A medical institution cooperation relationship identification method for medical image processing
By identifying collaborators for medical institutions and building models in groups, the problems of model performance and computational overhead under the heterogeneity of medical image data were solved, realizing an efficient medical image processing model suitable for medical image-assisted diagnosis.
Patent Information
- Application Number
- CN202211477561.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-23
- Publication Date
- 2026-02-06
- Estimated Expiration
- 2042-11-23
AI Technical Summary
In collaborative learning environments where medical image data is scattered and heterogeneous, how can we balance model performance and computational overhead to enable medical institutions to benefit from the data from others while avoiding the high computational cost of personalized models?
By detecting collaborators for each medical institution, an adaptive grouping medical image processing model is constructed. Institutions are divided using contribution vectors and the K-Means algorithm or semidefinite programming. A machine learning model is then built to label institutions that do not meet the criteria, thereby reducing computational overhead.
It enables various institutions to benefit from high-quality models with strong generalization capabilities, with computational overhead far less than personalized models, improving the efficiency of collaborative learning and making it suitable for assisted diagnosis in medical image processing.
Smart Images
Figure CN115714005B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to a medical institution cooperation relationship identification method for medical image processing, which detects the cooperation partners of each medical institution based on the data scattered in each medical institution, and can adaptively group the medical institutions with similar cooperation partners to construct a high-quality medical image processing model, so that each institution can benefit from the gain brought by the different data of other institutions, and is applied to the field of image processing technology. BACKGROUND
[0002] The medical image processing system based on machine learning technology can converge and analyze a large amount of medical image and digital pathology data in a short time, and the medical image processing system has great application value in assisting medical image classification.
[0003] However, medical image data is usually scattered in image storage systems of different institutions and different departments, forming a "data island" that lacks effective interconnection, and each medical institution has relatively less data. Although machine learning technology has shown good results in medical applications, it highly depends on the amount and diversity of training data, and due to the number of medical image data or the type of pathology, the training data required to train a high-performance medical image classification model may not be available in a single medical institution. As a new machine learning concept, collaborative learning can build a powerful machine learning model by utilizing scattered data, thereby realizing the value sharing of multi-source data.
[0004] The current collaborative learning paradigm mainly includes personalized collaborative learning and centralized collaborative learning. Among them, personalized collaborative learning allows different institutions to provide different models, while centralized collaborative learning requires the same model to be provided to all institutions. In many practical applications, data from different institutions is usually heterogeneous, so a single centralized collaborative learning model inevitably performs poorly on the data of some institutions. In the case of data heterogeneity, the most ideal way for each institution is to customize a personalized model for each institution, but when the number of institutions is large, this approach faces an unbearable computational overhead. Therefore, in the collaborative learning environment with data heterogeneity, how to balance model performance and computational overhead is a very key technology, and it is of great significance for the landing of collaborative learning in practical applications including the medical field. SUMMARY
[0005] The present application provides a medical institution cooperation relationship identification method for medical image processing to solve the problems and deficiencies in the prior art.
[0006] The present application is based on medical image data dispersed in each medical institution, and by detecting the cooperation partners of each medical institution, medical institutions with similar cooperation partners can be adaptively grouped to construct a medical image processing model, so as to fully tap the potential and deep value of multi-institution data, so that each institution can benefit from the gain brought by the medical image data of the other institution, and thus learn a high-quality medical image processing model with strong generalization ability. In addition, the present application marks the medical institutions that do not meet the preset conditions, so that the machine learning model constructed by grouping has a performance comparable to that of the personalized model with the optimal performance, and the computational overhead of the present application is much smaller than that required for training the personalized model, greatly improving the efficiency of collaborative learning in practical applications.
[0007] Technical scheme: A medical institution cooperation relationship identification method for medical image processing, comprising the following contents:
[0008] (I) Collecting training data required for medical image classification processing of each medical institution, taking the training data as the source data set of the corresponding medical institution, and determining the sample quantity and statistical information contained in the source data set of each medical institution;
[0009] (II) According to the source data set of each of the plurality of medical institutions, the sample quantity and statistical information contained in the source data set of each of the plurality of medical institutions, the contribution degree between each two of the plurality of medical institutions is calculated, and for each medical institution in the plurality of medical institutions, a contribution degree vector of the plurality of medical institutions is constructed according to the contribution degree; the contribution degree vector of each of the plurality of medical institutions is set as a multi-dimensional vector, which is used to represent the contribution degree of the plurality of medical institutions to the medical institution in the optimal cooperation state;
[0010] (III) According to the contribution degree vector of each of the plurality of medical institutions, the plurality of medical institutions are divided into several groups.
[0011] Further, in the (I), collecting training data required for medical image processing of each medical institution, taking the training data as the source data set of the corresponding medical institution, and determining the sample quantity and statistical information contained in the source data set of each medical institution, comprising the following specific steps:
[0012] 100, determining the target group and the number of medical institutions referred to in the specific cooperative learning task of the plurality of medical institutions;
[0013] 101. Collect the training data required by the multiple medical institutions for medical image processing, use the training data as the source dataset of the corresponding medical institutions, and determine the number of samples contained in the source dataset of each of the multiple medical institutions based on the source dataset of each of the multiple medical institutions.
[0014] 102. Determine the statistical information of the multiple medical institutions based on their respective source datasets and sample sizes.
[0015] Further, in step (ii), based on the source datasets of each of the multiple medical institutions, the number of samples contained in the source datasets of each of the multiple medical institutions, and statistical information, the contribution degree between every two medical institutions is calculated. For each of the multiple medical institutions, a contribution degree vector for each of the multiple medical institutions is constructed based on the contribution degree, including the following specific steps:
[0016] 200. Based on the statistical information corresponding to the source datasets of each of the multiple medical institutions, calculate the source dataset difference value between every two medical institutions.
[0017] 201. For each of the plurality of medical institutions, sort the plurality of medical institutions in ascending order according to the difference value of the source dataset between the medical institution and the source dataset.
[0018] 202. Based on preset weights, the number of samples and statistical information contained in the source datasets of each of the multiple medical institutions, and the difference value of the source datasets between each pair of medical institutions, determine the difference thresholds of each of the multiple medical institutions, and determine the collaborating partners of each of the multiple medical institutions; the collaborating partners of each of the multiple medical institutions are used to characterize the group of medical institutions that have a benefit to the medical institution among the multiple medical institutions.
[0019] 203. Based on the number of samples and statistical information contained in the source datasets of each of the multiple medical institutions, the difference value of the source datasets between each pair of medical institutions, the difference threshold of each of the multiple medical institutions, and the collaborating partners of each of the multiple medical institutions, calculate the contribution between each pair of medical institutions among the multiple medical institutions.
[0020] 204. For each of the plurality of medical institutions, construct a contribution vector for each of the plurality of medical institutions based on the contribution degree.
[0021] Furthermore, in step (iii), the multiple medical institutions are divided into several groups based on their respective contribution vectors, including the following specific steps:
[0022] 300、determining the number of groups needed to be divided in the specific collaborative learning task;
[0023] 301、running a K-Means algorithm on the contribution vector set of the plurality of medical institutions according to the number of groups, and determining the error sum of squares value of the contribution vector set;
[0024] 302、dividing the plurality of medical institutions into the number of groups according to the error sum of squares value.
[0025] Further, dividing the plurality of medical institutions into the number of groups according to the error sum of squares value comprises:
[0026] 302-11、when the error sum of squares value is less than a preset threshold, dividing the plurality of medical institutions into corresponding groups according to the division result of the contribution vector set by the K-Means algorithm;
[0027] 302-12、determining a key distance value according to the error sum of squares value and a preset parameter;
[0028] 302-13、according to the key distance value, if the contribution vector corresponding to a medical institution in the plurality of medical institutions is greater than the key distance value from the cluster center vector obtained by the K-Means algorithm, marking the medical institution.
[0029] Further, dividing the plurality of medical institutions into the number of groups according to the error sum of squares value comprises:
[0030] 302-21、when the error sum of squares value is greater than a preset threshold, constructing a collaboration similarity matrix according to the contribution vector set of the plurality of medical institutions; the collaboration similarity matrix is used to represent the collaboration partner similarity degree between each two medical institutions in the plurality of medical institutions;
[0031] 302-22、constructing a semi-definite programming problem according to the collaboration similarity matrix and the number of groups;
[0032] 302-23、solving the semi-definite programming problem by using an integer algorithm, and dividing the plurality of medical institutions into corresponding groups according to the solving result;
[0033] 302-24、determining a weak edge inside the group according to the collaboration similarity matrix, the division result obtained by solving the semi-definite programming problem, and a preset parameter; the weak edge is used to represent weak collaboration similarity between medical institutions;
[0034] 302-25, according to the weak edge, mark the specific medical institutions in the plurality of medical institutions that meet the preset conditions.
[0035] A medical image processing model construction method based on medical institution cooperation relationship, according to the several groups, a machine learning model is constructed for each group for identification and classification of medical images.
[0036] Further, according to the several groups divided, a machine learning model is constructed for each group, including the following specific steps:
[0037] 600, according to the grouping of the plurality of medical institutions and the marked medical institutions, selecting unmarked medical institutions in the plurality of medical institutions;
[0038] 601, according to the grouping of the plurality of medical institutions and the unmarked medical institutions, calculating the average contribution degree vector of the contribution degree vector corresponding to the unmarked medical institutions in each group;
[0039] 602, according to the average contribution degree vector of each group, constructing a weighted loss function;
[0040] 603, according to the weighted loss function, constructing a machine learning model for medical image identification and classification for each group.
[0041] A computer device, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, when the processor executes the above-mentioned computer program, the medical institution cooperation relationship identification method for medical image processing is realized.
[0042] A computer readable storage medium, the computer readable storage medium stores a computer program for executing the medical institution cooperation relationship identification method for medical image processing.
[0043] The beneficial aspects of the present application are:
[0044] The present application can adaptively group medical institutions with similar collaboration partners to construct a model, based on the data dispersed in each medical institution, by detecting collaboration partners of each medical institution, so as to fully tap the potential and deep value of multi-institution data, so that each institution can benefit from the gain brought by the different data of other institutions, thereby learning a high-quality medical image processing model with strong generalization ability; in addition, the present application marks medical institutions that do not meet the preset conditions, so that the machine learning model constructed by grouping has a performance comparable to that of the personalized model with optimal performance, and the computational overhead of the present application is much smaller than that required for training a personalized model, greatly improving the efficiency of collaborative learning in practical applications, thereby playing a timely and active auxiliary role in the diagnosis process of doctors. The present application is based on collaborative modeling of data dispersed in each medical institution, taking into account the model performance and computational overhead of collaborative learning in medical image processing, and has important clinical significance, and is of great significance for the landing of collaborative learning in practical applications including the medical field. BRIEF DESCRIPTION OF DRAWINGS
[0045] Figure 1 is a method flowchart of an embodiment of the present application;
[0046] Figure 2 is a flowchart of determining the sample size and statistical information according to the source data set of each of the plurality of medical institutions in an embodiment of the present application;
[0047] Figure 3 is a flowchart of calculating the contribution degree between each two of the plurality of medical institutions and constructing the contribution degree vector of each of the plurality of medical institutions in an embodiment of the present application;
[0048] Figure 4 is a flowchart of dividing the plurality of medical institutions into a plurality of groups according to the contribution degree vector of each of the plurality of medical institutions in an embodiment of the present application;
[0049] Figure 5 is a flowchart of dividing the plurality of medical institutions into a plurality of groups and marking a specific medical institution when the sum of squared errors is less than a preset threshold in an embodiment of the present application;
[0050] Figure 6 is a flowchart of dividing the plurality of medical institutions into a plurality of groups and marking a specific medical institution when the sum of squared errors is greater than a preset threshold in an embodiment of the present application;
[0051] Figure 7 is a flowchart of constructing a machine learning model for each group according to the group division of the plurality of medical institutions in an embodiment of the present application. DETAILED DESCRIPTION
[0052] The application will be further illustrated by the following specific examples, which are only used to illustrate the application and not used to limit the scope of the application. After reading the application, those skilled in the art can make various modifications to the application, and the modifications fall within the scope defined by the appended claims.
[0053] A medical institution cooperation relationship identification method and a medical image processing model construction method for medical image processing, comprising the following steps:
[0054] (I) Collecting medical image data required by each medical institution for medical image processing, and taking the medical image data as training data, taking the training data as the source data set of the corresponding medical institution, determining the sample quantity contained in the source data set of each medical institution and the statistical information;
[0055] (II) According to the source data set of each medical institution, the sample quantity contained in the source data set and the statistical information, calculating the contribution degree between each two medical institutions in the plurality of medical institutions, and for each medical institution in the plurality of medical institutions, constructing a contribution degree vector of the plurality of medical institutions according to the contribution degree; the contribution degree vector of each medical institution is set as a multi-dimensional vector, which is used to represent the contribution degree of the plurality of medical institutions to the medical institution in the optimal cooperation state;
[0056] (III) According to the contribution degree vector of each medical institution, the plurality of medical institutions are divided into several groups;
[0057] (IV) According to the several groups, a machine learning model is constructed for each group for identification and classification of medical images.
[0058] In combination with Figure 1 , collecting training data required by each medical institution for medical image processing, taking the training data as the source data set of the corresponding medical institution, determining the sample quantity contained in the source data set of each medical institution and the statistical information, comprising the following specific steps:
[0059] 100, determining the target group {C1,…,C N} and the number N of medical institutions referred to by the plurality of medical institutions in a specific cooperative learning task.
[0060] 101, collecting training data required by the plurality of medical institutions for medical image processing, specifically, the training data i of the medical institution C is m i samples sampled from the corresponding local distribution image data set , taking the training data as the source data set {S1,…,S N}, based on the source datasets {S1,…,S} from multiple medical institutions. N}, determine the number of samples {m1,…,m} contained in the source datasets of multiple medical institutions. N}
[0061] 102. Based on the source datasets and sample sizes of multiple medical institutions, determine the statistical information for each institution. For medical institution C... i Its statistical information corresponds Where h i In the source dataset S i The discriminative machine learning model trained on it, It is model h i For the sample Output marked as The predicted probability.
[0062] Combination Figure 2 Based on the source datasets of multiple medical institutions, the number of samples in each source dataset, and statistical information, the contribution between every two medical institutions is calculated. For each medical institution, a contribution vector is constructed based on the contribution. The specific steps include:
[0063] 200. Based on the statistical information corresponding to the source datasets of multiple medical institutions, calculate the difference value of the source datasets between any two medical institutions. For any two medical institutions C i and C j In this context, the difference between the source datasets of the two medical institutions is denoted as... Where d Js (S i ||S j ) is the source dataset S i and S j The Jensen–Shannon (JS) divergence between them is calculated as follows:
[0064] 201. For each of multiple medical institutions, sort the medical institutions in ascending order according to the difference values between them and the source datasets of that medical institution. For medical institution C... i The medical institution group {C1,…,C N According to C i The source dataset's difference values are sorted in ascending order, and the order of the medical institutions after ascending sort is denoted as {σ}. i (1),…,σ i (N)}, then we have
[0065] 202. Based on the preset weights λ and the number of samples {m1,…,m} contained in the source datasets of multiple medical institutions. N} and statistical information, as well as the difference values of the source datasets between each pair of medical institutions, determine the difference thresholds for each of the multiple medical institutions, for medical institution C i In this regard, the corresponding difference threshold is denoted as ζ. i , here ζ i It is an equation The larger of the two roots, where q i Representing the sorted medical institutions, medical institution C i The maximum healthcare institution index value with gain. It is an organization The number of samples contained in the source dataset, specifically,
[0066] Where t is an enumeration of the institutional index in the arg max problem.
[0067] And to identify the respective collaborative partners of the multiple medical institutions, for medical institution C i In other words, its collaborating partners are those in the sorted order. The training data of this group of medical institutions is used for medical institution C. i Building its machine learning model can have a beneficial effect.
[0068] 203. Based on the number of samples {m1,…,m} contained in the source datasets of multiple medical institutions. N} and statistical information, the source dataset difference values between each pair of medical institutions, the difference thresholds for each of the multiple medical institutions, and the collaborating partners of each of the multiple medical institutions, to calculate the contribution between each pair of medical institutions. Specifically, medical institution C j For medical institution C i Contribution The calculation method is as follows: Among them, the operator [·] + =max(·,0), contribution level Used to characterize the situation in medical institution C i In terms of the optimal collaborative state, medical institution C j Should be C i The degree of contribution made is that of medical institution C. i In collaborative scenarios where personalized models are customized, healthcare institution C j Should be C i The degree of contribution made.
[0069] 204、For each medical institution in the plurality of medical institutions, a contribution degree vector of the plurality of medical institutions is constructed according to the contribution degree, in particular, for any medical institution C i , the corresponding contribution degree vector is denoted as
[0070] In combination with Figure 3 , the plurality of medical institutions is divided into several groups according to the contribution degree vectors of the plurality of medical institutions, including the following specific steps:
[0071] 300、Determine the number K of groups to be divided in the specific collaborative learning task, usually 1≤K<<N.
[0072] 301、According to the number K of groups, run the K-Means algorithm on the set of contribution degree vectors of the plurality of medical institutions , and determine the sum of squared errors of the set of contribution degree vectors, in particular, denote as the cluster center vector of the kth cluster P k obtained by the K-Means algorithm, and the sum of squared errors corresponding to the cluster division {P1,…,P K} obtained by the K-Means algorithm is denoted as The calculation method of is where is the distance function between vectors .
[0073] 302、According to the sum of squared errors , the plurality of medical institutions is divided into K groups.
[0074] In combination with Figure 4 , according to the sum of squared errors, the plurality of medical institutions is divided into K groups, including:
[0075] 400、When the sum of squared errors is less than a preset threshold τ, according to the division result of the set of contribution degree vectors using the K-Means algorithm, the plurality of medical institutions {C1,…,C N} is divided into the corresponding groups {P1,…,P K}.
[0076] 401、According to the sum of squared errors and a preset parameter β, determine a key distance value d ＊ , the specific calculation method is
[0077] 402、According to the key distance value d ＊, if the contribution degree vector corresponding to a medical institution in the plurality of medical institutions is greater than the key distance value from the cluster center vector obtained by the K-Means algorithm, the medical institution is marked, in particular, for any medical institution C k in any cluster P i , if , the medical institution C i is marked.
[0078] In combination Figure 5 , according to the error sum of squares, the plurality of medical institutions is divided into the group number of groups, including:
[0079] 500, when the error sum of squares is greater than a preset threshold τ, a cooperation similarity matrix U is constructed according to the contribution degree vector set of the plurality of medical institutions, in particular, first, a directed graph is constructed according to the plurality of medical institutions {C1,…,C N} and the corresponding contribution degree vector set . The N nodes {1,…,N} in the directed graph correspond to the medical institutions {C1,…,C N}, and the weight of the directed edge from node j to node i is Let matrix A be the adjacency matrix of the directed graph , and matrix D in be a diagonal matrix composed of the in-degree of all nodes in the directed graph , and the cooperation similarity matrix U is calculated as follows: The cooperation similarity matrix U is used to represent the cooperation partner similarity degree between each two medical institutions in the plurality of medical institutions.
[0080] 501, according to the cooperation similarity matrix U and the group number K, a semi-definite programming problem is constructed, in particular, let w ij be the (i,j)th element in the cooperation similarity matrix U, d i =∑ j w ij be the degree of node i, be the sum of the weights of all edges, let be a K-dimensional vector with the jth dimension equal to 1 and other dimensions equal to 0, and let represent the affiliation group index of medical institution C i , be set to a K-dimensional vector, if , it means that the medical institution C i belongs to the group G j , and let the symbol be Then the semi-definite programming problem is expressed as follows:
[0081]
[0082]
[0083]
[0084]
[0085] 502. Solve the positive semidefinite programming problem using the rounding algorithm. Based on the solution, divide multiple medical institutions into corresponding groups. Specifically, randomly select t planes in a K-dimensional space and use projection to transform the vector set in the positive semidefinite programming problem. Divided into 2 t A cluster, denoted as H0 for the objective function value in the semidefinite programming problem produced by this approach. t The set of vectors corresponding to the maximum values in H2 and H3. The output is the set of vectors corresponding to best(H2,H3). Medical institutions {C1,…,C N} Divide into different groups {G1,…,G K}
[0086] 503. Based on the cooperation similarity matrix U, the partitioning result obtained from solving the semidefinite programming problem, and the preset parameter η, determine the group {G1,…,G...} K The weak edges inside}, specifically, if w ij <η, then w ij The corresponding edge is called a weak edge, i.e., medical institution C. ki and medical institution C j The collaboration similarity between them is less than the preset parameter η, and weak edges are used to characterize the weak collaboration similarity between medical institutions.
[0087] 504. Based on weak edges, mark specific medical institutions among multiple medical institutions that meet preset conditions. Specifically, for any group G k Remove G k After all the weak edges mentioned inside, G k If the internal nodes are divided into several disjoint sets, the largest set is retained, and the medical institutions corresponding to the nodes that do not belong to the largest set are marked.
[0088] Combination Figure 6 Based on the aforementioned groups, a machine learning model is constructed for each group, including the following specific steps:
[0089] 600. Based on the grouping of multiple medical institutions {P1,…,P…}K} or {G1,…,G K} and the labeled medical institutions, denoted as Selecting the unlabeled medical institutions from the plurality of medical institutions, denoted as
[0090] 601. According to the grouping of the plurality of medical institutions and the unlabeled medical institutions Calculating the average contribution degree vector of the contribution degree vector corresponding to the unlabeled medical institutions in each group, for any group G k , the average contribution degree vector corresponding to it is The calculation method is as follows:
[0091] 602. According to the average contribution degree vector of each group, constructing a weighted loss function for any group G k , assuming that h corresponds to the weighted loss function of the group G k wherein is the empirical loss of the hypothesis h on the source data set S i corresponding to the medical institution C i .
[0092] 603. According to the weighted loss function, constructing a machine learning model for medical image recognition classification for each group, taking a neural network as an example, for any group G k , first initializing the neural network, according to its corresponding weighted loss function optimizing all parameters in the neural network through a gradient optimization method until the neural network converges, obtaining a machine learning model for medical image recognition classification.
[0093] In summary, based on the data dispersed in each medical institution, the medical image processing model based on the cooperation relationship of the medical institution can adaptively group the medical institutions with similar cooperation partners to construct a model, fully tap the potential and deep value of multi-institution data, so that each institution can benefit from the gain brought by the different data of other institutions, and learn a high-quality medical image processing model with strong generalization ability. In addition, the medical institution that does not meet the preset condition is marked, so that the machine learning model constructed by grouping has a performance comparable to that of the personalized model with optimal performance. The computing overhead of the present application is much smaller than that required for training a personalized model, greatly improving the efficiency of collaborative learning in practical applications, thereby playing a timely and active auxiliary role in the diagnosis process of doctors. The present application is based on the data dispersed in each medical institution to collaboratively model, taking into account the model performance and computing overhead of collaborative learning in medical image processing, which has great significance for the landing of collaborative learning in practical applications including the medical field.
[0094] Obviously, those skilled in the art should understand that each step of the medical image processing model construction method based on the cooperation relationship of the medical institution of the above-mentioned embodiments of the present application or the medical image processing model construction method based on the cooperation relationship of the medical institution can be realized by a general computing device, which can be concentrated on a single computing device or distributed on a network composed of multiple computing devices. Alternatively, they can be realized by program codes executable by a computing device, so that they can be stored in a storage device and executed by a computing device, and in some cases, the steps shown or described can be executed in a different order, or they can be made into individual integrated circuit modules, or multiple modules or steps thereof can be made into a single integrated circuit module. Thus, the embodiments of the present application are not limited to any specific combination of hardware and software.
[0095] Although the embodiments of the present application have been shown and described, various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.
Claims
1. A medical institution collaboration relationship identification method for medical image processing, characterized by, The method comprises the following steps: (I) collecting training data required by a plurality of medical institutions for medical image classification processing, taking the training data as a source data set of a corresponding medical institution, determining a sample quantity and statistical information contained in the source data set of each medical institution; (II) calculating a contribution degree between each two medical institutions in the plurality of medical institutions according to the source data set of each medical institution, the sample quantity and statistical information contained in the source data set of each medical institution, and constructing a contribution degree vector of each medical institution in the plurality of medical institutions according to the contribution degree; the contribution degree vector of each medical institution is set as a multi-dimensional vector, which is used to represent a contribution degree of the plurality of medical institutions to the medical institution in an optimal cooperation state; (III) dividing the plurality of medical institutions into a plurality of groups according to the contribution degree vector of each medical institution; In the step (II), the contribution degree between each two medical institutions in the plurality of medical institutions is calculated according to the source data set of each medical institution, the sample quantity and statistical information contained in the source data set of each medical institution, and the contribution degree vector of each medical institution in the plurality of medical institutions is constructed according to the contribution degree, comprising the following specific steps:
200. calculating a source data set difference value between each two medical institutions in the plurality of medical institutions according to the statistical information corresponding to the source data set of each medical institution; 201. arranging the plurality of medical institutions in ascending order according to the source data set difference value between the medical institution and each medical institution in the plurality of medical institutions; 202. determining a difference threshold of each medical institution in the plurality of medical institutions and a cooperation partner of each medical institution according to a preset weight, the sample quantity and statistical information contained in the source data set of each medical institution, and the source data set difference value between each two medical institutions; the cooperation partner of each medical institution is used to represent a medical institution group having gain to the medical institution in the plurality of medical institutions; 203. calculating a contribution degree between each two medical institutions in the plurality of medical institutions according to the sample quantity and statistical information contained in the source data set of each medical institution, the source data set difference value between each two medical institutions, the difference threshold of each medical institution in the plurality of medical institutions, and the cooperation partner of each medical institution; 204. constructing a contribution degree vector of each medical institution in the plurality of medical institutions according to the contribution degree.
2. The medical institution collaboration relationship identification method for medical image processing according to claim 1, characterized by, In the step (I), the training data required by each medical institution for medical image processing is collected, the training data is taken as a source data set of a corresponding medical institution, and a sample quantity and statistical information contained in the source data set of each medical institution are determined, comprising the following specific steps:
100. determining the target groups and the number of medical institutions in specific collaborative learning tasks; 101. collecting training data required by the medical institutions for medical image processing, taking the training data as the source data set of the corresponding medical institution, and determining the number of samples contained in the source data set of each medical institution according to the source data set of each medical institution; 102. determining the statistical information of the medical institutions according to the source data set and the number of samples of each medical institution.
3. The medical institution collaboration relationship identification method for medical image processing according to claim 1, characterized by, In the third aspect, the plurality of medical institutions are divided into several groups according to the contribution degree vectors of the plurality of medical institutions, including the following specific steps:
300. determining the number of groups to be divided in specific collaborative learning tasks; 301. running the K-Means algorithm on the contribution degree vector set of the plurality of medical institutions according to the number of groups, and determining the error sum of squares value of the contribution degree vector set; 302. dividing the plurality of medical institutions into the number of groups according to the error sum of squares value.
4. The medical institution collaboration relationship identification method for medical image processing according to claim 3, characterized by, According to the error sum of squares value, the plurality of medical institutions are divided into the number of groups, including: 302-11. when the error sum of squares value is less than a preset threshold, the plurality of medical institutions are divided into corresponding groups according to the division result of the contribution degree vector set by the K-Means algorithm; 302-12. determining a key distance value according to the error sum of squares value and a preset parameter; 302-13. according to the key distance value, if the contribution degree vector corresponding to a medical institution in the plurality of medical institutions is greater than the key distance value, the medical institution is marked.
5. The medical institution collaboration relationship identification method for medical image processing according to claim 3, characterized by, According to the error sum of squares value, the plurality of medical institutions are divided into the number of groups, including: 302-21. when the error sum of squares value is greater than a preset threshold, a collaborative similarity matrix is constructed according to the contribution degree vector set of the plurality of medical institutions; the collaborative similarity matrix is used to represent the similarity degree of collaborative partners between each two medical institutions in the plurality of medical institutions; 302-22. constructing a semi-definite programming problem according to the collaborative similarity matrix and the number of groups; 302-23. solving the semi-definite programming problem by using an integer algorithm, and dividing the plurality of medical institutions into corresponding groups according to the solving result; 302-24. determining the weak edges within the groups according to the collaborative similarity matrix, the division result obtained by solving the semi-definite programming problem, and a preset parameter; the weak edges are used to represent the weak collaborative similarity between medical institutions; 302-25. marking specific medical institutions in the plurality of medical institutions that meet a preset condition according to the weak edges.
6. A method of constructing a medical image processing model based on the medical institution collaboration relationship according to any one of claims 1 to 5, characterized by, According to the groups of medical institutions, a machine learning model is constructed for each group for the identification and classification of medical images. 7.The method of claim 6, wherein, According to the division of the plurality of groups, a machine learning model is constructed for each group, including the following specific steps:
600. According to the grouping of the plurality of medical institutions and the labeled medical institutions, selecting unlabeled medical institutions in the plurality of medical institutions; 601. According to the grouping of the plurality of medical institutions and the unlabeled medical institutions, calculating the average contribution degree vector of the contribution degree vector corresponding to the unlabeled medical institutions in each group; 602. According to the average contribution degree vector of each group, a weighted loss function is constructed; 603. According to the weighted loss function, a machine learning model for medical image recognition classification is constructed for each group.
8. A computer device, comprising: The computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the above computer program to realize the medical institution cooperation relationship identification method for medical image processing according to any one of claims 1-5.
9. A computer-readable storage medium, characterized in that: The computer readable storage medium stores a computer program for executing the medical institution cooperation relationship identification method for medical image processing according to any one of claims 1-5.