Machine learning based ct image quantification analysis system

Abstract

The application discloses a CT image quantitative analysis system based on machine learning and relates to the technical field of image processing. The application comprises an information acquisition module for acquiring CT image data and analysis requirements; and an information processing module for performing specification uniform processing on the CT image data based on the CT image data through a specification processing method to obtain to-be-analyzed image data. The application can cross-verify and optimize multiple sub-information obtained from the to-be-analyzed image data through preset verification quantity and a subsequent cross-verification module, so as to improve the accuracy of characteristic region name and quantity information reflecting the to-be-analyzed image data, and further improve the accuracy of subsequent model processing. A single characteristic recognition model can analyze only a single characteristic region, so that the efficiency of characteristic recognition analysis can be improved, and multiple different recognition analysis requirements can be processed, multi-task processing is realized, and the efficiency of information processing is improved, so as to meet the increasing demand for precise diagnosis in clinical practice.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a CT image quantitative analysis system based on machine learning. Background Technology

[0002] The CT image quantitative analysis system based on machine learning is an intelligent platform that combines medical image processing, artificial intelligence and machine learning technologies to improve the diagnostic efficiency and accuracy of medical images. It extracts rich feature information from CT images through automated and quantitative methods to support early diagnosis, treatment evaluation and disease monitoring.

[0003] A method for quantitative analysis of novel coronavirus infection based on CT images, disclosed in patent publication number CN112037218A, relates to the field of image processing technology and includes the following steps: S1: Establish a machine learning model; S2: Input CT images for processing; S3: Generate 3D lesion data and calculate a quantitative factor for novel coronavirus infection; S4: Determine whether the quantitative factor score is higher than a critical threshold; if so, the patient is a COVID-19 patient, and S5 is executed; otherwise, the patient is not infected with the novel coronavirus; S5: Repeat S2 to S4, and after accumulating the quantitative factor of multiple CT images of the patient, obtain the patient's disease trend. This invention provides a simple and convenient method for quantitative analysis of novel coronavirus infection based on CT images. By generating 3D lesion data from CT images and calculating and quantifying the lesions for scoring, doctors can intuitively observe the location, size, and changes of the lesions, effectively determining whether a patient is infected with the novel coronavirus and further assessing the disease trend of COVID-19 patients.

[0004] In existing CT image analysis, although most systems have automatic identification and quantification functions, most of them rely on a single model or rule. When faced with complex and ambiguous image areas, they are prone to misjudgment and missed detection. This method, which relies on a single information source, lacks a multi-angle and multi-level verification mechanism, making it difficult to guarantee high accuracy and robustness. At the same time, it is inefficient when analyzing and processing feature information of multiple feature regions with a single model. In the absence of cross-validation and multi-task processing, it is difficult to meet the growing clinical demand for precision diagnosis. Therefore, this invention is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-based CT image quantitative analysis system to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a CT image quantitative analysis system based on machine learning, comprising:

[0007] Information acquisition module: Acquires CT image data and analyzes the required information;

[0008] Information processing module: Based on CT image data, the module performs standardization processing on the CT image data to obtain the image data to be analyzed.

[0009] Its features include:

[0010] Region acquisition module: Preset verification quantity, based on the verification quantity, acquire the feature region name and quantity information of the image data to be analyzed to obtain the result information. The result information includes several sub-information, and the number of sub-information is consistent with the verification quantity;

[0011] Cross-validation module: Based on the results, the cross-validation method is used to determine whether there are differences in the sub-information of the image data to be analyzed, and the validation results are obtained. Based on the validation results, auxiliary methods are used to judge the sub-information with differences and generate feedback information and optimization information.

[0012] Multi-task processing module: It presets basic feature regions, establishes a repository for storing different feature regions and their feature information through information acquisition methods, builds feature recognition models for analyzing different feature regions based on the repository, establishes the association between basic feature regions and feature recognition models, and establishes a relationship directory to store basic feature regions, feature recognition models and association relationships;

[0013] Demand matching module: Based on the analysis requirements, the module searches for the feature recognition model corresponding to the analysis requirements in the relational directory using a matching method to obtain the target recognition model. Based on the optimization information, the module labels the image data related to the analysis requirements in the image data to be analyzed to obtain the target image data. The target image data is then imported into the target recognition model to obtain the analysis information.

[0014] Furthermore, the cross-validation method includes: splitting the result information into several sub-information items, extracting feature region information from the sub-information items to obtain sub-feature information, extracting the number of feature regions from the sub-information items to obtain sub-quantity information, comparing whether the sub-feature information in the several sub-information items is consistent to obtain a first comparison result, comparing whether the sub-quantity information in the several sub-information items is consistent to obtain a second comparison result, obtaining a comparison result based on the first comparison result and the second comparison result, when the first comparison result and the second comparison result indicate that there is inconsistent information in the several sub-information items, the comparison result indicates that there is a difference, extracting the inconsistent information items in the several sub-information items to obtain the information to be verified, and integrating the comparison result and the information to be verified to obtain the verification result.

[0015] Furthermore, the auxiliary method includes: splitting the verification result to obtain the information to be verified; pre-selecting objects, including quantity selection, specified selection and auxiliary selection; determining the inconsistent information items in the information to be verified based on the selected objects to obtain replacement information items; and replacing the part of the sub-information corresponding to the replacement information item with the replacement information item to obtain the optimized information.

[0016] Furthermore, the process of establishing a feature recognition model based on the repository to analyze the features of different basic parts is as follows: Data acquisition and processing: Determine the recognition object of the feature recognition model to obtain the target object, traverse the repository based on the target object to select the sub-repository corresponding to the target object to obtain the target repository, and extract the reference information in the target repository to obtain the training data;

[0017] Model selection and training: Select a deep learning model as the model base, and import the training data into the model base for training to obtain the initial model;

[0018] Model Adjustment and Output: Preset verification images and verification information, import the verification images into the initial model to obtain test information, compare the test information and verification information to obtain the difference results, and adjust and optimize the initial model based on the difference results to obtain the feature recognition model of the target object;

[0019] Integrated Output: Integrate the feature recognition models of all target objects to complete the establishment of feature recognition models for analysis of different feature regions.

[0020] Furthermore, the information acquisition method includes: establishing a sub-repository named after the basic feature region, establishing a repository for storing the sub-repository, acquiring CT image information and feature information of the basic feature region to obtain reference information, and storing the reference information in the sub-repository.

[0021] Furthermore, the matching method includes: splitting and analyzing the requirements to obtain the target feature region and information requirements; traversing the relational directory based on the target feature region to obtain the traversal result; selecting a feature recognition model based on the traversal result and the correlation relationship to obtain the selected recognition model; and adjusting the output mode of the selected recognition model based on the information requirements to obtain the target recognition model.

[0022] Furthermore, the process of obtaining the feature region names and quantity information of the image data to be analyzed based on the number of verifications is as follows: determine the method for obtaining the feature region names and quantity information of the image data to be analyzed to obtain the target method; obtain the recognition accuracy of the target method; sort the target methods from high to low recognition accuracy to obtain the sorting result; select the target methods that are consistent with the number of verifications from the sorting result from high to low recognition accuracy to obtain several selection methods; combine the several selection methods with the image data to be analyzed to obtain sub-information reflecting the feature region names and quantity information in the image data to be analyzed; integrate the several sub-information to obtain the result information.

[0023] Furthermore, the specification processing method includes: pre-setting a target format, pre-processing the CT image data to obtain pre-processed CT image data, determining whether the CT image data is consistent with the target format to obtain a judgment result, when the judgment result indicates that the CT image data is inconsistent with the target format, horizontally scaling and vertically scaling the CT image data to be consistent with the target format to obtain the image data to be analyzed and the target scaling information, and obtaining the actual information based on the target scaling information and the analysis information.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] This machine learning-based CT image quantitative analysis system, through a preset number of validations and a subsequent cross-validation module, can cross-validate multiple sub-information obtained from the image data to be analyzed, and optimize the sub-information to improve the accuracy of the feature region names and quantity information of the image data to be analyzed, thereby improving the accuracy of subsequent model processing. By analyzing only a single feature region through a single feature recognition model, the efficiency of feature recognition analysis can be improved. At the same time, it can handle multiple different recognition analysis needs, reflecting multi-task processing, so as to improve the efficiency of information processing and meet the growing clinical demand for precision diagnosis.

[0026] Meanwhile, the cross-validation method can be used to determine whether there are differences in the sub-information of the image data to be analyzed. When the validation results indicate that there are differences in several sub-information, the auxiliary method can be used to judge and optimize the sub-information to generate feedback information and optimization information. The optimization information is the optimized sub-information, and the feedback information is the sub-information with differences. By feeding the feedback information back to the manager, the manager can make corresponding optimization and adjustment to the target path of generating sub-information to improve the accuracy of data processing.

[0027] Meanwhile, by setting up a standard processing method, CT image data of different specifications are processed to unify the specifications, so as to improve the analysis efficiency of the analysis system. Standardized image data can improve the stability and generalization ability of the model and avoid errors caused by data differences. The process of adjusting the output mode of the target recognition model based on information needs to obtain the target recognition model means that the information format output by the target recognition model is adjusted according to the information format requirements to meet the needs of the staff. Attached Figure Description

[0028] Figure 1 This is a simplified flowchart of the present invention;

[0029] Figure 2 This is a schematic diagram of the main process structure of the present invention;

[0030] Figure 3 This is a schematic diagram of the cross-validation method of the present invention;

[0031] Figure 4 This is a schematic diagram of the annotation process structure of the present invention;

[0032] Figure 5 This is a schematic diagram of the training structure of the feature recognition model of the present invention;

[0033] Figure 6 This is a schematic diagram of the specification processing method of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Quantitative analysis of CT images is a process that uses computer technology to quantitatively measure and extract features from CT (computed tomography) images. By converting information in the images into numerical indicators, it helps doctors more objectively and accurately assess the state of tissues, organs, and lesions, thereby supporting diagnostic and treatment decisions.

[0036] like Figures 1-6 As shown, the present invention provides a technical solution: a CT image quantitative analysis system based on machine learning, comprising:

[0037] Information acquisition module: Acquires CT image data and analyzes the required information;

[0038] It is important to note that the process of acquiring CT image data includes connecting to the diagnostic and treatment system, screening target images, standardizing the format and collecting data after ensuring quality, and obtaining analysis requirements through clinical communication or system input, clarifying the target area and indicators, and ensuring that the requirements are expressed accurately.

[0039] Information processing module: Based on CT image data, the module performs standardization processing on the CT image data to obtain the image data to be analyzed.

[0040] It is important to note that by setting a standardization method, CT image data of different specifications can be processed to unify their specifications, thereby improving the analysis efficiency of the analysis system. Standardized image data can enhance the stability and generalization ability of the model and avoid errors caused by data differences.

[0041] Its features include:

[0042] Region acquisition module: Preset verification quantity, based on the verification quantity, acquire the feature region name and quantity information of the image data to be analyzed to obtain the result information. The result information includes several sub-information, and the number of sub-information is consistent with the verification quantity;

[0043] It is important to note that the number of validations refers to the total number of sub-information. Each sub-information is obtained from a single approach in conjunction with the image data to be analyzed. By presetting the number of validations and using the subsequent cross-validation module, the sub-information obtained from the image data to be analyzed can be cross-validated and optimized to improve the accuracy of reflecting the feature region names and quantity information of the image data to be analyzed, thereby improving the accuracy of subsequent model processing.

[0044] Cross-validation module: Based on the results, the cross-validation method is used to determine whether there are differences in the sub-information of the image data to be analyzed, and the validation results are obtained. Based on the validation results, auxiliary methods are used to judge the sub-information with differences and generate feedback information and optimization information.

[0045] It is important to note that the cross-validation method can be used to determine whether there are differences in the sub-information of the image data to be analyzed. When the validation results indicate that there are differences in several sub-information, the auxiliary method can be used to judge and optimize the sub-information to generate feedback information and optimization information. The optimization information is the optimized sub-information, and the feedback information is the sub-information that has differences. By feeding the feedback information back to the administrator, the administrator can make corresponding optimizations and adjustments to the way the sub-information is generated, thereby improving the accuracy of data processing.

[0046] Multi-task processing module: Establishes a repository for storing different feature regions and their feature information through information acquisition methods; builds feature recognition models for analyzing different feature regions based on the repository; establishes the association between basic feature regions and feature recognition models; and establishes a relational directory to store basic feature regions, feature recognition models, and associations.

[0047] It is important to note that a repository is established by setting the information acquisition method. The repository includes several sub-repositories, which store information about individual feature regions. This allows for the creation of feature recognition models for individual feature regions based on individual sub-repositories. By analyzing only a single feature region using a single feature recognition model, the efficiency of feature recognition analysis can be improved. At the same time, multiple different recognition analysis needs can be handled, reflecting multi-tasking to improve the efficiency of information processing. By establishing relationships, it is easy to locate the corresponding feature recognition model based on the basic feature region. By establishing a relational directory, it is easy to search within the relational directory.

[0048] Demand matching module: Based on the analysis requirements, the module searches for the feature recognition model corresponding to the analysis requirements in the relational directory using a matching method to obtain the target recognition model. Based on the optimization information, the module labels the image data related to the analysis requirements in the image data to be analyzed to obtain the target image data. The target image data is then imported into the target recognition model to obtain the analysis information.

[0049] It is important to note that, based on the analysis requirements and in conjunction with the matching method, a feature recognition model that meets the analysis requirements is found in the relation directory to obtain the target recognition model. After optimizing the information of the feature regions in the information and labeling the image data to be analyzed in accordance with the analysis requirements, the data is imported into the target recognition model to obtain the analysis information. The analysis information can then be used to assist medical staff in making judgments.

[0050] In the specific implementation process, such as Figure 4As shown, firstly, the CT image data to be processed is loaded into the analysis environment to obtain or predefine the feature information of the target region. The feature information of the target region is defined according to the analysis requirements, such as intensity range, texture features, boundary characteristics, and shape parameters. This feature information supports subsequent filtering and matching. Multiple candidate regions are extracted from the original image using region segmentation techniques (such as thresholding, region growing, clustering, and multi-threshold methods). The purpose of this step is to distinguish potential target regions from the background, providing a basis for feature comparison. Feature parameters are extracted for each candidate region and then compared with known target feature information. Regions that meet specific conditions are filtered based on feature similarity; these regions are identified as potential target regions. Regions that meet the conditions are marked on the original image. Marking methods can include adding boundary contours, changing region colors, and adding marker points or labels at the center of the region. Finally, the target region is clearly marked in the image, providing a basis for subsequent quantitative analysis, report generation, or assisted diagnosis.

[0051] like Figure 3 As shown, the cross-validation method includes: splitting the result information into several sub-information, extracting feature region information from the sub-information to obtain sub-feature information, extracting the number of feature regions from the sub-information to obtain sub-quantity information, comparing whether the sub-feature information in the several sub-information is consistent to obtain a first comparison result, comparing whether the sub-quantity information in the several sub-information is consistent to obtain a second comparison result, obtaining a comparison result based on the first comparison result and the second comparison result, when the first comparison result and the second comparison result indicate that there is inconsistent information in the several sub-information, the comparison result indicates that there is a difference, extracting the inconsistent information items in the several sub-information to obtain the information to be verified, and integrating the comparison result and the information to be verified to obtain the verification result.

[0052] It is important to note that the cross-validation method is used to cross-validate the consistency of feature region information generated from multiple methods to represent CT image data. This improves the accuracy of obtaining feature region information representing CT image data. Furthermore, staff can optimize corresponding methods based on the cross-validation results to enhance the accuracy of subsequent information generation. During the cross-validation process, feature region information and quantity information in the sub-information are compared in two parts to improve comparison efficiency. The quantity information represents the number of feature regions, which are the names of the regions identified in the CT image data, such as lung nodules. The result information includes sub-information 1, sub-information 2, sub-information 3... sub-information n. Similarly, sub-feature information and sub-quantity information correspond to each other.

[0053] like Figure 2As shown, the auxiliary method includes: splitting the verification result to obtain the information to be verified, pre-selecting objects, including quantity selection, specified selection and auxiliary selection, determining the inconsistent information items in the information to be verified based on the selected objects to obtain the replacement information items, and replacing the part of the sub-information corresponding to the replacement information items with the replacement information items to obtain the optimization information.

[0054] It is important to note that when the selection object is quantity selection, the information item with the most inconsistent information items in the information to be verified is extracted to obtain the replacement information item. When the selection object is specified selection, the information item corresponding to the inconsistent information item in the information to be verified and the specified selection is extracted to obtain the replacement information item. When the selection object is auxiliary selection, the information to be verified is integrated to generate feedback information and fed back to the user. The user feedback is listened to to obtain the replacement information item. For quantity selection, the sub-feature information and sub-quantity information with the most common quantity among several sub-information are selected as replacement information items. For specified selection, the sub-feature information and sub-quantity information generated by the specified method are selected as replacement information items when there are differences, as determined by the administrator. For auxiliary selection, the verification result is fed back to the administrator, who determines the corresponding sub-feature information and quantity information as replacement information items.

[0055] like Figure 2 and Figure 5 As shown, the process of establishing a feature recognition model based on the repository to analyze the features of different basic parts is as follows: Data acquisition and processing: Determine the recognition object of the feature recognition model to obtain the target object, traverse the repository based on the target object to select the sub-repository corresponding to the target object to obtain the target repository, and extract the reference information in the target repository to obtain the training data;

[0056] It is important to note that the process of determining the target object for the feature recognition model is to identify the object that the feature recognition model will analyze, specifically a single feature region. The training data for model training can be obtained by finding the sub-repository corresponding to the target object in the repository and extracting the reference information. In actual use, the training data can be preprocessed, including but not limited to data cleaning, deduplication, and enhanced display.

[0057] Model selection and training: Select a deep learning model as the model base, and import the training data into the model base for training to obtain the initial model;

[0058] It is important to note that deep learning models include, but are not limited to, convolutional neural networks, which are trained by importing training data into a model base to obtain an initial model.

[0059] Model Adjustment and Output: Preset verification images and verification information, import the verification images into the initial model to obtain test information, compare the test information and verification information to obtain the difference results, and adjust and optimize the initial model based on the difference results to obtain the feature recognition model of the target object;

[0060] It is important to note that the verification image is the feature region corresponding to the target object, and the verification information is the specific information of the feature region, including but not limited to pathological information. The difference results are obtained by comparing the test information generated by the initial model with the verification information. The feature recognition model of the target object can be obtained by adjusting and optimizing the initial model based on the difference results.

[0061] Integrated Output: Integrate the feature recognition models of all target objects to complete the establishment of feature recognition models for analysis of different feature regions.

[0062] It is worth noting that by using a single feature region to correspond to a single feature recognition model, the efficiency of analyzing feature regions can be significantly improved, reflecting multi-task processing and improving the efficiency of acquiring information from feature regions.

[0063] like Figure 2 As shown, the information acquisition method includes: pre-setting a basic feature region, establishing a sub-repository named after the basic feature region, establishing a repository to store the sub-repository, acquiring CT image information and feature information of the basic feature region to obtain reference information, and storing the reference information in the sub-repository.

[0064] It is important to note that the process of presetting basic feature regions involves presetting feature regions based on the CT images that would appear in actual use, such as nodules. The process of obtaining reference information by acquiring CT image information and feature information of basic feature regions involves acquiring image information of basic feature regions and feature information that expresses that image information from previous CT image information. There are multiple basic feature regions. By storing the reference information in a sub-repository and naming it after the basic feature regions, it is easy to find the corresponding training data when training the model, thereby improving the efficiency of model training.

[0065] like Figure 2 As shown, the matching method includes: splitting and analyzing the requirements to obtain the target feature region and information requirements; traversing the relational directory based on the target feature region to obtain the traversal results; selecting a feature recognition model based on the traversal results and the correlation relationship to obtain the selected recognition model; and adjusting the output mode of the selected recognition model based on the information requirements to obtain the target recognition model.

[0066] It is important to note that by setting the matching method, selecting the corresponding feature recognition model for the target feature region based on the needs analysis, and adjusting the output mode of the selected recognition model based on information needs to obtain the target recognition model, the process involves adjusting the information format output by the target recognition model according to the required information format so that the output information format meets the needs of the staff.

[0067] In the specific implementation process, different output modes can be added during the model training phase. Different output modes correspond to different information needs. For example, Requirement 1: output only the basic pathological category information of the region, such as: "This region is a malignant tumor, characterized by irregular edges, with an area of ​​approximately 12 square centimeters." Requirement 2: output the pathological feature parameters of the region (such as morphology, tissue type, density, etc.), such as: "This region is a well-differentiated gastric tumor with unclear boundaries, a density of +40 HU, an internal heterogeneous structure, and abundant blood vessels." Requirement 3: output detailed pathological related indicators and parameters (such as size, density, blood flow information, tissue type parameters, etc.), such as: Pathological category: malignant adenocarcinoma, size parameter: maximum radial diameter 15 mm, area 20 square millimeters, density parameter: average density of +45 HU, internal heterogeneity, vascular richness: highly vascularized with obvious blood flow signals, tissue characteristics: cells are tightly arranged, and nuclear atypia is obvious, indicating high-grade malignant changes.

[0068] The process of obtaining the feature region names and quantity information of the image data to be analyzed based on the number of verifications is as follows: determine the method for obtaining the feature region names and quantity information of the image data to be analyzed to obtain the target method; obtain the recognition accuracy of the target method; sort the target methods from high to low recognition accuracy to obtain the sorting result; select the target methods that are consistent with the number of verifications from the sorting result from high to low recognition accuracy to obtain several selection methods; combine the several selection methods with the image data to be analyzed to obtain sub-information reflecting the feature region names and quantity information in the image data to be analyzed; integrate the several sub-information to obtain the result information.

[0069] It should be noted that the process of determining the method for obtaining the name and quantity information of the feature regions of the image data to be analyzed, and obtaining the target method, can be done in the existing technology according to the usage requirements. The specific function of the target method is to obtain the name and quantity information of the feature regions of the image data to be analyzed.

[0070] In the specific implementation process, the target approach can be as follows: Approach 1: Automatically extract different regions through image segmentation algorithms (such as threshold segmentation, region growing, watershed, clustering, etc.). The specific process involves setting a threshold in advance, segmenting according to pixel intensity or texture features, using continuous region extraction algorithms to obtain the spatial definition of each feature region, and counting the name (or number) and quantity of each region and storing it in an information structure; Approach 2: Identify by matching the features of image regions with features in known templates or dictionaries. The specific process involves constructing a feature dictionary (such as typical lesion features, parameters of specific structures), calculating the similarity between the regional features in the image and the templates in the dictionary, identifying regions that match the features, and naming them the corresponding region types; Approach 3: Train a model to identify different feature regions (such as using a convolutional neural network). The specific process involves training an object detection or segmentation model (such as Mask R-CNN) using a labeled dataset, learning regional features, running the model to detect target regions, and the model outputting the region category (name) and quantity. The regional information output by the model is extracted as the "feature region name and quantity"; Approach 4: Medical experts manually or semi-automatically annotate region information. The specific process involves experts manually annotating region names and ranges based on image features, and using knowledge bases or rule systems to assist in identification and statistics. Approach 5: Set rules (such as intensity range, texture parameters, geometric features) to automatically identify regions. The specific process involves applying filtering, edge detection, and connected component analysis to the image, filtering regions based on parameters, counting the number of regions that meet the conditions, and naming the regions (such as "tumor region 1", "vascular cluster 2", etc.) according to the rules.

[0071] like Figure 6 As shown, the specification processing method includes: setting a target format, preprocessing the CT image data to obtain preprocessed CT image data, judging whether the CT image data is consistent with the target format to obtain a judgment result, when the judgment result is that the CT image data is inconsistent with the target format, scaling the CT image data horizontally and vertically until it is consistent with the target format to obtain the image data to be analyzed and the target scaling information, and obtaining the actual information based on the target scaling information and the analysis information.

[0072] It is important to note that the target format is determined by the staff, specifically the size information of the CT image data. A standardized format is established for the CT image data through a set specification processing method, improving the efficiency of subsequent model processing. The process of obtaining the actual information based on the target scaling information, combined with the analysis information, involves scaling up and down the target information according to the area and other unit information in the obtained analysis information to obtain the actual information of the analysis. Figure 6 As shown, Figure 6 The actual format is the original format of the CT image data, and the target scaling information is the scaling size required to adjust the actual format to the target format.

[0073] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. A machine learning-based CT image quantitative analysis system, including: Information acquisition module: Acquires CT image data and analyzes the required information; Information processing module: Based on CT image data, the module performs standardization processing on the CT image data to obtain the image data to be analyzed. Its features include: Region acquisition module: Preset verification quantity, based on the verification quantity, acquire the feature region name and quantity information of the image data to be analyzed to obtain the result information. The result information includes several sub-information, and the number of sub-information is consistent with the verification quantity; Cross-validation module: Based on the results, the cross-validation method is used to determine whether there are differences in the sub-information of the image data to be analyzed, and the validation results are obtained. Based on the validation results, auxiliary methods are used to judge the sub-information with differences and generate feedback information and optimization information. Multi-task processing module: It presets basic feature regions, establishes a repository for storing different feature regions and their feature information through information acquisition methods, builds feature recognition models for analyzing different feature regions based on the repository, establishes the association between basic feature regions and feature recognition models, and establishes a relationship directory to store basic feature regions, feature recognition models and association relationships; Demand matching module: Based on the analysis requirements, the module searches for the feature recognition model corresponding to the analysis requirements in the relational directory using a matching method to obtain the target recognition model. Based on the optimization information, the module labels the image data related to the analysis requirements in the image data to be analyzed to obtain the target image data. The target image data is then imported into the target recognition model to obtain the analysis information.

2. The CT image quantitative analysis system based on machine learning according to claim 1, characterized in that: The cross-validation method includes: splitting the result information into several sub-information items; extracting feature region information from the sub-information items to obtain sub-feature information; extracting the number of feature regions from the sub-information items to obtain sub-quantity information; comparing whether the sub-feature information in the several sub-information items is consistent to obtain a first comparison result; comparing whether the sub-quantity information in the several sub-information items is consistent to obtain a second comparison result; obtaining a comparison result based on the first comparison result and the second comparison result; when the first comparison result and the second comparison result indicate that there is inconsistent information in the several sub-information items, the comparison result indicates that there is a difference; extracting the inconsistent information items in the several sub-information items to obtain the information to be verified; and integrating the comparison result and the information to be verified to obtain the verification result.

3. The CT image quantitative analysis system based on machine learning according to claim 1, characterized in that: The auxiliary method includes: splitting the verification result to obtain the information to be verified; pre-selecting objects, including quantity selection, specified selection and auxiliary selection; determining the inconsistent information items in the information to be verified based on the selected objects to obtain replacement information items; and replacing the part of the sub-information corresponding to the replacement information item with the replacement information item to obtain the optimized information.

4. The CT image quantitative analysis system based on machine learning according to claim 1, characterized in that: The process of establishing a feature recognition model based on a repository to analyze the features of different basic parts is as follows: Data acquisition and processing: Determine the recognition object of the feature recognition model to obtain the target object, traverse the repository based on the target object to select the sub-repository corresponding to the target object to obtain the target repository, and extract the reference information in the target repository to obtain the training data. Model selection and training: Select a deep learning model as the model base, and import the training data into the model base for training to obtain the initial model; Model Adjustment and Output: Preset verification images and verification information, import the verification images into the initial model to obtain test information, compare the test information and verification information to obtain the difference results, and adjust and optimize the initial model based on the difference results to obtain the feature recognition model of the target object; Integrated Output: Integrate the feature recognition models of all target objects to complete the establishment of feature recognition models for analysis of different feature regions.

5. The CT image quantitative analysis system based on machine learning according to claim 1, characterized in that: The information acquisition method includes: establishing a sub-repository named after the basic feature region, establishing a repository to store the sub-repository, acquiring CT image information and feature information of the basic feature region to obtain reference information, and storing the reference information in the sub-repository.

6. The CT image quantitative analysis system based on machine learning according to claim 1, characterized in that: The matching method includes: splitting and analyzing the requirements to obtain the target feature region and information requirements; traversing the relational directory based on the target feature region to obtain the traversal result; selecting a feature recognition model based on the traversal result and the correlation relationship to obtain the selected recognition model; and adjusting the output mode of the selected recognition model based on the information requirements to obtain the target recognition model.

7. The CT image quantitative analysis system based on machine learning according to claim 1, characterized in that: The process of obtaining the feature region names and quantity information of the image data to be analyzed based on the number of verifications is as follows: determine the method for obtaining the feature region names and quantity information of the image data to be analyzed to obtain the target method; obtain the recognition accuracy of the target method; sort the target methods from high to low recognition accuracy to obtain the sorting result; select the target methods that are consistent with the number of verifications from the sorting result from high to low recognition accuracy to obtain several selection methods; combine the several selection methods with the image data to be analyzed to obtain sub-information reflecting the feature region names and quantity information in the image data to be analyzed; integrate the several sub-information to obtain the result information.

8. The CT image quantitative analysis system based on machine learning according to claim 1, characterized in that: The specification processing method includes: setting a target format, preprocessing CT image data to obtain preprocessed CT image data, determining whether the CT image data is consistent with the target format to obtain a judgment result, when the judgment result indicates that the CT image data is inconsistent with the target format, scaling the CT image data horizontally and vertically until it is consistent with the target format to obtain the image data to be analyzed and the target scaling information, and obtaining the actual information based on the target scaling information and the analysis information.

Images