Characterization, blood data and medical image multi-data approximation analysis and multiple suspected disease comprehensive identification method
Patent Information
- Application Number
- CN202180087224.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-14
- Publication Date
- 2025-03-21
AI Technical Summary
In the current medical data analysis, the accuracy of intelligent identification of single data is low, time-consuming and labor-intensive, and disease identification is inaccurate. In particular, the identification efficiency of multiple types of diseases and similar diseases is low, easy to miss, and difficult to accurately distinguish.
Through multi-data proximity analysis of characterization, blood data, and medical images, we use improved deep neural network methods and Apriori algorithms to create disease-feature-eigenvalue association models, calculate disease feature similarity, and combine with tabu search methods to dynamically adjust parameters. , to identify multiple types of suspected diseases.
It improves the efficiency of intelligent identification of medical data, accurately identifies multiple types and highly similar suspected diseases, solves the problem of inaccurate identification and omissions, and achieves efficient multi-data comprehensive identification.
Smart Images
Figure 00000009_0000 
Figure 00000010_0000 
Figure 00000011_0000
Abstract
Description
Representation, blood data, medical image multi-data approximation analysis and comprehensive identification methods for multiple suspected diseases Technical field:
[0001] The present invention belongs to the field of artificial intelligence technology, and relates to data analysis technology, image intelligent recognition technology, and technology and methods combining artificial intelligence with medical data analysis. Background technology:
[0002] With the advancement of medical examinations and medical technology, disease analysis and identification require a wide range of medical data, including disease feature data such as medical images, blood data, microscopic blood images, and organ representations. Medical image data includes CT, ultrasound, and MRI images. Current intelligent recognition of single-item data suffers from low accuracy, time-consuming and labor-intensive tasks, and inaccurate manual analysis and identification. Recognition efficiency is low for blood test and image data, leading to the emergence of multiple suspected diseases. In particular, high disease similarity leads to inaccurate disease identification, missed disease identification, and difficulty in identifying multiple diseases. Multi-data fusion is crucial for high-precision intelligent identification of multiple and similar diseases. By fusing representations, blood data, and medical images, disease-feature correlations are determined and association rules are extracted for multiple diseases across multiple feature types. Based on the association rules between multiple diseases and disease features (blood, representations, and images), a disease-feature-feature value association model is created for these multiple data types. Based on the correlation between diseases and disease features, extraction rules are used to determine the compatibility of diseases with blood, representations, and image features. Based on the degree of fit, the disease and feature similarity method is calculated to distinguish and identify suspected diseases with high similarity.
[0003] For various diseases, we identify highly similar features from the representation set, blood set, and image feature set, and place them into a candidate set. We then swap similar features within the candidate set, set a taboo table, and place the swapped features into the taboo table, setting the length of the taboo table. Dynamic features are used as neural network inputs, along with their probabilities, and fed into the input layer. Using an improved deep neural network approach, we adjust parameters for intelligent disease identification. This dynamic input is fed into the improved neural network disease recognition framework, improving the efficiency of intelligent disease identification.
[0004] Summary of the invention:
[0005] The present invention addresses the problems of inaccurate analysis and recognition of medical data, inability to identify multiple types of suspected diseases, many omissions in identification, time-consuming and labor-intensive problems, and comprehensively identifies multiple types of suspected diseases. It solves the problems of existing technologies such as identification of single diseases, omission of suspected diseases, difficulty in identifying multiple data, and inaccurate data identification judgment. Comprehensive identification of multiple types of suspected diseases has become an important effect.
[0006] A comprehensive identification method for multiple suspected diseases utilizes multi-data similarity analysis using representations, blood data, and medical images. For each disease, highly similar features are found in the representation set, blood data set, and image feature set, and these are placed into a candidate set. Similar features in the candidate set are swapped, a taboo table is established, and the swapped set is placed into the taboo table, with a specific length. Dynamic features are used as neural network inputs, along with their probabilities, and fed into the input layer. Using an improved deep neural network approach, parameters are adjusted for intelligent disease identification. This dynamic input is fed into the improved neural network disease recognition framework, improving the efficiency of intelligent disease identification.
[0007] The technical solution adopted in the present invention is:
[0008] Characterization, blood data, medical image multi-data similarity analysis and comprehensive identification of multiple suspected diseases. The multi-feature data includes: blood data, organ characterization data, blood cell images, organ tissue medical image data and other data. The improved Apriori method is used to determine the disease-feature correlation, and the association rules of multiple diseases for multiple types of features are extracted. Based on the association rules of multiple diseases and disease features (blood, characterization, images), a multi-class data disease-feature-feature value association model is created. The improved Apriori method is used as follows:
[0009] Step 1. Disease is represented by Du, u∈{0, 1...U}.
[0010] Step 2. The characteristics of the disease are represented by Fi, i∈{0, 1...I}.
[0011] Step 3. The details of the disease are represented by 1-2 digit codes.
[0012] Step 4. The characteristic field is represented by a code of 3 to 20 digits.
[0013] Step 5. The disease feature-feature value combination represents the accuracy of the feature item in identifying the disease.
[0014] Characterization, blood data, medical image multi-data similarity analysis and comprehensive identification method of multiple suspected diseases, characterized by comprising a similar disease - disease feature similarity method and result analysis model, the disease feature similarity similarity calculation method steps are as follows:
[0015] Based on the correlation between the disease and disease features, the extraction rules are used to determine the degree of fit between the disease and blood, manifestation, image and other features. Based on the fit, the disease and feature similarity method is calculated to distinguish and identify suspected diseases with high similarity.
[0016] Assuming disease u and disease v, let N(u) represent the feature set of disease u, let N(v) be the feature set of disease v, and use cosine similarity to calculate the formula of disease similarity:
[0017]
[0018] After obtaining the similarity between disease u and disease v, the recognition accuracy of disease u for feature i is measured again, P(u, i):
[0019]
[0020] Where S(u, K) is the set of K disease features that are closest to disease u and disease v using the Apriori method, and whose matching degree is greater than 0.5, N(i) is the set of diseases with feature i, Wuv is the similarity between disease u and disease v, and rvi is the implicit feedback information, which represents the accuracy of disease v's recognition of feature i. To simplify the calculation, rvi can be set to 1.
[0021] After calculating Pu for all disease features, Pu can be sorted in descending order, and the top N disease features can be assigned to disease u. Based on the number of features, the disease is calculated to determine whether the disease has a high degree of similarity. Suspected diseases with a high degree of similarity based on disease features are found.
[0022] Characterization, blood data, medical image multi-data similarity analysis and comprehensive identification method for multiple suspected diseases, characterized by comprising a multi-category data multi-disease feature fusion model, using an improved tabu search (Tabu-Search) method, for multiple diseases, respectively, in the characterization set, blood set, image feature set, find features with high similarity and put them into the candidate set, exchange the same type of feature items in the candidate set, set a tabu table, put the exchanged set into the tabu table, and set the length of the tabu table. The dynamic feature items are read into the input layer as the input items of the neural network and their probabilities, and the parameters are adjusted through the improved deep neural network method to intelligently identify the disease. Dynamic input is input into the improved neural network disease recognition framework, and the recognition method steps are as follows:
[0023] Step 1: Initialize the disease model Du{F1, F2, F3, ..., Fi}.
[0024] Step 2: Calculate similarity based on disease-feature-feature value.
[0025] Step 3: For disease Du, in the representation set {C1, C2...Cn}, the blood set {B1, B2,...Bn}, and the image feature set {P1, P2,...Pn}, for features CX1, BY1, and PZ1, find highly similar features CX2, BY2, and PZ2 to swap with them. Find Du-Fi and put it into the {CX1-CX2, BY1-BY2, PZ1-PZ2} Adjustment-Fi candidate solution swap set.
[0026] Step 4: Exchange {CX1-CX2, BY1-BY2, PZ1-PZ2} in the Adjustment-Fi set and put them into the feature input set {F1, F2, F3.....Fi}.
[0027] Step 5: Set up a tabu table, put the exchange set into the tabu table, and set the length of the tabu table.
[0028] Step 6: Read the dynamic feature items as the input items of the neural network and their probabilities into the input layer.
[0029] Step 6: Determine whether the input item is a hard input item or a soft input item.
[0030] Step 7: Set the hard disease recognition probability as the hard indicator and the soft disease recognition probability as the soft indicator.
[0031] Step 8: Use the improved deep neural network algorithm to train samples, adaptively learn, adjust parameters, and identify diseases. The activation function is the sum of the weight of each hard input item * the hard disease recognition probability and the weight of each soft input item * the soft disease recognition probability.
[0032] Step 9: Use deep neural network method to adaptively adjust weight parameters.
[0033] Step 10. Output the results and determine the disease.
[0034] Step 11: Output various diseases and their disease identification probability, and increase iterations.
[0035] Step 12. Objective function: maximize the total probability.
[0036] In summary, the beneficial effects of the present invention are:
[0037] The present invention addresses the problems of inaccurate medical data analysis and recognition, a large number of suspected diseases, long identification and judgment time, inability to identify, many omissions in identification, time-consuming and labor-intensive problems. The present invention can comprehensively identify multiple suspected diseases by finding diseases with high similarity, effectively solving the problem of missing suspected diseases. Multi-data recognition, multiple high similarity, and comprehensive identification of suspected diseases become important effects.
[0038] All diseases are difficult to identify effectively with the naked eye, especially those with high similarity, which can delay diagnosis and treatment. A comprehensive identification method for multiple suspected diseases that integrates representational data, blood data, and medical image data can efficiently identify multiple diseases and multiple suspected diseases with high similarity, achieving highly efficient comprehensive identification of multiple diseases. Description of the drawings:
[0039] FIG1 is a flow chart of the method for matching diseases with similar characteristics in this application specification.
[0040] FIG2 is a flowchart of a method for calculating the similarity matrix of diseases a, b, and c in this application specification.
[0041] FIG3 is a schematic diagram of the calculation flow of the disease, feature, and feature value association model for various types of data in this application specification.
[0042] Figure 4 is a flow chart of the method for identifying suspected diseases using high-similarity feature input and improved deep neural network in this application specification.
[0043] Example 1:
[0044] As shown in Figure 1 and Figure 2, a comprehensive identification method for multiple types of suspected diseases is integrated with characterization data, blood data, and medical image data. The steps and implementation methods are taken as an example, but the implementation methods are not limited to this. The steps and methods of the embodiment are as follows:
[0045] The Apriori method is used to determine the disease-feature correlation and extract the association rules between multiple diseases and multiple types of features. Based on the association rules between multiple diseases and disease features (blood, manifestation, image), a disease-feature-feature value association model for multiple types of data is created using the Apriori method. The following steps are taken:
[0046] S1. Disease is represented by Du, u∈{0, 1...U}.
[0047] S2. The characteristics of the disease are represented by Fi, i∈{0, 1...I}.
[0048] S3. The details of the disease are represented by 1-2 digit codes.
[0049] S4. The characteristic field to which it belongs is represented by a code of 3 to 20 digits.
[0050] S5. The disease feature-feature value combination represents the accuracy of the feature item in identifying the disease.
[0051] According to the designed model and algorithm, the association rule extraction calculation example is as follows:
[0052] Disease codes in the calculation example: 1, 2, 3
[0053] Disease characteristic codes: 11356, 22456, 33678, 44890, 5590, 6672, 7765, 8843.
[0054] Calculate the parameters generateRules(L, supportData, minConf=0.05)
[0055] Output the association relationship results:
[0056] frozenset({2})-->frozenset({22456})conf: 0.4
[0057] ......
[0058] frozenset({3})-->frozenset({33678})conf: 0.2758620689655173
[0059] Based on the correlation between the disease and disease features, the degree of fit between the disease and blood, representation, image and other features is determined according to the extraction rules. Based on the degree of fit, the disease and feature similarity method is calculated. Assuming disease u and disease v, let N(u) represent the feature set of disease u, and let N(v) be the feature set of disease v. Then use cosine similarity to calculate the disease similarity formula:
[0060]
[0061] After obtaining the similarity between disease u and disease v, the recognition accuracy of disease u for feature i is measured again, P(u, i):
[0062]
[0063] Where S(u, K) is the set of K disease features that are closest to disease u and disease v with a matching degree greater than 0.5, N(i) is the disease set of feature i, Wuv is the similarity between disease u and disease v, and rvi is the implicit feedback information, which represents the accuracy of disease v's recognition of feature i. To simplify the calculation, rvi can be set to 1.
[0064] After calculating Pui for all disease features, Pui is sorted in descending order and the top N disease features are taken as the recommendation results to be displayed to disease u, where N ranges from 1 to 200. The calculation process of the disease feature similarity matrix for disease a, disease b, and disease c is as follows:
[0065] Consider the similarity calculation example for three diseases and eight diseases. The similarity method uses the following calculation example, which uses an 11x11 matrix. Rows and columns indexed 0, 1, and 2 represent the disease source, while rows and columns indexed 3-10 represent similar diseases. A matrix of disease n and disease m is input, with the intersection of row m and column n representing the disease feature count. Based on the feature count, the similarity between the two diseases is calculated. Suspected diseases with a high degree of similarity are identified based on disease features.
[0066] As shown in Figures 3 and 4, the disease model Du {F1, F2, F3, ..., Fi} is initialized. Similarity is calculated based on disease-feature-feature value. For each disease, the highly similar features {F1, F2, F3, ..., Fi} are found in the representation set {C1, C2...Cn}, the blood set {B1, B2, ...Bn}, and the image feature set {P1, P2, ...Pn}. The dynamic feature items are used as inputs to the neural network, along with their probabilities, and are read into the input layer. The pairs {CX1-CX2, BY1-BY2, PZ1-PZ2} that need to be swapped are found and placed in the Adjustment-Fi set. The {CX1-CX2, BY1-BY2, PZ1-PZ2} in the Adjustment-Fi set are swapped and placed in the features {F1, F2, F3, ..., Fi}. A tabu table is set for the input set, and the swapped pairs are placed in the tabu table. The length of the tabu table is set. All feature inputs are determined to be hard or soft. The hard disease recognition probability is set as a hard indicator. The soft disease recognition probability is set as a soft indicator. Using an improved deep neural network algorithm, training samples, adaptive learning, and parameter adjustment are used to identify diseases. The activation function is: the sum of the weight of each hard input item * the hard disease recognition probability and the weight of each soft input item * the soft disease recognition probability. Using a deep neural network method, the weight parameters are adaptively adjusted. The results are output and the disease is determined. The objective function is to maximize the sum of probabilities. Various diseases and their identification probabilities are output, and iterations are increased.
Claims
1. Characterization, blood data, medical image multi-data approximation analysis and comprehensive identification methods of multiple suspected diseases, multi-feature data include: Blood data, organ characterization data, blood cell images, organ tissue medical image data and other data are used to determine the disease-feature correlation using an improved association method, which includes but is not limited to the Apriori method. The association rules of multiple diseases for multiple types of features are extracted. Based on the association rules of multiple diseases and disease features (blood, characterization, images), a disease-feature-feature value association model of multiple types of data is created. Using the improved association Apriori method, the steps are as follows: Step 1: Disease is represented by Du, u∈{0, 1...U}; Step2: The characteristics of the disease are represented by Fi, i∈{0, 1...I}; Step 3: The details of the disease are expressed by 1-2 digit codes; Step 4: The feature domain is represented by a 3- to 20-digit code; Step 5: The disease feature-feature value combination represents the accuracy of the feature item in identifying the disease.
2. Characterization, blood data, medical image multi-data similarity analysis and comprehensive identification methods for multiple suspected diseases, It is characterized in that It includes a similar disease-disease feature similarity method and a result analysis model. The steps of the disease feature similarity calculation method are as follows: According to the correlation between the disease and the disease characteristics, the degree of fit between the disease and the blood, representation, image and other characteristics is determined according to the extraction rules. According to the degree of fit, the method based on the similarity between the disease and the characteristics is calculated to distinguish and identify suspected diseases with high similarity. Assuming disease u and disease v, let N(u) represent the feature set of disease u, let N(v) be the feature set of disease v, then the formula for calculating disease similarity using cosine similarity is: After obtaining the similarity between disease u and disease v, the recognition accuracy of disease u for feature i is measured again, P(u, i): Where S(u, K) is the set of K disease features that are closest to disease u and whose matching degree with disease v using the Apriori method is greater than 0.5, N(i) is the set of diseases with feature i, Wuv is the similarity between disease u and disease v, rvi is implicit feedback information, representing the accuracy of disease v's recognition of feature i. To simplify the calculation, rvi=1. After calculating Pui for all disease features, Pui can be processed in descending order, and the first N disease features can be taken for disease u. Based on the number of features, it is calculated whether the disease and the disease have a high similarity, and suspected diseases with high similarity based on disease features can be found.
3. Characterization, blood data, medical image multi-data similarity analysis and comprehensive identification methods for multiple suspected diseases, Features It includes a multi-type data multi-disease feature fusion model, using an optimization method, the optimization method includes an improved tabu search (Tabu-Search) method, but is not limited to the improved tabu search method, for multiple diseases, respectively, in the representation set, blood set, image feature set, find features with high similarity and put them into the candidate set, exchange the same type of feature items in the candidate set, set a taboo table, put the exchange set into the taboo table, set the length of the taboo table, and read the dynamic feature items as the input items of the neural network and their probabilities into the input layer, adjust the parameters through the improved deep neural network method, intelligently identify the disease, and dynamically input it into the improved neural network disease recognition framework. The recognition method steps are as follows: Step 1: Initialize the disease model Du{F1, F2, F3, ..., Fi}; Step 2: Similarity calculation based on disease-feature-feature value; Step 3: For the disease Du, in the representation set {C1, C2...Cn}, blood set {B1, B2,...,Bn}, image feature set {P1, P2,...,Pn}, for the CX1, BY1, PZ1 features, respectively find the features CX2, BY2, PZ2 with high similarity to exchange with them, find Du-Fi and put it into the {CX1-CX2, BY1-BY2, PZ1-PZ2} Adjustment-Fi candidate solution exchange set; Step 4: Exchange {CX1-CX2, BY1-BY2, PZ1-PZ2} in the Adjustment-Fi candidate set and put them into the feature input set {F1, F2, F3, ..., Fi}; Step 5: Set up a taboo table, put the exchange set into the taboo table, and set the length of the taboo table; Step 6: Read the dynamic feature items as the input items of the neural network and their probabilities into the input layer; Step 7: Determine whether the input item is a hard input item or a soft input item; Step 8: Set the hard disease recognition probability as a hard indicator and the soft disease recognition probability as a soft indicator; Step 9: Use the improved deep neural network algorithm, training samples, adaptive learning, adjusting parameters, and identifying diseases. The activation function is: the sum of the weight of each hard input item * the hard disease recognition probability and the weight of each soft input item * the soft disease recognition probability; Step 10: Use deep neural network method to adaptively adjust weight parameters; Step 11: Output the results and determine the disease; Step 12: Output various diseases and their disease identification probability, and increase iterations; Step 13: Objective function: maximize the total probability.