AI-based severe pneumonia diagnosis method
Through the federated learning architecture of servers and clients, conditions are used to generate adversarial networks and transfer learning, synthetic data is generated and feature calibration is performed, which solves the problem of data heterogeneity in federal medical scenarios and improves the accuracy of pneumonia-assisted diagnosis.
Patent Information
- Application Number
- CN202510469619.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-15
- Publication Date
- 2025-08-29
AI Technical Summary
In the federal medical scenario, data silos in various medical institutions lead to data heterogeneity in the diagnosis of pneumonia, affecting model performance and diagnostic accuracy.
The federated learning architecture of servers and clients is adopted to generate synthetic data through conditional generation adversarial networks, and the model parameter weighted average and feature calibration is performed. Transfer learning is used to calibrate private data feature information to improve model performance.
It improves the accuracy of pneumonia-assisted diagnosis results in federal medical scenarios, solves the problem of data heterogeneity, and improves the overall performance of the model.
Smart Images
Figure CN120565022A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of auxiliary medical judgment, and specifically relates to an AI-based diagnosis method for severe pneumonia. Background Art
[0002] With the rapid maturity of artificial intelligence technology and the deepening of medical informatization and digitization, the use of machine learning to diagnose pneumonia has become possible. Automated diagnosis of severe pneumonia through AI can better assist doctors in the diagnostic process. However, machine learning requires data as a learning foundation, and medical institutions have limited data. Large numbers of clients collaborate to train shared models without sharing their private data, leading to data silos in the diagnostic process. This creates a degree of subjectivity in diagnosing within a single client. Therefore, addressing the heterogeneity of pneumonia data in federated healthcare scenarios, thereby improving model performance and, consequently, the accuracy of pneumonia-assisted diagnostic results, has become a pressing issue for those skilled in the art. Summary of the Invention
[0003] The purpose of the present invention is to propose an AI-based method for diagnosing severe pneumonia to solve one or more technical problems existing in the prior art and at least provide a beneficial option or create conditions.
[0004] An AI-based method for diagnosing severe pneumonia, the method is based on a hospital-assisted diagnosis system, the system includes a server and multiple clients; the client is the hospital data management department;
[0005] The server adopts a global model, and the client adopts a local model; the server and each client adopt federated learning to perform model training;
[0006] The server is used to broadcast the initialized global parameters and the synthetic data generated by the conditional generative adversarial network to each of the clients; the synthetic data is a pneumonia dataset; the pneumonia dataset includes different lung image samples and a label corresponding to each of the lung image samples; the label is positive or negative;
[0007] Furthermore, the server is also used to perform a weighted averaging operation on the model parameters after receiving them uploaded by each of the clients to obtain aggregated global parameters, and broadcast the aggregated global parameters to each of the clients until the performance of the global model reaches the expected target; the aggregated global parameters are used by each of the clients to continue local model classification training locally until the performance of the global model reaches the expected target; the global model whose performance reaches the expected target is used to perform pneumonia auxiliary diagnosis on the input lung image to obtain a pneumonia auxiliary diagnosis result.
[0008] Furthermore, the specific steps of performing weighted averaging on the model parameters include:
[0009] Obtain the image sample weight value, standardize the model parameters, agree on the model parameter format, the model parameters include the pneumonia data set and the weighting parameter, the weighting parameter is to set the weight value of the image sample in the training process, the pneumonia data set is defined as k, the data set includes k including: image sample set S d and the corresponding weighted parameter set Q d , S d =(Sa d , f d ), the Sa d is the image sample, f d is the label corresponding to the image sample, the subscript d is the sequence number of the image sample set, and the model parameters are given weights. According to the given weight parameter set Q d For the image sample set S d Assign weights, where W d is the corresponding image sample set S d Weight value, mean() is the mean function, (Sa d , f d ) is performed by using the corresponding label f in the image sample d Participate in the operation, note f d If it is Yang, then f d =1, note f d If it is Yin, then f d =0;
[0010] Perform weighted averaging on the model parameters, integrate the image sample weights into a data stream, and construct a sequence data, data = [data1, data2..., data n ], n is the total number of image samples, weighted mean L,
[0011]
[0012] data i is the value of the i-th element in the sequence data, max(data i ) is the maximum value in the sequence data, min(data i ) is the minimum value in the sequence data, exp() is the exponential function, and the weighted mean L is set as the lowest weight during model training. The weighted mean L and the model parameters are aggregated into global data, and the global data is trained through a convolutional neural network, and the global data is output as a global model.
[0013] Furthermore, the network loss function is used to solve the impact of the heterogeneity of pneumonia data from multiple hospitals after the server and each client use federated learning for model training. A conditional distribution mismatch penalty is introduced in the objective function of client training to achieve the effect of feature calibration. By introducing transfer learning, the feature information of the synthetic data is calibrated with private data to improve accuracy, thereby solving the heterogeneity of pneumonia data in the federated medical scenario, thereby improving model performance and further improving the accuracy of pneumonia auxiliary diagnosis results.
[0014] Furthermore, the aggregated global parameters are subjected to local model classification training, the global model is output, and it is determined whether the predicted value of the global model meets the standard, and the model parameters are defined as C d , create a blank array Rs, denote Rsij as the element in the i-th row and j-th column of the array Rs, and denote the value of Rsij as the RSSI value between the i-th node and the j-th node in the model, where i, j = 1, 2, ..., N, and when i = j, Rsij = 0; create an array Ra, denote Rak as the k-th element in the array Ra, and denote the value of Rak as the mean of the elements in each column of Rs, k = 1, 2, ..., N, and obtain the signal coordination array C by the following formula:
[0015]
[0016] Among them, Rsk1 is the k1th column element in Rs, k1=1,2,…,N, and the data in the array Rs and the array Ra are all model parameters C d ;
[0017] make Where Cpq is the element in the p-th row and q-th column of the array Rs, and Cpp is the element in the p-th row and p-th column of the array Ra. The value range of p and q is (1, N). Create the model training array Ass:
[0018]
[0019] Ass array is a pq The permutation array of Ass is recorded as Ar and Ac, and the row and column where the element with the smallest value in Ass is located are stored. The elements of other columns in Ass that intersect with the Arth row and the Acth row in turn are stored. Where i1 is an integer and i1∈[1,N-2]. The other columns in Ass are the columns of Ass excluding the Arth column and the Acth column. Traverse i1 within the value range of i1, update Ass to the array obtained by removing the Arth row, the Acth row, the Arth column, and the Acth column from Ass, and create an array Aave, recorded as Aave j1 is the j1th element in the array Aave, and Aave is recorded in sequence j1The value of is the mean of each column element in Ass, j1=1,2,…,N-2, and the array Aave is normalized to obtain the array Aave n , record the node security value Aave m For Aave T , the Aave T For array Aave n The transpose of the model parameter C in the model d The corresponding indicator is greater than Aave m is marked as an important weight, and the model parameter C d The corresponding indicator is greater than Aave m The mark is a secondary weight. When the number of important weights is less than the number of secondary weights, mark the current model parameter C d The weight values of the image samples in the training are relatively low. When the number of important weights is greater than the number of secondary weights, the current model prediction value is marked as high, and the prediction model is output. When the current model prediction value is high, the data is encrypted.
[0020] The beneficial effects of the present invention are: calibrating the feature information of synthetic data with private data to improve accuracy, thereby solving the heterogeneity of pneumonia data in federal medical scenarios, thereby improving model performance and further improving the accuracy of pneumonia auxiliary diagnosis results. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] The above and other features of the present disclosure will become more apparent through a detailed description of the embodiments shown in conjunction with the accompanying drawings. The same reference numerals in the drawings of the present disclosure represent the same or similar elements. Obviously, the drawings described below are only some embodiments of the present disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.
[0022] In the picture:
[0023] Figure 1 Shown is a structural diagram of an AI-based severe pneumonia diagnosis system;
[0024] Figure 2 Shown is a flowchart of an AI-based method for diagnosing severe pneumonia. DETAILED DESCRIPTION
[0025] The following will be combined with the embodiments and drawings to clearly and completely describe the concept, specific structure and technical effects of the present disclosure to fully understand the purpose, scheme and effect of the present disclosure. It should be noted that the embodiments and features in the embodiments of this application can be combined with each other unless there is a conflict.
[0026] Pain points of existing technologies: Traditional diagnosis relies on physician experience, resulting in a 38% misdiagnosis rate for early-stage severe pneumonia; CT image analysis is time-consuming (average 15-20 minutes per case); and there is a lack of dynamic correlation analysis between clinical indicators (blood oxygen, inflammatory factors, etc.) and imaging features.
[0027] like Figure 1 As shown, an AI-based method for diagnosing severe pneumonia is based on a hospital-assisted diagnosis system, which includes a server and multiple clients; the client is the hospital data management department;
[0028] The server adopts a global model, and the client adopts a local model; the server and each client adopt federated learning to perform model training;
[0029] The server is used to broadcast the initialized global parameters and the synthetic data generated by the conditional generative adversarial network to each of the clients; the synthetic data is a pneumonia dataset; the pneumonia dataset includes different lung image samples and a label corresponding to each of the lung image samples; the label is positive or negative;
[0030] Furthermore, if Figure 2 As shown, the server is further used to perform a weighted average operation on the model parameters after receiving the model parameters uploaded by each of the clients to obtain aggregated global parameters, and broadcast the aggregated global parameters to each of the clients until the performance of the global model reaches the expected target; the aggregated global parameters are used by each of the clients to continue local model classification training locally until the performance of the global model reaches the expected target; the global model whose performance reaches the expected target is used to perform pneumonia auxiliary diagnosis on the input lung image to obtain pneumonia auxiliary diagnosis results.
[0031] Furthermore, the specific steps of performing weighted averaging on the model parameters include:
[0032] Obtain the image sample weight value, standardize the model parameters, agree on the model parameter format, the model parameters include the pneumonia data set and the weighting parameter, the weighting parameter is to set the weight value of the image sample in the training process, the pneumonia data set is defined as k, the data set includes k including: image sample set S d and the corresponding weighted parameter set Q d , S d =(Sa d , f d ), the Sa d is the image sample, f d is the label corresponding to the image sample, the subscript d is the sequence number of the image sample set, and the model parameters are given weights. According to the given weight parameter set Q d For the image sample set S d Assign weights, where W d is the corresponding image sample set S d Weight value, mean() is the mean function, (Sa d , f d ) is performed by using the corresponding label f in the image sample d Participate in the operation, note f d If it is Yang, then f d =1, note f d If it is Yin, then f d =0;
[0033] Perform weighted averaging on the model parameters, integrate the image sample weights into a data stream, and construct a sequence data, data = [data1, data2..., data n ], n is the total number of image samples, weighted mean L,
[0034]
[0035] data i is the value of the i-th element in the sequence data, max(data i ) is the maximum value in the sequence data, min(data i ) is the minimum value in the sequence data, exp() is the exponential function, and the weighted mean L is set as the lowest weight during model training. The weighted mean L and the model parameters are aggregated into global data, and the global data is trained through a convolutional neural network, and the global data is output as a global model.
[0036] Furthermore, the network loss function is used to solve the impact of the heterogeneity of pneumonia data from multiple hospitals after the server and each client use federated learning for model training. A conditional distribution mismatch penalty is introduced in the objective function of client training to achieve the effect of feature calibration. By introducing transfer learning, the feature information of the synthetic data is calibrated with private data to improve accuracy, thereby solving the heterogeneity of pneumonia data in the federated medical scenario, thereby improving model performance and further improving the accuracy of pneumonia auxiliary diagnosis results.
[0037] Furthermore, the aggregated global parameters are subjected to local model classification training, the global model is output, and it is determined whether the predicted value of the global model meets the standard, and the model parameters are defined as C d, create a blank array Rs, denote Rsij as the element in the i-th row and j-th column of the array Rs, and denote the value of Rsij as the RSSI value between the i-th node and the j-th node in the model, where i, j = 1, 2, ..., N, and when i = j, Rsij = 0; create an array Ra, denote Rak as the k-th element in the array Ra, and denote the value of Rak as the mean of the elements in each column of Rs, k = 1, 2, ..., N, and obtain the signal coordination array C by the following formula:
[0038]
[0039] Among them, Rsk1 is the k1th column element in Rs, k1=1,2,…,N, and the data in the array Rs and the array Ra are all model parameters C d ;
[0040] make Where Cpq is the element in the p-th row and q-th column of the array Rs, and Cpp is the element in the p-th row and p-th column of the array Ra. The value range of p and q is (1, N). Create the model training array Ass:
[0041]
[0042] Ass array is a pq The permutation array of Ass is recorded as Ar and Ac, and the row and column where the element with the smallest value in Ass is located are stored. The elements of other columns in Ass that intersect with the Arth row and the Acth row in turn are stored. Where i1 is an integer and i1∈[1,N-2]. The other columns in Ass are the columns of Ass excluding the Arth column and the Acth column. Traverse i1 within the value range of i1, update Ass to the array obtained by removing the Arth row, the Acth row, the Arth column, and the Acth column from Ass, and create an array Aave, recorded as Aave j1 is the j1th element in the array Aave, and Aave is recorded in sequence j1 The value of is the mean of each column element in Ass, j1=1,2,…,N-2, and the array Aave is normalized to obtain the array Aave n , record the node security value Aave m For Aave T , the Aave T For array Aave n The transpose of the model parameter C in the model d The corresponding indicator is greater than Aave m is marked as an important weight, and the model parameter C d The corresponding indicator is greater than Aave m The mark is a secondary weight. When the number of important weights is less than the number of secondary weights, mark the current model parameter C dThe weight values of the image samples in the training are relatively low. When the number of important weights is greater than the number of secondary weights, the current model prediction value is marked as high, and the prediction model is output. When the current model prediction value is high, the data is encrypted.
[0043] 1. Experimental Methods
[0044] Common infection biomarkers in ICU patients: procalcitonin (PCT) and high-sensitivity C-reactive protein (hsCRP). Peripheral blood samples were obtained by peripheral venipuncture. Testing was performed using a Hitachi (Japan) Model 600 fully automated biochemical analyzer.
[0045] RT-qPCR was performed to detect the expression of exosomal miR-193a-5p in the bronchoalveolar lavage fluid of patients using the detection kit of Example 2. The diagnostic effects of the three biomarkers were compared using a receiver operating characteristic (ROC) curve.
[0046] 2. Experimental Results
[0047] Results showed that exosomal miR-193a-5p in bronchoalveolar lavage fluid (BALF) had a high diagnostic efficacy, with an area under the curve (AUC) of 0.875, higher than that of commonly used clinical markers such as procalcitonin (PCT) and high-sensitivity C-reactive protein (hsCRP). The AUC for multi-marker combined diagnosis was 0.933, greater than that of other single marker tests.
[0048] The specific operations of the three combined are as follows:
[0049] Medcalc software was used to perform receiver operating characteristic (ROC) curve analysis of multiple indicators combined diagnosis.
[0050] Step 1: Define variables and create a table: for example, the first column (testa) is the value of procalcitonin (PCT); the second column (testb) is the value of high-sensitivity C-reactive protein (hsCRP); the third column (testc) is the value of exosomal miR-193a-5p, which is a numerical variable; disease is the disease status, 0 means no disease, and 1 means severe pneumonia.
[0051] Step 2: Import the data into Medcalc
[0052] Step 3: Logistic regression comprehensively reflects the diagnostic capabilities of testa, testb, and testc. The receiver operating characteristic (ROC) curve drawn by LOGREGR_Pred1 is the joint diagnostic result of the three.
[0053] Although the description of the present disclosure has been quite detailed and specifically describes several embodiments, it is not intended to be limited to any of these details or embodiments or any particular embodiment, so as to effectively cover the intended scope of the present disclosure. In addition, the above description of the present disclosure is based on the embodiments foreseen by the inventors, which is intended to provide a useful description, and those non-substantial changes to the present disclosure that have not yet been foreseen may still represent equivalent changes to the present disclosure.
Claims
1. A method for diagnosing severe pneumonia based on AI, characterized in that: The method is based on a hospital auxiliary diagnosis system, which includes a server and multiple clients; the client is a hospital data management department; The server adopts a global model, and the client adopts a local model; the server and each client adopt federated learning to perform model training; The server is used to broadcast the initialized global parameters and the synthetic data generated by the conditional generative adversarial network to each of the clients; the synthetic data is a pneumonia dataset; the pneumonia dataset includes different lung image samples and a label corresponding to each of the lung image samples; the label is positive or negative.
2. The AI-based severe pneumonia diagnosis method according to claim 1, characterized in that: The server is further configured to perform a weighted average operation on the model parameters after receiving the model parameters uploaded by each of the clients to obtain aggregated global parameters, and broadcast the aggregated global parameters to each of the clients until the performance of the global model reaches the expected target; the aggregated global parameters are used by each of the clients to continue local model classification training locally until the performance of the global model reaches the expected target; the global model whose performance reaches the expected target is used to perform pneumonia-assisted diagnosis on the input lung image to obtain a pneumonia-assisted diagnosis result.
3. The AI-based severe pneumonia diagnosis method according to claim 2, characterized in that: The specific steps of performing weighted averaging on the model parameters include: Obtain the weight value of the image sample, standardize the model parameters, and unify the model parameter format. The model parameters include the pneumonia data set and the weighting parameter. The weighting parameter is the weight value of the image sample in the training process. The pneumonia data set is defined as k. The data set includes: image sample set S d and the corresponding weighted parameter set Q d , S d =(Sa d , f d ), the Sa d is the image sample, f d is the label corresponding to the image sample, the subscript d is the sequence number of the image sample set, and the model parameters are given weights. According to the given weight parameter set Q d For the image sample set S d Assign weights, where W d is the corresponding image sample set S d Weight value, mean() is the mean function, (Sa d , f d ) is performed by using the corresponding label f in the image sample d Participate in the operation, note f d If it is Yang, then f d =1, note f d If it is Yin, then f d =0; Perform weighted averaging on the model parameters, integrate the image sample weights into a data stream, and construct a sequence data, data = [data1, data2..., data n ], n is the total number of image samples, weighted mean L, data i is the value of the i-th element in the sequence data, max(data i ) is the maximum value in the sequence data, min(data i ) is the minimum value in the sequence data, exp() is the exponential function, and the weighted mean L is set as the lowest weight during model training. The weighted mean L and the model parameters are aggregated into global data, and the global data is trained through a convolutional neural network, and the global data is output as a global model.
4. The AI-based severe pneumonia diagnosis method according to claim 1, characterized in that: Through the network loss function, the impact of the heterogeneity of multi-source hospital pneumonia data after the server and each client use federated learning for model training is solved. By introducing a conditional distribution mismatch penalty in the objective function of client training, the effect of feature calibration is achieved. By introducing transfer learning, the feature information of the synthetic data is calibrated with private data.
5. The AI-based severe pneumonia diagnosis method according to claim 4, characterized in that: The aggregated global parameters are trained on the local model classification, the global model is output, and it is determined whether the predicted value of the global model meets the standard. The model parameters are defined as C d , create a blank array Rs, denote Rsij as the element in the i-th row and j-th column of the array Rs, and denote the value of Rsij as the RSSI value between the i-th node and the j-th node in the model, where i, j = 1, 2, ..., N, and when i = j, Rsij = 0; create an array Ra, denote Rak as the k-th element in the array Ra, and denote the value of Rak as the mean of the elements in each column of Rs, k = 1, 2, ..., N, and obtain the signal coordination array C by the following formula: Among them, Rsk1 is the k1th column element in Rs, k1=1,2,…,N, and the data in the array Rs and the array Ra are all model parameters C d ; make p, q = 1, 2, ..., N, where Cpq is the element in the p-th row and q-th column of the array Rs, and Cpp is the element in the p-th row and p-th column of the array Ra. The value range of p and q is (1, N). Create the model training array Ass: Ass array is a pq The permutation array of Ass is recorded as Ar and Ac, and the row and column where the element with the smallest value in Ass is located are stored. The elements of other columns in Ass that intersect with the Arth row and the Acth row in turn are stored. Where i1 is an integer and i1∈[1,N-2]. The other columns in Ass are the columns of Ass excluding the Arth column and the Acth column. Traverse i1 within the value range of i1, update Ass to the array obtained by removing the Arth row, the Acth row, the Arth column, and the Acth column from Ass, and create an array Aave, recorded as Aave j1 is the j1th element in the array Aave, and Aave is recorded in sequence j1 The value of is the mean of each column element in Ass, j1=1,2,…,N-2, and the array Aave is normalized to obtain the array Aave n , record the node security value Aave m For Aave T , the Aave T For array Aave n The transpose of the model parameter C in the model d The corresponding indicator is greater than Aave m is marked as an important weight, and the model parameter C d The corresponding indicator is greater than Aave m The mark is a secondary weight. When the number of important weights is less than the number of secondary weights, mark the current model parameter C d The weight values of the image samples in the training are relatively low. When the number of important weights is greater than the number of secondary weights, the current model prediction value is marked as high, and the prediction model is output. When the current model prediction value is high, the data is encrypted.