining an initial node according to the basic project information.
6. The method for labeling and training medical data according to claim 3, the medical data labeling and training method is characterized by further comprising the following steps: And monitoring the data quantity of the corresponding sub-data in the database aiming at the preset reasoning model, expanding the sub-data in the database to the corresponding training set when the data quantity is larger than the corresponding preset numerical value, and retraining and updating the preset reasoning model according to the training set.
7. A medical data labeling and training apparatus, the medical data labeling and training apparatus comprising: the system comprises an acquisition module, a judgment module and a judgment module, wherein the acquisition module is used for acquiring patient instance information, and acquiring an inference annotation model instance and a corresponding initial node as execution nodes according to the patient instance information, wherein the inference annotation model instance comprises an initial node, an intermediate node and a terminal node; The system comprises an inference module, a processing module and a processing module, wherein the inference module is used for judging whether an execution node accords with a corresponding preset execution condition or not, and acquiring an execution result according to a corresponding preset execution step if the execution node accords with the preset execution condition, wherein the corresponding preset execution step is used for processing input data according to a preset inference model to acquire the execution result when the execution node is a preset basic task node/a preset multi-level task node, and the execution result is used for being used as a label of the input data; And taking the next level node of the execution node as the execution node, and repeating the step of judging whether the execution node meets the corresponding preset execution condition or not until the execution node does not meet the preset execution condition or the execution node is the terminal node.
8. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor executes the computer program to carry out the steps of the method of labeling and training of medical data according to any of claims 1-6.
9. A computer readable storage medium having stored thereon a computer program, characterized in that the computer program, when executed by a processor, implements the steps of the medical data labeling and training method of any of claims 1-6.
10. A computer program product comprising a computer program, characterized in that the computer program, when executed by a processor, implements the steps of the medical data labeling and training method of any of claims 1-6.

Description

Medical data labeling and training method, device, equipment, medium and product Technical Field The application relates to the technical field of medical data processing, in particular to a method, a device, equipment, a medium and a product for marking and training medical data. Background There are a large number of diagnostic and identification tasks in hospital systems that do not allow for deep analysis and intelligent processing of medical data (e.g., images, signals, text or time series recordings). These tasks typically rely on key model techniques such as feature extraction and pattern recognition. For example, in image diagnosis, the model is required to extract the shape, texture and density characteristics of a focus from medical images such as CT and MRI, in pathological analysis, the abnormal pattern of cells and tissues is required to be identified, and in physiological signal processing, characteristic points are required to be detected from waveforms such as electrocardiograph and electroencephalogram, and the abnormality is required to be classified. Through the models, the system can assist doctors to realize more accurate and efficient disease screening, classification and early warning, so that the objectivity and reliability of diagnosis and treatment are improved, and important data support is provided for clinical decision. Along with the continuous improvement of the medical informatization level, multi-source heterogeneous data such as medical images, electronic medical records, physiological monitoring data, genome information and the like are continuously accumulated in medical institutions, and the data volume is exponentially increased. However, these precious data resources are in a "deep sleep" state for a long time, and the traditional processing mode relying on manual arrangement, analysis and experience judgment has difficulty in meeting the requirements of clinical accurate diagnosis and treatment and medical scientific research on deep mining of data. In the prior art, the hospital system has difficulty in integrating data across departments and systems due to complex flow, ubiquitous information islands and non-uniform labeling standards. Large amounts of high quality medical data are difficult to label effectively and cannot be systematically fed back into existing machine learning and intelligent model techniques. The method not only affects the labeling quality and efficiency of the data, but also restricts the iterative optimization and retraining process of the model, so that the accuracy and generalization capability of a plurality of advanced algorithms are difficult to continuously improve in actual clinical scenes, and the comprehensive landing and development of the data-driven intelligent medical system are restricted. Disclosure of Invention The application aims to provide a medical data labeling and training method, device, equipment, medium and product, which can effectively integrate medical data and complete data labeling based on multi-source data parameters. In order to achieve the above object, the present application provides the following solutions: in a first aspect, the present application provides a method for labeling and training medical data, comprising: Acquiring patient instance information, and acquiring an inference annotation model instance and a corresponding initial node as executing nodes according to the patient instance information, wherein the inference annotation model instance comprises an initial node, an intermediate node and a terminal node, the initial node is the preset basic item node, the intermediate node is one or more of the preset feature extraction node, the preset basic task node and the preset multi-stage task node, and the terminal node is the preset basic task node and/or the preset multi-stage task node; Judging whether an executing node accords with a corresponding preset executing condition or not, and acquiring an executing result according to a corresponding preset executing step if the executing node accords with the preset executing condition, wherein if the executing node is a preset basic task node/a preset multi-level task node, the corresponding preset executing step is to process input data according to a preset reasoning model to acquire the executing result, and the executing result is used as a label of the input data; And taking the next level node of the execution node as the execution node, and repeating the step of judging whether the execution node meets the corresponding preset execution condition or not until the execution node does not meet the preset execution condition or the execution node is the terminal node. Optionally, the inference labeling model instance is a directed acyclic graph, and the initial node, the intermediate node and the end node are preset nodes in the directed acyclic graph, wherein the preset nodes comprise corresponding preset execution conditions and preset executi