KR102643869B1 - Pathology diagnosis apparatus using federated learning model and its processing method - Google Patents

Abstract

The present invention relates to a pathology diagnosis system using a federated learning model and a processing method thereof, which trains a deep learning model using pathology data according to a federated learning model and transmits model parameters corresponding to the learned deep learning model. Multiple local clients; And a federated learning server that builds a standard model by reflecting the learning weights through the transmitted model parameters. By including this, not only can leakage of personal information be prevented, but also a federated learning technique that shares only the model learning weights is applied. By doing so, a standardized model with high generalization performance can be built, thereby improving pathology diagnosis performance.

Description

Pathology diagnosis system and processing method using federated learning model {PATHOLOGY DIAGNOSIS APPARATUS USING FEDERATED LEARNING MODEL AND ITS PROCESSING METHOD}

The present invention trains a deep learning model using pathology data according to a federated learning model in a plurality of local clients, then transmits model parameters corresponding to the learned deep learning model, and sets learning weights through the model parameters in the federated learning server. By building a standard model that reflects the It is about a pathology diagnosis system using a federated learning model and its processing method.

As is well known, medical imaging is used as a very important tool in modern medicine for effective disease diagnosis and patient treatment, and the development of imaging technology allows the acquisition of more sophisticated medical imaging data (i.e., pathology data). In response, the amount of data is becoming increasingly vast, making it difficult to analyze pathology data relying on the perspective of a clinical pathologist.

In order to solve the above-mentioned problems, a large amount of pathological data is acquired through various medical devices (e.g., ultrasound, CT (computed tomography), MRI (magnetic resonance imaging), etc.), and artificial intelligence is used to obtain a large amount of pathological data. Pathology diagnosis model creation technology is also being researched to improve pathology diagnosis performance by analyzing and learning.

Meanwhile, the concept of user-centered artificial intelligence is being presented as a next-generation artificial intelligence model that protects user privacy and creates new transactions, enabling collaboration that maximizes user information protection while protecting individual user privacy. An artificial intelligence service that not only achieves the intended results but also achieves the intended results is called a user-centric artificial intelligence service.

This user-centered artificial intelligence not only protects user privacy while using user information to a minimum, but also helps businesses collaborate safely without sharing or integrating user personal information held by companies.

In order to provide user-centered artificial intelligence services as described above, many medical institutions are actively researching CDM (common data model), a data system with a unified structure that can be shared by each institution, depending on medical purposes. Various technological developments are underway to efficiently utilize data from multiple hospitals.

However, in order to effectively secure pathology data, learn the pathology diagnosis model, and improve the performance of pathology diagnosis using the pathology diagnosis model, multiple hospitals must be able to share and analyze the pathology data and learn the pathology diagnosis model. In the case of pathology data provided, there is a problem in that learning and diagnosis performance is deteriorated due to differences in data quality by institution.

In addition, there is a problem in directly sharing medical data (i.e. pathology data) due to the leakage of medical information by institution, and accordingly, there is a need to improve generalization performance for learning deep learning models that require large amounts of learning data. There is a problem in that it is difficult to secure pathological data.

1. Korean Patent Publication No. 10-2021-0098048 (published on August 10, 2021)

The present invention trains a deep learning model using pathology data according to a federated learning model in a plurality of local clients, then transmits model parameters corresponding to the learned deep learning model, and sets learning weights through the model parameters in the federated learning server. By building a standard model that reflects the We aim to provide a pathology diagnosis system and processing method using a federated learning model that can

The purposes of the embodiments of the present invention are not limited to the purposes mentioned above, and other purposes not mentioned will be clearly understood by those skilled in the art from the description below. .

According to one aspect of the present invention, a plurality of local clients train a deep learning model using pathology data according to a federated learning model and transmit model parameters corresponding to the learned deep learning model; A pathology diagnosis system using a federated learning model including a federated learning server that builds a standard model by reflecting learning weights through the transmitted model parameters can be provided.

Additionally, according to one aspect of the present invention, the federated learning server includes a second communication module unit that receives the model parameters transmitted from the plurality of local clients; And a standard model building unit that builds the standard model by reflecting the learning weights for building the standard model through the model parameters. A pathology diagnosis system using a federated learning model including a.

In addition, according to one aspect of the present invention, the federated learning model is a local model that transmits a plurality of model parameters corresponding to the results of learning and pathology diagnosis of a plurality of deep learning models in the plurality of local clients ( A pathology diagnosis system using a federated learning model that applies a local model can be provided.

In addition, according to one aspect of the present invention, the federated learning model sets one model parameter corresponding to the result of learning one deep learning model and performing pathology diagnosis using the entire pathology data of the plurality of local clients. A pathology diagnosis system using a federated learning model that applies a data-shard global model can be provided.

In addition, according to one aspect of the present invention, the federated learning model learns the average value of a plurality of model parameters corresponding to the results of learning and performing pathology diagnosis on a plurality of deep learning models in the plurality of local clients. A pathology diagnosis system using a federated learning model that applies a monolithic average model that is transmitted to be set as a weight can be provided.

In addition, according to one aspect of the present invention, the federated learning model performs learning and pathology diagnosis of a plurality of deep learning models in the plurality of local clients, and applies the weight of the deep learning model set in a pair to the deep learning model of the next pair. A pathology diagnosis system using a federated learning model that applies an Iterative Parallel Average Model that transmits a plurality of final model parameters corresponding to the results of repeating the learning cycle applied to the learning model can be provided.

According to another aspect of the present invention, training a deep learning model using pathology data according to a federated learning model in a plurality of local clients; Transmitting model parameters corresponding to the learned deep learning model from a plurality of local clients; And a step of constructing a standard model by reflecting the learning weights through the model parameters transmitted from the federated learning server. A processing method of a pathology diagnosis system using a federated learning model including a step can be provided.

In addition, according to another aspect of the present invention, the step of constructing the standard model includes receiving the model parameters transmitted from the plurality of local clients in a second communication module unit provided in the federated learning server; And a step of constructing the standard model by reflecting the learning weights for building the standard model through the model parameters in a standard model building unit provided in the federated learning server. Pathology diagnosis system using a federated learning model comprising a. A processing method may be provided.

In addition, according to another aspect of the present invention, the federated learning model is a local model that transmits a plurality of model parameters corresponding to the results of learning and pathology diagnosis of a plurality of deep learning models in the plurality of local clients ( A processing method for a pathology diagnosis system using a federated learning model that applies a local model can be provided.

In addition, according to another aspect of the present invention, the federated learning model sets one model parameter corresponding to the result of learning one deep learning model and performing pathology diagnosis using the entire pathology data of the plurality of local clients. A processing method for a pathology diagnosis system using a federated learning model that applies a data-shard global model can be provided.

In addition, according to another aspect of the present invention, the federated learning model learns the average value of a plurality of model parameters respectively corresponding to the results of learning and performing pathology diagnosis on a plurality of deep learning models in the plurality of local clients. A processing method for a pathology diagnosis system using a federated learning model that applies an average model (Monolithic Average Model) that is transmitted to be set as a weight can be provided.

In addition, according to another aspect of the present invention, the federated learning model performs learning and pathology diagnosis of a plurality of deep learning models in the plurality of local clients, and applies the weight of the deep learning model set in a pair to the deep learning model of the next pair. A processing method for a pathology diagnosis system using a federated learning model that applies an Iterative Parallel Average Model that transmits a plurality of final model parameters corresponding to the results of repeating the learning cycle applied to the learning model can be provided. .

The present invention trains a deep learning model using pathology data according to a federated learning model in a plurality of local clients, then transmits model parameters corresponding to the learned deep learning model, and sets learning weights through the model parameters in the federated learning server. By building a standard model that reflects the You can do it.

1 to 3 are block diagrams of a pathology diagnosis system using a federated learning model according to an embodiment of the present invention;
4 to 10 are diagrams for explaining the detailed configuration of a pathology diagnosis system using a federated learning model according to an embodiment of the present invention;
Figures 11 and 12 are flow charts showing the processing process of building a standard model in a pathology diagnosis system using a federated learning model according to another embodiment of the present invention.

Advantages and features of the embodiments of the present invention and methods for achieving them will become clear by referring to the embodiments described in detail below along with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below and may be implemented in various different forms. The present embodiments are merely provided to ensure that the disclosure of the present invention is complete and to be understood by those skilled in the art. It is provided to fully inform those who have the scope of the invention, and the present invention is only defined by the scope of the claims. Like reference numerals refer to like elements throughout the specification.

In describing embodiments of the present invention, if a detailed description of a known function or configuration is judged to unnecessarily obscure the gist of the present invention, the detailed description will be omitted. The terms described below are terms defined in consideration of functions in the embodiments of the present invention, and may vary depending on the intention or custom of the user or operator. Therefore, the definition should be made based on the contents throughout this specification.

Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings.

1 to 3 are block diagrams of a pathology diagnosis system using a federated learning model according to an embodiment of the present invention, and Figures 4 to 10 are a pathology diagnosis system using a federated learning model according to an embodiment of the present invention. This is a drawing to explain the detailed configuration of.

1 to 10, the pathology diagnosis system using a federated learning model according to an embodiment of the present invention may include a plurality of local clients (100/1-100/N), a federated learning server 200, etc. You can.

Multiple local clients (100/1-100/N) learn a deep learning model using pathology data according to a federated learning model and transmit model parameters corresponding to the learned deep learning model, and each of them learns the model. It may include a unit 110, a pathology diagnosis unit 120, a first communication module unit 130, etc.

Here, the model learning unit 110 is a component that learns a deep learning model using pathology data collected for each local client. The deep learning model is a CNN (Convolutional Neural Network) and GAN that have excellent performance in learning and identifying images. At least one selected from (Generative Adversarial Network), AutoEncorder, etc. can be applied.

This model learning unit 110 is, for example, a local model (Local Model), a global model (Data-Shard Data-Shard Global Model), an average model (Monolithic Average Model), and an iterative average model (Iterative Parallel Average Model). A deep learning model can be trained according to the federated learning model that applies any of the selected ones.

The pathology diagnosis unit 120 is a component that performs pathology diagnosis by inputting and outputting pathological diagnosis data to a deep learning model when pathological diagnosis data to be diagnosed is input. It identifies and detects the pathology area of interest when pathological diagnosis data to be diagnosed is input. After inputting the corresponding pathology diagnosis data into the deep learning model learned through the model learning unit 110, the pathology diagnosis can be effectively performed as an output result.

The first communication module unit 130 is a component that transmits model parameters corresponding to the deep learning model according to the federated learning model to the federated learning server 200, and can provide a communication environment for a communication method corresponding to the communication network described later. It may include a communication module, which can extract model parameters of the deep learning model used for pathology diagnosis through the pathology diagnosis unit 120 after being learned in the model learning unit 110 and transmit them to the federated learning server 200. there is.

Here, model parameters may include, for example, the number of learning data, learning data ratio, learning data resolution, weights, biases, etc. The number of learning data is the number used to train the deep learning model. It refers to the number of data, and the learning data ratio refers to the ratio representing the number of learning data used for learning in response to the total number of learning data. The learning data resolution refers to the resolution of the learning data (i.e., pathology image), and the weight refers to the weight of each neuron that processes pathology data in a deep learning model, and bias refers to a parameter that is added to the input sum of neurons in a deep learning model to adjust the output.

Model parameters as described above can be selectively extracted and transmitted according to the federated learning model.

The federated learning model as described above may apply, for example, any one selected from a local model, a global model, an average model, and an iterative average model. Hereinafter, for a clear explanation of the federated learning model, it will be described assuming that five local clients (100/1-100/N) are provided (that is, N is 5).

As shown in Figure 5, the local model is a model that transmits a plurality of model parameters corresponding to the results of learning and performing pathology diagnosis on a plurality of deep learning models in a plurality of local clients (100/1-100/N). As such, deep learning model A, deep learning model B, deep learning model C, deep learning model D, and deep learning model E can each start learning according to the initial model parameters provided from the federated learning server 200.

Here, in the case of the local model, local client A (100/a) trains deep learning model A using pathology data A collected from local client A (100/a), and then uses the learned deep learning model A. Pathology diagnosis can be performed, and local client B (100/b) trains deep learning model B using pathology data B collected from local client B (100/b), and then learns deep learning model B. Pathology diagnosis can be performed using, and local client C (100/c) trains deep learning model C using pathology data C collected from local client C (100/c), and then uses the learned deep learning Pathology diagnosis can be performed using model C, and local client D (100/d) trains deep learning model D using pathology data D collected from local client D (100/d), and then uses the learned Pathology diagnosis can be performed using deep learning model D, and local client E (100/e) learns deep learning model E using pathology data E collected from local client E (100/e). Pathological diagnosis can be performed using the learned deep learning model E.

In addition, local client A (100/a) can transmit model parameter A corresponding to the learned deep learning model A to the federated learning server 200 through the first communication module unit 160, and local client B (100 /b) can transmit model parameter B corresponding to the learned deep learning model B to the federated learning server 200 through the first communication module unit 160, and local client C (100/c) can transmit the learned deep learning model B to the federated learning server 200. Model parameters C corresponding to learning model C can be transmitted to the federated learning server 200 through the first communication module unit 160, and local client D (100/d) is a model corresponding to the learned deep learning model D. Parameter D can be transmitted to the federated learning server 200 through the first communication module unit 160, and the local client E (100/e) sends the model parameter E corresponding to the learned deep learning model E to the first communication module. It can be transmitted to the federated learning server 200 through the unit 160.

As shown in Figure 6, the global model is a model that corresponds to the results of learning and performing pathology diagnosis on a deep learning model using the entire pathology data of multiple local clients (100/1-100/N). As a model that transmits parameters, deep learning model 1 can start learning according to the initial model parameters provided from the federated learning server 200.

Here, in the case of the global model, in one local client, pathology data A collected from local client A (100/a), pathology data B collected from local client B (100/b), and local client C (100/b) Learn deep learning model 1 using pathology data C collected from /c), pathology data D collected from local client D (100/d), and pathology data E collected from local client E (100/e). After that, pathological diagnosis can be performed using the learned deep learning model 1.

And, any one local client can transmit model parameter 1 corresponding to the learned deep learning model 1 to the federated learning server 200 through the first communication module unit 160.

As shown in FIG. 7, the average model is the average value of a plurality of model parameters corresponding to the results of learning and performing pathology diagnosis on a plurality of deep learning models on a plurality of local clients (100/1-100/N). As a model transmitted to be set as a learning weight, deep learning model A, deep learning model B, deep learning model C, deep learning model D, and deep learning model E are each based on the initial model parameters provided from the federated learning server 200. You can start learning each.

Here, in the case of the average model, similar to the local model, local client A (100/a) trains deep learning model A using pathology data A collected from local client A (100/a), and then uses the learned deep Pathology diagnosis can be performed using learning model A, and local client B (100/b) trains deep learning model B using pathology data B collected from local client B (100/b). Pathology diagnosis can be performed using deep learning model B, and local client C (100/c) learns deep learning model C using pathology data C collected from local client C (100/c). , Pathology diagnosis can be performed using the learned deep learning model C, and local client D (100/d) learns deep learning model D using pathology data D collected from local client D (100/d). After doing so, pathology diagnosis can be performed using the learned deep learning model D, and local client E (100/e) uses the pathology data E collected from local client E (100/e) to perform deep learning model E. After learning, pathology diagnosis can be performed using the learned deep learning model E.

In addition, local client A (100/a) can transmit model parameter A corresponding to the learned deep learning model A to the federated learning server 200 through the first communication module unit 160, and local client B (100 /b) can transmit model parameter B corresponding to the learned deep learning model B to the federated learning server 200 through the first communication module unit 160, and local client C (100/c) can transmit the learned deep learning model B to the federated learning server 200. Model parameters C corresponding to learning model C can be transmitted to the federated learning server 200 through the first communication module unit 160, and local client D (100/d) is a model corresponding to the learned deep learning model D. Parameter D can be transmitted to the federated learning server 200 through the first communication module unit 160, and the local client E (100/e) sends the model parameter E corresponding to the learned deep learning model E to the first communication module. It can be transmitted to the federated learning server 200 through the unit 160.

And, as shown in FIG. 8, the repetition average model learns and diagnoses pathology with a plurality of deep learning models on a plurality of local clients (100/1-100/N), and the weights of the deep learning models set in pairs A model that transmits a plurality of final model parameters corresponding to the results of repeating the learning cycle applied to the next pair of deep learning models, deep learning model A, deep learning model B, deep learning model C, deep learning model D and Deep learning model E can each start learning according to the initial model parameters provided from the federated learning server 200.

Here, in the case of the repetition average model, local client A (100/a) divides the pathology data A collected from local client A (100/a) into, for example, 5 pieces (20% of the total data each), After pairing deep learning model A1 and deep learning model A2, deep learning model A3 and deep learning model A4, deep learning model A1 is learned and pathology diagnosis is performed using pathology data A1, and pathology data A2 is used to learn deep learning model A1 and perform pathology diagnosis. After learning deep learning model A2 and performing pathology diagnosis, the average value of model parameters A1 and model parameter A2 corresponding to the learned deep learning model A1 and deep learning model A2 is set as average model parameter 1 to create deep learning model A3. It can be applied to and deep learning model A4.

Next, using pathology data A3, learn deep learning model A3 with average model parameter 1 applied and perform pathology diagnosis, and use pathology data A4 to learn deep learning model A4 with average model parameter 1 applied and perform pathology diagnosis. After execution, the average value of model parameters A3 and model parameters A4 corresponding to the learned deep learning model A3 and deep learning model A4 can be set as average model parameter 2 and applied to deep learning model A5.

In addition, after learning deep learning model A5 to which average model parameter 2 is applied and performing pathology diagnosis using pathology data A5, model parameter A5 corresponding to deep learning model A5 is set as the final model parameter A, and the first communication module unit It can be transmitted to the federated learning server (200) through (130).

As described above, local client A (100/a) classifies and trains a plurality of deep learning models to learn each of the pathological data divided according to the data ratio, and sets two deep learning models as a pair, and through this After learning the deep learning model sequentially and diagnosing pathology according to a learning cycle in which the average value of the obtained model parameters is applied to learning the next pair of deep learning models, the final model parameter A is transmitted to the federated learning server 200. You can.

Meanwhile, local client B (100/b), local client C (100/c), local client D (100/d), and local client E (100/e) also follow a similar process as local client A (100/a). By performing this, the final model parameter B, final model parameter C, final model parameter D, and final model parameter E can be transmitted to the federated learning server 200 through the first communication module unit 130, respectively.

The federated learning server 200 is a server that builds a standard model by reflecting learning weights through model parameters transmitted from a plurality of local clients (100/1-100/N). The second communication module unit 210, standard It may include a model construction unit 220, etc.

Here, the second communication module unit 210 is a component that receives model parameters transmitted from a plurality of local clients, and may include a communication module that can provide a communication environment for a communication method corresponding to a communication network to be described later. .

The standard model building unit 220 is a component that builds a standard model by reflecting the learning weights for building the standard model through model parameters transmitted from the second communication module unit 210, and has different learning weights depending on the federated learning model. A standard model can be built by reflecting this.

This standard model building unit 220 can reflect learning weights differently depending on the federated learning model. When a local model among the federated learning models is applied, the data transmitted from a plurality of local clients (100/1-100/N) A plurality of model parameters may be transmitted to the standard model building unit 220 through the second communication module unit 210.

For example, the standard model building unit 220 includes model parameter A transmitted from local client A (100/a), model parameter B transmitted from local client B (100/b), and local client C in the local model. Learning by adding model parameter C transmitted from (100/c), model parameter D transmitted from local client D (100/d), and model parameter E transmitted from local client E (100/e). After setting the weights, you can build a standard model by reflecting the set learning weights in the deep learning model.

In addition, when the global model among the federated learning models is applied, one model parameter transmitted from a plurality of local clients (100/1-100/N) is sent to the standard model building unit 220 through the second communication module unit 210. The standard model building unit 220 sets model parameter 1 transmitted from any one local client as a learning weight and then reflects the set learning weight to the deep learning model to build a standard model.

When the average model among these federated learning models is applied, the average value of a plurality of model parameters from a plurality of local clients (100/1-100/N) is transmitted to the standard model construction unit 220 through the second communication module unit 210. The standard model building unit 220 includes model parameter A transmitted from local client A (100/a), model parameter B transmitted from local client B (100/b), and local client C (100/c). Learning by calculating the average value after adding up the model parameter C transmitted from , model parameter D transmitted from local client D (100/d), and model parameter E transmitted from local client E (100/e). After setting the weights, you can build a standard model by reflecting the set learning weights in the deep learning model.

When the repetition average model among these federated learning models is applied, a plurality of final model parameters corresponding to the results of repeating the learning cycle from a plurality of local clients (100/1-100/N) are transmitted through the second communication module unit 210. The standard model building unit 220 transmits final model parameter A transmitted from local client A (100/a), final model parameter B transmitted from local client B (100/b), and local client C (100/b). The average value after adding up the final model parameter C transmitted from /c), the final model parameter D transmitted from local client D (100/d), and the final model parameter E transmitted from local client E (100/e). After setting the learning weight by calculating , a standard model can be built by reflecting the set learning weight in the deep learning model.

A communication network for providing a communication environment between a plurality of local clients (100/1-100/N) and the federated learning server (200) in the pathology diagnosis system using the federated learning model according to an embodiment of the present invention as described above. This can be provided.

These communication networks may include, for example, a mobile communication network, a wired Internet communication network, a wireless Internet communication network, a broadcasting network, an administrative network, etc., and specifically include personal area network (PAN), local area network (LAN), and CAN ( It may include one or more networks among networks such as campus area network), MAN (metropolitan area network), WAN (wide area network), BBN (broadband network), Internet, LoRaWAN, bus network, star network, ring network, It may include one or more of network topologies including mesh networks, star-bus networks, tree networks, hierarchical networks, etc.

Performance experiments performed according to the federated learning model in the pathology diagnosis system using the federated learning model according to an embodiment of the present invention as described above and the evaluation results thereof will be described.

First, the data used was pathology data provided by the hospital pathology department, and in learning the deep learning model, the segmentation model was tested using the structure shown in Figure 9, and the evaluation indicators were loss rate and pixel accuracy. (Pixel Accuracy) was used.

As a result, the central model (Data-Shard Data-Shard Global Model), local model (Local Model), FedAvg (Monolithic Average Model), and Iterative parallel average (IPA) (Iterative Parallel Average) It can be seen that the loss rate and pixel accuracy for the model are shown in Figure 10.

Through these evaluation results, it can be seen that the performance of the FedAvg model, which is a simple combination of local information, has improved compared to the central model, and the Iterative parallel average (IPA) model has also improved performance compared to the central model, which results in federated learning It can be seen that as the model is applied, generality is secured and model performance is improved.

Therefore, according to an embodiment of the present invention, after learning a deep learning model using pathology data according to a federated learning model in a plurality of local clients, model parameters corresponding to the learned deep learning model are transmitted, and federated learning is performed. By building a standard model by reflecting the learning weights through model parameters on the server, not only can leakage of personal information be prevented, but also a federated learning technique that shares only the model learning weights is applied to build a standardized model with high generalization performance. This can improve pathology diagnosis performance.

Figures 11 and 12 are flow charts showing the processing process of building a standard model in a pathology diagnosis system using a federated learning model according to another embodiment of the present invention.

Referring to Figures 11 and 12, a deep learning model can be trained using pathology data according to a federated learning model in a plurality of local clients (100/1-100/N) (step 310).

This deep learning model can, for example, apply at least one selected from CNN, GAN, autoencoder, etc., which have excellent performance in learning and identifying images.

In addition, for learning and pathology diagnosis of a deep learning model using pathology data, a plurality of local clients (100/1-100/N) are equipped with a model learning unit 110, a pathology diagnosis unit 120, and a first communication module unit, respectively. It may include (130), etc., and the detailed description thereof will be omitted since it is described in an embodiment of the present invention.

The deep learning model as described above is performed by applying a federated learning model, and after learning according to the local model, global model, average model, and iterative average model, pathology diagnosis, verification, etc. can be performed.

Here, the local model is a model that transmits a plurality of model parameters corresponding to the results of learning and performing pathology diagnosis on a plurality of deep learning models in a plurality of local clients (100/1-100/N), and the global model is a model that transmits one model parameter corresponding to the results of learning one deep learning model and performing pathology diagnosis using the entire pathology data of multiple local clients (100/1-100/N).

In addition, the average model is a model that transmits multiple model parameters corresponding to the results of learning and performing pathology diagnosis on multiple deep learning models on multiple local clients (100/1-100/N), and the repetition average The model learns and performs pathology diagnosis on multiple deep learning models on multiple local clients (100/1-100/N), and applies the weights of the deep learning models set in pairs to the deep learning models of the next pair. This is a model that transmits multiple final model parameters corresponding to the results of repeating the cycle.

Next, model parameters corresponding to the learned deep learning model can be transmitted from a plurality of local clients (100/1-100/N) (step 320).

Next, a standard model can be built by reflecting the learning weights through model parameters transmitted from the federated learning server 200 (step 330).

In the step 330 of building the standard model, the second communication module unit 210 provided in the federated learning server 200 receives model parameters transmitted from a plurality of local clients (100/1-100/N). It includes a step (331) of constructing a standard model by reflecting the learning weights for building a standard model through model parameters in the standard model building unit (220) provided in the federated learning server (200) (332). can do.

For example, in the local model among federated learning models, model parameter A transmitted from local client A (100/a), model parameter B transmitted from local client B (100/b), and local client C (100/c) ), model parameter C transmitted from local client D (100/d), and model parameter E transmitted from local client E (100/e) are set to set the learning weight. Later, a standard model can be built by reflecting the set learning weights in the deep learning model.

In addition, in the global model among federated learning models, model parameter 1 transmitted from any one local client is set as a learning weight, and then the set learning weight is reflected in the deep learning model to build a standard model.

Among these federated learning models, in the average model, model parameter A is transmitted from local client A (100/a), model parameter B is transmitted from local client B (100/b), and model parameter B is transmitted from local client C (100/c) The learning weight is calculated by adding up the model parameter C, the model parameter D transmitted from local client D (100/d), and the model parameter E transmitted from local client E (100/e), and then calculating the average value. After setting, you can build a standard model by reflecting the set learning weights in the deep learning model.

Among these federated learning models, in the iterative average model, final model parameter A transmitted from local client A (100/a), final model parameter B transmitted from local client B (100/b), and local client C (100/c ), final model parameter C transmitted from local client D (100/d), final model parameter E transmitted from local client E (100/e), and calculate the average value. After setting the learning weights in this way, a standard model can be built by reflecting the set learning weights in the deep learning model.

Therefore, according to another embodiment of the present invention, after learning a deep learning model using pathology data according to a federated learning model in a plurality of local clients, model parameters corresponding to the learned deep learning model are transmitted, and federated learning is performed. By building a standard model by reflecting the learning weights through model parameters on the server, not only can leakage of personal information be prevented, but also a federated learning technique that shares only the model learning weights is applied to build a standardized model with high generalization performance. This can improve pathology diagnosis performance.

In the above description, various embodiments of the present invention have been presented and explained, but the present invention is not necessarily limited thereto, and those skilled in the art will understand various embodiments without departing from the technical spirit of the present invention. It will be easy to see that branch substitutions, transformations, and changes are possible.

100/1-100/N: Multiple local clients
110: Model learning department
120: Pathological diagnosis department
130: 1st communication module unit
200: Union learning server
210: 2nd communication module unit
220: Standard model construction department

Claims (12)

A plurality of local clients that train a deep learning model using pathology data according to a federated learning model and transmit model parameters corresponding to the learned deep learning model; and
It includes a federated learning server that builds a standard model by reflecting learning weights through the transmitted model parameters,
The federated learning model is,
An iterative average model that transmits a plurality of final model parameters corresponding to the results of repeating the learning cycle of setting two deep learning models as a pair and applying the weight of the deep learning model set as a pair to the next pair of deep learning models ( Apply the Iterative Parallel Average Model,
The plurality of local clients are:
Using a plurality of deep learning models according to the federated learning model, the plurality of deep learning models are converted to the repetition average model according to the initial model parameters provided from the federated learning server to learn each of the pathological data divided according to the data ratio. Accordingly, transmitting the plurality of final model parameters corresponding to each learned result,
The federated learning server,
After setting the learning weights by calculating the average value after adding up the plurality of final model parameters, the set learning weights are reflected in the deep learning model to build the standard model.
Pathology diagnosis system using federated learning model.
In claim 1,
The federated learning server,
a second communication module unit receiving the model parameters transmitted from the plurality of local clients; and
a standard model construction unit that builds the standard model by reflecting the learning weights for building the standard model through the model parameters;
Pathology diagnosis system using a federated learning model including.
delete delete delete delete Learning a deep learning model using pathology data according to a federated learning model in a plurality of local clients;
Transmitting model parameters corresponding to the learned deep learning model from a plurality of local clients; and
It includes the step of building a standard model by reflecting the learning weights through the model parameters transmitted from the federated learning server,
The step of training the deep learning model is,
In the case of the federated learning model, two deep learning models are set as a pair, and a plurality of final model parameters corresponding to the results of repeating the learning cycle of applying the weights of the deep learning model set as a pair to the next pair of deep learning models are set. Apply the Iterative Parallel Average Model to transmit,
The plurality of deep learning models according to the initial model parameters provided from the federated learning server so that the plurality of local clients learn each of the pathological data divided according to the data ratio using a plurality of deep learning models according to the federated learning model. Transmitting the plurality of final model parameters corresponding to the results of each learning according to the repetition average model,
The step of building the standard model is,
After setting the learning weights by calculating the average value after adding up the plurality of final model parameters in the federated learning server, the set learning weights are reflected in the deep learning model to build the standard model.
Processing method of pathology diagnosis system using federated learning model.
In claim 7,
The step of building the standard model is,
Receiving the model parameters transmitted from the plurality of local clients at a second communication module provided in the federated learning server; and
Constructing the standard model by reflecting the learning weights for building the standard model through the model parameters in a standard model building unit provided in the federated learning server;
Processing method of a pathology diagnosis system using a federated learning model including.
delete delete delete delete