Medical data classification method and apparatus based on mutual correction and information fusion
The method addresses inefficiencies in medical data classification by correcting and fusing decision information from CT and MRI images, improving accuracy and interpretability through a fuzzy classification model.
Patent Information
- Application Number
- JP2024160788
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2024-06-19
- Filing Date
- 2024-09-18
- Publication Date
- 2024-12-12
- Estimated Expiration
- 2044-09-18
AI Technical Summary
Existing medical data classification methods using CT and MRI images lack effective interpretation of feature learning from multiple views, leading to inefficient automatic detection and the need for human intervention due to varying imaging resolutions and insufficient data integration.
A medical data classification method utilizing mutual correction and information fusion, employing a fuzzy classification model with CT and MRI images, where decision information from different viewing angles is corrected and fused to enhance interpretability and accuracy.
The method improves classification accuracy by correcting and fusing decision information from CT and MRI views, reducing feature dimensionality and enhancing interpretability, thus minimizing the need for human intervention.
Smart Images

Figure 0007598597000001_ABST
Abstract
Description
[Technical field]
[0001] The present application relates to the field of medical data processing technology, and in particular to a method and apparatus for medical data classification based on mutual correction and information fusion. [Background technology]
[0002] With the development of artificial intelligence technology, more and more medical data can be used to find relevant pathological features, for example, computed tomography data (CT) or magnetic resonance imaging data (MRI) can be used to find digitized features of relevant tumors. In clinical diagnosis, many attempts have been made to improve the medical data classification capabilities of artificial intelligence tools. For example, more advanced networks such as three-dimensional convolutional neural networks (CNNs), recurrent neural networks based on courtesy gate control recurrent units, support vector machines (SVMs), U-net networks, etc. have been successfully applied to classify various medical data, such as nasopharyngeal carcinoma.
[0003] The key to classifying the above medical data is to obtain favorable key feature information that influences the final decision. The prior art uses the idea of feature selection by genetic algorithm, or the idea of combining clinical features and radiological features based on MRI. However, the above-mentioned method uses the endoscopic image as a single viewing angle, and does not consider the different resolution of the imaging technology of different imaging methods. Considering the actual situation, if the subsequent prior art adopts the multi-view model of CT and MRI images, such a model lacks clear interpretation for feature learning, and data images from multiple views often have a large amount of image information, so automatic detection takes a long time, and in some scenes, due to the limited model classification effect, the human intervention of the doctor is always required to determine the classification result, and advanced automatic classification cannot be realized. Summary of the Invention [Problem to be solved by the invention]
[0004] In view of this, embodiments of the present application are directed to a method and apparatus for medical data classification based on mutual correction and information fusion. [Means for solving the problem]
[0005] In a first aspect, an embodiment of the present invention provides a method for medical data classification based on mutual correction and information fusion, comprising: acquiring one or more CT images and / or one or more MRI images of the target object; inputting the plurality of CT images and the plurality of MRI images into a fuzzy classification model, and obtaining a medical data classification result output from the fuzzy classification model; Wherein, the training process of the fuzzy classification model is training decision information of a CT view based on sample data in a CT image sample data set and corresponding classification tags, the CT image sample data set being respectively used for a scan sub-view, a first contrast-enhanced sub-view and a second contrast-enhanced sub-view in the CT view to perform mutual correction of the decision information; training decision information for an MRI view based on sample data in a MRI image sample data set and corresponding classification tags, the MRI image sample data being respectively used for a number of different organ sub-views in the MRI view to perform mutual correction of the decision information; and outputting and obtaining a medical data classification result of the target object after performing information fusion on the trained decision information of the CT view and the MRI view, respectively.
[0006] Optionally, the step of training CT view determination information based on sample data in the CT image sample dataset and corresponding classification tags specifically includes: inputting any sample in the CT image sample dataset into a first target sub-view corresponding to the CT view, and obtaining a first actual classification result output from the first target sub-view; obtaining a judgment tag from the first actual classification result and a classification tag corresponding to any of the samples; correcting decision information corresponding to the CT view based on the decision tag until training of all samples is completed; and obtaining a final decision of the CT view by a minimum learner method based on the decision information corresponding to the CT view after correction.
[0007] Optionally, the step of correcting the decision information corresponding to the CT view based on the decision tag specifically includes: If the decision tag indicates that any of the samples is correctly classified into the first target subview, the method includes a step of retaining decision information in the first target subview, and otherwise correcting the decision information in the first target subview based on decision information of other subviews other than the first target subview.
[0008] The step of selectively correcting decision information in the first target subview based on decision information of other subviews other than the first target subview specifically comprises: if the decision tag indicates that any of the samples is correctly classified into a candidate subview other than the first target subview, correcting decision information in the first target subview based on decision information of the candidate subview; and if the decision tag indicates that any of the samples is correctly classified into multiple candidate subviews other than the first target subview, correcting decision information in the first target subview based on decision information of the multiple candidate subviews by a separation degree calculation.
[0009] Optionally, the step of training MRI view decision information based on sample data in the MRI image sample dataset and corresponding classification tags specifically includes: inputting any sample in the MRI image sample dataset into a second target organ sub-view corresponding to the MRI view, and obtaining a second actual classification result output from the second target organ sub-view; obtaining a decision tag from the second actual classification result and a classification tag corresponding to any of the samples; correcting decision information corresponding to the MRI view based on the decision tag until training of all samples is completed; and obtaining a final decision for the MRI view by a minimal learner method based on the decision information corresponding to the corrected MRI view.
[0010] Optionally, the step of correcting the decision information corresponding to the MRI view based on the decision tag specifically includes: If the decision tag indicates that any of the samples is correctly classified into the second target organ subview, the decision information in the second target organ subview is retained, and otherwise, the decision information in the second target organ subview is corrected based on decision information of other organ subviews other than the second target organ subview.
[0011] The step of selectively correcting the determination information in the second target organ subview based on the determination information of another subview other than the second target organ subview specifically comprises: if the decision tag indicates that any of the samples is correctly classified into a candidate subview other than the second target organ subview, correcting decision information in the second target organ subview based on decision information of the candidate subview; and if the decision tag indicates that any of the samples is correctly classified into a plurality of candidate subviews other than the second target organ subview, correcting the decision information in the second target organ subview based on the decision information of the plurality of candidate subviews by a separation degree calculation.
[0012] Optionally, the step of outputting a medical data classification result of a target object after information fusion of the trained decision information of the CT view and the MRI view is specifically as follows: determining optimal decision information for each of the CT view and the MRI view based on the corrected outputs and corresponding classification tags for each of the CT view and the MRI view; determining one of the CT view and the MRI view as a main view based on a separation error, and determining the other as an auxiliary view, and fusing the determination information of the auxiliary view with the determination information of the main view based on a preset synthesis parameter; and calculating and obtaining a medical data classification result of the target object based on the fused decision information.
[0013] Optionally, the plurality of different organ sub-views are a lateral parotid gland view, a unilateral parotid gland view, a primary nasopharyngeal carcinoma tumor total volume view, and a metastatic lymph node view.
[0014] In a second aspect, an embodiment of the present invention provides a medical data classification apparatus based on mutual correction and information fusion, comprising: an input module for acquiring one or more CT images and / or one or more MRI images of a target object; a classification module for inputting the plurality of CT images and the plurality of MRI images into a fuzzy classification model and obtaining a medical data classification result output from the fuzzy classification model; a training module for training the fuzzy classification model, a CT view correction submodule for training decision information of a CT view according to sample data in a CT image sample data set and corresponding classification tags, the CT image sample data set being respectively used for a scan subview, a first contrast-enhanced subview and a second contrast-enhanced subview in the CT view to perform mutual correction of decision information; an MRI view correction submodule for training decision information of an MRI view based on sample data in an MRI image sample data set and corresponding classification tags, the MRI image sample data being respectively used for a plurality of different organ subviews in the MRI view to perform mutual correction of decision information; and an information fusion submodule for outputting and obtaining a medical data classification result of a target object after information fusion of the trained decision information of the CT view and the MRI view. Effect of the Invention
[0015] The medical data classification method and apparatus based on mutual correction and information fusion provided in the embodiment of the present invention provides a new fuzzy classifier to ensure high interpretability by implementing the constructed training model, and takes the interpretable zero-order TS fuzzy classifier as the basic training unit, and makes the data generated in the model training process such as fuzzy rules, rule output and model output interpretable; based on the characteristics of the high-dimensional features of medical data such as CT and MRI images, the feature dimension is reduced by KL divergence to select more key features from high-dimensional medical data, which further makes the decision more convenient and relieves the pressure of the decision process; from the two views of CT and MRI, the decision information required for classification is corrected by multiple sub-views in each view, and the mutual complementation of the advantages of different observation angles can improve the decision information, thereby improving the accuracy of classification. Furthermore, the CT view and the MRI view are information fused to capture the diverse decision information, and the advantages of each view in the decision aspect are utilized to realize further optimization of the decision information. [Brief description of the drawings]
[0016] In order to more clearly describe the technical solutions in the embodiments of the present application, the drawings used in the embodiments of the present application will be briefly introduced below.
[0017] [Figure 1] 1 is a flowchart of a medical data classification method based on mutual correction and information fusion provided in an embodiment of the present invention. [Diagram 2] FIG. 2 is a schematic diagram of a CT view determination information correction method provided in an embodiment of the present invention; [Diagram 3] 1 is a flowchart of an MRI view determination information correction method provided in an embodiment of the present invention. [Figure 4] 2 is a flowchart of a decision information information fusion method provided in an embodiment of the present invention; [Diagram 5] FIG. 2 is a schematic diagram of the structure of a medical data classification device based on mutual correction and information fusion provided in an embodiment of the present invention; [Figure 6] FIG. 2 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0018] Hereinafter, the technical solutions in the embodiments of the present application will be described with reference to the drawings in the embodiments of the present application.
[0019] In the following drawings, like symbols and letters indicate like items, and therefore, once an item is defined in one drawing, it does not need to be further defined and interpreted in the following drawings. Also, in the description of this application, the terms "first", "second", etc. are merely for distinguishing the description, and cannot be understood as indicating or implying relative importance.
[0020] With the development of artificial intelligence technology, more and more medical data can be used to find relevant pathological features, for example, computed tomography data (CT) or magnetic resonance imaging data (MRI) can be used to find digitized features of relevant tumors. In clinical diagnosis, many attempts have been made to improve the medical data classification capabilities of artificial intelligence tools. For example, more advanced networks such as three-dimensional convolutional neural networks (CNNs), recurrent neural networks based on courtesy gate control recurrent units, support vector machines (SVMs), U-net networks, etc. have been successfully applied to classify various medical data, such as nasopharyngeal carcinoma.
[0021] The key to classifying the above medical data is to obtain favorable key feature information that influences the final decision. The prior art uses the idea of feature selection by genetic algorithm, or the idea of combining clinical features and radiological features based on MRI. However, the above-mentioned method uses the endoscopic image as a single viewing angle, and does not consider the different resolution of the imaging technology of different imaging methods. Considering the actual situation, if the subsequent prior art adopts the multi-view model of CT and MR images, such a model lacks clear interpretation for feature learning, and data images from multiple views often have a large amount of image information, so automatic detection takes a long time, and in some scenes, due to the limitation of the model classification effect, the human intervention of the doctor is always required to determine the classification result, and advanced automatic classification cannot be realized.
[0022] On this basis, an embodiment of the present invention provides a medical data classification method based on mutual correction and information fusion. Figure 1 shows a flowchart of the medical data classification method based on mutual correction and information fusion provided in an embodiment of the present invention.
[0023] In step S110, one or more CT images and / or one or more MRI images of the target object are acquired.
[0024] In the embodiment of the present invention, the target subject is a subject from which pathological information is extracted, and the pathological information is specifically CT images and MRI images. CT has an advantage in observing bones, lungs, bleeding, etc., and MRI has high resolution for soft tissue synovium, blood vessels, nerves, muscles, tendons, ligaments, and transparent cartilage. Therefore, when classifying some pathological information, the above two can complement each other to provide more comprehensive diagnostic information. The pathology in the embodiment of the present invention may be a variety of situations such as tumors, tuberculosis, deformities, and inflammation in the head, chest, abdomen, spine, and limbs, as long as both the CT image and the MRI image can contribute to the pathological information classification. In the subsequent embodiments, the pathological information of nasopharyngeal carcinoma is mainly taken as an example, and overlapping descriptions of the application of other pathological information are omitted.
[0025] Specifically, the CT image and the MRI image output in this step may be either one or multiple, and the two types of image data may be input either individually or jointly.
[0026] When performing medical data classification, the observation information obtained from a single viewing angle of CT or MRI is quite one-sided. Due to the different viewing angles, it is often difficult to obtain disease information that originates from both image features and feedback imaging sites. In an embodiment of the present invention, the information at two observation viewing angles is fine-tuned, i.e., the decision information at the two viewing angles is modified to obtain more comprehensive decision support. In addition, imaging of pathologies such as nasopharyngeal tumors is greatly influenced by the image technology and the organ itself. However, the decision information obtained from the imaging technology lacks focus on the key features of the disease site. In addition, the decision information obtained from the organ perspective does not include important features in the imaging technology. By fully utilizing the key features at the two viewing angles and organically combining the decision information obtained at the two viewing angles, a higher effect can be obtained.
[0027] In step S120, the plurality of CT images and the plurality of MRI images are input into a fuzzy classification model, and a medical data classification result is output from the fuzzy classification model.
[0028] In order to ensure high interpretability by implementing the constructed training model, the embodiment of the present invention provides a new fuzzy classifier, which takes an interpretable zero-order Takagi-Sugeno-Kang (TS) fuzzy classifier as the basic training unit, and realizes the mutual correction and fusion of two kinds of viewing angles. The data generated from the fuzzy rules, rule output and model output training model must be interpretable. The TS fuzzy classifier has natural nonlinear approximation ability and concise language interpretability. Therefore, it is widely used in trajectory tracking control, fault prediction, medicine and other fields.
[0029] As those skilled in the art will appreciate, a TS fuzzy classifier may include an input layer, a membership layer, a rule layer including a multi-layer structure, and an output layer. A zero-order TS fuzzy classifier has excellent nonlinear approximation capabilities and uses "IF-THEN" statements to define rules in the rule layer of a fuzzy system. This classifier always has linguistically interpretable fuzzy rules. The kth rule is
number
number
number
number
number
[0030] Also,
number
number
number
number
number
number
number
number
number
number
[0031] In the embodiment of the present invention, CT and MR images have high-dimensional features, which are directly related to their huge processing requirements. Therefore, the TS fuzzy classifier uses KL divergence (KLIC) as a measurement index to evaluate the similarity of two probability distributions, to reduce the feature dimension and select more key features from high-dimensional medical images, further making the decision more convenient and reducing the pressure of the decision-making process.
[0032] Specifically, the KLIC used in the embodiment of the present invention is the degree to which the uncertainty of the data set information is reduced under the condition of a certain feature, which reflects the feature's capacity to distinguish data samples. It is a common feature selection method and can be expressed as follows:
number
[0033] Data set
number
number
number
number
number
number
number
number
number
number
number
number
[0034] The key to the embodiment of the present invention is the training process of the fuzzy classification model, which specifically includes the following steps S130 to S150:
[0035] In step S130, decision information of a CT view is trained based on sample data in the CT image sample dataset and its corresponding classification tag, and the CT image sample dataset is respectively used for a scan subview, a first contrast-enhanced subview and a second contrast-enhanced subview in the CT view to perform mutual correction of the decision information.
[0036] As shown in the flow chart of the CT view determination information correction method provided in the embodiment of the present invention in FIG. 2, the step S130 specifically includes: Step S131: inputting any sample in the CT image sample data set into a first target sub-view corresponding to the CT view, and obtaining a first actual classification result output from the first target sub-view; A step S132 of obtaining a judgment tag from the first actual classification result and a classification tag corresponding to any of the samples; Step S133: correcting the decision information corresponding to the CT view according to the decision tag until training of all samples is completed; and step S134 of obtaining a final decision of the CT view by a minimum learner method based on the decision information corresponding to the CT view after correction.
[0037] Further, the step S134 specifically includes: if the decision tag indicates that any of the samples is correctly classified into the first target subview, retaining decision information in the first target subview; Otherwise, if the decision tag indicates that any of the samples are correctly classified into a candidate subview other than the first target subview, correct the decision information in the first target subview based on the decision information of the candidate subview; if the decision tag indicates that any of the samples are correctly classified into multiple candidate subviews other than the first target subview, correct the decision information in the first target subview based on the decision information of the multiple candidate subviews by separation degree calculation.
[0038] In step S140, the decision information of the MRI view is trained based on the sample data in the MRI image sample dataset and the corresponding classification tags, and the MRI image sample data is respectively used for multiple different organ sub-views in the MRI view to perform mutual correction of the decision information.
[0039] As shown in the flow chart of the MRI view determination information correction method provided in the embodiment of the present invention in FIG. 3, the step S140 specifically includes: a step S141 of inputting any sample in the MRI image sample data set into a second target organ sub-view corresponding to the MRI view, and obtaining a second actual classification result output from the second target organ sub-view; A step S142 of obtaining a judgment tag from the second actual classification result and a classification tag corresponding to any of the samples; Step S143 of correcting the decision information corresponding to the MRI view based on the decision tag until training of all samples is completed; and step S144 of obtaining a final decision for the MRI view by a minimal learner method based on the decision information corresponding to the corrected MRI view.
[0040] Specifically, step S144 further includes, if the decision tag indicates that any of the samples is correctly classified into the second target organ subview, retaining decision information in the second target organ subview; otherwise, if the decision tag indicates that any of the samples is correctly classified into a candidate subview other than the second target organ subview, correcting decision information in the second target organ subview based on decision information of the candidate subview; If the decision tag indicates that any of the samples is correctly classified into multiple candidate subviews other than the second target organ subview, a separation degree calculation is performed to correct the decision information in the second target organ subview based on the decision information of the multiple candidate subviews.
[0041] From the specific execution process of steps S130 and S140, the embodiment of the present invention trains two training views constituting a fuzzy classification model, that is, a CT view and an MRI view, and both views use the TS fuzzy classifier as a basic training unit. The purpose of the training is to complete the mutual correction of each sub-view in each of the CT view and the MRI view at a single viewing angle, so as to generate decision information for use in the rule layer in the TS fuzzy classifier after correction in each view.
[0042] Here, the CT view includes a scan subview, a first contrast-enhanced subview and a second contrast-enhanced subview, where the scan subview corresponds to medical data obtained from an original CT scan, the first contrast-enhanced subview corresponds to medical data from the CT scan after enhancement with a first contrast parameter, and the second contrast-enhanced subview corresponds to medical data from the CT scan after enhancement with a second contrast parameter.
[0043] The MRI view includes multiple different organ subviews, and the number of organ subviews is determined according to a specific pathology. For example, for a nasopharyngeal carcinoma pathology, the multiple different organ subviews may be four organ subviews, such as a lateral parotid gland view, a unilateral parotid gland view, a primary nasopharyngeal carcinoma tumor total volume view, and a metastatic lymph node view.
[0044] Next, a CT view determination information correction method in specific steps S131 to S134 included in step S130 will be described in detail.
[0045] This step performs decision compensation for three subviews from the perspective of CT imaging technology, and calculates computed tomography (CT), contrast enhanced t1 weighted (CET1) and contrast enhanced t2 weighted (CET2). The TS fuzzy classifier trains these three modules to generate filtering conditions and decision compensation information. The training dataset is
number
number
number
number
number
number
[0046] First, we use FCM clustering to find the training set
number
number
number
number
number
number
number
[0047] Second, select a fuzzy rule based on Eq. 8. ω ik Calculate and set it as the antecedent of rule k.
number
[0048] Next, the output parameters of the hidden layer are calculated using the method described above.
number
number
[0049] As mentioned above, the minimum learning matrix (LLM) method is used to calculate the output weight vector
number
number
number
number
number
number
[0050] Based on the above method, the training is completed with the viewing angle of CT imaging technology. The decision information is the key to fuzzy classification, and the optimized classifier is the decision information
number
number
number
[0051] H (1,1) , H (1,2) and H (1,3) are the output matrix parameters of the CT, CET1 and CET2 models, respectively, and Y (1,1) , Y (1,2) and Y (1,3) are the corresponding real outputs. The output class tag set is
number
number
number
number
[0052] T (1,1) , T(1,2) and T (1,3) are the output classification tag sets of the CT, CET1, and CET2 models, respectively. Taking the CT module as an example, the class tags output by the CT model are
number
number
number
[0053] The decision information is processed sequentially to obtain the decision parameter Z (1,1) Calculate and correct.
number
[0054] Case A: If a single training unit is correctly identified (i.e.,
number
number
[0055] Case B: When a single training unit is misidentified (i.e.,
number
number
number
[0056] Case B-1: When one training unit accurately recognizes one view (i.e., n (1) =1) Decision information selected only if classification p1′ of image technology tag matches
number
number
number
[0057] Case B-2: When the training unit's accurate recognition of a certain view is not unique (i.e., n (1) ≠1) Furthermore, a set of image technologies p1 that meet the conditions
number
number
number
number
[0058] degree of separation
number
number
number
number
number
number
[0059] After correcting the decision information, the consequent parameters are generated using LLM.
[0060] Similarly, an MRI view is similar to a CT view, and next, a method for correcting CT view determination information in specific steps S131 to S134 included in step S130 will be described in detail.
[0061] Each sample is observed using four organ subviews: lateral parotid gland subview (CPG), unilateral parotid gland subview (IPG), total primary nasopharyngeal carcinoma tumor volume subview (GTVnp), and metastatic lymph node subview (GTVn). Again, nasopharyngeal carcinoma pathology is used as the standard, and other pathologies are similar and will not be described.
[0062] Then, by comparing the classification results, the decision information of each subview is corrected and the decision information of each module is optimized.
number
number
number
number
number
[0063] According to equations (7) to (14), TS fuzzy training is performed on the MRI views, and the decision information
number
number
number
number
[0064] Output matrix parameters H of CPG, IPG, GTVnp, and GTVn modules (2,1) , H (2,2) , H (2,3) andH (2,4) and the actual output Y (2,1) , Y (2,2) , Y (2,3) andY (2,4) , and the output class tag set T (2,1) , T (2,2) , T (2,3) andT (2,4) It is known that the output class tag results of CPG modules are
number
[0065] Case C: If a single training unit is correctly identified (i.e.,
number
[0066] Case D: A single training unit is misidentified (i.e.,
number
[0067] Case D-1: When one training unit accurately identifies one view (i.e., n (2) =1) Decision information selected only if the classification of organ tag p′2 matches
number
number
[0068] Case D-2: The training unit is not unique in correctly identifying one view (i.e., n (2) ≠1) Furthermore, the set of organ subviews p2 that meet the conditions
number
number
number
number
number
[0069] After correcting the decision information, the subsequent parameters are calculated by LLM.
number
number
[0070] In step S150, after the trained decision information of the CT view and the MRI view is information fused, the medical data classification result of the target object is obtained as an output.
[0071] Specifically, as shown in the flowchart of the information fusion method for decision information in FIG. 4, step S150 is further divided into steps S151 to S153.
[0072] In step S151, optimal decision information of each of the CT view and the MRI view is determined based on the corrected outputs of each of the CT view and the MRI view and corresponding classification tags.
[0073] In step S152, one of the CT view and the MRI view is determined as a main view based on a separation error, and the other is determined as an auxiliary view, and the determination information of the auxiliary view is fused with the determination information of the main view based on a preset synthesis parameter.
[0074] In step S153, a medical data classification result of the target object is calculated based on the fused decision information.
[0075] In the embodiment of the present invention, the classification of various pathologies relies on CT image information and MRI organ tissue structure information. In the absence of doctor intervention, the two perspectives interact with each other and cannot provide mutual support for decision information. From the perspective of CT imaging technology alone, the pathology-specific characteristics of the examined organ are ignored. Similarly, from the perspective of MRI organ image data alone, the functional space differences of each CT imaging technology are not taken into account. After performing internal correction for the two perspectives, the decision information is adjusted alternately. The specific implementation steps are as follows:
[0076] First, before synthesizing the information of two views, it is necessary to perform internal selection on the optimization decision information provided by each view, and determine the synthesis parameters of each view to simplify the subsequent synthesis process. Taking the CT view as an example, the optimized output
number
number
number
number
number
number
[0077] Considering the representation of the final decision from the two views, the model takes the decision information of one view as the center and coordinates the information of the other view with the center. The separation error v is a key selection parameter applied to the main decision information. The phase separation errors of the two views are
number
number
number
[0078] Further, more was obtained
number
number
number
number
[0079] After correcting the decision information, consequent parameters are generated and are the key parameters of the testing process that subsequently direct the synthesis of the training algorithm of the embodiment of the present invention.
[0080] The medical data classification method based on mutual correction and information fusion provided in the embodiment of the present invention provides a new fuzzy classifier to ensure high interpretability by implementing the constructed training model, and takes the interpretable zero-order TS fuzzy classifier as the basic training unit, and makes the data generated in the model training process such as fuzzy rules, rule output and model output interpretable. Based on the characteristics of the high-dimensional features of medical data such as CT and MRI images, the feature dimension is reduced by KL divergence to select more key features from high-dimensional medical data, which further makes the decision more convenient and relieves the pressure of the decision process. From the two views of CT and MRI, the decision information required for classification is corrected by multiple sub-views in each view, and the mutual complementation of the advantages of different observation angles can improve the decision information, thereby improving the accuracy of classification. Furthermore, the CT view and the MRI view are information fused to capture the diverse decision information, and the advantages of each view in the decision aspect are utilized to realize further optimization of the decision information.
[0081] Based on any of the above embodiments, FIG. 5 illustrates a medical data classification apparatus based on mutual correction and information fusion provided in an embodiment of the present invention, the apparatus comprising: an input module 510 for acquiring one or more CT images and / or one or more MRI images of the target object; a classification module 520 for inputting the plurality of CT images and the plurality of MRI images into a fuzzy classification model and obtaining a medical data classification result output from the fuzzy classification model; A training module 530 for training the fuzzy classification model, comprising: a CT view correction submodule 531 for training decision information of a CT view according to sample data in a CT image sample data set and its corresponding classification tag, the CT image sample data set being respectively used for a scan subview, a first contrast-enhanced subview and a second contrast-enhanced subview in the CT view to perform mutual correction of decision information; an MRI view correction submodule 532 for training decision information of an MRI view based on sample data in an MRI image sample dataset and corresponding classification tags, the MRI image sample data being respectively used for a number of different organ subviews in the MRI view to perform mutual correction of decision information; and a training module 530 including an information fusion sub-module 533 for outputting and obtaining a medical data classification result of a target object after information fusion of the trained decision information of the CT view and the MRI view.
[0082] The medical data classification device based on mutual correction and information fusion provided in the embodiment of the present invention provides a new fuzzy classifier to ensure high interpretability by implementing the constructed training model, and takes the interpretable zero-order TS fuzzy classifier as the basic training unit, and makes the data generated in the model training process such as fuzzy rules, rule output and model output interpretable. Based on the characteristics of the high-dimensional features of medical data such as CT and MRI images, the feature dimension is reduced by KL divergence to select more key features from high-dimensional medical data, which further makes the decision more convenient and relieves the pressure of the decision process. From the two views of CT and MRI, the decision information required for classification is corrected by multiple sub-views in each view, and the mutual complementation of the advantages of different observation angles can improve the decision information, thereby improving the accuracy of classification. Furthermore, the CT view and the MRI view are information fused to capture the diverse decision information, and the advantages of each view in the decision aspect are utilized to realize further optimization of the decision information.
[0083] Based on any of the above embodiments, Figure 6 shows a schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention, which may include a processor 610, a communications interface 620, a memory 630, and a communications bus 640, and the processor 610, the communications interface 620, and the memory 630 communicate with each other via the communications bus 640. The processor 610 can call the logic instructions in the memory 630 to perform the following method:
[0084] acquiring one or more CT images and / or one or more MRI images of the target object; inputting the plurality of CT images and the plurality of MRI images into a fuzzy classification model; and obtaining a medical data classification result output from the fuzzy classification model; Wherein, the training process of the fuzzy classification model is training decision information of a CT view based on sample data in a CT image sample data set and corresponding classification tags, the CT image sample data set being respectively used for a scan sub-view, a first contrast-enhanced sub-view and a second contrast-enhanced sub-view in the CT view to perform mutual correction of the decision information; training decision information for an MRI view based on sample data in a MRI image sample data set and corresponding classification tags, the MRI image sample data being respectively used for a number of different organ sub-views in the MRI view to perform mutual correction of the decision information; and outputting and obtaining a medical data classification result of a target object after information fusion of the trained decision information of the CT view and the MRI view respectively.
[0085] Also, the logic instructions in the memory 630 can be stored in a computer-readable storage medium when the logic instructions are realized in the form of a software functional unit and sold or used as an independent product. Based on this understanding, the technical solutions of the embodiments of the present invention can be essentially or in part of the contribution to the prior art, or a part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method according to the embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a U disk, a portable hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0086] On the other hand, an embodiment of the present invention further provides a non-transitory computer-readable storage medium storing a computer program, which when executed by a processor, performs the method provided in the above embodiment, e.g. acquiring one or more CT images and / or one or more MRI images of the target object; inputting the plurality of CT images and the plurality of MRI images into a fuzzy classification model, and obtaining a medical data classification result output from the fuzzy classification model; Wherein, the training process of the fuzzy classification model is training decision information of a CT view based on sample data in a CT image sample data set and corresponding classification tags, the CT image sample data set being respectively used for a scan sub-view, a first contrast-enhanced sub-view and a second contrast-enhanced sub-view in the CT view to perform mutual correction of the decision information; training decision information for an MRI view based on sample data in a MRI image sample data set and corresponding classification tags, the MRI image sample data being respectively used for a number of different organ sub-views in the MRI view to perform mutual correction of the decision information; and outputting the medical data classification result of the target object after information fusion of the trained decision information of the CT view and the MRI view.
[0087] The above-described embodiment of the device is merely illustrative, in which the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., they may be located in one place or distributed among multiple network units. Some or all of the modules may be selected according to actual needs to achieve the objective of the solution of the embodiment. Those skilled in the art can understand and implement it without any creative effort.
[0088] From the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be realized by adding a required general-purpose hardware platform to software, and of course by hardware. Based on such understanding, the above technical solutions can be essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product is stored in a computer-readable storage medium such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions for making a computer device (which may be a personal computer, a server, or a network device, etc.) execute the method described in each embodiment or a certain part of the embodiment.
[0089] Finally, it should be noted that the above embodiments are only for illustrating the technical solutions of the present invention, and are not intended to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still modify the technical solutions described in the above embodiments or replace some of the technical features with equivalents, and such modifications or replacements do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.
[0090] (Additional Note) (Appendix 1) acquiring one or more CT images and / or one or more MRI images of the target object; inputting the plurality of CT images and the plurality of MRI images into a fuzzy classification model, and obtaining a medical data classification result output from the fuzzy classification model; Wherein, the training process of the fuzzy classification model is training decision information of a CT view based on sample data in a CT image sample data set and corresponding classification tags, the CT image sample data set being respectively used for a scan sub-view, a first contrast-enhanced sub-view and a second contrast-enhanced sub-view in the CT view to perform mutual correction of the decision information; training decision information for an MRI view based on sample data in a MRI image sample data set and corresponding classification tags, the MRI image sample data being respectively used for a number of different organ sub-views in the MRI view to perform mutual correction of the decision information; and outputting and obtaining a medical data classification result of the target object after performing information fusion on the trained decision information of the CT view and the MRI view.
[0091] (Appendix 2) The step of training CT view determination information based on sample data in the CT image sample dataset and corresponding classification tags specifically includes: inputting any sample in the CT image sample dataset into a first target sub-view corresponding to the CT view, and obtaining a first actual classification result output from the first target sub-view; obtaining a judgment tag from the first actual classification result and a classification tag corresponding to any of the samples; correcting decision information corresponding to the CT view based on the decision tag until training of all samples is completed; and obtaining a final decision for the CT view by a minimal learner method based on decision information corresponding to the corrected CT view.
[0092] (Appendix 3) The step of correcting the decision information corresponding to the CT view based on the decision tag specifically includes: The medical data classification method of claim 2, further comprising the step of retaining decision information in the first target subview if the decision tag indicates that any of the samples is correctly classified into the first target subview, and otherwise correcting the decision information in the first target subview based on decision information of other subviews other than the first target subview.
[0093] (Appendix 4) Specifically, the step of correcting the determination information in the first target subview based on the determination information of another subview other than the first target subview includes: if the decision tag indicates that any of the samples is correctly classified into a candidate subview other than the first target subview, correcting decision information in the first target subview based on decision information of the candidate subview; and correcting decision information in the first target subview based on decision information of the multiple candidate subviews by separation degree calculation when the decision tag indicates that any of the samples is correctly classified into multiple candidate subviews other than the first target subview.
[0094] (Appendix 5) Specifically, the step of training MRI view decision information based on sample data in the MRI image sample dataset and corresponding classification tags includes: inputting any sample in the MRI image sample dataset into a second target organ sub-view corresponding to the MRI view, and obtaining a second actual classification result output from the second target organ sub-view; obtaining a decision tag from the second actual classification result and a classification tag corresponding to any of the samples; correcting decision information corresponding to the MRI view based on the decision tag until training of all samples is completed; and obtaining a final decision for the MRI view by a minimal learner method based on decision information corresponding to the corrected MRI view.
[0095] (Appendix 6) Specifically, the step of correcting the decision information corresponding to the MRI view based on the decision tag includes: The medical data classification method of claim 5, further comprising the step of: retaining decision information in the second target organ subview if the decision tag indicates that any of the samples is correctly classified into the second target organ subview; and otherwise correcting the decision information in the second target organ subview based on decision information of other organ subviews other than the second target organ subview.
[0096] (Appendix 7) Specifically, the step of correcting the determination information in the second target organ subview based on the determination information of another subview other than the second target organ subview includes: if the decision tag indicates that any of the samples is correctly classified into a candidate subview other than the second target organ subview, correcting decision information in the second target organ subview based on decision information of the candidate subview; The medical data classification method of claim 6, further comprising a step of correcting decision information in the second target organ subview based on decision information of the multiple candidate subviews by separation degree calculation when the decision tag indicates that any of the samples is accurately classified into multiple candidate subviews other than the second target organ subview.
[0097] (Appendix 8) The step of outputting the medical data classification result of the target object after information fusion of the trained decision information of the CT view and the MRI view is specifically as follows: determining optimal decision information for each of the CT view and the MRI view based on the corrected outputs and corresponding classification tags for each of the CT view and the MRI view; determining one of the CT view and the MRI view as a main view based on a separation error, and determining the other as an auxiliary view, and fusing the determination information of the auxiliary view with the determination information of the main view based on a preset synthesis parameter; and calculating and obtaining a medical data classification result for the target object based on the fused decision information.
[0098] (Appendix 9) 2. The medical data classification method of claim 1, wherein the plurality of different organ subviews are a lateral parotid gland view, a unilateral parotid gland view, a primary nasopharyngeal carcinoma tumor total volume view, and a metastatic lymph node view.
[0099] (Appendix 10) an input module for acquiring one or more CT images and / or one or more MRI images of a target object; a classification module for inputting the plurality of CT images and the plurality of MRI images into a fuzzy classification model and obtaining a medical data classification result output from the fuzzy classification model; a training module for training the fuzzy classification model, a CT view correction submodule for training decision information of a CT view according to sample data in a CT image sample data set and corresponding classification tags, the CT image sample data set being respectively used for a scan subview, a first contrast-enhanced subview and a second contrast-enhanced subview in the CT view to perform mutual correction of decision information; an MRI view correction submodule for training decision information of an MRI view based on sample data in an MRI image sample data set and corresponding classification tags, the MRI image sample data being respectively used for a plurality of different organ subviews in the MRI view to perform mutual correction of decision information; and an information fusion submodule for outputting and obtaining a medical data classification result of a target object after information fusion of the trained decision information of the CT view and the MRI view.
Claims
1. acquiring a plurality of CT images and a plurality of MRI images of the target object; inputting the plurality of CT images and the plurality of MRI images into a fuzzy classification model, and obtaining a medical data classification result output from the fuzzy classification model; Wherein, the training process of the fuzzy classification model is training decision information of a CT view based on sample data in a CT image sample data set and corresponding classification tags, the CT image sample data set being respectively used for a scan sub-view, a first contrast-enhanced sub-view and a second contrast-enhanced sub-view in the CT view to perform mutual correction of the decision information; training decision information for an MRI view based on sample data in an MRI image sample data set and corresponding classification tags, the MRI image sample data being respectively used for a number of different organ sub-views in the MRI view to perform mutual correction of the decision information; and outputting a medical data classification result of the target object after information fusion of the trained decision information of the CT view and the MRI view. The step of training CT view decision information based on sample data in the CT image sample dataset and its corresponding classification tag specifically includes: inputting any sample in the CT image sample dataset into a first target sub-view corresponding to the CT view, and obtaining a first actual classification result output from the first target sub-view; obtaining a decision tag from the first actual classification result and a classification tag corresponding to any of the samples; correcting decision information corresponding to the CT view based on the decision tag until training of all samples is completed; and obtaining a final decision of the CT view by a minimal learner method based on the decision information corresponding to the corrected CT view; The step of training MRI view decision information based on sample data in the MRI image sample dataset and corresponding classification tags specifically includes: inputting any sample in the MRI image sample dataset into a second target organ sub-view corresponding to the MRI view, and obtaining a second actual classification result output from the second target organ sub-view; obtaining a decision tag from the second actual classification result and a classification tag corresponding to any of the samples; correcting decision information corresponding to the MRI view based on the decision tag until training of all samples is completed; and obtaining a final decision of the MRI view by a minimal learner method based on the decision information corresponding to the corrected MRI view; The step of outputting a medical data classification result of a target object after information fusion of the trained decision information of the CT view and the MRI view is specifically as follows: determining optimal decision information for each of the CT view and the MRI view based on the corrected outputs and corresponding classification tags for each of the CT view and the MRI view; determining one of the CT view and the MRI view as a main view based on a separation error, and determining the other as an auxiliary view, and fusing the determination information of the auxiliary view with the determination information of the main view based on a preset synthesis parameter; and calculating and obtaining a medical data classification result of the target object based on the fused decision information.
2. Specifically, the step of correcting the decision information corresponding to the CT view based on the decision tag includes:
2. The method of claim 1, further comprising the step of: retaining decision information in the first target subview if the decision tag indicates that any of the samples is correctly classified into the first target subview; and otherwise correcting the decision information in the first target subview based on decision information of other subviews other than the first target subview.
3. Specifically, the step of correcting the determination information in the first target subview based on the determination information of another subview other than the first target subview includes: if the decision tag indicates that any of the samples is correctly classified into a candidate subview other than the first target subview, correcting decision information in the first target subview based on decision information of the candidate subview; 3. The method of claim 2, further comprising: if the decision tag indicates that any of the samples is correctly classified into a plurality of candidate subviews other than the first target subview, correcting decision information in the first target subview based on decision information of the plurality of candidate subviews by a separation degree calculation.
4. The step of correcting the decision information corresponding to the MRI view based on the decision tag specifically includes:
2. The method of claim 1, further comprising the step of: retaining decision information in the second target organ subview if the decision tag indicates that any of the samples is correctly classified into the second target organ subview; and otherwise correcting the decision information in the second target organ subview based on decision information of other organ subviews other than the second target organ subview.
5. Specifically, the step of correcting the determination information in the second target organ subview based on the determination information of another subview other than the second target organ subview includes: if the decision tag indicates that any of the samples is correctly classified into a candidate subview other than the second target organ subview, correcting decision information in the second target organ subview based on decision information of the candidate subview; 5. The medical data classification method of claim 4, further comprising: if the decision tag indicates that any of the samples is correctly classified into a plurality of candidate subviews other than the second target organ subview, correcting the decision information in the second target organ subview based on the decision information of the plurality of candidate subviews by a separation degree calculation.
6. 2. The method of claim 1, wherein the plurality of different organ sub-views are a lateral parotid gland view, a unilateral parotid gland view, a primary nasopharyngeal carcinoma tumor total volume view, and a metastatic lymph node view.
7. an input module for acquiring a plurality of CT images and a plurality of MRI images of a target object; a classification module for inputting the plurality of CT images and the plurality of MRI images into a fuzzy classification model and obtaining a medical data classification result output from the fuzzy classification model; a training module for training the fuzzy classification model, a CT view correction submodule for training decision information of a CT view based on sample data in a CT image sample data set and its corresponding classification tag, the CT image sample data set being respectively used for a scan subview, a first contrast-enhanced subview and a second contrast-enhanced subview in the CT view to perform mutual correction of decision information; an MRI view correction sub-module for training decision information of an MRI view based on sample data in an MRI image sample dataset and its corresponding classification tag, the MRI image sample data being respectively used for a plurality of different organ sub-views in the MRI view to perform mutual correction of decision information; and an information fusion sub-module for outputting and obtaining a medical data classification result of a target object after information fusion of the trained decision information of the CT view and the MRI view, Specifically, training the CT view decision information based on the sample data in the CT image sample dataset and the corresponding classification tags includes: inputting any sample in the CT image sample dataset into a first target sub-view corresponding to the CT view, and obtaining a first actual classification result output from the first target sub-view; obtaining a judgment tag from the first actual classification result and a classification tag corresponding to any of the samples; correcting decision information corresponding to the CT view based on the decision tag until training of all samples is completed; and obtaining a final decision of the CT view by a minimum learner method based on the decision information corresponding to the corrected CT view; Specifically, training the MRI view decision information based on the sample data in the MRI image sample dataset and the corresponding classification tags includes: inputting any sample in the MRI image sample dataset into a second target organ sub-view corresponding to the MRI view, and obtaining a second actual classification result output from the second target organ sub-view; obtaining a judgment tag from the second actual classification result and a classification tag corresponding to any of the samples; correcting decision information corresponding to the MRI view based on the decision tag until training of all samples is completed; and obtaining a final decision of the MRI view by a minimal learner method based on the decision information corresponding to the corrected MRI view; The trained decision information of the CT view and the MRI view are respectively information fused, and then the medical data classification result of the target object is outputted to obtain the result, specifically, determining optimal decision information for each of the CT view and the MRI view based on the corrected outputs and corresponding classification tags for each of the CT view and the MRI view; determining one of the CT view and the MRI view as a main view based on a separation error, and determining the other as an auxiliary view, and fusing the determination information of the auxiliary view with the determination information of the main view based on a preset synthesis parameter; and calculating and obtaining a medical data classification result of the target object based on the fused decision information.
Citation Information
Patent Citations
Universal medical image report generation method and system, electronic equipment and readable storage medium
CN116631584A
Automatic classification of liver disease severity from noninvasive radiological imaging
JP2023501282A
